diff --git a/.gitattributes b/.gitattributes
index a6344aac8c09253b3b630fb776ae94478aa0275b..a2ffb96ab9e11b0ccad6e4db0e076bb0e0dc0cc6 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
diff --git a/models/Robometer-4B/.gitattributes b/models/Robometer-4B/.gitattributes
new file mode 100644
index 0000000000000000000000000000000000000000..a6344aac8c09253b3b630fb776ae94478aa0275b
--- /dev/null
+++ b/models/Robometer-4B/.gitattributes
@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
diff --git a/models/Robometer-4B/README.md b/models/Robometer-4B/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..56bba2ae87aaf12df16875608debec55fb42771c
--- /dev/null
+++ b/models/Robometer-4B/README.md
@@ -0,0 +1,66 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-VL-4B-Instruct
+tags:
+  - reward model
+  - robot learning
+  - foundation models
+library_name: transformers
+---
+
+# Robometer 4B
+
+**Paper:** [arXiv (Coming Soon)](https://arxiv.org/)
+
+**Robometer** is a general-purpose vision-language reward model for robotics. It is trained on [RBM-1M](https://huggingface.co/datasets/) with **Qwen3-VL-4B** to predict **per-frame progress**, **per-frame success**, and **trajectory preferences** from rollout videos. The model combines (1) frame-level progress supervision on expert data and (2) trajectory-comparison preference supervision, so it can learn from both successful and failed rollouts and generalize across diverse robot embodiments and tasks.
+
+Given a **task instruction** and a **rollout video** (or frame sequence), the model predicts:
+
+- **Per-frame progress** — continuous progress values over time (e.g. 0–1 or binned).
+- **Per-frame success** — success probability (or binary) at each timestep.
+- **Preference / ranking** — which of two trajectories is better for the task.
+
+### Usage
+
+For full setup, example scripts, and configs, see the **GitHub repo**: [github.com/robometer/robometer](https://github.com/robometer/robometer).
+
+**Option 1 — Run the model locally** (loads this checkpoint from Hugging Face):
+
+```bash
+uv run python scripts/example_inference_local.py \
+  --model-path robometer/Robometer-4B \
+  --video /path/to/video.mp4 \
+  --task "your task description"
+```
+
+**Option 2 — Use the evaluation server** (start server, then run client):
+
+```bash
+# Start server
+uv run python robometer/evals/eval_server.py \
+  --config-path=robometer/configs \
+  --config-name=eval_config_server \
+  model_path=robometer/Robometer-4B \
+  server_url=0.0.0.0 \
+  server_port=8000
+
+# Client (no robometer dependency)
+uv run python scripts/example_inference.py \
+  --eval-server-url http://localhost:8000 \
+  --video /path/to/video.mp4 \
+  --task "your task description"
+```
+
+## Citation
+
+If you use this model, please cite:
+
+```bibtex
+@misc{liang2026robometer,
+  title={Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
+  author={Anthony Liang* and Yigit Korkmaz* and Jiahui Zhang and Minyoung Hwang and Abrar Anwar and Sidhant Kaushik and Aditya Shah and Alex S. Huang and Luke Zettlemoyer and Dieter Fox and Yu Xiang and Anqi Li and Andreea Bobu and Abhishek Gupta and Stephen Tu† and Erdem B{\i}y{\i}k† and Jesse Zhang†},
+  year={2026},
+  url={https://github.com/robometer/robometer},
+  note={arXiv coming soon}
+}
+```
diff --git a/models/Robometer-4B/config.json b/models/Robometer-4B/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..33c5aa8c1d3a0bee0164762f9a10d19f6675978b
--- /dev/null
+++ b/models/Robometer-4B/config.json
@@ -0,0 +1,68 @@
+{
+  "architectures": [
+    "RFM"
+  ],
+  "dtype": "bfloat16",
+  "image_token_id": 151655,
+  "model_type": "qwen3_vl",
+  "pad_token_id": 151654,
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "dtype": "bfloat16",
+    "eos_token_id": 151645,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9728,
+    "max_position_embeddings": 262144,
+    "model_type": "qwen3_vl_text",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 36,
+    "num_key_value_heads": 8,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        24,
+        20,
+        20
+      ],
+      "rope_type": "default"
+    },
+    "rope_theta": 5000000,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 151674
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.2",
+  "unsloth_fixed": true,
+  "unsloth_version": "2025.10.11",
+  "video_token_id": 151656,
+  "vision_config": {
+    "deepstack_visual_indexes": [
+      5,
+      11,
+      17
+    ],
+    "depth": 24,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1024,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "model_type": "qwen3_vl",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 2560,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652
+}
diff --git a/models/Robometer-4B/config.yaml b/models/Robometer-4B/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d17e3b626ac795163d498bea33a735296f3cb886
--- /dev/null
+++ b/models/Robometer-4B/config.yaml
@@ -0,0 +1,187 @@
+custom_eval:
+  comparisons_per_task: 5
+  confusion_matrix:
+  - mw
+  custom_eval_random_seed: 42
+  eval_types:
+  - reward_alignment
+  - policy_ranking
+  num_examples_per_quality_pr: 5
+  num_partial_successes: 5
+  pad_frames: true
+  policy_ranking:
+  - rbm-1m-ood
+  policy_ranking_max_tasks: 100
+  quality_preference:
+  - mw
+  reward_alignment:
+  - rbm-1m-id
+  - rbm-1m-ood
+  reward_alignment_max_trajectories: 10
+  subsample_n_frames: null
+  use_frame_steps: true
+data:
+  data_source_weights:
+    metaworld_train: 1.0
+    molmoact_dataset_household: 1.0
+    molmoact_dataset_tabletop: 1.0
+    oxe_droid: 1.0
+    roboarena: 1.0
+  dataloader_num_workers: 8
+  dataloader_persistent_workers: true
+  dataloader_pin_memory: true
+  dataset_preference_ratio: 0.7
+  dataset_success_cutoff_file: robometer/data/dataset_success_cutoff.txt
+  dataset_type: strategy_first
+  eval_datasets:
+  - mw
+  eval_subset_size: null
+  load_embeddings: false
+  max_frames: 8
+  max_frames_after_preprocessing: 64
+  max_success: 1.0
+  max_trajectories: -1
+  min_frames_per_trajectory: 5
+  min_success: 0.5
+  partial_success_threshold: 0.2
+  predict_last_frame_partial_progress: false
+  preference_strategy_ratio:
+  - 1.0
+  - 1.0
+  - 1.0
+  - 1.0
+  progress_discrete_bins: 10
+  progress_loss_type: discrete
+  progress_pred_type: absolute_wrt_total_frames
+  progress_strategy_ratio:
+  - 1.0
+  - 1.0
+  - 1.0
+  - 1.0
+  resized_height: null
+  resized_width: null
+  sample_type_ratio:
+  - 1.0
+  - 0.0
+  - 0.0
+  seed: 42
+  shuffle: true
+  shuffle_progress_frames: false
+  train_datasets:
+  - rbm-1m-id
+  traj_same_source_prob: 0.5
+  use_multi_image: true
+  use_per_frame_progress_token: true
+debug: false
+logging:
+  log_level: debug
+  log_to:
+  - wandb
+  save_best:
+    greater_is_better:
+    - true
+    - true
+    - true
+    - true
+    - true
+    hub_private: false
+    hub_save_every: 1000
+    hub_token: null
+    keep_top_k: 5
+    metric_names:
+    - eval_p_rank/kendall_last_utd_so101_clean_top
+    - eval_p_rank/kendall_last_usc_xarm
+    - eval_p_rank/kendall_last_usc_franka
+    - eval_p_rank/kendall_last_rfm_new_mit_franka_nowrist
+    - eval_p_rank/kendall_last_usc_trossen
+    save_every: 250
+    upload_to_hub: false
+  save_model: true
+  save_processor: true
+  wandb_entity: clvr
+  wandb_mode: null
+  wandb_notes: all run with prog_token per frame, qwen 4b, discrete progress, 10 bins
+  wandb_project: robometer
+loss:
+  predict_last_frame_progress: false
+  progress_discrete_bins: 10
+  progress_loss_type: discrete
+  success_positive_weight: 1.0
+mode: train
+model:
+  average_temporal_patches: true
+  base_model_id: Qwen/Qwen3-VL-4B-Instruct
+  frame_pooling: mean
+  frame_pooling_attn_temperature: 1.0
+  model_type: default
+  peft_vision_encoder: false
+  progress_discrete_bins: 10
+  progress_loss_type: discrete
+  quantization: false
+  rewind: null
+  torch_dtype: bfloat16
+  train_language_model: true
+  train_preference_head: true
+  train_progress_head: true
+  train_success_head: true
+  train_vision_encoder: false
+  trust_remote_code: true
+  use_multi_image: true
+  use_peft: false
+  use_per_frame_progress_token: true
+  use_unsloth: true
+peft:
+  bias: none
+  lora_alpha: 64
+  lora_dropout: 0.05
+  peft_vision_encoder: false
+  r: 32
+  target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - gate_proj
+  - up_proj
+  - down_proj
+trainer_cls: rbm_heads
+training:
+  beta: 0.1
+  bf16: true
+  custom_eval_steps: 250
+  dataloader_num_workers: 8
+  dataloader_persistent_workers: true
+  dataloader_pin_memory: true
+  ddp_bucket_cap_mb: 25
+  ddp_find_unused_parameters: false
+  do_eval: true
+  eval_steps: 250
+  evaluation_strategy: steps
+  exp_name: ant_rfm_qwen4b_4gpu_bs16_pref_prog_succ_8_frames_all_discrete_10_bins_part2
+  fp16: false
+  gradient_accumulation_steps: 1
+  gradient_checkpointing: true
+  learning_rate: 2.0e-05
+  logging_steps: 1
+  lr_scheduler_type: cosine
+  max_grad_norm: 10.0
+  max_seq_length: 1024
+  max_steps: 15000
+  num_gpus: 2
+  num_train_epochs: -1
+  output_dir: ./logs
+  overwrite_output_dir: true
+  per_device_eval_batch_size: 16
+  per_device_train_batch_size: 16
+  predict_pref_progress: true
+  prediction_loss_only: true
+  remove_unused_columns: false
+  resume_from_checkpoint: /gpfs/home/jessezha/scrubbed_storage/reward_fm/logs/ant_rfm_qwen4b_4gpu_bs16_pref_prog_succ_8_frames_all_discrete_10_bins/ckpt-avg-5metrics=0.6973_step=3000
+  run_default_eval: false
+  save_steps: 200
+  save_strategy: 'no'
+  vision_encoder_lr: 5.0e-06
+  vision_encoder_num_layers: 3
+  warmup_ratio: 0.1
+  warmup_steps: 0
+  weight_decay: 0.01
diff --git a/models/Robometer-4B/dataset_random_state.json b/models/Robometer-4B/dataset_random_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..9b4d82bc375c4846a8ddbc0c1d77021f8be68276
--- /dev/null
+++ b/models/Robometer-4B/dataset_random_state.json
@@ -0,0 +1,1266 @@
+{
+  "dataset": [
+    3,
+    [
+      2147483648,
+      3564348608,
+      1266698288,
+      4212342371,
+      3595291661,
+      3180588708,
+      3037210256,
+      946923017,
+      2565409715,
+      2900535780,
+      924383152,
+      4180157270,
+      4230508198,
+      2039675917,
+      3755350407,
+      2362848650,
+      2818100609,
+      2097423432,
+      524478045,
+      540883378,
+      281170210,
+      1485176884,
+      1493190386,
+      1773214509,
+      380915208,
+      3667698522,
+      2648371337,
+      2961234806,
+      3857480267,
+      1582950522,
+      246289694,
+      3322185604,
+      1944574775,
+      302623699,
+      169865066,
+      1143540808,
+      3733177770,
+      513116636,
+      1411153081,
+      3205493053,
+      768926902,
+      549624109,
+      1470655403,
+      59539609,
+      3678480009,
+      3087139671,
+      1176835859,
+      2078491503,
+      2299934332,
+      1592059249,
+      1062716176,
+      2654193596,
+      3531838733,
+      2661260596,
+      3881209635,
+      2106865768,
+      4154287292,
+      2082185616,
+      2301197011,
+      2177349827,
+      3082181756,
+      1787663536,
+      3714670796,
+      3018262113,
+      1670056238,
+      1856738750,
+      99824592,
+      2279837081,
+      1414647942,
+      3416675731,
+      3458782472,
+      3997022236,
+      468762002,
+      2666158583,
+      953353270,
+      1788980658,
+      3802061067,
+      407586584,
+      1844776834,
+      1906917274,
+      3154715663,
+      3028370222,
+      4156024188,
+      3996363428,
+      80495456,
+      2659800972,
+      2005649973,
+      3818358673,
+      3952623596,
+      2506862371,
+      3282302532,
+      263923435,
+      3384662671,
+      3292439172,
+      3119957588,
+      1224426111,
+      899864150,
+      215262826,
+      1619647231,
+      3347694949,
+      3497868538,
+      2029552053,
+      2992804824,
+      4080010250,
+      2023513186,
+      1885979437,
+      3564622190,
+      3775424270,
+      2297810139,
+      3549449169,
+      2664856277,
+      3274801974,
+      2794883969,
+      980412666,
+      2980215653,
+      2794389321,
+      2816521934,
+      1266970739,
+      542306338,
+      3646225311,
+      3598997630,
+      2111980720,
+      2949252482,
+      2489027658,
+      352815024,
+      11610683,
+      1386663624,
+      2004196796,
+      1161461546,
+      1921293780,
+      2463949525,
+      1647009713,
+      3550093655,
+      2563894064,
+      3486310554,
+      1506105865,
+      243092931,
+      2659437476,
+      4200687059,
+      2284345122,
+      1974438610,
+      3591096528,
+      967119212,
+      3362401375,
+      140678365,
+      311602112,
+      2361740275,
+      2139598582,
+      3632873481,
+      2762232439,
+      4156482318,
+      381637792,
+      3253346525,
+      2492118775,
+      1502434558,
+      3164497290,
+      3550998357,
+      2412448305,
+      2223955385,
+      4122879535,
+      350121793,
+      1835149778,
+      2175117867,
+      989674750,
+      3178241202,
+      3553093569,
+      3470650311,
+      2829698151,
+      3209427769,
+      1779174943,
+      275388428,
+      4044574515,
+      715447260,
+      3180940440,
+      4020772289,
+      1322708567,
+      3189868792,
+      4250485633,
+      716970023,
+      2307550151,
+      1074996711,
+      1217573599,
+      197006094,
+      2178394212,
+      1255233746,
+      4164251484,
+      1405608772,
+      2808160475,
+      1304736088,
+      1796071066,
+      2761748078,
+      3570739698,
+      1616118556,
+      2232868135,
+      3567541936,
+      3470600401,
+      3031621994,
+      3351764214,
+      1359785149,
+      2617497797,
+      3340028190,
+      356162828,
+      2083806068,
+      2503635608,
+      4024838996,
+      2577080371,
+      2897993505,
+      3120733934,
+      905794891,
+      2506078507,
+      4211618666,
+      3777871979,
+      809751414,
+      4080874167,
+      1562977008,
+      3917373055,
+      2132779194,
+      4014249473,
+      4067327082,
+      2582869847,
+      1780081876,
+      1842619106,
+      3381761227,
+      921004274,
+      1393256920,
+      1883566732,
+      2702071861,
+      865327389,
+      1622085203,
+      3021825820,
+      2687061406,
+      1748902923,
+      689023977,
+      308399650,
+      2377287978,
+      1646969411,
+      1051806316,
+      4277884230,
+      2041056290,
+      101134519,
+      2032472116,
+      4112521069,
+      151202901,
+      2773743461,
+      551348559,
+      3476836808,
+      510935951,
+      625057077,
+      3757450756,
+      2977698135,
+      3027776859,
+      2616998041,
+      2773430005,
+      544190486,
+      2241368212,
+      1141105829,
+      1452816309,
+      4199229235,
+      3218013033,
+      4229475816,
+      1659576351,
+      3020348754,
+      1193400518,
+      3208584597,
+      1151197733,
+      2597187966,
+      503065140,
+      2421841572,
+      1437291709,
+      1909275895,
+      2872630545,
+      793588217,
+      3792934707,
+      1784451785,
+      2921385648,
+      1669902526,
+      4189978976,
+      1196986251,
+      434805516,
+      1907541826,
+      2624415034,
+      1687778718,
+      650746582,
+      1949153382,
+      4148493093,
+      841300520,
+      1164202054,
+      4203468658,
+      4106300911,
+      850346789,
+      1715730760,
+      3114661489,
+      2866524548,
+      1360448945,
+      3601318775,
+      1743078223,
+      2413855408,
+      1211895622,
+      325117146,
+      2721152875,
+      1284334485,
+      2446538832,
+      739014618,
+      2237045115,
+      842553465,
+      2538598293,
+      746460793,
+      4010387366,
+      2002655192,
+      4193733112,
+      1194380773,
+      3918217378,
+      1447487475,
+      5659228,
+      3408847694,
+      4190318700,
+      1862549564,
+      781683719,
+      1194618118,
+      755053413,
+      3436011942,
+      2885435303,
+      3081151348,
+      2017642831,
+      1053816502,
+      1086627485,
+      2157296554,
+      110650022,
+      965352898,
+      1003174194,
+      1288956241,
+      4057404871,
+      2965068465,
+      2897064481,
+      2457377317,
+      1879872545,
+      358455290,
+      375086701,
+      3015902095,
+      1676249984,
+      924455526,
+      2084169389,
+      1989014644,
+      1993749926,
+      2009424973,
+      2113340508,
+      3980883273,
+      2915977458,
+      203328382,
+      3020815229,
+      2415050113,
+      4103009585,
+      3700885489,
+      2916647550,
+      1523006503,
+      174302338,
+      2476909338,
+      1969322490,
+      4285741984,
+      1528449097,
+      3355315515,
+      4217241278,
+      599579127,
+      2572243673,
+      3035856735,
+      1539140489,
+      1782314913,
+      4238644287,
+      1746424142,
+      1978148312,
+      2380746849,
+      184941882,
+      1106717981,
+      1720750349,
+      981701307,
+      3953154731,
+      3257809181,
+      2892339376,
+      3339778166,
+      3676936849,
+      87425948,
+      3029257381,
+      2037942523,
+      3807628706,
+      2861474706,
+      1058852346,
+      1322765211,
+      2686046342,
+      2689342655,
+      2303436168,
+      2571627181,
+      1986057734,
+      1183564308,
+      2829677523,
+      1295563975,
+      503126586,
+      2025890348,
+      4179277821,
+      1735262467,
+      981331774,
+      1613447066,
+      1011606109,
+      2000062246,
+      3581448390,
+      3477731384,
+      3641307373,
+      3508544379,
+      2327233491,
+      3931944343,
+      4189052882,
+      2990416380,
+      422406169,
+      202291313,
+      2531006461,
+      4277024116,
+      3815144003,
+      821314585,
+      1344175168,
+      3562834071,
+      1339615445,
+      1831545190,
+      3115548822,
+      743512780,
+      4006999448,
+      3720181735,
+      1012033521,
+      919931041,
+      2628967879,
+      1151876565,
+      1268107129,
+      3674829936,
+      834977846,
+      743987006,
+      3947536548,
+      3706529695,
+      4121073678,
+      2507605742,
+      1595636918,
+      2708047833,
+      2427507331,
+      3868216331,
+      3254240010,
+      2097683411,
+      3279710596,
+      3686819053,
+      1843541720,
+      1683793619,
+      3245287285,
+      3571828776,
+      3733296431,
+      3806747478,
+      1390930605,
+      3860422228,
+      114397037,
+      1931519825,
+      2770684378,
+      1556101783,
+      1436111731,
+      4031950081,
+      562876656,
+      1775895782,
+      612364620,
+      1313509772,
+      4283410242,
+      3252958463,
+      2176555836,
+      3933073367,
+      3013277102,
+      1444071961,
+      3120949516,
+      2824578890,
+      325676929,
+      943677134,
+      1800649256,
+      1721927060,
+      347498719,
+      1435221321,
+      2623572981,
+      1408548470,
+      4145586315,
+      2901889237,
+      1849377952,
+      1239144551,
+      3382598266,
+      2992893897,
+      3738297588,
+      611280106,
+      3897415338,
+      2370299241,
+      1772308583,
+      3697465753,
+      354508058,
+      2702360134,
+      591308331,
+      3524072501,
+      976616000,
+      2563717192,
+      3078266097,
+      1376594703,
+      4209795919,
+      2454412767,
+      2712206031,
+      2963860163,
+      3734324882,
+      2248653800,
+      324872786,
+      3789837448,
+      3779000146,
+      527733939,
+      2844165793,
+      576499681,
+      1618787435,
+      2638888650,
+      57511068,
+      2804627518,
+      2993670030,
+      481402236,
+      2810124845,
+      1416045214,
+      1723694191,
+      1214944572,
+      3188123783,
+      1139185907,
+      3851015362,
+      1719652470,
+      1661343029,
+      3644307578,
+      3564178709,
+      1256656955,
+      46631590,
+      4231317929,
+      3098958589,
+      1834956625,
+      2206185428,
+      3695688374,
+      3647957317,
+      1064098871,
+      1739100906,
+      2579568980,
+      27974051,
+      2617466775,
+      964075233,
+      907049942,
+      4164146575,
+      3377168066,
+      2524828266,
+      1083546008,
+      2992960953,
+      2260789066,
+      1543742095,
+      2843842831,
+      1375722284,
+      3574521313,
+      110842534,
+      2310998251,
+      3076511734,
+      783145600,
+      1287776608,
+      3087144146,
+      305559823,
+      2356293719,
+      3228441476,
+      1678938122,
+      3775814061,
+      1620283952,
+      2512027726,
+      1031432407,
+      962295099,
+      3877418501,
+      968669928,
+      304126693,
+      3711291137,
+      3847527101,
+      494066767,
+      4050229756,
+      4169448589,
+      671763915,
+      1095747781,
+      4006132710,
+      394725957,
+      200521654,
+      2715998750,
+      1477567673,
+      895171901,
+      3370105999,
+      2684157455,
+      4153990023,
+      3966076501,
+      2043374409,
+      144443759,
+      6764556,
+      1611650045,
+      1480956755,
+      1388276468,
+      4136518438,
+      1538041336,
+      266773992,
+      1623357516,
+      2267298390,
+      3183919402,
+      1084292424,
+      2796136160,
+      2413448816,
+      2850375199,
+      3510894040,
+      2644778623,
+      3317288284,
+      3697317540,
+      1465776787,
+      1843489446,
+      1416711171,
+      744701117,
+      1286781349,
+      3748640476,
+      861982119,
+      2377742909,
+      1171768136,
+      2701877439,
+      3839724288,
+      2869791015,
+      2386067954,
+      2629214347,
+      955801623,
+      3831079317,
+      624
+    ],
+    null
+  ],
+  "pref_sampler": [
+    3,
+    [
+      2147483648,
+      3564348608,
+      1266698288,
+      4212342371,
+      3595291661,
+      3180588708,
+      3037210256,
+      946923017,
+      2565409715,
+      2900535780,
+      924383152,
+      4180157270,
+      4230508198,
+      2039675917,
+      3755350407,
+      2362848650,
+      2818100609,
+      2097423432,
+      524478045,
+      540883378,
+      281170210,
+      1485176884,
+      1493190386,
+      1773214509,
+      380915208,
+      3667698522,
+      2648371337,
+      2961234806,
+      3857480267,
+      1582950522,
+      246289694,
+      3322185604,
+      1944574775,
+      302623699,
+      169865066,
+      1143540808,
+      3733177770,
+      513116636,
+      1411153081,
+      3205493053,
+      768926902,
+      549624109,
+      1470655403,
+      59539609,
+      3678480009,
+      3087139671,
+      1176835859,
+      2078491503,
+      2299934332,
+      1592059249,
+      1062716176,
+      2654193596,
+      3531838733,
+      2661260596,
+      3881209635,
+      2106865768,
+      4154287292,
+      2082185616,
+      2301197011,
+      2177349827,
+      3082181756,
+      1787663536,
+      3714670796,
+      3018262113,
+      1670056238,
+      1856738750,
+      99824592,
+      2279837081,
+      1414647942,
+      3416675731,
+      3458782472,
+      3997022236,
+      468762002,
+      2666158583,
+      953353270,
+      1788980658,
+      3802061067,
+      407586584,
+      1844776834,
+      1906917274,
+      3154715663,
+      3028370222,
+      4156024188,
+      3996363428,
+      80495456,
+      2659800972,
+      2005649973,
+      3818358673,
+      3952623596,
+      2506862371,
+      3282302532,
+      263923435,
+      3384662671,
+      3292439172,
+      3119957588,
+      1224426111,
+      899864150,
+      215262826,
+      1619647231,
+      3347694949,
+      3497868538,
+      2029552053,
+      2992804824,
+      4080010250,
+      2023513186,
+      1885979437,
+      3564622190,
+      3775424270,
+      2297810139,
+      3549449169,
+      2664856277,
+      3274801974,
+      2794883969,
+      980412666,
+      2980215653,
+      2794389321,
+      2816521934,
+      1266970739,
+      542306338,
+      3646225311,
+      3598997630,
+      2111980720,
+      2949252482,
+      2489027658,
+      352815024,
+      11610683,
+      1386663624,
+      2004196796,
+      1161461546,
+      1921293780,
+      2463949525,
+      1647009713,
+      3550093655,
+      2563894064,
+      3486310554,
+      1506105865,
+      243092931,
+      2659437476,
+      4200687059,
+      2284345122,
+      1974438610,
+      3591096528,
+      967119212,
+      3362401375,
+      140678365,
+      311602112,
+      2361740275,
+      2139598582,
+      3632873481,
+      2762232439,
+      4156482318,
+      381637792,
+      3253346525,
+      2492118775,
+      1502434558,
+      3164497290,
+      3550998357,
+      2412448305,
+      2223955385,
+      4122879535,
+      350121793,
+      1835149778,
+      2175117867,
+      989674750,
+      3178241202,
+      3553093569,
+      3470650311,
+      2829698151,
+      3209427769,
+      1779174943,
+      275388428,
+      4044574515,
+      715447260,
+      3180940440,
+      4020772289,
+      1322708567,
+      3189868792,
+      4250485633,
+      716970023,
+      2307550151,
+      1074996711,
+      1217573599,
+      197006094,
+      2178394212,
+      1255233746,
+      4164251484,
+      1405608772,
+      2808160475,
+      1304736088,
+      1796071066,
+      2761748078,
+      3570739698,
+      1616118556,
+      2232868135,
+      3567541936,
+      3470600401,
+      3031621994,
+      3351764214,
+      1359785149,
+      2617497797,
+      3340028190,
+      356162828,
+      2083806068,
+      2503635608,
+      4024838996,
+      2577080371,
+      2897993505,
+      3120733934,
+      905794891,
+      2506078507,
+      4211618666,
+      3777871979,
+      809751414,
+      4080874167,
+      1562977008,
+      3917373055,
+      2132779194,
+      4014249473,
+      4067327082,
+      2582869847,
+      1780081876,
+      1842619106,
+      3381761227,
+      921004274,
+      1393256920,
+      1883566732,
+      2702071861,
+      865327389,
+      1622085203,
+      3021825820,
+      2687061406,
+      1748902923,
+      689023977,
+      308399650,
+      2377287978,
+      1646969411,
+      1051806316,
+      4277884230,
+      2041056290,
+      101134519,
+      2032472116,
+      4112521069,
+      151202901,
+      2773743461,
+      551348559,
+      3476836808,
+      510935951,
+      625057077,
+      3757450756,
+      2977698135,
+      3027776859,
+      2616998041,
+      2773430005,
+      544190486,
+      2241368212,
+      1141105829,
+      1452816309,
+      4199229235,
+      3218013033,
+      4229475816,
+      1659576351,
+      3020348754,
+      1193400518,
+      3208584597,
+      1151197733,
+      2597187966,
+      503065140,
+      2421841572,
+      1437291709,
+      1909275895,
+      2872630545,
+      793588217,
+      3792934707,
+      1784451785,
+      2921385648,
+      1669902526,
+      4189978976,
+      1196986251,
+      434805516,
+      1907541826,
+      2624415034,
+      1687778718,
+      650746582,
+      1949153382,
+      4148493093,
+      841300520,
+      1164202054,
+      4203468658,
+      4106300911,
+      850346789,
+      1715730760,
+      3114661489,
+      2866524548,
+      1360448945,
+      3601318775,
+      1743078223,
+      2413855408,
+      1211895622,
+      325117146,
+      2721152875,
+      1284334485,
+      2446538832,
+      739014618,
+      2237045115,
+      842553465,
+      2538598293,
+      746460793,
+      4010387366,
+      2002655192,
+      4193733112,
+      1194380773,
+      3918217378,
+      1447487475,
+      5659228,
+      3408847694,
+      4190318700,
+      1862549564,
+      781683719,
+      1194618118,
+      755053413,
+      3436011942,
+      2885435303,
+      3081151348,
+      2017642831,
+      1053816502,
+      1086627485,
+      2157296554,
+      110650022,
+      965352898,
+      1003174194,
+      1288956241,
+      4057404871,
+      2965068465,
+      2897064481,
+      2457377317,
+      1879872545,
+      358455290,
+      375086701,
+      3015902095,
+      1676249984,
+      924455526,
+      2084169389,
+      1989014644,
+      1993749926,
+      2009424973,
+      2113340508,
+      3980883273,
+      2915977458,
+      203328382,
+      3020815229,
+      2415050113,
+      4103009585,
+      3700885489,
+      2916647550,
+      1523006503,
+      174302338,
+      2476909338,
+      1969322490,
+      4285741984,
+      1528449097,
+      3355315515,
+      4217241278,
+      599579127,
+      2572243673,
+      3035856735,
+      1539140489,
+      1782314913,
+      4238644287,
+      1746424142,
+      1978148312,
+      2380746849,
+      184941882,
+      1106717981,
+      1720750349,
+      981701307,
+      3953154731,
+      3257809181,
+      2892339376,
+      3339778166,
+      3676936849,
+      87425948,
+      3029257381,
+      2037942523,
+      3807628706,
+      2861474706,
+      1058852346,
+      1322765211,
+      2686046342,
+      2689342655,
+      2303436168,
+      2571627181,
+      1986057734,
+      1183564308,
+      2829677523,
+      1295563975,
+      503126586,
+      2025890348,
+      4179277821,
+      1735262467,
+      981331774,
+      1613447066,
+      1011606109,
+      2000062246,
+      3581448390,
+      3477731384,
+      3641307373,
+      3508544379,
+      2327233491,
+      3931944343,
+      4189052882,
+      2990416380,
+      422406169,
+      202291313,
+      2531006461,
+      4277024116,
+      3815144003,
+      821314585,
+      1344175168,
+      3562834071,
+      1339615445,
+      1831545190,
+      3115548822,
+      743512780,
+      4006999448,
+      3720181735,
+      1012033521,
+      919931041,
+      2628967879,
+      1151876565,
+      1268107129,
+      3674829936,
+      834977846,
+      743987006,
+      3947536548,
+      3706529695,
+      4121073678,
+      2507605742,
+      1595636918,
+      2708047833,
+      2427507331,
+      3868216331,
+      3254240010,
+      2097683411,
+      3279710596,
+      3686819053,
+      1843541720,
+      1683793619,
+      3245287285,
+      3571828776,
+      3733296431,
+      3806747478,
+      1390930605,
+      3860422228,
+      114397037,
+      1931519825,
+      2770684378,
+      1556101783,
+      1436111731,
+      4031950081,
+      562876656,
+      1775895782,
+      612364620,
+      1313509772,
+      4283410242,
+      3252958463,
+      2176555836,
+      3933073367,
+      3013277102,
+      1444071961,
+      3120949516,
+      2824578890,
+      325676929,
+      943677134,
+      1800649256,
+      1721927060,
+      347498719,
+      1435221321,
+      2623572981,
+      1408548470,
+      4145586315,
+      2901889237,
+      1849377952,
+      1239144551,
+      3382598266,
+      2992893897,
+      3738297588,
+      611280106,
+      3897415338,
+      2370299241,
+      1772308583,
+      3697465753,
+      354508058,
+      2702360134,
+      591308331,
+      3524072501,
+      976616000,
+      2563717192,
+      3078266097,
+      1376594703,
+      4209795919,
+      2454412767,
+      2712206031,
+      2963860163,
+      3734324882,
+      2248653800,
+      324872786,
+      3789837448,
+      3779000146,
+      527733939,
+      2844165793,
+      576499681,
+      1618787435,
+      2638888650,
+      57511068,
+      2804627518,
+      2993670030,
+      481402236,
+      2810124845,
+      1416045214,
+      1723694191,
+      1214944572,
+      3188123783,
+      1139185907,
+      3851015362,
+      1719652470,
+      1661343029,
+      3644307578,
+      3564178709,
+      1256656955,
+      46631590,
+      4231317929,
+      3098958589,
+      1834956625,
+      2206185428,
+      3695688374,
+      3647957317,
+      1064098871,
+      1739100906,
+      2579568980,
+      27974051,
+      2617466775,
+      964075233,
+      907049942,
+      4164146575,
+      3377168066,
+      2524828266,
+      1083546008,
+      2992960953,
+      2260789066,
+      1543742095,
+      2843842831,
+      1375722284,
+      3574521313,
+      110842534,
+      2310998251,
+      3076511734,
+      783145600,
+      1287776608,
+      3087144146,
+      305559823,
+      2356293719,
+      3228441476,
+      1678938122,
+      3775814061,
+      1620283952,
+      2512027726,
+      1031432407,
+      962295099,
+      3877418501,
+      968669928,
+      304126693,
+      3711291137,
+      3847527101,
+      494066767,
+      4050229756,
+      4169448589,
+      671763915,
+      1095747781,
+      4006132710,
+      394725957,
+      200521654,
+      2715998750,
+      1477567673,
+      895171901,
+      3370105999,
+      2684157455,
+      4153990023,
+      3966076501,
+      2043374409,
+      144443759,
+      6764556,
+      1611650045,
+      1480956755,
+      1388276468,
+      4136518438,
+      1538041336,
+      266773992,
+      1623357516,
+      2267298390,
+      3183919402,
+      1084292424,
+      2796136160,
+      2413448816,
+      2850375199,
+      3510894040,
+      2644778623,
+      3317288284,
+      3697317540,
+      1465776787,
+      1843489446,
+      1416711171,
+      744701117,
+      1286781349,
+      3748640476,
+      861982119,
+      2377742909,
+      1171768136,
+      2701877439,
+      3839724288,
+      2869791015,
+      2386067954,
+      2629214347,
+      955801623,
+      3831079317,
+      624
+    ],
+    null
+  ],
+  "progress_sampler": null,
+  "similarity_sampler": null
+}
\ No newline at end of file
diff --git a/models/Robometer-4B/metrics.json b/models/Robometer-4B/metrics.json
new file mode 100644
index 0000000000000000000000000000000000000000..9054e563c37d8a51985256f80b87d7f320945176
--- /dev/null
+++ b/models/Robometer-4B/metrics.json
@@ -0,0 +1,369 @@
+{
+  "step": 4500,
+  "metrics": {
+    "eval_rew_align/success_auprc_racer_val": 0.5972598636691593,
+    "eval_rew_align/positive_success_acc_racer_val": 0.5238095238095238,
+    "eval_rew_align/negative_success_acc_racer_val": 0.9725363489499192,
+    "eval_rew_align/loss_racer_val": 1.5039077520370483,
+    "eval_rew_align/pearson_racer_val": 0.8166853465988891,
+    "eval_rew_align/success_auprc_oxe_bc_z_eval": 0.054588487359398905,
+    "eval_rew_align/positive_success_acc_oxe_bc_z_eval": 0.7,
+    "eval_rew_align/negative_success_acc_oxe_bc_z_eval": 0.9432314410480349,
+    "eval_rew_align/loss_oxe_bc_z_eval": 1.7649718403816224,
+    "eval_rew_align/pearson_oxe_bc_z_eval": 0.5611694184881661,
+    "eval_rew_align/success_auprc_oxe_berkeley_cable_eval": 0.12320737550700828,
+    "eval_rew_align/positive_success_acc_oxe_berkeley_cable_eval": 0.7,
+    "eval_rew_align/negative_success_acc_oxe_berkeley_cable_eval": 0.9396299902629016,
+    "eval_rew_align/loss_oxe_berkeley_cable_eval": 1.6676030993461608,
+    "eval_rew_align/pearson_oxe_berkeley_cable_eval": 0.7626281468321523,
+    "eval_rew_align/success_auprc_oxe_bridge_v2_eval": 0.2226129586383097,
+    "eval_rew_align/positive_success_acc_oxe_bridge_v2_eval": 0.7,
+    "eval_rew_align/negative_success_acc_oxe_bridge_v2_eval": 0.9700440528634361,
+    "eval_rew_align/loss_oxe_bridge_v2_eval": 1.5779191851615906,
+    "eval_rew_align/pearson_oxe_bridge_v2_eval": 0.8196023502220793,
+    "eval_rew_align/success_auprc_oxe_jaco_eval": 0.05703350629550197,
+    "eval_rew_align/positive_success_acc_oxe_jaco_eval": 0.8,
+    "eval_rew_align/negative_success_acc_oxe_jaco_eval": 0.9796816087138668,
+    "eval_rew_align/loss_oxe_jaco_eval": 1.701886808872223,
+    "eval_rew_align/pearson_oxe_jaco_eval": 0.7369627561402344,
+    "eval_rew_align/success_auprc_oxe_toto_eval": 0.10819046102805713,
+    "eval_rew_align/positive_success_acc_oxe_toto_eval": 1.0,
+    "eval_rew_align/negative_success_acc_oxe_toto_eval": 0.9452054794520548,
+    "eval_rew_align/loss_oxe_toto_eval": 1.5248035669326783,
+    "eval_rew_align/pearson_oxe_toto_eval": 0.9275399402861348,
+    "eval_rew_align/success_auprc_oxe_viola_eval": 0.3924038961069135,
+    "eval_rew_align/positive_success_acc_oxe_viola_eval": 1.0,
+    "eval_rew_align/negative_success_acc_oxe_viola_eval": 0.9430528375733855,
+    "eval_rew_align/loss_oxe_viola_eval": 1.5757618188858031,
+    "eval_rew_align/pearson_oxe_viola_eval": 0.8978344352364431,
+    "eval_rew_align/success_auprc_mw_eval": 0.14365004363589842,
+    "eval_rew_align/positive_success_acc_mw_eval": 0.8,
+    "eval_rew_align/negative_success_acc_mw_eval": 0.9627450980392157,
+    "eval_rew_align/loss_mw_eval": 1.7702434301376342,
+    "eval_rew_align/pearson_mw_eval": 0.7687541228936258,
+    "eval_rew_align/success_auprc_libero_90": 0.1795092166845774,
+    "eval_rew_align/positive_success_acc_libero_90": 0.9,
+    "eval_rew_align/negative_success_acc_libero_90": 0.9682352941176471,
+    "eval_rew_align/loss_libero_90": 1.5339298248291016,
+    "eval_rew_align/pearson_libero_90": 0.8980980150621931,
+    "eval_rew_align/success_auprc_usc_trossen": 0.2819898652527857,
+    "eval_rew_align/positive_success_acc_usc_trossen": 0.5,
+    "eval_rew_align/negative_success_acc_usc_trossen": 0.98,
+    "eval_rew_align/loss_usc_trossen": 1.5562334299087524,
+    "eval_rew_align/pearson_usc_trossen": 0.7085253582776633,
+    "eval_p_rank/kendall_last_usc_trossen": 0.8333333333333333,
+    "eval_p_rank/kendall_rewind_last_usc_trossen": 1.0,
+    "eval_p_rank/avg_succ_subopt_diff_last_usc_trossen": 0.14124762515227,
+    "eval_p_rank/min_succ_subopt_diff_last_usc_trossen": 0.040902674198150635,
+    "eval_p_rank/max_succ_subopt_diff_last_usc_trossen": 0.2803109735250473,
+    "eval_p_rank/avg_subopt_fail_diff_last_usc_trossen": 0.19397936016321182,
+    "eval_p_rank/min_subopt_fail_diff_last_usc_trossen": 0.026902765035629272,
+    "eval_p_rank/max_subopt_fail_diff_last_usc_trossen": 0.3610559552907944,
+    "eval_p_rank/avg_succ_fail_diff_last_usc_trossen": 0.28165244973368114,
+    "eval_p_rank/min_succ_fail_diff_last_usc_trossen": 0.06780543923377991,
+    "eval_p_rank/max_succ_fail_diff_last_usc_trossen": 0.46358518302440643,
+    "eval_p_rank/ranking_acc_last_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_all_pairs_last_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_usc_trossen": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_usc_trossen": 0.75,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_usc_trossen": 0.8888888888888888,
+    "eval_p_rank/kendall_avg_usc_trossen": 0.8333333333333333,
+    "eval_p_rank/kendall_rewind_avg_usc_trossen": 1.0,
+    "eval_p_rank/avg_succ_subopt_diff_avg_usc_trossen": 0.14124762515227,
+    "eval_p_rank/min_succ_subopt_diff_avg_usc_trossen": 0.040902674198150635,
+    "eval_p_rank/max_succ_subopt_diff_avg_usc_trossen": 0.2803109735250473,
+    "eval_p_rank/avg_subopt_fail_diff_avg_usc_trossen": 0.19397936016321182,
+    "eval_p_rank/min_subopt_fail_diff_avg_usc_trossen": 0.026902765035629272,
+    "eval_p_rank/max_subopt_fail_diff_avg_usc_trossen": 0.3610559552907944,
+    "eval_p_rank/avg_succ_fail_diff_avg_usc_trossen": 0.28165244973368114,
+    "eval_p_rank/min_succ_fail_diff_avg_usc_trossen": 0.06780543923377991,
+    "eval_p_rank/max_succ_fail_diff_avg_usc_trossen": 0.46358518302440643,
+    "eval_p_rank/ranking_acc_avg_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_all_pairs_avg_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_usc_trossen": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_usc_trossen": 0.75,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_usc_trossen": 0.8888888888888888,
+    "eval_p_rank/kendall_sum_usc_trossen": 0.8333333333333333,
+    "eval_p_rank/kendall_rewind_sum_usc_trossen": 1.0,
+    "eval_p_rank/avg_succ_subopt_diff_sum_usc_trossen": 0.14124762515227,
+    "eval_p_rank/min_succ_subopt_diff_sum_usc_trossen": 0.040902674198150635,
+    "eval_p_rank/max_succ_subopt_diff_sum_usc_trossen": 0.2803109735250473,
+    "eval_p_rank/avg_subopt_fail_diff_sum_usc_trossen": 0.19397936016321182,
+    "eval_p_rank/min_subopt_fail_diff_sum_usc_trossen": 0.026902765035629272,
+    "eval_p_rank/max_subopt_fail_diff_sum_usc_trossen": 0.3610559552907944,
+    "eval_p_rank/avg_succ_fail_diff_sum_usc_trossen": 0.28165244973368114,
+    "eval_p_rank/min_succ_fail_diff_sum_usc_trossen": 0.06780543923377991,
+    "eval_p_rank/max_succ_fail_diff_sum_usc_trossen": 0.46358518302440643,
+    "eval_p_rank/ranking_acc_sum_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_all_pairs_sum_usc_trossen": 0.8809523809523809,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_usc_trossen": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_usc_trossen": 0.75,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_usc_trossen": 0.8888888888888888,
+    "eval_rew_align/success_auprc_rfm_new_mit_franka_nowrist": 0.13878492377635082,
+    "eval_rew_align/positive_success_acc_rfm_new_mit_franka_nowrist": 0.9,
+    "eval_rew_align/negative_success_acc_rfm_new_mit_franka_nowrist": 0.9635294117647059,
+    "eval_rew_align/loss_rfm_new_mit_franka_nowrist": 1.3595333456993104,
+    "eval_rew_align/pearson_rfm_new_mit_franka_nowrist": 0.9332205211882452,
+    "eval_p_rank/kendall_last_rfm_new_mit_franka_nowrist": 0.46904761904761905,
+    "eval_p_rank/kendall_rewind_last_rfm_new_mit_franka_nowrist": 0.8095238095238095,
+    "eval_p_rank/avg_succ_subopt_diff_last_rfm_new_mit_franka_nowrist": 0.1011232117811839,
+    "eval_p_rank/min_succ_subopt_diff_last_rfm_new_mit_franka_nowrist": 0.022794996698697445,
+    "eval_p_rank/max_succ_subopt_diff_last_rfm_new_mit_franka_nowrist": 0.21488183736801147,
+    "eval_p_rank/avg_subopt_fail_diff_last_rfm_new_mit_franka_nowrist": 0.14513030257962997,
+    "eval_p_rank/min_subopt_fail_diff_last_rfm_new_mit_franka_nowrist": -0.14316336512565614,
+    "eval_p_rank/max_subopt_fail_diff_last_rfm_new_mit_franka_nowrist": 0.34648392796516414,
+    "eval_p_rank/avg_succ_fail_diff_last_rfm_new_mit_franka_nowrist": 0.24625351436081383,
+    "eval_p_rank/min_succ_fail_diff_last_rfm_new_mit_franka_nowrist": -0.00810291568438215,
+    "eval_p_rank/max_succ_fail_diff_last_rfm_new_mit_franka_nowrist": 0.49043338249127066,
+    "eval_p_rank/ranking_acc_last_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_all_pairs_last_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_rfm_new_mit_franka_nowrist": 0.8482142857142857,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_rfm_new_mit_franka_nowrist": 0.7523809523809524,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_rfm_new_mit_franka_nowrist": 0.6551724137931034,
+    "eval_p_rank/kendall_avg_rfm_new_mit_franka_nowrist": 0.46904761904761905,
+    "eval_p_rank/kendall_rewind_avg_rfm_new_mit_franka_nowrist": 0.8095238095238095,
+    "eval_p_rank/avg_succ_subopt_diff_avg_rfm_new_mit_franka_nowrist": 0.1011232117811839,
+    "eval_p_rank/min_succ_subopt_diff_avg_rfm_new_mit_franka_nowrist": 0.022794996698697445,
+    "eval_p_rank/max_succ_subopt_diff_avg_rfm_new_mit_franka_nowrist": 0.21488183736801147,
+    "eval_p_rank/avg_subopt_fail_diff_avg_rfm_new_mit_franka_nowrist": 0.14513030257962997,
+    "eval_p_rank/min_subopt_fail_diff_avg_rfm_new_mit_franka_nowrist": -0.14316336512565614,
+    "eval_p_rank/max_subopt_fail_diff_avg_rfm_new_mit_franka_nowrist": 0.34648392796516414,
+    "eval_p_rank/avg_succ_fail_diff_avg_rfm_new_mit_franka_nowrist": 0.24625351436081383,
+    "eval_p_rank/min_succ_fail_diff_avg_rfm_new_mit_franka_nowrist": -0.00810291568438215,
+    "eval_p_rank/max_succ_fail_diff_avg_rfm_new_mit_franka_nowrist": 0.49043338249127066,
+    "eval_p_rank/ranking_acc_avg_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_all_pairs_avg_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_rfm_new_mit_franka_nowrist": 0.8482142857142857,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_rfm_new_mit_franka_nowrist": 0.7523809523809524,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_rfm_new_mit_franka_nowrist": 0.6551724137931034,
+    "eval_p_rank/kendall_sum_rfm_new_mit_franka_nowrist": 0.46904761904761905,
+    "eval_p_rank/kendall_rewind_sum_rfm_new_mit_franka_nowrist": 0.8095238095238095,
+    "eval_p_rank/avg_succ_subopt_diff_sum_rfm_new_mit_franka_nowrist": 0.1011232117811839,
+    "eval_p_rank/min_succ_subopt_diff_sum_rfm_new_mit_franka_nowrist": 0.022794996698697445,
+    "eval_p_rank/max_succ_subopt_diff_sum_rfm_new_mit_franka_nowrist": 0.21488183736801147,
+    "eval_p_rank/avg_subopt_fail_diff_sum_rfm_new_mit_franka_nowrist": 0.14513030257962997,
+    "eval_p_rank/min_subopt_fail_diff_sum_rfm_new_mit_franka_nowrist": -0.14316336512565614,
+    "eval_p_rank/max_subopt_fail_diff_sum_rfm_new_mit_franka_nowrist": 0.34648392796516414,
+    "eval_p_rank/avg_succ_fail_diff_sum_rfm_new_mit_franka_nowrist": 0.24625351436081383,
+    "eval_p_rank/min_succ_fail_diff_sum_rfm_new_mit_franka_nowrist": -0.00810291568438215,
+    "eval_p_rank/max_succ_fail_diff_sum_rfm_new_mit_franka_nowrist": 0.49043338249127066,
+    "eval_p_rank/ranking_acc_sum_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_all_pairs_sum_rfm_new_mit_franka_nowrist": 0.7598684210526315,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_rfm_new_mit_franka_nowrist": 0.8482142857142857,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_rfm_new_mit_franka_nowrist": 0.7523809523809524,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_rfm_new_mit_franka_nowrist": 0.6551724137931034,
+    "eval_rew_align/success_auprc_utd_so101_clean_top": 0.1594673014952464,
+    "eval_rew_align/positive_success_acc_utd_so101_clean_top": 0.8,
+    "eval_rew_align/negative_success_acc_utd_so101_clean_top": 0.9796078431372549,
+    "eval_rew_align/loss_utd_so101_clean_top": 1.422999668121338,
+    "eval_rew_align/pearson_utd_so101_clean_top": 0.9214771733077172,
+    "eval_p_rank/kendall_last_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/kendall_rewind_last_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_last_utd_so101_clean_top": 0.1281689941883087,
+    "eval_p_rank/min_succ_subopt_diff_last_utd_so101_clean_top": -0.6224770694971085,
+    "eval_p_rank/max_succ_subopt_diff_last_utd_so101_clean_top": 0.4432547390460968,
+    "eval_p_rank/avg_subopt_fail_diff_last_utd_so101_clean_top": 0.2357720375061035,
+    "eval_p_rank/min_subopt_fail_diff_last_utd_so101_clean_top": -0.012576103210449219,
+    "eval_p_rank/max_subopt_fail_diff_last_utd_so101_clean_top": 0.5894219428300858,
+    "eval_p_rank/avg_succ_fail_diff_last_utd_so101_clean_top": 0.3639410316944122,
+    "eval_p_rank/min_succ_fail_diff_last_utd_so101_clean_top": -0.033055126667022705,
+    "eval_p_rank/max_succ_fail_diff_last_utd_so101_clean_top": 0.6650743782520294,
+    "eval_p_rank/ranking_acc_last_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_all_pairs_last_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_utd_so101_clean_top": 0.9,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_utd_so101_clean_top": 0.8,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_utd_so101_clean_top": 0.9,
+    "eval_p_rank/kendall_avg_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/kendall_rewind_avg_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_avg_utd_so101_clean_top": 0.1281689941883087,
+    "eval_p_rank/min_succ_subopt_diff_avg_utd_so101_clean_top": -0.6224770694971085,
+    "eval_p_rank/max_succ_subopt_diff_avg_utd_so101_clean_top": 0.4432547390460968,
+    "eval_p_rank/avg_subopt_fail_diff_avg_utd_so101_clean_top": 0.2357720375061035,
+    "eval_p_rank/min_subopt_fail_diff_avg_utd_so101_clean_top": -0.012576103210449219,
+    "eval_p_rank/max_subopt_fail_diff_avg_utd_so101_clean_top": 0.5894219428300858,
+    "eval_p_rank/avg_succ_fail_diff_avg_utd_so101_clean_top": 0.3639410316944122,
+    "eval_p_rank/min_succ_fail_diff_avg_utd_so101_clean_top": -0.033055126667022705,
+    "eval_p_rank/max_succ_fail_diff_avg_utd_so101_clean_top": 0.6650743782520294,
+    "eval_p_rank/ranking_acc_avg_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_all_pairs_avg_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_utd_so101_clean_top": 0.9,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_utd_so101_clean_top": 0.8,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_utd_so101_clean_top": 0.9,
+    "eval_p_rank/kendall_sum_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/kendall_rewind_sum_utd_so101_clean_top": 0.7333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_sum_utd_so101_clean_top": 0.1281689941883087,
+    "eval_p_rank/min_succ_subopt_diff_sum_utd_so101_clean_top": -0.6224770694971085,
+    "eval_p_rank/max_succ_subopt_diff_sum_utd_so101_clean_top": 0.4432547390460968,
+    "eval_p_rank/avg_subopt_fail_diff_sum_utd_so101_clean_top": 0.2357720375061035,
+    "eval_p_rank/min_subopt_fail_diff_sum_utd_so101_clean_top": -0.012576103210449219,
+    "eval_p_rank/max_subopt_fail_diff_sum_utd_so101_clean_top": 0.5894219428300858,
+    "eval_p_rank/avg_succ_fail_diff_sum_utd_so101_clean_top": 0.3639410316944122,
+    "eval_p_rank/min_succ_fail_diff_sum_utd_so101_clean_top": -0.033055126667022705,
+    "eval_p_rank/max_succ_fail_diff_sum_utd_so101_clean_top": 0.6650743782520294,
+    "eval_p_rank/ranking_acc_sum_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_all_pairs_sum_utd_so101_clean_top": 0.8666666666666667,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_utd_so101_clean_top": 0.9,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_utd_so101_clean_top": 0.8,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_utd_so101_clean_top": 0.9,
+    "eval_rew_align/success_auprc_usc_xarm": 0.3298253598253598,
+    "eval_rew_align/positive_success_acc_usc_xarm": 1.0,
+    "eval_rew_align/negative_success_acc_usc_xarm": 0.971764705882353,
+    "eval_rew_align/loss_usc_xarm": 1.459894859790802,
+    "eval_rew_align/pearson_usc_xarm": 0.9290145264370201,
+    "eval_p_rank/kendall_last_usc_xarm": 0.75,
+    "eval_p_rank/kendall_rewind_last_usc_xarm": 0.8888888888888888,
+    "eval_p_rank/avg_succ_subopt_diff_last_usc_xarm": 0.11288829644521077,
+    "eval_p_rank/min_succ_subopt_diff_last_usc_xarm": 0.0029833614826202393,
+    "eval_p_rank/max_succ_subopt_diff_last_usc_xarm": 0.217641681432724,
+    "eval_p_rank/avg_subopt_fail_diff_last_usc_xarm": 0.11646403868993123,
+    "eval_p_rank/min_subopt_fail_diff_last_usc_xarm": -0.03846535086631775,
+    "eval_p_rank/max_subopt_fail_diff_last_usc_xarm": 0.34273654222488403,
+    "eval_p_rank/avg_succ_fail_diff_last_usc_xarm": 0.229352335135142,
+    "eval_p_rank/min_succ_fail_diff_last_usc_xarm": 0.0956188440322876,
+    "eval_p_rank/max_succ_fail_diff_last_usc_xarm": 0.35525771975517273,
+    "eval_p_rank/ranking_acc_last_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_all_pairs_last_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_usc_xarm": 0.7083333333333334,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_usc_xarm": 1.0,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_usc_xarm": 0.9166666666666666,
+    "eval_p_rank/kendall_avg_usc_xarm": 0.75,
+    "eval_p_rank/kendall_rewind_avg_usc_xarm": 0.8888888888888888,
+    "eval_p_rank/avg_succ_subopt_diff_avg_usc_xarm": 0.11288829644521077,
+    "eval_p_rank/min_succ_subopt_diff_avg_usc_xarm": 0.0029833614826202393,
+    "eval_p_rank/max_succ_subopt_diff_avg_usc_xarm": 0.217641681432724,
+    "eval_p_rank/avg_subopt_fail_diff_avg_usc_xarm": 0.11646403868993123,
+    "eval_p_rank/min_subopt_fail_diff_avg_usc_xarm": -0.03846535086631775,
+    "eval_p_rank/max_subopt_fail_diff_avg_usc_xarm": 0.34273654222488403,
+    "eval_p_rank/avg_succ_fail_diff_avg_usc_xarm": 0.229352335135142,
+    "eval_p_rank/min_succ_fail_diff_avg_usc_xarm": 0.0956188440322876,
+    "eval_p_rank/max_succ_fail_diff_avg_usc_xarm": 0.35525771975517273,
+    "eval_p_rank/ranking_acc_avg_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_all_pairs_avg_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_usc_xarm": 0.7083333333333334,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_usc_xarm": 1.0,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_usc_xarm": 0.9166666666666666,
+    "eval_p_rank/kendall_sum_usc_xarm": 0.75,
+    "eval_p_rank/kendall_rewind_sum_usc_xarm": 0.8888888888888888,
+    "eval_p_rank/avg_succ_subopt_diff_sum_usc_xarm": 0.11288829644521077,
+    "eval_p_rank/min_succ_subopt_diff_sum_usc_xarm": 0.0029833614826202393,
+    "eval_p_rank/max_succ_subopt_diff_sum_usc_xarm": 0.217641681432724,
+    "eval_p_rank/avg_subopt_fail_diff_sum_usc_xarm": 0.11646403868993123,
+    "eval_p_rank/min_subopt_fail_diff_sum_usc_xarm": -0.03846535086631775,
+    "eval_p_rank/max_subopt_fail_diff_sum_usc_xarm": 0.34273654222488403,
+    "eval_p_rank/avg_succ_fail_diff_sum_usc_xarm": 0.229352335135142,
+    "eval_p_rank/min_succ_fail_diff_sum_usc_xarm": 0.0956188440322876,
+    "eval_p_rank/max_succ_fail_diff_sum_usc_xarm": 0.35525771975517273,
+    "eval_p_rank/ranking_acc_sum_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_all_pairs_sum_usc_xarm": 0.875,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_usc_xarm": 0.7083333333333334,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_usc_xarm": 1.0,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_usc_xarm": 0.9166666666666666,
+    "eval_rew_align/success_auprc_usc_franka": 0.178648534454372,
+    "eval_rew_align/positive_success_acc_usc_franka": 0.75,
+    "eval_rew_align/negative_success_acc_usc_franka": 0.9436274509803921,
+    "eval_rew_align/loss_usc_franka": 1.4907499551773071,
+    "eval_rew_align/pearson_usc_franka": 0.9115594502071923,
+    "eval_p_rank/kendall_last_usc_franka": 0.7916666666666666,
+    "eval_p_rank/kendall_rewind_last_usc_franka": 0.8333333333333334,
+    "eval_p_rank/avg_succ_subopt_diff_last_usc_franka": 0.06161930412054062,
+    "eval_p_rank/min_succ_subopt_diff_last_usc_franka": -0.010589927434921265,
+    "eval_p_rank/max_succ_subopt_diff_last_usc_franka": 0.17146822810173035,
+    "eval_p_rank/avg_subopt_fail_diff_last_usc_franka": 0.18651490285992622,
+    "eval_p_rank/min_subopt_fail_diff_last_usc_franka": 0.0136566162109375,
+    "eval_p_rank/max_subopt_fail_diff_last_usc_franka": 0.3522116541862488,
+    "eval_p_rank/avg_succ_fail_diff_last_usc_franka": 0.24813420698046684,
+    "eval_p_rank/min_succ_fail_diff_last_usc_franka": 0.032290756702423096,
+    "eval_p_rank/max_succ_fail_diff_last_usc_franka": 0.4191764295101166,
+    "eval_p_rank/ranking_acc_last_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_all_pairs_last_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_usc_franka": 0.8125,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_usc_franka": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_usc_franka": 0.9375,
+    "eval_p_rank/kendall_avg_usc_franka": 0.7916666666666666,
+    "eval_p_rank/kendall_rewind_avg_usc_franka": 0.8333333333333334,
+    "eval_p_rank/avg_succ_subopt_diff_avg_usc_franka": 0.06161930412054062,
+    "eval_p_rank/min_succ_subopt_diff_avg_usc_franka": -0.010589927434921265,
+    "eval_p_rank/max_succ_subopt_diff_avg_usc_franka": 0.17146822810173035,
+    "eval_p_rank/avg_subopt_fail_diff_avg_usc_franka": 0.18651490285992622,
+    "eval_p_rank/min_subopt_fail_diff_avg_usc_franka": 0.0136566162109375,
+    "eval_p_rank/max_subopt_fail_diff_avg_usc_franka": 0.3522116541862488,
+    "eval_p_rank/avg_succ_fail_diff_avg_usc_franka": 0.24813420698046684,
+    "eval_p_rank/min_succ_fail_diff_avg_usc_franka": 0.032290756702423096,
+    "eval_p_rank/max_succ_fail_diff_avg_usc_franka": 0.4191764295101166,
+    "eval_p_rank/ranking_acc_avg_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_all_pairs_avg_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_usc_franka": 0.8125,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_usc_franka": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_usc_franka": 0.9375,
+    "eval_p_rank/kendall_sum_usc_franka": 0.7916666666666666,
+    "eval_p_rank/kendall_rewind_sum_usc_franka": 0.8333333333333334,
+    "eval_p_rank/avg_succ_subopt_diff_sum_usc_franka": 0.06161930412054062,
+    "eval_p_rank/min_succ_subopt_diff_sum_usc_franka": -0.010589927434921265,
+    "eval_p_rank/max_succ_subopt_diff_sum_usc_franka": 0.17146822810173035,
+    "eval_p_rank/avg_subopt_fail_diff_sum_usc_franka": 0.18651490285992622,
+    "eval_p_rank/min_subopt_fail_diff_sum_usc_franka": 0.0136566162109375,
+    "eval_p_rank/max_subopt_fail_diff_sum_usc_franka": 0.3522116541862488,
+    "eval_p_rank/avg_succ_fail_diff_sum_usc_franka": 0.24813420698046684,
+    "eval_p_rank/min_succ_fail_diff_sum_usc_franka": 0.032290756702423096,
+    "eval_p_rank/max_succ_fail_diff_sum_usc_franka": 0.4191764295101166,
+    "eval_p_rank/ranking_acc_sum_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_all_pairs_sum_usc_franka": 0.8958333333333334,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_usc_franka": 0.8125,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_usc_franka": 0.9375,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_usc_franka": 0.9375,
+    "eval_rew_align/success_auprc_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.08961776352588778,
+    "eval_rew_align/positive_success_acc_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.9,
+    "eval_rew_align/negative_success_acc_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.9552941176470588,
+    "eval_rew_align/loss_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 1.5833105087280273,
+    "eval_rew_align/pearson_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.8841339237987327,
+    "eval_p_rank/kendall_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.3093333333333333,
+    "eval_p_rank/kendall_rewind_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.5333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.02896499633789066,
+    "eval_p_rank/min_succ_subopt_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.23819030523300178,
+    "eval_p_rank/max_succ_subopt_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.0715120196342468,
+    "eval_p_rank/avg_subopt_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.20831456266343595,
+    "eval_p_rank/min_subopt_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.16911142468452456,
+    "eval_p_rank/max_subopt_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4413899302482605,
+    "eval_p_rank/avg_succ_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.17934956632554527,
+    "eval_p_rank/min_succ_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.09759940505027775,
+    "eval_p_rank/max_succ_fail_diff_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4535657167434692,
+    "eval_p_rank/ranking_acc_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_all_pairs_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.76,
+    "eval_p_rank/ranking_acc_failure_vs_successful_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.744,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_last_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.46,
+    "eval_p_rank/kendall_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.3093333333333333,
+    "eval_p_rank/kendall_rewind_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.5333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.02896499633789066,
+    "eval_p_rank/min_succ_subopt_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.23819030523300178,
+    "eval_p_rank/max_succ_subopt_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.0715120196342468,
+    "eval_p_rank/avg_subopt_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.20831456266343595,
+    "eval_p_rank/min_subopt_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.16911142468452456,
+    "eval_p_rank/max_subopt_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4413899302482605,
+    "eval_p_rank/avg_succ_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.17934956632554527,
+    "eval_p_rank/min_succ_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.09759940505027775,
+    "eval_p_rank/max_succ_fail_diff_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4535657167434692,
+    "eval_p_rank/ranking_acc_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_all_pairs_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.76,
+    "eval_p_rank/ranking_acc_failure_vs_successful_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.744,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_avg_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.46,
+    "eval_p_rank/kendall_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.3093333333333333,
+    "eval_p_rank/kendall_rewind_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.5333333333333333,
+    "eval_p_rank/avg_succ_subopt_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.02896499633789066,
+    "eval_p_rank/min_succ_subopt_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.23819030523300178,
+    "eval_p_rank/max_succ_subopt_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.0715120196342468,
+    "eval_p_rank/avg_subopt_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.20831456266343595,
+    "eval_p_rank/min_subopt_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.16911142468452456,
+    "eval_p_rank/max_subopt_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4413899302482605,
+    "eval_p_rank/avg_succ_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.17934956632554527,
+    "eval_p_rank/min_succ_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": -0.09759940505027775,
+    "eval_p_rank/max_succ_fail_diff_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.4535657167434692,
+    "eval_p_rank/ranking_acc_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_all_pairs_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.6546666666666666,
+    "eval_p_rank/ranking_acc_failure_vs_suboptimal_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.76,
+    "eval_p_rank/ranking_acc_failure_vs_successful_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.744,
+    "eval_p_rank/ranking_acc_suboptimal_vs_successful_sum_jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": 0.46,
+    "time/custom_evaluations": 227.5345072869677
+  }
+}
\ No newline at end of file
diff --git a/models/Robometer-4B/model-00001-of-00002.safetensors b/models/Robometer-4B/model-00001-of-00002.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e45dcf2d98236089c61705f013cf10baec6ec5bb
--- /dev/null
+++ b/models/Robometer-4B/model-00001-of-00002.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cfa81761595f6b10bd079736eb32bdeeeb0b49224619411bd511d9d1f1171a4
+size 4965629698
diff --git a/models/Robometer-4B/model-00002-of-00002.safetensors b/models/Robometer-4B/model-00002-of-00002.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1be8a5ccfe859630938dc0f212374411d1d35ed8
--- /dev/null
+++ b/models/Robometer-4B/model-00002-of-00002.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a25be9713ccc3cb14738be76ef6665742093d70323ec026ea70e8fa843a342be
+size 3935034144
diff --git a/models/Robometer-4B/model.safetensors.index.json b/models/Robometer-4B/model.safetensors.index.json
new file mode 100644
index 0000000000000000000000000000000000000000..3274c65863b237a8f04ddb2a67753f61d010768b
--- /dev/null
+++ b/models/Robometer-4B/model.safetensors.index.json
@@ -0,0 +1,746 @@
+{
+  "metadata": {
+    "total_parameters": 4450286861,
+    "total_size": 8900573722
+  },
+  "weight_map": {
+    "frame_pool_attn.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.embed_tokens.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.16.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.16.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.17.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.17.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.18.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.19.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.20.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.21.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.22.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.24.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.26.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.28.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.32.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.33.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.34.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.35.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.language_model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.language_model.norm.weight": "model-00002-of-00002.safetensors",
+    "model.visual.blocks.0.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.0.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.1.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.10.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.11.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.12.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.13.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.14.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.15.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.16.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.17.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.18.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.19.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.2.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.20.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.21.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.22.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.23.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.3.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.4.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.5.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.6.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.7.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.8.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.attn.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.attn.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.attn.qkv.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.attn.qkv.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.mlp.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.mlp.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.mlp.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.mlp.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.norm1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.norm1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.norm2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.blocks.9.norm2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.norm.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.0.norm.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.norm.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.1.norm.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.norm.bias": "model-00001-of-00002.safetensors",
+    "model.visual.deepstack_merger_list.2.norm.weight": "model-00001-of-00002.safetensors",
+    "model.visual.merger.linear_fc1.bias": "model-00001-of-00002.safetensors",
+    "model.visual.merger.linear_fc1.weight": "model-00001-of-00002.safetensors",
+    "model.visual.merger.linear_fc2.bias": "model-00001-of-00002.safetensors",
+    "model.visual.merger.linear_fc2.weight": "model-00001-of-00002.safetensors",
+    "model.visual.merger.norm.bias": "model-00001-of-00002.safetensors",
+    "model.visual.merger.norm.weight": "model-00001-of-00002.safetensors",
+    "model.visual.patch_embed.proj.bias": "model-00001-of-00002.safetensors",
+    "model.visual.patch_embed.proj.weight": "model-00001-of-00002.safetensors",
+    "model.visual.pos_embed.weight": "model-00001-of-00002.safetensors",
+    "preference_head.0.bias": "model-00001-of-00002.safetensors",
+    "preference_head.0.weight": "model-00001-of-00002.safetensors",
+    "preference_head.1.bias": "model-00001-of-00002.safetensors",
+    "preference_head.1.weight": "model-00001-of-00002.safetensors",
+    "preference_head.4.bias": "model-00001-of-00002.safetensors",
+    "preference_head.4.weight": "model-00001-of-00002.safetensors",
+    "progress_head.0.bias": "model-00001-of-00002.safetensors",
+    "progress_head.0.weight": "model-00001-of-00002.safetensors",
+    "progress_head.1.bias": "model-00001-of-00002.safetensors",
+    "progress_head.1.weight": "model-00001-of-00002.safetensors",
+    "progress_head.4.bias": "model-00001-of-00002.safetensors",
+    "progress_head.4.weight": "model-00001-of-00002.safetensors",
+    "similarity_head.0.bias": "model-00001-of-00002.safetensors",
+    "similarity_head.0.weight": "model-00001-of-00002.safetensors",
+    "similarity_head.1.bias": "model-00001-of-00002.safetensors",
+    "similarity_head.1.weight": "model-00001-of-00002.safetensors",
+    "similarity_head.4.bias": "model-00001-of-00002.safetensors",
+    "similarity_head.4.weight": "model-00001-of-00002.safetensors",
+    "success_head.0.bias": "model-00001-of-00002.safetensors",
+    "success_head.0.weight": "model-00001-of-00002.safetensors",
+    "success_head.1.bias": "model-00001-of-00002.safetensors",
+    "success_head.1.weight": "model-00001-of-00002.safetensors",
+    "success_head.4.bias": "model-00001-of-00002.safetensors",
+    "success_head.4.weight": "model-00001-of-00002.safetensors"
+  }
+}
diff --git a/models/Robometer-4B/trainer_state.json b/models/Robometer-4B/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..b778001459c11fcdfa893971c730208f5cf08923
--- /dev/null
+++ b/models/Robometer-4B/trainer_state.json
@@ -0,0 +1,31534 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.27654867256637167,
+  "eval_steps": 250,
+  "global_step": 4500,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 6.145526057030482e-05,
+      "grad_norm": 29.75,
+      "learning_rate": 0.0,
+      "loss": 1.4108,
+      "step": 1
+    },
+    {
+      "epoch": 0.00012291052114060964,
+      "grad_norm": 28.75,
+      "learning_rate": 1.3333333333333334e-08,
+      "loss": 1.4446,
+      "step": 2
+    },
+    {
+      "epoch": 0.00018436578171091445,
+      "grad_norm": 29.875,
+      "learning_rate": 2.6666666666666667e-08,
+      "loss": 1.4226,
+      "step": 3
+    },
+    {
+      "epoch": 0.0002458210422812193,
+      "grad_norm": 28.5,
+      "learning_rate": 4e-08,
+      "loss": 1.4369,
+      "step": 4
+    },
+    {
+      "epoch": 0.00030727630285152407,
+      "grad_norm": 34.0,
+      "learning_rate": 5.3333333333333334e-08,
+      "loss": 1.4434,
+      "step": 5
+    },
+    {
+      "epoch": 0.0003687315634218289,
+      "grad_norm": 35.25,
+      "learning_rate": 6.666666666666668e-08,
+      "loss": 1.4788,
+      "step": 6
+    },
+    {
+      "epoch": 0.00043018682399213374,
+      "grad_norm": 36.25,
+      "learning_rate": 8e-08,
+      "loss": 1.4425,
+      "step": 7
+    },
+    {
+      "epoch": 0.0004916420845624386,
+      "grad_norm": 50.25,
+      "learning_rate": 9.333333333333335e-08,
+      "loss": 1.4805,
+      "step": 8
+    },
+    {
+      "epoch": 0.0005530973451327434,
+      "grad_norm": 33.25,
+      "learning_rate": 1.0666666666666667e-07,
+      "loss": 1.4837,
+      "step": 9
+    },
+    {
+      "epoch": 0.0006145526057030481,
+      "grad_norm": 32.75,
+      "learning_rate": 1.2000000000000002e-07,
+      "loss": 1.472,
+      "step": 10
+    },
+    {
+      "epoch": 0.000676007866273353,
+      "grad_norm": 33.5,
+      "learning_rate": 1.3333333333333336e-07,
+      "loss": 1.5112,
+      "step": 11
+    },
+    {
+      "epoch": 0.0007374631268436578,
+      "grad_norm": 37.5,
+      "learning_rate": 1.4666666666666668e-07,
+      "loss": 1.4344,
+      "step": 12
+    },
+    {
+      "epoch": 0.0007989183874139626,
+      "grad_norm": 41.5,
+      "learning_rate": 1.6e-07,
+      "loss": 1.4732,
+      "step": 13
+    },
+    {
+      "epoch": 0.0008603736479842675,
+      "grad_norm": 30.5,
+      "learning_rate": 1.7333333333333335e-07,
+      "loss": 1.4347,
+      "step": 14
+    },
+    {
+      "epoch": 0.0009218289085545723,
+      "grad_norm": 35.25,
+      "learning_rate": 1.866666666666667e-07,
+      "loss": 1.5008,
+      "step": 15
+    },
+    {
+      "epoch": 0.0009832841691248771,
+      "grad_norm": 33.25,
+      "learning_rate": 2.0000000000000002e-07,
+      "loss": 1.437,
+      "step": 16
+    },
+    {
+      "epoch": 0.0010447394296951818,
+      "grad_norm": 29.875,
+      "learning_rate": 2.1333333333333334e-07,
+      "loss": 1.5193,
+      "step": 17
+    },
+    {
+      "epoch": 0.0011061946902654867,
+      "grad_norm": 34.0,
+      "learning_rate": 2.266666666666667e-07,
+      "loss": 1.4585,
+      "step": 18
+    },
+    {
+      "epoch": 0.0011676499508357916,
+      "grad_norm": 33.25,
+      "learning_rate": 2.4000000000000003e-07,
+      "loss": 1.5157,
+      "step": 19
+    },
+    {
+      "epoch": 0.0012291052114060963,
+      "grad_norm": 31.875,
+      "learning_rate": 2.533333333333333e-07,
+      "loss": 1.4815,
+      "step": 20
+    },
+    {
+      "epoch": 0.0012905604719764012,
+      "grad_norm": 27.0,
+      "learning_rate": 2.666666666666667e-07,
+      "loss": 1.4444,
+      "step": 21
+    },
+    {
+      "epoch": 0.001352015732546706,
+      "grad_norm": 28.0,
+      "learning_rate": 2.8e-07,
+      "loss": 1.4639,
+      "step": 22
+    },
+    {
+      "epoch": 0.0014134709931170107,
+      "grad_norm": 26.125,
+      "learning_rate": 2.9333333333333337e-07,
+      "loss": 1.4939,
+      "step": 23
+    },
+    {
+      "epoch": 0.0014749262536873156,
+      "grad_norm": 35.5,
+      "learning_rate": 3.0666666666666666e-07,
+      "loss": 1.5076,
+      "step": 24
+    },
+    {
+      "epoch": 0.0015363815142576205,
+      "grad_norm": 32.75,
+      "learning_rate": 3.2e-07,
+      "loss": 1.4532,
+      "step": 25
+    },
+    {
+      "epoch": 0.0015978367748279252,
+      "grad_norm": 27.375,
+      "learning_rate": 3.3333333333333335e-07,
+      "loss": 1.5085,
+      "step": 26
+    },
+    {
+      "epoch": 0.00165929203539823,
+      "grad_norm": 26.625,
+      "learning_rate": 3.466666666666667e-07,
+      "loss": 1.4565,
+      "step": 27
+    },
+    {
+      "epoch": 0.001720747295968535,
+      "grad_norm": 37.25,
+      "learning_rate": 3.6e-07,
+      "loss": 1.5237,
+      "step": 28
+    },
+    {
+      "epoch": 0.0017822025565388396,
+      "grad_norm": 24.5,
+      "learning_rate": 3.733333333333334e-07,
+      "loss": 1.4745,
+      "step": 29
+    },
+    {
+      "epoch": 0.0018436578171091445,
+      "grad_norm": 30.375,
+      "learning_rate": 3.8666666666666674e-07,
+      "loss": 1.484,
+      "step": 30
+    },
+    {
+      "epoch": 0.0019051130776794494,
+      "grad_norm": 25.25,
+      "learning_rate": 4.0000000000000003e-07,
+      "loss": 1.4895,
+      "step": 31
+    },
+    {
+      "epoch": 0.0019665683382497543,
+      "grad_norm": 28.0,
+      "learning_rate": 4.133333333333334e-07,
+      "loss": 1.4245,
+      "step": 32
+    },
+    {
+      "epoch": 0.002028023598820059,
+      "grad_norm": 30.875,
+      "learning_rate": 4.266666666666667e-07,
+      "loss": 1.5198,
+      "step": 33
+    },
+    {
+      "epoch": 0.0020894788593903636,
+      "grad_norm": 30.375,
+      "learning_rate": 4.4e-07,
+      "loss": 1.4599,
+      "step": 34
+    },
+    {
+      "epoch": 0.0021509341199606687,
+      "grad_norm": 27.875,
+      "learning_rate": 4.533333333333334e-07,
+      "loss": 1.4397,
+      "step": 35
+    },
+    {
+      "epoch": 0.0022123893805309734,
+      "grad_norm": 27.875,
+      "learning_rate": 4.666666666666667e-07,
+      "loss": 1.4923,
+      "step": 36
+    },
+    {
+      "epoch": 0.002273844641101278,
+      "grad_norm": 24.5,
+      "learning_rate": 4.800000000000001e-07,
+      "loss": 1.468,
+      "step": 37
+    },
+    {
+      "epoch": 0.002335299901671583,
+      "grad_norm": 29.125,
+      "learning_rate": 4.933333333333334e-07,
+      "loss": 1.4585,
+      "step": 38
+    },
+    {
+      "epoch": 0.002396755162241888,
+      "grad_norm": 39.0,
+      "learning_rate": 5.066666666666667e-07,
+      "loss": 1.4864,
+      "step": 39
+    },
+    {
+      "epoch": 0.0024582104228121925,
+      "grad_norm": 28.625,
+      "learning_rate": 5.2e-07,
+      "loss": 1.5024,
+      "step": 40
+    },
+    {
+      "epoch": 0.0025196656833824977,
+      "grad_norm": 29.5,
+      "learning_rate": 5.333333333333335e-07,
+      "loss": 1.4955,
+      "step": 41
+    },
+    {
+      "epoch": 0.0025811209439528023,
+      "grad_norm": 35.25,
+      "learning_rate": 5.466666666666667e-07,
+      "loss": 1.4087,
+      "step": 42
+    },
+    {
+      "epoch": 0.002642576204523107,
+      "grad_norm": 33.5,
+      "learning_rate": 5.6e-07,
+      "loss": 1.4448,
+      "step": 43
+    },
+    {
+      "epoch": 0.002704031465093412,
+      "grad_norm": 32.5,
+      "learning_rate": 5.733333333333334e-07,
+      "loss": 1.4495,
+      "step": 44
+    },
+    {
+      "epoch": 0.0027654867256637168,
+      "grad_norm": 31.75,
+      "learning_rate": 5.866666666666667e-07,
+      "loss": 1.4562,
+      "step": 45
+    },
+    {
+      "epoch": 0.0028269419862340214,
+      "grad_norm": 27.0,
+      "learning_rate": 6.000000000000001e-07,
+      "loss": 1.4601,
+      "step": 46
+    },
+    {
+      "epoch": 0.0028883972468043266,
+      "grad_norm": 29.5,
+      "learning_rate": 6.133333333333333e-07,
+      "loss": 1.4944,
+      "step": 47
+    },
+    {
+      "epoch": 0.0029498525073746312,
+      "grad_norm": 27.25,
+      "learning_rate": 6.266666666666667e-07,
+      "loss": 1.4711,
+      "step": 48
+    },
+    {
+      "epoch": 0.003011307767944936,
+      "grad_norm": 29.75,
+      "learning_rate": 6.4e-07,
+      "loss": 1.4145,
+      "step": 49
+    },
+    {
+      "epoch": 0.003072763028515241,
+      "grad_norm": 23.25,
+      "learning_rate": 6.533333333333334e-07,
+      "loss": 1.4167,
+      "step": 50
+    },
+    {
+      "epoch": 0.0031342182890855457,
+      "grad_norm": 26.5,
+      "learning_rate": 6.666666666666667e-07,
+      "loss": 1.4101,
+      "step": 51
+    },
+    {
+      "epoch": 0.0031956735496558504,
+      "grad_norm": 34.5,
+      "learning_rate": 6.800000000000001e-07,
+      "loss": 1.4635,
+      "step": 52
+    },
+    {
+      "epoch": 0.0032571288102261555,
+      "grad_norm": 25.25,
+      "learning_rate": 6.933333333333334e-07,
+      "loss": 1.4237,
+      "step": 53
+    },
+    {
+      "epoch": 0.00331858407079646,
+      "grad_norm": 26.375,
+      "learning_rate": 7.066666666666667e-07,
+      "loss": 1.4127,
+      "step": 54
+    },
+    {
+      "epoch": 0.003380039331366765,
+      "grad_norm": 30.5,
+      "learning_rate": 7.2e-07,
+      "loss": 1.4324,
+      "step": 55
+    },
+    {
+      "epoch": 0.00344149459193707,
+      "grad_norm": 23.25,
+      "learning_rate": 7.333333333333334e-07,
+      "loss": 1.4182,
+      "step": 56
+    },
+    {
+      "epoch": 0.0035029498525073746,
+      "grad_norm": 36.75,
+      "learning_rate": 7.466666666666668e-07,
+      "loss": 1.4212,
+      "step": 57
+    },
+    {
+      "epoch": 0.0035644051130776793,
+      "grad_norm": 31.125,
+      "learning_rate": 7.6e-07,
+      "loss": 1.4183,
+      "step": 58
+    },
+    {
+      "epoch": 0.0036258603736479844,
+      "grad_norm": 32.5,
+      "learning_rate": 7.733333333333335e-07,
+      "loss": 1.5004,
+      "step": 59
+    },
+    {
+      "epoch": 0.003687315634218289,
+      "grad_norm": 29.125,
+      "learning_rate": 7.866666666666667e-07,
+      "loss": 1.4667,
+      "step": 60
+    },
+    {
+      "epoch": 0.0037487708947885937,
+      "grad_norm": 30.5,
+      "learning_rate": 8.000000000000001e-07,
+      "loss": 1.4404,
+      "step": 61
+    },
+    {
+      "epoch": 0.003810226155358899,
+      "grad_norm": 27.875,
+      "learning_rate": 8.133333333333333e-07,
+      "loss": 1.4657,
+      "step": 62
+    },
+    {
+      "epoch": 0.0038716814159292035,
+      "grad_norm": 40.75,
+      "learning_rate": 8.266666666666668e-07,
+      "loss": 1.4001,
+      "step": 63
+    },
+    {
+      "epoch": 0.003933136676499509,
+      "grad_norm": 27.0,
+      "learning_rate": 8.400000000000001e-07,
+      "loss": 1.4232,
+      "step": 64
+    },
+    {
+      "epoch": 0.003994591937069813,
+      "grad_norm": 27.25,
+      "learning_rate": 8.533333333333334e-07,
+      "loss": 1.3774,
+      "step": 65
+    },
+    {
+      "epoch": 0.004056047197640118,
+      "grad_norm": 29.625,
+      "learning_rate": 8.666666666666668e-07,
+      "loss": 1.3763,
+      "step": 66
+    },
+    {
+      "epoch": 0.004117502458210423,
+      "grad_norm": 27.5,
+      "learning_rate": 8.8e-07,
+      "loss": 1.3875,
+      "step": 67
+    },
+    {
+      "epoch": 0.004178957718780727,
+      "grad_norm": 34.25,
+      "learning_rate": 8.933333333333334e-07,
+      "loss": 1.404,
+      "step": 68
+    },
+    {
+      "epoch": 0.004240412979351032,
+      "grad_norm": 33.5,
+      "learning_rate": 9.066666666666668e-07,
+      "loss": 1.3785,
+      "step": 69
+    },
+    {
+      "epoch": 0.0043018682399213375,
+      "grad_norm": 31.875,
+      "learning_rate": 9.200000000000001e-07,
+      "loss": 1.3663,
+      "step": 70
+    },
+    {
+      "epoch": 0.004363323500491642,
+      "grad_norm": 22.75,
+      "learning_rate": 9.333333333333334e-07,
+      "loss": 1.3843,
+      "step": 71
+    },
+    {
+      "epoch": 0.004424778761061947,
+      "grad_norm": 22.875,
+      "learning_rate": 9.466666666666667e-07,
+      "loss": 1.3844,
+      "step": 72
+    },
+    {
+      "epoch": 0.004486234021632252,
+      "grad_norm": 19.375,
+      "learning_rate": 9.600000000000001e-07,
+      "loss": 1.4336,
+      "step": 73
+    },
+    {
+      "epoch": 0.004547689282202556,
+      "grad_norm": 25.5,
+      "learning_rate": 9.733333333333333e-07,
+      "loss": 1.4111,
+      "step": 74
+    },
+    {
+      "epoch": 0.004609144542772861,
+      "grad_norm": 31.875,
+      "learning_rate": 9.866666666666668e-07,
+      "loss": 1.3456,
+      "step": 75
+    },
+    {
+      "epoch": 0.004670599803343166,
+      "grad_norm": 25.375,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 1.3526,
+      "step": 76
+    },
+    {
+      "epoch": 0.004732055063913471,
+      "grad_norm": 28.625,
+      "learning_rate": 1.0133333333333333e-06,
+      "loss": 1.3306,
+      "step": 77
+    },
+    {
+      "epoch": 0.004793510324483776,
+      "grad_norm": 22.375,
+      "learning_rate": 1.0266666666666669e-06,
+      "loss": 1.3395,
+      "step": 78
+    },
+    {
+      "epoch": 0.004854965585054081,
+      "grad_norm": 23.125,
+      "learning_rate": 1.04e-06,
+      "loss": 1.3588,
+      "step": 79
+    },
+    {
+      "epoch": 0.004916420845624385,
+      "grad_norm": 22.0,
+      "learning_rate": 1.0533333333333333e-06,
+      "loss": 1.3521,
+      "step": 80
+    },
+    {
+      "epoch": 0.00497787610619469,
+      "grad_norm": 27.25,
+      "learning_rate": 1.066666666666667e-06,
+      "loss": 1.3255,
+      "step": 81
+    },
+    {
+      "epoch": 0.005039331366764995,
+      "grad_norm": 25.375,
+      "learning_rate": 1.08e-06,
+      "loss": 1.3141,
+      "step": 82
+    },
+    {
+      "epoch": 0.0051007866273352995,
+      "grad_norm": 19.0,
+      "learning_rate": 1.0933333333333334e-06,
+      "loss": 1.3346,
+      "step": 83
+    },
+    {
+      "epoch": 0.005162241887905605,
+      "grad_norm": 27.75,
+      "learning_rate": 1.1066666666666667e-06,
+      "loss": 1.2773,
+      "step": 84
+    },
+    {
+      "epoch": 0.00522369714847591,
+      "grad_norm": 26.25,
+      "learning_rate": 1.12e-06,
+      "loss": 1.2768,
+      "step": 85
+    },
+    {
+      "epoch": 0.005285152409046214,
+      "grad_norm": 26.0,
+      "learning_rate": 1.1333333333333334e-06,
+      "loss": 1.3128,
+      "step": 86
+    },
+    {
+      "epoch": 0.005346607669616519,
+      "grad_norm": 19.875,
+      "learning_rate": 1.1466666666666668e-06,
+      "loss": 1.3132,
+      "step": 87
+    },
+    {
+      "epoch": 0.005408062930186824,
+      "grad_norm": 30.875,
+      "learning_rate": 1.1600000000000001e-06,
+      "loss": 1.3141,
+      "step": 88
+    },
+    {
+      "epoch": 0.0054695181907571284,
+      "grad_norm": 22.0,
+      "learning_rate": 1.1733333333333335e-06,
+      "loss": 1.2636,
+      "step": 89
+    },
+    {
+      "epoch": 0.0055309734513274336,
+      "grad_norm": 25.125,
+      "learning_rate": 1.1866666666666668e-06,
+      "loss": 1.2607,
+      "step": 90
+    },
+    {
+      "epoch": 0.005592428711897739,
+      "grad_norm": 17.875,
+      "learning_rate": 1.2000000000000002e-06,
+      "loss": 1.2474,
+      "step": 91
+    },
+    {
+      "epoch": 0.005653883972468043,
+      "grad_norm": 21.25,
+      "learning_rate": 1.2133333333333335e-06,
+      "loss": 1.2857,
+      "step": 92
+    },
+    {
+      "epoch": 0.005715339233038348,
+      "grad_norm": 31.125,
+      "learning_rate": 1.2266666666666666e-06,
+      "loss": 1.2651,
+      "step": 93
+    },
+    {
+      "epoch": 0.005776794493608653,
+      "grad_norm": 22.625,
+      "learning_rate": 1.2400000000000002e-06,
+      "loss": 1.2309,
+      "step": 94
+    },
+    {
+      "epoch": 0.005838249754178957,
+      "grad_norm": 26.375,
+      "learning_rate": 1.2533333333333333e-06,
+      "loss": 1.2152,
+      "step": 95
+    },
+    {
+      "epoch": 0.0058997050147492625,
+      "grad_norm": 31.75,
+      "learning_rate": 1.2666666666666669e-06,
+      "loss": 1.2495,
+      "step": 96
+    },
+    {
+      "epoch": 0.0059611602753195676,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.28e-06,
+      "loss": 1.2448,
+      "step": 97
+    },
+    {
+      "epoch": 0.006022615535889872,
+      "grad_norm": 15.875,
+      "learning_rate": 1.2933333333333334e-06,
+      "loss": 1.2347,
+      "step": 98
+    },
+    {
+      "epoch": 0.006084070796460177,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.3066666666666667e-06,
+      "loss": 1.2758,
+      "step": 99
+    },
+    {
+      "epoch": 0.006145526057030482,
+      "grad_norm": 15.25,
+      "learning_rate": 1.32e-06,
+      "loss": 1.3125,
+      "step": 100
+    },
+    {
+      "epoch": 0.006206981317600786,
+      "grad_norm": 28.375,
+      "learning_rate": 1.3333333333333334e-06,
+      "loss": 1.254,
+      "step": 101
+    },
+    {
+      "epoch": 0.006268436578171091,
+      "grad_norm": 21.0,
+      "learning_rate": 1.3466666666666668e-06,
+      "loss": 1.2636,
+      "step": 102
+    },
+    {
+      "epoch": 0.0063298918387413965,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.3600000000000001e-06,
+      "loss": 1.2687,
+      "step": 103
+    },
+    {
+      "epoch": 0.006391347099311701,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.3733333333333335e-06,
+      "loss": 1.2174,
+      "step": 104
+    },
+    {
+      "epoch": 0.006452802359882006,
+      "grad_norm": 18.75,
+      "learning_rate": 1.3866666666666668e-06,
+      "loss": 1.2783,
+      "step": 105
+    },
+    {
+      "epoch": 0.006514257620452311,
+      "grad_norm": 15.875,
+      "learning_rate": 1.4000000000000001e-06,
+      "loss": 1.2473,
+      "step": 106
+    },
+    {
+      "epoch": 0.006575712881022615,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.4133333333333335e-06,
+      "loss": 1.2454,
+      "step": 107
+    },
+    {
+      "epoch": 0.00663716814159292,
+      "grad_norm": 41.25,
+      "learning_rate": 1.4266666666666668e-06,
+      "loss": 1.3315,
+      "step": 108
+    },
+    {
+      "epoch": 0.006698623402163225,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.44e-06,
+      "loss": 1.2593,
+      "step": 109
+    },
+    {
+      "epoch": 0.00676007866273353,
+      "grad_norm": 16.75,
+      "learning_rate": 1.4533333333333335e-06,
+      "loss": 1.2575,
+      "step": 110
+    },
+    {
+      "epoch": 0.006821533923303835,
+      "grad_norm": 15.0,
+      "learning_rate": 1.4666666666666669e-06,
+      "loss": 1.265,
+      "step": 111
+    },
+    {
+      "epoch": 0.00688298918387414,
+      "grad_norm": 23.125,
+      "learning_rate": 1.48e-06,
+      "loss": 1.2893,
+      "step": 112
+    },
+    {
+      "epoch": 0.006944444444444444,
+      "grad_norm": 23.0,
+      "learning_rate": 1.4933333333333336e-06,
+      "loss": 1.2425,
+      "step": 113
+    },
+    {
+      "epoch": 0.007005899705014749,
+      "grad_norm": 18.375,
+      "learning_rate": 1.506666666666667e-06,
+      "loss": 1.3032,
+      "step": 114
+    },
+    {
+      "epoch": 0.007067354965585054,
+      "grad_norm": 24.625,
+      "learning_rate": 1.52e-06,
+      "loss": 1.2893,
+      "step": 115
+    },
+    {
+      "epoch": 0.0071288102261553585,
+      "grad_norm": 14.375,
+      "learning_rate": 1.5333333333333334e-06,
+      "loss": 1.1966,
+      "step": 116
+    },
+    {
+      "epoch": 0.007190265486725664,
+      "grad_norm": 21.0,
+      "learning_rate": 1.546666666666667e-06,
+      "loss": 1.23,
+      "step": 117
+    },
+    {
+      "epoch": 0.007251720747295969,
+      "grad_norm": 18.25,
+      "learning_rate": 1.56e-06,
+      "loss": 1.294,
+      "step": 118
+    },
+    {
+      "epoch": 0.007313176007866273,
+      "grad_norm": 18.125,
+      "learning_rate": 1.5733333333333334e-06,
+      "loss": 1.2289,
+      "step": 119
+    },
+    {
+      "epoch": 0.007374631268436578,
+      "grad_norm": 13.875,
+      "learning_rate": 1.586666666666667e-06,
+      "loss": 1.3106,
+      "step": 120
+    },
+    {
+      "epoch": 0.007436086529006883,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.6000000000000001e-06,
+      "loss": 1.238,
+      "step": 121
+    },
+    {
+      "epoch": 0.007497541789577187,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.6133333333333335e-06,
+      "loss": 1.2299,
+      "step": 122
+    },
+    {
+      "epoch": 0.0075589970501474925,
+      "grad_norm": 13.125,
+      "learning_rate": 1.6266666666666666e-06,
+      "loss": 1.2363,
+      "step": 123
+    },
+    {
+      "epoch": 0.007620452310717798,
+      "grad_norm": 10.1875,
+      "learning_rate": 1.6400000000000002e-06,
+      "loss": 1.1714,
+      "step": 124
+    },
+    {
+      "epoch": 0.007681907571288102,
+      "grad_norm": 19.75,
+      "learning_rate": 1.6533333333333335e-06,
+      "loss": 1.191,
+      "step": 125
+    },
+    {
+      "epoch": 0.007743362831858407,
+      "grad_norm": 18.625,
+      "learning_rate": 1.6666666666666667e-06,
+      "loss": 1.1915,
+      "step": 126
+    },
+    {
+      "epoch": 0.007804818092428712,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.6800000000000002e-06,
+      "loss": 1.1701,
+      "step": 127
+    },
+    {
+      "epoch": 0.007866273352999017,
+      "grad_norm": 19.625,
+      "learning_rate": 1.6933333333333336e-06,
+      "loss": 1.173,
+      "step": 128
+    },
+    {
+      "epoch": 0.007927728613569322,
+      "grad_norm": 10.625,
+      "learning_rate": 1.7066666666666667e-06,
+      "loss": 1.2696,
+      "step": 129
+    },
+    {
+      "epoch": 0.007989183874139626,
+      "grad_norm": 15.5625,
+      "learning_rate": 1.72e-06,
+      "loss": 1.2133,
+      "step": 130
+    },
+    {
+      "epoch": 0.00805063913470993,
+      "grad_norm": 13.0,
+      "learning_rate": 1.7333333333333336e-06,
+      "loss": 1.242,
+      "step": 131
+    },
+    {
+      "epoch": 0.008112094395280236,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.7466666666666667e-06,
+      "loss": 1.2084,
+      "step": 132
+    },
+    {
+      "epoch": 0.008173549655850541,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.76e-06,
+      "loss": 1.192,
+      "step": 133
+    },
+    {
+      "epoch": 0.008235004916420846,
+      "grad_norm": 14.625,
+      "learning_rate": 1.7733333333333336e-06,
+      "loss": 1.1556,
+      "step": 134
+    },
+    {
+      "epoch": 0.008296460176991151,
+      "grad_norm": 13.125,
+      "learning_rate": 1.7866666666666668e-06,
+      "loss": 1.2567,
+      "step": 135
+    },
+    {
+      "epoch": 0.008357915437561455,
+      "grad_norm": 23.5,
+      "learning_rate": 1.8000000000000001e-06,
+      "loss": 1.243,
+      "step": 136
+    },
+    {
+      "epoch": 0.00841937069813176,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.8133333333333337e-06,
+      "loss": 1.1579,
+      "step": 137
+    },
+    {
+      "epoch": 0.008480825958702065,
+      "grad_norm": 10.0,
+      "learning_rate": 1.8266666666666668e-06,
+      "loss": 1.0972,
+      "step": 138
+    },
+    {
+      "epoch": 0.00854228121927237,
+      "grad_norm": 20.875,
+      "learning_rate": 1.8400000000000002e-06,
+      "loss": 1.1399,
+      "step": 139
+    },
+    {
+      "epoch": 0.008603736479842675,
+      "grad_norm": 27.625,
+      "learning_rate": 1.8533333333333333e-06,
+      "loss": 1.1345,
+      "step": 140
+    },
+    {
+      "epoch": 0.00866519174041298,
+      "grad_norm": 9.1875,
+      "learning_rate": 1.8666666666666669e-06,
+      "loss": 1.1324,
+      "step": 141
+    },
+    {
+      "epoch": 0.008726647000983283,
+      "grad_norm": 15.625,
+      "learning_rate": 1.8800000000000002e-06,
+      "loss": 1.1115,
+      "step": 142
+    },
+    {
+      "epoch": 0.008788102261553589,
+      "grad_norm": 21.125,
+      "learning_rate": 1.8933333333333333e-06,
+      "loss": 1.0983,
+      "step": 143
+    },
+    {
+      "epoch": 0.008849557522123894,
+      "grad_norm": 32.0,
+      "learning_rate": 1.906666666666667e-06,
+      "loss": 1.1227,
+      "step": 144
+    },
+    {
+      "epoch": 0.008911012782694199,
+      "grad_norm": 20.125,
+      "learning_rate": 1.9200000000000003e-06,
+      "loss": 1.2015,
+      "step": 145
+    },
+    {
+      "epoch": 0.008972468043264504,
+      "grad_norm": 20.5,
+      "learning_rate": 1.9333333333333336e-06,
+      "loss": 1.2647,
+      "step": 146
+    },
+    {
+      "epoch": 0.009033923303834809,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9466666666666665e-06,
+      "loss": 1.2075,
+      "step": 147
+    },
+    {
+      "epoch": 0.009095378564405112,
+      "grad_norm": 17.0,
+      "learning_rate": 1.9600000000000003e-06,
+      "loss": 1.2263,
+      "step": 148
+    },
+    {
+      "epoch": 0.009156833824975417,
+      "grad_norm": 25.25,
+      "learning_rate": 1.9733333333333336e-06,
+      "loss": 1.313,
+      "step": 149
+    },
+    {
+      "epoch": 0.009218289085545723,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.9866666666666666e-06,
+      "loss": 1.1606,
+      "step": 150
+    },
+    {
+      "epoch": 0.009279744346116028,
+      "grad_norm": 47.5,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 1.3511,
+      "step": 151
+    },
+    {
+      "epoch": 0.009341199606686333,
+      "grad_norm": 10.6875,
+      "learning_rate": 2.0133333333333337e-06,
+      "loss": 1.2138,
+      "step": 152
+    },
+    {
+      "epoch": 0.009402654867256638,
+      "grad_norm": 14.125,
+      "learning_rate": 2.0266666666666666e-06,
+      "loss": 1.1739,
+      "step": 153
+    },
+    {
+      "epoch": 0.009464110127826941,
+      "grad_norm": 24.0,
+      "learning_rate": 2.04e-06,
+      "loss": 1.1725,
+      "step": 154
+    },
+    {
+      "epoch": 0.009525565388397246,
+      "grad_norm": 21.625,
+      "learning_rate": 2.0533333333333337e-06,
+      "loss": 1.1152,
+      "step": 155
+    },
+    {
+      "epoch": 0.009587020648967551,
+      "grad_norm": 13.375,
+      "learning_rate": 2.0666666666666666e-06,
+      "loss": 1.1109,
+      "step": 156
+    },
+    {
+      "epoch": 0.009648475909537857,
+      "grad_norm": 20.25,
+      "learning_rate": 2.08e-06,
+      "loss": 1.1053,
+      "step": 157
+    },
+    {
+      "epoch": 0.009709931170108162,
+      "grad_norm": 10.375,
+      "learning_rate": 2.0933333333333338e-06,
+      "loss": 1.1225,
+      "step": 158
+    },
+    {
+      "epoch": 0.009771386430678467,
+      "grad_norm": 11.4375,
+      "learning_rate": 2.1066666666666667e-06,
+      "loss": 1.1424,
+      "step": 159
+    },
+    {
+      "epoch": 0.00983284169124877,
+      "grad_norm": 28.5,
+      "learning_rate": 2.12e-06,
+      "loss": 1.1637,
+      "step": 160
+    },
+    {
+      "epoch": 0.009894296951819075,
+      "grad_norm": 19.5,
+      "learning_rate": 2.133333333333334e-06,
+      "loss": 1.1104,
+      "step": 161
+    },
+    {
+      "epoch": 0.00995575221238938,
+      "grad_norm": 30.75,
+      "learning_rate": 2.1466666666666667e-06,
+      "loss": 1.1314,
+      "step": 162
+    },
+    {
+      "epoch": 0.010017207472959685,
+      "grad_norm": 13.625,
+      "learning_rate": 2.16e-06,
+      "loss": 1.0777,
+      "step": 163
+    },
+    {
+      "epoch": 0.01007866273352999,
+      "grad_norm": 14.125,
+      "learning_rate": 2.1733333333333334e-06,
+      "loss": 1.1029,
+      "step": 164
+    },
+    {
+      "epoch": 0.010140117994100296,
+      "grad_norm": 28.625,
+      "learning_rate": 2.1866666666666668e-06,
+      "loss": 1.0845,
+      "step": 165
+    },
+    {
+      "epoch": 0.010201573254670599,
+      "grad_norm": 16.375,
+      "learning_rate": 2.2e-06,
+      "loss": 1.1821,
+      "step": 166
+    },
+    {
+      "epoch": 0.010263028515240904,
+      "grad_norm": 21.25,
+      "learning_rate": 2.2133333333333335e-06,
+      "loss": 1.141,
+      "step": 167
+    },
+    {
+      "epoch": 0.01032448377581121,
+      "grad_norm": 17.0,
+      "learning_rate": 2.226666666666667e-06,
+      "loss": 1.1344,
+      "step": 168
+    },
+    {
+      "epoch": 0.010385939036381514,
+      "grad_norm": 13.125,
+      "learning_rate": 2.24e-06,
+      "loss": 1.1131,
+      "step": 169
+    },
+    {
+      "epoch": 0.01044739429695182,
+      "grad_norm": 16.125,
+      "learning_rate": 2.2533333333333335e-06,
+      "loss": 1.0608,
+      "step": 170
+    },
+    {
+      "epoch": 0.010508849557522125,
+      "grad_norm": 15.0,
+      "learning_rate": 2.266666666666667e-06,
+      "loss": 1.1074,
+      "step": 171
+    },
+    {
+      "epoch": 0.010570304818092428,
+      "grad_norm": 16.75,
+      "learning_rate": 2.28e-06,
+      "loss": 1.1097,
+      "step": 172
+    },
+    {
+      "epoch": 0.010631760078662733,
+      "grad_norm": 12.4375,
+      "learning_rate": 2.2933333333333335e-06,
+      "loss": 1.1263,
+      "step": 173
+    },
+    {
+      "epoch": 0.010693215339233038,
+      "grad_norm": 36.25,
+      "learning_rate": 2.306666666666667e-06,
+      "loss": 1.1327,
+      "step": 174
+    },
+    {
+      "epoch": 0.010754670599803343,
+      "grad_norm": 13.375,
+      "learning_rate": 2.3200000000000002e-06,
+      "loss": 1.1072,
+      "step": 175
+    },
+    {
+      "epoch": 0.010816125860373648,
+      "grad_norm": 16.875,
+      "learning_rate": 2.3333333333333336e-06,
+      "loss": 1.1052,
+      "step": 176
+    },
+    {
+      "epoch": 0.010877581120943954,
+      "grad_norm": 12.5625,
+      "learning_rate": 2.346666666666667e-06,
+      "loss": 1.1051,
+      "step": 177
+    },
+    {
+      "epoch": 0.010939036381514257,
+      "grad_norm": 19.0,
+      "learning_rate": 2.3600000000000003e-06,
+      "loss": 1.2418,
+      "step": 178
+    },
+    {
+      "epoch": 0.011000491642084562,
+      "grad_norm": 18.25,
+      "learning_rate": 2.3733333333333336e-06,
+      "loss": 1.0645,
+      "step": 179
+    },
+    {
+      "epoch": 0.011061946902654867,
+      "grad_norm": 9.25,
+      "learning_rate": 2.386666666666667e-06,
+      "loss": 1.1151,
+      "step": 180
+    },
+    {
+      "epoch": 0.011123402163225172,
+      "grad_norm": 34.25,
+      "learning_rate": 2.4000000000000003e-06,
+      "loss": 1.1364,
+      "step": 181
+    },
+    {
+      "epoch": 0.011184857423795477,
+      "grad_norm": 21.375,
+      "learning_rate": 2.4133333333333337e-06,
+      "loss": 1.1714,
+      "step": 182
+    },
+    {
+      "epoch": 0.011246312684365782,
+      "grad_norm": 17.875,
+      "learning_rate": 2.426666666666667e-06,
+      "loss": 1.1789,
+      "step": 183
+    },
+    {
+      "epoch": 0.011307767944936086,
+      "grad_norm": 39.0,
+      "learning_rate": 2.4400000000000004e-06,
+      "loss": 1.0766,
+      "step": 184
+    },
+    {
+      "epoch": 0.011369223205506391,
+      "grad_norm": 18.25,
+      "learning_rate": 2.4533333333333333e-06,
+      "loss": 1.2983,
+      "step": 185
+    },
+    {
+      "epoch": 0.011430678466076696,
+      "grad_norm": 12.375,
+      "learning_rate": 2.466666666666667e-06,
+      "loss": 1.1043,
+      "step": 186
+    },
+    {
+      "epoch": 0.011492133726647001,
+      "grad_norm": 22.5,
+      "learning_rate": 2.4800000000000004e-06,
+      "loss": 1.1895,
+      "step": 187
+    },
+    {
+      "epoch": 0.011553588987217306,
+      "grad_norm": 21.75,
+      "learning_rate": 2.4933333333333333e-06,
+      "loss": 1.1812,
+      "step": 188
+    },
+    {
+      "epoch": 0.011615044247787611,
+      "grad_norm": 11.1875,
+      "learning_rate": 2.5066666666666667e-06,
+      "loss": 1.1458,
+      "step": 189
+    },
+    {
+      "epoch": 0.011676499508357915,
+      "grad_norm": 20.625,
+      "learning_rate": 2.52e-06,
+      "loss": 1.3059,
+      "step": 190
+    },
+    {
+      "epoch": 0.01173795476892822,
+      "grad_norm": 16.625,
+      "learning_rate": 2.5333333333333338e-06,
+      "loss": 1.1491,
+      "step": 191
+    },
+    {
+      "epoch": 0.011799410029498525,
+      "grad_norm": 17.25,
+      "learning_rate": 2.5466666666666667e-06,
+      "loss": 1.1945,
+      "step": 192
+    },
+    {
+      "epoch": 0.01186086529006883,
+      "grad_norm": 23.375,
+      "learning_rate": 2.56e-06,
+      "loss": 1.1187,
+      "step": 193
+    },
+    {
+      "epoch": 0.011922320550639135,
+      "grad_norm": 32.5,
+      "learning_rate": 2.573333333333334e-06,
+      "loss": 1.1354,
+      "step": 194
+    },
+    {
+      "epoch": 0.01198377581120944,
+      "grad_norm": 18.875,
+      "learning_rate": 2.5866666666666667e-06,
+      "loss": 1.1791,
+      "step": 195
+    },
+    {
+      "epoch": 0.012045231071779744,
+      "grad_norm": 20.625,
+      "learning_rate": 2.6e-06,
+      "loss": 1.1322,
+      "step": 196
+    },
+    {
+      "epoch": 0.012106686332350049,
+      "grad_norm": 32.0,
+      "learning_rate": 2.6133333333333334e-06,
+      "loss": 1.1313,
+      "step": 197
+    },
+    {
+      "epoch": 0.012168141592920354,
+      "grad_norm": 45.75,
+      "learning_rate": 2.6266666666666668e-06,
+      "loss": 1.1602,
+      "step": 198
+    },
+    {
+      "epoch": 0.012229596853490659,
+      "grad_norm": 14.375,
+      "learning_rate": 2.64e-06,
+      "loss": 1.1752,
+      "step": 199
+    },
+    {
+      "epoch": 0.012291052114060964,
+      "grad_norm": 30.875,
+      "learning_rate": 2.6533333333333335e-06,
+      "loss": 1.2652,
+      "step": 200
+    },
+    {
+      "epoch": 0.012352507374631269,
+      "grad_norm": 23.75,
+      "learning_rate": 2.666666666666667e-06,
+      "loss": 1.1371,
+      "step": 201
+    },
+    {
+      "epoch": 0.012413962635201573,
+      "grad_norm": 21.25,
+      "learning_rate": 2.68e-06,
+      "loss": 1.0729,
+      "step": 202
+    },
+    {
+      "epoch": 0.012475417895771878,
+      "grad_norm": 17.0,
+      "learning_rate": 2.6933333333333335e-06,
+      "loss": 1.0273,
+      "step": 203
+    },
+    {
+      "epoch": 0.012536873156342183,
+      "grad_norm": 20.0,
+      "learning_rate": 2.706666666666667e-06,
+      "loss": 1.1027,
+      "step": 204
+    },
+    {
+      "epoch": 0.012598328416912488,
+      "grad_norm": 21.625,
+      "learning_rate": 2.7200000000000002e-06,
+      "loss": 1.149,
+      "step": 205
+    },
+    {
+      "epoch": 0.012659783677482793,
+      "grad_norm": 19.375,
+      "learning_rate": 2.7333333333333336e-06,
+      "loss": 1.0611,
+      "step": 206
+    },
+    {
+      "epoch": 0.012721238938053098,
+      "grad_norm": 17.125,
+      "learning_rate": 2.746666666666667e-06,
+      "loss": 1.055,
+      "step": 207
+    },
+    {
+      "epoch": 0.012782694198623401,
+      "grad_norm": 9.9375,
+      "learning_rate": 2.7600000000000003e-06,
+      "loss": 1.0826,
+      "step": 208
+    },
+    {
+      "epoch": 0.012844149459193707,
+      "grad_norm": 39.25,
+      "learning_rate": 2.7733333333333336e-06,
+      "loss": 1.0489,
+      "step": 209
+    },
+    {
+      "epoch": 0.012905604719764012,
+      "grad_norm": 11.1875,
+      "learning_rate": 2.786666666666667e-06,
+      "loss": 1.1321,
+      "step": 210
+    },
+    {
+      "epoch": 0.012967059980334317,
+      "grad_norm": 24.75,
+      "learning_rate": 2.8000000000000003e-06,
+      "loss": 1.1996,
+      "step": 211
+    },
+    {
+      "epoch": 0.013028515240904622,
+      "grad_norm": 21.875,
+      "learning_rate": 2.8133333333333336e-06,
+      "loss": 1.1491,
+      "step": 212
+    },
+    {
+      "epoch": 0.013089970501474927,
+      "grad_norm": 20.375,
+      "learning_rate": 2.826666666666667e-06,
+      "loss": 1.0555,
+      "step": 213
+    },
+    {
+      "epoch": 0.01315142576204523,
+      "grad_norm": 25.125,
+      "learning_rate": 2.84e-06,
+      "loss": 1.1075,
+      "step": 214
+    },
+    {
+      "epoch": 0.013212881022615535,
+      "grad_norm": 10.125,
+      "learning_rate": 2.8533333333333337e-06,
+      "loss": 1.0926,
+      "step": 215
+    },
+    {
+      "epoch": 0.01327433628318584,
+      "grad_norm": 17.375,
+      "learning_rate": 2.866666666666667e-06,
+      "loss": 1.1172,
+      "step": 216
+    },
+    {
+      "epoch": 0.013335791543756146,
+      "grad_norm": 30.875,
+      "learning_rate": 2.88e-06,
+      "loss": 1.1143,
+      "step": 217
+    },
+    {
+      "epoch": 0.01339724680432645,
+      "grad_norm": 14.875,
+      "learning_rate": 2.8933333333333337e-06,
+      "loss": 1.1945,
+      "step": 218
+    },
+    {
+      "epoch": 0.013458702064896756,
+      "grad_norm": 20.0,
+      "learning_rate": 2.906666666666667e-06,
+      "loss": 1.111,
+      "step": 219
+    },
+    {
+      "epoch": 0.01352015732546706,
+      "grad_norm": 12.25,
+      "learning_rate": 2.92e-06,
+      "loss": 1.1282,
+      "step": 220
+    },
+    {
+      "epoch": 0.013581612586037364,
+      "grad_norm": 29.5,
+      "learning_rate": 2.9333333333333338e-06,
+      "loss": 1.0827,
+      "step": 221
+    },
+    {
+      "epoch": 0.01364306784660767,
+      "grad_norm": 10.0625,
+      "learning_rate": 2.946666666666667e-06,
+      "loss": 1.1045,
+      "step": 222
+    },
+    {
+      "epoch": 0.013704523107177975,
+      "grad_norm": 40.5,
+      "learning_rate": 2.96e-06,
+      "loss": 1.1818,
+      "step": 223
+    },
+    {
+      "epoch": 0.01376597836774828,
+      "grad_norm": 38.25,
+      "learning_rate": 2.973333333333334e-06,
+      "loss": 1.1809,
+      "step": 224
+    },
+    {
+      "epoch": 0.013827433628318585,
+      "grad_norm": 20.5,
+      "learning_rate": 2.986666666666667e-06,
+      "loss": 1.0895,
+      "step": 225
+    },
+    {
+      "epoch": 0.013888888888888888,
+      "grad_norm": 13.5,
+      "learning_rate": 3e-06,
+      "loss": 1.0573,
+      "step": 226
+    },
+    {
+      "epoch": 0.013950344149459193,
+      "grad_norm": 22.625,
+      "learning_rate": 3.013333333333334e-06,
+      "loss": 1.1354,
+      "step": 227
+    },
+    {
+      "epoch": 0.014011799410029498,
+      "grad_norm": 21.75,
+      "learning_rate": 3.0266666666666668e-06,
+      "loss": 1.0277,
+      "step": 228
+    },
+    {
+      "epoch": 0.014073254670599803,
+      "grad_norm": 15.375,
+      "learning_rate": 3.04e-06,
+      "loss": 1.0699,
+      "step": 229
+    },
+    {
+      "epoch": 0.014134709931170109,
+      "grad_norm": 28.0,
+      "learning_rate": 3.053333333333334e-06,
+      "loss": 1.0487,
+      "step": 230
+    },
+    {
+      "epoch": 0.014196165191740414,
+      "grad_norm": 43.75,
+      "learning_rate": 3.066666666666667e-06,
+      "loss": 1.0535,
+      "step": 231
+    },
+    {
+      "epoch": 0.014257620452310717,
+      "grad_norm": 9.5,
+      "learning_rate": 3.08e-06,
+      "loss": 1.0433,
+      "step": 232
+    },
+    {
+      "epoch": 0.014319075712881022,
+      "grad_norm": 35.5,
+      "learning_rate": 3.093333333333334e-06,
+      "loss": 1.2087,
+      "step": 233
+    },
+    {
+      "epoch": 0.014380530973451327,
+      "grad_norm": 34.25,
+      "learning_rate": 3.106666666666667e-06,
+      "loss": 1.25,
+      "step": 234
+    },
+    {
+      "epoch": 0.014441986234021632,
+      "grad_norm": 9.75,
+      "learning_rate": 3.12e-06,
+      "loss": 1.1464,
+      "step": 235
+    },
+    {
+      "epoch": 0.014503441494591937,
+      "grad_norm": 12.25,
+      "learning_rate": 3.133333333333334e-06,
+      "loss": 1.1258,
+      "step": 236
+    },
+    {
+      "epoch": 0.014564896755162243,
+      "grad_norm": 13.75,
+      "learning_rate": 3.146666666666667e-06,
+      "loss": 1.2378,
+      "step": 237
+    },
+    {
+      "epoch": 0.014626352015732546,
+      "grad_norm": 52.75,
+      "learning_rate": 3.1600000000000002e-06,
+      "loss": 1.2047,
+      "step": 238
+    },
+    {
+      "epoch": 0.014687807276302851,
+      "grad_norm": 12.25,
+      "learning_rate": 3.173333333333334e-06,
+      "loss": 1.0881,
+      "step": 239
+    },
+    {
+      "epoch": 0.014749262536873156,
+      "grad_norm": 13.3125,
+      "learning_rate": 3.186666666666667e-06,
+      "loss": 1.1176,
+      "step": 240
+    },
+    {
+      "epoch": 0.014810717797443461,
+      "grad_norm": 24.125,
+      "learning_rate": 3.2000000000000003e-06,
+      "loss": 1.1777,
+      "step": 241
+    },
+    {
+      "epoch": 0.014872173058013766,
+      "grad_norm": 15.6875,
+      "learning_rate": 3.213333333333334e-06,
+      "loss": 1.2388,
+      "step": 242
+    },
+    {
+      "epoch": 0.014933628318584071,
+      "grad_norm": 37.0,
+      "learning_rate": 3.226666666666667e-06,
+      "loss": 1.1639,
+      "step": 243
+    },
+    {
+      "epoch": 0.014995083579154375,
+      "grad_norm": 26.375,
+      "learning_rate": 3.2400000000000003e-06,
+      "loss": 1.2209,
+      "step": 244
+    },
+    {
+      "epoch": 0.01505653883972468,
+      "grad_norm": 13.0,
+      "learning_rate": 3.2533333333333332e-06,
+      "loss": 1.1995,
+      "step": 245
+    },
+    {
+      "epoch": 0.015117994100294985,
+      "grad_norm": 18.625,
+      "learning_rate": 3.266666666666667e-06,
+      "loss": 1.1579,
+      "step": 246
+    },
+    {
+      "epoch": 0.01517944936086529,
+      "grad_norm": 19.875,
+      "learning_rate": 3.2800000000000004e-06,
+      "loss": 1.1907,
+      "step": 247
+    },
+    {
+      "epoch": 0.015240904621435595,
+      "grad_norm": 18.375,
+      "learning_rate": 3.2933333333333333e-06,
+      "loss": 1.2377,
+      "step": 248
+    },
+    {
+      "epoch": 0.0153023598820059,
+      "grad_norm": 16.5,
+      "learning_rate": 3.306666666666667e-06,
+      "loss": 1.0934,
+      "step": 249
+    },
+    {
+      "epoch": 0.015363815142576204,
+      "grad_norm": 15.0,
+      "learning_rate": 3.3200000000000004e-06,
+      "loss": 1.0654,
+      "step": 250
+    },
+    {
+      "epoch": 0.015425270403146509,
+      "grad_norm": 20.625,
+      "learning_rate": 3.3333333333333333e-06,
+      "loss": 1.1563,
+      "step": 251
+    },
+    {
+      "epoch": 0.015486725663716814,
+      "grad_norm": 16.625,
+      "learning_rate": 3.346666666666667e-06,
+      "loss": 1.099,
+      "step": 252
+    },
+    {
+      "epoch": 0.015548180924287119,
+      "grad_norm": 12.6875,
+      "learning_rate": 3.3600000000000004e-06,
+      "loss": 1.0963,
+      "step": 253
+    },
+    {
+      "epoch": 0.015609636184857424,
+      "grad_norm": 18.625,
+      "learning_rate": 3.3733333333333334e-06,
+      "loss": 1.0406,
+      "step": 254
+    },
+    {
+      "epoch": 0.015671091445427728,
+      "grad_norm": 8.5625,
+      "learning_rate": 3.386666666666667e-06,
+      "loss": 1.0701,
+      "step": 255
+    },
+    {
+      "epoch": 0.015732546705998034,
+      "grad_norm": 14.1875,
+      "learning_rate": 3.4000000000000005e-06,
+      "loss": 1.077,
+      "step": 256
+    },
+    {
+      "epoch": 0.015794001966568338,
+      "grad_norm": 17.75,
+      "learning_rate": 3.4133333333333334e-06,
+      "loss": 1.1132,
+      "step": 257
+    },
+    {
+      "epoch": 0.015855457227138645,
+      "grad_norm": 23.25,
+      "learning_rate": 3.426666666666667e-06,
+      "loss": 1.1191,
+      "step": 258
+    },
+    {
+      "epoch": 0.015916912487708948,
+      "grad_norm": 26.625,
+      "learning_rate": 3.44e-06,
+      "loss": 1.322,
+      "step": 259
+    },
+    {
+      "epoch": 0.01597836774827925,
+      "grad_norm": 14.6875,
+      "learning_rate": 3.4533333333333334e-06,
+      "loss": 1.1019,
+      "step": 260
+    },
+    {
+      "epoch": 0.016039823008849558,
+      "grad_norm": 11.5,
+      "learning_rate": 3.4666666666666672e-06,
+      "loss": 1.1694,
+      "step": 261
+    },
+    {
+      "epoch": 0.01610127826941986,
+      "grad_norm": 17.25,
+      "learning_rate": 3.48e-06,
+      "loss": 1.1177,
+      "step": 262
+    },
+    {
+      "epoch": 0.01616273352999017,
+      "grad_norm": 13.25,
+      "learning_rate": 3.4933333333333335e-06,
+      "loss": 1.0603,
+      "step": 263
+    },
+    {
+      "epoch": 0.016224188790560472,
+      "grad_norm": 23.5,
+      "learning_rate": 3.5066666666666673e-06,
+      "loss": 1.1767,
+      "step": 264
+    },
+    {
+      "epoch": 0.016285644051130775,
+      "grad_norm": 35.0,
+      "learning_rate": 3.52e-06,
+      "loss": 1.2425,
+      "step": 265
+    },
+    {
+      "epoch": 0.016347099311701082,
+      "grad_norm": 10.875,
+      "learning_rate": 3.5333333333333335e-06,
+      "loss": 1.1201,
+      "step": 266
+    },
+    {
+      "epoch": 0.016408554572271385,
+      "grad_norm": 11.0,
+      "learning_rate": 3.5466666666666673e-06,
+      "loss": 1.1221,
+      "step": 267
+    },
+    {
+      "epoch": 0.016470009832841692,
+      "grad_norm": 16.5,
+      "learning_rate": 3.5600000000000002e-06,
+      "loss": 1.27,
+      "step": 268
+    },
+    {
+      "epoch": 0.016531465093411996,
+      "grad_norm": 13.375,
+      "learning_rate": 3.5733333333333336e-06,
+      "loss": 1.1907,
+      "step": 269
+    },
+    {
+      "epoch": 0.016592920353982302,
+      "grad_norm": 11.125,
+      "learning_rate": 3.5866666666666673e-06,
+      "loss": 1.1787,
+      "step": 270
+    },
+    {
+      "epoch": 0.016654375614552606,
+      "grad_norm": 22.25,
+      "learning_rate": 3.6000000000000003e-06,
+      "loss": 1.232,
+      "step": 271
+    },
+    {
+      "epoch": 0.01671583087512291,
+      "grad_norm": 32.75,
+      "learning_rate": 3.6133333333333336e-06,
+      "loss": 1.1656,
+      "step": 272
+    },
+    {
+      "epoch": 0.016777286135693216,
+      "grad_norm": 17.25,
+      "learning_rate": 3.6266666666666674e-06,
+      "loss": 1.1151,
+      "step": 273
+    },
+    {
+      "epoch": 0.01683874139626352,
+      "grad_norm": 19.0,
+      "learning_rate": 3.6400000000000003e-06,
+      "loss": 1.0429,
+      "step": 274
+    },
+    {
+      "epoch": 0.016900196656833826,
+      "grad_norm": 16.75,
+      "learning_rate": 3.6533333333333336e-06,
+      "loss": 1.1054,
+      "step": 275
+    },
+    {
+      "epoch": 0.01696165191740413,
+      "grad_norm": 16.125,
+      "learning_rate": 3.6666666666666666e-06,
+      "loss": 1.1221,
+      "step": 276
+    },
+    {
+      "epoch": 0.017023107177974433,
+      "grad_norm": 19.375,
+      "learning_rate": 3.6800000000000003e-06,
+      "loss": 1.1401,
+      "step": 277
+    },
+    {
+      "epoch": 0.01708456243854474,
+      "grad_norm": 17.375,
+      "learning_rate": 3.6933333333333337e-06,
+      "loss": 1.1432,
+      "step": 278
+    },
+    {
+      "epoch": 0.017146017699115043,
+      "grad_norm": 25.75,
+      "learning_rate": 3.7066666666666666e-06,
+      "loss": 1.0428,
+      "step": 279
+    },
+    {
+      "epoch": 0.01720747295968535,
+      "grad_norm": 15.6875,
+      "learning_rate": 3.7200000000000004e-06,
+      "loss": 1.0607,
+      "step": 280
+    },
+    {
+      "epoch": 0.017268928220255653,
+      "grad_norm": 10.0,
+      "learning_rate": 3.7333333333333337e-06,
+      "loss": 1.1018,
+      "step": 281
+    },
+    {
+      "epoch": 0.01733038348082596,
+      "grad_norm": 22.875,
+      "learning_rate": 3.7466666666666667e-06,
+      "loss": 1.1253,
+      "step": 282
+    },
+    {
+      "epoch": 0.017391838741396264,
+      "grad_norm": 15.8125,
+      "learning_rate": 3.7600000000000004e-06,
+      "loss": 1.1724,
+      "step": 283
+    },
+    {
+      "epoch": 0.017453294001966567,
+      "grad_norm": 34.25,
+      "learning_rate": 3.7733333333333338e-06,
+      "loss": 1.1102,
+      "step": 284
+    },
+    {
+      "epoch": 0.017514749262536874,
+      "grad_norm": 21.375,
+      "learning_rate": 3.7866666666666667e-06,
+      "loss": 1.1293,
+      "step": 285
+    },
+    {
+      "epoch": 0.017576204523107177,
+      "grad_norm": 29.5,
+      "learning_rate": 3.8000000000000005e-06,
+      "loss": 1.2461,
+      "step": 286
+    },
+    {
+      "epoch": 0.017637659783677484,
+      "grad_norm": 38.75,
+      "learning_rate": 3.813333333333334e-06,
+      "loss": 1.1235,
+      "step": 287
+    },
+    {
+      "epoch": 0.017699115044247787,
+      "grad_norm": 9.3125,
+      "learning_rate": 3.826666666666667e-06,
+      "loss": 1.1524,
+      "step": 288
+    },
+    {
+      "epoch": 0.01776057030481809,
+      "grad_norm": 21.75,
+      "learning_rate": 3.8400000000000005e-06,
+      "loss": 1.0407,
+      "step": 289
+    },
+    {
+      "epoch": 0.017822025565388398,
+      "grad_norm": 13.8125,
+      "learning_rate": 3.853333333333334e-06,
+      "loss": 1.0842,
+      "step": 290
+    },
+    {
+      "epoch": 0.0178834808259587,
+      "grad_norm": 11.625,
+      "learning_rate": 3.866666666666667e-06,
+      "loss": 1.0728,
+      "step": 291
+    },
+    {
+      "epoch": 0.017944936086529008,
+      "grad_norm": 13.3125,
+      "learning_rate": 3.88e-06,
+      "loss": 1.0725,
+      "step": 292
+    },
+    {
+      "epoch": 0.01800639134709931,
+      "grad_norm": 36.5,
+      "learning_rate": 3.893333333333333e-06,
+      "loss": 1.0338,
+      "step": 293
+    },
+    {
+      "epoch": 0.018067846607669618,
+      "grad_norm": 11.75,
+      "learning_rate": 3.906666666666667e-06,
+      "loss": 1.0135,
+      "step": 294
+    },
+    {
+      "epoch": 0.01812930186823992,
+      "grad_norm": 27.25,
+      "learning_rate": 3.920000000000001e-06,
+      "loss": 1.0667,
+      "step": 295
+    },
+    {
+      "epoch": 0.018190757128810225,
+      "grad_norm": 15.25,
+      "learning_rate": 3.9333333333333335e-06,
+      "loss": 1.068,
+      "step": 296
+    },
+    {
+      "epoch": 0.01825221238938053,
+      "grad_norm": 46.75,
+      "learning_rate": 3.946666666666667e-06,
+      "loss": 1.2074,
+      "step": 297
+    },
+    {
+      "epoch": 0.018313667649950835,
+      "grad_norm": 24.125,
+      "learning_rate": 3.96e-06,
+      "loss": 1.1072,
+      "step": 298
+    },
+    {
+      "epoch": 0.018375122910521142,
+      "grad_norm": 41.75,
+      "learning_rate": 3.973333333333333e-06,
+      "loss": 1.2243,
+      "step": 299
+    },
+    {
+      "epoch": 0.018436578171091445,
+      "grad_norm": 36.0,
+      "learning_rate": 3.986666666666667e-06,
+      "loss": 1.1642,
+      "step": 300
+    },
+    {
+      "epoch": 0.01849803343166175,
+      "grad_norm": 22.875,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 1.1821,
+      "step": 301
+    },
+    {
+      "epoch": 0.018559488692232055,
+      "grad_norm": 20.625,
+      "learning_rate": 4.013333333333334e-06,
+      "loss": 1.2046,
+      "step": 302
+    },
+    {
+      "epoch": 0.01862094395280236,
+      "grad_norm": 22.625,
+      "learning_rate": 4.026666666666667e-06,
+      "loss": 1.2256,
+      "step": 303
+    },
+    {
+      "epoch": 0.018682399213372666,
+      "grad_norm": 20.0,
+      "learning_rate": 4.04e-06,
+      "loss": 1.2917,
+      "step": 304
+    },
+    {
+      "epoch": 0.01874385447394297,
+      "grad_norm": 34.0,
+      "learning_rate": 4.053333333333333e-06,
+      "loss": 1.1289,
+      "step": 305
+    },
+    {
+      "epoch": 0.018805309734513276,
+      "grad_norm": 37.0,
+      "learning_rate": 4.066666666666667e-06,
+      "loss": 1.1572,
+      "step": 306
+    },
+    {
+      "epoch": 0.01886676499508358,
+      "grad_norm": 24.375,
+      "learning_rate": 4.08e-06,
+      "loss": 1.106,
+      "step": 307
+    },
+    {
+      "epoch": 0.018928220255653883,
+      "grad_norm": 27.0,
+      "learning_rate": 4.093333333333334e-06,
+      "loss": 1.1808,
+      "step": 308
+    },
+    {
+      "epoch": 0.01898967551622419,
+      "grad_norm": 13.5625,
+      "learning_rate": 4.1066666666666674e-06,
+      "loss": 1.1536,
+      "step": 309
+    },
+    {
+      "epoch": 0.019051130776794493,
+      "grad_norm": 25.75,
+      "learning_rate": 4.12e-06,
+      "loss": 1.1452,
+      "step": 310
+    },
+    {
+      "epoch": 0.0191125860373648,
+      "grad_norm": 41.5,
+      "learning_rate": 4.133333333333333e-06,
+      "loss": 1.0979,
+      "step": 311
+    },
+    {
+      "epoch": 0.019174041297935103,
+      "grad_norm": 11.6875,
+      "learning_rate": 4.146666666666667e-06,
+      "loss": 1.1598,
+      "step": 312
+    },
+    {
+      "epoch": 0.019235496558505406,
+      "grad_norm": 18.375,
+      "learning_rate": 4.16e-06,
+      "loss": 1.0453,
+      "step": 313
+    },
+    {
+      "epoch": 0.019296951819075713,
+      "grad_norm": 21.75,
+      "learning_rate": 4.173333333333334e-06,
+      "loss": 1.0437,
+      "step": 314
+    },
+    {
+      "epoch": 0.019358407079646017,
+      "grad_norm": 13.375,
+      "learning_rate": 4.1866666666666675e-06,
+      "loss": 1.0363,
+      "step": 315
+    },
+    {
+      "epoch": 0.019419862340216323,
+      "grad_norm": 8.125,
+      "learning_rate": 4.2000000000000004e-06,
+      "loss": 1.0984,
+      "step": 316
+    },
+    {
+      "epoch": 0.019481317600786627,
+      "grad_norm": 30.125,
+      "learning_rate": 4.213333333333333e-06,
+      "loss": 1.0599,
+      "step": 317
+    },
+    {
+      "epoch": 0.019542772861356934,
+      "grad_norm": 15.375,
+      "learning_rate": 4.226666666666667e-06,
+      "loss": 1.076,
+      "step": 318
+    },
+    {
+      "epoch": 0.019604228121927237,
+      "grad_norm": 19.375,
+      "learning_rate": 4.24e-06,
+      "loss": 1.0374,
+      "step": 319
+    },
+    {
+      "epoch": 0.01966568338249754,
+      "grad_norm": 11.25,
+      "learning_rate": 4.253333333333334e-06,
+      "loss": 1.0459,
+      "step": 320
+    },
+    {
+      "epoch": 0.019727138643067847,
+      "grad_norm": 8.625,
+      "learning_rate": 4.266666666666668e-06,
+      "loss": 1.0229,
+      "step": 321
+    },
+    {
+      "epoch": 0.01978859390363815,
+      "grad_norm": 19.25,
+      "learning_rate": 4.2800000000000005e-06,
+      "loss": 1.0628,
+      "step": 322
+    },
+    {
+      "epoch": 0.019850049164208457,
+      "grad_norm": 14.4375,
+      "learning_rate": 4.2933333333333334e-06,
+      "loss": 1.0568,
+      "step": 323
+    },
+    {
+      "epoch": 0.01991150442477876,
+      "grad_norm": 9.8125,
+      "learning_rate": 4.306666666666666e-06,
+      "loss": 1.037,
+      "step": 324
+    },
+    {
+      "epoch": 0.019972959685349064,
+      "grad_norm": 26.125,
+      "learning_rate": 4.32e-06,
+      "loss": 1.0578,
+      "step": 325
+    },
+    {
+      "epoch": 0.02003441494591937,
+      "grad_norm": 11.8125,
+      "learning_rate": 4.333333333333334e-06,
+      "loss": 1.0025,
+      "step": 326
+    },
+    {
+      "epoch": 0.020095870206489674,
+      "grad_norm": 13.375,
+      "learning_rate": 4.346666666666667e-06,
+      "loss": 1.0442,
+      "step": 327
+    },
+    {
+      "epoch": 0.02015732546705998,
+      "grad_norm": 21.25,
+      "learning_rate": 4.360000000000001e-06,
+      "loss": 0.9953,
+      "step": 328
+    },
+    {
+      "epoch": 0.020218780727630285,
+      "grad_norm": 11.75,
+      "learning_rate": 4.3733333333333335e-06,
+      "loss": 1.1609,
+      "step": 329
+    },
+    {
+      "epoch": 0.02028023598820059,
+      "grad_norm": 23.75,
+      "learning_rate": 4.3866666666666665e-06,
+      "loss": 1.1098,
+      "step": 330
+    },
+    {
+      "epoch": 0.020341691248770895,
+      "grad_norm": 37.25,
+      "learning_rate": 4.4e-06,
+      "loss": 1.0725,
+      "step": 331
+    },
+    {
+      "epoch": 0.020403146509341198,
+      "grad_norm": 32.25,
+      "learning_rate": 4.413333333333334e-06,
+      "loss": 1.0925,
+      "step": 332
+    },
+    {
+      "epoch": 0.020464601769911505,
+      "grad_norm": 14.1875,
+      "learning_rate": 4.426666666666667e-06,
+      "loss": 1.0454,
+      "step": 333
+    },
+    {
+      "epoch": 0.02052605703048181,
+      "grad_norm": 12.375,
+      "learning_rate": 4.440000000000001e-06,
+      "loss": 1.1047,
+      "step": 334
+    },
+    {
+      "epoch": 0.020587512291052115,
+      "grad_norm": 22.125,
+      "learning_rate": 4.453333333333334e-06,
+      "loss": 1.096,
+      "step": 335
+    },
+    {
+      "epoch": 0.02064896755162242,
+      "grad_norm": 14.875,
+      "learning_rate": 4.4666666666666665e-06,
+      "loss": 1.1137,
+      "step": 336
+    },
+    {
+      "epoch": 0.020710422812192722,
+      "grad_norm": 20.75,
+      "learning_rate": 4.48e-06,
+      "loss": 1.121,
+      "step": 337
+    },
+    {
+      "epoch": 0.02077187807276303,
+      "grad_norm": 19.625,
+      "learning_rate": 4.493333333333333e-06,
+      "loss": 1.1034,
+      "step": 338
+    },
+    {
+      "epoch": 0.020833333333333332,
+      "grad_norm": 11.625,
+      "learning_rate": 4.506666666666667e-06,
+      "loss": 1.1099,
+      "step": 339
+    },
+    {
+      "epoch": 0.02089478859390364,
+      "grad_norm": 18.875,
+      "learning_rate": 4.520000000000001e-06,
+      "loss": 1.0467,
+      "step": 340
+    },
+    {
+      "epoch": 0.020956243854473942,
+      "grad_norm": 23.5,
+      "learning_rate": 4.533333333333334e-06,
+      "loss": 1.2196,
+      "step": 341
+    },
+    {
+      "epoch": 0.02101769911504425,
+      "grad_norm": 24.0,
+      "learning_rate": 4.546666666666667e-06,
+      "loss": 1.1509,
+      "step": 342
+    },
+    {
+      "epoch": 0.021079154375614553,
+      "grad_norm": 27.5,
+      "learning_rate": 4.56e-06,
+      "loss": 1.1707,
+      "step": 343
+    },
+    {
+      "epoch": 0.021140609636184856,
+      "grad_norm": 13.375,
+      "learning_rate": 4.573333333333333e-06,
+      "loss": 1.0749,
+      "step": 344
+    },
+    {
+      "epoch": 0.021202064896755163,
+      "grad_norm": 26.75,
+      "learning_rate": 4.586666666666667e-06,
+      "loss": 1.1022,
+      "step": 345
+    },
+    {
+      "epoch": 0.021263520157325466,
+      "grad_norm": 10.1875,
+      "learning_rate": 4.600000000000001e-06,
+      "loss": 1.0706,
+      "step": 346
+    },
+    {
+      "epoch": 0.021324975417895773,
+      "grad_norm": 23.0,
+      "learning_rate": 4.613333333333334e-06,
+      "loss": 1.113,
+      "step": 347
+    },
+    {
+      "epoch": 0.021386430678466076,
+      "grad_norm": 26.625,
+      "learning_rate": 4.626666666666667e-06,
+      "loss": 1.0932,
+      "step": 348
+    },
+    {
+      "epoch": 0.021447885939036383,
+      "grad_norm": 43.5,
+      "learning_rate": 4.6400000000000005e-06,
+      "loss": 1.1524,
+      "step": 349
+    },
+    {
+      "epoch": 0.021509341199606687,
+      "grad_norm": 22.125,
+      "learning_rate": 4.653333333333333e-06,
+      "loss": 1.1053,
+      "step": 350
+    },
+    {
+      "epoch": 0.02157079646017699,
+      "grad_norm": 29.875,
+      "learning_rate": 4.666666666666667e-06,
+      "loss": 1.1645,
+      "step": 351
+    },
+    {
+      "epoch": 0.021632251720747297,
+      "grad_norm": 17.0,
+      "learning_rate": 4.680000000000001e-06,
+      "loss": 1.0259,
+      "step": 352
+    },
+    {
+      "epoch": 0.0216937069813176,
+      "grad_norm": 47.75,
+      "learning_rate": 4.693333333333334e-06,
+      "loss": 1.1654,
+      "step": 353
+    },
+    {
+      "epoch": 0.021755162241887907,
+      "grad_norm": 21.75,
+      "learning_rate": 4.706666666666667e-06,
+      "loss": 1.1345,
+      "step": 354
+    },
+    {
+      "epoch": 0.02181661750245821,
+      "grad_norm": 21.125,
+      "learning_rate": 4.7200000000000005e-06,
+      "loss": 1.1081,
+      "step": 355
+    },
+    {
+      "epoch": 0.021878072763028514,
+      "grad_norm": 17.25,
+      "learning_rate": 4.7333333333333335e-06,
+      "loss": 1.108,
+      "step": 356
+    },
+    {
+      "epoch": 0.02193952802359882,
+      "grad_norm": 20.875,
+      "learning_rate": 4.746666666666667e-06,
+      "loss": 1.1756,
+      "step": 357
+    },
+    {
+      "epoch": 0.022000983284169124,
+      "grad_norm": 14.5625,
+      "learning_rate": 4.76e-06,
+      "loss": 1.1457,
+      "step": 358
+    },
+    {
+      "epoch": 0.02206243854473943,
+      "grad_norm": 18.375,
+      "learning_rate": 4.773333333333334e-06,
+      "loss": 1.0998,
+      "step": 359
+    },
+    {
+      "epoch": 0.022123893805309734,
+      "grad_norm": 52.25,
+      "learning_rate": 4.786666666666667e-06,
+      "loss": 1.2224,
+      "step": 360
+    },
+    {
+      "epoch": 0.02218534906588004,
+      "grad_norm": 15.6875,
+      "learning_rate": 4.800000000000001e-06,
+      "loss": 1.184,
+      "step": 361
+    },
+    {
+      "epoch": 0.022246804326450344,
+      "grad_norm": 13.25,
+      "learning_rate": 4.8133333333333336e-06,
+      "loss": 1.1848,
+      "step": 362
+    },
+    {
+      "epoch": 0.022308259587020648,
+      "grad_norm": 16.5,
+      "learning_rate": 4.826666666666667e-06,
+      "loss": 1.0546,
+      "step": 363
+    },
+    {
+      "epoch": 0.022369714847590955,
+      "grad_norm": 77.5,
+      "learning_rate": 4.84e-06,
+      "loss": 1.3082,
+      "step": 364
+    },
+    {
+      "epoch": 0.022431170108161258,
+      "grad_norm": 32.25,
+      "learning_rate": 4.853333333333334e-06,
+      "loss": 1.0865,
+      "step": 365
+    },
+    {
+      "epoch": 0.022492625368731565,
+      "grad_norm": 40.5,
+      "learning_rate": 4.866666666666667e-06,
+      "loss": 1.1381,
+      "step": 366
+    },
+    {
+      "epoch": 0.022554080629301868,
+      "grad_norm": 12.125,
+      "learning_rate": 4.880000000000001e-06,
+      "loss": 1.1044,
+      "step": 367
+    },
+    {
+      "epoch": 0.02261553588987217,
+      "grad_norm": 18.625,
+      "learning_rate": 4.893333333333334e-06,
+      "loss": 1.1129,
+      "step": 368
+    },
+    {
+      "epoch": 0.02267699115044248,
+      "grad_norm": 13.1875,
+      "learning_rate": 4.9066666666666666e-06,
+      "loss": 1.168,
+      "step": 369
+    },
+    {
+      "epoch": 0.022738446411012782,
+      "grad_norm": 14.25,
+      "learning_rate": 4.92e-06,
+      "loss": 1.0833,
+      "step": 370
+    },
+    {
+      "epoch": 0.02279990167158309,
+      "grad_norm": 20.625,
+      "learning_rate": 4.933333333333334e-06,
+      "loss": 1.1194,
+      "step": 371
+    },
+    {
+      "epoch": 0.022861356932153392,
+      "grad_norm": 21.25,
+      "learning_rate": 4.946666666666667e-06,
+      "loss": 1.0971,
+      "step": 372
+    },
+    {
+      "epoch": 0.0229228121927237,
+      "grad_norm": 21.25,
+      "learning_rate": 4.960000000000001e-06,
+      "loss": 1.0615,
+      "step": 373
+    },
+    {
+      "epoch": 0.022984267453294002,
+      "grad_norm": 18.25,
+      "learning_rate": 4.973333333333334e-06,
+      "loss": 1.0367,
+      "step": 374
+    },
+    {
+      "epoch": 0.023045722713864306,
+      "grad_norm": 12.375,
+      "learning_rate": 4.986666666666667e-06,
+      "loss": 1.0912,
+      "step": 375
+    },
+    {
+      "epoch": 0.023107177974434612,
+      "grad_norm": 14.375,
+      "learning_rate": 5e-06,
+      "loss": 1.0483,
+      "step": 376
+    },
+    {
+      "epoch": 0.023168633235004916,
+      "grad_norm": 28.0,
+      "learning_rate": 5.013333333333333e-06,
+      "loss": 1.2354,
+      "step": 377
+    },
+    {
+      "epoch": 0.023230088495575223,
+      "grad_norm": 25.625,
+      "learning_rate": 5.026666666666667e-06,
+      "loss": 1.1683,
+      "step": 378
+    },
+    {
+      "epoch": 0.023291543756145526,
+      "grad_norm": 30.0,
+      "learning_rate": 5.04e-06,
+      "loss": 1.2038,
+      "step": 379
+    },
+    {
+      "epoch": 0.02335299901671583,
+      "grad_norm": 18.875,
+      "learning_rate": 5.053333333333334e-06,
+      "loss": 1.1381,
+      "step": 380
+    },
+    {
+      "epoch": 0.023414454277286136,
+      "grad_norm": 19.5,
+      "learning_rate": 5.0666666666666676e-06,
+      "loss": 1.2141,
+      "step": 381
+    },
+    {
+      "epoch": 0.02347590953785644,
+      "grad_norm": 17.5,
+      "learning_rate": 5.0800000000000005e-06,
+      "loss": 1.0728,
+      "step": 382
+    },
+    {
+      "epoch": 0.023537364798426746,
+      "grad_norm": 27.875,
+      "learning_rate": 5.093333333333333e-06,
+      "loss": 1.2395,
+      "step": 383
+    },
+    {
+      "epoch": 0.02359882005899705,
+      "grad_norm": 8.375,
+      "learning_rate": 5.106666666666667e-06,
+      "loss": 1.1278,
+      "step": 384
+    },
+    {
+      "epoch": 0.023660275319567357,
+      "grad_norm": 10.375,
+      "learning_rate": 5.12e-06,
+      "loss": 1.1472,
+      "step": 385
+    },
+    {
+      "epoch": 0.02372173058013766,
+      "grad_norm": 23.75,
+      "learning_rate": 5.133333333333334e-06,
+      "loss": 1.0952,
+      "step": 386
+    },
+    {
+      "epoch": 0.023783185840707963,
+      "grad_norm": 11.375,
+      "learning_rate": 5.146666666666668e-06,
+      "loss": 1.1354,
+      "step": 387
+    },
+    {
+      "epoch": 0.02384464110127827,
+      "grad_norm": 13.5,
+      "learning_rate": 5.1600000000000006e-06,
+      "loss": 1.0963,
+      "step": 388
+    },
+    {
+      "epoch": 0.023906096361848574,
+      "grad_norm": 19.75,
+      "learning_rate": 5.1733333333333335e-06,
+      "loss": 1.1244,
+      "step": 389
+    },
+    {
+      "epoch": 0.02396755162241888,
+      "grad_norm": 15.6875,
+      "learning_rate": 5.186666666666667e-06,
+      "loss": 1.0766,
+      "step": 390
+    },
+    {
+      "epoch": 0.024029006882989184,
+      "grad_norm": 13.375,
+      "learning_rate": 5.2e-06,
+      "loss": 1.1214,
+      "step": 391
+    },
+    {
+      "epoch": 0.024090462143559487,
+      "grad_norm": 14.25,
+      "learning_rate": 5.213333333333334e-06,
+      "loss": 1.0921,
+      "step": 392
+    },
+    {
+      "epoch": 0.024151917404129794,
+      "grad_norm": 10.0625,
+      "learning_rate": 5.226666666666667e-06,
+      "loss": 1.0561,
+      "step": 393
+    },
+    {
+      "epoch": 0.024213372664700097,
+      "grad_norm": 10.375,
+      "learning_rate": 5.240000000000001e-06,
+      "loss": 1.0064,
+      "step": 394
+    },
+    {
+      "epoch": 0.024274827925270404,
+      "grad_norm": 9.875,
+      "learning_rate": 5.2533333333333336e-06,
+      "loss": 1.084,
+      "step": 395
+    },
+    {
+      "epoch": 0.024336283185840708,
+      "grad_norm": 13.375,
+      "learning_rate": 5.2666666666666665e-06,
+      "loss": 1.1051,
+      "step": 396
+    },
+    {
+      "epoch": 0.024397738446411014,
+      "grad_norm": 36.25,
+      "learning_rate": 5.28e-06,
+      "loss": 1.0979,
+      "step": 397
+    },
+    {
+      "epoch": 0.024459193706981318,
+      "grad_norm": 15.6875,
+      "learning_rate": 5.293333333333334e-06,
+      "loss": 1.1049,
+      "step": 398
+    },
+    {
+      "epoch": 0.02452064896755162,
+      "grad_norm": 14.8125,
+      "learning_rate": 5.306666666666667e-06,
+      "loss": 1.1048,
+      "step": 399
+    },
+    {
+      "epoch": 0.024582104228121928,
+      "grad_norm": 23.375,
+      "learning_rate": 5.320000000000001e-06,
+      "loss": 1.0895,
+      "step": 400
+    },
+    {
+      "epoch": 0.02464355948869223,
+      "grad_norm": 17.0,
+      "learning_rate": 5.333333333333334e-06,
+      "loss": 1.032,
+      "step": 401
+    },
+    {
+      "epoch": 0.024705014749262538,
+      "grad_norm": 7.46875,
+      "learning_rate": 5.346666666666667e-06,
+      "loss": 1.0176,
+      "step": 402
+    },
+    {
+      "epoch": 0.02476647000983284,
+      "grad_norm": 8.4375,
+      "learning_rate": 5.36e-06,
+      "loss": 1.1405,
+      "step": 403
+    },
+    {
+      "epoch": 0.024827925270403145,
+      "grad_norm": 26.0,
+      "learning_rate": 5.373333333333334e-06,
+      "loss": 1.0573,
+      "step": 404
+    },
+    {
+      "epoch": 0.024889380530973452,
+      "grad_norm": 10.1875,
+      "learning_rate": 5.386666666666667e-06,
+      "loss": 1.1466,
+      "step": 405
+    },
+    {
+      "epoch": 0.024950835791543755,
+      "grad_norm": 24.25,
+      "learning_rate": 5.400000000000001e-06,
+      "loss": 1.1189,
+      "step": 406
+    },
+    {
+      "epoch": 0.025012291052114062,
+      "grad_norm": 19.625,
+      "learning_rate": 5.413333333333334e-06,
+      "loss": 1.0372,
+      "step": 407
+    },
+    {
+      "epoch": 0.025073746312684365,
+      "grad_norm": 36.0,
+      "learning_rate": 5.426666666666667e-06,
+      "loss": 1.0903,
+      "step": 408
+    },
+    {
+      "epoch": 0.025135201573254672,
+      "grad_norm": 41.5,
+      "learning_rate": 5.4400000000000004e-06,
+      "loss": 1.2313,
+      "step": 409
+    },
+    {
+      "epoch": 0.025196656833824976,
+      "grad_norm": 19.0,
+      "learning_rate": 5.453333333333334e-06,
+      "loss": 1.0858,
+      "step": 410
+    },
+    {
+      "epoch": 0.02525811209439528,
+      "grad_norm": 14.9375,
+      "learning_rate": 5.466666666666667e-06,
+      "loss": 1.1539,
+      "step": 411
+    },
+    {
+      "epoch": 0.025319567354965586,
+      "grad_norm": 14.4375,
+      "learning_rate": 5.480000000000001e-06,
+      "loss": 1.0802,
+      "step": 412
+    },
+    {
+      "epoch": 0.02538102261553589,
+      "grad_norm": 26.875,
+      "learning_rate": 5.493333333333334e-06,
+      "loss": 1.0796,
+      "step": 413
+    },
+    {
+      "epoch": 0.025442477876106196,
+      "grad_norm": 17.0,
+      "learning_rate": 5.506666666666667e-06,
+      "loss": 1.0887,
+      "step": 414
+    },
+    {
+      "epoch": 0.0255039331366765,
+      "grad_norm": 22.875,
+      "learning_rate": 5.5200000000000005e-06,
+      "loss": 1.0662,
+      "step": 415
+    },
+    {
+      "epoch": 0.025565388397246803,
+      "grad_norm": 16.5,
+      "learning_rate": 5.533333333333334e-06,
+      "loss": 1.0819,
+      "step": 416
+    },
+    {
+      "epoch": 0.02562684365781711,
+      "grad_norm": 15.5625,
+      "learning_rate": 5.546666666666667e-06,
+      "loss": 1.1185,
+      "step": 417
+    },
+    {
+      "epoch": 0.025688298918387413,
+      "grad_norm": 26.125,
+      "learning_rate": 5.560000000000001e-06,
+      "loss": 1.1053,
+      "step": 418
+    },
+    {
+      "epoch": 0.02574975417895772,
+      "grad_norm": 19.75,
+      "learning_rate": 5.573333333333334e-06,
+      "loss": 1.1059,
+      "step": 419
+    },
+    {
+      "epoch": 0.025811209439528023,
+      "grad_norm": 14.0625,
+      "learning_rate": 5.586666666666667e-06,
+      "loss": 1.0988,
+      "step": 420
+    },
+    {
+      "epoch": 0.02587266470009833,
+      "grad_norm": 9.4375,
+      "learning_rate": 5.600000000000001e-06,
+      "loss": 1.1474,
+      "step": 421
+    },
+    {
+      "epoch": 0.025934119960668633,
+      "grad_norm": 11.0625,
+      "learning_rate": 5.613333333333334e-06,
+      "loss": 1.1135,
+      "step": 422
+    },
+    {
+      "epoch": 0.025995575221238937,
+      "grad_norm": 9.0,
+      "learning_rate": 5.626666666666667e-06,
+      "loss": 1.0204,
+      "step": 423
+    },
+    {
+      "epoch": 0.026057030481809244,
+      "grad_norm": 16.0,
+      "learning_rate": 5.64e-06,
+      "loss": 0.9981,
+      "step": 424
+    },
+    {
+      "epoch": 0.026118485742379547,
+      "grad_norm": 9.0625,
+      "learning_rate": 5.653333333333334e-06,
+      "loss": 1.0163,
+      "step": 425
+    },
+    {
+      "epoch": 0.026179941002949854,
+      "grad_norm": 21.125,
+      "learning_rate": 5.666666666666667e-06,
+      "loss": 1.0589,
+      "step": 426
+    },
+    {
+      "epoch": 0.026241396263520157,
+      "grad_norm": 16.5,
+      "learning_rate": 5.68e-06,
+      "loss": 1.1238,
+      "step": 427
+    },
+    {
+      "epoch": 0.02630285152409046,
+      "grad_norm": 21.25,
+      "learning_rate": 5.6933333333333344e-06,
+      "loss": 1.2123,
+      "step": 428
+    },
+    {
+      "epoch": 0.026364306784660767,
+      "grad_norm": 13.875,
+      "learning_rate": 5.706666666666667e-06,
+      "loss": 1.1196,
+      "step": 429
+    },
+    {
+      "epoch": 0.02642576204523107,
+      "grad_norm": 20.625,
+      "learning_rate": 5.72e-06,
+      "loss": 1.0555,
+      "step": 430
+    },
+    {
+      "epoch": 0.026487217305801378,
+      "grad_norm": 25.375,
+      "learning_rate": 5.733333333333334e-06,
+      "loss": 1.0522,
+      "step": 431
+    },
+    {
+      "epoch": 0.02654867256637168,
+      "grad_norm": 8.5625,
+      "learning_rate": 5.746666666666667e-06,
+      "loss": 1.0864,
+      "step": 432
+    },
+    {
+      "epoch": 0.026610127826941988,
+      "grad_norm": 37.0,
+      "learning_rate": 5.76e-06,
+      "loss": 1.1354,
+      "step": 433
+    },
+    {
+      "epoch": 0.02667158308751229,
+      "grad_norm": 25.125,
+      "learning_rate": 5.7733333333333345e-06,
+      "loss": 1.0033,
+      "step": 434
+    },
+    {
+      "epoch": 0.026733038348082595,
+      "grad_norm": 43.0,
+      "learning_rate": 5.7866666666666674e-06,
+      "loss": 1.047,
+      "step": 435
+    },
+    {
+      "epoch": 0.0267944936086529,
+      "grad_norm": 18.125,
+      "learning_rate": 5.8e-06,
+      "loss": 1.1274,
+      "step": 436
+    },
+    {
+      "epoch": 0.026855948869223205,
+      "grad_norm": 15.4375,
+      "learning_rate": 5.813333333333334e-06,
+      "loss": 1.0806,
+      "step": 437
+    },
+    {
+      "epoch": 0.02691740412979351,
+      "grad_norm": 10.0625,
+      "learning_rate": 5.826666666666667e-06,
+      "loss": 1.0263,
+      "step": 438
+    },
+    {
+      "epoch": 0.026978859390363815,
+      "grad_norm": 16.75,
+      "learning_rate": 5.84e-06,
+      "loss": 1.1131,
+      "step": 439
+    },
+    {
+      "epoch": 0.02704031465093412,
+      "grad_norm": 22.75,
+      "learning_rate": 5.853333333333335e-06,
+      "loss": 1.1802,
+      "step": 440
+    },
+    {
+      "epoch": 0.027101769911504425,
+      "grad_norm": 10.5,
+      "learning_rate": 5.8666666666666675e-06,
+      "loss": 1.0123,
+      "step": 441
+    },
+    {
+      "epoch": 0.02716322517207473,
+      "grad_norm": 13.4375,
+      "learning_rate": 5.8800000000000005e-06,
+      "loss": 1.0406,
+      "step": 442
+    },
+    {
+      "epoch": 0.027224680432645035,
+      "grad_norm": 18.375,
+      "learning_rate": 5.893333333333334e-06,
+      "loss": 1.1492,
+      "step": 443
+    },
+    {
+      "epoch": 0.02728613569321534,
+      "grad_norm": 12.75,
+      "learning_rate": 5.906666666666667e-06,
+      "loss": 1.1121,
+      "step": 444
+    },
+    {
+      "epoch": 0.027347590953785646,
+      "grad_norm": 12.4375,
+      "learning_rate": 5.92e-06,
+      "loss": 1.1758,
+      "step": 445
+    },
+    {
+      "epoch": 0.02740904621435595,
+      "grad_norm": 16.5,
+      "learning_rate": 5.933333333333335e-06,
+      "loss": 1.0911,
+      "step": 446
+    },
+    {
+      "epoch": 0.027470501474926252,
+      "grad_norm": 20.25,
+      "learning_rate": 5.946666666666668e-06,
+      "loss": 1.054,
+      "step": 447
+    },
+    {
+      "epoch": 0.02753195673549656,
+      "grad_norm": 10.4375,
+      "learning_rate": 5.9600000000000005e-06,
+      "loss": 1.0425,
+      "step": 448
+    },
+    {
+      "epoch": 0.027593411996066863,
+      "grad_norm": 17.875,
+      "learning_rate": 5.973333333333334e-06,
+      "loss": 1.224,
+      "step": 449
+    },
+    {
+      "epoch": 0.02765486725663717,
+      "grad_norm": 16.375,
+      "learning_rate": 5.986666666666667e-06,
+      "loss": 1.1029,
+      "step": 450
+    },
+    {
+      "epoch": 0.027716322517207473,
+      "grad_norm": 19.5,
+      "learning_rate": 6e-06,
+      "loss": 1.1007,
+      "step": 451
+    },
+    {
+      "epoch": 0.027777777777777776,
+      "grad_norm": 28.375,
+      "learning_rate": 6.013333333333335e-06,
+      "loss": 1.0575,
+      "step": 452
+    },
+    {
+      "epoch": 0.027839233038348083,
+      "grad_norm": 22.125,
+      "learning_rate": 6.026666666666668e-06,
+      "loss": 1.1137,
+      "step": 453
+    },
+    {
+      "epoch": 0.027900688298918386,
+      "grad_norm": 36.25,
+      "learning_rate": 6.040000000000001e-06,
+      "loss": 1.0707,
+      "step": 454
+    },
+    {
+      "epoch": 0.027962143559488693,
+      "grad_norm": 13.8125,
+      "learning_rate": 6.0533333333333335e-06,
+      "loss": 1.0358,
+      "step": 455
+    },
+    {
+      "epoch": 0.028023598820058997,
+      "grad_norm": 12.5,
+      "learning_rate": 6.066666666666667e-06,
+      "loss": 1.0705,
+      "step": 456
+    },
+    {
+      "epoch": 0.028085054080629304,
+      "grad_norm": 14.75,
+      "learning_rate": 6.08e-06,
+      "loss": 1.0621,
+      "step": 457
+    },
+    {
+      "epoch": 0.028146509341199607,
+      "grad_norm": 14.9375,
+      "learning_rate": 6.093333333333333e-06,
+      "loss": 1.0041,
+      "step": 458
+    },
+    {
+      "epoch": 0.02820796460176991,
+      "grad_norm": 15.6875,
+      "learning_rate": 6.106666666666668e-06,
+      "loss": 0.9826,
+      "step": 459
+    },
+    {
+      "epoch": 0.028269419862340217,
+      "grad_norm": 18.0,
+      "learning_rate": 6.120000000000001e-06,
+      "loss": 1.0493,
+      "step": 460
+    },
+    {
+      "epoch": 0.02833087512291052,
+      "grad_norm": 23.0,
+      "learning_rate": 6.133333333333334e-06,
+      "loss": 0.9997,
+      "step": 461
+    },
+    {
+      "epoch": 0.028392330383480827,
+      "grad_norm": 16.125,
+      "learning_rate": 6.146666666666667e-06,
+      "loss": 1.0271,
+      "step": 462
+    },
+    {
+      "epoch": 0.02845378564405113,
+      "grad_norm": 28.625,
+      "learning_rate": 6.16e-06,
+      "loss": 1.0693,
+      "step": 463
+    },
+    {
+      "epoch": 0.028515240904621434,
+      "grad_norm": 26.25,
+      "learning_rate": 6.173333333333333e-06,
+      "loss": 1.0104,
+      "step": 464
+    },
+    {
+      "epoch": 0.02857669616519174,
+      "grad_norm": 8.9375,
+      "learning_rate": 6.186666666666668e-06,
+      "loss": 0.9769,
+      "step": 465
+    },
+    {
+      "epoch": 0.028638151425762044,
+      "grad_norm": 14.25,
+      "learning_rate": 6.200000000000001e-06,
+      "loss": 0.9755,
+      "step": 466
+    },
+    {
+      "epoch": 0.02869960668633235,
+      "grad_norm": 10.6875,
+      "learning_rate": 6.213333333333334e-06,
+      "loss": 1.0195,
+      "step": 467
+    },
+    {
+      "epoch": 0.028761061946902654,
+      "grad_norm": 8.0,
+      "learning_rate": 6.2266666666666675e-06,
+      "loss": 0.9316,
+      "step": 468
+    },
+    {
+      "epoch": 0.02882251720747296,
+      "grad_norm": 10.9375,
+      "learning_rate": 6.24e-06,
+      "loss": 0.8514,
+      "step": 469
+    },
+    {
+      "epoch": 0.028883972468043265,
+      "grad_norm": 24.375,
+      "learning_rate": 6.253333333333333e-06,
+      "loss": 1.0113,
+      "step": 470
+    },
+    {
+      "epoch": 0.028945427728613568,
+      "grad_norm": 20.125,
+      "learning_rate": 6.266666666666668e-06,
+      "loss": 1.0327,
+      "step": 471
+    },
+    {
+      "epoch": 0.029006882989183875,
+      "grad_norm": 35.5,
+      "learning_rate": 6.280000000000001e-06,
+      "loss": 0.885,
+      "step": 472
+    },
+    {
+      "epoch": 0.029068338249754178,
+      "grad_norm": 15.5,
+      "learning_rate": 6.293333333333334e-06,
+      "loss": 1.1404,
+      "step": 473
+    },
+    {
+      "epoch": 0.029129793510324485,
+      "grad_norm": 14.3125,
+      "learning_rate": 6.3066666666666676e-06,
+      "loss": 1.0994,
+      "step": 474
+    },
+    {
+      "epoch": 0.02919124877089479,
+      "grad_norm": 23.5,
+      "learning_rate": 6.3200000000000005e-06,
+      "loss": 1.0879,
+      "step": 475
+    },
+    {
+      "epoch": 0.029252704031465092,
+      "grad_norm": 12.8125,
+      "learning_rate": 6.333333333333333e-06,
+      "loss": 0.8972,
+      "step": 476
+    },
+    {
+      "epoch": 0.0293141592920354,
+      "grad_norm": 11.0,
+      "learning_rate": 6.346666666666668e-06,
+      "loss": 1.0301,
+      "step": 477
+    },
+    {
+      "epoch": 0.029375614552605702,
+      "grad_norm": 28.0,
+      "learning_rate": 6.360000000000001e-06,
+      "loss": 1.0921,
+      "step": 478
+    },
+    {
+      "epoch": 0.02943706981317601,
+      "grad_norm": 16.125,
+      "learning_rate": 6.373333333333334e-06,
+      "loss": 0.9853,
+      "step": 479
+    },
+    {
+      "epoch": 0.029498525073746312,
+      "grad_norm": 11.1875,
+      "learning_rate": 6.386666666666668e-06,
+      "loss": 0.9663,
+      "step": 480
+    },
+    {
+      "epoch": 0.02955998033431662,
+      "grad_norm": 11.75,
+      "learning_rate": 6.4000000000000006e-06,
+      "loss": 1.0874,
+      "step": 481
+    },
+    {
+      "epoch": 0.029621435594886922,
+      "grad_norm": 12.75,
+      "learning_rate": 6.4133333333333335e-06,
+      "loss": 0.9843,
+      "step": 482
+    },
+    {
+      "epoch": 0.029682890855457226,
+      "grad_norm": 29.625,
+      "learning_rate": 6.426666666666668e-06,
+      "loss": 1.0121,
+      "step": 483
+    },
+    {
+      "epoch": 0.029744346116027533,
+      "grad_norm": 31.625,
+      "learning_rate": 6.440000000000001e-06,
+      "loss": 1.0971,
+      "step": 484
+    },
+    {
+      "epoch": 0.029805801376597836,
+      "grad_norm": 18.0,
+      "learning_rate": 6.453333333333334e-06,
+      "loss": 1.0744,
+      "step": 485
+    },
+    {
+      "epoch": 0.029867256637168143,
+      "grad_norm": 23.125,
+      "learning_rate": 6.466666666666667e-06,
+      "loss": 1.0479,
+      "step": 486
+    },
+    {
+      "epoch": 0.029928711897738446,
+      "grad_norm": 12.6875,
+      "learning_rate": 6.480000000000001e-06,
+      "loss": 0.9901,
+      "step": 487
+    },
+    {
+      "epoch": 0.02999016715830875,
+      "grad_norm": 46.25,
+      "learning_rate": 6.4933333333333336e-06,
+      "loss": 1.038,
+      "step": 488
+    },
+    {
+      "epoch": 0.030051622418879056,
+      "grad_norm": 33.75,
+      "learning_rate": 6.5066666666666665e-06,
+      "loss": 1.0754,
+      "step": 489
+    },
+    {
+      "epoch": 0.03011307767944936,
+      "grad_norm": 14.3125,
+      "learning_rate": 6.520000000000001e-06,
+      "loss": 1.0229,
+      "step": 490
+    },
+    {
+      "epoch": 0.030174532940019667,
+      "grad_norm": 23.75,
+      "learning_rate": 6.533333333333334e-06,
+      "loss": 1.0013,
+      "step": 491
+    },
+    {
+      "epoch": 0.03023598820058997,
+      "grad_norm": 31.75,
+      "learning_rate": 6.546666666666667e-06,
+      "loss": 1.1724,
+      "step": 492
+    },
+    {
+      "epoch": 0.030297443461160277,
+      "grad_norm": 19.5,
+      "learning_rate": 6.560000000000001e-06,
+      "loss": 0.9667,
+      "step": 493
+    },
+    {
+      "epoch": 0.03035889872173058,
+      "grad_norm": 14.125,
+      "learning_rate": 6.573333333333334e-06,
+      "loss": 1.1136,
+      "step": 494
+    },
+    {
+      "epoch": 0.030420353982300884,
+      "grad_norm": 41.25,
+      "learning_rate": 6.5866666666666666e-06,
+      "loss": 1.0007,
+      "step": 495
+    },
+    {
+      "epoch": 0.03048180924287119,
+      "grad_norm": 15.5,
+      "learning_rate": 6.600000000000001e-06,
+      "loss": 1.0133,
+      "step": 496
+    },
+    {
+      "epoch": 0.030543264503441494,
+      "grad_norm": 21.0,
+      "learning_rate": 6.613333333333334e-06,
+      "loss": 0.8813,
+      "step": 497
+    },
+    {
+      "epoch": 0.0306047197640118,
+      "grad_norm": 16.375,
+      "learning_rate": 6.626666666666667e-06,
+      "loss": 0.9524,
+      "step": 498
+    },
+    {
+      "epoch": 0.030666175024582104,
+      "grad_norm": 17.75,
+      "learning_rate": 6.640000000000001e-06,
+      "loss": 0.9767,
+      "step": 499
+    },
+    {
+      "epoch": 0.030727630285152407,
+      "grad_norm": 18.125,
+      "learning_rate": 6.653333333333334e-06,
+      "loss": 1.0693,
+      "step": 500
+    },
+    {
+      "epoch": 0.030789085545722714,
+      "grad_norm": 20.125,
+      "learning_rate": 6.666666666666667e-06,
+      "loss": 0.8604,
+      "step": 501
+    },
+    {
+      "epoch": 0.030850540806293018,
+      "grad_norm": 12.0,
+      "learning_rate": 6.680000000000001e-06,
+      "loss": 0.8634,
+      "step": 502
+    },
+    {
+      "epoch": 0.030911996066863325,
+      "grad_norm": 13.5,
+      "learning_rate": 6.693333333333334e-06,
+      "loss": 1.0665,
+      "step": 503
+    },
+    {
+      "epoch": 0.030973451327433628,
+      "grad_norm": 24.375,
+      "learning_rate": 6.706666666666667e-06,
+      "loss": 0.921,
+      "step": 504
+    },
+    {
+      "epoch": 0.031034906588003935,
+      "grad_norm": 28.5,
+      "learning_rate": 6.720000000000001e-06,
+      "loss": 0.935,
+      "step": 505
+    },
+    {
+      "epoch": 0.031096361848574238,
+      "grad_norm": 14.75,
+      "learning_rate": 6.733333333333334e-06,
+      "loss": 0.9524,
+      "step": 506
+    },
+    {
+      "epoch": 0.03115781710914454,
+      "grad_norm": 18.75,
+      "learning_rate": 6.746666666666667e-06,
+      "loss": 0.9773,
+      "step": 507
+    },
+    {
+      "epoch": 0.03121927236971485,
+      "grad_norm": 23.375,
+      "learning_rate": 6.760000000000001e-06,
+      "loss": 0.9076,
+      "step": 508
+    },
+    {
+      "epoch": 0.03128072763028515,
+      "grad_norm": 30.75,
+      "learning_rate": 6.773333333333334e-06,
+      "loss": 0.9671,
+      "step": 509
+    },
+    {
+      "epoch": 0.031342182890855455,
+      "grad_norm": 29.625,
+      "learning_rate": 6.786666666666667e-06,
+      "loss": 0.965,
+      "step": 510
+    },
+    {
+      "epoch": 0.031403638151425765,
+      "grad_norm": 22.5,
+      "learning_rate": 6.800000000000001e-06,
+      "loss": 0.9741,
+      "step": 511
+    },
+    {
+      "epoch": 0.03146509341199607,
+      "grad_norm": 28.0,
+      "learning_rate": 6.813333333333334e-06,
+      "loss": 0.7974,
+      "step": 512
+    },
+    {
+      "epoch": 0.03152654867256637,
+      "grad_norm": 35.25,
+      "learning_rate": 6.826666666666667e-06,
+      "loss": 0.844,
+      "step": 513
+    },
+    {
+      "epoch": 0.031588003933136675,
+      "grad_norm": 25.75,
+      "learning_rate": 6.8400000000000014e-06,
+      "loss": 0.8082,
+      "step": 514
+    },
+    {
+      "epoch": 0.03164945919370698,
+      "grad_norm": 54.75,
+      "learning_rate": 6.853333333333334e-06,
+      "loss": 0.8529,
+      "step": 515
+    },
+    {
+      "epoch": 0.03171091445427729,
+      "grad_norm": 48.0,
+      "learning_rate": 6.866666666666667e-06,
+      "loss": 0.7857,
+      "step": 516
+    },
+    {
+      "epoch": 0.03177236971484759,
+      "grad_norm": 53.5,
+      "learning_rate": 6.88e-06,
+      "loss": 0.8327,
+      "step": 517
+    },
+    {
+      "epoch": 0.031833824975417896,
+      "grad_norm": 36.75,
+      "learning_rate": 6.893333333333334e-06,
+      "loss": 0.7622,
+      "step": 518
+    },
+    {
+      "epoch": 0.0318952802359882,
+      "grad_norm": 32.0,
+      "learning_rate": 6.906666666666667e-06,
+      "loss": 0.8164,
+      "step": 519
+    },
+    {
+      "epoch": 0.0319567354965585,
+      "grad_norm": 20.875,
+      "learning_rate": 6.92e-06,
+      "loss": 0.799,
+      "step": 520
+    },
+    {
+      "epoch": 0.03201819075712881,
+      "grad_norm": 36.5,
+      "learning_rate": 6.9333333333333344e-06,
+      "loss": 0.7903,
+      "step": 521
+    },
+    {
+      "epoch": 0.032079646017699116,
+      "grad_norm": 35.25,
+      "learning_rate": 6.946666666666667e-06,
+      "loss": 1.0056,
+      "step": 522
+    },
+    {
+      "epoch": 0.03214110127826942,
+      "grad_norm": 27.5,
+      "learning_rate": 6.96e-06,
+      "loss": 0.8031,
+      "step": 523
+    },
+    {
+      "epoch": 0.03220255653883972,
+      "grad_norm": 71.0,
+      "learning_rate": 6.973333333333334e-06,
+      "loss": 1.1609,
+      "step": 524
+    },
+    {
+      "epoch": 0.032264011799410026,
+      "grad_norm": 23.625,
+      "learning_rate": 6.986666666666667e-06,
+      "loss": 0.8413,
+      "step": 525
+    },
+    {
+      "epoch": 0.03232546705998034,
+      "grad_norm": 30.125,
+      "learning_rate": 7e-06,
+      "loss": 0.7228,
+      "step": 526
+    },
+    {
+      "epoch": 0.03238692232055064,
+      "grad_norm": 33.0,
+      "learning_rate": 7.0133333333333345e-06,
+      "loss": 0.8425,
+      "step": 527
+    },
+    {
+      "epoch": 0.032448377581120944,
+      "grad_norm": 46.75,
+      "learning_rate": 7.0266666666666674e-06,
+      "loss": 0.8788,
+      "step": 528
+    },
+    {
+      "epoch": 0.03250983284169125,
+      "grad_norm": 61.5,
+      "learning_rate": 7.04e-06,
+      "loss": 0.9771,
+      "step": 529
+    },
+    {
+      "epoch": 0.03257128810226155,
+      "grad_norm": 70.5,
+      "learning_rate": 7.053333333333334e-06,
+      "loss": 0.8288,
+      "step": 530
+    },
+    {
+      "epoch": 0.03263274336283186,
+      "grad_norm": 35.0,
+      "learning_rate": 7.066666666666667e-06,
+      "loss": 0.9278,
+      "step": 531
+    },
+    {
+      "epoch": 0.032694198623402164,
+      "grad_norm": 88.0,
+      "learning_rate": 7.08e-06,
+      "loss": 0.8835,
+      "step": 532
+    },
+    {
+      "epoch": 0.03275565388397247,
+      "grad_norm": 56.5,
+      "learning_rate": 7.093333333333335e-06,
+      "loss": 0.7722,
+      "step": 533
+    },
+    {
+      "epoch": 0.03281710914454277,
+      "grad_norm": 49.25,
+      "learning_rate": 7.1066666666666675e-06,
+      "loss": 0.7818,
+      "step": 534
+    },
+    {
+      "epoch": 0.03287856440511308,
+      "grad_norm": 69.5,
+      "learning_rate": 7.1200000000000004e-06,
+      "loss": 0.7787,
+      "step": 535
+    },
+    {
+      "epoch": 0.032940019665683384,
+      "grad_norm": 49.25,
+      "learning_rate": 7.133333333333334e-06,
+      "loss": 0.8444,
+      "step": 536
+    },
+    {
+      "epoch": 0.03300147492625369,
+      "grad_norm": 46.0,
+      "learning_rate": 7.146666666666667e-06,
+      "loss": 0.8784,
+      "step": 537
+    },
+    {
+      "epoch": 0.03306293018682399,
+      "grad_norm": 73.0,
+      "learning_rate": 7.16e-06,
+      "loss": 0.8475,
+      "step": 538
+    },
+    {
+      "epoch": 0.033124385447394294,
+      "grad_norm": 32.75,
+      "learning_rate": 7.173333333333335e-06,
+      "loss": 0.9105,
+      "step": 539
+    },
+    {
+      "epoch": 0.033185840707964605,
+      "grad_norm": 91.0,
+      "learning_rate": 7.186666666666668e-06,
+      "loss": 0.9562,
+      "step": 540
+    },
+    {
+      "epoch": 0.03324729596853491,
+      "grad_norm": 49.5,
+      "learning_rate": 7.2000000000000005e-06,
+      "loss": 0.8389,
+      "step": 541
+    },
+    {
+      "epoch": 0.03330875122910521,
+      "grad_norm": 33.5,
+      "learning_rate": 7.213333333333334e-06,
+      "loss": 0.838,
+      "step": 542
+    },
+    {
+      "epoch": 0.033370206489675515,
+      "grad_norm": 43.25,
+      "learning_rate": 7.226666666666667e-06,
+      "loss": 0.9548,
+      "step": 543
+    },
+    {
+      "epoch": 0.03343166175024582,
+      "grad_norm": 32.0,
+      "learning_rate": 7.24e-06,
+      "loss": 0.8057,
+      "step": 544
+    },
+    {
+      "epoch": 0.03349311701081613,
+      "grad_norm": 33.0,
+      "learning_rate": 7.253333333333335e-06,
+      "loss": 0.8826,
+      "step": 545
+    },
+    {
+      "epoch": 0.03355457227138643,
+      "grad_norm": 40.0,
+      "learning_rate": 7.266666666666668e-06,
+      "loss": 0.802,
+      "step": 546
+    },
+    {
+      "epoch": 0.033616027531956735,
+      "grad_norm": 29.5,
+      "learning_rate": 7.280000000000001e-06,
+      "loss": 0.7812,
+      "step": 547
+    },
+    {
+      "epoch": 0.03367748279252704,
+      "grad_norm": 56.75,
+      "learning_rate": 7.2933333333333335e-06,
+      "loss": 0.8719,
+      "step": 548
+    },
+    {
+      "epoch": 0.03373893805309734,
+      "grad_norm": 33.75,
+      "learning_rate": 7.306666666666667e-06,
+      "loss": 0.9626,
+      "step": 549
+    },
+    {
+      "epoch": 0.03380039331366765,
+      "grad_norm": 51.5,
+      "learning_rate": 7.32e-06,
+      "loss": 1.0222,
+      "step": 550
+    },
+    {
+      "epoch": 0.033861848574237956,
+      "grad_norm": 38.75,
+      "learning_rate": 7.333333333333333e-06,
+      "loss": 0.7883,
+      "step": 551
+    },
+    {
+      "epoch": 0.03392330383480826,
+      "grad_norm": 77.5,
+      "learning_rate": 7.346666666666668e-06,
+      "loss": 0.8014,
+      "step": 552
+    },
+    {
+      "epoch": 0.03398475909537856,
+      "grad_norm": 70.5,
+      "learning_rate": 7.360000000000001e-06,
+      "loss": 0.862,
+      "step": 553
+    },
+    {
+      "epoch": 0.034046214355948866,
+      "grad_norm": 40.25,
+      "learning_rate": 7.373333333333334e-06,
+      "loss": 0.7617,
+      "step": 554
+    },
+    {
+      "epoch": 0.034107669616519176,
+      "grad_norm": 46.25,
+      "learning_rate": 7.386666666666667e-06,
+      "loss": 0.7726,
+      "step": 555
+    },
+    {
+      "epoch": 0.03416912487708948,
+      "grad_norm": 49.25,
+      "learning_rate": 7.4e-06,
+      "loss": 0.9869,
+      "step": 556
+    },
+    {
+      "epoch": 0.03423058013765978,
+      "grad_norm": 34.5,
+      "learning_rate": 7.413333333333333e-06,
+      "loss": 0.7539,
+      "step": 557
+    },
+    {
+      "epoch": 0.034292035398230086,
+      "grad_norm": 30.5,
+      "learning_rate": 7.426666666666668e-06,
+      "loss": 0.6686,
+      "step": 558
+    },
+    {
+      "epoch": 0.0343534906588004,
+      "grad_norm": 49.75,
+      "learning_rate": 7.440000000000001e-06,
+      "loss": 0.8961,
+      "step": 559
+    },
+    {
+      "epoch": 0.0344149459193707,
+      "grad_norm": 95.5,
+      "learning_rate": 7.453333333333334e-06,
+      "loss": 0.8031,
+      "step": 560
+    },
+    {
+      "epoch": 0.034476401179941,
+      "grad_norm": 54.25,
+      "learning_rate": 7.4666666666666675e-06,
+      "loss": 0.915,
+      "step": 561
+    },
+    {
+      "epoch": 0.03453785644051131,
+      "grad_norm": 102.0,
+      "learning_rate": 7.48e-06,
+      "loss": 1.0234,
+      "step": 562
+    },
+    {
+      "epoch": 0.03459931170108161,
+      "grad_norm": 70.5,
+      "learning_rate": 7.493333333333333e-06,
+      "loss": 0.9048,
+      "step": 563
+    },
+    {
+      "epoch": 0.03466076696165192,
+      "grad_norm": 101.0,
+      "learning_rate": 7.506666666666668e-06,
+      "loss": 0.989,
+      "step": 564
+    },
+    {
+      "epoch": 0.034722222222222224,
+      "grad_norm": 73.0,
+      "learning_rate": 7.520000000000001e-06,
+      "loss": 1.1432,
+      "step": 565
+    },
+    {
+      "epoch": 0.03478367748279253,
+      "grad_norm": 53.5,
+      "learning_rate": 7.533333333333334e-06,
+      "loss": 1.048,
+      "step": 566
+    },
+    {
+      "epoch": 0.03484513274336283,
+      "grad_norm": 34.75,
+      "learning_rate": 7.5466666666666675e-06,
+      "loss": 0.9056,
+      "step": 567
+    },
+    {
+      "epoch": 0.034906588003933134,
+      "grad_norm": 60.0,
+      "learning_rate": 7.5600000000000005e-06,
+      "loss": 0.9027,
+      "step": 568
+    },
+    {
+      "epoch": 0.034968043264503444,
+      "grad_norm": 32.25,
+      "learning_rate": 7.573333333333333e-06,
+      "loss": 0.7684,
+      "step": 569
+    },
+    {
+      "epoch": 0.03502949852507375,
+      "grad_norm": 57.0,
+      "learning_rate": 7.586666666666668e-06,
+      "loss": 0.7446,
+      "step": 570
+    },
+    {
+      "epoch": 0.03509095378564405,
+      "grad_norm": 52.75,
+      "learning_rate": 7.600000000000001e-06,
+      "loss": 0.8055,
+      "step": 571
+    },
+    {
+      "epoch": 0.035152409046214354,
+      "grad_norm": 36.75,
+      "learning_rate": 7.613333333333334e-06,
+      "loss": 0.8156,
+      "step": 572
+    },
+    {
+      "epoch": 0.03521386430678466,
+      "grad_norm": 27.0,
+      "learning_rate": 7.626666666666668e-06,
+      "loss": 0.7767,
+      "step": 573
+    },
+    {
+      "epoch": 0.03527531956735497,
+      "grad_norm": 32.5,
+      "learning_rate": 7.640000000000001e-06,
+      "loss": 0.812,
+      "step": 574
+    },
+    {
+      "epoch": 0.03533677482792527,
+      "grad_norm": 28.0,
+      "learning_rate": 7.653333333333333e-06,
+      "loss": 0.8056,
+      "step": 575
+    },
+    {
+      "epoch": 0.035398230088495575,
+      "grad_norm": 22.75,
+      "learning_rate": 7.666666666666667e-06,
+      "loss": 0.6915,
+      "step": 576
+    },
+    {
+      "epoch": 0.03545968534906588,
+      "grad_norm": 34.5,
+      "learning_rate": 7.680000000000001e-06,
+      "loss": 0.7822,
+      "step": 577
+    },
+    {
+      "epoch": 0.03552114060963618,
+      "grad_norm": 58.25,
+      "learning_rate": 7.693333333333333e-06,
+      "loss": 0.8356,
+      "step": 578
+    },
+    {
+      "epoch": 0.03558259587020649,
+      "grad_norm": 40.25,
+      "learning_rate": 7.706666666666669e-06,
+      "loss": 0.8046,
+      "step": 579
+    },
+    {
+      "epoch": 0.035644051130776795,
+      "grad_norm": 42.5,
+      "learning_rate": 7.72e-06,
+      "loss": 0.642,
+      "step": 580
+    },
+    {
+      "epoch": 0.0357055063913471,
+      "grad_norm": 36.75,
+      "learning_rate": 7.733333333333334e-06,
+      "loss": 0.7037,
+      "step": 581
+    },
+    {
+      "epoch": 0.0357669616519174,
+      "grad_norm": 53.25,
+      "learning_rate": 7.746666666666666e-06,
+      "loss": 0.9178,
+      "step": 582
+    },
+    {
+      "epoch": 0.03582841691248771,
+      "grad_norm": 61.5,
+      "learning_rate": 7.76e-06,
+      "loss": 0.8404,
+      "step": 583
+    },
+    {
+      "epoch": 0.035889872173058016,
+      "grad_norm": 34.5,
+      "learning_rate": 7.773333333333334e-06,
+      "loss": 0.7966,
+      "step": 584
+    },
+    {
+      "epoch": 0.03595132743362832,
+      "grad_norm": 40.25,
+      "learning_rate": 7.786666666666666e-06,
+      "loss": 0.8787,
+      "step": 585
+    },
+    {
+      "epoch": 0.03601278269419862,
+      "grad_norm": 53.0,
+      "learning_rate": 7.800000000000002e-06,
+      "loss": 1.002,
+      "step": 586
+    },
+    {
+      "epoch": 0.036074237954768926,
+      "grad_norm": 39.75,
+      "learning_rate": 7.813333333333334e-06,
+      "loss": 1.1473,
+      "step": 587
+    },
+    {
+      "epoch": 0.036135693215339236,
+      "grad_norm": 38.75,
+      "learning_rate": 7.826666666666667e-06,
+      "loss": 1.088,
+      "step": 588
+    },
+    {
+      "epoch": 0.03619714847590954,
+      "grad_norm": 26.375,
+      "learning_rate": 7.840000000000001e-06,
+      "loss": 0.9988,
+      "step": 589
+    },
+    {
+      "epoch": 0.03625860373647984,
+      "grad_norm": 45.25,
+      "learning_rate": 7.853333333333333e-06,
+      "loss": 0.9803,
+      "step": 590
+    },
+    {
+      "epoch": 0.036320058997050146,
+      "grad_norm": 36.75,
+      "learning_rate": 7.866666666666667e-06,
+      "loss": 1.039,
+      "step": 591
+    },
+    {
+      "epoch": 0.03638151425762045,
+      "grad_norm": 40.5,
+      "learning_rate": 7.88e-06,
+      "loss": 0.8799,
+      "step": 592
+    },
+    {
+      "epoch": 0.03644296951819076,
+      "grad_norm": 39.0,
+      "learning_rate": 7.893333333333335e-06,
+      "loss": 0.8865,
+      "step": 593
+    },
+    {
+      "epoch": 0.03650442477876106,
+      "grad_norm": 30.125,
+      "learning_rate": 7.906666666666667e-06,
+      "loss": 0.8302,
+      "step": 594
+    },
+    {
+      "epoch": 0.03656588003933137,
+      "grad_norm": 29.375,
+      "learning_rate": 7.92e-06,
+      "loss": 0.9001,
+      "step": 595
+    },
+    {
+      "epoch": 0.03662733529990167,
+      "grad_norm": 59.25,
+      "learning_rate": 7.933333333333334e-06,
+      "loss": 0.915,
+      "step": 596
+    },
+    {
+      "epoch": 0.03668879056047197,
+      "grad_norm": 36.75,
+      "learning_rate": 7.946666666666666e-06,
+      "loss": 0.7387,
+      "step": 597
+    },
+    {
+      "epoch": 0.036750245821042284,
+      "grad_norm": 48.75,
+      "learning_rate": 7.960000000000002e-06,
+      "loss": 0.8842,
+      "step": 598
+    },
+    {
+      "epoch": 0.03681170108161259,
+      "grad_norm": 16.875,
+      "learning_rate": 7.973333333333334e-06,
+      "loss": 0.6863,
+      "step": 599
+    },
+    {
+      "epoch": 0.03687315634218289,
+      "grad_norm": 35.75,
+      "learning_rate": 7.986666666666668e-06,
+      "loss": 0.796,
+      "step": 600
+    },
+    {
+      "epoch": 0.036934611602753194,
+      "grad_norm": 36.25,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 0.7564,
+      "step": 601
+    },
+    {
+      "epoch": 0.0369960668633235,
+      "grad_norm": 52.25,
+      "learning_rate": 8.013333333333333e-06,
+      "loss": 0.8301,
+      "step": 602
+    },
+    {
+      "epoch": 0.03705752212389381,
+      "grad_norm": 27.75,
+      "learning_rate": 8.026666666666667e-06,
+      "loss": 0.6874,
+      "step": 603
+    },
+    {
+      "epoch": 0.03711897738446411,
+      "grad_norm": 50.75,
+      "learning_rate": 8.040000000000001e-06,
+      "loss": 0.6375,
+      "step": 604
+    },
+    {
+      "epoch": 0.037180432645034414,
+      "grad_norm": 54.25,
+      "learning_rate": 8.053333333333335e-06,
+      "loss": 0.6989,
+      "step": 605
+    },
+    {
+      "epoch": 0.03724188790560472,
+      "grad_norm": 84.0,
+      "learning_rate": 8.066666666666667e-06,
+      "loss": 0.9357,
+      "step": 606
+    },
+    {
+      "epoch": 0.03730334316617503,
+      "grad_norm": 33.0,
+      "learning_rate": 8.08e-06,
+      "loss": 0.9372,
+      "step": 607
+    },
+    {
+      "epoch": 0.03736479842674533,
+      "grad_norm": 32.25,
+      "learning_rate": 8.093333333333334e-06,
+      "loss": 0.7379,
+      "step": 608
+    },
+    {
+      "epoch": 0.037426253687315635,
+      "grad_norm": 34.25,
+      "learning_rate": 8.106666666666666e-06,
+      "loss": 0.776,
+      "step": 609
+    },
+    {
+      "epoch": 0.03748770894788594,
+      "grad_norm": 25.125,
+      "learning_rate": 8.120000000000002e-06,
+      "loss": 0.7643,
+      "step": 610
+    },
+    {
+      "epoch": 0.03754916420845624,
+      "grad_norm": 55.75,
+      "learning_rate": 8.133333333333334e-06,
+      "loss": 0.8392,
+      "step": 611
+    },
+    {
+      "epoch": 0.03761061946902655,
+      "grad_norm": 51.25,
+      "learning_rate": 8.146666666666668e-06,
+      "loss": 0.726,
+      "step": 612
+    },
+    {
+      "epoch": 0.037672074729596855,
+      "grad_norm": 78.0,
+      "learning_rate": 8.16e-06,
+      "loss": 0.99,
+      "step": 613
+    },
+    {
+      "epoch": 0.03773352999016716,
+      "grad_norm": 53.5,
+      "learning_rate": 8.173333333333334e-06,
+      "loss": 0.8275,
+      "step": 614
+    },
+    {
+      "epoch": 0.03779498525073746,
+      "grad_norm": 48.0,
+      "learning_rate": 8.186666666666667e-06,
+      "loss": 0.827,
+      "step": 615
+    },
+    {
+      "epoch": 0.037856440511307765,
+      "grad_norm": 44.5,
+      "learning_rate": 8.2e-06,
+      "loss": 0.751,
+      "step": 616
+    },
+    {
+      "epoch": 0.037917895771878075,
+      "grad_norm": 58.25,
+      "learning_rate": 8.213333333333335e-06,
+      "loss": 0.8964,
+      "step": 617
+    },
+    {
+      "epoch": 0.03797935103244838,
+      "grad_norm": 37.5,
+      "learning_rate": 8.226666666666667e-06,
+      "loss": 1.0725,
+      "step": 618
+    },
+    {
+      "epoch": 0.03804080629301868,
+      "grad_norm": 73.0,
+      "learning_rate": 8.24e-06,
+      "loss": 0.8148,
+      "step": 619
+    },
+    {
+      "epoch": 0.038102261553588986,
+      "grad_norm": 45.25,
+      "learning_rate": 8.253333333333334e-06,
+      "loss": 0.7827,
+      "step": 620
+    },
+    {
+      "epoch": 0.03816371681415929,
+      "grad_norm": 69.0,
+      "learning_rate": 8.266666666666667e-06,
+      "loss": 0.9836,
+      "step": 621
+    },
+    {
+      "epoch": 0.0382251720747296,
+      "grad_norm": 56.0,
+      "learning_rate": 8.28e-06,
+      "loss": 0.9259,
+      "step": 622
+    },
+    {
+      "epoch": 0.0382866273352999,
+      "grad_norm": 42.5,
+      "learning_rate": 8.293333333333334e-06,
+      "loss": 0.8056,
+      "step": 623
+    },
+    {
+      "epoch": 0.038348082595870206,
+      "grad_norm": 40.25,
+      "learning_rate": 8.306666666666668e-06,
+      "loss": 0.8676,
+      "step": 624
+    },
+    {
+      "epoch": 0.03840953785644051,
+      "grad_norm": 61.5,
+      "learning_rate": 8.32e-06,
+      "loss": 1.0395,
+      "step": 625
+    },
+    {
+      "epoch": 0.03847099311701081,
+      "grad_norm": 42.75,
+      "learning_rate": 8.333333333333334e-06,
+      "loss": 0.8258,
+      "step": 626
+    },
+    {
+      "epoch": 0.03853244837758112,
+      "grad_norm": 65.5,
+      "learning_rate": 8.346666666666668e-06,
+      "loss": 0.7707,
+      "step": 627
+    },
+    {
+      "epoch": 0.038593903638151426,
+      "grad_norm": 44.75,
+      "learning_rate": 8.36e-06,
+      "loss": 0.8274,
+      "step": 628
+    },
+    {
+      "epoch": 0.03865535889872173,
+      "grad_norm": 66.5,
+      "learning_rate": 8.373333333333335e-06,
+      "loss": 0.8568,
+      "step": 629
+    },
+    {
+      "epoch": 0.03871681415929203,
+      "grad_norm": 68.5,
+      "learning_rate": 8.386666666666667e-06,
+      "loss": 0.7702,
+      "step": 630
+    },
+    {
+      "epoch": 0.03877826941986234,
+      "grad_norm": 46.5,
+      "learning_rate": 8.400000000000001e-06,
+      "loss": 0.7266,
+      "step": 631
+    },
+    {
+      "epoch": 0.03883972468043265,
+      "grad_norm": 24.75,
+      "learning_rate": 8.413333333333335e-06,
+      "loss": 0.6446,
+      "step": 632
+    },
+    {
+      "epoch": 0.03890117994100295,
+      "grad_norm": 74.0,
+      "learning_rate": 8.426666666666667e-06,
+      "loss": 0.9642,
+      "step": 633
+    },
+    {
+      "epoch": 0.038962635201573254,
+      "grad_norm": 66.0,
+      "learning_rate": 8.44e-06,
+      "loss": 1.119,
+      "step": 634
+    },
+    {
+      "epoch": 0.03902409046214356,
+      "grad_norm": 57.25,
+      "learning_rate": 8.453333333333334e-06,
+      "loss": 0.9793,
+      "step": 635
+    },
+    {
+      "epoch": 0.03908554572271387,
+      "grad_norm": 51.75,
+      "learning_rate": 8.466666666666668e-06,
+      "loss": 1.1962,
+      "step": 636
+    },
+    {
+      "epoch": 0.03914700098328417,
+      "grad_norm": 44.5,
+      "learning_rate": 8.48e-06,
+      "loss": 1.0263,
+      "step": 637
+    },
+    {
+      "epoch": 0.039208456243854474,
+      "grad_norm": 46.0,
+      "learning_rate": 8.493333333333334e-06,
+      "loss": 0.9972,
+      "step": 638
+    },
+    {
+      "epoch": 0.03926991150442478,
+      "grad_norm": 54.5,
+      "learning_rate": 8.506666666666668e-06,
+      "loss": 0.9303,
+      "step": 639
+    },
+    {
+      "epoch": 0.03933136676499508,
+      "grad_norm": 24.625,
+      "learning_rate": 8.52e-06,
+      "loss": 0.8192,
+      "step": 640
+    },
+    {
+      "epoch": 0.03939282202556539,
+      "grad_norm": 57.25,
+      "learning_rate": 8.533333333333335e-06,
+      "loss": 0.9214,
+      "step": 641
+    },
+    {
+      "epoch": 0.039454277286135694,
+      "grad_norm": 37.25,
+      "learning_rate": 8.546666666666667e-06,
+      "loss": 0.9645,
+      "step": 642
+    },
+    {
+      "epoch": 0.039515732546706,
+      "grad_norm": 26.125,
+      "learning_rate": 8.560000000000001e-06,
+      "loss": 0.971,
+      "step": 643
+    },
+    {
+      "epoch": 0.0395771878072763,
+      "grad_norm": 44.5,
+      "learning_rate": 8.573333333333333e-06,
+      "loss": 0.8452,
+      "step": 644
+    },
+    {
+      "epoch": 0.039638643067846605,
+      "grad_norm": 52.25,
+      "learning_rate": 8.586666666666667e-06,
+      "loss": 0.967,
+      "step": 645
+    },
+    {
+      "epoch": 0.039700098328416915,
+      "grad_norm": 79.0,
+      "learning_rate": 8.6e-06,
+      "loss": 1.0605,
+      "step": 646
+    },
+    {
+      "epoch": 0.03976155358898722,
+      "grad_norm": 19.125,
+      "learning_rate": 8.613333333333333e-06,
+      "loss": 0.8204,
+      "step": 647
+    },
+    {
+      "epoch": 0.03982300884955752,
+      "grad_norm": 28.25,
+      "learning_rate": 8.626666666666668e-06,
+      "loss": 0.8179,
+      "step": 648
+    },
+    {
+      "epoch": 0.039884464110127825,
+      "grad_norm": 30.375,
+      "learning_rate": 8.64e-06,
+      "loss": 0.8463,
+      "step": 649
+    },
+    {
+      "epoch": 0.03994591937069813,
+      "grad_norm": 28.375,
+      "learning_rate": 8.653333333333334e-06,
+      "loss": 0.7462,
+      "step": 650
+    },
+    {
+      "epoch": 0.04000737463126844,
+      "grad_norm": 64.5,
+      "learning_rate": 8.666666666666668e-06,
+      "loss": 0.7727,
+      "step": 651
+    },
+    {
+      "epoch": 0.04006882989183874,
+      "grad_norm": 80.5,
+      "learning_rate": 8.68e-06,
+      "loss": 0.9155,
+      "step": 652
+    },
+    {
+      "epoch": 0.040130285152409045,
+      "grad_norm": 47.25,
+      "learning_rate": 8.693333333333334e-06,
+      "loss": 0.728,
+      "step": 653
+    },
+    {
+      "epoch": 0.04019174041297935,
+      "grad_norm": 46.0,
+      "learning_rate": 8.706666666666667e-06,
+      "loss": 0.85,
+      "step": 654
+    },
+    {
+      "epoch": 0.04025319567354966,
+      "grad_norm": 30.0,
+      "learning_rate": 8.720000000000001e-06,
+      "loss": 0.8259,
+      "step": 655
+    },
+    {
+      "epoch": 0.04031465093411996,
+      "grad_norm": 51.25,
+      "learning_rate": 8.733333333333333e-06,
+      "loss": 0.729,
+      "step": 656
+    },
+    {
+      "epoch": 0.040376106194690266,
+      "grad_norm": 35.0,
+      "learning_rate": 8.746666666666667e-06,
+      "loss": 0.7594,
+      "step": 657
+    },
+    {
+      "epoch": 0.04043756145526057,
+      "grad_norm": 43.5,
+      "learning_rate": 8.76e-06,
+      "loss": 0.7039,
+      "step": 658
+    },
+    {
+      "epoch": 0.04049901671583087,
+      "grad_norm": 38.5,
+      "learning_rate": 8.773333333333333e-06,
+      "loss": 0.6428,
+      "step": 659
+    },
+    {
+      "epoch": 0.04056047197640118,
+      "grad_norm": 33.0,
+      "learning_rate": 8.786666666666668e-06,
+      "loss": 0.8518,
+      "step": 660
+    },
+    {
+      "epoch": 0.040621927236971486,
+      "grad_norm": 39.0,
+      "learning_rate": 8.8e-06,
+      "loss": 0.6758,
+      "step": 661
+    },
+    {
+      "epoch": 0.04068338249754179,
+      "grad_norm": 34.5,
+      "learning_rate": 8.813333333333334e-06,
+      "loss": 0.6863,
+      "step": 662
+    },
+    {
+      "epoch": 0.04074483775811209,
+      "grad_norm": 48.75,
+      "learning_rate": 8.826666666666668e-06,
+      "loss": 0.7819,
+      "step": 663
+    },
+    {
+      "epoch": 0.040806293018682396,
+      "grad_norm": 33.5,
+      "learning_rate": 8.84e-06,
+      "loss": 0.7992,
+      "step": 664
+    },
+    {
+      "epoch": 0.04086774827925271,
+      "grad_norm": 39.0,
+      "learning_rate": 8.853333333333334e-06,
+      "loss": 0.8031,
+      "step": 665
+    },
+    {
+      "epoch": 0.04092920353982301,
+      "grad_norm": 36.75,
+      "learning_rate": 8.866666666666668e-06,
+      "loss": 0.7268,
+      "step": 666
+    },
+    {
+      "epoch": 0.04099065880039331,
+      "grad_norm": 22.25,
+      "learning_rate": 8.880000000000001e-06,
+      "loss": 0.8118,
+      "step": 667
+    },
+    {
+      "epoch": 0.04105211406096362,
+      "grad_norm": 30.875,
+      "learning_rate": 8.893333333333333e-06,
+      "loss": 0.6847,
+      "step": 668
+    },
+    {
+      "epoch": 0.04111356932153392,
+      "grad_norm": 47.5,
+      "learning_rate": 8.906666666666667e-06,
+      "loss": 0.7691,
+      "step": 669
+    },
+    {
+      "epoch": 0.04117502458210423,
+      "grad_norm": 46.25,
+      "learning_rate": 8.920000000000001e-06,
+      "loss": 0.743,
+      "step": 670
+    },
+    {
+      "epoch": 0.041236479842674534,
+      "grad_norm": 34.25,
+      "learning_rate": 8.933333333333333e-06,
+      "loss": 0.6177,
+      "step": 671
+    },
+    {
+      "epoch": 0.04129793510324484,
+      "grad_norm": 37.75,
+      "learning_rate": 8.946666666666669e-06,
+      "loss": 0.7844,
+      "step": 672
+    },
+    {
+      "epoch": 0.04135939036381514,
+      "grad_norm": 49.0,
+      "learning_rate": 8.96e-06,
+      "loss": 0.7841,
+      "step": 673
+    },
+    {
+      "epoch": 0.041420845624385444,
+      "grad_norm": 50.5,
+      "learning_rate": 8.973333333333334e-06,
+      "loss": 0.7439,
+      "step": 674
+    },
+    {
+      "epoch": 0.041482300884955754,
+      "grad_norm": 37.75,
+      "learning_rate": 8.986666666666666e-06,
+      "loss": 0.7848,
+      "step": 675
+    },
+    {
+      "epoch": 0.04154375614552606,
+      "grad_norm": 59.0,
+      "learning_rate": 9e-06,
+      "loss": 0.7377,
+      "step": 676
+    },
+    {
+      "epoch": 0.04160521140609636,
+      "grad_norm": 27.125,
+      "learning_rate": 9.013333333333334e-06,
+      "loss": 0.7076,
+      "step": 677
+    },
+    {
+      "epoch": 0.041666666666666664,
+      "grad_norm": 42.0,
+      "learning_rate": 9.026666666666666e-06,
+      "loss": 0.7818,
+      "step": 678
+    },
+    {
+      "epoch": 0.041728121927236975,
+      "grad_norm": 35.0,
+      "learning_rate": 9.040000000000002e-06,
+      "loss": 0.6463,
+      "step": 679
+    },
+    {
+      "epoch": 0.04178957718780728,
+      "grad_norm": 51.25,
+      "learning_rate": 9.053333333333334e-06,
+      "loss": 0.7036,
+      "step": 680
+    },
+    {
+      "epoch": 0.04185103244837758,
+      "grad_norm": 39.25,
+      "learning_rate": 9.066666666666667e-06,
+      "loss": 0.7589,
+      "step": 681
+    },
+    {
+      "epoch": 0.041912487708947885,
+      "grad_norm": 44.0,
+      "learning_rate": 9.080000000000001e-06,
+      "loss": 0.7547,
+      "step": 682
+    },
+    {
+      "epoch": 0.04197394296951819,
+      "grad_norm": 48.0,
+      "learning_rate": 9.093333333333333e-06,
+      "loss": 0.7718,
+      "step": 683
+    },
+    {
+      "epoch": 0.0420353982300885,
+      "grad_norm": 31.625,
+      "learning_rate": 9.106666666666667e-06,
+      "loss": 0.816,
+      "step": 684
+    },
+    {
+      "epoch": 0.0420968534906588,
+      "grad_norm": 26.125,
+      "learning_rate": 9.12e-06,
+      "loss": 0.7412,
+      "step": 685
+    },
+    {
+      "epoch": 0.042158308751229105,
+      "grad_norm": 25.25,
+      "learning_rate": 9.133333333333335e-06,
+      "loss": 0.8087,
+      "step": 686
+    },
+    {
+      "epoch": 0.04221976401179941,
+      "grad_norm": 21.875,
+      "learning_rate": 9.146666666666667e-06,
+      "loss": 0.6321,
+      "step": 687
+    },
+    {
+      "epoch": 0.04228121927236971,
+      "grad_norm": 34.5,
+      "learning_rate": 9.16e-06,
+      "loss": 0.7168,
+      "step": 688
+    },
+    {
+      "epoch": 0.04234267453294002,
+      "grad_norm": 28.0,
+      "learning_rate": 9.173333333333334e-06,
+      "loss": 0.7317,
+      "step": 689
+    },
+    {
+      "epoch": 0.042404129793510326,
+      "grad_norm": 45.0,
+      "learning_rate": 9.186666666666666e-06,
+      "loss": 0.7807,
+      "step": 690
+    },
+    {
+      "epoch": 0.04246558505408063,
+      "grad_norm": 30.125,
+      "learning_rate": 9.200000000000002e-06,
+      "loss": 0.735,
+      "step": 691
+    },
+    {
+      "epoch": 0.04252704031465093,
+      "grad_norm": 35.75,
+      "learning_rate": 9.213333333333334e-06,
+      "loss": 0.8574,
+      "step": 692
+    },
+    {
+      "epoch": 0.042588495575221236,
+      "grad_norm": 41.25,
+      "learning_rate": 9.226666666666668e-06,
+      "loss": 0.6181,
+      "step": 693
+    },
+    {
+      "epoch": 0.042649950835791546,
+      "grad_norm": 39.0,
+      "learning_rate": 9.240000000000001e-06,
+      "loss": 0.8465,
+      "step": 694
+    },
+    {
+      "epoch": 0.04271140609636185,
+      "grad_norm": 26.25,
+      "learning_rate": 9.253333333333333e-06,
+      "loss": 0.7085,
+      "step": 695
+    },
+    {
+      "epoch": 0.04277286135693215,
+      "grad_norm": 35.25,
+      "learning_rate": 9.266666666666667e-06,
+      "loss": 0.7554,
+      "step": 696
+    },
+    {
+      "epoch": 0.042834316617502456,
+      "grad_norm": 82.5,
+      "learning_rate": 9.280000000000001e-06,
+      "loss": 1.2424,
+      "step": 697
+    },
+    {
+      "epoch": 0.042895771878072766,
+      "grad_norm": 33.75,
+      "learning_rate": 9.293333333333335e-06,
+      "loss": 0.9356,
+      "step": 698
+    },
+    {
+      "epoch": 0.04295722713864307,
+      "grad_norm": 30.25,
+      "learning_rate": 9.306666666666667e-06,
+      "loss": 0.9749,
+      "step": 699
+    },
+    {
+      "epoch": 0.04301868239921337,
+      "grad_norm": 52.0,
+      "learning_rate": 9.32e-06,
+      "loss": 1.1634,
+      "step": 700
+    },
+    {
+      "epoch": 0.04308013765978368,
+      "grad_norm": 32.25,
+      "learning_rate": 9.333333333333334e-06,
+      "loss": 1.1435,
+      "step": 701
+    },
+    {
+      "epoch": 0.04314159292035398,
+      "grad_norm": 44.0,
+      "learning_rate": 9.346666666666666e-06,
+      "loss": 1.0926,
+      "step": 702
+    },
+    {
+      "epoch": 0.04320304818092429,
+      "grad_norm": 32.75,
+      "learning_rate": 9.360000000000002e-06,
+      "loss": 0.8124,
+      "step": 703
+    },
+    {
+      "epoch": 0.043264503441494594,
+      "grad_norm": 25.125,
+      "learning_rate": 9.373333333333334e-06,
+      "loss": 0.978,
+      "step": 704
+    },
+    {
+      "epoch": 0.0433259587020649,
+      "grad_norm": 26.75,
+      "learning_rate": 9.386666666666668e-06,
+      "loss": 0.7538,
+      "step": 705
+    },
+    {
+      "epoch": 0.0433874139626352,
+      "grad_norm": 28.25,
+      "learning_rate": 9.4e-06,
+      "loss": 0.5974,
+      "step": 706
+    },
+    {
+      "epoch": 0.043448869223205504,
+      "grad_norm": 25.75,
+      "learning_rate": 9.413333333333334e-06,
+      "loss": 0.7366,
+      "step": 707
+    },
+    {
+      "epoch": 0.043510324483775814,
+      "grad_norm": 27.625,
+      "learning_rate": 9.426666666666667e-06,
+      "loss": 0.7622,
+      "step": 708
+    },
+    {
+      "epoch": 0.04357177974434612,
+      "grad_norm": 44.25,
+      "learning_rate": 9.440000000000001e-06,
+      "loss": 0.6432,
+      "step": 709
+    },
+    {
+      "epoch": 0.04363323500491642,
+      "grad_norm": 49.0,
+      "learning_rate": 9.453333333333335e-06,
+      "loss": 0.7007,
+      "step": 710
+    },
+    {
+      "epoch": 0.043694690265486724,
+      "grad_norm": 38.25,
+      "learning_rate": 9.466666666666667e-06,
+      "loss": 0.6815,
+      "step": 711
+    },
+    {
+      "epoch": 0.04375614552605703,
+      "grad_norm": 39.25,
+      "learning_rate": 9.48e-06,
+      "loss": 0.6481,
+      "step": 712
+    },
+    {
+      "epoch": 0.04381760078662734,
+      "grad_norm": 39.0,
+      "learning_rate": 9.493333333333334e-06,
+      "loss": 0.8907,
+      "step": 713
+    },
+    {
+      "epoch": 0.04387905604719764,
+      "grad_norm": 27.875,
+      "learning_rate": 9.506666666666667e-06,
+      "loss": 0.7847,
+      "step": 714
+    },
+    {
+      "epoch": 0.043940511307767945,
+      "grad_norm": 33.0,
+      "learning_rate": 9.52e-06,
+      "loss": 0.7244,
+      "step": 715
+    },
+    {
+      "epoch": 0.04400196656833825,
+      "grad_norm": 42.75,
+      "learning_rate": 9.533333333333334e-06,
+      "loss": 0.7727,
+      "step": 716
+    },
+    {
+      "epoch": 0.04406342182890855,
+      "grad_norm": 49.5,
+      "learning_rate": 9.546666666666668e-06,
+      "loss": 0.7163,
+      "step": 717
+    },
+    {
+      "epoch": 0.04412487708947886,
+      "grad_norm": 66.0,
+      "learning_rate": 9.56e-06,
+      "loss": 0.7222,
+      "step": 718
+    },
+    {
+      "epoch": 0.044186332350049165,
+      "grad_norm": 69.5,
+      "learning_rate": 9.573333333333334e-06,
+      "loss": 0.8451,
+      "step": 719
+    },
+    {
+      "epoch": 0.04424778761061947,
+      "grad_norm": 42.5,
+      "learning_rate": 9.586666666666667e-06,
+      "loss": 0.8206,
+      "step": 720
+    },
+    {
+      "epoch": 0.04430924287118977,
+      "grad_norm": 64.0,
+      "learning_rate": 9.600000000000001e-06,
+      "loss": 0.775,
+      "step": 721
+    },
+    {
+      "epoch": 0.04437069813176008,
+      "grad_norm": 30.25,
+      "learning_rate": 9.613333333333335e-06,
+      "loss": 0.6167,
+      "step": 722
+    },
+    {
+      "epoch": 0.044432153392330385,
+      "grad_norm": 49.0,
+      "learning_rate": 9.626666666666667e-06,
+      "loss": 0.5596,
+      "step": 723
+    },
+    {
+      "epoch": 0.04449360865290069,
+      "grad_norm": 39.5,
+      "learning_rate": 9.640000000000001e-06,
+      "loss": 0.7102,
+      "step": 724
+    },
+    {
+      "epoch": 0.04455506391347099,
+      "grad_norm": 41.25,
+      "learning_rate": 9.653333333333335e-06,
+      "loss": 0.79,
+      "step": 725
+    },
+    {
+      "epoch": 0.044616519174041296,
+      "grad_norm": 36.0,
+      "learning_rate": 9.666666666666667e-06,
+      "loss": 0.7565,
+      "step": 726
+    },
+    {
+      "epoch": 0.044677974434611606,
+      "grad_norm": 59.75,
+      "learning_rate": 9.68e-06,
+      "loss": 0.7336,
+      "step": 727
+    },
+    {
+      "epoch": 0.04473942969518191,
+      "grad_norm": 30.875,
+      "learning_rate": 9.693333333333334e-06,
+      "loss": 0.5786,
+      "step": 728
+    },
+    {
+      "epoch": 0.04480088495575221,
+      "grad_norm": 37.25,
+      "learning_rate": 9.706666666666668e-06,
+      "loss": 0.8139,
+      "step": 729
+    },
+    {
+      "epoch": 0.044862340216322516,
+      "grad_norm": 30.75,
+      "learning_rate": 9.72e-06,
+      "loss": 0.8694,
+      "step": 730
+    },
+    {
+      "epoch": 0.04492379547689282,
+      "grad_norm": 28.5,
+      "learning_rate": 9.733333333333334e-06,
+      "loss": 0.738,
+      "step": 731
+    },
+    {
+      "epoch": 0.04498525073746313,
+      "grad_norm": 26.5,
+      "learning_rate": 9.746666666666668e-06,
+      "loss": 0.6922,
+      "step": 732
+    },
+    {
+      "epoch": 0.04504670599803343,
+      "grad_norm": 50.75,
+      "learning_rate": 9.760000000000001e-06,
+      "loss": 0.7852,
+      "step": 733
+    },
+    {
+      "epoch": 0.045108161258603736,
+      "grad_norm": 39.5,
+      "learning_rate": 9.773333333333335e-06,
+      "loss": 0.8342,
+      "step": 734
+    },
+    {
+      "epoch": 0.04516961651917404,
+      "grad_norm": 27.5,
+      "learning_rate": 9.786666666666667e-06,
+      "loss": 0.6431,
+      "step": 735
+    },
+    {
+      "epoch": 0.04523107177974434,
+      "grad_norm": 27.375,
+      "learning_rate": 9.800000000000001e-06,
+      "loss": 0.757,
+      "step": 736
+    },
+    {
+      "epoch": 0.045292527040314653,
+      "grad_norm": 15.625,
+      "learning_rate": 9.813333333333333e-06,
+      "loss": 0.5835,
+      "step": 737
+    },
+    {
+      "epoch": 0.04535398230088496,
+      "grad_norm": 49.0,
+      "learning_rate": 9.826666666666667e-06,
+      "loss": 0.737,
+      "step": 738
+    },
+    {
+      "epoch": 0.04541543756145526,
+      "grad_norm": 54.0,
+      "learning_rate": 9.84e-06,
+      "loss": 0.661,
+      "step": 739
+    },
+    {
+      "epoch": 0.045476892822025564,
+      "grad_norm": 41.25,
+      "learning_rate": 9.853333333333334e-06,
+      "loss": 0.5831,
+      "step": 740
+    },
+    {
+      "epoch": 0.04553834808259587,
+      "grad_norm": 27.5,
+      "learning_rate": 9.866666666666668e-06,
+      "loss": 0.5602,
+      "step": 741
+    },
+    {
+      "epoch": 0.04559980334316618,
+      "grad_norm": 52.25,
+      "learning_rate": 9.88e-06,
+      "loss": 0.6011,
+      "step": 742
+    },
+    {
+      "epoch": 0.04566125860373648,
+      "grad_norm": 60.25,
+      "learning_rate": 9.893333333333334e-06,
+      "loss": 0.5949,
+      "step": 743
+    },
+    {
+      "epoch": 0.045722713864306784,
+      "grad_norm": 46.0,
+      "learning_rate": 9.906666666666668e-06,
+      "loss": 0.6577,
+      "step": 744
+    },
+    {
+      "epoch": 0.04578416912487709,
+      "grad_norm": 24.5,
+      "learning_rate": 9.920000000000002e-06,
+      "loss": 0.6808,
+      "step": 745
+    },
+    {
+      "epoch": 0.0458456243854474,
+      "grad_norm": 42.25,
+      "learning_rate": 9.933333333333334e-06,
+      "loss": 0.7542,
+      "step": 746
+    },
+    {
+      "epoch": 0.0459070796460177,
+      "grad_norm": 32.0,
+      "learning_rate": 9.946666666666667e-06,
+      "loss": 0.8687,
+      "step": 747
+    },
+    {
+      "epoch": 0.045968534906588004,
+      "grad_norm": 54.0,
+      "learning_rate": 9.960000000000001e-06,
+      "loss": 1.0703,
+      "step": 748
+    },
+    {
+      "epoch": 0.04602999016715831,
+      "grad_norm": 36.75,
+      "learning_rate": 9.973333333333333e-06,
+      "loss": 0.8712,
+      "step": 749
+    },
+    {
+      "epoch": 0.04609144542772861,
+      "grad_norm": 125.5,
+      "learning_rate": 9.986666666666667e-06,
+      "loss": 1.0925,
+      "step": 750
+    },
+    {
+      "epoch": 0.04615290068829892,
+      "grad_norm": 38.5,
+      "learning_rate": 1e-05,
+      "loss": 0.7888,
+      "step": 751
+    },
+    {
+      "epoch": 0.046214355948869225,
+      "grad_norm": 49.0,
+      "learning_rate": 1.0013333333333335e-05,
+      "loss": 0.6424,
+      "step": 752
+    },
+    {
+      "epoch": 0.04627581120943953,
+      "grad_norm": 48.5,
+      "learning_rate": 1.0026666666666667e-05,
+      "loss": 0.7062,
+      "step": 753
+    },
+    {
+      "epoch": 0.04633726647000983,
+      "grad_norm": 53.25,
+      "learning_rate": 1.004e-05,
+      "loss": 0.7203,
+      "step": 754
+    },
+    {
+      "epoch": 0.046398721730580135,
+      "grad_norm": 49.0,
+      "learning_rate": 1.0053333333333334e-05,
+      "loss": 0.8338,
+      "step": 755
+    },
+    {
+      "epoch": 0.046460176991150445,
+      "grad_norm": 77.0,
+      "learning_rate": 1.0066666666666666e-05,
+      "loss": 0.8558,
+      "step": 756
+    },
+    {
+      "epoch": 0.04652163225172075,
+      "grad_norm": 58.0,
+      "learning_rate": 1.008e-05,
+      "loss": 0.7071,
+      "step": 757
+    },
+    {
+      "epoch": 0.04658308751229105,
+      "grad_norm": 51.25,
+      "learning_rate": 1.0093333333333336e-05,
+      "loss": 0.6844,
+      "step": 758
+    },
+    {
+      "epoch": 0.046644542772861355,
+      "grad_norm": 36.0,
+      "learning_rate": 1.0106666666666668e-05,
+      "loss": 0.721,
+      "step": 759
+    },
+    {
+      "epoch": 0.04670599803343166,
+      "grad_norm": 46.75,
+      "learning_rate": 1.0120000000000001e-05,
+      "loss": 0.7316,
+      "step": 760
+    },
+    {
+      "epoch": 0.04676745329400197,
+      "grad_norm": 47.75,
+      "learning_rate": 1.0133333333333335e-05,
+      "loss": 0.7692,
+      "step": 761
+    },
+    {
+      "epoch": 0.04682890855457227,
+      "grad_norm": 37.5,
+      "learning_rate": 1.0146666666666667e-05,
+      "loss": 0.7745,
+      "step": 762
+    },
+    {
+      "epoch": 0.046890363815142576,
+      "grad_norm": 28.75,
+      "learning_rate": 1.0160000000000001e-05,
+      "loss": 0.7728,
+      "step": 763
+    },
+    {
+      "epoch": 0.04695181907571288,
+      "grad_norm": 36.25,
+      "learning_rate": 1.0173333333333335e-05,
+      "loss": 0.6003,
+      "step": 764
+    },
+    {
+      "epoch": 0.04701327433628318,
+      "grad_norm": 20.875,
+      "learning_rate": 1.0186666666666667e-05,
+      "loss": 0.7126,
+      "step": 765
+    },
+    {
+      "epoch": 0.04707472959685349,
+      "grad_norm": 22.75,
+      "learning_rate": 1.02e-05,
+      "loss": 0.5851,
+      "step": 766
+    },
+    {
+      "epoch": 0.047136184857423796,
+      "grad_norm": 32.75,
+      "learning_rate": 1.0213333333333334e-05,
+      "loss": 0.5361,
+      "step": 767
+    },
+    {
+      "epoch": 0.0471976401179941,
+      "grad_norm": 60.75,
+      "learning_rate": 1.0226666666666666e-05,
+      "loss": 0.6993,
+      "step": 768
+    },
+    {
+      "epoch": 0.0472590953785644,
+      "grad_norm": 40.5,
+      "learning_rate": 1.024e-05,
+      "loss": 0.7402,
+      "step": 769
+    },
+    {
+      "epoch": 0.04732055063913471,
+      "grad_norm": 32.75,
+      "learning_rate": 1.0253333333333336e-05,
+      "loss": 0.6583,
+      "step": 770
+    },
+    {
+      "epoch": 0.04738200589970502,
+      "grad_norm": 30.0,
+      "learning_rate": 1.0266666666666668e-05,
+      "loss": 0.6254,
+      "step": 771
+    },
+    {
+      "epoch": 0.04744346116027532,
+      "grad_norm": 51.0,
+      "learning_rate": 1.0280000000000002e-05,
+      "loss": 0.6938,
+      "step": 772
+    },
+    {
+      "epoch": 0.04750491642084562,
+      "grad_norm": 37.75,
+      "learning_rate": 1.0293333333333335e-05,
+      "loss": 0.697,
+      "step": 773
+    },
+    {
+      "epoch": 0.04756637168141593,
+      "grad_norm": 30.0,
+      "learning_rate": 1.0306666666666667e-05,
+      "loss": 0.7263,
+      "step": 774
+    },
+    {
+      "epoch": 0.04762782694198624,
+      "grad_norm": 49.0,
+      "learning_rate": 1.0320000000000001e-05,
+      "loss": 0.6839,
+      "step": 775
+    },
+    {
+      "epoch": 0.04768928220255654,
+      "grad_norm": 50.25,
+      "learning_rate": 1.0333333333333335e-05,
+      "loss": 0.6247,
+      "step": 776
+    },
+    {
+      "epoch": 0.047750737463126844,
+      "grad_norm": 53.0,
+      "learning_rate": 1.0346666666666667e-05,
+      "loss": 0.9384,
+      "step": 777
+    },
+    {
+      "epoch": 0.04781219272369715,
+      "grad_norm": 33.75,
+      "learning_rate": 1.036e-05,
+      "loss": 0.8235,
+      "step": 778
+    },
+    {
+      "epoch": 0.04787364798426745,
+      "grad_norm": 21.875,
+      "learning_rate": 1.0373333333333335e-05,
+      "loss": 0.6401,
+      "step": 779
+    },
+    {
+      "epoch": 0.04793510324483776,
+      "grad_norm": 31.875,
+      "learning_rate": 1.0386666666666667e-05,
+      "loss": 0.6479,
+      "step": 780
+    },
+    {
+      "epoch": 0.047996558505408064,
+      "grad_norm": 39.25,
+      "learning_rate": 1.04e-05,
+      "loss": 0.5009,
+      "step": 781
+    },
+    {
+      "epoch": 0.04805801376597837,
+      "grad_norm": 47.25,
+      "learning_rate": 1.0413333333333336e-05,
+      "loss": 0.7009,
+      "step": 782
+    },
+    {
+      "epoch": 0.04811946902654867,
+      "grad_norm": 28.375,
+      "learning_rate": 1.0426666666666668e-05,
+      "loss": 0.8483,
+      "step": 783
+    },
+    {
+      "epoch": 0.048180924287118974,
+      "grad_norm": 46.5,
+      "learning_rate": 1.0440000000000002e-05,
+      "loss": 0.6512,
+      "step": 784
+    },
+    {
+      "epoch": 0.048242379547689285,
+      "grad_norm": 40.0,
+      "learning_rate": 1.0453333333333334e-05,
+      "loss": 0.7431,
+      "step": 785
+    },
+    {
+      "epoch": 0.04830383480825959,
+      "grad_norm": 55.75,
+      "learning_rate": 1.0466666666666668e-05,
+      "loss": 0.8059,
+      "step": 786
+    },
+    {
+      "epoch": 0.04836529006882989,
+      "grad_norm": 22.875,
+      "learning_rate": 1.0480000000000001e-05,
+      "loss": 0.6876,
+      "step": 787
+    },
+    {
+      "epoch": 0.048426745329400195,
+      "grad_norm": 45.25,
+      "learning_rate": 1.0493333333333333e-05,
+      "loss": 0.6354,
+      "step": 788
+    },
+    {
+      "epoch": 0.0484882005899705,
+      "grad_norm": 36.25,
+      "learning_rate": 1.0506666666666667e-05,
+      "loss": 0.8947,
+      "step": 789
+    },
+    {
+      "epoch": 0.04854965585054081,
+      "grad_norm": 39.5,
+      "learning_rate": 1.0520000000000001e-05,
+      "loss": 0.593,
+      "step": 790
+    },
+    {
+      "epoch": 0.04861111111111111,
+      "grad_norm": 16.125,
+      "learning_rate": 1.0533333333333333e-05,
+      "loss": 0.6071,
+      "step": 791
+    },
+    {
+      "epoch": 0.048672566371681415,
+      "grad_norm": 60.75,
+      "learning_rate": 1.0546666666666667e-05,
+      "loss": 0.6698,
+      "step": 792
+    },
+    {
+      "epoch": 0.04873402163225172,
+      "grad_norm": 41.0,
+      "learning_rate": 1.056e-05,
+      "loss": 0.9218,
+      "step": 793
+    },
+    {
+      "epoch": 0.04879547689282203,
+      "grad_norm": 36.25,
+      "learning_rate": 1.0573333333333333e-05,
+      "loss": 0.8542,
+      "step": 794
+    },
+    {
+      "epoch": 0.04885693215339233,
+      "grad_norm": 35.25,
+      "learning_rate": 1.0586666666666668e-05,
+      "loss": 0.6295,
+      "step": 795
+    },
+    {
+      "epoch": 0.048918387413962636,
+      "grad_norm": 31.5,
+      "learning_rate": 1.0600000000000002e-05,
+      "loss": 0.7025,
+      "step": 796
+    },
+    {
+      "epoch": 0.04897984267453294,
+      "grad_norm": 32.75,
+      "learning_rate": 1.0613333333333334e-05,
+      "loss": 0.6968,
+      "step": 797
+    },
+    {
+      "epoch": 0.04904129793510324,
+      "grad_norm": 31.375,
+      "learning_rate": 1.0626666666666668e-05,
+      "loss": 0.7318,
+      "step": 798
+    },
+    {
+      "epoch": 0.04910275319567355,
+      "grad_norm": 44.5,
+      "learning_rate": 1.0640000000000001e-05,
+      "loss": 0.7562,
+      "step": 799
+    },
+    {
+      "epoch": 0.049164208456243856,
+      "grad_norm": 18.125,
+      "learning_rate": 1.0653333333333334e-05,
+      "loss": 0.6267,
+      "step": 800
+    },
+    {
+      "epoch": 0.04922566371681416,
+      "grad_norm": 36.75,
+      "learning_rate": 1.0666666666666667e-05,
+      "loss": 0.665,
+      "step": 801
+    },
+    {
+      "epoch": 0.04928711897738446,
+      "grad_norm": 75.0,
+      "learning_rate": 1.0680000000000001e-05,
+      "loss": 0.6933,
+      "step": 802
+    },
+    {
+      "epoch": 0.049348574237954766,
+      "grad_norm": 57.75,
+      "learning_rate": 1.0693333333333333e-05,
+      "loss": 0.5902,
+      "step": 803
+    },
+    {
+      "epoch": 0.049410029498525077,
+      "grad_norm": 65.5,
+      "learning_rate": 1.0706666666666667e-05,
+      "loss": 0.6891,
+      "step": 804
+    },
+    {
+      "epoch": 0.04947148475909538,
+      "grad_norm": 51.75,
+      "learning_rate": 1.072e-05,
+      "loss": 0.7041,
+      "step": 805
+    },
+    {
+      "epoch": 0.04953294001966568,
+      "grad_norm": 35.0,
+      "learning_rate": 1.0733333333333333e-05,
+      "loss": 0.6583,
+      "step": 806
+    },
+    {
+      "epoch": 0.04959439528023599,
+      "grad_norm": 37.25,
+      "learning_rate": 1.0746666666666668e-05,
+      "loss": 0.7138,
+      "step": 807
+    },
+    {
+      "epoch": 0.04965585054080629,
+      "grad_norm": 52.25,
+      "learning_rate": 1.0760000000000002e-05,
+      "loss": 0.7472,
+      "step": 808
+    },
+    {
+      "epoch": 0.0497173058013766,
+      "grad_norm": 49.25,
+      "learning_rate": 1.0773333333333334e-05,
+      "loss": 0.6569,
+      "step": 809
+    },
+    {
+      "epoch": 0.049778761061946904,
+      "grad_norm": 50.5,
+      "learning_rate": 1.0786666666666668e-05,
+      "loss": 0.8802,
+      "step": 810
+    },
+    {
+      "epoch": 0.04984021632251721,
+      "grad_norm": 41.25,
+      "learning_rate": 1.0800000000000002e-05,
+      "loss": 0.7542,
+      "step": 811
+    },
+    {
+      "epoch": 0.04990167158308751,
+      "grad_norm": 27.25,
+      "learning_rate": 1.0813333333333334e-05,
+      "loss": 0.5113,
+      "step": 812
+    },
+    {
+      "epoch": 0.049963126843657814,
+      "grad_norm": 23.625,
+      "learning_rate": 1.0826666666666667e-05,
+      "loss": 0.677,
+      "step": 813
+    },
+    {
+      "epoch": 0.050024582104228124,
+      "grad_norm": 44.0,
+      "learning_rate": 1.0840000000000001e-05,
+      "loss": 0.6348,
+      "step": 814
+    },
+    {
+      "epoch": 0.05008603736479843,
+      "grad_norm": 22.375,
+      "learning_rate": 1.0853333333333333e-05,
+      "loss": 0.5962,
+      "step": 815
+    },
+    {
+      "epoch": 0.05014749262536873,
+      "grad_norm": 93.0,
+      "learning_rate": 1.0866666666666667e-05,
+      "loss": 0.6353,
+      "step": 816
+    },
+    {
+      "epoch": 0.050208947885939034,
+      "grad_norm": 65.0,
+      "learning_rate": 1.0880000000000001e-05,
+      "loss": 0.603,
+      "step": 817
+    },
+    {
+      "epoch": 0.050270403146509345,
+      "grad_norm": 44.0,
+      "learning_rate": 1.0893333333333333e-05,
+      "loss": 0.9362,
+      "step": 818
+    },
+    {
+      "epoch": 0.05033185840707965,
+      "grad_norm": 21.75,
+      "learning_rate": 1.0906666666666668e-05,
+      "loss": 0.7394,
+      "step": 819
+    },
+    {
+      "epoch": 0.05039331366764995,
+      "grad_norm": 43.0,
+      "learning_rate": 1.0920000000000002e-05,
+      "loss": 0.6651,
+      "step": 820
+    },
+    {
+      "epoch": 0.050454768928220255,
+      "grad_norm": 45.25,
+      "learning_rate": 1.0933333333333334e-05,
+      "loss": 0.6671,
+      "step": 821
+    },
+    {
+      "epoch": 0.05051622418879056,
+      "grad_norm": 69.0,
+      "learning_rate": 1.0946666666666668e-05,
+      "loss": 0.7588,
+      "step": 822
+    },
+    {
+      "epoch": 0.05057767944936087,
+      "grad_norm": 59.25,
+      "learning_rate": 1.0960000000000002e-05,
+      "loss": 0.7313,
+      "step": 823
+    },
+    {
+      "epoch": 0.05063913470993117,
+      "grad_norm": 44.25,
+      "learning_rate": 1.0973333333333334e-05,
+      "loss": 0.63,
+      "step": 824
+    },
+    {
+      "epoch": 0.050700589970501475,
+      "grad_norm": 45.25,
+      "learning_rate": 1.0986666666666668e-05,
+      "loss": 0.7048,
+      "step": 825
+    },
+    {
+      "epoch": 0.05076204523107178,
+      "grad_norm": 42.25,
+      "learning_rate": 1.1000000000000001e-05,
+      "loss": 0.5822,
+      "step": 826
+    },
+    {
+      "epoch": 0.05082350049164208,
+      "grad_norm": 51.75,
+      "learning_rate": 1.1013333333333333e-05,
+      "loss": 0.7066,
+      "step": 827
+    },
+    {
+      "epoch": 0.05088495575221239,
+      "grad_norm": 55.0,
+      "learning_rate": 1.1026666666666667e-05,
+      "loss": 0.5021,
+      "step": 828
+    },
+    {
+      "epoch": 0.050946411012782696,
+      "grad_norm": 43.75,
+      "learning_rate": 1.1040000000000001e-05,
+      "loss": 0.7969,
+      "step": 829
+    },
+    {
+      "epoch": 0.051007866273353,
+      "grad_norm": 54.5,
+      "learning_rate": 1.1053333333333333e-05,
+      "loss": 0.6311,
+      "step": 830
+    },
+    {
+      "epoch": 0.0510693215339233,
+      "grad_norm": 25.875,
+      "learning_rate": 1.1066666666666669e-05,
+      "loss": 0.6224,
+      "step": 831
+    },
+    {
+      "epoch": 0.051130776794493606,
+      "grad_norm": 49.75,
+      "learning_rate": 1.1080000000000002e-05,
+      "loss": 0.9938,
+      "step": 832
+    },
+    {
+      "epoch": 0.051192232055063916,
+      "grad_norm": 35.0,
+      "learning_rate": 1.1093333333333334e-05,
+      "loss": 0.6986,
+      "step": 833
+    },
+    {
+      "epoch": 0.05125368731563422,
+      "grad_norm": 34.75,
+      "learning_rate": 1.1106666666666668e-05,
+      "loss": 0.7152,
+      "step": 834
+    },
+    {
+      "epoch": 0.05131514257620452,
+      "grad_norm": 20.25,
+      "learning_rate": 1.1120000000000002e-05,
+      "loss": 0.5822,
+      "step": 835
+    },
+    {
+      "epoch": 0.051376597836774826,
+      "grad_norm": 30.875,
+      "learning_rate": 1.1133333333333334e-05,
+      "loss": 0.6785,
+      "step": 836
+    },
+    {
+      "epoch": 0.05143805309734513,
+      "grad_norm": 22.25,
+      "learning_rate": 1.1146666666666668e-05,
+      "loss": 0.5322,
+      "step": 837
+    },
+    {
+      "epoch": 0.05149950835791544,
+      "grad_norm": 47.0,
+      "learning_rate": 1.1160000000000002e-05,
+      "loss": 0.7106,
+      "step": 838
+    },
+    {
+      "epoch": 0.05156096361848574,
+      "grad_norm": 32.75,
+      "learning_rate": 1.1173333333333334e-05,
+      "loss": 0.7764,
+      "step": 839
+    },
+    {
+      "epoch": 0.051622418879056046,
+      "grad_norm": 63.75,
+      "learning_rate": 1.1186666666666667e-05,
+      "loss": 0.9453,
+      "step": 840
+    },
+    {
+      "epoch": 0.05168387413962635,
+      "grad_norm": 27.875,
+      "learning_rate": 1.1200000000000001e-05,
+      "loss": 0.7396,
+      "step": 841
+    },
+    {
+      "epoch": 0.05174532940019666,
+      "grad_norm": 25.75,
+      "learning_rate": 1.1213333333333333e-05,
+      "loss": 0.8482,
+      "step": 842
+    },
+    {
+      "epoch": 0.051806784660766964,
+      "grad_norm": 18.375,
+      "learning_rate": 1.1226666666666669e-05,
+      "loss": 0.6308,
+      "step": 843
+    },
+    {
+      "epoch": 0.05186823992133727,
+      "grad_norm": 22.25,
+      "learning_rate": 1.1240000000000002e-05,
+      "loss": 0.7905,
+      "step": 844
+    },
+    {
+      "epoch": 0.05192969518190757,
+      "grad_norm": 26.0,
+      "learning_rate": 1.1253333333333335e-05,
+      "loss": 0.6793,
+      "step": 845
+    },
+    {
+      "epoch": 0.051991150442477874,
+      "grad_norm": 24.375,
+      "learning_rate": 1.1266666666666668e-05,
+      "loss": 0.6244,
+      "step": 846
+    },
+    {
+      "epoch": 0.052052605703048184,
+      "grad_norm": 44.25,
+      "learning_rate": 1.128e-05,
+      "loss": 0.8591,
+      "step": 847
+    },
+    {
+      "epoch": 0.05211406096361849,
+      "grad_norm": 43.5,
+      "learning_rate": 1.1293333333333334e-05,
+      "loss": 0.7352,
+      "step": 848
+    },
+    {
+      "epoch": 0.05217551622418879,
+      "grad_norm": 41.5,
+      "learning_rate": 1.1306666666666668e-05,
+      "loss": 0.8791,
+      "step": 849
+    },
+    {
+      "epoch": 0.052236971484759094,
+      "grad_norm": 31.5,
+      "learning_rate": 1.132e-05,
+      "loss": 0.8298,
+      "step": 850
+    },
+    {
+      "epoch": 0.0522984267453294,
+      "grad_norm": 43.25,
+      "learning_rate": 1.1333333333333334e-05,
+      "loss": 0.7726,
+      "step": 851
+    },
+    {
+      "epoch": 0.05235988200589971,
+      "grad_norm": 44.75,
+      "learning_rate": 1.1346666666666668e-05,
+      "loss": 0.9036,
+      "step": 852
+    },
+    {
+      "epoch": 0.05242133726647001,
+      "grad_norm": 40.25,
+      "learning_rate": 1.136e-05,
+      "loss": 0.8183,
+      "step": 853
+    },
+    {
+      "epoch": 0.052482792527040315,
+      "grad_norm": 30.375,
+      "learning_rate": 1.1373333333333333e-05,
+      "loss": 0.7978,
+      "step": 854
+    },
+    {
+      "epoch": 0.05254424778761062,
+      "grad_norm": 28.0,
+      "learning_rate": 1.1386666666666669e-05,
+      "loss": 0.9153,
+      "step": 855
+    },
+    {
+      "epoch": 0.05260570304818092,
+      "grad_norm": 29.0,
+      "learning_rate": 1.14e-05,
+      "loss": 0.6611,
+      "step": 856
+    },
+    {
+      "epoch": 0.05266715830875123,
+      "grad_norm": 39.25,
+      "learning_rate": 1.1413333333333335e-05,
+      "loss": 0.7199,
+      "step": 857
+    },
+    {
+      "epoch": 0.052728613569321535,
+      "grad_norm": 40.25,
+      "learning_rate": 1.1426666666666669e-05,
+      "loss": 0.8384,
+      "step": 858
+    },
+    {
+      "epoch": 0.05279006882989184,
+      "grad_norm": 29.0,
+      "learning_rate": 1.144e-05,
+      "loss": 0.7427,
+      "step": 859
+    },
+    {
+      "epoch": 0.05285152409046214,
+      "grad_norm": 30.375,
+      "learning_rate": 1.1453333333333334e-05,
+      "loss": 0.7447,
+      "step": 860
+    },
+    {
+      "epoch": 0.052912979351032445,
+      "grad_norm": 55.0,
+      "learning_rate": 1.1466666666666668e-05,
+      "loss": 0.7126,
+      "step": 861
+    },
+    {
+      "epoch": 0.052974434611602755,
+      "grad_norm": 20.25,
+      "learning_rate": 1.148e-05,
+      "loss": 0.6529,
+      "step": 862
+    },
+    {
+      "epoch": 0.05303588987217306,
+      "grad_norm": 28.625,
+      "learning_rate": 1.1493333333333334e-05,
+      "loss": 0.5448,
+      "step": 863
+    },
+    {
+      "epoch": 0.05309734513274336,
+      "grad_norm": 34.25,
+      "learning_rate": 1.1506666666666668e-05,
+      "loss": 0.5725,
+      "step": 864
+    },
+    {
+      "epoch": 0.053158800393313665,
+      "grad_norm": 41.25,
+      "learning_rate": 1.152e-05,
+      "loss": 0.7562,
+      "step": 865
+    },
+    {
+      "epoch": 0.053220255653883976,
+      "grad_norm": 33.25,
+      "learning_rate": 1.1533333333333334e-05,
+      "loss": 0.9486,
+      "step": 866
+    },
+    {
+      "epoch": 0.05328171091445428,
+      "grad_norm": 26.25,
+      "learning_rate": 1.1546666666666669e-05,
+      "loss": 0.7683,
+      "step": 867
+    },
+    {
+      "epoch": 0.05334316617502458,
+      "grad_norm": 15.375,
+      "learning_rate": 1.156e-05,
+      "loss": 0.5429,
+      "step": 868
+    },
+    {
+      "epoch": 0.053404621435594886,
+      "grad_norm": 30.125,
+      "learning_rate": 1.1573333333333335e-05,
+      "loss": 0.5291,
+      "step": 869
+    },
+    {
+      "epoch": 0.05346607669616519,
+      "grad_norm": 29.5,
+      "learning_rate": 1.1586666666666669e-05,
+      "loss": 0.5687,
+      "step": 870
+    },
+    {
+      "epoch": 0.0535275319567355,
+      "grad_norm": 17.5,
+      "learning_rate": 1.16e-05,
+      "loss": 0.5489,
+      "step": 871
+    },
+    {
+      "epoch": 0.0535889872173058,
+      "grad_norm": 30.125,
+      "learning_rate": 1.1613333333333335e-05,
+      "loss": 0.6197,
+      "step": 872
+    },
+    {
+      "epoch": 0.053650442477876106,
+      "grad_norm": 24.5,
+      "learning_rate": 1.1626666666666668e-05,
+      "loss": 0.7239,
+      "step": 873
+    },
+    {
+      "epoch": 0.05371189773844641,
+      "grad_norm": 29.625,
+      "learning_rate": 1.164e-05,
+      "loss": 0.7267,
+      "step": 874
+    },
+    {
+      "epoch": 0.05377335299901671,
+      "grad_norm": 43.0,
+      "learning_rate": 1.1653333333333334e-05,
+      "loss": 0.6004,
+      "step": 875
+    },
+    {
+      "epoch": 0.05383480825958702,
+      "grad_norm": 34.0,
+      "learning_rate": 1.1666666666666668e-05,
+      "loss": 0.7228,
+      "step": 876
+    },
+    {
+      "epoch": 0.05389626352015733,
+      "grad_norm": 29.125,
+      "learning_rate": 1.168e-05,
+      "loss": 0.6849,
+      "step": 877
+    },
+    {
+      "epoch": 0.05395771878072763,
+      "grad_norm": 29.75,
+      "learning_rate": 1.1693333333333334e-05,
+      "loss": 0.6881,
+      "step": 878
+    },
+    {
+      "epoch": 0.054019174041297933,
+      "grad_norm": 35.0,
+      "learning_rate": 1.170666666666667e-05,
+      "loss": 0.7448,
+      "step": 879
+    },
+    {
+      "epoch": 0.05408062930186824,
+      "grad_norm": 39.0,
+      "learning_rate": 1.172e-05,
+      "loss": 0.8501,
+      "step": 880
+    },
+    {
+      "epoch": 0.05414208456243855,
+      "grad_norm": 35.5,
+      "learning_rate": 1.1733333333333335e-05,
+      "loss": 0.7999,
+      "step": 881
+    },
+    {
+      "epoch": 0.05420353982300885,
+      "grad_norm": 38.5,
+      "learning_rate": 1.1746666666666669e-05,
+      "loss": 0.579,
+      "step": 882
+    },
+    {
+      "epoch": 0.054264995083579154,
+      "grad_norm": 16.25,
+      "learning_rate": 1.1760000000000001e-05,
+      "loss": 0.5202,
+      "step": 883
+    },
+    {
+      "epoch": 0.05432645034414946,
+      "grad_norm": 21.625,
+      "learning_rate": 1.1773333333333335e-05,
+      "loss": 0.5582,
+      "step": 884
+    },
+    {
+      "epoch": 0.05438790560471976,
+      "grad_norm": 32.25,
+      "learning_rate": 1.1786666666666668e-05,
+      "loss": 0.5072,
+      "step": 885
+    },
+    {
+      "epoch": 0.05444936086529007,
+      "grad_norm": 26.375,
+      "learning_rate": 1.18e-05,
+      "loss": 0.5798,
+      "step": 886
+    },
+    {
+      "epoch": 0.054510816125860374,
+      "grad_norm": 30.75,
+      "learning_rate": 1.1813333333333334e-05,
+      "loss": 0.4824,
+      "step": 887
+    },
+    {
+      "epoch": 0.05457227138643068,
+      "grad_norm": 27.375,
+      "learning_rate": 1.1826666666666668e-05,
+      "loss": 0.5179,
+      "step": 888
+    },
+    {
+      "epoch": 0.05463372664700098,
+      "grad_norm": 62.0,
+      "learning_rate": 1.184e-05,
+      "loss": 0.7642,
+      "step": 889
+    },
+    {
+      "epoch": 0.05469518190757129,
+      "grad_norm": 23.25,
+      "learning_rate": 1.1853333333333334e-05,
+      "loss": 0.692,
+      "step": 890
+    },
+    {
+      "epoch": 0.054756637168141595,
+      "grad_norm": 50.0,
+      "learning_rate": 1.186666666666667e-05,
+      "loss": 0.7653,
+      "step": 891
+    },
+    {
+      "epoch": 0.0548180924287119,
+      "grad_norm": 82.5,
+      "learning_rate": 1.188e-05,
+      "loss": 0.7653,
+      "step": 892
+    },
+    {
+      "epoch": 0.0548795476892822,
+      "grad_norm": 68.0,
+      "learning_rate": 1.1893333333333335e-05,
+      "loss": 0.742,
+      "step": 893
+    },
+    {
+      "epoch": 0.054941002949852505,
+      "grad_norm": 69.0,
+      "learning_rate": 1.1906666666666669e-05,
+      "loss": 0.7095,
+      "step": 894
+    },
+    {
+      "epoch": 0.055002458210422815,
+      "grad_norm": 45.25,
+      "learning_rate": 1.1920000000000001e-05,
+      "loss": 0.5097,
+      "step": 895
+    },
+    {
+      "epoch": 0.05506391347099312,
+      "grad_norm": 48.5,
+      "learning_rate": 1.1933333333333335e-05,
+      "loss": 0.5493,
+      "step": 896
+    },
+    {
+      "epoch": 0.05512536873156342,
+      "grad_norm": 40.75,
+      "learning_rate": 1.1946666666666669e-05,
+      "loss": 0.6108,
+      "step": 897
+    },
+    {
+      "epoch": 0.055186823992133725,
+      "grad_norm": 52.25,
+      "learning_rate": 1.196e-05,
+      "loss": 0.6981,
+      "step": 898
+    },
+    {
+      "epoch": 0.05524827925270403,
+      "grad_norm": 42.25,
+      "learning_rate": 1.1973333333333334e-05,
+      "loss": 0.7181,
+      "step": 899
+    },
+    {
+      "epoch": 0.05530973451327434,
+      "grad_norm": 22.625,
+      "learning_rate": 1.1986666666666668e-05,
+      "loss": 0.581,
+      "step": 900
+    },
+    {
+      "epoch": 0.05537118977384464,
+      "grad_norm": 34.75,
+      "learning_rate": 1.2e-05,
+      "loss": 0.6108,
+      "step": 901
+    },
+    {
+      "epoch": 0.055432645034414946,
+      "grad_norm": 29.875,
+      "learning_rate": 1.2013333333333334e-05,
+      "loss": 0.5962,
+      "step": 902
+    },
+    {
+      "epoch": 0.05549410029498525,
+      "grad_norm": 38.5,
+      "learning_rate": 1.202666666666667e-05,
+      "loss": 0.5995,
+      "step": 903
+    },
+    {
+      "epoch": 0.05555555555555555,
+      "grad_norm": 27.625,
+      "learning_rate": 1.204e-05,
+      "loss": 0.6027,
+      "step": 904
+    },
+    {
+      "epoch": 0.05561701081612586,
+      "grad_norm": 27.875,
+      "learning_rate": 1.2053333333333335e-05,
+      "loss": 0.771,
+      "step": 905
+    },
+    {
+      "epoch": 0.055678466076696166,
+      "grad_norm": 33.25,
+      "learning_rate": 1.206666666666667e-05,
+      "loss": 0.9046,
+      "step": 906
+    },
+    {
+      "epoch": 0.05573992133726647,
+      "grad_norm": 31.875,
+      "learning_rate": 1.2080000000000001e-05,
+      "loss": 0.8869,
+      "step": 907
+    },
+    {
+      "epoch": 0.05580137659783677,
+      "grad_norm": 38.25,
+      "learning_rate": 1.2093333333333335e-05,
+      "loss": 0.964,
+      "step": 908
+    },
+    {
+      "epoch": 0.055862831858407076,
+      "grad_norm": 30.5,
+      "learning_rate": 1.2106666666666667e-05,
+      "loss": 1.1805,
+      "step": 909
+    },
+    {
+      "epoch": 0.05592428711897739,
+      "grad_norm": 23.75,
+      "learning_rate": 1.2120000000000001e-05,
+      "loss": 0.74,
+      "step": 910
+    },
+    {
+      "epoch": 0.05598574237954769,
+      "grad_norm": 24.625,
+      "learning_rate": 1.2133333333333335e-05,
+      "loss": 0.8718,
+      "step": 911
+    },
+    {
+      "epoch": 0.05604719764011799,
+      "grad_norm": 25.625,
+      "learning_rate": 1.2146666666666667e-05,
+      "loss": 1.0713,
+      "step": 912
+    },
+    {
+      "epoch": 0.0561086529006883,
+      "grad_norm": 46.0,
+      "learning_rate": 1.216e-05,
+      "loss": 0.8758,
+      "step": 913
+    },
+    {
+      "epoch": 0.05617010816125861,
+      "grad_norm": 26.875,
+      "learning_rate": 1.2173333333333334e-05,
+      "loss": 0.771,
+      "step": 914
+    },
+    {
+      "epoch": 0.05623156342182891,
+      "grad_norm": 22.625,
+      "learning_rate": 1.2186666666666666e-05,
+      "loss": 0.8002,
+      "step": 915
+    },
+    {
+      "epoch": 0.056293018682399214,
+      "grad_norm": 32.25,
+      "learning_rate": 1.22e-05,
+      "loss": 0.6459,
+      "step": 916
+    },
+    {
+      "epoch": 0.05635447394296952,
+      "grad_norm": 25.125,
+      "learning_rate": 1.2213333333333336e-05,
+      "loss": 0.5845,
+      "step": 917
+    },
+    {
+      "epoch": 0.05641592920353982,
+      "grad_norm": 21.25,
+      "learning_rate": 1.2226666666666666e-05,
+      "loss": 0.6145,
+      "step": 918
+    },
+    {
+      "epoch": 0.05647738446411013,
+      "grad_norm": 23.875,
+      "learning_rate": 1.2240000000000001e-05,
+      "loss": 0.6707,
+      "step": 919
+    },
+    {
+      "epoch": 0.056538839724680434,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.2253333333333335e-05,
+      "loss": 0.5965,
+      "step": 920
+    },
+    {
+      "epoch": 0.05660029498525074,
+      "grad_norm": 33.5,
+      "learning_rate": 1.2266666666666667e-05,
+      "loss": 0.5909,
+      "step": 921
+    },
+    {
+      "epoch": 0.05666175024582104,
+      "grad_norm": 28.625,
+      "learning_rate": 1.2280000000000001e-05,
+      "loss": 0.7442,
+      "step": 922
+    },
+    {
+      "epoch": 0.056723205506391344,
+      "grad_norm": 41.25,
+      "learning_rate": 1.2293333333333335e-05,
+      "loss": 0.7375,
+      "step": 923
+    },
+    {
+      "epoch": 0.056784660766961655,
+      "grad_norm": 25.25,
+      "learning_rate": 1.2306666666666667e-05,
+      "loss": 0.5163,
+      "step": 924
+    },
+    {
+      "epoch": 0.05684611602753196,
+      "grad_norm": 37.75,
+      "learning_rate": 1.232e-05,
+      "loss": 0.6513,
+      "step": 925
+    },
+    {
+      "epoch": 0.05690757128810226,
+      "grad_norm": 34.5,
+      "learning_rate": 1.2333333333333334e-05,
+      "loss": 0.7151,
+      "step": 926
+    },
+    {
+      "epoch": 0.056969026548672565,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.2346666666666666e-05,
+      "loss": 0.6596,
+      "step": 927
+    },
+    {
+      "epoch": 0.05703048180924287,
+      "grad_norm": 34.0,
+      "learning_rate": 1.236e-05,
+      "loss": 0.6131,
+      "step": 928
+    },
+    {
+      "epoch": 0.05709193706981318,
+      "grad_norm": 39.25,
+      "learning_rate": 1.2373333333333336e-05,
+      "loss": 0.759,
+      "step": 929
+    },
+    {
+      "epoch": 0.05715339233038348,
+      "grad_norm": 29.25,
+      "learning_rate": 1.2386666666666666e-05,
+      "loss": 0.662,
+      "step": 930
+    },
+    {
+      "epoch": 0.057214847590953785,
+      "grad_norm": 72.0,
+      "learning_rate": 1.2400000000000002e-05,
+      "loss": 1.1421,
+      "step": 931
+    },
+    {
+      "epoch": 0.05727630285152409,
+      "grad_norm": 40.5,
+      "learning_rate": 1.2413333333333335e-05,
+      "loss": 0.7545,
+      "step": 932
+    },
+    {
+      "epoch": 0.05733775811209439,
+      "grad_norm": 50.5,
+      "learning_rate": 1.2426666666666667e-05,
+      "loss": 0.6465,
+      "step": 933
+    },
+    {
+      "epoch": 0.0573992133726647,
+      "grad_norm": 27.125,
+      "learning_rate": 1.2440000000000001e-05,
+      "loss": 0.811,
+      "step": 934
+    },
+    {
+      "epoch": 0.057460668633235006,
+      "grad_norm": 27.875,
+      "learning_rate": 1.2453333333333335e-05,
+      "loss": 0.6222,
+      "step": 935
+    },
+    {
+      "epoch": 0.05752212389380531,
+      "grad_norm": 31.375,
+      "learning_rate": 1.2466666666666667e-05,
+      "loss": 0.6505,
+      "step": 936
+    },
+    {
+      "epoch": 0.05758357915437561,
+      "grad_norm": 27.375,
+      "learning_rate": 1.248e-05,
+      "loss": 0.6831,
+      "step": 937
+    },
+    {
+      "epoch": 0.05764503441494592,
+      "grad_norm": 40.25,
+      "learning_rate": 1.2493333333333335e-05,
+      "loss": 0.7704,
+      "step": 938
+    },
+    {
+      "epoch": 0.057706489675516226,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.2506666666666667e-05,
+      "loss": 0.6004,
+      "step": 939
+    },
+    {
+      "epoch": 0.05776794493608653,
+      "grad_norm": 67.0,
+      "learning_rate": 1.252e-05,
+      "loss": 0.9434,
+      "step": 940
+    },
+    {
+      "epoch": 0.05782940019665683,
+      "grad_norm": 36.25,
+      "learning_rate": 1.2533333333333336e-05,
+      "loss": 0.8295,
+      "step": 941
+    },
+    {
+      "epoch": 0.057890855457227136,
+      "grad_norm": 47.75,
+      "learning_rate": 1.2546666666666666e-05,
+      "loss": 0.6951,
+      "step": 942
+    },
+    {
+      "epoch": 0.057952310717797446,
+      "grad_norm": 17.625,
+      "learning_rate": 1.2560000000000002e-05,
+      "loss": 0.7429,
+      "step": 943
+    },
+    {
+      "epoch": 0.05801376597836775,
+      "grad_norm": 22.25,
+      "learning_rate": 1.2573333333333335e-05,
+      "loss": 0.4897,
+      "step": 944
+    },
+    {
+      "epoch": 0.05807522123893805,
+      "grad_norm": 45.75,
+      "learning_rate": 1.2586666666666668e-05,
+      "loss": 0.7351,
+      "step": 945
+    },
+    {
+      "epoch": 0.058136676499508357,
+      "grad_norm": 34.5,
+      "learning_rate": 1.2600000000000001e-05,
+      "loss": 0.8585,
+      "step": 946
+    },
+    {
+      "epoch": 0.05819813176007866,
+      "grad_norm": 22.875,
+      "learning_rate": 1.2613333333333335e-05,
+      "loss": 0.6328,
+      "step": 947
+    },
+    {
+      "epoch": 0.05825958702064897,
+      "grad_norm": 26.75,
+      "learning_rate": 1.2626666666666667e-05,
+      "loss": 0.7067,
+      "step": 948
+    },
+    {
+      "epoch": 0.058321042281219274,
+      "grad_norm": 23.0,
+      "learning_rate": 1.2640000000000001e-05,
+      "loss": 0.7098,
+      "step": 949
+    },
+    {
+      "epoch": 0.05838249754178958,
+      "grad_norm": 27.375,
+      "learning_rate": 1.2653333333333335e-05,
+      "loss": 0.7714,
+      "step": 950
+    },
+    {
+      "epoch": 0.05844395280235988,
+      "grad_norm": 40.75,
+      "learning_rate": 1.2666666666666667e-05,
+      "loss": 0.6829,
+      "step": 951
+    },
+    {
+      "epoch": 0.058505408062930184,
+      "grad_norm": 22.5,
+      "learning_rate": 1.268e-05,
+      "loss": 0.7709,
+      "step": 952
+    },
+    {
+      "epoch": 0.058566863323500494,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.2693333333333336e-05,
+      "loss": 0.7671,
+      "step": 953
+    },
+    {
+      "epoch": 0.0586283185840708,
+      "grad_norm": 28.75,
+      "learning_rate": 1.2706666666666666e-05,
+      "loss": 0.8198,
+      "step": 954
+    },
+    {
+      "epoch": 0.0586897738446411,
+      "grad_norm": 39.0,
+      "learning_rate": 1.2720000000000002e-05,
+      "loss": 0.754,
+      "step": 955
+    },
+    {
+      "epoch": 0.058751229105211404,
+      "grad_norm": 38.0,
+      "learning_rate": 1.2733333333333336e-05,
+      "loss": 0.8153,
+      "step": 956
+    },
+    {
+      "epoch": 0.05881268436578171,
+      "grad_norm": 19.75,
+      "learning_rate": 1.2746666666666668e-05,
+      "loss": 0.8053,
+      "step": 957
+    },
+    {
+      "epoch": 0.05887413962635202,
+      "grad_norm": 28.0,
+      "learning_rate": 1.2760000000000001e-05,
+      "loss": 0.8798,
+      "step": 958
+    },
+    {
+      "epoch": 0.05893559488692232,
+      "grad_norm": 29.375,
+      "learning_rate": 1.2773333333333335e-05,
+      "loss": 0.6067,
+      "step": 959
+    },
+    {
+      "epoch": 0.058997050147492625,
+      "grad_norm": 46.25,
+      "learning_rate": 1.2786666666666667e-05,
+      "loss": 0.7138,
+      "step": 960
+    },
+    {
+      "epoch": 0.05905850540806293,
+      "grad_norm": 21.375,
+      "learning_rate": 1.2800000000000001e-05,
+      "loss": 0.8096,
+      "step": 961
+    },
+    {
+      "epoch": 0.05911996066863324,
+      "grad_norm": 35.5,
+      "learning_rate": 1.2813333333333335e-05,
+      "loss": 0.9131,
+      "step": 962
+    },
+    {
+      "epoch": 0.05918141592920354,
+      "grad_norm": 27.0,
+      "learning_rate": 1.2826666666666667e-05,
+      "loss": 0.7825,
+      "step": 963
+    },
+    {
+      "epoch": 0.059242871189773845,
+      "grad_norm": 28.5,
+      "learning_rate": 1.284e-05,
+      "loss": 0.7802,
+      "step": 964
+    },
+    {
+      "epoch": 0.05930432645034415,
+      "grad_norm": 29.25,
+      "learning_rate": 1.2853333333333336e-05,
+      "loss": 0.7461,
+      "step": 965
+    },
+    {
+      "epoch": 0.05936578171091445,
+      "grad_norm": 22.0,
+      "learning_rate": 1.2866666666666667e-05,
+      "loss": 0.5613,
+      "step": 966
+    },
+    {
+      "epoch": 0.05942723697148476,
+      "grad_norm": 21.25,
+      "learning_rate": 1.2880000000000002e-05,
+      "loss": 0.6754,
+      "step": 967
+    },
+    {
+      "epoch": 0.059488692232055065,
+      "grad_norm": 18.125,
+      "learning_rate": 1.2893333333333336e-05,
+      "loss": 0.6465,
+      "step": 968
+    },
+    {
+      "epoch": 0.05955014749262537,
+      "grad_norm": 21.125,
+      "learning_rate": 1.2906666666666668e-05,
+      "loss": 0.6107,
+      "step": 969
+    },
+    {
+      "epoch": 0.05961160275319567,
+      "grad_norm": 26.25,
+      "learning_rate": 1.2920000000000002e-05,
+      "loss": 0.7043,
+      "step": 970
+    },
+    {
+      "epoch": 0.059673058013765976,
+      "grad_norm": 20.25,
+      "learning_rate": 1.2933333333333334e-05,
+      "loss": 0.7324,
+      "step": 971
+    },
+    {
+      "epoch": 0.059734513274336286,
+      "grad_norm": 25.0,
+      "learning_rate": 1.2946666666666668e-05,
+      "loss": 0.5054,
+      "step": 972
+    },
+    {
+      "epoch": 0.05979596853490659,
+      "grad_norm": 45.75,
+      "learning_rate": 1.2960000000000001e-05,
+      "loss": 0.7858,
+      "step": 973
+    },
+    {
+      "epoch": 0.05985742379547689,
+      "grad_norm": 34.0,
+      "learning_rate": 1.2973333333333333e-05,
+      "loss": 0.654,
+      "step": 974
+    },
+    {
+      "epoch": 0.059918879056047196,
+      "grad_norm": 43.0,
+      "learning_rate": 1.2986666666666667e-05,
+      "loss": 0.6117,
+      "step": 975
+    },
+    {
+      "epoch": 0.0599803343166175,
+      "grad_norm": 17.75,
+      "learning_rate": 1.3000000000000001e-05,
+      "loss": 0.5343,
+      "step": 976
+    },
+    {
+      "epoch": 0.06004178957718781,
+      "grad_norm": 62.5,
+      "learning_rate": 1.3013333333333333e-05,
+      "loss": 0.6651,
+      "step": 977
+    },
+    {
+      "epoch": 0.06010324483775811,
+      "grad_norm": 62.0,
+      "learning_rate": 1.3026666666666667e-05,
+      "loss": 0.7106,
+      "step": 978
+    },
+    {
+      "epoch": 0.060164700098328416,
+      "grad_norm": 21.5,
+      "learning_rate": 1.3040000000000002e-05,
+      "loss": 0.7053,
+      "step": 979
+    },
+    {
+      "epoch": 0.06022615535889872,
+      "grad_norm": 21.625,
+      "learning_rate": 1.3053333333333333e-05,
+      "loss": 0.7573,
+      "step": 980
+    },
+    {
+      "epoch": 0.06028761061946902,
+      "grad_norm": 17.0,
+      "learning_rate": 1.3066666666666668e-05,
+      "loss": 0.5481,
+      "step": 981
+    },
+    {
+      "epoch": 0.06034906588003933,
+      "grad_norm": 31.0,
+      "learning_rate": 1.3080000000000002e-05,
+      "loss": 0.4952,
+      "step": 982
+    },
+    {
+      "epoch": 0.06041052114060964,
+      "grad_norm": 28.25,
+      "learning_rate": 1.3093333333333334e-05,
+      "loss": 0.5908,
+      "step": 983
+    },
+    {
+      "epoch": 0.06047197640117994,
+      "grad_norm": 11.0,
+      "learning_rate": 1.3106666666666668e-05,
+      "loss": 0.7005,
+      "step": 984
+    },
+    {
+      "epoch": 0.060533431661750244,
+      "grad_norm": 66.5,
+      "learning_rate": 1.3120000000000001e-05,
+      "loss": 0.6408,
+      "step": 985
+    },
+    {
+      "epoch": 0.060594886922320554,
+      "grad_norm": 31.625,
+      "learning_rate": 1.3133333333333334e-05,
+      "loss": 0.7732,
+      "step": 986
+    },
+    {
+      "epoch": 0.06065634218289086,
+      "grad_norm": 38.5,
+      "learning_rate": 1.3146666666666667e-05,
+      "loss": 0.5178,
+      "step": 987
+    },
+    {
+      "epoch": 0.06071779744346116,
+      "grad_norm": 46.75,
+      "learning_rate": 1.3160000000000001e-05,
+      "loss": 0.679,
+      "step": 988
+    },
+    {
+      "epoch": 0.060779252704031464,
+      "grad_norm": 36.5,
+      "learning_rate": 1.3173333333333333e-05,
+      "loss": 0.7539,
+      "step": 989
+    },
+    {
+      "epoch": 0.06084070796460177,
+      "grad_norm": 43.5,
+      "learning_rate": 1.3186666666666667e-05,
+      "loss": 0.5826,
+      "step": 990
+    },
+    {
+      "epoch": 0.06090216322517208,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.3200000000000002e-05,
+      "loss": 0.6359,
+      "step": 991
+    },
+    {
+      "epoch": 0.06096361848574238,
+      "grad_norm": 34.75,
+      "learning_rate": 1.3213333333333333e-05,
+      "loss": 0.6015,
+      "step": 992
+    },
+    {
+      "epoch": 0.061025073746312684,
+      "grad_norm": 20.625,
+      "learning_rate": 1.3226666666666668e-05,
+      "loss": 0.5874,
+      "step": 993
+    },
+    {
+      "epoch": 0.06108652900688299,
+      "grad_norm": 27.25,
+      "learning_rate": 1.3240000000000002e-05,
+      "loss": 0.6563,
+      "step": 994
+    },
+    {
+      "epoch": 0.06114798426745329,
+      "grad_norm": 34.0,
+      "learning_rate": 1.3253333333333334e-05,
+      "loss": 0.6165,
+      "step": 995
+    },
+    {
+      "epoch": 0.0612094395280236,
+      "grad_norm": 22.5,
+      "learning_rate": 1.3266666666666668e-05,
+      "loss": 0.508,
+      "step": 996
+    },
+    {
+      "epoch": 0.061270894788593905,
+      "grad_norm": 32.75,
+      "learning_rate": 1.3280000000000002e-05,
+      "loss": 0.5479,
+      "step": 997
+    },
+    {
+      "epoch": 0.06133235004916421,
+      "grad_norm": 17.375,
+      "learning_rate": 1.3293333333333334e-05,
+      "loss": 0.5173,
+      "step": 998
+    },
+    {
+      "epoch": 0.06139380530973451,
+      "grad_norm": 29.625,
+      "learning_rate": 1.3306666666666667e-05,
+      "loss": 0.5615,
+      "step": 999
+    },
+    {
+      "epoch": 0.061455260570304815,
+      "grad_norm": 51.75,
+      "learning_rate": 1.3320000000000001e-05,
+      "loss": 0.6375,
+      "step": 1000
+    },
+    {
+      "epoch": 0.061516715830875125,
+      "grad_norm": 66.0,
+      "learning_rate": 1.3333333333333333e-05,
+      "loss": 0.8289,
+      "step": 1001
+    },
+    {
+      "epoch": 0.06157817109144543,
+      "grad_norm": 55.75,
+      "learning_rate": 1.3346666666666667e-05,
+      "loss": 0.7942,
+      "step": 1002
+    },
+    {
+      "epoch": 0.06163962635201573,
+      "grad_norm": 31.625,
+      "learning_rate": 1.3360000000000003e-05,
+      "loss": 0.7145,
+      "step": 1003
+    },
+    {
+      "epoch": 0.061701081612586035,
+      "grad_norm": 21.875,
+      "learning_rate": 1.3373333333333333e-05,
+      "loss": 0.5583,
+      "step": 1004
+    },
+    {
+      "epoch": 0.06176253687315634,
+      "grad_norm": 30.75,
+      "learning_rate": 1.3386666666666668e-05,
+      "loss": 0.8524,
+      "step": 1005
+    },
+    {
+      "epoch": 0.06182399213372665,
+      "grad_norm": 54.0,
+      "learning_rate": 1.3400000000000002e-05,
+      "loss": 0.6351,
+      "step": 1006
+    },
+    {
+      "epoch": 0.06188544739429695,
+      "grad_norm": 23.25,
+      "learning_rate": 1.3413333333333334e-05,
+      "loss": 0.4771,
+      "step": 1007
+    },
+    {
+      "epoch": 0.061946902654867256,
+      "grad_norm": 40.25,
+      "learning_rate": 1.3426666666666668e-05,
+      "loss": 0.5642,
+      "step": 1008
+    },
+    {
+      "epoch": 0.06200835791543756,
+      "grad_norm": 29.625,
+      "learning_rate": 1.3440000000000002e-05,
+      "loss": 0.6781,
+      "step": 1009
+    },
+    {
+      "epoch": 0.06206981317600787,
+      "grad_norm": 41.5,
+      "learning_rate": 1.3453333333333334e-05,
+      "loss": 0.7648,
+      "step": 1010
+    },
+    {
+      "epoch": 0.06213126843657817,
+      "grad_norm": 26.625,
+      "learning_rate": 1.3466666666666668e-05,
+      "loss": 0.8041,
+      "step": 1011
+    },
+    {
+      "epoch": 0.062192723697148476,
+      "grad_norm": 19.5,
+      "learning_rate": 1.3480000000000001e-05,
+      "loss": 0.5447,
+      "step": 1012
+    },
+    {
+      "epoch": 0.06225417895771878,
+      "grad_norm": 19.0,
+      "learning_rate": 1.3493333333333333e-05,
+      "loss": 0.5638,
+      "step": 1013
+    },
+    {
+      "epoch": 0.06231563421828908,
+      "grad_norm": 37.0,
+      "learning_rate": 1.3506666666666667e-05,
+      "loss": 0.5458,
+      "step": 1014
+    },
+    {
+      "epoch": 0.06237708947885939,
+      "grad_norm": 51.75,
+      "learning_rate": 1.3520000000000003e-05,
+      "loss": 0.6322,
+      "step": 1015
+    },
+    {
+      "epoch": 0.0624385447394297,
+      "grad_norm": 27.875,
+      "learning_rate": 1.3533333333333333e-05,
+      "loss": 0.6656,
+      "step": 1016
+    },
+    {
+      "epoch": 0.0625,
+      "grad_norm": 30.0,
+      "learning_rate": 1.3546666666666669e-05,
+      "loss": 0.8446,
+      "step": 1017
+    },
+    {
+      "epoch": 0.0625614552605703,
+      "grad_norm": 19.75,
+      "learning_rate": 1.3560000000000002e-05,
+      "loss": 0.777,
+      "step": 1018
+    },
+    {
+      "epoch": 0.0626229105211406,
+      "grad_norm": 46.0,
+      "learning_rate": 1.3573333333333334e-05,
+      "loss": 0.7143,
+      "step": 1019
+    },
+    {
+      "epoch": 0.06268436578171091,
+      "grad_norm": 41.5,
+      "learning_rate": 1.3586666666666668e-05,
+      "loss": 0.7387,
+      "step": 1020
+    },
+    {
+      "epoch": 0.06274582104228121,
+      "grad_norm": 31.875,
+      "learning_rate": 1.3600000000000002e-05,
+      "loss": 0.7029,
+      "step": 1021
+    },
+    {
+      "epoch": 0.06280727630285153,
+      "grad_norm": 34.5,
+      "learning_rate": 1.3613333333333334e-05,
+      "loss": 0.7499,
+      "step": 1022
+    },
+    {
+      "epoch": 0.06286873156342183,
+      "grad_norm": 23.0,
+      "learning_rate": 1.3626666666666668e-05,
+      "loss": 0.6823,
+      "step": 1023
+    },
+    {
+      "epoch": 0.06293018682399214,
+      "grad_norm": 24.625,
+      "learning_rate": 1.3640000000000002e-05,
+      "loss": 0.6845,
+      "step": 1024
+    },
+    {
+      "epoch": 0.06299164208456244,
+      "grad_norm": 35.0,
+      "learning_rate": 1.3653333333333334e-05,
+      "loss": 0.7603,
+      "step": 1025
+    },
+    {
+      "epoch": 0.06305309734513274,
+      "grad_norm": 28.375,
+      "learning_rate": 1.3666666666666667e-05,
+      "loss": 0.7808,
+      "step": 1026
+    },
+    {
+      "epoch": 0.06311455260570305,
+      "grad_norm": 20.75,
+      "learning_rate": 1.3680000000000003e-05,
+      "loss": 0.6171,
+      "step": 1027
+    },
+    {
+      "epoch": 0.06317600786627335,
+      "grad_norm": 30.125,
+      "learning_rate": 1.3693333333333333e-05,
+      "loss": 0.549,
+      "step": 1028
+    },
+    {
+      "epoch": 0.06323746312684365,
+      "grad_norm": 23.5,
+      "learning_rate": 1.3706666666666669e-05,
+      "loss": 0.7087,
+      "step": 1029
+    },
+    {
+      "epoch": 0.06329891838741396,
+      "grad_norm": 18.875,
+      "learning_rate": 1.3720000000000002e-05,
+      "loss": 0.7751,
+      "step": 1030
+    },
+    {
+      "epoch": 0.06336037364798426,
+      "grad_norm": 18.125,
+      "learning_rate": 1.3733333333333335e-05,
+      "loss": 0.6164,
+      "step": 1031
+    },
+    {
+      "epoch": 0.06342182890855458,
+      "grad_norm": 19.0,
+      "learning_rate": 1.3746666666666668e-05,
+      "loss": 0.5633,
+      "step": 1032
+    },
+    {
+      "epoch": 0.06348328416912488,
+      "grad_norm": 21.75,
+      "learning_rate": 1.376e-05,
+      "loss": 0.649,
+      "step": 1033
+    },
+    {
+      "epoch": 0.06354473942969519,
+      "grad_norm": 24.5,
+      "learning_rate": 1.3773333333333334e-05,
+      "loss": 0.764,
+      "step": 1034
+    },
+    {
+      "epoch": 0.06360619469026549,
+      "grad_norm": 46.5,
+      "learning_rate": 1.3786666666666668e-05,
+      "loss": 0.5924,
+      "step": 1035
+    },
+    {
+      "epoch": 0.06366764995083579,
+      "grad_norm": 32.5,
+      "learning_rate": 1.38e-05,
+      "loss": 0.6562,
+      "step": 1036
+    },
+    {
+      "epoch": 0.0637291052114061,
+      "grad_norm": 43.0,
+      "learning_rate": 1.3813333333333334e-05,
+      "loss": 0.731,
+      "step": 1037
+    },
+    {
+      "epoch": 0.0637905604719764,
+      "grad_norm": 25.375,
+      "learning_rate": 1.3826666666666668e-05,
+      "loss": 0.4474,
+      "step": 1038
+    },
+    {
+      "epoch": 0.0638520157325467,
+      "grad_norm": 29.5,
+      "learning_rate": 1.384e-05,
+      "loss": 0.5407,
+      "step": 1039
+    },
+    {
+      "epoch": 0.063913470993117,
+      "grad_norm": 17.75,
+      "learning_rate": 1.3853333333333333e-05,
+      "loss": 0.562,
+      "step": 1040
+    },
+    {
+      "epoch": 0.06397492625368732,
+      "grad_norm": 73.0,
+      "learning_rate": 1.3866666666666669e-05,
+      "loss": 0.6734,
+      "step": 1041
+    },
+    {
+      "epoch": 0.06403638151425763,
+      "grad_norm": 53.5,
+      "learning_rate": 1.3880000000000001e-05,
+      "loss": 0.9431,
+      "step": 1042
+    },
+    {
+      "epoch": 0.06409783677482793,
+      "grad_norm": 13.625,
+      "learning_rate": 1.3893333333333335e-05,
+      "loss": 0.5406,
+      "step": 1043
+    },
+    {
+      "epoch": 0.06415929203539823,
+      "grad_norm": 60.5,
+      "learning_rate": 1.3906666666666668e-05,
+      "loss": 0.7386,
+      "step": 1044
+    },
+    {
+      "epoch": 0.06422074729596854,
+      "grad_norm": 74.5,
+      "learning_rate": 1.392e-05,
+      "loss": 0.6873,
+      "step": 1045
+    },
+    {
+      "epoch": 0.06428220255653884,
+      "grad_norm": 45.5,
+      "learning_rate": 1.3933333333333334e-05,
+      "loss": 0.5061,
+      "step": 1046
+    },
+    {
+      "epoch": 0.06434365781710914,
+      "grad_norm": 41.5,
+      "learning_rate": 1.3946666666666668e-05,
+      "loss": 0.6824,
+      "step": 1047
+    },
+    {
+      "epoch": 0.06440511307767945,
+      "grad_norm": 34.0,
+      "learning_rate": 1.396e-05,
+      "loss": 0.6991,
+      "step": 1048
+    },
+    {
+      "epoch": 0.06446656833824975,
+      "grad_norm": 25.125,
+      "learning_rate": 1.3973333333333334e-05,
+      "loss": 0.7031,
+      "step": 1049
+    },
+    {
+      "epoch": 0.06452802359882005,
+      "grad_norm": 40.0,
+      "learning_rate": 1.3986666666666668e-05,
+      "loss": 0.7898,
+      "step": 1050
+    },
+    {
+      "epoch": 0.06458947885939037,
+      "grad_norm": 17.25,
+      "learning_rate": 1.4e-05,
+      "loss": 0.6144,
+      "step": 1051
+    },
+    {
+      "epoch": 0.06465093411996067,
+      "grad_norm": 11.25,
+      "learning_rate": 1.4013333333333334e-05,
+      "loss": 0.5525,
+      "step": 1052
+    },
+    {
+      "epoch": 0.06471238938053098,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.4026666666666669e-05,
+      "loss": 0.4986,
+      "step": 1053
+    },
+    {
+      "epoch": 0.06477384464110128,
+      "grad_norm": 26.625,
+      "learning_rate": 1.4040000000000001e-05,
+      "loss": 0.5708,
+      "step": 1054
+    },
+    {
+      "epoch": 0.06483529990167158,
+      "grad_norm": 61.0,
+      "learning_rate": 1.4053333333333335e-05,
+      "loss": 0.5904,
+      "step": 1055
+    },
+    {
+      "epoch": 0.06489675516224189,
+      "grad_norm": 16.125,
+      "learning_rate": 1.4066666666666669e-05,
+      "loss": 0.5787,
+      "step": 1056
+    },
+    {
+      "epoch": 0.06495821042281219,
+      "grad_norm": 32.75,
+      "learning_rate": 1.408e-05,
+      "loss": 0.5654,
+      "step": 1057
+    },
+    {
+      "epoch": 0.0650196656833825,
+      "grad_norm": 40.5,
+      "learning_rate": 1.4093333333333334e-05,
+      "loss": 0.4767,
+      "step": 1058
+    },
+    {
+      "epoch": 0.0650811209439528,
+      "grad_norm": 7.71875,
+      "learning_rate": 1.4106666666666668e-05,
+      "loss": 0.5948,
+      "step": 1059
+    },
+    {
+      "epoch": 0.0651425762045231,
+      "grad_norm": 22.5,
+      "learning_rate": 1.412e-05,
+      "loss": 0.5722,
+      "step": 1060
+    },
+    {
+      "epoch": 0.06520403146509342,
+      "grad_norm": 40.0,
+      "learning_rate": 1.4133333333333334e-05,
+      "loss": 0.5792,
+      "step": 1061
+    },
+    {
+      "epoch": 0.06526548672566372,
+      "grad_norm": 21.625,
+      "learning_rate": 1.4146666666666668e-05,
+      "loss": 0.5644,
+      "step": 1062
+    },
+    {
+      "epoch": 0.06532694198623402,
+      "grad_norm": 61.75,
+      "learning_rate": 1.416e-05,
+      "loss": 0.5862,
+      "step": 1063
+    },
+    {
+      "epoch": 0.06538839724680433,
+      "grad_norm": 91.0,
+      "learning_rate": 1.4173333333333334e-05,
+      "loss": 0.6101,
+      "step": 1064
+    },
+    {
+      "epoch": 0.06544985250737463,
+      "grad_norm": 69.5,
+      "learning_rate": 1.418666666666667e-05,
+      "loss": 0.6225,
+      "step": 1065
+    },
+    {
+      "epoch": 0.06551130776794493,
+      "grad_norm": 38.25,
+      "learning_rate": 1.4200000000000001e-05,
+      "loss": 0.7773,
+      "step": 1066
+    },
+    {
+      "epoch": 0.06557276302851524,
+      "grad_norm": 48.0,
+      "learning_rate": 1.4213333333333335e-05,
+      "loss": 0.7213,
+      "step": 1067
+    },
+    {
+      "epoch": 0.06563421828908554,
+      "grad_norm": 47.25,
+      "learning_rate": 1.4226666666666669e-05,
+      "loss": 0.7991,
+      "step": 1068
+    },
+    {
+      "epoch": 0.06569567354965584,
+      "grad_norm": 89.0,
+      "learning_rate": 1.4240000000000001e-05,
+      "loss": 0.7871,
+      "step": 1069
+    },
+    {
+      "epoch": 0.06575712881022616,
+      "grad_norm": 73.5,
+      "learning_rate": 1.4253333333333335e-05,
+      "loss": 0.7182,
+      "step": 1070
+    },
+    {
+      "epoch": 0.06581858407079647,
+      "grad_norm": 41.75,
+      "learning_rate": 1.4266666666666668e-05,
+      "loss": 0.6139,
+      "step": 1071
+    },
+    {
+      "epoch": 0.06588003933136677,
+      "grad_norm": 66.0,
+      "learning_rate": 1.428e-05,
+      "loss": 0.8248,
+      "step": 1072
+    },
+    {
+      "epoch": 0.06594149459193707,
+      "grad_norm": 72.5,
+      "learning_rate": 1.4293333333333334e-05,
+      "loss": 0.8806,
+      "step": 1073
+    },
+    {
+      "epoch": 0.06600294985250738,
+      "grad_norm": 54.75,
+      "learning_rate": 1.4306666666666668e-05,
+      "loss": 1.0585,
+      "step": 1074
+    },
+    {
+      "epoch": 0.06606440511307768,
+      "grad_norm": 24.625,
+      "learning_rate": 1.432e-05,
+      "loss": 0.6862,
+      "step": 1075
+    },
+    {
+      "epoch": 0.06612586037364798,
+      "grad_norm": 35.25,
+      "learning_rate": 1.4333333333333334e-05,
+      "loss": 0.7819,
+      "step": 1076
+    },
+    {
+      "epoch": 0.06618731563421829,
+      "grad_norm": 32.0,
+      "learning_rate": 1.434666666666667e-05,
+      "loss": 0.7898,
+      "step": 1077
+    },
+    {
+      "epoch": 0.06624877089478859,
+      "grad_norm": 29.625,
+      "learning_rate": 1.4360000000000001e-05,
+      "loss": 0.601,
+      "step": 1078
+    },
+    {
+      "epoch": 0.06631022615535889,
+      "grad_norm": 28.5,
+      "learning_rate": 1.4373333333333335e-05,
+      "loss": 0.6504,
+      "step": 1079
+    },
+    {
+      "epoch": 0.06637168141592921,
+      "grad_norm": 30.25,
+      "learning_rate": 1.4386666666666669e-05,
+      "loss": 0.6407,
+      "step": 1080
+    },
+    {
+      "epoch": 0.06643313667649951,
+      "grad_norm": 21.625,
+      "learning_rate": 1.4400000000000001e-05,
+      "loss": 0.6019,
+      "step": 1081
+    },
+    {
+      "epoch": 0.06649459193706982,
+      "grad_norm": 18.75,
+      "learning_rate": 1.4413333333333335e-05,
+      "loss": 0.7393,
+      "step": 1082
+    },
+    {
+      "epoch": 0.06655604719764012,
+      "grad_norm": 21.75,
+      "learning_rate": 1.4426666666666669e-05,
+      "loss": 0.5901,
+      "step": 1083
+    },
+    {
+      "epoch": 0.06661750245821042,
+      "grad_norm": 26.25,
+      "learning_rate": 1.444e-05,
+      "loss": 0.5716,
+      "step": 1084
+    },
+    {
+      "epoch": 0.06667895771878073,
+      "grad_norm": 20.375,
+      "learning_rate": 1.4453333333333334e-05,
+      "loss": 0.5058,
+      "step": 1085
+    },
+    {
+      "epoch": 0.06674041297935103,
+      "grad_norm": 18.375,
+      "learning_rate": 1.4466666666666668e-05,
+      "loss": 0.6311,
+      "step": 1086
+    },
+    {
+      "epoch": 0.06680186823992133,
+      "grad_norm": 11.25,
+      "learning_rate": 1.448e-05,
+      "loss": 0.4169,
+      "step": 1087
+    },
+    {
+      "epoch": 0.06686332350049164,
+      "grad_norm": 10.625,
+      "learning_rate": 1.4493333333333334e-05,
+      "loss": 0.4954,
+      "step": 1088
+    },
+    {
+      "epoch": 0.06692477876106195,
+      "grad_norm": 32.0,
+      "learning_rate": 1.450666666666667e-05,
+      "loss": 0.7037,
+      "step": 1089
+    },
+    {
+      "epoch": 0.06698623402163226,
+      "grad_norm": 21.375,
+      "learning_rate": 1.4520000000000002e-05,
+      "loss": 0.7401,
+      "step": 1090
+    },
+    {
+      "epoch": 0.06704768928220256,
+      "grad_norm": 7.625,
+      "learning_rate": 1.4533333333333335e-05,
+      "loss": 0.6295,
+      "step": 1091
+    },
+    {
+      "epoch": 0.06710914454277286,
+      "grad_norm": 56.0,
+      "learning_rate": 1.4546666666666669e-05,
+      "loss": 0.637,
+      "step": 1092
+    },
+    {
+      "epoch": 0.06717059980334317,
+      "grad_norm": 56.75,
+      "learning_rate": 1.4560000000000001e-05,
+      "loss": 0.6713,
+      "step": 1093
+    },
+    {
+      "epoch": 0.06723205506391347,
+      "grad_norm": 31.125,
+      "learning_rate": 1.4573333333333335e-05,
+      "loss": 0.4808,
+      "step": 1094
+    },
+    {
+      "epoch": 0.06729351032448377,
+      "grad_norm": 16.5,
+      "learning_rate": 1.4586666666666667e-05,
+      "loss": 0.3933,
+      "step": 1095
+    },
+    {
+      "epoch": 0.06735496558505408,
+      "grad_norm": 36.0,
+      "learning_rate": 1.46e-05,
+      "loss": 0.4986,
+      "step": 1096
+    },
+    {
+      "epoch": 0.06741642084562438,
+      "grad_norm": 50.75,
+      "learning_rate": 1.4613333333333335e-05,
+      "loss": 0.5888,
+      "step": 1097
+    },
+    {
+      "epoch": 0.06747787610619468,
+      "grad_norm": 91.0,
+      "learning_rate": 1.4626666666666667e-05,
+      "loss": 0.7183,
+      "step": 1098
+    },
+    {
+      "epoch": 0.067539331366765,
+      "grad_norm": 64.5,
+      "learning_rate": 1.464e-05,
+      "loss": 0.7257,
+      "step": 1099
+    },
+    {
+      "epoch": 0.0676007866273353,
+      "grad_norm": 48.75,
+      "learning_rate": 1.4653333333333334e-05,
+      "loss": 0.8161,
+      "step": 1100
+    },
+    {
+      "epoch": 0.06766224188790561,
+      "grad_norm": 34.25,
+      "learning_rate": 1.4666666666666666e-05,
+      "loss": 0.7636,
+      "step": 1101
+    },
+    {
+      "epoch": 0.06772369714847591,
+      "grad_norm": 45.75,
+      "learning_rate": 1.4680000000000002e-05,
+      "loss": 0.6096,
+      "step": 1102
+    },
+    {
+      "epoch": 0.06778515240904621,
+      "grad_norm": 38.75,
+      "learning_rate": 1.4693333333333336e-05,
+      "loss": 0.7142,
+      "step": 1103
+    },
+    {
+      "epoch": 0.06784660766961652,
+      "grad_norm": 34.0,
+      "learning_rate": 1.4706666666666668e-05,
+      "loss": 0.6443,
+      "step": 1104
+    },
+    {
+      "epoch": 0.06790806293018682,
+      "grad_norm": 51.5,
+      "learning_rate": 1.4720000000000001e-05,
+      "loss": 0.7447,
+      "step": 1105
+    },
+    {
+      "epoch": 0.06796951819075712,
+      "grad_norm": 37.25,
+      "learning_rate": 1.4733333333333335e-05,
+      "loss": 0.8298,
+      "step": 1106
+    },
+    {
+      "epoch": 0.06803097345132743,
+      "grad_norm": 40.75,
+      "learning_rate": 1.4746666666666667e-05,
+      "loss": 0.7638,
+      "step": 1107
+    },
+    {
+      "epoch": 0.06809242871189773,
+      "grad_norm": 58.75,
+      "learning_rate": 1.4760000000000001e-05,
+      "loss": 0.7746,
+      "step": 1108
+    },
+    {
+      "epoch": 0.06815388397246805,
+      "grad_norm": 43.5,
+      "learning_rate": 1.4773333333333335e-05,
+      "loss": 0.9173,
+      "step": 1109
+    },
+    {
+      "epoch": 0.06821533923303835,
+      "grad_norm": 32.0,
+      "learning_rate": 1.4786666666666667e-05,
+      "loss": 0.7881,
+      "step": 1110
+    },
+    {
+      "epoch": 0.06827679449360866,
+      "grad_norm": 35.5,
+      "learning_rate": 1.48e-05,
+      "loss": 0.735,
+      "step": 1111
+    },
+    {
+      "epoch": 0.06833824975417896,
+      "grad_norm": 21.25,
+      "learning_rate": 1.4813333333333334e-05,
+      "loss": 0.6729,
+      "step": 1112
+    },
+    {
+      "epoch": 0.06839970501474926,
+      "grad_norm": 20.5,
+      "learning_rate": 1.4826666666666666e-05,
+      "loss": 1.0147,
+      "step": 1113
+    },
+    {
+      "epoch": 0.06846116027531957,
+      "grad_norm": 30.125,
+      "learning_rate": 1.4840000000000002e-05,
+      "loss": 0.9717,
+      "step": 1114
+    },
+    {
+      "epoch": 0.06852261553588987,
+      "grad_norm": 36.5,
+      "learning_rate": 1.4853333333333336e-05,
+      "loss": 0.7154,
+      "step": 1115
+    },
+    {
+      "epoch": 0.06858407079646017,
+      "grad_norm": 44.25,
+      "learning_rate": 1.4866666666666668e-05,
+      "loss": 1.0762,
+      "step": 1116
+    },
+    {
+      "epoch": 0.06864552605703048,
+      "grad_norm": 42.5,
+      "learning_rate": 1.4880000000000002e-05,
+      "loss": 0.8782,
+      "step": 1117
+    },
+    {
+      "epoch": 0.0687069813176008,
+      "grad_norm": 39.75,
+      "learning_rate": 1.4893333333333335e-05,
+      "loss": 0.9667,
+      "step": 1118
+    },
+    {
+      "epoch": 0.0687684365781711,
+      "grad_norm": 32.75,
+      "learning_rate": 1.4906666666666667e-05,
+      "loss": 0.8376,
+      "step": 1119
+    },
+    {
+      "epoch": 0.0688298918387414,
+      "grad_norm": 38.25,
+      "learning_rate": 1.4920000000000001e-05,
+      "loss": 0.8856,
+      "step": 1120
+    },
+    {
+      "epoch": 0.0688913470993117,
+      "grad_norm": 40.5,
+      "learning_rate": 1.4933333333333335e-05,
+      "loss": 0.7576,
+      "step": 1121
+    },
+    {
+      "epoch": 0.068952802359882,
+      "grad_norm": 31.5,
+      "learning_rate": 1.4946666666666667e-05,
+      "loss": 0.5704,
+      "step": 1122
+    },
+    {
+      "epoch": 0.06901425762045231,
+      "grad_norm": 29.625,
+      "learning_rate": 1.496e-05,
+      "loss": 0.6238,
+      "step": 1123
+    },
+    {
+      "epoch": 0.06907571288102261,
+      "grad_norm": 28.375,
+      "learning_rate": 1.4973333333333335e-05,
+      "loss": 0.6668,
+      "step": 1124
+    },
+    {
+      "epoch": 0.06913716814159292,
+      "grad_norm": 35.5,
+      "learning_rate": 1.4986666666666667e-05,
+      "loss": 0.8984,
+      "step": 1125
+    },
+    {
+      "epoch": 0.06919862340216322,
+      "grad_norm": 27.125,
+      "learning_rate": 1.5000000000000002e-05,
+      "loss": 0.5605,
+      "step": 1126
+    },
+    {
+      "epoch": 0.06926007866273352,
+      "grad_norm": 33.75,
+      "learning_rate": 1.5013333333333336e-05,
+      "loss": 0.6231,
+      "step": 1127
+    },
+    {
+      "epoch": 0.06932153392330384,
+      "grad_norm": 27.5,
+      "learning_rate": 1.5026666666666668e-05,
+      "loss": 0.5572,
+      "step": 1128
+    },
+    {
+      "epoch": 0.06938298918387414,
+      "grad_norm": 41.0,
+      "learning_rate": 1.5040000000000002e-05,
+      "loss": 0.938,
+      "step": 1129
+    },
+    {
+      "epoch": 0.06944444444444445,
+      "grad_norm": 30.125,
+      "learning_rate": 1.5053333333333335e-05,
+      "loss": 0.6742,
+      "step": 1130
+    },
+    {
+      "epoch": 0.06950589970501475,
+      "grad_norm": 7.71875,
+      "learning_rate": 1.5066666666666668e-05,
+      "loss": 0.657,
+      "step": 1131
+    },
+    {
+      "epoch": 0.06956735496558505,
+      "grad_norm": 46.0,
+      "learning_rate": 1.5080000000000001e-05,
+      "loss": 0.7191,
+      "step": 1132
+    },
+    {
+      "epoch": 0.06962881022615536,
+      "grad_norm": 25.625,
+      "learning_rate": 1.5093333333333335e-05,
+      "loss": 0.7605,
+      "step": 1133
+    },
+    {
+      "epoch": 0.06969026548672566,
+      "grad_norm": 38.0,
+      "learning_rate": 1.5106666666666667e-05,
+      "loss": 0.6979,
+      "step": 1134
+    },
+    {
+      "epoch": 0.06975172074729596,
+      "grad_norm": 34.25,
+      "learning_rate": 1.5120000000000001e-05,
+      "loss": 0.7091,
+      "step": 1135
+    },
+    {
+      "epoch": 0.06981317600786627,
+      "grad_norm": 32.25,
+      "learning_rate": 1.5133333333333335e-05,
+      "loss": 0.5476,
+      "step": 1136
+    },
+    {
+      "epoch": 0.06987463126843659,
+      "grad_norm": 26.375,
+      "learning_rate": 1.5146666666666667e-05,
+      "loss": 0.6273,
+      "step": 1137
+    },
+    {
+      "epoch": 0.06993608652900689,
+      "grad_norm": 30.875,
+      "learning_rate": 1.516e-05,
+      "loss": 0.8992,
+      "step": 1138
+    },
+    {
+      "epoch": 0.06999754178957719,
+      "grad_norm": 30.25,
+      "learning_rate": 1.5173333333333336e-05,
+      "loss": 0.87,
+      "step": 1139
+    },
+    {
+      "epoch": 0.0700589970501475,
+      "grad_norm": 53.0,
+      "learning_rate": 1.5186666666666668e-05,
+      "loss": 0.5385,
+      "step": 1140
+    },
+    {
+      "epoch": 0.0701204523107178,
+      "grad_norm": 55.25,
+      "learning_rate": 1.5200000000000002e-05,
+      "loss": 0.6953,
+      "step": 1141
+    },
+    {
+      "epoch": 0.0701819075712881,
+      "grad_norm": 21.25,
+      "learning_rate": 1.5213333333333336e-05,
+      "loss": 0.4616,
+      "step": 1142
+    },
+    {
+      "epoch": 0.0702433628318584,
+      "grad_norm": 32.25,
+      "learning_rate": 1.5226666666666668e-05,
+      "loss": 0.6059,
+      "step": 1143
+    },
+    {
+      "epoch": 0.07030481809242871,
+      "grad_norm": 36.25,
+      "learning_rate": 1.5240000000000001e-05,
+      "loss": 0.6232,
+      "step": 1144
+    },
+    {
+      "epoch": 0.07036627335299901,
+      "grad_norm": 51.5,
+      "learning_rate": 1.5253333333333335e-05,
+      "loss": 0.6747,
+      "step": 1145
+    },
+    {
+      "epoch": 0.07042772861356932,
+      "grad_norm": 23.625,
+      "learning_rate": 1.5266666666666667e-05,
+      "loss": 0.6227,
+      "step": 1146
+    },
+    {
+      "epoch": 0.07048918387413963,
+      "grad_norm": 19.375,
+      "learning_rate": 1.5280000000000003e-05,
+      "loss": 0.6565,
+      "step": 1147
+    },
+    {
+      "epoch": 0.07055063913470994,
+      "grad_norm": 48.5,
+      "learning_rate": 1.5293333333333335e-05,
+      "loss": 0.5863,
+      "step": 1148
+    },
+    {
+      "epoch": 0.07061209439528024,
+      "grad_norm": 37.25,
+      "learning_rate": 1.5306666666666667e-05,
+      "loss": 0.5731,
+      "step": 1149
+    },
+    {
+      "epoch": 0.07067354965585054,
+      "grad_norm": 48.0,
+      "learning_rate": 1.5320000000000002e-05,
+      "loss": 0.6599,
+      "step": 1150
+    },
+    {
+      "epoch": 0.07073500491642085,
+      "grad_norm": 42.0,
+      "learning_rate": 1.5333333333333334e-05,
+      "loss": 0.6357,
+      "step": 1151
+    },
+    {
+      "epoch": 0.07079646017699115,
+      "grad_norm": 35.25,
+      "learning_rate": 1.5346666666666667e-05,
+      "loss": 0.7379,
+      "step": 1152
+    },
+    {
+      "epoch": 0.07085791543756145,
+      "grad_norm": 23.125,
+      "learning_rate": 1.5360000000000002e-05,
+      "loss": 0.7713,
+      "step": 1153
+    },
+    {
+      "epoch": 0.07091937069813176,
+      "grad_norm": 35.75,
+      "learning_rate": 1.5373333333333334e-05,
+      "loss": 0.7325,
+      "step": 1154
+    },
+    {
+      "epoch": 0.07098082595870206,
+      "grad_norm": 35.0,
+      "learning_rate": 1.5386666666666666e-05,
+      "loss": 0.7437,
+      "step": 1155
+    },
+    {
+      "epoch": 0.07104228121927236,
+      "grad_norm": 42.5,
+      "learning_rate": 1.54e-05,
+      "loss": 0.7071,
+      "step": 1156
+    },
+    {
+      "epoch": 0.07110373647984268,
+      "grad_norm": 45.0,
+      "learning_rate": 1.5413333333333337e-05,
+      "loss": 0.6112,
+      "step": 1157
+    },
+    {
+      "epoch": 0.07116519174041298,
+      "grad_norm": 54.0,
+      "learning_rate": 1.542666666666667e-05,
+      "loss": 0.685,
+      "step": 1158
+    },
+    {
+      "epoch": 0.07122664700098329,
+      "grad_norm": 43.0,
+      "learning_rate": 1.544e-05,
+      "loss": 0.6106,
+      "step": 1159
+    },
+    {
+      "epoch": 0.07128810226155359,
+      "grad_norm": 26.75,
+      "learning_rate": 1.5453333333333333e-05,
+      "loss": 0.534,
+      "step": 1160
+    },
+    {
+      "epoch": 0.0713495575221239,
+      "grad_norm": 46.25,
+      "learning_rate": 1.546666666666667e-05,
+      "loss": 0.6219,
+      "step": 1161
+    },
+    {
+      "epoch": 0.0714110127826942,
+      "grad_norm": 16.625,
+      "learning_rate": 1.548e-05,
+      "loss": 0.4795,
+      "step": 1162
+    },
+    {
+      "epoch": 0.0714724680432645,
+      "grad_norm": 58.5,
+      "learning_rate": 1.5493333333333333e-05,
+      "loss": 0.6535,
+      "step": 1163
+    },
+    {
+      "epoch": 0.0715339233038348,
+      "grad_norm": 21.0,
+      "learning_rate": 1.550666666666667e-05,
+      "loss": 0.6997,
+      "step": 1164
+    },
+    {
+      "epoch": 0.07159537856440511,
+      "grad_norm": 29.25,
+      "learning_rate": 1.552e-05,
+      "loss": 0.6972,
+      "step": 1165
+    },
+    {
+      "epoch": 0.07165683382497542,
+      "grad_norm": 32.0,
+      "learning_rate": 1.5533333333333333e-05,
+      "loss": 0.7933,
+      "step": 1166
+    },
+    {
+      "epoch": 0.07171828908554573,
+      "grad_norm": 12.875,
+      "learning_rate": 1.5546666666666668e-05,
+      "loss": 0.6228,
+      "step": 1167
+    },
+    {
+      "epoch": 0.07177974434611603,
+      "grad_norm": 36.0,
+      "learning_rate": 1.556e-05,
+      "loss": 0.5528,
+      "step": 1168
+    },
+    {
+      "epoch": 0.07184119960668633,
+      "grad_norm": 20.375,
+      "learning_rate": 1.5573333333333332e-05,
+      "loss": 0.6039,
+      "step": 1169
+    },
+    {
+      "epoch": 0.07190265486725664,
+      "grad_norm": 22.625,
+      "learning_rate": 1.5586666666666668e-05,
+      "loss": 0.5609,
+      "step": 1170
+    },
+    {
+      "epoch": 0.07196411012782694,
+      "grad_norm": 7.78125,
+      "learning_rate": 1.5600000000000003e-05,
+      "loss": 0.4003,
+      "step": 1171
+    },
+    {
+      "epoch": 0.07202556538839724,
+      "grad_norm": 25.0,
+      "learning_rate": 1.5613333333333335e-05,
+      "loss": 0.3726,
+      "step": 1172
+    },
+    {
+      "epoch": 0.07208702064896755,
+      "grad_norm": 46.25,
+      "learning_rate": 1.5626666666666667e-05,
+      "loss": 0.462,
+      "step": 1173
+    },
+    {
+      "epoch": 0.07214847590953785,
+      "grad_norm": 50.25,
+      "learning_rate": 1.5640000000000003e-05,
+      "loss": 0.4438,
+      "step": 1174
+    },
+    {
+      "epoch": 0.07220993117010815,
+      "grad_norm": 16.875,
+      "learning_rate": 1.5653333333333335e-05,
+      "loss": 0.4034,
+      "step": 1175
+    },
+    {
+      "epoch": 0.07227138643067847,
+      "grad_norm": 27.125,
+      "learning_rate": 1.5666666666666667e-05,
+      "loss": 0.4843,
+      "step": 1176
+    },
+    {
+      "epoch": 0.07233284169124878,
+      "grad_norm": 60.5,
+      "learning_rate": 1.5680000000000002e-05,
+      "loss": 0.6719,
+      "step": 1177
+    },
+    {
+      "epoch": 0.07239429695181908,
+      "grad_norm": 34.5,
+      "learning_rate": 1.5693333333333334e-05,
+      "loss": 0.7294,
+      "step": 1178
+    },
+    {
+      "epoch": 0.07245575221238938,
+      "grad_norm": 30.625,
+      "learning_rate": 1.5706666666666666e-05,
+      "loss": 0.5936,
+      "step": 1179
+    },
+    {
+      "epoch": 0.07251720747295969,
+      "grad_norm": 68.0,
+      "learning_rate": 1.5720000000000002e-05,
+      "loss": 0.67,
+      "step": 1180
+    },
+    {
+      "epoch": 0.07257866273352999,
+      "grad_norm": 38.25,
+      "learning_rate": 1.5733333333333334e-05,
+      "loss": 0.5604,
+      "step": 1181
+    },
+    {
+      "epoch": 0.07264011799410029,
+      "grad_norm": 67.0,
+      "learning_rate": 1.574666666666667e-05,
+      "loss": 0.5524,
+      "step": 1182
+    },
+    {
+      "epoch": 0.0727015732546706,
+      "grad_norm": 67.0,
+      "learning_rate": 1.576e-05,
+      "loss": 0.4872,
+      "step": 1183
+    },
+    {
+      "epoch": 0.0727630285152409,
+      "grad_norm": 26.875,
+      "learning_rate": 1.5773333333333334e-05,
+      "loss": 0.495,
+      "step": 1184
+    },
+    {
+      "epoch": 0.07282448377581122,
+      "grad_norm": 72.0,
+      "learning_rate": 1.578666666666667e-05,
+      "loss": 0.8087,
+      "step": 1185
+    },
+    {
+      "epoch": 0.07288593903638152,
+      "grad_norm": 40.5,
+      "learning_rate": 1.58e-05,
+      "loss": 0.8484,
+      "step": 1186
+    },
+    {
+      "epoch": 0.07294739429695182,
+      "grad_norm": 30.75,
+      "learning_rate": 1.5813333333333333e-05,
+      "loss": 0.8589,
+      "step": 1187
+    },
+    {
+      "epoch": 0.07300884955752213,
+      "grad_norm": 39.75,
+      "learning_rate": 1.582666666666667e-05,
+      "loss": 0.7581,
+      "step": 1188
+    },
+    {
+      "epoch": 0.07307030481809243,
+      "grad_norm": 23.375,
+      "learning_rate": 1.584e-05,
+      "loss": 0.729,
+      "step": 1189
+    },
+    {
+      "epoch": 0.07313176007866273,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.5853333333333333e-05,
+      "loss": 0.6175,
+      "step": 1190
+    },
+    {
+      "epoch": 0.07319321533923304,
+      "grad_norm": 37.5,
+      "learning_rate": 1.586666666666667e-05,
+      "loss": 0.5987,
+      "step": 1191
+    },
+    {
+      "epoch": 0.07325467059980334,
+      "grad_norm": 49.0,
+      "learning_rate": 1.588e-05,
+      "loss": 0.7068,
+      "step": 1192
+    },
+    {
+      "epoch": 0.07331612586037364,
+      "grad_norm": 42.5,
+      "learning_rate": 1.5893333333333333e-05,
+      "loss": 0.8008,
+      "step": 1193
+    },
+    {
+      "epoch": 0.07337758112094395,
+      "grad_norm": 50.25,
+      "learning_rate": 1.5906666666666668e-05,
+      "loss": 0.7017,
+      "step": 1194
+    },
+    {
+      "epoch": 0.07343903638151426,
+      "grad_norm": 28.5,
+      "learning_rate": 1.5920000000000003e-05,
+      "loss": 0.6605,
+      "step": 1195
+    },
+    {
+      "epoch": 0.07350049164208457,
+      "grad_norm": 21.875,
+      "learning_rate": 1.5933333333333336e-05,
+      "loss": 0.527,
+      "step": 1196
+    },
+    {
+      "epoch": 0.07356194690265487,
+      "grad_norm": 24.25,
+      "learning_rate": 1.5946666666666668e-05,
+      "loss": 0.5601,
+      "step": 1197
+    },
+    {
+      "epoch": 0.07362340216322517,
+      "grad_norm": 42.0,
+      "learning_rate": 1.5960000000000003e-05,
+      "loss": 0.5686,
+      "step": 1198
+    },
+    {
+      "epoch": 0.07368485742379548,
+      "grad_norm": 26.125,
+      "learning_rate": 1.5973333333333335e-05,
+      "loss": 0.5352,
+      "step": 1199
+    },
+    {
+      "epoch": 0.07374631268436578,
+      "grad_norm": 12.25,
+      "learning_rate": 1.5986666666666667e-05,
+      "loss": 0.4315,
+      "step": 1200
+    },
+    {
+      "epoch": 0.07380776794493608,
+      "grad_norm": 26.0,
+      "learning_rate": 1.6000000000000003e-05,
+      "loss": 0.7356,
+      "step": 1201
+    },
+    {
+      "epoch": 0.07386922320550639,
+      "grad_norm": 33.75,
+      "learning_rate": 1.6013333333333335e-05,
+      "loss": 0.6635,
+      "step": 1202
+    },
+    {
+      "epoch": 0.07393067846607669,
+      "grad_norm": 19.25,
+      "learning_rate": 1.6026666666666667e-05,
+      "loss": 0.6158,
+      "step": 1203
+    },
+    {
+      "epoch": 0.073992133726647,
+      "grad_norm": 28.0,
+      "learning_rate": 1.6040000000000002e-05,
+      "loss": 0.7113,
+      "step": 1204
+    },
+    {
+      "epoch": 0.07405358898721731,
+      "grad_norm": 26.75,
+      "learning_rate": 1.6053333333333334e-05,
+      "loss": 0.6945,
+      "step": 1205
+    },
+    {
+      "epoch": 0.07411504424778761,
+      "grad_norm": 17.75,
+      "learning_rate": 1.606666666666667e-05,
+      "loss": 0.5829,
+      "step": 1206
+    },
+    {
+      "epoch": 0.07417649950835792,
+      "grad_norm": 23.75,
+      "learning_rate": 1.6080000000000002e-05,
+      "loss": 0.6567,
+      "step": 1207
+    },
+    {
+      "epoch": 0.07423795476892822,
+      "grad_norm": 21.75,
+      "learning_rate": 1.6093333333333334e-05,
+      "loss": 0.5667,
+      "step": 1208
+    },
+    {
+      "epoch": 0.07429941002949852,
+      "grad_norm": 40.75,
+      "learning_rate": 1.610666666666667e-05,
+      "loss": 1.02,
+      "step": 1209
+    },
+    {
+      "epoch": 0.07436086529006883,
+      "grad_norm": 32.25,
+      "learning_rate": 1.612e-05,
+      "loss": 0.8284,
+      "step": 1210
+    },
+    {
+      "epoch": 0.07442232055063913,
+      "grad_norm": 16.375,
+      "learning_rate": 1.6133333333333334e-05,
+      "loss": 0.6401,
+      "step": 1211
+    },
+    {
+      "epoch": 0.07448377581120944,
+      "grad_norm": 21.125,
+      "learning_rate": 1.614666666666667e-05,
+      "loss": 0.6497,
+      "step": 1212
+    },
+    {
+      "epoch": 0.07454523107177974,
+      "grad_norm": 18.0,
+      "learning_rate": 1.616e-05,
+      "loss": 0.6779,
+      "step": 1213
+    },
+    {
+      "epoch": 0.07460668633235006,
+      "grad_norm": 38.0,
+      "learning_rate": 1.6173333333333333e-05,
+      "loss": 0.944,
+      "step": 1214
+    },
+    {
+      "epoch": 0.07466814159292036,
+      "grad_norm": 19.0,
+      "learning_rate": 1.618666666666667e-05,
+      "loss": 0.7312,
+      "step": 1215
+    },
+    {
+      "epoch": 0.07472959685349066,
+      "grad_norm": 13.25,
+      "learning_rate": 1.62e-05,
+      "loss": 0.4905,
+      "step": 1216
+    },
+    {
+      "epoch": 0.07479105211406097,
+      "grad_norm": 41.5,
+      "learning_rate": 1.6213333333333333e-05,
+      "loss": 0.8054,
+      "step": 1217
+    },
+    {
+      "epoch": 0.07485250737463127,
+      "grad_norm": 29.625,
+      "learning_rate": 1.6226666666666668e-05,
+      "loss": 0.6629,
+      "step": 1218
+    },
+    {
+      "epoch": 0.07491396263520157,
+      "grad_norm": 16.375,
+      "learning_rate": 1.6240000000000004e-05,
+      "loss": 0.661,
+      "step": 1219
+    },
+    {
+      "epoch": 0.07497541789577188,
+      "grad_norm": 33.25,
+      "learning_rate": 1.6253333333333336e-05,
+      "loss": 0.5849,
+      "step": 1220
+    },
+    {
+      "epoch": 0.07503687315634218,
+      "grad_norm": 54.75,
+      "learning_rate": 1.6266666666666668e-05,
+      "loss": 0.7616,
+      "step": 1221
+    },
+    {
+      "epoch": 0.07509832841691248,
+      "grad_norm": 38.25,
+      "learning_rate": 1.628e-05,
+      "loss": 0.6739,
+      "step": 1222
+    },
+    {
+      "epoch": 0.07515978367748279,
+      "grad_norm": 39.5,
+      "learning_rate": 1.6293333333333335e-05,
+      "loss": 0.7668,
+      "step": 1223
+    },
+    {
+      "epoch": 0.0752212389380531,
+      "grad_norm": 31.125,
+      "learning_rate": 1.6306666666666668e-05,
+      "loss": 0.4427,
+      "step": 1224
+    },
+    {
+      "epoch": 0.0752826941986234,
+      "grad_norm": 25.125,
+      "learning_rate": 1.632e-05,
+      "loss": 0.5461,
+      "step": 1225
+    },
+    {
+      "epoch": 0.07534414945919371,
+      "grad_norm": 29.25,
+      "learning_rate": 1.6333333333333335e-05,
+      "loss": 0.7374,
+      "step": 1226
+    },
+    {
+      "epoch": 0.07540560471976401,
+      "grad_norm": 17.75,
+      "learning_rate": 1.6346666666666667e-05,
+      "loss": 0.5886,
+      "step": 1227
+    },
+    {
+      "epoch": 0.07546705998033432,
+      "grad_norm": 23.25,
+      "learning_rate": 1.636e-05,
+      "loss": 0.54,
+      "step": 1228
+    },
+    {
+      "epoch": 0.07552851524090462,
+      "grad_norm": 18.625,
+      "learning_rate": 1.6373333333333335e-05,
+      "loss": 0.5391,
+      "step": 1229
+    },
+    {
+      "epoch": 0.07558997050147492,
+      "grad_norm": 18.5,
+      "learning_rate": 1.638666666666667e-05,
+      "loss": 0.569,
+      "step": 1230
+    },
+    {
+      "epoch": 0.07565142576204523,
+      "grad_norm": 25.375,
+      "learning_rate": 1.64e-05,
+      "loss": 0.5415,
+      "step": 1231
+    },
+    {
+      "epoch": 0.07571288102261553,
+      "grad_norm": 21.125,
+      "learning_rate": 1.6413333333333334e-05,
+      "loss": 0.5971,
+      "step": 1232
+    },
+    {
+      "epoch": 0.07577433628318585,
+      "grad_norm": 33.0,
+      "learning_rate": 1.642666666666667e-05,
+      "loss": 0.7882,
+      "step": 1233
+    },
+    {
+      "epoch": 0.07583579154375615,
+      "grad_norm": 35.25,
+      "learning_rate": 1.6440000000000002e-05,
+      "loss": 1.0296,
+      "step": 1234
+    },
+    {
+      "epoch": 0.07589724680432645,
+      "grad_norm": 29.0,
+      "learning_rate": 1.6453333333333334e-05,
+      "loss": 1.0611,
+      "step": 1235
+    },
+    {
+      "epoch": 0.07595870206489676,
+      "grad_norm": 35.0,
+      "learning_rate": 1.646666666666667e-05,
+      "loss": 0.8196,
+      "step": 1236
+    },
+    {
+      "epoch": 0.07602015732546706,
+      "grad_norm": 57.25,
+      "learning_rate": 1.648e-05,
+      "loss": 0.9906,
+      "step": 1237
+    },
+    {
+      "epoch": 0.07608161258603736,
+      "grad_norm": 17.125,
+      "learning_rate": 1.6493333333333334e-05,
+      "loss": 0.72,
+      "step": 1238
+    },
+    {
+      "epoch": 0.07614306784660767,
+      "grad_norm": 19.5,
+      "learning_rate": 1.650666666666667e-05,
+      "loss": 0.6491,
+      "step": 1239
+    },
+    {
+      "epoch": 0.07620452310717797,
+      "grad_norm": 18.875,
+      "learning_rate": 1.652e-05,
+      "loss": 0.6304,
+      "step": 1240
+    },
+    {
+      "epoch": 0.07626597836774827,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.6533333333333333e-05,
+      "loss": 0.7159,
+      "step": 1241
+    },
+    {
+      "epoch": 0.07632743362831858,
+      "grad_norm": 22.125,
+      "learning_rate": 1.654666666666667e-05,
+      "loss": 0.6239,
+      "step": 1242
+    },
+    {
+      "epoch": 0.0763888888888889,
+      "grad_norm": 24.375,
+      "learning_rate": 1.656e-05,
+      "loss": 0.7342,
+      "step": 1243
+    },
+    {
+      "epoch": 0.0764503441494592,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.6573333333333336e-05,
+      "loss": 0.5269,
+      "step": 1244
+    },
+    {
+      "epoch": 0.0765117994100295,
+      "grad_norm": 18.25,
+      "learning_rate": 1.6586666666666668e-05,
+      "loss": 0.4487,
+      "step": 1245
+    },
+    {
+      "epoch": 0.0765732546705998,
+      "grad_norm": 10.5,
+      "learning_rate": 1.66e-05,
+      "loss": 0.5805,
+      "step": 1246
+    },
+    {
+      "epoch": 0.07663470993117011,
+      "grad_norm": 27.625,
+      "learning_rate": 1.6613333333333336e-05,
+      "loss": 0.551,
+      "step": 1247
+    },
+    {
+      "epoch": 0.07669616519174041,
+      "grad_norm": 7.28125,
+      "learning_rate": 1.6626666666666668e-05,
+      "loss": 0.3734,
+      "step": 1248
+    },
+    {
+      "epoch": 0.07675762045231072,
+      "grad_norm": 27.0,
+      "learning_rate": 1.664e-05,
+      "loss": 0.5799,
+      "step": 1249
+    },
+    {
+      "epoch": 0.07681907571288102,
+      "grad_norm": 25.0,
+      "learning_rate": 1.6653333333333335e-05,
+      "loss": 0.4928,
+      "step": 1250
+    },
+    {
+      "epoch": 0.07688053097345132,
+      "grad_norm": 39.0,
+      "learning_rate": 1.6666666666666667e-05,
+      "loss": 0.5519,
+      "step": 1251
+    },
+    {
+      "epoch": 0.07694198623402163,
+      "grad_norm": 28.125,
+      "learning_rate": 1.668e-05,
+      "loss": 0.5392,
+      "step": 1252
+    },
+    {
+      "epoch": 0.07700344149459194,
+      "grad_norm": 20.375,
+      "learning_rate": 1.6693333333333335e-05,
+      "loss": 0.5647,
+      "step": 1253
+    },
+    {
+      "epoch": 0.07706489675516225,
+      "grad_norm": 40.25,
+      "learning_rate": 1.670666666666667e-05,
+      "loss": 0.6763,
+      "step": 1254
+    },
+    {
+      "epoch": 0.07712635201573255,
+      "grad_norm": 34.25,
+      "learning_rate": 1.672e-05,
+      "loss": 0.4502,
+      "step": 1255
+    },
+    {
+      "epoch": 0.07718780727630285,
+      "grad_norm": 8.75,
+      "learning_rate": 1.6733333333333335e-05,
+      "loss": 0.5004,
+      "step": 1256
+    },
+    {
+      "epoch": 0.07724926253687316,
+      "grad_norm": 94.5,
+      "learning_rate": 1.674666666666667e-05,
+      "loss": 0.9472,
+      "step": 1257
+    },
+    {
+      "epoch": 0.07731071779744346,
+      "grad_norm": 74.0,
+      "learning_rate": 1.6760000000000002e-05,
+      "loss": 0.8513,
+      "step": 1258
+    },
+    {
+      "epoch": 0.07737217305801376,
+      "grad_norm": 34.75,
+      "learning_rate": 1.6773333333333334e-05,
+      "loss": 0.8647,
+      "step": 1259
+    },
+    {
+      "epoch": 0.07743362831858407,
+      "grad_norm": 37.5,
+      "learning_rate": 1.678666666666667e-05,
+      "loss": 0.7916,
+      "step": 1260
+    },
+    {
+      "epoch": 0.07749508357915437,
+      "grad_norm": 35.0,
+      "learning_rate": 1.6800000000000002e-05,
+      "loss": 0.7195,
+      "step": 1261
+    },
+    {
+      "epoch": 0.07755653883972469,
+      "grad_norm": 45.25,
+      "learning_rate": 1.6813333333333334e-05,
+      "loss": 0.8831,
+      "step": 1262
+    },
+    {
+      "epoch": 0.07761799410029499,
+      "grad_norm": 35.25,
+      "learning_rate": 1.682666666666667e-05,
+      "loss": 0.5772,
+      "step": 1263
+    },
+    {
+      "epoch": 0.0776794493608653,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.684e-05,
+      "loss": 0.5639,
+      "step": 1264
+    },
+    {
+      "epoch": 0.0777409046214356,
+      "grad_norm": 39.0,
+      "learning_rate": 1.6853333333333333e-05,
+      "loss": 0.6101,
+      "step": 1265
+    },
+    {
+      "epoch": 0.0778023598820059,
+      "grad_norm": 20.75,
+      "learning_rate": 1.686666666666667e-05,
+      "loss": 0.8422,
+      "step": 1266
+    },
+    {
+      "epoch": 0.0778638151425762,
+      "grad_norm": 36.25,
+      "learning_rate": 1.688e-05,
+      "loss": 0.7107,
+      "step": 1267
+    },
+    {
+      "epoch": 0.07792527040314651,
+      "grad_norm": 32.0,
+      "learning_rate": 1.6893333333333336e-05,
+      "loss": 0.6155,
+      "step": 1268
+    },
+    {
+      "epoch": 0.07798672566371681,
+      "grad_norm": 36.75,
+      "learning_rate": 1.690666666666667e-05,
+      "loss": 0.6383,
+      "step": 1269
+    },
+    {
+      "epoch": 0.07804818092428711,
+      "grad_norm": 29.125,
+      "learning_rate": 1.692e-05,
+      "loss": 0.6434,
+      "step": 1270
+    },
+    {
+      "epoch": 0.07810963618485742,
+      "grad_norm": 27.125,
+      "learning_rate": 1.6933333333333336e-05,
+      "loss": 0.4912,
+      "step": 1271
+    },
+    {
+      "epoch": 0.07817109144542773,
+      "grad_norm": 37.0,
+      "learning_rate": 1.6946666666666668e-05,
+      "loss": 0.5515,
+      "step": 1272
+    },
+    {
+      "epoch": 0.07823254670599804,
+      "grad_norm": 42.0,
+      "learning_rate": 1.696e-05,
+      "loss": 0.768,
+      "step": 1273
+    },
+    {
+      "epoch": 0.07829400196656834,
+      "grad_norm": 22.625,
+      "learning_rate": 1.6973333333333336e-05,
+      "loss": 0.8686,
+      "step": 1274
+    },
+    {
+      "epoch": 0.07835545722713864,
+      "grad_norm": 31.75,
+      "learning_rate": 1.6986666666666668e-05,
+      "loss": 0.8938,
+      "step": 1275
+    },
+    {
+      "epoch": 0.07841691248770895,
+      "grad_norm": 26.875,
+      "learning_rate": 1.7e-05,
+      "loss": 0.9088,
+      "step": 1276
+    },
+    {
+      "epoch": 0.07847836774827925,
+      "grad_norm": 34.5,
+      "learning_rate": 1.7013333333333335e-05,
+      "loss": 0.6139,
+      "step": 1277
+    },
+    {
+      "epoch": 0.07853982300884955,
+      "grad_norm": 17.375,
+      "learning_rate": 1.702666666666667e-05,
+      "loss": 0.4993,
+      "step": 1278
+    },
+    {
+      "epoch": 0.07860127826941986,
+      "grad_norm": 15.625,
+      "learning_rate": 1.704e-05,
+      "loss": 0.7102,
+      "step": 1279
+    },
+    {
+      "epoch": 0.07866273352999016,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.7053333333333335e-05,
+      "loss": 0.6416,
+      "step": 1280
+    },
+    {
+      "epoch": 0.07872418879056048,
+      "grad_norm": 25.625,
+      "learning_rate": 1.706666666666667e-05,
+      "loss": 0.707,
+      "step": 1281
+    },
+    {
+      "epoch": 0.07878564405113078,
+      "grad_norm": 29.0,
+      "learning_rate": 1.7080000000000002e-05,
+      "loss": 0.8199,
+      "step": 1282
+    },
+    {
+      "epoch": 0.07884709931170109,
+      "grad_norm": 22.5,
+      "learning_rate": 1.7093333333333335e-05,
+      "loss": 0.7795,
+      "step": 1283
+    },
+    {
+      "epoch": 0.07890855457227139,
+      "grad_norm": 40.25,
+      "learning_rate": 1.7106666666666667e-05,
+      "loss": 0.5938,
+      "step": 1284
+    },
+    {
+      "epoch": 0.07897000983284169,
+      "grad_norm": 18.375,
+      "learning_rate": 1.7120000000000002e-05,
+      "loss": 0.6192,
+      "step": 1285
+    },
+    {
+      "epoch": 0.079031465093412,
+      "grad_norm": 34.75,
+      "learning_rate": 1.7133333333333334e-05,
+      "loss": 0.7148,
+      "step": 1286
+    },
+    {
+      "epoch": 0.0790929203539823,
+      "grad_norm": 10.375,
+      "learning_rate": 1.7146666666666666e-05,
+      "loss": 0.6218,
+      "step": 1287
+    },
+    {
+      "epoch": 0.0791543756145526,
+      "grad_norm": 33.75,
+      "learning_rate": 1.7160000000000002e-05,
+      "loss": 0.5906,
+      "step": 1288
+    },
+    {
+      "epoch": 0.0792158308751229,
+      "grad_norm": 23.75,
+      "learning_rate": 1.7173333333333334e-05,
+      "loss": 0.803,
+      "step": 1289
+    },
+    {
+      "epoch": 0.07927728613569321,
+      "grad_norm": 27.75,
+      "learning_rate": 1.7186666666666666e-05,
+      "loss": 0.5048,
+      "step": 1290
+    },
+    {
+      "epoch": 0.07933874139626353,
+      "grad_norm": 28.875,
+      "learning_rate": 1.72e-05,
+      "loss": 0.6709,
+      "step": 1291
+    },
+    {
+      "epoch": 0.07940019665683383,
+      "grad_norm": 24.5,
+      "learning_rate": 1.7213333333333337e-05,
+      "loss": 0.6672,
+      "step": 1292
+    },
+    {
+      "epoch": 0.07946165191740413,
+      "grad_norm": 23.5,
+      "learning_rate": 1.7226666666666665e-05,
+      "loss": 0.4858,
+      "step": 1293
+    },
+    {
+      "epoch": 0.07952310717797444,
+      "grad_norm": 18.875,
+      "learning_rate": 1.724e-05,
+      "loss": 0.5971,
+      "step": 1294
+    },
+    {
+      "epoch": 0.07958456243854474,
+      "grad_norm": 16.25,
+      "learning_rate": 1.7253333333333336e-05,
+      "loss": 0.5369,
+      "step": 1295
+    },
+    {
+      "epoch": 0.07964601769911504,
+      "grad_norm": 28.375,
+      "learning_rate": 1.726666666666667e-05,
+      "loss": 0.4849,
+      "step": 1296
+    },
+    {
+      "epoch": 0.07970747295968535,
+      "grad_norm": 29.25,
+      "learning_rate": 1.728e-05,
+      "loss": 0.5118,
+      "step": 1297
+    },
+    {
+      "epoch": 0.07976892822025565,
+      "grad_norm": 11.25,
+      "learning_rate": 1.7293333333333336e-05,
+      "loss": 0.4618,
+      "step": 1298
+    },
+    {
+      "epoch": 0.07983038348082595,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.7306666666666668e-05,
+      "loss": 0.5113,
+      "step": 1299
+    },
+    {
+      "epoch": 0.07989183874139626,
+      "grad_norm": 31.875,
+      "learning_rate": 1.732e-05,
+      "loss": 0.4626,
+      "step": 1300
+    },
+    {
+      "epoch": 0.07995329400196657,
+      "grad_norm": 24.5,
+      "learning_rate": 1.7333333333333336e-05,
+      "loss": 0.5453,
+      "step": 1301
+    },
+    {
+      "epoch": 0.08001474926253688,
+      "grad_norm": 26.375,
+      "learning_rate": 1.7346666666666668e-05,
+      "loss": 0.4727,
+      "step": 1302
+    },
+    {
+      "epoch": 0.08007620452310718,
+      "grad_norm": 17.25,
+      "learning_rate": 1.736e-05,
+      "loss": 0.5201,
+      "step": 1303
+    },
+    {
+      "epoch": 0.08013765978367748,
+      "grad_norm": 10.25,
+      "learning_rate": 1.7373333333333335e-05,
+      "loss": 0.3878,
+      "step": 1304
+    },
+    {
+      "epoch": 0.08019911504424779,
+      "grad_norm": 54.5,
+      "learning_rate": 1.7386666666666667e-05,
+      "loss": 0.6364,
+      "step": 1305
+    },
+    {
+      "epoch": 0.08026057030481809,
+      "grad_norm": 18.5,
+      "learning_rate": 1.7400000000000003e-05,
+      "loss": 0.6689,
+      "step": 1306
+    },
+    {
+      "epoch": 0.0803220255653884,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.7413333333333335e-05,
+      "loss": 0.5948,
+      "step": 1307
+    },
+    {
+      "epoch": 0.0803834808259587,
+      "grad_norm": 16.25,
+      "learning_rate": 1.7426666666666667e-05,
+      "loss": 0.7083,
+      "step": 1308
+    },
+    {
+      "epoch": 0.080444936086529,
+      "grad_norm": 36.5,
+      "learning_rate": 1.7440000000000002e-05,
+      "loss": 0.6281,
+      "step": 1309
+    },
+    {
+      "epoch": 0.08050639134709932,
+      "grad_norm": 26.25,
+      "learning_rate": 1.7453333333333335e-05,
+      "loss": 0.5585,
+      "step": 1310
+    },
+    {
+      "epoch": 0.08056784660766962,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.7466666666666667e-05,
+      "loss": 0.386,
+      "step": 1311
+    },
+    {
+      "epoch": 0.08062930186823992,
+      "grad_norm": 24.125,
+      "learning_rate": 1.7480000000000002e-05,
+      "loss": 0.3992,
+      "step": 1312
+    },
+    {
+      "epoch": 0.08069075712881023,
+      "grad_norm": 45.0,
+      "learning_rate": 1.7493333333333334e-05,
+      "loss": 0.9013,
+      "step": 1313
+    },
+    {
+      "epoch": 0.08075221238938053,
+      "grad_norm": 29.875,
+      "learning_rate": 1.7506666666666666e-05,
+      "loss": 1.0279,
+      "step": 1314
+    },
+    {
+      "epoch": 0.08081366764995083,
+      "grad_norm": 21.125,
+      "learning_rate": 1.752e-05,
+      "loss": 0.7779,
+      "step": 1315
+    },
+    {
+      "epoch": 0.08087512291052114,
+      "grad_norm": 16.375,
+      "learning_rate": 1.7533333333333337e-05,
+      "loss": 0.6813,
+      "step": 1316
+    },
+    {
+      "epoch": 0.08093657817109144,
+      "grad_norm": 44.25,
+      "learning_rate": 1.7546666666666666e-05,
+      "loss": 0.9382,
+      "step": 1317
+    },
+    {
+      "epoch": 0.08099803343166175,
+      "grad_norm": 36.0,
+      "learning_rate": 1.756e-05,
+      "loss": 0.9156,
+      "step": 1318
+    },
+    {
+      "epoch": 0.08105948869223205,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.7573333333333337e-05,
+      "loss": 0.746,
+      "step": 1319
+    },
+    {
+      "epoch": 0.08112094395280237,
+      "grad_norm": 12.5,
+      "learning_rate": 1.758666666666667e-05,
+      "loss": 0.5663,
+      "step": 1320
+    },
+    {
+      "epoch": 0.08118239921337267,
+      "grad_norm": 18.625,
+      "learning_rate": 1.76e-05,
+      "loss": 0.6666,
+      "step": 1321
+    },
+    {
+      "epoch": 0.08124385447394297,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.7613333333333336e-05,
+      "loss": 0.5154,
+      "step": 1322
+    },
+    {
+      "epoch": 0.08130530973451328,
+      "grad_norm": 32.75,
+      "learning_rate": 1.762666666666667e-05,
+      "loss": 0.6638,
+      "step": 1323
+    },
+    {
+      "epoch": 0.08136676499508358,
+      "grad_norm": 15.625,
+      "learning_rate": 1.764e-05,
+      "loss": 0.5732,
+      "step": 1324
+    },
+    {
+      "epoch": 0.08142822025565388,
+      "grad_norm": 20.5,
+      "learning_rate": 1.7653333333333336e-05,
+      "loss": 0.5806,
+      "step": 1325
+    },
+    {
+      "epoch": 0.08148967551622419,
+      "grad_norm": 32.75,
+      "learning_rate": 1.7666666666666668e-05,
+      "loss": 0.6079,
+      "step": 1326
+    },
+    {
+      "epoch": 0.08155113077679449,
+      "grad_norm": 43.0,
+      "learning_rate": 1.768e-05,
+      "loss": 0.6406,
+      "step": 1327
+    },
+    {
+      "epoch": 0.08161258603736479,
+      "grad_norm": 25.5,
+      "learning_rate": 1.7693333333333336e-05,
+      "loss": 0.6069,
+      "step": 1328
+    },
+    {
+      "epoch": 0.08167404129793511,
+      "grad_norm": 26.25,
+      "learning_rate": 1.7706666666666668e-05,
+      "loss": 0.7652,
+      "step": 1329
+    },
+    {
+      "epoch": 0.08173549655850541,
+      "grad_norm": 21.125,
+      "learning_rate": 1.7720000000000003e-05,
+      "loss": 0.7136,
+      "step": 1330
+    },
+    {
+      "epoch": 0.08179695181907572,
+      "grad_norm": 46.0,
+      "learning_rate": 1.7733333333333335e-05,
+      "loss": 0.7269,
+      "step": 1331
+    },
+    {
+      "epoch": 0.08185840707964602,
+      "grad_norm": 58.5,
+      "learning_rate": 1.7746666666666667e-05,
+      "loss": 0.8524,
+      "step": 1332
+    },
+    {
+      "epoch": 0.08191986234021632,
+      "grad_norm": 19.375,
+      "learning_rate": 1.7760000000000003e-05,
+      "loss": 0.8192,
+      "step": 1333
+    },
+    {
+      "epoch": 0.08198131760078663,
+      "grad_norm": 35.25,
+      "learning_rate": 1.7773333333333335e-05,
+      "loss": 0.7183,
+      "step": 1334
+    },
+    {
+      "epoch": 0.08204277286135693,
+      "grad_norm": 24.5,
+      "learning_rate": 1.7786666666666667e-05,
+      "loss": 0.6622,
+      "step": 1335
+    },
+    {
+      "epoch": 0.08210422812192723,
+      "grad_norm": 27.375,
+      "learning_rate": 1.7800000000000002e-05,
+      "loss": 0.7104,
+      "step": 1336
+    },
+    {
+      "epoch": 0.08216568338249754,
+      "grad_norm": 39.0,
+      "learning_rate": 1.7813333333333334e-05,
+      "loss": 0.5864,
+      "step": 1337
+    },
+    {
+      "epoch": 0.08222713864306784,
+      "grad_norm": 21.875,
+      "learning_rate": 1.7826666666666667e-05,
+      "loss": 0.7185,
+      "step": 1338
+    },
+    {
+      "epoch": 0.08228859390363816,
+      "grad_norm": 22.5,
+      "learning_rate": 1.7840000000000002e-05,
+      "loss": 0.4773,
+      "step": 1339
+    },
+    {
+      "epoch": 0.08235004916420846,
+      "grad_norm": 31.375,
+      "learning_rate": 1.7853333333333337e-05,
+      "loss": 0.6331,
+      "step": 1340
+    },
+    {
+      "epoch": 0.08241150442477876,
+      "grad_norm": 25.75,
+      "learning_rate": 1.7866666666666666e-05,
+      "loss": 0.5024,
+      "step": 1341
+    },
+    {
+      "epoch": 0.08247295968534907,
+      "grad_norm": 24.125,
+      "learning_rate": 1.788e-05,
+      "loss": 0.5157,
+      "step": 1342
+    },
+    {
+      "epoch": 0.08253441494591937,
+      "grad_norm": 24.25,
+      "learning_rate": 1.7893333333333337e-05,
+      "loss": 0.6028,
+      "step": 1343
+    },
+    {
+      "epoch": 0.08259587020648967,
+      "grad_norm": 21.875,
+      "learning_rate": 1.790666666666667e-05,
+      "loss": 0.3299,
+      "step": 1344
+    },
+    {
+      "epoch": 0.08265732546705998,
+      "grad_norm": 25.5,
+      "learning_rate": 1.792e-05,
+      "loss": 0.4917,
+      "step": 1345
+    },
+    {
+      "epoch": 0.08271878072763028,
+      "grad_norm": 19.375,
+      "learning_rate": 1.7933333333333333e-05,
+      "loss": 0.6454,
+      "step": 1346
+    },
+    {
+      "epoch": 0.08278023598820058,
+      "grad_norm": 15.5,
+      "learning_rate": 1.794666666666667e-05,
+      "loss": 0.6313,
+      "step": 1347
+    },
+    {
+      "epoch": 0.08284169124877089,
+      "grad_norm": 27.25,
+      "learning_rate": 1.796e-05,
+      "loss": 0.7153,
+      "step": 1348
+    },
+    {
+      "epoch": 0.0829031465093412,
+      "grad_norm": 30.625,
+      "learning_rate": 1.7973333333333333e-05,
+      "loss": 0.5216,
+      "step": 1349
+    },
+    {
+      "epoch": 0.08296460176991151,
+      "grad_norm": 33.75,
+      "learning_rate": 1.798666666666667e-05,
+      "loss": 0.497,
+      "step": 1350
+    },
+    {
+      "epoch": 0.08302605703048181,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.8e-05,
+      "loss": 0.4383,
+      "step": 1351
+    },
+    {
+      "epoch": 0.08308751229105212,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.8013333333333333e-05,
+      "loss": 0.5304,
+      "step": 1352
+    },
+    {
+      "epoch": 0.08314896755162242,
+      "grad_norm": 30.125,
+      "learning_rate": 1.8026666666666668e-05,
+      "loss": 0.542,
+      "step": 1353
+    },
+    {
+      "epoch": 0.08321042281219272,
+      "grad_norm": 24.5,
+      "learning_rate": 1.8040000000000003e-05,
+      "loss": 0.6674,
+      "step": 1354
+    },
+    {
+      "epoch": 0.08327187807276303,
+      "grad_norm": 10.0,
+      "learning_rate": 1.8053333333333332e-05,
+      "loss": 0.6858,
+      "step": 1355
+    },
+    {
+      "epoch": 0.08333333333333333,
+      "grad_norm": 16.125,
+      "learning_rate": 1.8066666666666668e-05,
+      "loss": 0.6673,
+      "step": 1356
+    },
+    {
+      "epoch": 0.08339478859390363,
+      "grad_norm": 9.75,
+      "learning_rate": 1.8080000000000003e-05,
+      "loss": 0.4523,
+      "step": 1357
+    },
+    {
+      "epoch": 0.08345624385447395,
+      "grad_norm": 52.25,
+      "learning_rate": 1.8093333333333335e-05,
+      "loss": 0.7677,
+      "step": 1358
+    },
+    {
+      "epoch": 0.08351769911504425,
+      "grad_norm": 33.0,
+      "learning_rate": 1.8106666666666667e-05,
+      "loss": 0.4808,
+      "step": 1359
+    },
+    {
+      "epoch": 0.08357915437561456,
+      "grad_norm": 51.75,
+      "learning_rate": 1.8120000000000003e-05,
+      "loss": 0.8363,
+      "step": 1360
+    },
+    {
+      "epoch": 0.08364060963618486,
+      "grad_norm": 33.0,
+      "learning_rate": 1.8133333333333335e-05,
+      "loss": 0.636,
+      "step": 1361
+    },
+    {
+      "epoch": 0.08370206489675516,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.8146666666666667e-05,
+      "loss": 0.6303,
+      "step": 1362
+    },
+    {
+      "epoch": 0.08376352015732547,
+      "grad_norm": 21.25,
+      "learning_rate": 1.8160000000000002e-05,
+      "loss": 0.6055,
+      "step": 1363
+    },
+    {
+      "epoch": 0.08382497541789577,
+      "grad_norm": 43.0,
+      "learning_rate": 1.8173333333333334e-05,
+      "loss": 0.6292,
+      "step": 1364
+    },
+    {
+      "epoch": 0.08388643067846607,
+      "grad_norm": 29.375,
+      "learning_rate": 1.8186666666666666e-05,
+      "loss": 0.6603,
+      "step": 1365
+    },
+    {
+      "epoch": 0.08394788593903638,
+      "grad_norm": 29.75,
+      "learning_rate": 1.8200000000000002e-05,
+      "loss": 0.5737,
+      "step": 1366
+    },
+    {
+      "epoch": 0.08400934119960668,
+      "grad_norm": 50.75,
+      "learning_rate": 1.8213333333333334e-05,
+      "loss": 0.7035,
+      "step": 1367
+    },
+    {
+      "epoch": 0.084070796460177,
+      "grad_norm": 76.5,
+      "learning_rate": 1.822666666666667e-05,
+      "loss": 0.5051,
+      "step": 1368
+    },
+    {
+      "epoch": 0.0841322517207473,
+      "grad_norm": 30.375,
+      "learning_rate": 1.824e-05,
+      "loss": 0.5556,
+      "step": 1369
+    },
+    {
+      "epoch": 0.0841937069813176,
+      "grad_norm": 6.84375,
+      "learning_rate": 1.8253333333333334e-05,
+      "loss": 0.4911,
+      "step": 1370
+    },
+    {
+      "epoch": 0.08425516224188791,
+      "grad_norm": 20.375,
+      "learning_rate": 1.826666666666667e-05,
+      "loss": 0.6677,
+      "step": 1371
+    },
+    {
+      "epoch": 0.08431661750245821,
+      "grad_norm": 31.5,
+      "learning_rate": 1.828e-05,
+      "loss": 0.5995,
+      "step": 1372
+    },
+    {
+      "epoch": 0.08437807276302851,
+      "grad_norm": 28.625,
+      "learning_rate": 1.8293333333333333e-05,
+      "loss": 0.5916,
+      "step": 1373
+    },
+    {
+      "epoch": 0.08443952802359882,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.830666666666667e-05,
+      "loss": 0.3366,
+      "step": 1374
+    },
+    {
+      "epoch": 0.08450098328416912,
+      "grad_norm": 20.0,
+      "learning_rate": 1.832e-05,
+      "loss": 0.6325,
+      "step": 1375
+    },
+    {
+      "epoch": 0.08456243854473942,
+      "grad_norm": 54.75,
+      "learning_rate": 1.8333333333333333e-05,
+      "loss": 0.4333,
+      "step": 1376
+    },
+    {
+      "epoch": 0.08462389380530974,
+      "grad_norm": 37.5,
+      "learning_rate": 1.834666666666667e-05,
+      "loss": 0.8487,
+      "step": 1377
+    },
+    {
+      "epoch": 0.08468534906588004,
+      "grad_norm": 39.5,
+      "learning_rate": 1.8360000000000004e-05,
+      "loss": 1.3229,
+      "step": 1378
+    },
+    {
+      "epoch": 0.08474680432645035,
+      "grad_norm": 26.25,
+      "learning_rate": 1.8373333333333332e-05,
+      "loss": 0.9511,
+      "step": 1379
+    },
+    {
+      "epoch": 0.08480825958702065,
+      "grad_norm": 27.125,
+      "learning_rate": 1.8386666666666668e-05,
+      "loss": 0.9488,
+      "step": 1380
+    },
+    {
+      "epoch": 0.08486971484759095,
+      "grad_norm": 20.25,
+      "learning_rate": 1.8400000000000003e-05,
+      "loss": 0.89,
+      "step": 1381
+    },
+    {
+      "epoch": 0.08493117010816126,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.8413333333333335e-05,
+      "loss": 0.7851,
+      "step": 1382
+    },
+    {
+      "epoch": 0.08499262536873156,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.8426666666666668e-05,
+      "loss": 0.5351,
+      "step": 1383
+    },
+    {
+      "epoch": 0.08505408062930186,
+      "grad_norm": 16.0,
+      "learning_rate": 1.8440000000000003e-05,
+      "loss": 0.7284,
+      "step": 1384
+    },
+    {
+      "epoch": 0.08511553588987217,
+      "grad_norm": 20.875,
+      "learning_rate": 1.8453333333333335e-05,
+      "loss": 0.6495,
+      "step": 1385
+    },
+    {
+      "epoch": 0.08517699115044247,
+      "grad_norm": 16.25,
+      "learning_rate": 1.8466666666666667e-05,
+      "loss": 0.7219,
+      "step": 1386
+    },
+    {
+      "epoch": 0.08523844641101279,
+      "grad_norm": 16.375,
+      "learning_rate": 1.8480000000000003e-05,
+      "loss": 0.795,
+      "step": 1387
+    },
+    {
+      "epoch": 0.08529990167158309,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.8493333333333335e-05,
+      "loss": 0.5516,
+      "step": 1388
+    },
+    {
+      "epoch": 0.0853613569321534,
+      "grad_norm": 57.25,
+      "learning_rate": 1.8506666666666667e-05,
+      "loss": 0.6733,
+      "step": 1389
+    },
+    {
+      "epoch": 0.0854228121927237,
+      "grad_norm": 11.8125,
+      "learning_rate": 1.8520000000000002e-05,
+      "loss": 0.5545,
+      "step": 1390
+    },
+    {
+      "epoch": 0.085484267453294,
+      "grad_norm": 24.875,
+      "learning_rate": 1.8533333333333334e-05,
+      "loss": 0.5134,
+      "step": 1391
+    },
+    {
+      "epoch": 0.0855457227138643,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.854666666666667e-05,
+      "loss": 0.4192,
+      "step": 1392
+    },
+    {
+      "epoch": 0.08560717797443461,
+      "grad_norm": 49.75,
+      "learning_rate": 1.8560000000000002e-05,
+      "loss": 0.6455,
+      "step": 1393
+    },
+    {
+      "epoch": 0.08566863323500491,
+      "grad_norm": 49.75,
+      "learning_rate": 1.8573333333333334e-05,
+      "loss": 0.5704,
+      "step": 1394
+    },
+    {
+      "epoch": 0.08573008849557522,
+      "grad_norm": 24.875,
+      "learning_rate": 1.858666666666667e-05,
+      "loss": 0.6552,
+      "step": 1395
+    },
+    {
+      "epoch": 0.08579154375614553,
+      "grad_norm": 28.125,
+      "learning_rate": 1.86e-05,
+      "loss": 0.5102,
+      "step": 1396
+    },
+    {
+      "epoch": 0.08585299901671584,
+      "grad_norm": 34.75,
+      "learning_rate": 1.8613333333333334e-05,
+      "loss": 0.8148,
+      "step": 1397
+    },
+    {
+      "epoch": 0.08591445427728614,
+      "grad_norm": 23.5,
+      "learning_rate": 1.862666666666667e-05,
+      "loss": 0.7509,
+      "step": 1398
+    },
+    {
+      "epoch": 0.08597590953785644,
+      "grad_norm": 36.0,
+      "learning_rate": 1.864e-05,
+      "loss": 0.5747,
+      "step": 1399
+    },
+    {
+      "epoch": 0.08603736479842675,
+      "grad_norm": 8.625,
+      "learning_rate": 1.8653333333333333e-05,
+      "loss": 0.4487,
+      "step": 1400
+    },
+    {
+      "epoch": 0.08609882005899705,
+      "grad_norm": 25.375,
+      "learning_rate": 1.866666666666667e-05,
+      "loss": 0.7071,
+      "step": 1401
+    },
+    {
+      "epoch": 0.08616027531956735,
+      "grad_norm": 83.0,
+      "learning_rate": 1.8680000000000004e-05,
+      "loss": 0.825,
+      "step": 1402
+    },
+    {
+      "epoch": 0.08622173058013766,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.8693333333333333e-05,
+      "loss": 0.5407,
+      "step": 1403
+    },
+    {
+      "epoch": 0.08628318584070796,
+      "grad_norm": 8.125,
+      "learning_rate": 1.8706666666666668e-05,
+      "loss": 0.5043,
+      "step": 1404
+    },
+    {
+      "epoch": 0.08634464110127826,
+      "grad_norm": 16.625,
+      "learning_rate": 1.8720000000000004e-05,
+      "loss": 0.5296,
+      "step": 1405
+    },
+    {
+      "epoch": 0.08640609636184858,
+      "grad_norm": 19.75,
+      "learning_rate": 1.8733333333333336e-05,
+      "loss": 0.4216,
+      "step": 1406
+    },
+    {
+      "epoch": 0.08646755162241888,
+      "grad_norm": 5.875,
+      "learning_rate": 1.8746666666666668e-05,
+      "loss": 0.3793,
+      "step": 1407
+    },
+    {
+      "epoch": 0.08652900688298919,
+      "grad_norm": 24.625,
+      "learning_rate": 1.876e-05,
+      "loss": 0.5236,
+      "step": 1408
+    },
+    {
+      "epoch": 0.08659046214355949,
+      "grad_norm": 19.25,
+      "learning_rate": 1.8773333333333335e-05,
+      "loss": 0.5596,
+      "step": 1409
+    },
+    {
+      "epoch": 0.0866519174041298,
+      "grad_norm": 16.25,
+      "learning_rate": 1.8786666666666667e-05,
+      "loss": 0.4285,
+      "step": 1410
+    },
+    {
+      "epoch": 0.0867133726647001,
+      "grad_norm": 19.125,
+      "learning_rate": 1.88e-05,
+      "loss": 0.6017,
+      "step": 1411
+    },
+    {
+      "epoch": 0.0867748279252704,
+      "grad_norm": 15.125,
+      "learning_rate": 1.8813333333333335e-05,
+      "loss": 0.5549,
+      "step": 1412
+    },
+    {
+      "epoch": 0.0868362831858407,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.8826666666666667e-05,
+      "loss": 0.5747,
+      "step": 1413
+    },
+    {
+      "epoch": 0.08689773844641101,
+      "grad_norm": 27.875,
+      "learning_rate": 1.884e-05,
+      "loss": 0.503,
+      "step": 1414
+    },
+    {
+      "epoch": 0.08695919370698131,
+      "grad_norm": 34.5,
+      "learning_rate": 1.8853333333333335e-05,
+      "loss": 0.4358,
+      "step": 1415
+    },
+    {
+      "epoch": 0.08702064896755163,
+      "grad_norm": 40.75,
+      "learning_rate": 1.886666666666667e-05,
+      "loss": 0.5821,
+      "step": 1416
+    },
+    {
+      "epoch": 0.08708210422812193,
+      "grad_norm": 21.875,
+      "learning_rate": 1.8880000000000002e-05,
+      "loss": 0.3984,
+      "step": 1417
+    },
+    {
+      "epoch": 0.08714355948869223,
+      "grad_norm": 26.125,
+      "learning_rate": 1.8893333333333334e-05,
+      "loss": 0.4793,
+      "step": 1418
+    },
+    {
+      "epoch": 0.08720501474926254,
+      "grad_norm": 26.0,
+      "learning_rate": 1.890666666666667e-05,
+      "loss": 0.4882,
+      "step": 1419
+    },
+    {
+      "epoch": 0.08726647000983284,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.8920000000000002e-05,
+      "loss": 0.7466,
+      "step": 1420
+    },
+    {
+      "epoch": 0.08732792527040315,
+      "grad_norm": 25.875,
+      "learning_rate": 1.8933333333333334e-05,
+      "loss": 0.4873,
+      "step": 1421
+    },
+    {
+      "epoch": 0.08738938053097345,
+      "grad_norm": 23.5,
+      "learning_rate": 1.894666666666667e-05,
+      "loss": 0.4732,
+      "step": 1422
+    },
+    {
+      "epoch": 0.08745083579154375,
+      "grad_norm": 49.5,
+      "learning_rate": 1.896e-05,
+      "loss": 0.5643,
+      "step": 1423
+    },
+    {
+      "epoch": 0.08751229105211406,
+      "grad_norm": 19.625,
+      "learning_rate": 1.8973333333333334e-05,
+      "loss": 0.4972,
+      "step": 1424
+    },
+    {
+      "epoch": 0.08757374631268437,
+      "grad_norm": 43.5,
+      "learning_rate": 1.898666666666667e-05,
+      "loss": 0.8297,
+      "step": 1425
+    },
+    {
+      "epoch": 0.08763520157325468,
+      "grad_norm": 44.0,
+      "learning_rate": 1.9e-05,
+      "loss": 0.8204,
+      "step": 1426
+    },
+    {
+      "epoch": 0.08769665683382498,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9013333333333333e-05,
+      "loss": 0.9316,
+      "step": 1427
+    },
+    {
+      "epoch": 0.08775811209439528,
+      "grad_norm": 18.125,
+      "learning_rate": 1.902666666666667e-05,
+      "loss": 0.6756,
+      "step": 1428
+    },
+    {
+      "epoch": 0.08781956735496559,
+      "grad_norm": 37.5,
+      "learning_rate": 1.904e-05,
+      "loss": 0.7311,
+      "step": 1429
+    },
+    {
+      "epoch": 0.08788102261553589,
+      "grad_norm": 42.75,
+      "learning_rate": 1.9053333333333336e-05,
+      "loss": 0.6094,
+      "step": 1430
+    },
+    {
+      "epoch": 0.08794247787610619,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9066666666666668e-05,
+      "loss": 0.5781,
+      "step": 1431
+    },
+    {
+      "epoch": 0.0880039331366765,
+      "grad_norm": 12.875,
+      "learning_rate": 1.908e-05,
+      "loss": 0.6201,
+      "step": 1432
+    },
+    {
+      "epoch": 0.0880653883972468,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9093333333333336e-05,
+      "loss": 0.7821,
+      "step": 1433
+    },
+    {
+      "epoch": 0.0881268436578171,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9106666666666668e-05,
+      "loss": 0.5977,
+      "step": 1434
+    },
+    {
+      "epoch": 0.08818829891838742,
+      "grad_norm": 36.75,
+      "learning_rate": 1.912e-05,
+      "loss": 0.6369,
+      "step": 1435
+    },
+    {
+      "epoch": 0.08824975417895772,
+      "grad_norm": 44.0,
+      "learning_rate": 1.9133333333333335e-05,
+      "loss": 0.7049,
+      "step": 1436
+    },
+    {
+      "epoch": 0.08831120943952803,
+      "grad_norm": 30.625,
+      "learning_rate": 1.9146666666666667e-05,
+      "loss": 0.5768,
+      "step": 1437
+    },
+    {
+      "epoch": 0.08837266470009833,
+      "grad_norm": 55.25,
+      "learning_rate": 1.916e-05,
+      "loss": 0.7788,
+      "step": 1438
+    },
+    {
+      "epoch": 0.08843411996066863,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9173333333333335e-05,
+      "loss": 0.668,
+      "step": 1439
+    },
+    {
+      "epoch": 0.08849557522123894,
+      "grad_norm": 28.25,
+      "learning_rate": 1.918666666666667e-05,
+      "loss": 0.5766,
+      "step": 1440
+    },
+    {
+      "epoch": 0.08855703048180924,
+      "grad_norm": 63.0,
+      "learning_rate": 1.9200000000000003e-05,
+      "loss": 0.6212,
+      "step": 1441
+    },
+    {
+      "epoch": 0.08861848574237954,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9213333333333335e-05,
+      "loss": 0.6886,
+      "step": 1442
+    },
+    {
+      "epoch": 0.08867994100294985,
+      "grad_norm": 50.75,
+      "learning_rate": 1.922666666666667e-05,
+      "loss": 0.7232,
+      "step": 1443
+    },
+    {
+      "epoch": 0.08874139626352016,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.9240000000000002e-05,
+      "loss": 0.4958,
+      "step": 1444
+    },
+    {
+      "epoch": 0.08880285152409047,
+      "grad_norm": 31.5,
+      "learning_rate": 1.9253333333333334e-05,
+      "loss": 0.4745,
+      "step": 1445
+    },
+    {
+      "epoch": 0.08886430678466077,
+      "grad_norm": 18.25,
+      "learning_rate": 1.926666666666667e-05,
+      "loss": 0.4712,
+      "step": 1446
+    },
+    {
+      "epoch": 0.08892576204523107,
+      "grad_norm": 14.375,
+      "learning_rate": 1.9280000000000002e-05,
+      "loss": 0.3748,
+      "step": 1447
+    },
+    {
+      "epoch": 0.08898721730580138,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.9293333333333334e-05,
+      "loss": 0.3747,
+      "step": 1448
+    },
+    {
+      "epoch": 0.08904867256637168,
+      "grad_norm": 41.5,
+      "learning_rate": 1.930666666666667e-05,
+      "loss": 0.7149,
+      "step": 1449
+    },
+    {
+      "epoch": 0.08911012782694198,
+      "grad_norm": 17.375,
+      "learning_rate": 1.932e-05,
+      "loss": 0.4438,
+      "step": 1450
+    },
+    {
+      "epoch": 0.08917158308751229,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.9333333333333333e-05,
+      "loss": 0.4511,
+      "step": 1451
+    },
+    {
+      "epoch": 0.08923303834808259,
+      "grad_norm": 54.0,
+      "learning_rate": 1.934666666666667e-05,
+      "loss": 0.4975,
+      "step": 1452
+    },
+    {
+      "epoch": 0.0892944936086529,
+      "grad_norm": 34.75,
+      "learning_rate": 1.936e-05,
+      "loss": 0.5387,
+      "step": 1453
+    },
+    {
+      "epoch": 0.08935594886922321,
+      "grad_norm": 55.5,
+      "learning_rate": 1.9373333333333336e-05,
+      "loss": 0.5473,
+      "step": 1454
+    },
+    {
+      "epoch": 0.08941740412979352,
+      "grad_norm": 36.0,
+      "learning_rate": 1.938666666666667e-05,
+      "loss": 0.5324,
+      "step": 1455
+    },
+    {
+      "epoch": 0.08947885939036382,
+      "grad_norm": 41.5,
+      "learning_rate": 1.94e-05,
+      "loss": 0.4644,
+      "step": 1456
+    },
+    {
+      "epoch": 0.08954031465093412,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9413333333333336e-05,
+      "loss": 0.9385,
+      "step": 1457
+    },
+    {
+      "epoch": 0.08960176991150443,
+      "grad_norm": 68.5,
+      "learning_rate": 1.9426666666666668e-05,
+      "loss": 0.5599,
+      "step": 1458
+    },
+    {
+      "epoch": 0.08966322517207473,
+      "grad_norm": 31.5,
+      "learning_rate": 1.944e-05,
+      "loss": 0.7873,
+      "step": 1459
+    },
+    {
+      "epoch": 0.08972468043264503,
+      "grad_norm": 40.25,
+      "learning_rate": 1.9453333333333336e-05,
+      "loss": 0.6027,
+      "step": 1460
+    },
+    {
+      "epoch": 0.08978613569321534,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9466666666666668e-05,
+      "loss": 0.7739,
+      "step": 1461
+    },
+    {
+      "epoch": 0.08984759095378564,
+      "grad_norm": 8.375,
+      "learning_rate": 1.948e-05,
+      "loss": 0.4431,
+      "step": 1462
+    },
+    {
+      "epoch": 0.08990904621435594,
+      "grad_norm": 30.875,
+      "learning_rate": 1.9493333333333335e-05,
+      "loss": 0.4555,
+      "step": 1463
+    },
+    {
+      "epoch": 0.08997050147492626,
+      "grad_norm": 10.25,
+      "learning_rate": 1.950666666666667e-05,
+      "loss": 0.3987,
+      "step": 1464
+    },
+    {
+      "epoch": 0.09003195673549656,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9520000000000003e-05,
+      "loss": 0.544,
+      "step": 1465
+    },
+    {
+      "epoch": 0.09009341199606687,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9533333333333335e-05,
+      "loss": 0.6123,
+      "step": 1466
+    },
+    {
+      "epoch": 0.09015486725663717,
+      "grad_norm": 25.125,
+      "learning_rate": 1.954666666666667e-05,
+      "loss": 0.6146,
+      "step": 1467
+    },
+    {
+      "epoch": 0.09021632251720747,
+      "grad_norm": 53.5,
+      "learning_rate": 1.9560000000000002e-05,
+      "loss": 0.5794,
+      "step": 1468
+    },
+    {
+      "epoch": 0.09027777777777778,
+      "grad_norm": 30.5,
+      "learning_rate": 1.9573333333333335e-05,
+      "loss": 0.5069,
+      "step": 1469
+    },
+    {
+      "epoch": 0.09033923303834808,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9586666666666667e-05,
+      "loss": 0.5018,
+      "step": 1470
+    },
+    {
+      "epoch": 0.09040068829891838,
+      "grad_norm": 16.5,
+      "learning_rate": 1.9600000000000002e-05,
+      "loss": 0.4914,
+      "step": 1471
+    },
+    {
+      "epoch": 0.09046214355948869,
+      "grad_norm": 73.5,
+      "learning_rate": 1.9613333333333334e-05,
+      "loss": 0.4178,
+      "step": 1472
+    },
+    {
+      "epoch": 0.090523598820059,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9626666666666666e-05,
+      "loss": 0.67,
+      "step": 1473
+    },
+    {
+      "epoch": 0.09058505408062931,
+      "grad_norm": 25.375,
+      "learning_rate": 1.9640000000000002e-05,
+      "loss": 0.6884,
+      "step": 1474
+    },
+    {
+      "epoch": 0.09064650934119961,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.9653333333333334e-05,
+      "loss": 0.3814,
+      "step": 1475
+    },
+    {
+      "epoch": 0.09070796460176991,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9666666666666666e-05,
+      "loss": 0.5632,
+      "step": 1476
+    },
+    {
+      "epoch": 0.09076941986234022,
+      "grad_norm": 22.625,
+      "learning_rate": 1.968e-05,
+      "loss": 0.6045,
+      "step": 1477
+    },
+    {
+      "epoch": 0.09083087512291052,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9693333333333337e-05,
+      "loss": 0.5875,
+      "step": 1478
+    },
+    {
+      "epoch": 0.09089233038348082,
+      "grad_norm": 10.375,
+      "learning_rate": 1.970666666666667e-05,
+      "loss": 0.4524,
+      "step": 1479
+    },
+    {
+      "epoch": 0.09095378564405113,
+      "grad_norm": 19.375,
+      "learning_rate": 1.972e-05,
+      "loss": 0.4269,
+      "step": 1480
+    },
+    {
+      "epoch": 0.09101524090462143,
+      "grad_norm": 20.125,
+      "learning_rate": 1.9733333333333336e-05,
+      "loss": 0.4363,
+      "step": 1481
+    },
+    {
+      "epoch": 0.09107669616519173,
+      "grad_norm": 26.625,
+      "learning_rate": 1.974666666666667e-05,
+      "loss": 0.6239,
+      "step": 1482
+    },
+    {
+      "epoch": 0.09113815142576205,
+      "grad_norm": 26.75,
+      "learning_rate": 1.976e-05,
+      "loss": 0.7371,
+      "step": 1483
+    },
+    {
+      "epoch": 0.09119960668633235,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.9773333333333336e-05,
+      "loss": 0.5799,
+      "step": 1484
+    },
+    {
+      "epoch": 0.09126106194690266,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9786666666666668e-05,
+      "loss": 0.3869,
+      "step": 1485
+    },
+    {
+      "epoch": 0.09132251720747296,
+      "grad_norm": 21.5,
+      "learning_rate": 1.98e-05,
+      "loss": 0.4029,
+      "step": 1486
+    },
+    {
+      "epoch": 0.09138397246804326,
+      "grad_norm": 7.625,
+      "learning_rate": 1.9813333333333336e-05,
+      "loss": 0.3257,
+      "step": 1487
+    },
+    {
+      "epoch": 0.09144542772861357,
+      "grad_norm": 31.25,
+      "learning_rate": 1.9826666666666668e-05,
+      "loss": 0.3925,
+      "step": 1488
+    },
+    {
+      "epoch": 0.09150688298918387,
+      "grad_norm": 25.375,
+      "learning_rate": 1.9840000000000003e-05,
+      "loss": 0.626,
+      "step": 1489
+    },
+    {
+      "epoch": 0.09156833824975417,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9853333333333335e-05,
+      "loss": 0.6065,
+      "step": 1490
+    },
+    {
+      "epoch": 0.09162979351032448,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9866666666666667e-05,
+      "loss": 0.4988,
+      "step": 1491
+    },
+    {
+      "epoch": 0.0916912487708948,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9880000000000003e-05,
+      "loss": 0.7633,
+      "step": 1492
+    },
+    {
+      "epoch": 0.0917527040314651,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.9893333333333335e-05,
+      "loss": 0.5236,
+      "step": 1493
+    },
+    {
+      "epoch": 0.0918141592920354,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9906666666666667e-05,
+      "loss": 0.744,
+      "step": 1494
+    },
+    {
+      "epoch": 0.0918756145526057,
+      "grad_norm": 60.25,
+      "learning_rate": 1.9920000000000002e-05,
+      "loss": 0.4128,
+      "step": 1495
+    },
+    {
+      "epoch": 0.09193706981317601,
+      "grad_norm": 11.0625,
+      "learning_rate": 1.9933333333333334e-05,
+      "loss": 0.4378,
+      "step": 1496
+    },
+    {
+      "epoch": 0.09199852507374631,
+      "grad_norm": 16.5,
+      "learning_rate": 1.9946666666666667e-05,
+      "loss": 0.5792,
+      "step": 1497
+    },
+    {
+      "epoch": 0.09205998033431662,
+      "grad_norm": 28.0,
+      "learning_rate": 1.9960000000000002e-05,
+      "loss": 0.7458,
+      "step": 1498
+    },
+    {
+      "epoch": 0.09212143559488692,
+      "grad_norm": 14.25,
+      "learning_rate": 1.9973333333333334e-05,
+      "loss": 0.8067,
+      "step": 1499
+    },
+    {
+      "epoch": 0.09218289085545722,
+      "grad_norm": 25.125,
+      "learning_rate": 1.9986666666666666e-05,
+      "loss": 0.6215,
+      "step": 1500
+    },
+    {
+      "epoch": 0.09224434611602753,
+      "grad_norm": 16.5,
+      "learning_rate": 2e-05,
+      "loss": 0.7529,
+      "step": 1501
+    },
+    {
+      "epoch": 0.09230580137659784,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.9999999729228962e-05,
+      "loss": 0.718,
+      "step": 1502
+    },
+    {
+      "epoch": 0.09236725663716815,
+      "grad_norm": 9.375,
+      "learning_rate": 1.9999998916915862e-05,
+      "loss": 0.5488,
+      "step": 1503
+    },
+    {
+      "epoch": 0.09242871189773845,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9999997563060744e-05,
+      "loss": 0.4644,
+      "step": 1504
+    },
+    {
+      "epoch": 0.09249016715830875,
+      "grad_norm": 14.75,
+      "learning_rate": 1.9999995667663676e-05,
+      "loss": 0.7382,
+      "step": 1505
+    },
+    {
+      "epoch": 0.09255162241887906,
+      "grad_norm": 22.125,
+      "learning_rate": 1.999999323072477e-05,
+      "loss": 0.6014,
+      "step": 1506
+    },
+    {
+      "epoch": 0.09261307767944936,
+      "grad_norm": 28.75,
+      "learning_rate": 1.9999990252244153e-05,
+      "loss": 0.6516,
+      "step": 1507
+    },
+    {
+      "epoch": 0.09267453294001966,
+      "grad_norm": 9.25,
+      "learning_rate": 1.999998673222198e-05,
+      "loss": 0.6359,
+      "step": 1508
+    },
+    {
+      "epoch": 0.09273598820058997,
+      "grad_norm": 49.0,
+      "learning_rate": 1.9999982670658457e-05,
+      "loss": 0.6436,
+      "step": 1509
+    },
+    {
+      "epoch": 0.09279744346116027,
+      "grad_norm": 11.625,
+      "learning_rate": 1.9999978067553796e-05,
+      "loss": 0.5704,
+      "step": 1510
+    },
+    {
+      "epoch": 0.09285889872173057,
+      "grad_norm": 25.0,
+      "learning_rate": 1.999997292290824e-05,
+      "loss": 0.6502,
+      "step": 1511
+    },
+    {
+      "epoch": 0.09292035398230089,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9999967236722075e-05,
+      "loss": 0.5587,
+      "step": 1512
+    },
+    {
+      "epoch": 0.0929818092428712,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9999961008995607e-05,
+      "loss": 0.4997,
+      "step": 1513
+    },
+    {
+      "epoch": 0.0930432645034415,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9999954239729175e-05,
+      "loss": 0.8159,
+      "step": 1514
+    },
+    {
+      "epoch": 0.0931047197640118,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.999994692892314e-05,
+      "loss": 0.5844,
+      "step": 1515
+    },
+    {
+      "epoch": 0.0931661750245821,
+      "grad_norm": 51.5,
+      "learning_rate": 1.9999939076577906e-05,
+      "loss": 0.5972,
+      "step": 1516
+    },
+    {
+      "epoch": 0.09322763028515241,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9999930682693892e-05,
+      "loss": 0.5903,
+      "step": 1517
+    },
+    {
+      "epoch": 0.09328908554572271,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9999921747271554e-05,
+      "loss": 0.6251,
+      "step": 1518
+    },
+    {
+      "epoch": 0.09335054080629301,
+      "grad_norm": 41.5,
+      "learning_rate": 1.9999912270311376e-05,
+      "loss": 0.4034,
+      "step": 1519
+    },
+    {
+      "epoch": 0.09341199606686332,
+      "grad_norm": 63.0,
+      "learning_rate": 1.9999902251813873e-05,
+      "loss": 0.4594,
+      "step": 1520
+    },
+    {
+      "epoch": 0.09347345132743363,
+      "grad_norm": 51.75,
+      "learning_rate": 1.999989169177959e-05,
+      "loss": 0.6752,
+      "step": 1521
+    },
+    {
+      "epoch": 0.09353490658800394,
+      "grad_norm": 34.0,
+      "learning_rate": 1.999988059020909e-05,
+      "loss": 0.8115,
+      "step": 1522
+    },
+    {
+      "epoch": 0.09359636184857424,
+      "grad_norm": 4.75,
+      "learning_rate": 1.9999868947102982e-05,
+      "loss": 0.6671,
+      "step": 1523
+    },
+    {
+      "epoch": 0.09365781710914454,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9999856762461892e-05,
+      "loss": 0.9567,
+      "step": 1524
+    },
+    {
+      "epoch": 0.09371927236971485,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9999844036286483e-05,
+      "loss": 0.6517,
+      "step": 1525
+    },
+    {
+      "epoch": 0.09378072763028515,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9999830768577445e-05,
+      "loss": 0.7101,
+      "step": 1526
+    },
+    {
+      "epoch": 0.09384218289085546,
+      "grad_norm": 29.25,
+      "learning_rate": 1.999981695933549e-05,
+      "loss": 0.5927,
+      "step": 1527
+    },
+    {
+      "epoch": 0.09390363815142576,
+      "grad_norm": 29.5,
+      "learning_rate": 1.999980260856137e-05,
+      "loss": 0.5613,
+      "step": 1528
+    },
+    {
+      "epoch": 0.09396509341199606,
+      "grad_norm": 21.0,
+      "learning_rate": 1.9999787716255867e-05,
+      "loss": 0.6787,
+      "step": 1529
+    },
+    {
+      "epoch": 0.09402654867256637,
+      "grad_norm": 10.0,
+      "learning_rate": 1.9999772282419783e-05,
+      "loss": 0.6078,
+      "step": 1530
+    },
+    {
+      "epoch": 0.09408800393313668,
+      "grad_norm": 23.625,
+      "learning_rate": 1.9999756307053947e-05,
+      "loss": 0.6432,
+      "step": 1531
+    },
+    {
+      "epoch": 0.09414945919370699,
+      "grad_norm": 30.5,
+      "learning_rate": 1.999973979015924e-05,
+      "loss": 0.7424,
+      "step": 1532
+    },
+    {
+      "epoch": 0.09421091445427729,
+      "grad_norm": 22.125,
+      "learning_rate": 1.999972273173654e-05,
+      "loss": 0.6113,
+      "step": 1533
+    },
+    {
+      "epoch": 0.09427236971484759,
+      "grad_norm": 48.25,
+      "learning_rate": 1.999970513178678e-05,
+      "loss": 0.6524,
+      "step": 1534
+    },
+    {
+      "epoch": 0.0943338249754179,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9999686990310914e-05,
+      "loss": 0.8329,
+      "step": 1535
+    },
+    {
+      "epoch": 0.0943952802359882,
+      "grad_norm": 36.5,
+      "learning_rate": 1.999966830730992e-05,
+      "loss": 0.8255,
+      "step": 1536
+    },
+    {
+      "epoch": 0.0944567354965585,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9999649082784807e-05,
+      "loss": 0.7094,
+      "step": 1537
+    },
+    {
+      "epoch": 0.0945181907571288,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9999629316736624e-05,
+      "loss": 0.7507,
+      "step": 1538
+    },
+    {
+      "epoch": 0.09457964601769911,
+      "grad_norm": 21.875,
+      "learning_rate": 1.999960900916644e-05,
+      "loss": 0.7474,
+      "step": 1539
+    },
+    {
+      "epoch": 0.09464110127826943,
+      "grad_norm": 17.875,
+      "learning_rate": 1.999958816007535e-05,
+      "loss": 0.5937,
+      "step": 1540
+    },
+    {
+      "epoch": 0.09470255653883973,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9999566769464483e-05,
+      "loss": 0.709,
+      "step": 1541
+    },
+    {
+      "epoch": 0.09476401179941003,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9999544837335007e-05,
+      "loss": 0.5635,
+      "step": 1542
+    },
+    {
+      "epoch": 0.09482546705998034,
+      "grad_norm": 18.625,
+      "learning_rate": 1.99995223636881e-05,
+      "loss": 0.53,
+      "step": 1543
+    },
+    {
+      "epoch": 0.09488692232055064,
+      "grad_norm": 10.125,
+      "learning_rate": 1.999949934852498e-05,
+      "loss": 0.5143,
+      "step": 1544
+    },
+    {
+      "epoch": 0.09494837758112094,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9999475791846896e-05,
+      "loss": 0.7373,
+      "step": 1545
+    },
+    {
+      "epoch": 0.09500983284169125,
+      "grad_norm": 43.5,
+      "learning_rate": 1.9999451693655125e-05,
+      "loss": 0.7662,
+      "step": 1546
+    },
+    {
+      "epoch": 0.09507128810226155,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9999427053950965e-05,
+      "loss": 0.6075,
+      "step": 1547
+    },
+    {
+      "epoch": 0.09513274336283185,
+      "grad_norm": 26.25,
+      "learning_rate": 1.999940187273576e-05,
+      "loss": 0.613,
+      "step": 1548
+    },
+    {
+      "epoch": 0.09519419862340216,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9999376150010868e-05,
+      "loss": 0.5371,
+      "step": 1549
+    },
+    {
+      "epoch": 0.09525565388397247,
+      "grad_norm": 33.25,
+      "learning_rate": 1.999934988577768e-05,
+      "loss": 0.6484,
+      "step": 1550
+    },
+    {
+      "epoch": 0.09531710914454278,
+      "grad_norm": 35.75,
+      "learning_rate": 1.9999323080037623e-05,
+      "loss": 0.6118,
+      "step": 1551
+    },
+    {
+      "epoch": 0.09537856440511308,
+      "grad_norm": 42.0,
+      "learning_rate": 1.9999295732792146e-05,
+      "loss": 0.4874,
+      "step": 1552
+    },
+    {
+      "epoch": 0.09544001966568338,
+      "grad_norm": 50.25,
+      "learning_rate": 1.9999267844042734e-05,
+      "loss": 0.6037,
+      "step": 1553
+    },
+    {
+      "epoch": 0.09550147492625369,
+      "grad_norm": 16.75,
+      "learning_rate": 1.9999239413790895e-05,
+      "loss": 0.475,
+      "step": 1554
+    },
+    {
+      "epoch": 0.09556293018682399,
+      "grad_norm": 23.75,
+      "learning_rate": 1.9999210442038164e-05,
+      "loss": 0.7353,
+      "step": 1555
+    },
+    {
+      "epoch": 0.0956243854473943,
+      "grad_norm": 35.75,
+      "learning_rate": 1.9999180928786113e-05,
+      "loss": 0.6392,
+      "step": 1556
+    },
+    {
+      "epoch": 0.0956858407079646,
+      "grad_norm": 31.375,
+      "learning_rate": 1.9999150874036344e-05,
+      "loss": 0.5754,
+      "step": 1557
+    },
+    {
+      "epoch": 0.0957472959685349,
+      "grad_norm": 19.0,
+      "learning_rate": 1.9999120277790477e-05,
+      "loss": 0.5501,
+      "step": 1558
+    },
+    {
+      "epoch": 0.0958087512291052,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9999089140050182e-05,
+      "loss": 0.4345,
+      "step": 1559
+    },
+    {
+      "epoch": 0.09587020648967552,
+      "grad_norm": 14.375,
+      "learning_rate": 1.999905746081713e-05,
+      "loss": 0.4183,
+      "step": 1560
+    },
+    {
+      "epoch": 0.09593166175024583,
+      "grad_norm": 23.5,
+      "learning_rate": 1.9999025240093045e-05,
+      "loss": 0.5495,
+      "step": 1561
+    },
+    {
+      "epoch": 0.09599311701081613,
+      "grad_norm": 16.5,
+      "learning_rate": 1.999899247787967e-05,
+      "loss": 0.7035,
+      "step": 1562
+    },
+    {
+      "epoch": 0.09605457227138643,
+      "grad_norm": 15.875,
+      "learning_rate": 1.9998959174178777e-05,
+      "loss": 0.539,
+      "step": 1563
+    },
+    {
+      "epoch": 0.09611602753195674,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9998925328992175e-05,
+      "loss": 0.4575,
+      "step": 1564
+    },
+    {
+      "epoch": 0.09617748279252704,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9998890942321693e-05,
+      "loss": 0.5762,
+      "step": 1565
+    },
+    {
+      "epoch": 0.09623893805309734,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.9998856014169193e-05,
+      "loss": 0.6934,
+      "step": 1566
+    },
+    {
+      "epoch": 0.09630039331366765,
+      "grad_norm": 11.75,
+      "learning_rate": 1.999882054453657e-05,
+      "loss": 0.629,
+      "step": 1567
+    },
+    {
+      "epoch": 0.09636184857423795,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9998784533425743e-05,
+      "loss": 0.5115,
+      "step": 1568
+    },
+    {
+      "epoch": 0.09642330383480827,
+      "grad_norm": 23.5,
+      "learning_rate": 1.9998747980838655e-05,
+      "loss": 0.4956,
+      "step": 1569
+    },
+    {
+      "epoch": 0.09648475909537857,
+      "grad_norm": 63.5,
+      "learning_rate": 1.9998710886777298e-05,
+      "loss": 0.864,
+      "step": 1570
+    },
+    {
+      "epoch": 0.09654621435594887,
+      "grad_norm": 11.125,
+      "learning_rate": 1.9998673251243672e-05,
+      "loss": 0.5724,
+      "step": 1571
+    },
+    {
+      "epoch": 0.09660766961651918,
+      "grad_norm": 41.25,
+      "learning_rate": 1.999863507423982e-05,
+      "loss": 0.7549,
+      "step": 1572
+    },
+    {
+      "epoch": 0.09666912487708948,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9998596355767805e-05,
+      "loss": 0.7428,
+      "step": 1573
+    },
+    {
+      "epoch": 0.09673058013765978,
+      "grad_norm": 47.5,
+      "learning_rate": 1.9998557095829727e-05,
+      "loss": 0.5957,
+      "step": 1574
+    },
+    {
+      "epoch": 0.09679203539823009,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.9998517294427712e-05,
+      "loss": 0.49,
+      "step": 1575
+    },
+    {
+      "epoch": 0.09685349065880039,
+      "grad_norm": 103.0,
+      "learning_rate": 1.9998476951563914e-05,
+      "loss": 0.5234,
+      "step": 1576
+    },
+    {
+      "epoch": 0.09691494591937069,
+      "grad_norm": 15.375,
+      "learning_rate": 1.9998436067240518e-05,
+      "loss": 0.5719,
+      "step": 1577
+    },
+    {
+      "epoch": 0.096976401179941,
+      "grad_norm": 58.25,
+      "learning_rate": 1.999839464145974e-05,
+      "loss": 0.6186,
+      "step": 1578
+    },
+    {
+      "epoch": 0.09703785644051131,
+      "grad_norm": 15.75,
+      "learning_rate": 1.9998352674223816e-05,
+      "loss": 0.5239,
+      "step": 1579
+    },
+    {
+      "epoch": 0.09709931170108162,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.999831016553503e-05,
+      "loss": 0.4837,
+      "step": 1580
+    },
+    {
+      "epoch": 0.09716076696165192,
+      "grad_norm": 35.75,
+      "learning_rate": 1.999826711539568e-05,
+      "loss": 0.3647,
+      "step": 1581
+    },
+    {
+      "epoch": 0.09722222222222222,
+      "grad_norm": 33.75,
+      "learning_rate": 1.9998223523808092e-05,
+      "loss": 0.426,
+      "step": 1582
+    },
+    {
+      "epoch": 0.09728367748279253,
+      "grad_norm": 70.5,
+      "learning_rate": 1.9998179390774632e-05,
+      "loss": 0.3515,
+      "step": 1583
+    },
+    {
+      "epoch": 0.09734513274336283,
+      "grad_norm": 28.0,
+      "learning_rate": 1.9998134716297688e-05,
+      "loss": 0.4279,
+      "step": 1584
+    },
+    {
+      "epoch": 0.09740658800393313,
+      "grad_norm": 20.75,
+      "learning_rate": 1.999808950037968e-05,
+      "loss": 0.4479,
+      "step": 1585
+    },
+    {
+      "epoch": 0.09746804326450344,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9998043743023056e-05,
+      "loss": 0.6468,
+      "step": 1586
+    },
+    {
+      "epoch": 0.09752949852507374,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9997997444230294e-05,
+      "loss": 0.6227,
+      "step": 1587
+    },
+    {
+      "epoch": 0.09759095378564406,
+      "grad_norm": 44.0,
+      "learning_rate": 1.99979506040039e-05,
+      "loss": 0.5317,
+      "step": 1588
+    },
+    {
+      "epoch": 0.09765240904621436,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.9997903222346418e-05,
+      "loss": 0.3899,
+      "step": 1589
+    },
+    {
+      "epoch": 0.09771386430678466,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.9997855299260403e-05,
+      "loss": 0.3965,
+      "step": 1590
+    },
+    {
+      "epoch": 0.09777531956735497,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9997806834748455e-05,
+      "loss": 0.449,
+      "step": 1591
+    },
+    {
+      "epoch": 0.09783677482792527,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.9997757828813203e-05,
+      "loss": 0.3195,
+      "step": 1592
+    },
+    {
+      "epoch": 0.09789823008849557,
+      "grad_norm": 27.875,
+      "learning_rate": 1.9997708281457293e-05,
+      "loss": 0.628,
+      "step": 1593
+    },
+    {
+      "epoch": 0.09795968534906588,
+      "grad_norm": 53.75,
+      "learning_rate": 1.9997658192683412e-05,
+      "loss": 0.4898,
+      "step": 1594
+    },
+    {
+      "epoch": 0.09802114060963618,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9997607562494277e-05,
+      "loss": 0.5189,
+      "step": 1595
+    },
+    {
+      "epoch": 0.09808259587020648,
+      "grad_norm": 20.25,
+      "learning_rate": 1.9997556390892623e-05,
+      "loss": 0.5038,
+      "step": 1596
+    },
+    {
+      "epoch": 0.09814405113077679,
+      "grad_norm": 31.5,
+      "learning_rate": 1.9997504677881224e-05,
+      "loss": 0.661,
+      "step": 1597
+    },
+    {
+      "epoch": 0.0982055063913471,
+      "grad_norm": 23.25,
+      "learning_rate": 1.999745242346288e-05,
+      "loss": 0.5747,
+      "step": 1598
+    },
+    {
+      "epoch": 0.09826696165191741,
+      "grad_norm": 42.5,
+      "learning_rate": 1.999739962764042e-05,
+      "loss": 0.4613,
+      "step": 1599
+    },
+    {
+      "epoch": 0.09832841691248771,
+      "grad_norm": 19.125,
+      "learning_rate": 1.9997346290416703e-05,
+      "loss": 0.4801,
+      "step": 1600
+    },
+    {
+      "epoch": 0.09838987217305802,
+      "grad_norm": 31.875,
+      "learning_rate": 1.999729241179462e-05,
+      "loss": 0.8283,
+      "step": 1601
+    },
+    {
+      "epoch": 0.09845132743362832,
+      "grad_norm": 39.25,
+      "learning_rate": 1.9997237991777085e-05,
+      "loss": 0.7647,
+      "step": 1602
+    },
+    {
+      "epoch": 0.09851278269419862,
+      "grad_norm": 25.625,
+      "learning_rate": 1.999718303036705e-05,
+      "loss": 0.7577,
+      "step": 1603
+    },
+    {
+      "epoch": 0.09857423795476893,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9997127527567487e-05,
+      "loss": 0.6192,
+      "step": 1604
+    },
+    {
+      "epoch": 0.09863569321533923,
+      "grad_norm": 58.25,
+      "learning_rate": 1.99970714833814e-05,
+      "loss": 0.6892,
+      "step": 1605
+    },
+    {
+      "epoch": 0.09869714847590953,
+      "grad_norm": 24.375,
+      "learning_rate": 1.9997014897811834e-05,
+      "loss": 0.5686,
+      "step": 1606
+    },
+    {
+      "epoch": 0.09875860373647984,
+      "grad_norm": 26.25,
+      "learning_rate": 1.9996957770861843e-05,
+      "loss": 0.5501,
+      "step": 1607
+    },
+    {
+      "epoch": 0.09882005899705015,
+      "grad_norm": 24.0,
+      "learning_rate": 1.9996900102534526e-05,
+      "loss": 0.4202,
+      "step": 1608
+    },
+    {
+      "epoch": 0.09888151425762046,
+      "grad_norm": 78.0,
+      "learning_rate": 1.9996841892833e-05,
+      "loss": 0.8203,
+      "step": 1609
+    },
+    {
+      "epoch": 0.09894296951819076,
+      "grad_norm": 38.0,
+      "learning_rate": 1.9996783141760425e-05,
+      "loss": 0.5394,
+      "step": 1610
+    },
+    {
+      "epoch": 0.09900442477876106,
+      "grad_norm": 45.25,
+      "learning_rate": 1.9996723849319978e-05,
+      "loss": 0.7046,
+      "step": 1611
+    },
+    {
+      "epoch": 0.09906588003933137,
+      "grad_norm": 27.125,
+      "learning_rate": 1.999666401551487e-05,
+      "loss": 0.6283,
+      "step": 1612
+    },
+    {
+      "epoch": 0.09912733529990167,
+      "grad_norm": 13.5,
+      "learning_rate": 1.999660364034835e-05,
+      "loss": 0.4878,
+      "step": 1613
+    },
+    {
+      "epoch": 0.09918879056047197,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9996542723823673e-05,
+      "loss": 0.4808,
+      "step": 1614
+    },
+    {
+      "epoch": 0.09925024582104228,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9996481265944146e-05,
+      "loss": 0.5646,
+      "step": 1615
+    },
+    {
+      "epoch": 0.09931170108161258,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9996419266713097e-05,
+      "loss": 0.6158,
+      "step": 1616
+    },
+    {
+      "epoch": 0.0993731563421829,
+      "grad_norm": 24.0,
+      "learning_rate": 1.999635672613388e-05,
+      "loss": 0.3868,
+      "step": 1617
+    },
+    {
+      "epoch": 0.0994346116027532,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9996293644209886e-05,
+      "loss": 0.4407,
+      "step": 1618
+    },
+    {
+      "epoch": 0.0994960668633235,
+      "grad_norm": 6.90625,
+      "learning_rate": 1.9996230020944535e-05,
+      "loss": 0.4726,
+      "step": 1619
+    },
+    {
+      "epoch": 0.09955752212389381,
+      "grad_norm": 13.625,
+      "learning_rate": 1.999616585634126e-05,
+      "loss": 0.5784,
+      "step": 1620
+    },
+    {
+      "epoch": 0.09961897738446411,
+      "grad_norm": 23.0,
+      "learning_rate": 1.9996101150403543e-05,
+      "loss": 0.5214,
+      "step": 1621
+    },
+    {
+      "epoch": 0.09968043264503441,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9996035903134893e-05,
+      "loss": 0.5937,
+      "step": 1622
+    },
+    {
+      "epoch": 0.09974188790560472,
+      "grad_norm": 6.0,
+      "learning_rate": 1.9995970114538835e-05,
+      "loss": 0.4794,
+      "step": 1623
+    },
+    {
+      "epoch": 0.09980334316617502,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9995903784618936e-05,
+      "loss": 0.4414,
+      "step": 1624
+    },
+    {
+      "epoch": 0.09986479842674532,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9995836913378786e-05,
+      "loss": 0.6542,
+      "step": 1625
+    },
+    {
+      "epoch": 0.09992625368731563,
+      "grad_norm": 16.875,
+      "learning_rate": 1.9995769500822007e-05,
+      "loss": 0.6602,
+      "step": 1626
+    },
+    {
+      "epoch": 0.09998770894788594,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9995701546952252e-05,
+      "loss": 0.5246,
+      "step": 1627
+    },
+    {
+      "epoch": 0.10004916420845625,
+      "grad_norm": 14.125,
+      "learning_rate": 1.9995633051773197e-05,
+      "loss": 0.5604,
+      "step": 1628
+    },
+    {
+      "epoch": 0.10011061946902655,
+      "grad_norm": 18.0,
+      "learning_rate": 1.9995564015288555e-05,
+      "loss": 0.5439,
+      "step": 1629
+    },
+    {
+      "epoch": 0.10017207472959685,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9995494437502064e-05,
+      "loss": 0.6075,
+      "step": 1630
+    },
+    {
+      "epoch": 0.10023352999016716,
+      "grad_norm": 23.875,
+      "learning_rate": 1.999542431841749e-05,
+      "loss": 0.6336,
+      "step": 1631
+    },
+    {
+      "epoch": 0.10029498525073746,
+      "grad_norm": 33.0,
+      "learning_rate": 1.999535365803863e-05,
+      "loss": 0.5192,
+      "step": 1632
+    },
+    {
+      "epoch": 0.10035644051130777,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9995282456369313e-05,
+      "loss": 0.502,
+      "step": 1633
+    },
+    {
+      "epoch": 0.10041789577187807,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9995210713413393e-05,
+      "loss": 0.5259,
+      "step": 1634
+    },
+    {
+      "epoch": 0.10047935103244837,
+      "grad_norm": 10.875,
+      "learning_rate": 1.9995138429174757e-05,
+      "loss": 0.4521,
+      "step": 1635
+    },
+    {
+      "epoch": 0.10054080629301869,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9995065603657317e-05,
+      "loss": 0.6616,
+      "step": 1636
+    },
+    {
+      "epoch": 0.10060226155358899,
+      "grad_norm": 22.875,
+      "learning_rate": 1.9994992236865017e-05,
+      "loss": 0.6891,
+      "step": 1637
+    },
+    {
+      "epoch": 0.1006637168141593,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.9994918328801835e-05,
+      "loss": 0.5476,
+      "step": 1638
+    },
+    {
+      "epoch": 0.1007251720747296,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.999484387947177e-05,
+      "loss": 0.4832,
+      "step": 1639
+    },
+    {
+      "epoch": 0.1007866273352999,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9994768888878848e-05,
+      "loss": 0.3903,
+      "step": 1640
+    },
+    {
+      "epoch": 0.1008480825958702,
+      "grad_norm": 31.125,
+      "learning_rate": 1.999469335702714e-05,
+      "loss": 0.7796,
+      "step": 1641
+    },
+    {
+      "epoch": 0.10090953785644051,
+      "grad_norm": 27.125,
+      "learning_rate": 1.999461728392073e-05,
+      "loss": 0.8646,
+      "step": 1642
+    },
+    {
+      "epoch": 0.10097099311701081,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9994540669563742e-05,
+      "loss": 0.6055,
+      "step": 1643
+    },
+    {
+      "epoch": 0.10103244837758112,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.999446351396032e-05,
+      "loss": 0.5912,
+      "step": 1644
+    },
+    {
+      "epoch": 0.10109390363815142,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9994385817114644e-05,
+      "loss": 0.5026,
+      "step": 1645
+    },
+    {
+      "epoch": 0.10115535889872174,
+      "grad_norm": 9.625,
+      "learning_rate": 1.9994307579030925e-05,
+      "loss": 0.5101,
+      "step": 1646
+    },
+    {
+      "epoch": 0.10121681415929204,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9994228799713396e-05,
+      "loss": 0.5674,
+      "step": 1647
+    },
+    {
+      "epoch": 0.10127826941986234,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.9994149479166324e-05,
+      "loss": 0.4756,
+      "step": 1648
+    },
+    {
+      "epoch": 0.10133972468043265,
+      "grad_norm": 30.375,
+      "learning_rate": 1.999406961739401e-05,
+      "loss": 0.7109,
+      "step": 1649
+    },
+    {
+      "epoch": 0.10140117994100295,
+      "grad_norm": 24.5,
+      "learning_rate": 1.9993989214400765e-05,
+      "loss": 0.943,
+      "step": 1650
+    },
+    {
+      "epoch": 0.10146263520157325,
+      "grad_norm": 22.625,
+      "learning_rate": 1.999390827019096e-05,
+      "loss": 0.6687,
+      "step": 1651
+    },
+    {
+      "epoch": 0.10152409046214356,
+      "grad_norm": 34.75,
+      "learning_rate": 1.9993826784768967e-05,
+      "loss": 0.6459,
+      "step": 1652
+    },
+    {
+      "epoch": 0.10158554572271386,
+      "grad_norm": 28.75,
+      "learning_rate": 1.9993744758139203e-05,
+      "loss": 0.7094,
+      "step": 1653
+    },
+    {
+      "epoch": 0.10164700098328416,
+      "grad_norm": 19.375,
+      "learning_rate": 1.999366219030611e-05,
+      "loss": 0.5492,
+      "step": 1654
+    },
+    {
+      "epoch": 0.10170845624385447,
+      "grad_norm": 30.25,
+      "learning_rate": 1.999357908127416e-05,
+      "loss": 0.5481,
+      "step": 1655
+    },
+    {
+      "epoch": 0.10176991150442478,
+      "grad_norm": 26.25,
+      "learning_rate": 1.999349543104785e-05,
+      "loss": 0.764,
+      "step": 1656
+    },
+    {
+      "epoch": 0.10183136676499509,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9993411239631713e-05,
+      "loss": 0.6592,
+      "step": 1657
+    },
+    {
+      "epoch": 0.10189282202556539,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9993326507030312e-05,
+      "loss": 0.9689,
+      "step": 1658
+    },
+    {
+      "epoch": 0.1019542772861357,
+      "grad_norm": 24.5,
+      "learning_rate": 1.9993241233248227e-05,
+      "loss": 0.6126,
+      "step": 1659
+    },
+    {
+      "epoch": 0.102015732546706,
+      "grad_norm": 17.125,
+      "learning_rate": 1.999315541829008e-05,
+      "loss": 0.5927,
+      "step": 1660
+    },
+    {
+      "epoch": 0.1020771878072763,
+      "grad_norm": 17.875,
+      "learning_rate": 1.999306906216052e-05,
+      "loss": 0.6562,
+      "step": 1661
+    },
+    {
+      "epoch": 0.1021386430678466,
+      "grad_norm": 30.375,
+      "learning_rate": 1.9992982164864223e-05,
+      "loss": 0.4678,
+      "step": 1662
+    },
+    {
+      "epoch": 0.10220009832841691,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9992894726405894e-05,
+      "loss": 0.4465,
+      "step": 1663
+    },
+    {
+      "epoch": 0.10226155358898721,
+      "grad_norm": 17.0,
+      "learning_rate": 1.9992806746790266e-05,
+      "loss": 0.4717,
+      "step": 1664
+    },
+    {
+      "epoch": 0.10232300884955753,
+      "grad_norm": 56.5,
+      "learning_rate": 1.999271822602211e-05,
+      "loss": 0.8307,
+      "step": 1665
+    },
+    {
+      "epoch": 0.10238446411012783,
+      "grad_norm": 41.25,
+      "learning_rate": 1.999262916410621e-05,
+      "loss": 0.9471,
+      "step": 1666
+    },
+    {
+      "epoch": 0.10244591937069814,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9992539561047403e-05,
+      "loss": 0.9567,
+      "step": 1667
+    },
+    {
+      "epoch": 0.10250737463126844,
+      "grad_norm": 31.625,
+      "learning_rate": 1.9992449416850527e-05,
+      "loss": 0.834,
+      "step": 1668
+    },
+    {
+      "epoch": 0.10256882989183874,
+      "grad_norm": 29.375,
+      "learning_rate": 1.999235873152047e-05,
+      "loss": 0.7224,
+      "step": 1669
+    },
+    {
+      "epoch": 0.10263028515240905,
+      "grad_norm": 30.5,
+      "learning_rate": 1.9992267505062143e-05,
+      "loss": 0.6981,
+      "step": 1670
+    },
+    {
+      "epoch": 0.10269174041297935,
+      "grad_norm": 19.125,
+      "learning_rate": 1.9992175737480487e-05,
+      "loss": 0.5298,
+      "step": 1671
+    },
+    {
+      "epoch": 0.10275319567354965,
+      "grad_norm": 17.25,
+      "learning_rate": 1.999208342878047e-05,
+      "loss": 0.5783,
+      "step": 1672
+    },
+    {
+      "epoch": 0.10281465093411996,
+      "grad_norm": 35.75,
+      "learning_rate": 1.999199057896709e-05,
+      "loss": 0.7088,
+      "step": 1673
+    },
+    {
+      "epoch": 0.10287610619469026,
+      "grad_norm": 30.875,
+      "learning_rate": 1.999189718804538e-05,
+      "loss": 0.6098,
+      "step": 1674
+    },
+    {
+      "epoch": 0.10293756145526058,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9991803256020393e-05,
+      "loss": 0.886,
+      "step": 1675
+    },
+    {
+      "epoch": 0.10299901671583088,
+      "grad_norm": 29.25,
+      "learning_rate": 1.9991708782897214e-05,
+      "loss": 0.504,
+      "step": 1676
+    },
+    {
+      "epoch": 0.10306047197640118,
+      "grad_norm": 24.375,
+      "learning_rate": 1.9991613768680965e-05,
+      "loss": 0.7169,
+      "step": 1677
+    },
+    {
+      "epoch": 0.10312192723697149,
+      "grad_norm": 26.75,
+      "learning_rate": 1.9991518213376787e-05,
+      "loss": 0.6876,
+      "step": 1678
+    },
+    {
+      "epoch": 0.10318338249754179,
+      "grad_norm": 43.0,
+      "learning_rate": 1.999142211698986e-05,
+      "loss": 0.4543,
+      "step": 1679
+    },
+    {
+      "epoch": 0.10324483775811209,
+      "grad_norm": 41.5,
+      "learning_rate": 1.999132547952538e-05,
+      "loss": 0.7736,
+      "step": 1680
+    },
+    {
+      "epoch": 0.1033062930186824,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9991228300988586e-05,
+      "loss": 0.5768,
+      "step": 1681
+    },
+    {
+      "epoch": 0.1033677482792527,
+      "grad_norm": 46.75,
+      "learning_rate": 1.9991130581384737e-05,
+      "loss": 0.6743,
+      "step": 1682
+    },
+    {
+      "epoch": 0.103429203539823,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9991032320719132e-05,
+      "loss": 0.5048,
+      "step": 1683
+    },
+    {
+      "epoch": 0.10349065880039332,
+      "grad_norm": 27.75,
+      "learning_rate": 1.9990933518997086e-05,
+      "loss": 0.539,
+      "step": 1684
+    },
+    {
+      "epoch": 0.10355211406096362,
+      "grad_norm": 30.25,
+      "learning_rate": 1.999083417622395e-05,
+      "loss": 0.4855,
+      "step": 1685
+    },
+    {
+      "epoch": 0.10361356932153393,
+      "grad_norm": 15.75,
+      "learning_rate": 1.9990734292405102e-05,
+      "loss": 0.4706,
+      "step": 1686
+    },
+    {
+      "epoch": 0.10367502458210423,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.9990633867545956e-05,
+      "loss": 0.4691,
+      "step": 1687
+    },
+    {
+      "epoch": 0.10373647984267453,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.9990532901651948e-05,
+      "loss": 0.4734,
+      "step": 1688
+    },
+    {
+      "epoch": 0.10379793510324484,
+      "grad_norm": 10.5,
+      "learning_rate": 1.9990431394728547e-05,
+      "loss": 0.4649,
+      "step": 1689
+    },
+    {
+      "epoch": 0.10385939036381514,
+      "grad_norm": 10.75,
+      "learning_rate": 1.999032934678125e-05,
+      "loss": 0.5539,
+      "step": 1690
+    },
+    {
+      "epoch": 0.10392084562438544,
+      "grad_norm": 6.8125,
+      "learning_rate": 1.9990226757815582e-05,
+      "loss": 0.39,
+      "step": 1691
+    },
+    {
+      "epoch": 0.10398230088495575,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.9990123627837094e-05,
+      "loss": 0.3595,
+      "step": 1692
+    },
+    {
+      "epoch": 0.10404375614552605,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.9990019956851384e-05,
+      "loss": 0.5008,
+      "step": 1693
+    },
+    {
+      "epoch": 0.10410521140609637,
+      "grad_norm": 19.75,
+      "learning_rate": 1.998991574486405e-05,
+      "loss": 0.5157,
+      "step": 1694
+    },
+    {
+      "epoch": 0.10416666666666667,
+      "grad_norm": 18.875,
+      "learning_rate": 1.998981099188075e-05,
+      "loss": 0.3759,
+      "step": 1695
+    },
+    {
+      "epoch": 0.10422812192723697,
+      "grad_norm": 28.875,
+      "learning_rate": 1.998970569790715e-05,
+      "loss": 0.3985,
+      "step": 1696
+    },
+    {
+      "epoch": 0.10428957718780728,
+      "grad_norm": 50.5,
+      "learning_rate": 1.998959986294895e-05,
+      "loss": 0.6884,
+      "step": 1697
+    },
+    {
+      "epoch": 0.10435103244837758,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9989493487011883e-05,
+      "loss": 0.713,
+      "step": 1698
+    },
+    {
+      "epoch": 0.10441248770894788,
+      "grad_norm": 54.25,
+      "learning_rate": 1.9989386570101716e-05,
+      "loss": 0.7491,
+      "step": 1699
+    },
+    {
+      "epoch": 0.10447394296951819,
+      "grad_norm": 39.0,
+      "learning_rate": 1.9989279112224226e-05,
+      "loss": 0.6194,
+      "step": 1700
+    },
+    {
+      "epoch": 0.10453539823008849,
+      "grad_norm": 26.125,
+      "learning_rate": 1.998917111338525e-05,
+      "loss": 0.4844,
+      "step": 1701
+    },
+    {
+      "epoch": 0.1045968534906588,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9989062573590618e-05,
+      "loss": 0.6842,
+      "step": 1702
+    },
+    {
+      "epoch": 0.10465830875122911,
+      "grad_norm": 8.5625,
+      "learning_rate": 1.9988953492846218e-05,
+      "loss": 0.3785,
+      "step": 1703
+    },
+    {
+      "epoch": 0.10471976401179942,
+      "grad_norm": 29.375,
+      "learning_rate": 1.998884387115796e-05,
+      "loss": 0.4311,
+      "step": 1704
+    },
+    {
+      "epoch": 0.10478121927236972,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9988733708531772e-05,
+      "loss": 0.7294,
+      "step": 1705
+    },
+    {
+      "epoch": 0.10484267453294002,
+      "grad_norm": 28.375,
+      "learning_rate": 1.9988623004973625e-05,
+      "loss": 0.7229,
+      "step": 1706
+    },
+    {
+      "epoch": 0.10490412979351033,
+      "grad_norm": 6.28125,
+      "learning_rate": 1.9988511760489514e-05,
+      "loss": 0.6447,
+      "step": 1707
+    },
+    {
+      "epoch": 0.10496558505408063,
+      "grad_norm": 20.25,
+      "learning_rate": 1.998839997508546e-05,
+      "loss": 0.6963,
+      "step": 1708
+    },
+    {
+      "epoch": 0.10502704031465093,
+      "grad_norm": 63.0,
+      "learning_rate": 1.9988287648767523e-05,
+      "loss": 0.6341,
+      "step": 1709
+    },
+    {
+      "epoch": 0.10508849557522124,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.9988174781541774e-05,
+      "loss": 0.3937,
+      "step": 1710
+    },
+    {
+      "epoch": 0.10514995083579154,
+      "grad_norm": 22.5,
+      "learning_rate": 1.9988061373414342e-05,
+      "loss": 0.4364,
+      "step": 1711
+    },
+    {
+      "epoch": 0.10521140609636184,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9987947424391358e-05,
+      "loss": 0.4652,
+      "step": 1712
+    },
+    {
+      "epoch": 0.10527286135693216,
+      "grad_norm": 28.625,
+      "learning_rate": 1.9987832934478992e-05,
+      "loss": 0.6301,
+      "step": 1713
+    },
+    {
+      "epoch": 0.10533431661750246,
+      "grad_norm": 22.625,
+      "learning_rate": 1.9987717903683447e-05,
+      "loss": 0.6495,
+      "step": 1714
+    },
+    {
+      "epoch": 0.10539577187807277,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9987602332010952e-05,
+      "loss": 0.668,
+      "step": 1715
+    },
+    {
+      "epoch": 0.10545722713864307,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9987486219467764e-05,
+      "loss": 0.5514,
+      "step": 1716
+    },
+    {
+      "epoch": 0.10551868239921337,
+      "grad_norm": 17.0,
+      "learning_rate": 1.998736956606018e-05,
+      "loss": 0.5,
+      "step": 1717
+    },
+    {
+      "epoch": 0.10558013765978368,
+      "grad_norm": 35.25,
+      "learning_rate": 1.99872523717945e-05,
+      "loss": 0.5595,
+      "step": 1718
+    },
+    {
+      "epoch": 0.10564159292035398,
+      "grad_norm": 11.25,
+      "learning_rate": 1.998713463667709e-05,
+      "loss": 0.3511,
+      "step": 1719
+    },
+    {
+      "epoch": 0.10570304818092428,
+      "grad_norm": 3.78125,
+      "learning_rate": 1.9987016360714307e-05,
+      "loss": 0.3298,
+      "step": 1720
+    },
+    {
+      "epoch": 0.10576450344149459,
+      "grad_norm": 42.5,
+      "learning_rate": 1.998689754391257e-05,
+      "loss": 0.6871,
+      "step": 1721
+    },
+    {
+      "epoch": 0.10582595870206489,
+      "grad_norm": 33.0,
+      "learning_rate": 1.998677818627831e-05,
+      "loss": 0.7382,
+      "step": 1722
+    },
+    {
+      "epoch": 0.10588741396263521,
+      "grad_norm": 24.625,
+      "learning_rate": 1.998665828781799e-05,
+      "loss": 0.7832,
+      "step": 1723
+    },
+    {
+      "epoch": 0.10594886922320551,
+      "grad_norm": 37.5,
+      "learning_rate": 1.99865378485381e-05,
+      "loss": 0.6788,
+      "step": 1724
+    },
+    {
+      "epoch": 0.10601032448377581,
+      "grad_norm": 34.5,
+      "learning_rate": 1.9986416868445167e-05,
+      "loss": 0.7302,
+      "step": 1725
+    },
+    {
+      "epoch": 0.10607177974434612,
+      "grad_norm": 50.0,
+      "learning_rate": 1.9986295347545738e-05,
+      "loss": 0.6242,
+      "step": 1726
+    },
+    {
+      "epoch": 0.10613323500491642,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.99861732858464e-05,
+      "loss": 0.3836,
+      "step": 1727
+    },
+    {
+      "epoch": 0.10619469026548672,
+      "grad_norm": 26.0,
+      "learning_rate": 1.998605068335376e-05,
+      "loss": 0.6436,
+      "step": 1728
+    },
+    {
+      "epoch": 0.10625614552605703,
+      "grad_norm": 46.0,
+      "learning_rate": 1.9985927540074453e-05,
+      "loss": 0.5314,
+      "step": 1729
+    },
+    {
+      "epoch": 0.10631760078662733,
+      "grad_norm": 40.0,
+      "learning_rate": 1.9985803856015155e-05,
+      "loss": 0.623,
+      "step": 1730
+    },
+    {
+      "epoch": 0.10637905604719763,
+      "grad_norm": 10.25,
+      "learning_rate": 1.998567963118256e-05,
+      "loss": 0.606,
+      "step": 1731
+    },
+    {
+      "epoch": 0.10644051130776795,
+      "grad_norm": 29.375,
+      "learning_rate": 1.9985554865583394e-05,
+      "loss": 0.5354,
+      "step": 1732
+    },
+    {
+      "epoch": 0.10650196656833825,
+      "grad_norm": 29.125,
+      "learning_rate": 1.998542955922442e-05,
+      "loss": 0.4986,
+      "step": 1733
+    },
+    {
+      "epoch": 0.10656342182890856,
+      "grad_norm": 10.875,
+      "learning_rate": 1.9985303712112415e-05,
+      "loss": 0.4722,
+      "step": 1734
+    },
+    {
+      "epoch": 0.10662487708947886,
+      "grad_norm": 48.75,
+      "learning_rate": 1.99851773242542e-05,
+      "loss": 0.4711,
+      "step": 1735
+    },
+    {
+      "epoch": 0.10668633235004917,
+      "grad_norm": 6.71875,
+      "learning_rate": 1.9985050395656617e-05,
+      "loss": 0.4718,
+      "step": 1736
+    },
+    {
+      "epoch": 0.10674778761061947,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9984922926326543e-05,
+      "loss": 0.4948,
+      "step": 1737
+    },
+    {
+      "epoch": 0.10680924287118977,
+      "grad_norm": 53.25,
+      "learning_rate": 1.9984794916270876e-05,
+      "loss": 0.6574,
+      "step": 1738
+    },
+    {
+      "epoch": 0.10687069813176008,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9984666365496555e-05,
+      "loss": 0.5539,
+      "step": 1739
+    },
+    {
+      "epoch": 0.10693215339233038,
+      "grad_norm": 15.25,
+      "learning_rate": 1.9984537274010536e-05,
+      "loss": 0.6014,
+      "step": 1740
+    },
+    {
+      "epoch": 0.10699360865290068,
+      "grad_norm": 11.75,
+      "learning_rate": 1.9984407641819812e-05,
+      "loss": 0.4793,
+      "step": 1741
+    },
+    {
+      "epoch": 0.107055063913471,
+      "grad_norm": 11.25,
+      "learning_rate": 1.99842774689314e-05,
+      "loss": 0.4216,
+      "step": 1742
+    },
+    {
+      "epoch": 0.1071165191740413,
+      "grad_norm": 8.8125,
+      "learning_rate": 1.9984146755352354e-05,
+      "loss": 0.4465,
+      "step": 1743
+    },
+    {
+      "epoch": 0.1071779744346116,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.998401550108975e-05,
+      "loss": 0.4818,
+      "step": 1744
+    },
+    {
+      "epoch": 0.10723942969518191,
+      "grad_norm": 28.625,
+      "learning_rate": 1.99838837061507e-05,
+      "loss": 0.6439,
+      "step": 1745
+    },
+    {
+      "epoch": 0.10730088495575221,
+      "grad_norm": 106.0,
+      "learning_rate": 1.9983751370542334e-05,
+      "loss": 0.4896,
+      "step": 1746
+    },
+    {
+      "epoch": 0.10736234021632252,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.9983618494271825e-05,
+      "loss": 0.5001,
+      "step": 1747
+    },
+    {
+      "epoch": 0.10742379547689282,
+      "grad_norm": 45.75,
+      "learning_rate": 1.998348507734636e-05,
+      "loss": 0.4929,
+      "step": 1748
+    },
+    {
+      "epoch": 0.10748525073746312,
+      "grad_norm": 31.25,
+      "learning_rate": 1.998335111977318e-05,
+      "loss": 0.3756,
+      "step": 1749
+    },
+    {
+      "epoch": 0.10754670599803343,
+      "grad_norm": 35.25,
+      "learning_rate": 1.9983216621559525e-05,
+      "loss": 0.4998,
+      "step": 1750
+    },
+    {
+      "epoch": 0.10760816125860374,
+      "grad_norm": 30.75,
+      "learning_rate": 1.9983081582712684e-05,
+      "loss": 0.4718,
+      "step": 1751
+    },
+    {
+      "epoch": 0.10766961651917405,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.998294600323997e-05,
+      "loss": 0.3708,
+      "step": 1752
+    },
+    {
+      "epoch": 0.10773107177974435,
+      "grad_norm": 37.0,
+      "learning_rate": 1.998280988314872e-05,
+      "loss": 0.5807,
+      "step": 1753
+    },
+    {
+      "epoch": 0.10779252704031465,
+      "grad_norm": 58.5,
+      "learning_rate": 1.9982673222446312e-05,
+      "loss": 0.5386,
+      "step": 1754
+    },
+    {
+      "epoch": 0.10785398230088496,
+      "grad_norm": 5.34375,
+      "learning_rate": 1.998253602114015e-05,
+      "loss": 0.5628,
+      "step": 1755
+    },
+    {
+      "epoch": 0.10791543756145526,
+      "grad_norm": 65.5,
+      "learning_rate": 1.9982398279237657e-05,
+      "loss": 0.6032,
+      "step": 1756
+    },
+    {
+      "epoch": 0.10797689282202556,
+      "grad_norm": 60.5,
+      "learning_rate": 1.998225999674629e-05,
+      "loss": 0.5538,
+      "step": 1757
+    },
+    {
+      "epoch": 0.10803834808259587,
+      "grad_norm": 34.5,
+      "learning_rate": 1.9982121173673542e-05,
+      "loss": 0.5654,
+      "step": 1758
+    },
+    {
+      "epoch": 0.10809980334316617,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9981981810026932e-05,
+      "loss": 0.3955,
+      "step": 1759
+    },
+    {
+      "epoch": 0.10816125860373647,
+      "grad_norm": 26.0,
+      "learning_rate": 1.998184190581401e-05,
+      "loss": 0.4848,
+      "step": 1760
+    },
+    {
+      "epoch": 0.10822271386430679,
+      "grad_norm": 19.0,
+      "learning_rate": 1.998170146104234e-05,
+      "loss": 0.5413,
+      "step": 1761
+    },
+    {
+      "epoch": 0.1082841691248771,
+      "grad_norm": 10.875,
+      "learning_rate": 1.998156047571954e-05,
+      "loss": 0.5592,
+      "step": 1762
+    },
+    {
+      "epoch": 0.1083456243854474,
+      "grad_norm": 11.875,
+      "learning_rate": 1.998141894985324e-05,
+      "loss": 0.4617,
+      "step": 1763
+    },
+    {
+      "epoch": 0.1084070796460177,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9981276883451107e-05,
+      "loss": 0.486,
+      "step": 1764
+    },
+    {
+      "epoch": 0.108468534906588,
+      "grad_norm": 6.1875,
+      "learning_rate": 1.9981134276520828e-05,
+      "loss": 0.3595,
+      "step": 1765
+    },
+    {
+      "epoch": 0.10852999016715831,
+      "grad_norm": 47.75,
+      "learning_rate": 1.998099112907013e-05,
+      "loss": 0.5574,
+      "step": 1766
+    },
+    {
+      "epoch": 0.10859144542772861,
+      "grad_norm": 27.625,
+      "learning_rate": 1.998084744110677e-05,
+      "loss": 0.4474,
+      "step": 1767
+    },
+    {
+      "epoch": 0.10865290068829891,
+      "grad_norm": 39.25,
+      "learning_rate": 1.9980703212638522e-05,
+      "loss": 0.4586,
+      "step": 1768
+    },
+    {
+      "epoch": 0.10871435594886922,
+      "grad_norm": 22.375,
+      "learning_rate": 1.99805584436732e-05,
+      "loss": 0.6056,
+      "step": 1769
+    },
+    {
+      "epoch": 0.10877581120943952,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.998041313421864e-05,
+      "loss": 0.7046,
+      "step": 1770
+    },
+    {
+      "epoch": 0.10883726647000984,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9980267284282718e-05,
+      "loss": 0.7499,
+      "step": 1771
+    },
+    {
+      "epoch": 0.10889872173058014,
+      "grad_norm": 41.75,
+      "learning_rate": 1.9980120893873324e-05,
+      "loss": 0.7948,
+      "step": 1772
+    },
+    {
+      "epoch": 0.10896017699115045,
+      "grad_norm": 15.5625,
+      "learning_rate": 1.9979973962998396e-05,
+      "loss": 0.5272,
+      "step": 1773
+    },
+    {
+      "epoch": 0.10902163225172075,
+      "grad_norm": 39.75,
+      "learning_rate": 1.997982649166588e-05,
+      "loss": 0.7371,
+      "step": 1774
+    },
+    {
+      "epoch": 0.10908308751229105,
+      "grad_norm": 5.84375,
+      "learning_rate": 1.997967847988377e-05,
+      "loss": 0.4769,
+      "step": 1775
+    },
+    {
+      "epoch": 0.10914454277286136,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9979529927660076e-05,
+      "loss": 0.3689,
+      "step": 1776
+    },
+    {
+      "epoch": 0.10920599803343166,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9979380835002846e-05,
+      "loss": 0.5754,
+      "step": 1777
+    },
+    {
+      "epoch": 0.10926745329400196,
+      "grad_norm": 72.0,
+      "learning_rate": 1.9979231201920154e-05,
+      "loss": 0.5654,
+      "step": 1778
+    },
+    {
+      "epoch": 0.10932890855457227,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9979081028420105e-05,
+      "loss": 0.5503,
+      "step": 1779
+    },
+    {
+      "epoch": 0.10939036381514258,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9978930314510826e-05,
+      "loss": 0.5304,
+      "step": 1780
+    },
+    {
+      "epoch": 0.10945181907571289,
+      "grad_norm": 28.0,
+      "learning_rate": 1.9978779060200483e-05,
+      "loss": 0.5729,
+      "step": 1781
+    },
+    {
+      "epoch": 0.10951327433628319,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.9978627265497262e-05,
+      "loss": 0.4955,
+      "step": 1782
+    },
+    {
+      "epoch": 0.10957472959685349,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9978474930409396e-05,
+      "loss": 0.598,
+      "step": 1783
+    },
+    {
+      "epoch": 0.1096361848574238,
+      "grad_norm": 17.0,
+      "learning_rate": 1.997832205494512e-05,
+      "loss": 0.6184,
+      "step": 1784
+    },
+    {
+      "epoch": 0.1096976401179941,
+      "grad_norm": 16.625,
+      "learning_rate": 1.997816863911272e-05,
+      "loss": 0.5125,
+      "step": 1785
+    },
+    {
+      "epoch": 0.1097590953785644,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9978014682920503e-05,
+      "loss": 0.6683,
+      "step": 1786
+    },
+    {
+      "epoch": 0.1098205506391347,
+      "grad_norm": 50.75,
+      "learning_rate": 1.9977860186376804e-05,
+      "loss": 0.6064,
+      "step": 1787
+    },
+    {
+      "epoch": 0.10988200589970501,
+      "grad_norm": 9.0,
+      "learning_rate": 1.9977705149489996e-05,
+      "loss": 0.3939,
+      "step": 1788
+    },
+    {
+      "epoch": 0.10994346116027531,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.997754957226847e-05,
+      "loss": 0.3928,
+      "step": 1789
+    },
+    {
+      "epoch": 0.11000491642084563,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9977393454720646e-05,
+      "loss": 0.4789,
+      "step": 1790
+    },
+    {
+      "epoch": 0.11006637168141593,
+      "grad_norm": 12.5,
+      "learning_rate": 1.997723679685499e-05,
+      "loss": 0.4058,
+      "step": 1791
+    },
+    {
+      "epoch": 0.11012782694198624,
+      "grad_norm": 9.0,
+      "learning_rate": 1.9977079598679978e-05,
+      "loss": 0.4885,
+      "step": 1792
+    },
+    {
+      "epoch": 0.11018928220255654,
+      "grad_norm": 44.75,
+      "learning_rate": 1.997692186020413e-05,
+      "loss": 0.4216,
+      "step": 1793
+    },
+    {
+      "epoch": 0.11025073746312684,
+      "grad_norm": 39.0,
+      "learning_rate": 1.997676358143598e-05,
+      "loss": 0.4548,
+      "step": 1794
+    },
+    {
+      "epoch": 0.11031219272369715,
+      "grad_norm": 20.125,
+      "learning_rate": 1.99766047623841e-05,
+      "loss": 0.4144,
+      "step": 1795
+    },
+    {
+      "epoch": 0.11037364798426745,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9976445403057095e-05,
+      "loss": 0.283,
+      "step": 1796
+    },
+    {
+      "epoch": 0.11043510324483775,
+      "grad_norm": 9.0,
+      "learning_rate": 1.9976285503463593e-05,
+      "loss": 0.3986,
+      "step": 1797
+    },
+    {
+      "epoch": 0.11049655850540806,
+      "grad_norm": 10.75,
+      "learning_rate": 1.9976125063612254e-05,
+      "loss": 0.3914,
+      "step": 1798
+    },
+    {
+      "epoch": 0.11055801376597837,
+      "grad_norm": 37.5,
+      "learning_rate": 1.9975964083511768e-05,
+      "loss": 0.4547,
+      "step": 1799
+    },
+    {
+      "epoch": 0.11061946902654868,
+      "grad_norm": 14.75,
+      "learning_rate": 1.9975802563170846e-05,
+      "loss": 0.403,
+      "step": 1800
+    },
+    {
+      "epoch": 0.11068092428711898,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9975640502598243e-05,
+      "loss": 0.4122,
+      "step": 1801
+    },
+    {
+      "epoch": 0.11074237954768928,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9975477901802734e-05,
+      "loss": 0.5104,
+      "step": 1802
+    },
+    {
+      "epoch": 0.11080383480825959,
+      "grad_norm": 17.125,
+      "learning_rate": 1.997531476079312e-05,
+      "loss": 0.4515,
+      "step": 1803
+    },
+    {
+      "epoch": 0.11086529006882989,
+      "grad_norm": 78.0,
+      "learning_rate": 1.9975151079578238e-05,
+      "loss": 0.4663,
+      "step": 1804
+    },
+    {
+      "epoch": 0.1109267453294002,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9974986858166954e-05,
+      "loss": 0.4599,
+      "step": 1805
+    },
+    {
+      "epoch": 0.1109882005899705,
+      "grad_norm": 10.4375,
+      "learning_rate": 1.9974822096568157e-05,
+      "loss": 0.5938,
+      "step": 1806
+    },
+    {
+      "epoch": 0.1110496558505408,
+      "grad_norm": 58.5,
+      "learning_rate": 1.9974656794790777e-05,
+      "loss": 0.6857,
+      "step": 1807
+    },
+    {
+      "epoch": 0.1111111111111111,
+      "grad_norm": 25.25,
+      "learning_rate": 1.997449095284376e-05,
+      "loss": 0.5636,
+      "step": 1808
+    },
+    {
+      "epoch": 0.11117256637168142,
+      "grad_norm": 43.5,
+      "learning_rate": 1.9974324570736087e-05,
+      "loss": 0.7875,
+      "step": 1809
+    },
+    {
+      "epoch": 0.11123402163225173,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9974157648476768e-05,
+      "loss": 0.6296,
+      "step": 1810
+    },
+    {
+      "epoch": 0.11129547689282203,
+      "grad_norm": 73.5,
+      "learning_rate": 1.9973990186074844e-05,
+      "loss": 0.7205,
+      "step": 1811
+    },
+    {
+      "epoch": 0.11135693215339233,
+      "grad_norm": 43.25,
+      "learning_rate": 1.9973822183539384e-05,
+      "loss": 0.6143,
+      "step": 1812
+    },
+    {
+      "epoch": 0.11141838741396264,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9973653640879486e-05,
+      "loss": 0.7,
+      "step": 1813
+    },
+    {
+      "epoch": 0.11147984267453294,
+      "grad_norm": 44.25,
+      "learning_rate": 1.997348455810428e-05,
+      "loss": 0.5676,
+      "step": 1814
+    },
+    {
+      "epoch": 0.11154129793510324,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9973314935222917e-05,
+      "loss": 0.6433,
+      "step": 1815
+    },
+    {
+      "epoch": 0.11160275319567355,
+      "grad_norm": 45.0,
+      "learning_rate": 1.997314477224458e-05,
+      "loss": 0.5414,
+      "step": 1816
+    },
+    {
+      "epoch": 0.11166420845624385,
+      "grad_norm": 45.5,
+      "learning_rate": 1.9972974069178498e-05,
+      "loss": 0.4317,
+      "step": 1817
+    },
+    {
+      "epoch": 0.11172566371681415,
+      "grad_norm": 25.125,
+      "learning_rate": 1.99728028260339e-05,
+      "loss": 0.4421,
+      "step": 1818
+    },
+    {
+      "epoch": 0.11178711897738447,
+      "grad_norm": 24.625,
+      "learning_rate": 1.997263104282007e-05,
+      "loss": 0.4677,
+      "step": 1819
+    },
+    {
+      "epoch": 0.11184857423795477,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.9972458719546307e-05,
+      "loss": 0.3855,
+      "step": 1820
+    },
+    {
+      "epoch": 0.11191002949852508,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9972285856221944e-05,
+      "loss": 0.4087,
+      "step": 1821
+    },
+    {
+      "epoch": 0.11197148475909538,
+      "grad_norm": 6.40625,
+      "learning_rate": 1.997211245285634e-05,
+      "loss": 0.3288,
+      "step": 1822
+    },
+    {
+      "epoch": 0.11203294001966568,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9971938509458888e-05,
+      "loss": 0.3195,
+      "step": 1823
+    },
+    {
+      "epoch": 0.11209439528023599,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9971764026039005e-05,
+      "loss": 0.5184,
+      "step": 1824
+    },
+    {
+      "epoch": 0.11215585054080629,
+      "grad_norm": 17.125,
+      "learning_rate": 1.997158900260614e-05,
+      "loss": 0.5094,
+      "step": 1825
+    },
+    {
+      "epoch": 0.1122173058013766,
+      "grad_norm": 22.75,
+      "learning_rate": 1.9971413439169777e-05,
+      "loss": 0.5112,
+      "step": 1826
+    },
+    {
+      "epoch": 0.1122787610619469,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9971237335739416e-05,
+      "loss": 0.5052,
+      "step": 1827
+    },
+    {
+      "epoch": 0.11234021632251721,
+      "grad_norm": 18.25,
+      "learning_rate": 1.99710606923246e-05,
+      "loss": 0.5202,
+      "step": 1828
+    },
+    {
+      "epoch": 0.11240167158308752,
+      "grad_norm": 17.625,
+      "learning_rate": 1.997088350893489e-05,
+      "loss": 0.5362,
+      "step": 1829
+    },
+    {
+      "epoch": 0.11246312684365782,
+      "grad_norm": 26.75,
+      "learning_rate": 1.9970705785579882e-05,
+      "loss": 0.6664,
+      "step": 1830
+    },
+    {
+      "epoch": 0.11252458210422812,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9970527522269204e-05,
+      "loss": 0.5387,
+      "step": 1831
+    },
+    {
+      "epoch": 0.11258603736479843,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.997034871901251e-05,
+      "loss": 0.3272,
+      "step": 1832
+    },
+    {
+      "epoch": 0.11264749262536873,
+      "grad_norm": 18.625,
+      "learning_rate": 1.9970169375819475e-05,
+      "loss": 0.6959,
+      "step": 1833
+    },
+    {
+      "epoch": 0.11270894788593903,
+      "grad_norm": 17.625,
+      "learning_rate": 1.996998949269982e-05,
+      "loss": 0.6524,
+      "step": 1834
+    },
+    {
+      "epoch": 0.11277040314650934,
+      "grad_norm": 25.0,
+      "learning_rate": 1.996980906966328e-05,
+      "loss": 0.8149,
+      "step": 1835
+    },
+    {
+      "epoch": 0.11283185840707964,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.9969628106719632e-05,
+      "loss": 0.6222,
+      "step": 1836
+    },
+    {
+      "epoch": 0.11289331366764994,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.9969446603878673e-05,
+      "loss": 0.4338,
+      "step": 1837
+    },
+    {
+      "epoch": 0.11295476892822026,
+      "grad_norm": 25.75,
+      "learning_rate": 1.996926456115023e-05,
+      "loss": 0.6691,
+      "step": 1838
+    },
+    {
+      "epoch": 0.11301622418879056,
+      "grad_norm": 13.0,
+      "learning_rate": 1.9969081978544163e-05,
+      "loss": 0.5539,
+      "step": 1839
+    },
+    {
+      "epoch": 0.11307767944936087,
+      "grad_norm": 30.5,
+      "learning_rate": 1.996889885607036e-05,
+      "loss": 0.7009,
+      "step": 1840
+    },
+    {
+      "epoch": 0.11313913470993117,
+      "grad_norm": 43.5,
+      "learning_rate": 1.9968715193738738e-05,
+      "loss": 0.74,
+      "step": 1841
+    },
+    {
+      "epoch": 0.11320058997050148,
+      "grad_norm": 41.25,
+      "learning_rate": 1.9968530991559244e-05,
+      "loss": 0.592,
+      "step": 1842
+    },
+    {
+      "epoch": 0.11326204523107178,
+      "grad_norm": 39.75,
+      "learning_rate": 1.9968346249541848e-05,
+      "loss": 0.5829,
+      "step": 1843
+    },
+    {
+      "epoch": 0.11332350049164208,
+      "grad_norm": 30.875,
+      "learning_rate": 1.9968160967696565e-05,
+      "loss": 0.6194,
+      "step": 1844
+    },
+    {
+      "epoch": 0.11338495575221239,
+      "grad_norm": 35.75,
+      "learning_rate": 1.9967975146033417e-05,
+      "loss": 0.7347,
+      "step": 1845
+    },
+    {
+      "epoch": 0.11344641101278269,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.9967788784562474e-05,
+      "loss": 0.5696,
+      "step": 1846
+    },
+    {
+      "epoch": 0.113507866273353,
+      "grad_norm": 41.75,
+      "learning_rate": 1.9967601883293826e-05,
+      "loss": 0.5031,
+      "step": 1847
+    },
+    {
+      "epoch": 0.11356932153392331,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.9967414442237595e-05,
+      "loss": 0.4634,
+      "step": 1848
+    },
+    {
+      "epoch": 0.11363077679449361,
+      "grad_norm": 27.375,
+      "learning_rate": 1.9967226461403934e-05,
+      "loss": 0.5535,
+      "step": 1849
+    },
+    {
+      "epoch": 0.11369223205506392,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9967037940803016e-05,
+      "loss": 0.6167,
+      "step": 1850
+    },
+    {
+      "epoch": 0.11375368731563422,
+      "grad_norm": 18.75,
+      "learning_rate": 1.996684888044506e-05,
+      "loss": 0.5572,
+      "step": 1851
+    },
+    {
+      "epoch": 0.11381514257620452,
+      "grad_norm": 12.75,
+      "learning_rate": 1.99666592803403e-05,
+      "loss": 0.4397,
+      "step": 1852
+    },
+    {
+      "epoch": 0.11387659783677483,
+      "grad_norm": 8.1875,
+      "learning_rate": 1.9966469140499e-05,
+      "loss": 0.609,
+      "step": 1853
+    },
+    {
+      "epoch": 0.11393805309734513,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9966278460931463e-05,
+      "loss": 0.3777,
+      "step": 1854
+    },
+    {
+      "epoch": 0.11399950835791543,
+      "grad_norm": 20.125,
+      "learning_rate": 1.996608724164801e-05,
+      "loss": 0.4651,
+      "step": 1855
+    },
+    {
+      "epoch": 0.11406096361848574,
+      "grad_norm": 12.5,
+      "learning_rate": 1.9965895482659e-05,
+      "loss": 0.5072,
+      "step": 1856
+    },
+    {
+      "epoch": 0.11412241887905605,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9965703183974813e-05,
+      "loss": 0.6233,
+      "step": 1857
+    },
+    {
+      "epoch": 0.11418387413962636,
+      "grad_norm": 20.875,
+      "learning_rate": 1.9965510345605866e-05,
+      "loss": 0.6148,
+      "step": 1858
+    },
+    {
+      "epoch": 0.11424532940019666,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9965316967562606e-05,
+      "loss": 0.7652,
+      "step": 1859
+    },
+    {
+      "epoch": 0.11430678466076696,
+      "grad_norm": 42.75,
+      "learning_rate": 1.9965123049855498e-05,
+      "loss": 0.7911,
+      "step": 1860
+    },
+    {
+      "epoch": 0.11436823992133727,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9964928592495046e-05,
+      "loss": 0.7689,
+      "step": 1861
+    },
+    {
+      "epoch": 0.11442969518190757,
+      "grad_norm": 18.0,
+      "learning_rate": 1.996473359549178e-05,
+      "loss": 0.5049,
+      "step": 1862
+    },
+    {
+      "epoch": 0.11449115044247787,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9964538058856263e-05,
+      "loss": 0.6148,
+      "step": 1863
+    },
+    {
+      "epoch": 0.11455260570304818,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.996434198259908e-05,
+      "loss": 0.435,
+      "step": 1864
+    },
+    {
+      "epoch": 0.11461406096361848,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9964145366730854e-05,
+      "loss": 0.4826,
+      "step": 1865
+    },
+    {
+      "epoch": 0.11467551622418878,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9963948211262233e-05,
+      "loss": 0.5486,
+      "step": 1866
+    },
+    {
+      "epoch": 0.1147369714847591,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9963750516203887e-05,
+      "loss": 0.522,
+      "step": 1867
+    },
+    {
+      "epoch": 0.1147984267453294,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.9963552281566523e-05,
+      "loss": 0.5836,
+      "step": 1868
+    },
+    {
+      "epoch": 0.11485988200589971,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9963353507360884e-05,
+      "loss": 0.4916,
+      "step": 1869
+    },
+    {
+      "epoch": 0.11492133726647001,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9963154193597728e-05,
+      "loss": 0.5142,
+      "step": 1870
+    },
+    {
+      "epoch": 0.11498279252704031,
+      "grad_norm": 14.5,
+      "learning_rate": 1.996295434028785e-05,
+      "loss": 0.3301,
+      "step": 1871
+    },
+    {
+      "epoch": 0.11504424778761062,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.9962753947442072e-05,
+      "loss": 0.3774,
+      "step": 1872
+    },
+    {
+      "epoch": 0.11510570304818092,
+      "grad_norm": 29.875,
+      "learning_rate": 1.996255301507125e-05,
+      "loss": 0.6848,
+      "step": 1873
+    },
+    {
+      "epoch": 0.11516715830875122,
+      "grad_norm": 5.9375,
+      "learning_rate": 1.996235154318626e-05,
+      "loss": 0.717,
+      "step": 1874
+    },
+    {
+      "epoch": 0.11522861356932153,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.9962149531798015e-05,
+      "loss": 0.6016,
+      "step": 1875
+    },
+    {
+      "epoch": 0.11529006882989185,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9961946980917457e-05,
+      "loss": 0.5468,
+      "step": 1876
+    },
+    {
+      "epoch": 0.11535152409046215,
+      "grad_norm": 33.75,
+      "learning_rate": 1.9961743890555553e-05,
+      "loss": 0.4646,
+      "step": 1877
+    },
+    {
+      "epoch": 0.11541297935103245,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.99615402607233e-05,
+      "loss": 0.4641,
+      "step": 1878
+    },
+    {
+      "epoch": 0.11547443461160276,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9961336091431728e-05,
+      "loss": 0.5797,
+      "step": 1879
+    },
+    {
+      "epoch": 0.11553588987217306,
+      "grad_norm": 11.125,
+      "learning_rate": 1.996113138269189e-05,
+      "loss": 0.3914,
+      "step": 1880
+    },
+    {
+      "epoch": 0.11559734513274336,
+      "grad_norm": 40.5,
+      "learning_rate": 1.9960926134514875e-05,
+      "loss": 0.6931,
+      "step": 1881
+    },
+    {
+      "epoch": 0.11565880039331367,
+      "grad_norm": 53.0,
+      "learning_rate": 1.9960720346911798e-05,
+      "loss": 0.7989,
+      "step": 1882
+    },
+    {
+      "epoch": 0.11572025565388397,
+      "grad_norm": 48.25,
+      "learning_rate": 1.99605140198938e-05,
+      "loss": 0.8516,
+      "step": 1883
+    },
+    {
+      "epoch": 0.11578171091445427,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.996030715347206e-05,
+      "loss": 0.5719,
+      "step": 1884
+    },
+    {
+      "epoch": 0.11584316617502458,
+      "grad_norm": 12.375,
+      "learning_rate": 1.9960099747657774e-05,
+      "loss": 0.6435,
+      "step": 1885
+    },
+    {
+      "epoch": 0.11590462143559489,
+      "grad_norm": 12.75,
+      "learning_rate": 1.995989180246218e-05,
+      "loss": 0.509,
+      "step": 1886
+    },
+    {
+      "epoch": 0.1159660766961652,
+      "grad_norm": 22.25,
+      "learning_rate": 1.9959683317896535e-05,
+      "loss": 0.5274,
+      "step": 1887
+    },
+    {
+      "epoch": 0.1160275319567355,
+      "grad_norm": 47.75,
+      "learning_rate": 1.995947429397213e-05,
+      "loss": 0.7349,
+      "step": 1888
+    },
+    {
+      "epoch": 0.1160889872173058,
+      "grad_norm": 33.0,
+      "learning_rate": 1.995926473070029e-05,
+      "loss": 0.4721,
+      "step": 1889
+    },
+    {
+      "epoch": 0.1161504424778761,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9959054628092353e-05,
+      "loss": 0.5948,
+      "step": 1890
+    },
+    {
+      "epoch": 0.11621189773844641,
+      "grad_norm": 46.5,
+      "learning_rate": 1.9958843986159705e-05,
+      "loss": 0.5243,
+      "step": 1891
+    },
+    {
+      "epoch": 0.11627335299901671,
+      "grad_norm": 65.5,
+      "learning_rate": 1.9958632804913753e-05,
+      "loss": 0.5539,
+      "step": 1892
+    },
+    {
+      "epoch": 0.11633480825958702,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9958421084365928e-05,
+      "loss": 0.4856,
+      "step": 1893
+    },
+    {
+      "epoch": 0.11639626352015732,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9958208824527702e-05,
+      "loss": 0.5976,
+      "step": 1894
+    },
+    {
+      "epoch": 0.11645771878072764,
+      "grad_norm": 63.25,
+      "learning_rate": 1.9957996025410565e-05,
+      "loss": 0.4858,
+      "step": 1895
+    },
+    {
+      "epoch": 0.11651917404129794,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9957782687026046e-05,
+      "loss": 0.5822,
+      "step": 1896
+    },
+    {
+      "epoch": 0.11658062930186824,
+      "grad_norm": 46.25,
+      "learning_rate": 1.9957568809385693e-05,
+      "loss": 0.4152,
+      "step": 1897
+    },
+    {
+      "epoch": 0.11664208456243855,
+      "grad_norm": 45.25,
+      "learning_rate": 1.9957354392501092e-05,
+      "loss": 0.4972,
+      "step": 1898
+    },
+    {
+      "epoch": 0.11670353982300885,
+      "grad_norm": 44.75,
+      "learning_rate": 1.995713943638385e-05,
+      "loss": 0.5448,
+      "step": 1899
+    },
+    {
+      "epoch": 0.11676499508357915,
+      "grad_norm": 33.25,
+      "learning_rate": 1.9956923941045613e-05,
+      "loss": 0.4045,
+      "step": 1900
+    },
+    {
+      "epoch": 0.11682645034414946,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9956707906498046e-05,
+      "loss": 0.4992,
+      "step": 1901
+    },
+    {
+      "epoch": 0.11688790560471976,
+      "grad_norm": 20.75,
+      "learning_rate": 1.995649133275285e-05,
+      "loss": 0.4209,
+      "step": 1902
+    },
+    {
+      "epoch": 0.11694936086529006,
+      "grad_norm": 35.0,
+      "learning_rate": 1.995627421982176e-05,
+      "loss": 0.4744,
+      "step": 1903
+    },
+    {
+      "epoch": 0.11701081612586037,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.995605656771652e-05,
+      "loss": 0.4039,
+      "step": 1904
+    },
+    {
+      "epoch": 0.11707227138643068,
+      "grad_norm": 23.0,
+      "learning_rate": 1.9955838376448933e-05,
+      "loss": 0.4875,
+      "step": 1905
+    },
+    {
+      "epoch": 0.11713372664700099,
+      "grad_norm": 23.375,
+      "learning_rate": 1.99556196460308e-05,
+      "loss": 0.6234,
+      "step": 1906
+    },
+    {
+      "epoch": 0.11719518190757129,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9955400376473977e-05,
+      "loss": 0.5063,
+      "step": 1907
+    },
+    {
+      "epoch": 0.1172566371681416,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9955180567790332e-05,
+      "loss": 0.6176,
+      "step": 1908
+    },
+    {
+      "epoch": 0.1173180924287119,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.995496021999177e-05,
+      "loss": 0.4577,
+      "step": 1909
+    },
+    {
+      "epoch": 0.1173795476892822,
+      "grad_norm": 24.625,
+      "learning_rate": 1.995473933309023e-05,
+      "loss": 0.4337,
+      "step": 1910
+    },
+    {
+      "epoch": 0.1174410029498525,
+      "grad_norm": 7.25,
+      "learning_rate": 1.9954517907097663e-05,
+      "loss": 0.4849,
+      "step": 1911
+    },
+    {
+      "epoch": 0.11750245821042281,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9954295942026065e-05,
+      "loss": 0.585,
+      "step": 1912
+    },
+    {
+      "epoch": 0.11756391347099311,
+      "grad_norm": 45.25,
+      "learning_rate": 1.9954073437887457e-05,
+      "loss": 0.5748,
+      "step": 1913
+    },
+    {
+      "epoch": 0.11762536873156341,
+      "grad_norm": 38.25,
+      "learning_rate": 1.995385039469389e-05,
+      "loss": 0.4938,
+      "step": 1914
+    },
+    {
+      "epoch": 0.11768682399213373,
+      "grad_norm": 29.875,
+      "learning_rate": 1.995362681245744e-05,
+      "loss": 0.7121,
+      "step": 1915
+    },
+    {
+      "epoch": 0.11774827925270404,
+      "grad_norm": 34.5,
+      "learning_rate": 1.9953402691190218e-05,
+      "loss": 0.5834,
+      "step": 1916
+    },
+    {
+      "epoch": 0.11780973451327434,
+      "grad_norm": 9.125,
+      "learning_rate": 1.9953178030904357e-05,
+      "loss": 0.5756,
+      "step": 1917
+    },
+    {
+      "epoch": 0.11787118977384464,
+      "grad_norm": 8.125,
+      "learning_rate": 1.9952952831612027e-05,
+      "loss": 0.531,
+      "step": 1918
+    },
+    {
+      "epoch": 0.11793264503441495,
+      "grad_norm": 20.25,
+      "learning_rate": 1.9952727093325423e-05,
+      "loss": 0.3872,
+      "step": 1919
+    },
+    {
+      "epoch": 0.11799410029498525,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9952500816056765e-05,
+      "loss": 0.3882,
+      "step": 1920
+    },
+    {
+      "epoch": 0.11805555555555555,
+      "grad_norm": 5.21875,
+      "learning_rate": 1.9952273999818312e-05,
+      "loss": 0.527,
+      "step": 1921
+    },
+    {
+      "epoch": 0.11811701081612586,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.9952046644622346e-05,
+      "loss": 0.4709,
+      "step": 1922
+    },
+    {
+      "epoch": 0.11817846607669616,
+      "grad_norm": 27.5,
+      "learning_rate": 1.995181875048118e-05,
+      "loss": 0.3579,
+      "step": 1923
+    },
+    {
+      "epoch": 0.11823992133726648,
+      "grad_norm": 30.875,
+      "learning_rate": 1.9951590317407152e-05,
+      "loss": 0.4172,
+      "step": 1924
+    },
+    {
+      "epoch": 0.11830137659783678,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9951361345412636e-05,
+      "loss": 0.6696,
+      "step": 1925
+    },
+    {
+      "epoch": 0.11836283185840708,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9951131834510034e-05,
+      "loss": 0.3767,
+      "step": 1926
+    },
+    {
+      "epoch": 0.11842428711897739,
+      "grad_norm": 7.96875,
+      "learning_rate": 1.9950901784711765e-05,
+      "loss": 0.4291,
+      "step": 1927
+    },
+    {
+      "epoch": 0.11848574237954769,
+      "grad_norm": 69.0,
+      "learning_rate": 1.9950671196030298e-05,
+      "loss": 0.4201,
+      "step": 1928
+    },
+    {
+      "epoch": 0.118547197640118,
+      "grad_norm": 28.75,
+      "learning_rate": 1.9950440068478115e-05,
+      "loss": 0.7555,
+      "step": 1929
+    },
+    {
+      "epoch": 0.1186086529006883,
+      "grad_norm": 58.75,
+      "learning_rate": 1.9950208402067735e-05,
+      "loss": 0.7644,
+      "step": 1930
+    },
+    {
+      "epoch": 0.1186701081612586,
+      "grad_norm": 47.25,
+      "learning_rate": 1.99499761968117e-05,
+      "loss": 0.7416,
+      "step": 1931
+    },
+    {
+      "epoch": 0.1187315634218289,
+      "grad_norm": 38.25,
+      "learning_rate": 1.994974345272259e-05,
+      "loss": 0.498,
+      "step": 1932
+    },
+    {
+      "epoch": 0.1187930186823992,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9949510169813006e-05,
+      "loss": 0.6546,
+      "step": 1933
+    },
+    {
+      "epoch": 0.11885447394296952,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9949276348095576e-05,
+      "loss": 0.6238,
+      "step": 1934
+    },
+    {
+      "epoch": 0.11891592920353983,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9949041987582975e-05,
+      "loss": 0.5189,
+      "step": 1935
+    },
+    {
+      "epoch": 0.11897738446411013,
+      "grad_norm": 23.75,
+      "learning_rate": 1.9948807088287884e-05,
+      "loss": 0.4802,
+      "step": 1936
+    },
+    {
+      "epoch": 0.11903883972468043,
+      "grad_norm": 41.25,
+      "learning_rate": 1.9948571650223027e-05,
+      "loss": 0.6451,
+      "step": 1937
+    },
+    {
+      "epoch": 0.11910029498525074,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9948335673401157e-05,
+      "loss": 0.6864,
+      "step": 1938
+    },
+    {
+      "epoch": 0.11916175024582104,
+      "grad_norm": 38.0,
+      "learning_rate": 1.994809915783505e-05,
+      "loss": 0.5372,
+      "step": 1939
+    },
+    {
+      "epoch": 0.11922320550639134,
+      "grad_norm": 60.0,
+      "learning_rate": 1.994786210353751e-05,
+      "loss": 0.5996,
+      "step": 1940
+    },
+    {
+      "epoch": 0.11928466076696165,
+      "grad_norm": 35.75,
+      "learning_rate": 1.9947624510521385e-05,
+      "loss": 0.4957,
+      "step": 1941
+    },
+    {
+      "epoch": 0.11934611602753195,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9947386378799534e-05,
+      "loss": 0.4601,
+      "step": 1942
+    },
+    {
+      "epoch": 0.11940757128810227,
+      "grad_norm": 12.4375,
+      "learning_rate": 1.9947147708384858e-05,
+      "loss": 0.3769,
+      "step": 1943
+    },
+    {
+      "epoch": 0.11946902654867257,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.9946908499290275e-05,
+      "loss": 0.4424,
+      "step": 1944
+    },
+    {
+      "epoch": 0.11953048180924288,
+      "grad_norm": 21.0,
+      "learning_rate": 1.9946668751528745e-05,
+      "loss": 0.501,
+      "step": 1945
+    },
+    {
+      "epoch": 0.11959193706981318,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9946428465113244e-05,
+      "loss": 0.4752,
+      "step": 1946
+    },
+    {
+      "epoch": 0.11965339233038348,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9946187640056794e-05,
+      "loss": 0.5517,
+      "step": 1947
+    },
+    {
+      "epoch": 0.11971484759095379,
+      "grad_norm": 23.625,
+      "learning_rate": 1.9945946276372435e-05,
+      "loss": 0.4254,
+      "step": 1948
+    },
+    {
+      "epoch": 0.11977630285152409,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.994570437407323e-05,
+      "loss": 0.463,
+      "step": 1949
+    },
+    {
+      "epoch": 0.11983775811209439,
+      "grad_norm": 5.1875,
+      "learning_rate": 1.994546193317229e-05,
+      "loss": 0.4918,
+      "step": 1950
+    },
+    {
+      "epoch": 0.1198992133726647,
+      "grad_norm": 12.625,
+      "learning_rate": 1.9945218953682736e-05,
+      "loss": 0.3255,
+      "step": 1951
+    },
+    {
+      "epoch": 0.119960668633235,
+      "grad_norm": 20.0,
+      "learning_rate": 1.994497543561773e-05,
+      "loss": 0.2954,
+      "step": 1952
+    },
+    {
+      "epoch": 0.12002212389380532,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9944731378990452e-05,
+      "loss": 0.6378,
+      "step": 1953
+    },
+    {
+      "epoch": 0.12008357915437562,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9944486783814135e-05,
+      "loss": 0.526,
+      "step": 1954
+    },
+    {
+      "epoch": 0.12014503441494592,
+      "grad_norm": 15.25,
+      "learning_rate": 1.994424165010201e-05,
+      "loss": 0.3548,
+      "step": 1955
+    },
+    {
+      "epoch": 0.12020648967551623,
+      "grad_norm": 10.8125,
+      "learning_rate": 1.9943995977867358e-05,
+      "loss": 0.5702,
+      "step": 1956
+    },
+    {
+      "epoch": 0.12026794493608653,
+      "grad_norm": 9.375,
+      "learning_rate": 1.994374976712348e-05,
+      "loss": 0.4634,
+      "step": 1957
+    },
+    {
+      "epoch": 0.12032940019665683,
+      "grad_norm": 8.1875,
+      "learning_rate": 1.9943503017883716e-05,
+      "loss": 0.5298,
+      "step": 1958
+    },
+    {
+      "epoch": 0.12039085545722714,
+      "grad_norm": 3.890625,
+      "learning_rate": 1.994325573016142e-05,
+      "loss": 0.3792,
+      "step": 1959
+    },
+    {
+      "epoch": 0.12045231071779744,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.994300790396999e-05,
+      "loss": 0.5573,
+      "step": 1960
+    },
+    {
+      "epoch": 0.12051376597836774,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9942759539322845e-05,
+      "loss": 0.5288,
+      "step": 1961
+    },
+    {
+      "epoch": 0.12057522123893805,
+      "grad_norm": 13.375,
+      "learning_rate": 1.9942510636233435e-05,
+      "loss": 0.4077,
+      "step": 1962
+    },
+    {
+      "epoch": 0.12063667649950836,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9942261194715236e-05,
+      "loss": 0.6043,
+      "step": 1963
+    },
+    {
+      "epoch": 0.12069813176007867,
+      "grad_norm": 15.75,
+      "learning_rate": 1.994201121478176e-05,
+      "loss": 0.4062,
+      "step": 1964
+    },
+    {
+      "epoch": 0.12075958702064897,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.9941760696446546e-05,
+      "loss": 0.3833,
+      "step": 1965
+    },
+    {
+      "epoch": 0.12082104228121927,
+      "grad_norm": 40.0,
+      "learning_rate": 1.9941509639723155e-05,
+      "loss": 0.65,
+      "step": 1966
+    },
+    {
+      "epoch": 0.12088249754178958,
+      "grad_norm": 51.25,
+      "learning_rate": 1.994125804462519e-05,
+      "loss": 0.4981,
+      "step": 1967
+    },
+    {
+      "epoch": 0.12094395280235988,
+      "grad_norm": 24.625,
+      "learning_rate": 1.994100591116627e-05,
+      "loss": 0.476,
+      "step": 1968
+    },
+    {
+      "epoch": 0.12100540806293018,
+      "grad_norm": 60.5,
+      "learning_rate": 1.9940753239360047e-05,
+      "loss": 0.6537,
+      "step": 1969
+    },
+    {
+      "epoch": 0.12106686332350049,
+      "grad_norm": 33.5,
+      "learning_rate": 1.994050002922021e-05,
+      "loss": 0.4306,
+      "step": 1970
+    },
+    {
+      "epoch": 0.12112831858407079,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9940246280760473e-05,
+      "loss": 0.5095,
+      "step": 1971
+    },
+    {
+      "epoch": 0.12118977384464111,
+      "grad_norm": 58.25,
+      "learning_rate": 1.993999199399457e-05,
+      "loss": 0.506,
+      "step": 1972
+    },
+    {
+      "epoch": 0.12125122910521141,
+      "grad_norm": 36.5,
+      "learning_rate": 1.993973716893628e-05,
+      "loss": 0.4068,
+      "step": 1973
+    },
+    {
+      "epoch": 0.12131268436578171,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9939481805599398e-05,
+      "loss": 0.4259,
+      "step": 1974
+    },
+    {
+      "epoch": 0.12137413962635202,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9939225903997748e-05,
+      "loss": 0.4216,
+      "step": 1975
+    },
+    {
+      "epoch": 0.12143559488692232,
+      "grad_norm": 11.875,
+      "learning_rate": 1.99389694641452e-05,
+      "loss": 0.4085,
+      "step": 1976
+    },
+    {
+      "epoch": 0.12149705014749262,
+      "grad_norm": 48.25,
+      "learning_rate": 1.9938712486055632e-05,
+      "loss": 0.5013,
+      "step": 1977
+    },
+    {
+      "epoch": 0.12155850540806293,
+      "grad_norm": 22.5,
+      "learning_rate": 1.993845496974297e-05,
+      "loss": 0.418,
+      "step": 1978
+    },
+    {
+      "epoch": 0.12161996066863323,
+      "grad_norm": 7.65625,
+      "learning_rate": 1.9938196915221147e-05,
+      "loss": 0.3448,
+      "step": 1979
+    },
+    {
+      "epoch": 0.12168141592920353,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9937938322504144e-05,
+      "loss": 0.4219,
+      "step": 1980
+    },
+    {
+      "epoch": 0.12174287118977384,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9937679191605964e-05,
+      "loss": 0.3498,
+      "step": 1981
+    },
+    {
+      "epoch": 0.12180432645034416,
+      "grad_norm": 5.3125,
+      "learning_rate": 1.9937419522540644e-05,
+      "loss": 0.3105,
+      "step": 1982
+    },
+    {
+      "epoch": 0.12186578171091446,
+      "grad_norm": 43.5,
+      "learning_rate": 1.9937159315322242e-05,
+      "loss": 0.4721,
+      "step": 1983
+    },
+    {
+      "epoch": 0.12192723697148476,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.993689856996485e-05,
+      "loss": 0.3457,
+      "step": 1984
+    },
+    {
+      "epoch": 0.12198869223205507,
+      "grad_norm": 46.0,
+      "learning_rate": 1.993663728648259e-05,
+      "loss": 0.6441,
+      "step": 1985
+    },
+    {
+      "epoch": 0.12205014749262537,
+      "grad_norm": 51.75,
+      "learning_rate": 1.9936375464889608e-05,
+      "loss": 0.6362,
+      "step": 1986
+    },
+    {
+      "epoch": 0.12211160275319567,
+      "grad_norm": 27.5,
+      "learning_rate": 1.9936113105200085e-05,
+      "loss": 0.594,
+      "step": 1987
+    },
+    {
+      "epoch": 0.12217305801376598,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.993585020742823e-05,
+      "loss": 0.767,
+      "step": 1988
+    },
+    {
+      "epoch": 0.12223451327433628,
+      "grad_norm": 64.5,
+      "learning_rate": 1.993558677158828e-05,
+      "loss": 0.4731,
+      "step": 1989
+    },
+    {
+      "epoch": 0.12229596853490658,
+      "grad_norm": 37.25,
+      "learning_rate": 1.99353227976945e-05,
+      "loss": 0.6903,
+      "step": 1990
+    },
+    {
+      "epoch": 0.1223574237954769,
+      "grad_norm": 23.0,
+      "learning_rate": 1.9935058285761185e-05,
+      "loss": 0.3168,
+      "step": 1991
+    },
+    {
+      "epoch": 0.1224188790560472,
+      "grad_norm": 25.625,
+      "learning_rate": 1.993479323580266e-05,
+      "loss": 0.527,
+      "step": 1992
+    },
+    {
+      "epoch": 0.1224803343166175,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9934527647833276e-05,
+      "loss": 0.8246,
+      "step": 1993
+    },
+    {
+      "epoch": 0.12254178957718781,
+      "grad_norm": 30.5,
+      "learning_rate": 1.993426152186742e-05,
+      "loss": 0.7521,
+      "step": 1994
+    },
+    {
+      "epoch": 0.12260324483775811,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9933994857919505e-05,
+      "loss": 0.5836,
+      "step": 1995
+    },
+    {
+      "epoch": 0.12266470009832842,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9933727656003964e-05,
+      "loss": 0.5834,
+      "step": 1996
+    },
+    {
+      "epoch": 0.12272615535889872,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9933459916135277e-05,
+      "loss": 0.6181,
+      "step": 1997
+    },
+    {
+      "epoch": 0.12278761061946902,
+      "grad_norm": 16.375,
+      "learning_rate": 1.9933191638327935e-05,
+      "loss": 0.5982,
+      "step": 1998
+    },
+    {
+      "epoch": 0.12284906588003933,
+      "grad_norm": 22.625,
+      "learning_rate": 1.993292282259647e-05,
+      "loss": 0.5736,
+      "step": 1999
+    },
+    {
+      "epoch": 0.12291052114060963,
+      "grad_norm": 16.875,
+      "learning_rate": 1.9932653468955442e-05,
+      "loss": 0.4823,
+      "step": 2000
+    },
+    {
+      "epoch": 0.12297197640117995,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9932383577419432e-05,
+      "loss": 0.6774,
+      "step": 2001
+    },
+    {
+      "epoch": 0.12303343166175025,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9932113148003057e-05,
+      "loss": 0.5491,
+      "step": 2002
+    },
+    {
+      "epoch": 0.12309488692232055,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9931842180720967e-05,
+      "loss": 0.5375,
+      "step": 2003
+    },
+    {
+      "epoch": 0.12315634218289086,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9931570675587833e-05,
+      "loss": 0.4592,
+      "step": 2004
+    },
+    {
+      "epoch": 0.12321779744346116,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9931298632618355e-05,
+      "loss": 0.4537,
+      "step": 2005
+    },
+    {
+      "epoch": 0.12327925270403146,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.993102605182727e-05,
+      "loss": 0.4392,
+      "step": 2006
+    },
+    {
+      "epoch": 0.12334070796460177,
+      "grad_norm": 19.625,
+      "learning_rate": 1.9930752933229335e-05,
+      "loss": 0.5462,
+      "step": 2007
+    },
+    {
+      "epoch": 0.12340216322517207,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.9930479276839347e-05,
+      "loss": 0.4,
+      "step": 2008
+    },
+    {
+      "epoch": 0.12346361848574237,
+      "grad_norm": 24.0,
+      "learning_rate": 1.9930205082672115e-05,
+      "loss": 0.513,
+      "step": 2009
+    },
+    {
+      "epoch": 0.12352507374631268,
+      "grad_norm": 17.0,
+      "learning_rate": 1.99299303507425e-05,
+      "loss": 0.4483,
+      "step": 2010
+    },
+    {
+      "epoch": 0.123586529006883,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.992965508106537e-05,
+      "loss": 0.4814,
+      "step": 2011
+    },
+    {
+      "epoch": 0.1236479842674533,
+      "grad_norm": 31.5,
+      "learning_rate": 1.992937927365564e-05,
+      "loss": 0.507,
+      "step": 2012
+    },
+    {
+      "epoch": 0.1237094395280236,
+      "grad_norm": 22.75,
+      "learning_rate": 1.992910292852824e-05,
+      "loss": 0.4852,
+      "step": 2013
+    },
+    {
+      "epoch": 0.1237708947885939,
+      "grad_norm": 11.5,
+      "learning_rate": 1.9928826045698138e-05,
+      "loss": 0.4664,
+      "step": 2014
+    },
+    {
+      "epoch": 0.12383235004916421,
+      "grad_norm": 10.0,
+      "learning_rate": 1.992854862518033e-05,
+      "loss": 0.4074,
+      "step": 2015
+    },
+    {
+      "epoch": 0.12389380530973451,
+      "grad_norm": 15.75,
+      "learning_rate": 1.9928270666989835e-05,
+      "loss": 0.4631,
+      "step": 2016
+    },
+    {
+      "epoch": 0.12395526057030481,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9927992171141707e-05,
+      "loss": 0.7821,
+      "step": 2017
+    },
+    {
+      "epoch": 0.12401671583087512,
+      "grad_norm": 32.5,
+      "learning_rate": 1.992771313765103e-05,
+      "loss": 0.5627,
+      "step": 2018
+    },
+    {
+      "epoch": 0.12407817109144542,
+      "grad_norm": 50.25,
+      "learning_rate": 1.9927433566532913e-05,
+      "loss": 0.4954,
+      "step": 2019
+    },
+    {
+      "epoch": 0.12413962635201574,
+      "grad_norm": 17.25,
+      "learning_rate": 1.99271534578025e-05,
+      "loss": 0.7139,
+      "step": 2020
+    },
+    {
+      "epoch": 0.12420108161258604,
+      "grad_norm": 27.375,
+      "learning_rate": 1.9926872811474952e-05,
+      "loss": 0.7874,
+      "step": 2021
+    },
+    {
+      "epoch": 0.12426253687315635,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.9926591627565476e-05,
+      "loss": 0.4781,
+      "step": 2022
+    },
+    {
+      "epoch": 0.12432399213372665,
+      "grad_norm": 7.4375,
+      "learning_rate": 1.992630990608929e-05,
+      "loss": 0.3339,
+      "step": 2023
+    },
+    {
+      "epoch": 0.12438544739429695,
+      "grad_norm": 56.0,
+      "learning_rate": 1.9926027647061658e-05,
+      "loss": 0.668,
+      "step": 2024
+    },
+    {
+      "epoch": 0.12444690265486726,
+      "grad_norm": 28.375,
+      "learning_rate": 1.9925744850497866e-05,
+      "loss": 0.7087,
+      "step": 2025
+    },
+    {
+      "epoch": 0.12450835791543756,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9925461516413224e-05,
+      "loss": 0.5799,
+      "step": 2026
+    },
+    {
+      "epoch": 0.12456981317600786,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9925177644823073e-05,
+      "loss": 0.5245,
+      "step": 2027
+    },
+    {
+      "epoch": 0.12463126843657817,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9924893235742795e-05,
+      "loss": 0.4386,
+      "step": 2028
+    },
+    {
+      "epoch": 0.12469272369714847,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9924608289187786e-05,
+      "loss": 0.535,
+      "step": 2029
+    },
+    {
+      "epoch": 0.12475417895771879,
+      "grad_norm": 10.125,
+      "learning_rate": 1.992432280517348e-05,
+      "loss": 0.6802,
+      "step": 2030
+    },
+    {
+      "epoch": 0.12481563421828909,
+      "grad_norm": 16.5,
+      "learning_rate": 1.992403678371533e-05,
+      "loss": 0.3988,
+      "step": 2031
+    },
+    {
+      "epoch": 0.1248770894788594,
+      "grad_norm": 11.625,
+      "learning_rate": 1.9923750224828833e-05,
+      "loss": 0.4574,
+      "step": 2032
+    },
+    {
+      "epoch": 0.1249385447394297,
+      "grad_norm": 11.1875,
+      "learning_rate": 1.9923463128529507e-05,
+      "loss": 0.5725,
+      "step": 2033
+    },
+    {
+      "epoch": 0.125,
+      "grad_norm": 19.875,
+      "learning_rate": 1.9923175494832895e-05,
+      "loss": 0.6356,
+      "step": 2034
+    },
+    {
+      "epoch": 0.1250614552605703,
+      "grad_norm": 23.0,
+      "learning_rate": 1.992288732375458e-05,
+      "loss": 0.4285,
+      "step": 2035
+    },
+    {
+      "epoch": 0.1251229105211406,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9922598615310157e-05,
+      "loss": 0.5139,
+      "step": 2036
+    },
+    {
+      "epoch": 0.1251843657817109,
+      "grad_norm": 32.5,
+      "learning_rate": 1.9922309369515272e-05,
+      "loss": 0.528,
+      "step": 2037
+    },
+    {
+      "epoch": 0.1252458210422812,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9922019586385587e-05,
+      "loss": 0.4491,
+      "step": 2038
+    },
+    {
+      "epoch": 0.12530727630285152,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.992172926593679e-05,
+      "loss": 0.4352,
+      "step": 2039
+    },
+    {
+      "epoch": 0.12536873156342182,
+      "grad_norm": 5.28125,
+      "learning_rate": 1.99214384081846e-05,
+      "loss": 0.3763,
+      "step": 2040
+    },
+    {
+      "epoch": 0.12543018682399212,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9921147013144782e-05,
+      "loss": 0.6106,
+      "step": 2041
+    },
+    {
+      "epoch": 0.12549164208456243,
+      "grad_norm": 38.0,
+      "learning_rate": 1.99208550808331e-05,
+      "loss": 0.6626,
+      "step": 2042
+    },
+    {
+      "epoch": 0.12555309734513273,
+      "grad_norm": 27.5,
+      "learning_rate": 1.9920562611265378e-05,
+      "loss": 0.5614,
+      "step": 2043
+    },
+    {
+      "epoch": 0.12561455260570306,
+      "grad_norm": 32.5,
+      "learning_rate": 1.9920269604457444e-05,
+      "loss": 0.7316,
+      "step": 2044
+    },
+    {
+      "epoch": 0.12567600786627336,
+      "grad_norm": 36.25,
+      "learning_rate": 1.991997606042517e-05,
+      "loss": 0.5965,
+      "step": 2045
+    },
+    {
+      "epoch": 0.12573746312684367,
+      "grad_norm": 17.375,
+      "learning_rate": 1.9919681979184452e-05,
+      "loss": 0.484,
+      "step": 2046
+    },
+    {
+      "epoch": 0.12579891838741397,
+      "grad_norm": 62.5,
+      "learning_rate": 1.9919387360751216e-05,
+      "loss": 0.729,
+      "step": 2047
+    },
+    {
+      "epoch": 0.12586037364798427,
+      "grad_norm": 21.0,
+      "learning_rate": 1.9919092205141414e-05,
+      "loss": 0.4672,
+      "step": 2048
+    },
+    {
+      "epoch": 0.12592182890855458,
+      "grad_norm": 53.25,
+      "learning_rate": 1.9918796512371035e-05,
+      "loss": 0.8433,
+      "step": 2049
+    },
+    {
+      "epoch": 0.12598328416912488,
+      "grad_norm": 24.0,
+      "learning_rate": 1.991850028245609e-05,
+      "loss": 0.5589,
+      "step": 2050
+    },
+    {
+      "epoch": 0.12604473942969519,
+      "grad_norm": 27.625,
+      "learning_rate": 1.9918203515412616e-05,
+      "loss": 0.5541,
+      "step": 2051
+    },
+    {
+      "epoch": 0.1261061946902655,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9917906211256696e-05,
+      "loss": 0.6042,
+      "step": 2052
+    },
+    {
+      "epoch": 0.1261676499508358,
+      "grad_norm": 24.0,
+      "learning_rate": 1.9917608370004417e-05,
+      "loss": 0.5591,
+      "step": 2053
+    },
+    {
+      "epoch": 0.1262291052114061,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9917309991671915e-05,
+      "loss": 0.4642,
+      "step": 2054
+    },
+    {
+      "epoch": 0.1262905604719764,
+      "grad_norm": 53.25,
+      "learning_rate": 1.991701107627535e-05,
+      "loss": 0.6261,
+      "step": 2055
+    },
+    {
+      "epoch": 0.1263520157325467,
+      "grad_norm": 12.875,
+      "learning_rate": 1.9916711623830904e-05,
+      "loss": 0.5418,
+      "step": 2056
+    },
+    {
+      "epoch": 0.126413470993117,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.99164116343548e-05,
+      "loss": 0.4516,
+      "step": 2057
+    },
+    {
+      "epoch": 0.1264749262536873,
+      "grad_norm": 16.5,
+      "learning_rate": 1.991611110786328e-05,
+      "loss": 0.5993,
+      "step": 2058
+    },
+    {
+      "epoch": 0.1265363815142576,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9915810044372618e-05,
+      "loss": 0.4228,
+      "step": 2059
+    },
+    {
+      "epoch": 0.12659783677482792,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.991550844389912e-05,
+      "loss": 0.7607,
+      "step": 2060
+    },
+    {
+      "epoch": 0.12665929203539822,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9915206306459117e-05,
+      "loss": 0.4728,
+      "step": 2061
+    },
+    {
+      "epoch": 0.12672074729596852,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9914903632068975e-05,
+      "loss": 0.3913,
+      "step": 2062
+    },
+    {
+      "epoch": 0.12678220255653885,
+      "grad_norm": 11.0625,
+      "learning_rate": 1.991460042074508e-05,
+      "loss": 0.6293,
+      "step": 2063
+    },
+    {
+      "epoch": 0.12684365781710916,
+      "grad_norm": 29.125,
+      "learning_rate": 1.9914296672503854e-05,
+      "loss": 0.4878,
+      "step": 2064
+    },
+    {
+      "epoch": 0.12690511307767946,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9913992387361747e-05,
+      "loss": 0.4785,
+      "step": 2065
+    },
+    {
+      "epoch": 0.12696656833824976,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9913687565335237e-05,
+      "loss": 0.5729,
+      "step": 2066
+    },
+    {
+      "epoch": 0.12702802359882007,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9913382206440834e-05,
+      "loss": 0.5256,
+      "step": 2067
+    },
+    {
+      "epoch": 0.12708947885939037,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9913076310695068e-05,
+      "loss": 0.6305,
+      "step": 2068
+    },
+    {
+      "epoch": 0.12715093411996067,
+      "grad_norm": 24.5,
+      "learning_rate": 1.9912769878114508e-05,
+      "loss": 0.5512,
+      "step": 2069
+    },
+    {
+      "epoch": 0.12721238938053098,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9912462908715752e-05,
+      "loss": 0.5863,
+      "step": 2070
+    },
+    {
+      "epoch": 0.12727384464110128,
+      "grad_norm": 21.0,
+      "learning_rate": 1.991215540251542e-05,
+      "loss": 0.4194,
+      "step": 2071
+    },
+    {
+      "epoch": 0.12733529990167158,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9911847359530165e-05,
+      "loss": 0.5458,
+      "step": 2072
+    },
+    {
+      "epoch": 0.1273967551622419,
+      "grad_norm": 36.75,
+      "learning_rate": 1.991153877977667e-05,
+      "loss": 0.4815,
+      "step": 2073
+    },
+    {
+      "epoch": 0.1274582104228122,
+      "grad_norm": 34.25,
+      "learning_rate": 1.991122966327164e-05,
+      "loss": 0.5823,
+      "step": 2074
+    },
+    {
+      "epoch": 0.1275196656833825,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9910920010031825e-05,
+      "loss": 0.4819,
+      "step": 2075
+    },
+    {
+      "epoch": 0.1275811209439528,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.9910609820073986e-05,
+      "loss": 0.3158,
+      "step": 2076
+    },
+    {
+      "epoch": 0.1276425762045231,
+      "grad_norm": 23.875,
+      "learning_rate": 1.991029909341493e-05,
+      "loss": 0.4087,
+      "step": 2077
+    },
+    {
+      "epoch": 0.1277040314650934,
+      "grad_norm": 26.75,
+      "learning_rate": 1.9909987830071477e-05,
+      "loss": 0.4744,
+      "step": 2078
+    },
+    {
+      "epoch": 0.1277654867256637,
+      "grad_norm": 42.5,
+      "learning_rate": 1.990967603006048e-05,
+      "loss": 0.3962,
+      "step": 2079
+    },
+    {
+      "epoch": 0.127826941986234,
+      "grad_norm": 21.875,
+      "learning_rate": 1.9909363693398828e-05,
+      "loss": 0.3299,
+      "step": 2080
+    },
+    {
+      "epoch": 0.12788839724680431,
+      "grad_norm": 60.0,
+      "learning_rate": 1.990905082010344e-05,
+      "loss": 0.5674,
+      "step": 2081
+    },
+    {
+      "epoch": 0.12794985250737465,
+      "grad_norm": 62.75,
+      "learning_rate": 1.9908737410191255e-05,
+      "loss": 0.6488,
+      "step": 2082
+    },
+    {
+      "epoch": 0.12801130776794495,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9908423463679246e-05,
+      "loss": 0.4973,
+      "step": 2083
+    },
+    {
+      "epoch": 0.12807276302851525,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9908108980584417e-05,
+      "loss": 0.3893,
+      "step": 2084
+    },
+    {
+      "epoch": 0.12813421828908556,
+      "grad_norm": 15.8125,
+      "learning_rate": 1.9907793960923795e-05,
+      "loss": 0.4961,
+      "step": 2085
+    },
+    {
+      "epoch": 0.12819567354965586,
+      "grad_norm": 9.0,
+      "learning_rate": 1.9907478404714438e-05,
+      "loss": 0.3962,
+      "step": 2086
+    },
+    {
+      "epoch": 0.12825712881022616,
+      "grad_norm": 48.0,
+      "learning_rate": 1.990716231197344e-05,
+      "loss": 0.3513,
+      "step": 2087
+    },
+    {
+      "epoch": 0.12831858407079647,
+      "grad_norm": 5.53125,
+      "learning_rate": 1.9906845682717914e-05,
+      "loss": 0.4186,
+      "step": 2088
+    },
+    {
+      "epoch": 0.12838003933136677,
+      "grad_norm": 23.0,
+      "learning_rate": 1.990652851696501e-05,
+      "loss": 0.3368,
+      "step": 2089
+    },
+    {
+      "epoch": 0.12844149459193707,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9906210814731905e-05,
+      "loss": 0.6681,
+      "step": 2090
+    },
+    {
+      "epoch": 0.12850294985250738,
+      "grad_norm": 17.875,
+      "learning_rate": 1.9905892576035798e-05,
+      "loss": 0.4653,
+      "step": 2091
+    },
+    {
+      "epoch": 0.12856440511307768,
+      "grad_norm": 25.25,
+      "learning_rate": 1.990557380089393e-05,
+      "loss": 0.4199,
+      "step": 2092
+    },
+    {
+      "epoch": 0.12862586037364798,
+      "grad_norm": 21.375,
+      "learning_rate": 1.990525448932356e-05,
+      "loss": 0.4106,
+      "step": 2093
+    },
+    {
+      "epoch": 0.12868731563421829,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9904934641341983e-05,
+      "loss": 0.3241,
+      "step": 2094
+    },
+    {
+      "epoch": 0.1287487708947886,
+      "grad_norm": 30.875,
+      "learning_rate": 1.9904614256966514e-05,
+      "loss": 0.3988,
+      "step": 2095
+    },
+    {
+      "epoch": 0.1288102261553589,
+      "grad_norm": 7.71875,
+      "learning_rate": 1.9904293336214508e-05,
+      "loss": 0.3013,
+      "step": 2096
+    },
+    {
+      "epoch": 0.1288716814159292,
+      "grad_norm": 44.75,
+      "learning_rate": 1.9903971879103344e-05,
+      "loss": 0.5463,
+      "step": 2097
+    },
+    {
+      "epoch": 0.1289331366764995,
+      "grad_norm": 63.25,
+      "learning_rate": 1.990364988565043e-05,
+      "loss": 0.5981,
+      "step": 2098
+    },
+    {
+      "epoch": 0.1289945919370698,
+      "grad_norm": 49.25,
+      "learning_rate": 1.9903327355873205e-05,
+      "loss": 0.5066,
+      "step": 2099
+    },
+    {
+      "epoch": 0.1290560471976401,
+      "grad_norm": 26.5,
+      "learning_rate": 1.990300428978913e-05,
+      "loss": 0.7064,
+      "step": 2100
+    },
+    {
+      "epoch": 0.1291175024582104,
+      "grad_norm": 39.75,
+      "learning_rate": 1.9902680687415704e-05,
+      "loss": 0.594,
+      "step": 2101
+    },
+    {
+      "epoch": 0.12917895771878074,
+      "grad_norm": 75.5,
+      "learning_rate": 1.9902356548770455e-05,
+      "loss": 0.5776,
+      "step": 2102
+    },
+    {
+      "epoch": 0.12924041297935104,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9902031873870927e-05,
+      "loss": 0.3013,
+      "step": 2103
+    },
+    {
+      "epoch": 0.12930186823992135,
+      "grad_norm": 31.5,
+      "learning_rate": 1.990170666273471e-05,
+      "loss": 0.3609,
+      "step": 2104
+    },
+    {
+      "epoch": 0.12936332350049165,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9901380915379413e-05,
+      "loss": 0.4397,
+      "step": 2105
+    },
+    {
+      "epoch": 0.12942477876106195,
+      "grad_norm": 21.625,
+      "learning_rate": 1.990105463182268e-05,
+      "loss": 0.6121,
+      "step": 2106
+    },
+    {
+      "epoch": 0.12948623402163226,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9900727812082177e-05,
+      "loss": 0.5462,
+      "step": 2107
+    },
+    {
+      "epoch": 0.12954768928220256,
+      "grad_norm": 32.25,
+      "learning_rate": 1.99004004561756e-05,
+      "loss": 0.5745,
+      "step": 2108
+    },
+    {
+      "epoch": 0.12960914454277286,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9900072564120686e-05,
+      "loss": 0.7375,
+      "step": 2109
+    },
+    {
+      "epoch": 0.12967059980334317,
+      "grad_norm": 22.75,
+      "learning_rate": 1.989974413593518e-05,
+      "loss": 0.5015,
+      "step": 2110
+    },
+    {
+      "epoch": 0.12973205506391347,
+      "grad_norm": 8.375,
+      "learning_rate": 1.989941517163688e-05,
+      "loss": 0.4905,
+      "step": 2111
+    },
+    {
+      "epoch": 0.12979351032448377,
+      "grad_norm": 37.0,
+      "learning_rate": 1.989908567124359e-05,
+      "loss": 0.5282,
+      "step": 2112
+    },
+    {
+      "epoch": 0.12985496558505408,
+      "grad_norm": 41.25,
+      "learning_rate": 1.989875563477316e-05,
+      "loss": 0.719,
+      "step": 2113
+    },
+    {
+      "epoch": 0.12991642084562438,
+      "grad_norm": 31.625,
+      "learning_rate": 1.989842506224346e-05,
+      "loss": 0.5717,
+      "step": 2114
+    },
+    {
+      "epoch": 0.12997787610619468,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9898093953672392e-05,
+      "loss": 0.6358,
+      "step": 2115
+    },
+    {
+      "epoch": 0.130039331366765,
+      "grad_norm": 30.5,
+      "learning_rate": 1.989776230907789e-05,
+      "loss": 0.5398,
+      "step": 2116
+    },
+    {
+      "epoch": 0.1301007866273353,
+      "grad_norm": 17.0,
+      "learning_rate": 1.9897430128477912e-05,
+      "loss": 0.4648,
+      "step": 2117
+    },
+    {
+      "epoch": 0.1301622418879056,
+      "grad_norm": 10.8125,
+      "learning_rate": 1.989709741189045e-05,
+      "loss": 0.4408,
+      "step": 2118
+    },
+    {
+      "epoch": 0.1302236971484759,
+      "grad_norm": 31.25,
+      "learning_rate": 1.989676415933351e-05,
+      "loss": 0.5942,
+      "step": 2119
+    },
+    {
+      "epoch": 0.1302851524090462,
+      "grad_norm": 22.25,
+      "learning_rate": 1.9896430370825156e-05,
+      "loss": 0.5058,
+      "step": 2120
+    },
+    {
+      "epoch": 0.13034660766961653,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9896096046383456e-05,
+      "loss": 0.6241,
+      "step": 2121
+    },
+    {
+      "epoch": 0.13040806293018684,
+      "grad_norm": 20.25,
+      "learning_rate": 1.989576118602651e-05,
+      "loss": 0.4868,
+      "step": 2122
+    },
+    {
+      "epoch": 0.13046951819075714,
+      "grad_norm": 21.75,
+      "learning_rate": 1.989542578977246e-05,
+      "loss": 0.5702,
+      "step": 2123
+    },
+    {
+      "epoch": 0.13053097345132744,
+      "grad_norm": 33.75,
+      "learning_rate": 1.989508985763947e-05,
+      "loss": 0.4937,
+      "step": 2124
+    },
+    {
+      "epoch": 0.13059242871189775,
+      "grad_norm": 13.0,
+      "learning_rate": 1.9894753389645723e-05,
+      "loss": 0.4441,
+      "step": 2125
+    },
+    {
+      "epoch": 0.13065388397246805,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9894416385809444e-05,
+      "loss": 0.4078,
+      "step": 2126
+    },
+    {
+      "epoch": 0.13071533923303835,
+      "grad_norm": 19.875,
+      "learning_rate": 1.989407884614889e-05,
+      "loss": 0.3699,
+      "step": 2127
+    },
+    {
+      "epoch": 0.13077679449360866,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9893740770682334e-05,
+      "loss": 0.4088,
+      "step": 2128
+    },
+    {
+      "epoch": 0.13083824975417896,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9893402159428084e-05,
+      "loss": 0.5049,
+      "step": 2129
+    },
+    {
+      "epoch": 0.13089970501474926,
+      "grad_norm": 9.875,
+      "learning_rate": 1.989306301240448e-05,
+      "loss": 0.4817,
+      "step": 2130
+    },
+    {
+      "epoch": 0.13096116027531957,
+      "grad_norm": 9.625,
+      "learning_rate": 1.9892723329629885e-05,
+      "loss": 0.5094,
+      "step": 2131
+    },
+    {
+      "epoch": 0.13102261553588987,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9892383111122696e-05,
+      "loss": 0.4608,
+      "step": 2132
+    },
+    {
+      "epoch": 0.13108407079646017,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9892042356901337e-05,
+      "loss": 0.4225,
+      "step": 2133
+    },
+    {
+      "epoch": 0.13114552605703048,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9891701066984264e-05,
+      "loss": 0.3494,
+      "step": 2134
+    },
+    {
+      "epoch": 0.13120698131760078,
+      "grad_norm": 27.375,
+      "learning_rate": 1.9891359241389955e-05,
+      "loss": 0.5765,
+      "step": 2135
+    },
+    {
+      "epoch": 0.13126843657817108,
+      "grad_norm": 14.625,
+      "learning_rate": 1.9891016880136923e-05,
+      "loss": 0.5409,
+      "step": 2136
+    },
+    {
+      "epoch": 0.13132989183874139,
+      "grad_norm": 18.0,
+      "learning_rate": 1.9890673983243708e-05,
+      "loss": 0.4229,
+      "step": 2137
+    },
+    {
+      "epoch": 0.1313913470993117,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.989033055072888e-05,
+      "loss": 0.4937,
+      "step": 2138
+    },
+    {
+      "epoch": 0.131452802359882,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9889986582611036e-05,
+      "loss": 0.6144,
+      "step": 2139
+    },
+    {
+      "epoch": 0.13151425762045232,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.9889642078908805e-05,
+      "loss": 0.466,
+      "step": 2140
+    },
+    {
+      "epoch": 0.13157571288102263,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.988929703964084e-05,
+      "loss": 0.4866,
+      "step": 2141
+    },
+    {
+      "epoch": 0.13163716814159293,
+      "grad_norm": 27.75,
+      "learning_rate": 1.9888951464825835e-05,
+      "loss": 0.5341,
+      "step": 2142
+    },
+    {
+      "epoch": 0.13169862340216323,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.9888605354482494e-05,
+      "loss": 0.3986,
+      "step": 2143
+    },
+    {
+      "epoch": 0.13176007866273354,
+      "grad_norm": 11.75,
+      "learning_rate": 1.9888258708629565e-05,
+      "loss": 0.3388,
+      "step": 2144
+    },
+    {
+      "epoch": 0.13182153392330384,
+      "grad_norm": 32.75,
+      "learning_rate": 1.988791152728582e-05,
+      "loss": 0.704,
+      "step": 2145
+    },
+    {
+      "epoch": 0.13188298918387414,
+      "grad_norm": 17.625,
+      "learning_rate": 1.988756381047006e-05,
+      "loss": 0.6299,
+      "step": 2146
+    },
+    {
+      "epoch": 0.13194444444444445,
+      "grad_norm": 30.0,
+      "learning_rate": 1.9887215558201115e-05,
+      "loss": 0.6765,
+      "step": 2147
+    },
+    {
+      "epoch": 0.13200589970501475,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9886866770497847e-05,
+      "loss": 0.669,
+      "step": 2148
+    },
+    {
+      "epoch": 0.13206735496558505,
+      "grad_norm": 13.5,
+      "learning_rate": 1.988651744737914e-05,
+      "loss": 0.8379,
+      "step": 2149
+    },
+    {
+      "epoch": 0.13212881022615536,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.9886167588863918e-05,
+      "loss": 0.5354,
+      "step": 2150
+    },
+    {
+      "epoch": 0.13219026548672566,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.9885817194971116e-05,
+      "loss": 0.3848,
+      "step": 2151
+    },
+    {
+      "epoch": 0.13225172074729596,
+      "grad_norm": 46.25,
+      "learning_rate": 1.9885466265719723e-05,
+      "loss": 0.5101,
+      "step": 2152
+    },
+    {
+      "epoch": 0.13231317600786627,
+      "grad_norm": 13.875,
+      "learning_rate": 1.988511480112873e-05,
+      "loss": 0.4344,
+      "step": 2153
+    },
+    {
+      "epoch": 0.13237463126843657,
+      "grad_norm": 12.5,
+      "learning_rate": 1.9884762801217185e-05,
+      "loss": 0.5549,
+      "step": 2154
+    },
+    {
+      "epoch": 0.13243608652900687,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.9884410266004134e-05,
+      "loss": 0.3857,
+      "step": 2155
+    },
+    {
+      "epoch": 0.13249754178957718,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9884057195508683e-05,
+      "loss": 0.4965,
+      "step": 2156
+    },
+    {
+      "epoch": 0.13255899705014748,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.9883703589749938e-05,
+      "loss": 0.3709,
+      "step": 2157
+    },
+    {
+      "epoch": 0.13262045231071778,
+      "grad_norm": 25.875,
+      "learning_rate": 1.988334944874706e-05,
+      "loss": 0.3618,
+      "step": 2158
+    },
+    {
+      "epoch": 0.13268190757128812,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.9882994772519225e-05,
+      "loss": 0.3453,
+      "step": 2159
+    },
+    {
+      "epoch": 0.13274336283185842,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9882639561085634e-05,
+      "loss": 0.449,
+      "step": 2160
+    },
+    {
+      "epoch": 0.13280481809242872,
+      "grad_norm": 30.625,
+      "learning_rate": 1.988228381446553e-05,
+      "loss": 0.5761,
+      "step": 2161
+    },
+    {
+      "epoch": 0.13286627335299903,
+      "grad_norm": 28.75,
+      "learning_rate": 1.9881927532678174e-05,
+      "loss": 0.5464,
+      "step": 2162
+    },
+    {
+      "epoch": 0.13292772861356933,
+      "grad_norm": 52.0,
+      "learning_rate": 1.9881570715742865e-05,
+      "loss": 0.5055,
+      "step": 2163
+    },
+    {
+      "epoch": 0.13298918387413963,
+      "grad_norm": 9.625,
+      "learning_rate": 1.988121336367892e-05,
+      "loss": 0.5087,
+      "step": 2164
+    },
+    {
+      "epoch": 0.13305063913470994,
+      "grad_norm": 47.25,
+      "learning_rate": 1.9880855476505694e-05,
+      "loss": 0.5997,
+      "step": 2165
+    },
+    {
+      "epoch": 0.13311209439528024,
+      "grad_norm": 10.625,
+      "learning_rate": 1.9880497054242566e-05,
+      "loss": 0.5849,
+      "step": 2166
+    },
+    {
+      "epoch": 0.13317354965585054,
+      "grad_norm": 5.3125,
+      "learning_rate": 1.9880138096908955e-05,
+      "loss": 0.4537,
+      "step": 2167
+    },
+    {
+      "epoch": 0.13323500491642085,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.9879778604524286e-05,
+      "loss": 0.5885,
+      "step": 2168
+    },
+    {
+      "epoch": 0.13329646017699115,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9879418577108037e-05,
+      "loss": 0.5533,
+      "step": 2169
+    },
+    {
+      "epoch": 0.13335791543756145,
+      "grad_norm": 26.375,
+      "learning_rate": 1.9879058014679704e-05,
+      "loss": 0.4738,
+      "step": 2170
+    },
+    {
+      "epoch": 0.13341937069813176,
+      "grad_norm": 46.25,
+      "learning_rate": 1.987869691725881e-05,
+      "loss": 0.4987,
+      "step": 2171
+    },
+    {
+      "epoch": 0.13348082595870206,
+      "grad_norm": 73.5,
+      "learning_rate": 1.987833528486491e-05,
+      "loss": 0.5424,
+      "step": 2172
+    },
+    {
+      "epoch": 0.13354228121927236,
+      "grad_norm": 43.0,
+      "learning_rate": 1.987797311751759e-05,
+      "loss": 0.427,
+      "step": 2173
+    },
+    {
+      "epoch": 0.13360373647984267,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9877610415236464e-05,
+      "loss": 0.4269,
+      "step": 2174
+    },
+    {
+      "epoch": 0.13366519174041297,
+      "grad_norm": 53.0,
+      "learning_rate": 1.9877247178041167e-05,
+      "loss": 0.5603,
+      "step": 2175
+    },
+    {
+      "epoch": 0.13372664700098327,
+      "grad_norm": 5.875,
+      "learning_rate": 1.9876883405951378e-05,
+      "loss": 0.4074,
+      "step": 2176
+    },
+    {
+      "epoch": 0.13378810226155358,
+      "grad_norm": 21.875,
+      "learning_rate": 1.987651909898679e-05,
+      "loss": 0.4472,
+      "step": 2177
+    },
+    {
+      "epoch": 0.1338495575221239,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9876154257167143e-05,
+      "loss": 0.639,
+      "step": 2178
+    },
+    {
+      "epoch": 0.1339110127826942,
+      "grad_norm": 32.75,
+      "learning_rate": 1.9875788880512183e-05,
+      "loss": 0.6329,
+      "step": 2179
+    },
+    {
+      "epoch": 0.13397246804326451,
+      "grad_norm": 10.875,
+      "learning_rate": 1.9875422969041697e-05,
+      "loss": 0.5516,
+      "step": 2180
+    },
+    {
+      "epoch": 0.13403392330383482,
+      "grad_norm": 12.25,
+      "learning_rate": 1.9875056522775506e-05,
+      "loss": 0.5832,
+      "step": 2181
+    },
+    {
+      "epoch": 0.13409537856440512,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9874689541733455e-05,
+      "loss": 0.4347,
+      "step": 2182
+    },
+    {
+      "epoch": 0.13415683382497542,
+      "grad_norm": 37.5,
+      "learning_rate": 1.9874322025935414e-05,
+      "loss": 0.3835,
+      "step": 2183
+    },
+    {
+      "epoch": 0.13421828908554573,
+      "grad_norm": 28.125,
+      "learning_rate": 1.9873953975401286e-05,
+      "loss": 0.4577,
+      "step": 2184
+    },
+    {
+      "epoch": 0.13427974434611603,
+      "grad_norm": 42.75,
+      "learning_rate": 1.9873585390151003e-05,
+      "loss": 0.4348,
+      "step": 2185
+    },
+    {
+      "epoch": 0.13434119960668633,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.987321627020453e-05,
+      "loss": 0.5868,
+      "step": 2186
+    },
+    {
+      "epoch": 0.13440265486725664,
+      "grad_norm": 26.875,
+      "learning_rate": 1.9872846615581848e-05,
+      "loss": 0.5404,
+      "step": 2187
+    },
+    {
+      "epoch": 0.13446411012782694,
+      "grad_norm": 16.875,
+      "learning_rate": 1.9872476426302983e-05,
+      "loss": 0.4522,
+      "step": 2188
+    },
+    {
+      "epoch": 0.13452556538839724,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9872105702387976e-05,
+      "loss": 0.5009,
+      "step": 2189
+    },
+    {
+      "epoch": 0.13458702064896755,
+      "grad_norm": 19.75,
+      "learning_rate": 1.9871734443856907e-05,
+      "loss": 0.489,
+      "step": 2190
+    },
+    {
+      "epoch": 0.13464847590953785,
+      "grad_norm": 7.0,
+      "learning_rate": 1.987136265072988e-05,
+      "loss": 0.3451,
+      "step": 2191
+    },
+    {
+      "epoch": 0.13470993117010815,
+      "grad_norm": 27.75,
+      "learning_rate": 1.9870990323027032e-05,
+      "loss": 0.4976,
+      "step": 2192
+    },
+    {
+      "epoch": 0.13477138643067846,
+      "grad_norm": 23.75,
+      "learning_rate": 1.987061746076852e-05,
+      "loss": 0.4327,
+      "step": 2193
+    },
+    {
+      "epoch": 0.13483284169124876,
+      "grad_norm": 31.125,
+      "learning_rate": 1.987024406397454e-05,
+      "loss": 0.5763,
+      "step": 2194
+    },
+    {
+      "epoch": 0.13489429695181906,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9869870132665318e-05,
+      "loss": 0.5103,
+      "step": 2195
+    },
+    {
+      "epoch": 0.13495575221238937,
+      "grad_norm": 19.875,
+      "learning_rate": 1.9869495666861094e-05,
+      "loss": 0.5635,
+      "step": 2196
+    },
+    {
+      "epoch": 0.13501720747295967,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9869120666582153e-05,
+      "loss": 0.3279,
+      "step": 2197
+    },
+    {
+      "epoch": 0.13507866273353,
+      "grad_norm": 20.25,
+      "learning_rate": 1.9868745131848798e-05,
+      "loss": 0.4138,
+      "step": 2198
+    },
+    {
+      "epoch": 0.1351401179941003,
+      "grad_norm": 28.125,
+      "learning_rate": 1.9868369062681373e-05,
+      "loss": 0.4851,
+      "step": 2199
+    },
+    {
+      "epoch": 0.1352015732546706,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.986799245910024e-05,
+      "loss": 0.3865,
+      "step": 2200
+    },
+    {
+      "epoch": 0.1352630285152409,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9867615321125796e-05,
+      "loss": 0.5552,
+      "step": 2201
+    },
+    {
+      "epoch": 0.13532448377581122,
+      "grad_norm": 27.875,
+      "learning_rate": 1.986723764877846e-05,
+      "loss": 0.6568,
+      "step": 2202
+    },
+    {
+      "epoch": 0.13538593903638152,
+      "grad_norm": 21.375,
+      "learning_rate": 1.986685944207868e-05,
+      "loss": 0.574,
+      "step": 2203
+    },
+    {
+      "epoch": 0.13544739429695182,
+      "grad_norm": 41.5,
+      "learning_rate": 1.9866480701046953e-05,
+      "loss": 0.6094,
+      "step": 2204
+    },
+    {
+      "epoch": 0.13550884955752213,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9866101425703776e-05,
+      "loss": 0.6465,
+      "step": 2205
+    },
+    {
+      "epoch": 0.13557030481809243,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9865721616069695e-05,
+      "loss": 0.6278,
+      "step": 2206
+    },
+    {
+      "epoch": 0.13563176007866273,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9865341272165276e-05,
+      "loss": 0.393,
+      "step": 2207
+    },
+    {
+      "epoch": 0.13569321533923304,
+      "grad_norm": 16.625,
+      "learning_rate": 1.9864960394011115e-05,
+      "loss": 0.4922,
+      "step": 2208
+    },
+    {
+      "epoch": 0.13575467059980334,
+      "grad_norm": 28.625,
+      "learning_rate": 1.9864578981627844e-05,
+      "loss": 0.5771,
+      "step": 2209
+    },
+    {
+      "epoch": 0.13581612586037364,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9864197035036112e-05,
+      "loss": 0.5807,
+      "step": 2210
+    },
+    {
+      "epoch": 0.13587758112094395,
+      "grad_norm": 16.5,
+      "learning_rate": 1.98638145542566e-05,
+      "loss": 0.4254,
+      "step": 2211
+    },
+    {
+      "epoch": 0.13593903638151425,
+      "grad_norm": 13.375,
+      "learning_rate": 1.9863431539310033e-05,
+      "loss": 0.6729,
+      "step": 2212
+    },
+    {
+      "epoch": 0.13600049164208455,
+      "grad_norm": 16.5,
+      "learning_rate": 1.9863047990217143e-05,
+      "loss": 0.527,
+      "step": 2213
+    },
+    {
+      "epoch": 0.13606194690265486,
+      "grad_norm": 14.0625,
+      "learning_rate": 1.98626639069987e-05,
+      "loss": 0.3867,
+      "step": 2214
+    },
+    {
+      "epoch": 0.13612340216322516,
+      "grad_norm": 12.75,
+      "learning_rate": 1.986227928967551e-05,
+      "loss": 0.4007,
+      "step": 2215
+    },
+    {
+      "epoch": 0.13618485742379546,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9861894138268402e-05,
+      "loss": 0.4178,
+      "step": 2216
+    },
+    {
+      "epoch": 0.1362463126843658,
+      "grad_norm": 27.875,
+      "learning_rate": 1.9861508452798224e-05,
+      "loss": 0.5855,
+      "step": 2217
+    },
+    {
+      "epoch": 0.1363077679449361,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.9861122233285873e-05,
+      "loss": 0.4761,
+      "step": 2218
+    },
+    {
+      "epoch": 0.1363692232055064,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.986073547975226e-05,
+      "loss": 0.372,
+      "step": 2219
+    },
+    {
+      "epoch": 0.1364306784660767,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9860348192218327e-05,
+      "loss": 0.5724,
+      "step": 2220
+    },
+    {
+      "epoch": 0.136492133726647,
+      "grad_norm": 29.875,
+      "learning_rate": 1.985996037070505e-05,
+      "loss": 0.4056,
+      "step": 2221
+    },
+    {
+      "epoch": 0.1365535889872173,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.9859572015233436e-05,
+      "loss": 0.5368,
+      "step": 2222
+    },
+    {
+      "epoch": 0.13661504424778761,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9859183125824506e-05,
+      "loss": 0.3944,
+      "step": 2223
+    },
+    {
+      "epoch": 0.13667649950835792,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9858793702499322e-05,
+      "loss": 0.5275,
+      "step": 2224
+    },
+    {
+      "epoch": 0.13673795476892822,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9858403745278982e-05,
+      "loss": 0.5205,
+      "step": 2225
+    },
+    {
+      "epoch": 0.13679941002949852,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9858013254184597e-05,
+      "loss": 0.6841,
+      "step": 2226
+    },
+    {
+      "epoch": 0.13686086529006883,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9857622229237315e-05,
+      "loss": 0.7128,
+      "step": 2227
+    },
+    {
+      "epoch": 0.13692232055063913,
+      "grad_norm": 33.75,
+      "learning_rate": 1.985723067045831e-05,
+      "loss": 0.6001,
+      "step": 2228
+    },
+    {
+      "epoch": 0.13698377581120944,
+      "grad_norm": 33.5,
+      "learning_rate": 1.985683857786879e-05,
+      "loss": 0.6496,
+      "step": 2229
+    },
+    {
+      "epoch": 0.13704523107177974,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9856445951489984e-05,
+      "loss": 0.601,
+      "step": 2230
+    },
+    {
+      "epoch": 0.13710668633235004,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.9856052791343153e-05,
+      "loss": 0.3879,
+      "step": 2231
+    },
+    {
+      "epoch": 0.13716814159292035,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9855659097449597e-05,
+      "loss": 0.4743,
+      "step": 2232
+    },
+    {
+      "epoch": 0.13722959685349065,
+      "grad_norm": 17.5,
+      "learning_rate": 1.985526486983063e-05,
+      "loss": 0.7091,
+      "step": 2233
+    },
+    {
+      "epoch": 0.13729105211406095,
+      "grad_norm": 18.625,
+      "learning_rate": 1.98548701085076e-05,
+      "loss": 0.6984,
+      "step": 2234
+    },
+    {
+      "epoch": 0.13735250737463126,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.985447481350189e-05,
+      "loss": 0.5653,
+      "step": 2235
+    },
+    {
+      "epoch": 0.1374139626352016,
+      "grad_norm": 19.75,
+      "learning_rate": 1.9854078984834904e-05,
+      "loss": 0.4881,
+      "step": 2236
+    },
+    {
+      "epoch": 0.1374754178957719,
+      "grad_norm": 28.375,
+      "learning_rate": 1.9853682622528074e-05,
+      "loss": 0.6677,
+      "step": 2237
+    },
+    {
+      "epoch": 0.1375368731563422,
+      "grad_norm": 12.125,
+      "learning_rate": 1.9853285726602874e-05,
+      "loss": 0.4534,
+      "step": 2238
+    },
+    {
+      "epoch": 0.1375983284169125,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.985288829708079e-05,
+      "loss": 0.4135,
+      "step": 2239
+    },
+    {
+      "epoch": 0.1376597836774828,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9852490333983342e-05,
+      "loss": 0.5622,
+      "step": 2240
+    },
+    {
+      "epoch": 0.1377212389380531,
+      "grad_norm": 9.5625,
+      "learning_rate": 1.985209183733209e-05,
+      "loss": 0.3986,
+      "step": 2241
+    },
+    {
+      "epoch": 0.1377826941986234,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9851692807148612e-05,
+      "loss": 0.3636,
+      "step": 2242
+    },
+    {
+      "epoch": 0.1378441494591937,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9851293243454513e-05,
+      "loss": 0.3933,
+      "step": 2243
+    },
+    {
+      "epoch": 0.137905604719764,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.985089314627143e-05,
+      "loss": 0.42,
+      "step": 2244
+    },
+    {
+      "epoch": 0.13796705998033432,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.9850492515621038e-05,
+      "loss": 0.4015,
+      "step": 2245
+    },
+    {
+      "epoch": 0.13802851524090462,
+      "grad_norm": 14.75,
+      "learning_rate": 1.985009135152503e-05,
+      "loss": 0.3528,
+      "step": 2246
+    },
+    {
+      "epoch": 0.13808997050147492,
+      "grad_norm": 5.125,
+      "learning_rate": 1.9849689654005124e-05,
+      "loss": 0.2579,
+      "step": 2247
+    },
+    {
+      "epoch": 0.13815142576204523,
+      "grad_norm": 25.75,
+      "learning_rate": 1.984928742308308e-05,
+      "loss": 0.3416,
+      "step": 2248
+    },
+    {
+      "epoch": 0.13821288102261553,
+      "grad_norm": 12.8125,
+      "learning_rate": 1.984888465878068e-05,
+      "loss": 0.3824,
+      "step": 2249
+    },
+    {
+      "epoch": 0.13827433628318583,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9848481361119733e-05,
+      "loss": 0.5539,
+      "step": 2250
+    },
+    {
+      "epoch": 0.13833579154375614,
+      "grad_norm": 9.5625,
+      "learning_rate": 1.9848077530122083e-05,
+      "loss": 0.5354,
+      "step": 2251
+    },
+    {
+      "epoch": 0.13839724680432644,
+      "grad_norm": 26.25,
+      "learning_rate": 1.9847673165809595e-05,
+      "loss": 0.5823,
+      "step": 2252
+    },
+    {
+      "epoch": 0.13845870206489674,
+      "grad_norm": 63.0,
+      "learning_rate": 1.9847268268204168e-05,
+      "loss": 0.716,
+      "step": 2253
+    },
+    {
+      "epoch": 0.13852015732546705,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9846862837327733e-05,
+      "loss": 0.5441,
+      "step": 2254
+    },
+    {
+      "epoch": 0.13858161258603738,
+      "grad_norm": 27.125,
+      "learning_rate": 1.984645687320224e-05,
+      "loss": 0.5445,
+      "step": 2255
+    },
+    {
+      "epoch": 0.13864306784660768,
+      "grad_norm": 51.0,
+      "learning_rate": 1.9846050375849674e-05,
+      "loss": 0.5248,
+      "step": 2256
+    },
+    {
+      "epoch": 0.13870452310717798,
+      "grad_norm": 12.0,
+      "learning_rate": 1.9845643345292055e-05,
+      "loss": 0.6126,
+      "step": 2257
+    },
+    {
+      "epoch": 0.1387659783677483,
+      "grad_norm": 26.625,
+      "learning_rate": 1.984523578155142e-05,
+      "loss": 0.5107,
+      "step": 2258
+    },
+    {
+      "epoch": 0.1388274336283186,
+      "grad_norm": 29.25,
+      "learning_rate": 1.984482768464984e-05,
+      "loss": 0.506,
+      "step": 2259
+    },
+    {
+      "epoch": 0.1388888888888889,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9844419054609418e-05,
+      "loss": 0.595,
+      "step": 2260
+    },
+    {
+      "epoch": 0.1389503441494592,
+      "grad_norm": 28.375,
+      "learning_rate": 1.984400989145228e-05,
+      "loss": 0.4633,
+      "step": 2261
+    },
+    {
+      "epoch": 0.1390117994100295,
+      "grad_norm": 20.625,
+      "learning_rate": 1.984360019520059e-05,
+      "loss": 0.5401,
+      "step": 2262
+    },
+    {
+      "epoch": 0.1390732546705998,
+      "grad_norm": 9.125,
+      "learning_rate": 1.9843189965876525e-05,
+      "loss": 0.3445,
+      "step": 2263
+    },
+    {
+      "epoch": 0.1391347099311701,
+      "grad_norm": 29.625,
+      "learning_rate": 1.984277920350231e-05,
+      "loss": 0.3929,
+      "step": 2264
+    },
+    {
+      "epoch": 0.1391961651917404,
+      "grad_norm": 45.0,
+      "learning_rate": 1.9842367908100182e-05,
+      "loss": 0.5897,
+      "step": 2265
+    },
+    {
+      "epoch": 0.13925762045231072,
+      "grad_norm": 17.75,
+      "learning_rate": 1.984195607969242e-05,
+      "loss": 0.439,
+      "step": 2266
+    },
+    {
+      "epoch": 0.13931907571288102,
+      "grad_norm": 65.0,
+      "learning_rate": 1.9841543718301324e-05,
+      "loss": 0.5191,
+      "step": 2267
+    },
+    {
+      "epoch": 0.13938053097345132,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.9841130823949223e-05,
+      "loss": 0.5323,
+      "step": 2268
+    },
+    {
+      "epoch": 0.13944198623402163,
+      "grad_norm": 35.25,
+      "learning_rate": 1.9840717396658483e-05,
+      "loss": 0.5804,
+      "step": 2269
+    },
+    {
+      "epoch": 0.13950344149459193,
+      "grad_norm": 34.5,
+      "learning_rate": 1.9840303436451485e-05,
+      "loss": 0.5179,
+      "step": 2270
+    },
+    {
+      "epoch": 0.13956489675516223,
+      "grad_norm": 13.75,
+      "learning_rate": 1.9839888943350656e-05,
+      "loss": 0.3176,
+      "step": 2271
+    },
+    {
+      "epoch": 0.13962635201573254,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.9839473917378432e-05,
+      "loss": 0.5403,
+      "step": 2272
+    },
+    {
+      "epoch": 0.13968780727630284,
+      "grad_norm": 28.625,
+      "learning_rate": 1.9839058358557296e-05,
+      "loss": 0.8265,
+      "step": 2273
+    },
+    {
+      "epoch": 0.13974926253687317,
+      "grad_norm": 42.25,
+      "learning_rate": 1.983864226690975e-05,
+      "loss": 0.5885,
+      "step": 2274
+    },
+    {
+      "epoch": 0.13981071779744347,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.983822564245833e-05,
+      "loss": 0.4583,
+      "step": 2275
+    },
+    {
+      "epoch": 0.13987217305801378,
+      "grad_norm": 10.75,
+      "learning_rate": 1.983780848522559e-05,
+      "loss": 0.4722,
+      "step": 2276
+    },
+    {
+      "epoch": 0.13993362831858408,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9837390795234128e-05,
+      "loss": 0.5182,
+      "step": 2277
+    },
+    {
+      "epoch": 0.13999508357915438,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.9836972572506557e-05,
+      "loss": 0.4182,
+      "step": 2278
+    },
+    {
+      "epoch": 0.1400565388397247,
+      "grad_norm": 28.375,
+      "learning_rate": 1.983655381706554e-05,
+      "loss": 0.4289,
+      "step": 2279
+    },
+    {
+      "epoch": 0.140117994100295,
+      "grad_norm": 9.0,
+      "learning_rate": 1.9836134528933733e-05,
+      "loss": 0.4892,
+      "step": 2280
+    },
+    {
+      "epoch": 0.1401794493608653,
+      "grad_norm": 21.375,
+      "learning_rate": 1.983571470813386e-05,
+      "loss": 0.5442,
+      "step": 2281
+    },
+    {
+      "epoch": 0.1402409046214356,
+      "grad_norm": 24.875,
+      "learning_rate": 1.983529435468865e-05,
+      "loss": 0.7,
+      "step": 2282
+    },
+    {
+      "epoch": 0.1403023598820059,
+      "grad_norm": 9.5,
+      "learning_rate": 1.9834873468620864e-05,
+      "loss": 0.3771,
+      "step": 2283
+    },
+    {
+      "epoch": 0.1403638151425762,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.98344520499533e-05,
+      "loss": 0.4826,
+      "step": 2284
+    },
+    {
+      "epoch": 0.1404252704031465,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9834030098708773e-05,
+      "loss": 0.766,
+      "step": 2285
+    },
+    {
+      "epoch": 0.1404867256637168,
+      "grad_norm": 36.5,
+      "learning_rate": 1.983360761491014e-05,
+      "loss": 0.5972,
+      "step": 2286
+    },
+    {
+      "epoch": 0.1405481809242871,
+      "grad_norm": 20.5,
+      "learning_rate": 1.983318459858028e-05,
+      "loss": 0.5578,
+      "step": 2287
+    },
+    {
+      "epoch": 0.14060963618485742,
+      "grad_norm": 15.5625,
+      "learning_rate": 1.9832761049742092e-05,
+      "loss": 0.4374,
+      "step": 2288
+    },
+    {
+      "epoch": 0.14067109144542772,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9832336968418527e-05,
+      "loss": 0.5643,
+      "step": 2289
+    },
+    {
+      "epoch": 0.14073254670599802,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9831912354632537e-05,
+      "loss": 0.4419,
+      "step": 2290
+    },
+    {
+      "epoch": 0.14079400196656833,
+      "grad_norm": 10.25,
+      "learning_rate": 1.9831487208407126e-05,
+      "loss": 0.6119,
+      "step": 2291
+    },
+    {
+      "epoch": 0.14085545722713863,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9831061529765316e-05,
+      "loss": 0.6325,
+      "step": 2292
+    },
+    {
+      "epoch": 0.14091691248770893,
+      "grad_norm": 51.5,
+      "learning_rate": 1.9830635318730155e-05,
+      "loss": 0.5302,
+      "step": 2293
+    },
+    {
+      "epoch": 0.14097836774827927,
+      "grad_norm": 19.5,
+      "learning_rate": 1.983020857532473e-05,
+      "loss": 0.5718,
+      "step": 2294
+    },
+    {
+      "epoch": 0.14103982300884957,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.9829781299572147e-05,
+      "loss": 0.4449,
+      "step": 2295
+    },
+    {
+      "epoch": 0.14110127826941987,
+      "grad_norm": 30.375,
+      "learning_rate": 1.9829353491495545e-05,
+      "loss": 0.4569,
+      "step": 2296
+    },
+    {
+      "epoch": 0.14116273352999018,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9828925151118092e-05,
+      "loss": 0.7089,
+      "step": 2297
+    },
+    {
+      "epoch": 0.14122418879056048,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9828496278462987e-05,
+      "loss": 0.7536,
+      "step": 2298
+    },
+    {
+      "epoch": 0.14128564405113078,
+      "grad_norm": 25.875,
+      "learning_rate": 1.982806687355345e-05,
+      "loss": 0.5499,
+      "step": 2299
+    },
+    {
+      "epoch": 0.14134709931170109,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9827636936412736e-05,
+      "loss": 0.6142,
+      "step": 2300
+    },
+    {
+      "epoch": 0.1414085545722714,
+      "grad_norm": 6.9375,
+      "learning_rate": 1.9827206467064133e-05,
+      "loss": 0.5247,
+      "step": 2301
+    },
+    {
+      "epoch": 0.1414700098328417,
+      "grad_norm": 17.125,
+      "learning_rate": 1.982677546553095e-05,
+      "loss": 0.5337,
+      "step": 2302
+    },
+    {
+      "epoch": 0.141531465093412,
+      "grad_norm": 25.75,
+      "learning_rate": 1.982634393183653e-05,
+      "loss": 0.6573,
+      "step": 2303
+    },
+    {
+      "epoch": 0.1415929203539823,
+      "grad_norm": 6.65625,
+      "learning_rate": 1.9825911866004235e-05,
+      "loss": 0.5288,
+      "step": 2304
+    },
+    {
+      "epoch": 0.1416543756145526,
+      "grad_norm": 18.25,
+      "learning_rate": 1.982547926805747e-05,
+      "loss": 0.4985,
+      "step": 2305
+    },
+    {
+      "epoch": 0.1417158308751229,
+      "grad_norm": 12.1875,
+      "learning_rate": 1.9825046138019658e-05,
+      "loss": 0.607,
+      "step": 2306
+    },
+    {
+      "epoch": 0.1417772861356932,
+      "grad_norm": 18.0,
+      "learning_rate": 1.982461247591426e-05,
+      "loss": 0.4951,
+      "step": 2307
+    },
+    {
+      "epoch": 0.1418387413962635,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9824178281764753e-05,
+      "loss": 0.4146,
+      "step": 2308
+    },
+    {
+      "epoch": 0.14190019665683382,
+      "grad_norm": 27.125,
+      "learning_rate": 1.9823743555594655e-05,
+      "loss": 0.4983,
+      "step": 2309
+    },
+    {
+      "epoch": 0.14196165191740412,
+      "grad_norm": 26.875,
+      "learning_rate": 1.9823308297427514e-05,
+      "loss": 0.3965,
+      "step": 2310
+    },
+    {
+      "epoch": 0.14202310717797442,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.982287250728689e-05,
+      "loss": 0.3937,
+      "step": 2311
+    },
+    {
+      "epoch": 0.14208456243854473,
+      "grad_norm": 16.125,
+      "learning_rate": 1.982243618519639e-05,
+      "loss": 0.4282,
+      "step": 2312
+    },
+    {
+      "epoch": 0.14214601769911506,
+      "grad_norm": 39.25,
+      "learning_rate": 1.9821999331179635e-05,
+      "loss": 0.5139,
+      "step": 2313
+    },
+    {
+      "epoch": 0.14220747295968536,
+      "grad_norm": 26.125,
+      "learning_rate": 1.9821561945260292e-05,
+      "loss": 0.4733,
+      "step": 2314
+    },
+    {
+      "epoch": 0.14226892822025566,
+      "grad_norm": 15.75,
+      "learning_rate": 1.9821124027462043e-05,
+      "loss": 0.5178,
+      "step": 2315
+    },
+    {
+      "epoch": 0.14233038348082597,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9820685577808604e-05,
+      "loss": 0.6117,
+      "step": 2316
+    },
+    {
+      "epoch": 0.14239183874139627,
+      "grad_norm": 25.75,
+      "learning_rate": 1.982024659632372e-05,
+      "loss": 0.5936,
+      "step": 2317
+    },
+    {
+      "epoch": 0.14245329400196657,
+      "grad_norm": 27.625,
+      "learning_rate": 1.981980708303116e-05,
+      "loss": 0.4173,
+      "step": 2318
+    },
+    {
+      "epoch": 0.14251474926253688,
+      "grad_norm": 14.0,
+      "learning_rate": 1.9819367037954725e-05,
+      "loss": 0.4138,
+      "step": 2319
+    },
+    {
+      "epoch": 0.14257620452310718,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9818926461118254e-05,
+      "loss": 0.4707,
+      "step": 2320
+    },
+    {
+      "epoch": 0.14263765978367748,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9818485352545595e-05,
+      "loss": 0.7985,
+      "step": 2321
+    },
+    {
+      "epoch": 0.1426991150442478,
+      "grad_norm": 20.125,
+      "learning_rate": 1.981804371226064e-05,
+      "loss": 0.6754,
+      "step": 2322
+    },
+    {
+      "epoch": 0.1427605703048181,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.981760154028731e-05,
+      "loss": 0.6029,
+      "step": 2323
+    },
+    {
+      "epoch": 0.1428220255653884,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9817158836649542e-05,
+      "loss": 0.5088,
+      "step": 2324
+    },
+    {
+      "epoch": 0.1428834808259587,
+      "grad_norm": 22.25,
+      "learning_rate": 1.981671560137132e-05,
+      "loss": 0.4542,
+      "step": 2325
+    },
+    {
+      "epoch": 0.142944936086529,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.9816271834476642e-05,
+      "loss": 0.7558,
+      "step": 2326
+    },
+    {
+      "epoch": 0.1430063913470993,
+      "grad_norm": 22.5,
+      "learning_rate": 1.981582753598954e-05,
+      "loss": 0.5192,
+      "step": 2327
+    },
+    {
+      "epoch": 0.1430678466076696,
+      "grad_norm": 31.875,
+      "learning_rate": 1.981538270593407e-05,
+      "loss": 0.4768,
+      "step": 2328
+    },
+    {
+      "epoch": 0.1431293018682399,
+      "grad_norm": 37.5,
+      "learning_rate": 1.981493734433433e-05,
+      "loss": 0.4595,
+      "step": 2329
+    },
+    {
+      "epoch": 0.14319075712881021,
+      "grad_norm": 32.5,
+      "learning_rate": 1.9814491451214434e-05,
+      "loss": 0.6074,
+      "step": 2330
+    },
+    {
+      "epoch": 0.14325221238938052,
+      "grad_norm": 27.75,
+      "learning_rate": 1.981404502659853e-05,
+      "loss": 0.5599,
+      "step": 2331
+    },
+    {
+      "epoch": 0.14331366764995085,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.981359807051079e-05,
+      "loss": 0.5609,
+      "step": 2332
+    },
+    {
+      "epoch": 0.14337512291052115,
+      "grad_norm": 11.0625,
+      "learning_rate": 1.9813150582975425e-05,
+      "loss": 0.4313,
+      "step": 2333
+    },
+    {
+      "epoch": 0.14343657817109146,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9812702564016664e-05,
+      "loss": 0.4355,
+      "step": 2334
+    },
+    {
+      "epoch": 0.14349803343166176,
+      "grad_norm": 8.375,
+      "learning_rate": 1.981225401365877e-05,
+      "loss": 0.352,
+      "step": 2335
+    },
+    {
+      "epoch": 0.14355948869223206,
+      "grad_norm": 13.125,
+      "learning_rate": 1.981180493192603e-05,
+      "loss": 0.4398,
+      "step": 2336
+    },
+    {
+      "epoch": 0.14362094395280237,
+      "grad_norm": 25.375,
+      "learning_rate": 1.9811355318842776e-05,
+      "loss": 0.5923,
+      "step": 2337
+    },
+    {
+      "epoch": 0.14368239921337267,
+      "grad_norm": 16.5,
+      "learning_rate": 1.981090517443334e-05,
+      "loss": 0.4834,
+      "step": 2338
+    },
+    {
+      "epoch": 0.14374385447394297,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9810454498722114e-05,
+      "loss": 0.4312,
+      "step": 2339
+    },
+    {
+      "epoch": 0.14380530973451328,
+      "grad_norm": 6.375,
+      "learning_rate": 1.9810003291733492e-05,
+      "loss": 0.3149,
+      "step": 2340
+    },
+    {
+      "epoch": 0.14386676499508358,
+      "grad_norm": 49.25,
+      "learning_rate": 1.9809551553491918e-05,
+      "loss": 0.4829,
+      "step": 2341
+    },
+    {
+      "epoch": 0.14392822025565388,
+      "grad_norm": 5.90625,
+      "learning_rate": 1.980909928402185e-05,
+      "loss": 0.378,
+      "step": 2342
+    },
+    {
+      "epoch": 0.14398967551622419,
+      "grad_norm": 13.0,
+      "learning_rate": 1.980864648334778e-05,
+      "loss": 0.3847,
+      "step": 2343
+    },
+    {
+      "epoch": 0.1440511307767945,
+      "grad_norm": 5.03125,
+      "learning_rate": 1.9808193151494233e-05,
+      "loss": 0.3351,
+      "step": 2344
+    },
+    {
+      "epoch": 0.1441125860373648,
+      "grad_norm": 10.8125,
+      "learning_rate": 1.9807739288485758e-05,
+      "loss": 0.3433,
+      "step": 2345
+    },
+    {
+      "epoch": 0.1441740412979351,
+      "grad_norm": 49.75,
+      "learning_rate": 1.980728489434693e-05,
+      "loss": 0.5013,
+      "step": 2346
+    },
+    {
+      "epoch": 0.1442354965585054,
+      "grad_norm": 3.109375,
+      "learning_rate": 1.9806829969102356e-05,
+      "loss": 0.373,
+      "step": 2347
+    },
+    {
+      "epoch": 0.1442969518190757,
+      "grad_norm": 17.5,
+      "learning_rate": 1.980637451277668e-05,
+      "loss": 0.4202,
+      "step": 2348
+    },
+    {
+      "epoch": 0.144358407079646,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.980591852539456e-05,
+      "loss": 0.4677,
+      "step": 2349
+    },
+    {
+      "epoch": 0.1444198623402163,
+      "grad_norm": 20.125,
+      "learning_rate": 1.9805462006980688e-05,
+      "loss": 0.4328,
+      "step": 2350
+    },
+    {
+      "epoch": 0.14448131760078664,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9805004957559795e-05,
+      "loss": 0.4094,
+      "step": 2351
+    },
+    {
+      "epoch": 0.14454277286135694,
+      "grad_norm": 29.375,
+      "learning_rate": 1.9804547377156624e-05,
+      "loss": 0.3067,
+      "step": 2352
+    },
+    {
+      "epoch": 0.14460422812192725,
+      "grad_norm": 28.625,
+      "learning_rate": 1.980408926579596e-05,
+      "loss": 0.7151,
+      "step": 2353
+    },
+    {
+      "epoch": 0.14466568338249755,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9803630623502606e-05,
+      "loss": 0.5994,
+      "step": 2354
+    },
+    {
+      "epoch": 0.14472713864306785,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9803171450301405e-05,
+      "loss": 0.6675,
+      "step": 2355
+    },
+    {
+      "epoch": 0.14478859390363816,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9802711746217222e-05,
+      "loss": 0.6266,
+      "step": 2356
+    },
+    {
+      "epoch": 0.14485004916420846,
+      "grad_norm": 26.5,
+      "learning_rate": 1.9802251511274946e-05,
+      "loss": 0.7621,
+      "step": 2357
+    },
+    {
+      "epoch": 0.14491150442477876,
+      "grad_norm": 19.5,
+      "learning_rate": 1.980179074549951e-05,
+      "loss": 0.5386,
+      "step": 2358
+    },
+    {
+      "epoch": 0.14497295968534907,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9801329448915863e-05,
+      "loss": 0.4819,
+      "step": 2359
+    },
+    {
+      "epoch": 0.14503441494591937,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.980086762154898e-05,
+      "loss": 0.4122,
+      "step": 2360
+    },
+    {
+      "epoch": 0.14509587020648967,
+      "grad_norm": 16.625,
+      "learning_rate": 1.980040526342388e-05,
+      "loss": 0.608,
+      "step": 2361
+    },
+    {
+      "epoch": 0.14515732546705998,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9799942374565597e-05,
+      "loss": 0.5776,
+      "step": 2362
+    },
+    {
+      "epoch": 0.14521878072763028,
+      "grad_norm": 21.0,
+      "learning_rate": 1.9799478954999198e-05,
+      "loss": 0.7758,
+      "step": 2363
+    },
+    {
+      "epoch": 0.14528023598820058,
+      "grad_norm": 20.0,
+      "learning_rate": 1.979901500474978e-05,
+      "loss": 0.7533,
+      "step": 2364
+    },
+    {
+      "epoch": 0.1453416912487709,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.979855052384247e-05,
+      "loss": 0.679,
+      "step": 2365
+    },
+    {
+      "epoch": 0.1454031465093412,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.9798085512302418e-05,
+      "loss": 0.3526,
+      "step": 2366
+    },
+    {
+      "epoch": 0.1454646017699115,
+      "grad_norm": 17.375,
+      "learning_rate": 1.9797619970154807e-05,
+      "loss": 0.4703,
+      "step": 2367
+    },
+    {
+      "epoch": 0.1455260570304818,
+      "grad_norm": 15.625,
+      "learning_rate": 1.9797153897424854e-05,
+      "loss": 0.3675,
+      "step": 2368
+    },
+    {
+      "epoch": 0.1455875122910521,
+      "grad_norm": 19.0,
+      "learning_rate": 1.979668729413779e-05,
+      "loss": 0.4406,
+      "step": 2369
+    },
+    {
+      "epoch": 0.14564896755162243,
+      "grad_norm": 14.875,
+      "learning_rate": 1.9796220160318885e-05,
+      "loss": 0.5295,
+      "step": 2370
+    },
+    {
+      "epoch": 0.14571042281219274,
+      "grad_norm": 25.875,
+      "learning_rate": 1.979575249599344e-05,
+      "loss": 0.5895,
+      "step": 2371
+    },
+    {
+      "epoch": 0.14577187807276304,
+      "grad_norm": 14.625,
+      "learning_rate": 1.9795284301186782e-05,
+      "loss": 0.4618,
+      "step": 2372
+    },
+    {
+      "epoch": 0.14583333333333334,
+      "grad_norm": 28.625,
+      "learning_rate": 1.9794815575924264e-05,
+      "loss": 0.5992,
+      "step": 2373
+    },
+    {
+      "epoch": 0.14589478859390365,
+      "grad_norm": 12.1875,
+      "learning_rate": 1.9794346320231265e-05,
+      "loss": 0.3861,
+      "step": 2374
+    },
+    {
+      "epoch": 0.14595624385447395,
+      "grad_norm": 8.625,
+      "learning_rate": 1.9793876534133205e-05,
+      "loss": 0.3257,
+      "step": 2375
+    },
+    {
+      "epoch": 0.14601769911504425,
+      "grad_norm": 12.6875,
+      "learning_rate": 1.9793406217655516e-05,
+      "loss": 0.3326,
+      "step": 2376
+    },
+    {
+      "epoch": 0.14607915437561456,
+      "grad_norm": 18.125,
+      "learning_rate": 1.9792935370823676e-05,
+      "loss": 0.4156,
+      "step": 2377
+    },
+    {
+      "epoch": 0.14614060963618486,
+      "grad_norm": 18.75,
+      "learning_rate": 1.979246399366318e-05,
+      "loss": 0.4597,
+      "step": 2378
+    },
+    {
+      "epoch": 0.14620206489675516,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9791992086199552e-05,
+      "loss": 0.4928,
+      "step": 2379
+    },
+    {
+      "epoch": 0.14626352015732547,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9791519648458352e-05,
+      "loss": 0.3948,
+      "step": 2380
+    },
+    {
+      "epoch": 0.14632497541789577,
+      "grad_norm": 26.625,
+      "learning_rate": 1.979104668046516e-05,
+      "loss": 0.4023,
+      "step": 2381
+    },
+    {
+      "epoch": 0.14638643067846607,
+      "grad_norm": 15.75,
+      "learning_rate": 1.97905731822456e-05,
+      "loss": 0.4679,
+      "step": 2382
+    },
+    {
+      "epoch": 0.14644788593903638,
+      "grad_norm": 12.125,
+      "learning_rate": 1.97900991538253e-05,
+      "loss": 0.4376,
+      "step": 2383
+    },
+    {
+      "epoch": 0.14650934119960668,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9789624595229937e-05,
+      "loss": 0.4788,
+      "step": 2384
+    },
+    {
+      "epoch": 0.14657079646017698,
+      "grad_norm": 23.875,
+      "learning_rate": 1.978914950648521e-05,
+      "loss": 0.5495,
+      "step": 2385
+    },
+    {
+      "epoch": 0.1466322517207473,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.9788673887616852e-05,
+      "loss": 0.4129,
+      "step": 2386
+    },
+    {
+      "epoch": 0.1466937069813176,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.9788197738650613e-05,
+      "loss": 0.6316,
+      "step": 2387
+    },
+    {
+      "epoch": 0.1467551622418879,
+      "grad_norm": 12.125,
+      "learning_rate": 1.9787721059612276e-05,
+      "loss": 0.4781,
+      "step": 2388
+    },
+    {
+      "epoch": 0.14681661750245822,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9787243850527663e-05,
+      "loss": 0.3069,
+      "step": 2389
+    },
+    {
+      "epoch": 0.14687807276302853,
+      "grad_norm": 23.5,
+      "learning_rate": 1.9786766111422616e-05,
+      "loss": 0.3726,
+      "step": 2390
+    },
+    {
+      "epoch": 0.14693952802359883,
+      "grad_norm": 31.375,
+      "learning_rate": 1.9786287842323002e-05,
+      "loss": 0.3191,
+      "step": 2391
+    },
+    {
+      "epoch": 0.14700098328416913,
+      "grad_norm": 22.375,
+      "learning_rate": 1.978580904325472e-05,
+      "loss": 0.4062,
+      "step": 2392
+    },
+    {
+      "epoch": 0.14706243854473944,
+      "grad_norm": 49.25,
+      "learning_rate": 1.978532971424371e-05,
+      "loss": 0.4739,
+      "step": 2393
+    },
+    {
+      "epoch": 0.14712389380530974,
+      "grad_norm": 5.875,
+      "learning_rate": 1.9784849855315914e-05,
+      "loss": 0.3387,
+      "step": 2394
+    },
+    {
+      "epoch": 0.14718534906588004,
+      "grad_norm": 11.1875,
+      "learning_rate": 1.9784369466497333e-05,
+      "loss": 0.6701,
+      "step": 2395
+    },
+    {
+      "epoch": 0.14724680432645035,
+      "grad_norm": 23.375,
+      "learning_rate": 1.978388854781397e-05,
+      "loss": 0.5333,
+      "step": 2396
+    },
+    {
+      "epoch": 0.14730825958702065,
+      "grad_norm": 18.625,
+      "learning_rate": 1.978340709929188e-05,
+      "loss": 0.5456,
+      "step": 2397
+    },
+    {
+      "epoch": 0.14736971484759095,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9782925120957123e-05,
+      "loss": 0.3883,
+      "step": 2398
+    },
+    {
+      "epoch": 0.14743117010816126,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.9782442612835812e-05,
+      "loss": 0.3444,
+      "step": 2399
+    },
+    {
+      "epoch": 0.14749262536873156,
+      "grad_norm": 4.09375,
+      "learning_rate": 1.978195957495407e-05,
+      "loss": 0.3027,
+      "step": 2400
+    },
+    {
+      "epoch": 0.14755408062930186,
+      "grad_norm": 9.1875,
+      "learning_rate": 1.9781476007338058e-05,
+      "loss": 0.4413,
+      "step": 2401
+    },
+    {
+      "epoch": 0.14761553588987217,
+      "grad_norm": 30.25,
+      "learning_rate": 1.978099191001396e-05,
+      "loss": 0.5716,
+      "step": 2402
+    },
+    {
+      "epoch": 0.14767699115044247,
+      "grad_norm": 5.21875,
+      "learning_rate": 1.9780507283008e-05,
+      "loss": 0.4114,
+      "step": 2403
+    },
+    {
+      "epoch": 0.14773844641101277,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9780022126346413e-05,
+      "loss": 0.6259,
+      "step": 2404
+    },
+    {
+      "epoch": 0.14779990167158308,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9779536440055474e-05,
+      "loss": 0.5516,
+      "step": 2405
+    },
+    {
+      "epoch": 0.14786135693215338,
+      "grad_norm": 7.625,
+      "learning_rate": 1.9779050224161494e-05,
+      "loss": 0.2358,
+      "step": 2406
+    },
+    {
+      "epoch": 0.14792281219272368,
+      "grad_norm": 6.0,
+      "learning_rate": 1.977856347869079e-05,
+      "loss": 0.4169,
+      "step": 2407
+    },
+    {
+      "epoch": 0.147984267453294,
+      "grad_norm": 47.25,
+      "learning_rate": 1.977807620366973e-05,
+      "loss": 0.3913,
+      "step": 2408
+    },
+    {
+      "epoch": 0.14804572271386432,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9777588399124702e-05,
+      "loss": 0.7274,
+      "step": 2409
+    },
+    {
+      "epoch": 0.14810717797443462,
+      "grad_norm": 32.75,
+      "learning_rate": 1.977710006508212e-05,
+      "loss": 0.5983,
+      "step": 2410
+    },
+    {
+      "epoch": 0.14816863323500493,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9776611201568434e-05,
+      "loss": 0.7134,
+      "step": 2411
+    },
+    {
+      "epoch": 0.14823008849557523,
+      "grad_norm": 43.5,
+      "learning_rate": 1.977612180861011e-05,
+      "loss": 0.5775,
+      "step": 2412
+    },
+    {
+      "epoch": 0.14829154375614553,
+      "grad_norm": 23.625,
+      "learning_rate": 1.9775631886233655e-05,
+      "loss": 0.6386,
+      "step": 2413
+    },
+    {
+      "epoch": 0.14835299901671584,
+      "grad_norm": 18.625,
+      "learning_rate": 1.97751414344656e-05,
+      "loss": 0.4715,
+      "step": 2414
+    },
+    {
+      "epoch": 0.14841445427728614,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.9774650453332506e-05,
+      "loss": 0.6077,
+      "step": 2415
+    },
+    {
+      "epoch": 0.14847590953785644,
+      "grad_norm": 10.25,
+      "learning_rate": 1.9774158942860962e-05,
+      "loss": 0.6176,
+      "step": 2416
+    },
+    {
+      "epoch": 0.14853736479842675,
+      "grad_norm": 14.625,
+      "learning_rate": 1.977366690307758e-05,
+      "loss": 0.4794,
+      "step": 2417
+    },
+    {
+      "epoch": 0.14859882005899705,
+      "grad_norm": 8.75,
+      "learning_rate": 1.9773174334009016e-05,
+      "loss": 0.3786,
+      "step": 2418
+    },
+    {
+      "epoch": 0.14866027531956735,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9772681235681936e-05,
+      "loss": 0.5023,
+      "step": 2419
+    },
+    {
+      "epoch": 0.14872173058013766,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9772187608123047e-05,
+      "loss": 0.499,
+      "step": 2420
+    },
+    {
+      "epoch": 0.14878318584070796,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.977169345135908e-05,
+      "loss": 0.4382,
+      "step": 2421
+    },
+    {
+      "epoch": 0.14884464110127826,
+      "grad_norm": 30.75,
+      "learning_rate": 1.97711987654168e-05,
+      "loss": 0.4144,
+      "step": 2422
+    },
+    {
+      "epoch": 0.14890609636184857,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9770703550322986e-05,
+      "loss": 0.3927,
+      "step": 2423
+    },
+    {
+      "epoch": 0.14896755162241887,
+      "grad_norm": 6.875,
+      "learning_rate": 1.9770207806104466e-05,
+      "loss": 0.2667,
+      "step": 2424
+    },
+    {
+      "epoch": 0.14902900688298917,
+      "grad_norm": 17.0,
+      "learning_rate": 1.9769711532788083e-05,
+      "loss": 0.5883,
+      "step": 2425
+    },
+    {
+      "epoch": 0.14909046214355948,
+      "grad_norm": 4.71875,
+      "learning_rate": 1.976921473040071e-05,
+      "loss": 0.4883,
+      "step": 2426
+    },
+    {
+      "epoch": 0.14915191740412978,
+      "grad_norm": 40.5,
+      "learning_rate": 1.9768717398969258e-05,
+      "loss": 0.6213,
+      "step": 2427
+    },
+    {
+      "epoch": 0.1492133726647001,
+      "grad_norm": 60.0,
+      "learning_rate": 1.976821953852065e-05,
+      "loss": 0.5008,
+      "step": 2428
+    },
+    {
+      "epoch": 0.14927482792527041,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9767721149081852e-05,
+      "loss": 0.5369,
+      "step": 2429
+    },
+    {
+      "epoch": 0.14933628318584072,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9767222230679858e-05,
+      "loss": 0.473,
+      "step": 2430
+    },
+    {
+      "epoch": 0.14939773844641102,
+      "grad_norm": 47.25,
+      "learning_rate": 1.9766722783341682e-05,
+      "loss": 0.5133,
+      "step": 2431
+    },
+    {
+      "epoch": 0.14945919370698132,
+      "grad_norm": 34.25,
+      "learning_rate": 1.976622280709437e-05,
+      "loss": 0.5707,
+      "step": 2432
+    },
+    {
+      "epoch": 0.14952064896755163,
+      "grad_norm": 12.5,
+      "learning_rate": 1.9765722301965e-05,
+      "loss": 0.5615,
+      "step": 2433
+    },
+    {
+      "epoch": 0.14958210422812193,
+      "grad_norm": 22.625,
+      "learning_rate": 1.9765221267980675e-05,
+      "loss": 0.4909,
+      "step": 2434
+    },
+    {
+      "epoch": 0.14964355948869223,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9764719705168528e-05,
+      "loss": 0.3712,
+      "step": 2435
+    },
+    {
+      "epoch": 0.14970501474926254,
+      "grad_norm": 12.0,
+      "learning_rate": 1.976421761355572e-05,
+      "loss": 0.3455,
+      "step": 2436
+    },
+    {
+      "epoch": 0.14976647000983284,
+      "grad_norm": 29.625,
+      "learning_rate": 1.976371499316945e-05,
+      "loss": 0.4265,
+      "step": 2437
+    },
+    {
+      "epoch": 0.14982792527040315,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9763211844036928e-05,
+      "loss": 0.5158,
+      "step": 2438
+    },
+    {
+      "epoch": 0.14988938053097345,
+      "grad_norm": 22.875,
+      "learning_rate": 1.97627081661854e-05,
+      "loss": 0.5486,
+      "step": 2439
+    },
+    {
+      "epoch": 0.14995083579154375,
+      "grad_norm": 38.25,
+      "learning_rate": 1.976220395964215e-05,
+      "loss": 0.4846,
+      "step": 2440
+    },
+    {
+      "epoch": 0.15001229105211406,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9761699224434476e-05,
+      "loss": 0.6749,
+      "step": 2441
+    },
+    {
+      "epoch": 0.15007374631268436,
+      "grad_norm": 27.375,
+      "learning_rate": 1.976119396058972e-05,
+      "loss": 0.47,
+      "step": 2442
+    },
+    {
+      "epoch": 0.15013520157325466,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.9760688168135233e-05,
+      "loss": 0.4736,
+      "step": 2443
+    },
+    {
+      "epoch": 0.15019665683382497,
+      "grad_norm": 43.75,
+      "learning_rate": 1.9760181847098412e-05,
+      "loss": 0.4598,
+      "step": 2444
+    },
+    {
+      "epoch": 0.15025811209439527,
+      "grad_norm": 21.0,
+      "learning_rate": 1.975967499750668e-05,
+      "loss": 0.3586,
+      "step": 2445
+    },
+    {
+      "epoch": 0.15031956735496557,
+      "grad_norm": 37.75,
+      "learning_rate": 1.9759167619387474e-05,
+      "loss": 0.503,
+      "step": 2446
+    },
+    {
+      "epoch": 0.1503810226155359,
+      "grad_norm": 13.375,
+      "learning_rate": 1.9758659712768283e-05,
+      "loss": 0.4314,
+      "step": 2447
+    },
+    {
+      "epoch": 0.1504424778761062,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9758151277676603e-05,
+      "loss": 0.5163,
+      "step": 2448
+    },
+    {
+      "epoch": 0.1505039331366765,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9757642314139977e-05,
+      "loss": 0.458,
+      "step": 2449
+    },
+    {
+      "epoch": 0.1505653883972468,
+      "grad_norm": 27.375,
+      "learning_rate": 1.975713282218596e-05,
+      "loss": 0.5305,
+      "step": 2450
+    },
+    {
+      "epoch": 0.15062684365781712,
+      "grad_norm": 14.25,
+      "learning_rate": 1.9756622801842144e-05,
+      "loss": 0.5893,
+      "step": 2451
+    },
+    {
+      "epoch": 0.15068829891838742,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9756112253136154e-05,
+      "loss": 0.421,
+      "step": 2452
+    },
+    {
+      "epoch": 0.15074975417895772,
+      "grad_norm": 16.0,
+      "learning_rate": 1.975560117609563e-05,
+      "loss": 0.3991,
+      "step": 2453
+    },
+    {
+      "epoch": 0.15081120943952803,
+      "grad_norm": 16.875,
+      "learning_rate": 1.9755089570748255e-05,
+      "loss": 0.4059,
+      "step": 2454
+    },
+    {
+      "epoch": 0.15087266470009833,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.9754577437121733e-05,
+      "loss": 0.3474,
+      "step": 2455
+    },
+    {
+      "epoch": 0.15093411996066863,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9754064775243797e-05,
+      "loss": 0.5048,
+      "step": 2456
+    },
+    {
+      "epoch": 0.15099557522123894,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9753551585142214e-05,
+      "loss": 0.4424,
+      "step": 2457
+    },
+    {
+      "epoch": 0.15105703048180924,
+      "grad_norm": 31.0,
+      "learning_rate": 1.975303786684477e-05,
+      "loss": 0.5277,
+      "step": 2458
+    },
+    {
+      "epoch": 0.15111848574237954,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9752523620379288e-05,
+      "loss": 0.4678,
+      "step": 2459
+    },
+    {
+      "epoch": 0.15117994100294985,
+      "grad_norm": 14.75,
+      "learning_rate": 1.9752008845773618e-05,
+      "loss": 0.4845,
+      "step": 2460
+    },
+    {
+      "epoch": 0.15124139626352015,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9751493543055634e-05,
+      "loss": 0.6685,
+      "step": 2461
+    },
+    {
+      "epoch": 0.15130285152409045,
+      "grad_norm": 15.5,
+      "learning_rate": 1.9750977712253242e-05,
+      "loss": 0.4018,
+      "step": 2462
+    },
+    {
+      "epoch": 0.15136430678466076,
+      "grad_norm": 8.625,
+      "learning_rate": 1.975046135339438e-05,
+      "loss": 0.4305,
+      "step": 2463
+    },
+    {
+      "epoch": 0.15142576204523106,
+      "grad_norm": 9.875,
+      "learning_rate": 1.9749944466507007e-05,
+      "loss": 0.5154,
+      "step": 2464
+    },
+    {
+      "epoch": 0.15148721730580136,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9749427051619118e-05,
+      "loss": 0.3798,
+      "step": 2465
+    },
+    {
+      "epoch": 0.1515486725663717,
+      "grad_norm": 9.875,
+      "learning_rate": 1.9748909108758727e-05,
+      "loss": 0.5277,
+      "step": 2466
+    },
+    {
+      "epoch": 0.151610127826942,
+      "grad_norm": 17.375,
+      "learning_rate": 1.974839063795389e-05,
+      "loss": 0.5633,
+      "step": 2467
+    },
+    {
+      "epoch": 0.1516715830875123,
+      "grad_norm": 14.75,
+      "learning_rate": 1.974787163923268e-05,
+      "loss": 0.588,
+      "step": 2468
+    },
+    {
+      "epoch": 0.1517330383480826,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9747352112623207e-05,
+      "loss": 0.4127,
+      "step": 2469
+    },
+    {
+      "epoch": 0.1517944936086529,
+      "grad_norm": 7.84375,
+      "learning_rate": 1.9746832058153602e-05,
+      "loss": 0.3546,
+      "step": 2470
+    },
+    {
+      "epoch": 0.1518559488692232,
+      "grad_norm": 7.25,
+      "learning_rate": 1.9746311475852028e-05,
+      "loss": 0.3774,
+      "step": 2471
+    },
+    {
+      "epoch": 0.15191740412979352,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9745790365746678e-05,
+      "loss": 0.3119,
+      "step": 2472
+    },
+    {
+      "epoch": 0.15197885939036382,
+      "grad_norm": 28.0,
+      "learning_rate": 1.9745268727865774e-05,
+      "loss": 0.6003,
+      "step": 2473
+    },
+    {
+      "epoch": 0.15204031465093412,
+      "grad_norm": 32.75,
+      "learning_rate": 1.9744746562237563e-05,
+      "loss": 0.4319,
+      "step": 2474
+    },
+    {
+      "epoch": 0.15210176991150443,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9744223868890317e-05,
+      "loss": 0.4008,
+      "step": 2475
+    },
+    {
+      "epoch": 0.15216322517207473,
+      "grad_norm": 5.90625,
+      "learning_rate": 1.9743700647852356e-05,
+      "loss": 0.329,
+      "step": 2476
+    },
+    {
+      "epoch": 0.15222468043264503,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9743176899151998e-05,
+      "loss": 0.4819,
+      "step": 2477
+    },
+    {
+      "epoch": 0.15228613569321534,
+      "grad_norm": 26.125,
+      "learning_rate": 1.9742652622817622e-05,
+      "loss": 0.4721,
+      "step": 2478
+    },
+    {
+      "epoch": 0.15234759095378564,
+      "grad_norm": 12.8125,
+      "learning_rate": 1.9742127818877605e-05,
+      "loss": 0.5077,
+      "step": 2479
+    },
+    {
+      "epoch": 0.15240904621435594,
+      "grad_norm": 9.5,
+      "learning_rate": 1.9741602487360382e-05,
+      "loss": 0.3709,
+      "step": 2480
+    },
+    {
+      "epoch": 0.15247050147492625,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9741076628294387e-05,
+      "loss": 0.586,
+      "step": 2481
+    },
+    {
+      "epoch": 0.15253195673549655,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.974055024170811e-05,
+      "loss": 0.4819,
+      "step": 2482
+    },
+    {
+      "epoch": 0.15259341199606685,
+      "grad_norm": 35.5,
+      "learning_rate": 1.974002332763005e-05,
+      "loss": 0.4552,
+      "step": 2483
+    },
+    {
+      "epoch": 0.15265486725663716,
+      "grad_norm": 19.125,
+      "learning_rate": 1.9739495886088745e-05,
+      "loss": 0.4949,
+      "step": 2484
+    },
+    {
+      "epoch": 0.1527163225172075,
+      "grad_norm": 49.25,
+      "learning_rate": 1.9738967917112752e-05,
+      "loss": 0.5519,
+      "step": 2485
+    },
+    {
+      "epoch": 0.1527777777777778,
+      "grad_norm": 28.5,
+      "learning_rate": 1.9738439420730674e-05,
+      "loss": 0.4282,
+      "step": 2486
+    },
+    {
+      "epoch": 0.1528392330383481,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9737910396971122e-05,
+      "loss": 0.4504,
+      "step": 2487
+    },
+    {
+      "epoch": 0.1529006882989184,
+      "grad_norm": 39.5,
+      "learning_rate": 1.9737380845862745e-05,
+      "loss": 0.4778,
+      "step": 2488
+    },
+    {
+      "epoch": 0.1529621435594887,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9736850767434226e-05,
+      "loss": 0.4263,
+      "step": 2489
+    },
+    {
+      "epoch": 0.153023598820059,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9736320161714267e-05,
+      "loss": 0.4195,
+      "step": 2490
+    },
+    {
+      "epoch": 0.1530850540806293,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9735789028731603e-05,
+      "loss": 0.3562,
+      "step": 2491
+    },
+    {
+      "epoch": 0.1531465093411996,
+      "grad_norm": 31.375,
+      "learning_rate": 1.9735257368515e-05,
+      "loss": 0.5781,
+      "step": 2492
+    },
+    {
+      "epoch": 0.1532079646017699,
+      "grad_norm": 16.625,
+      "learning_rate": 1.9734725181093246e-05,
+      "loss": 0.4943,
+      "step": 2493
+    },
+    {
+      "epoch": 0.15326941986234022,
+      "grad_norm": 10.375,
+      "learning_rate": 1.9734192466495162e-05,
+      "loss": 0.4052,
+      "step": 2494
+    },
+    {
+      "epoch": 0.15333087512291052,
+      "grad_norm": 16.375,
+      "learning_rate": 1.9733659224749596e-05,
+      "loss": 0.365,
+      "step": 2495
+    },
+    {
+      "epoch": 0.15339233038348082,
+      "grad_norm": 18.625,
+      "learning_rate": 1.973312545588543e-05,
+      "loss": 0.4682,
+      "step": 2496
+    },
+    {
+      "epoch": 0.15345378564405113,
+      "grad_norm": 6.25,
+      "learning_rate": 1.9732591159931564e-05,
+      "loss": 0.3856,
+      "step": 2497
+    },
+    {
+      "epoch": 0.15351524090462143,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.9732056336916936e-05,
+      "loss": 0.6198,
+      "step": 2498
+    },
+    {
+      "epoch": 0.15357669616519173,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9731520986870505e-05,
+      "loss": 0.4293,
+      "step": 2499
+    },
+    {
+      "epoch": 0.15363815142576204,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9730985109821268e-05,
+      "loss": 0.6935,
+      "step": 2500
+    },
+    {
+      "epoch": 0.15369960668633234,
+      "grad_norm": 22.625,
+      "learning_rate": 1.973044870579824e-05,
+      "loss": 0.2953,
+      "step": 2501
+    },
+    {
+      "epoch": 0.15376106194690264,
+      "grad_norm": 40.25,
+      "learning_rate": 1.9729911774830472e-05,
+      "loss": 0.4655,
+      "step": 2502
+    },
+    {
+      "epoch": 0.15382251720747295,
+      "grad_norm": 9.5625,
+      "learning_rate": 1.972937431694704e-05,
+      "loss": 0.3424,
+      "step": 2503
+    },
+    {
+      "epoch": 0.15388397246804325,
+      "grad_norm": 31.625,
+      "learning_rate": 1.9728836332177054e-05,
+      "loss": 0.5138,
+      "step": 2504
+    },
+    {
+      "epoch": 0.15394542772861358,
+      "grad_norm": 28.25,
+      "learning_rate": 1.972829782054964e-05,
+      "loss": 0.6226,
+      "step": 2505
+    },
+    {
+      "epoch": 0.15400688298918389,
+      "grad_norm": 29.5,
+      "learning_rate": 1.972775878209397e-05,
+      "loss": 0.7932,
+      "step": 2506
+    },
+    {
+      "epoch": 0.1540683382497542,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9727219216839223e-05,
+      "loss": 0.5903,
+      "step": 2507
+    },
+    {
+      "epoch": 0.1541297935103245,
+      "grad_norm": 46.25,
+      "learning_rate": 1.9726679124814633e-05,
+      "loss": 0.7314,
+      "step": 2508
+    },
+    {
+      "epoch": 0.1541912487708948,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.9726138506049438e-05,
+      "loss": 0.6516,
+      "step": 2509
+    },
+    {
+      "epoch": 0.1542527040314651,
+      "grad_norm": 36.25,
+      "learning_rate": 1.9725597360572918e-05,
+      "loss": 0.8025,
+      "step": 2510
+    },
+    {
+      "epoch": 0.1543141592920354,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.9725055688414378e-05,
+      "loss": 0.6405,
+      "step": 2511
+    },
+    {
+      "epoch": 0.1543756145526057,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9724513489603153e-05,
+      "loss": 0.6724,
+      "step": 2512
+    },
+    {
+      "epoch": 0.154437069813176,
+      "grad_norm": 35.0,
+      "learning_rate": 1.9723970764168606e-05,
+      "loss": 0.4985,
+      "step": 2513
+    },
+    {
+      "epoch": 0.1544985250737463,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.9723427512140124e-05,
+      "loss": 0.4238,
+      "step": 2514
+    },
+    {
+      "epoch": 0.15455998033431662,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9722883733547128e-05,
+      "loss": 0.4777,
+      "step": 2515
+    },
+    {
+      "epoch": 0.15462143559488692,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.972233942841907e-05,
+      "loss": 0.3481,
+      "step": 2516
+    },
+    {
+      "epoch": 0.15468289085545722,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.9721794596785423e-05,
+      "loss": 0.4643,
+      "step": 2517
+    },
+    {
+      "epoch": 0.15474434611602753,
+      "grad_norm": 8.625,
+      "learning_rate": 1.9721249238675688e-05,
+      "loss": 0.3452,
+      "step": 2518
+    },
+    {
+      "epoch": 0.15480580137659783,
+      "grad_norm": 8.125,
+      "learning_rate": 1.9720703354119407e-05,
+      "loss": 0.3795,
+      "step": 2519
+    },
+    {
+      "epoch": 0.15486725663716813,
+      "grad_norm": 24.0,
+      "learning_rate": 1.9720156943146133e-05,
+      "loss": 0.4835,
+      "step": 2520
+    },
+    {
+      "epoch": 0.15492871189773844,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9719610005785466e-05,
+      "loss": 0.5871,
+      "step": 2521
+    },
+    {
+      "epoch": 0.15499016715830874,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.9719062542067014e-05,
+      "loss": 0.5449,
+      "step": 2522
+    },
+    {
+      "epoch": 0.15505162241887904,
+      "grad_norm": 20.375,
+      "learning_rate": 1.9718514552020434e-05,
+      "loss": 0.4589,
+      "step": 2523
+    },
+    {
+      "epoch": 0.15511307767944937,
+      "grad_norm": 13.75,
+      "learning_rate": 1.97179660356754e-05,
+      "loss": 0.5105,
+      "step": 2524
+    },
+    {
+      "epoch": 0.15517453294001968,
+      "grad_norm": 18.5,
+      "learning_rate": 1.971741699306161e-05,
+      "loss": 0.4686,
+      "step": 2525
+    },
+    {
+      "epoch": 0.15523598820058998,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9716867424208805e-05,
+      "loss": 0.539,
+      "step": 2526
+    },
+    {
+      "epoch": 0.15529744346116028,
+      "grad_norm": 20.75,
+      "learning_rate": 1.971631732914674e-05,
+      "loss": 0.4935,
+      "step": 2527
+    },
+    {
+      "epoch": 0.1553588987217306,
+      "grad_norm": 6.46875,
+      "learning_rate": 1.971576670790521e-05,
+      "loss": 0.3302,
+      "step": 2528
+    },
+    {
+      "epoch": 0.1554203539823009,
+      "grad_norm": 16.75,
+      "learning_rate": 1.9715215560514032e-05,
+      "loss": 0.5374,
+      "step": 2529
+    },
+    {
+      "epoch": 0.1554818092428712,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9714663887003055e-05,
+      "loss": 0.6986,
+      "step": 2530
+    },
+    {
+      "epoch": 0.1555432645034415,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9714111687402146e-05,
+      "loss": 0.7489,
+      "step": 2531
+    },
+    {
+      "epoch": 0.1556047197640118,
+      "grad_norm": 22.375,
+      "learning_rate": 1.971355896174122e-05,
+      "loss": 0.646,
+      "step": 2532
+    },
+    {
+      "epoch": 0.1556661750245821,
+      "grad_norm": 29.875,
+      "learning_rate": 1.9713005710050203e-05,
+      "loss": 0.4847,
+      "step": 2533
+    },
+    {
+      "epoch": 0.1557276302851524,
+      "grad_norm": 16.75,
+      "learning_rate": 1.9712451932359056e-05,
+      "loss": 0.4794,
+      "step": 2534
+    },
+    {
+      "epoch": 0.1557890855457227,
+      "grad_norm": 21.75,
+      "learning_rate": 1.971189762869777e-05,
+      "loss": 0.5296,
+      "step": 2535
+    },
+    {
+      "epoch": 0.15585054080629301,
+      "grad_norm": 13.5,
+      "learning_rate": 1.971134279909636e-05,
+      "loss": 0.4521,
+      "step": 2536
+    },
+    {
+      "epoch": 0.15591199606686332,
+      "grad_norm": 24.375,
+      "learning_rate": 1.971078744358488e-05,
+      "loss": 0.5604,
+      "step": 2537
+    },
+    {
+      "epoch": 0.15597345132743362,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9710231562193397e-05,
+      "loss": 0.6101,
+      "step": 2538
+    },
+    {
+      "epoch": 0.15603490658800392,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9709675154952017e-05,
+      "loss": 0.6847,
+      "step": 2539
+    },
+    {
+      "epoch": 0.15609636184857423,
+      "grad_norm": 14.75,
+      "learning_rate": 1.9709118221890872e-05,
+      "loss": 0.5018,
+      "step": 2540
+    },
+    {
+      "epoch": 0.15615781710914453,
+      "grad_norm": 36.25,
+      "learning_rate": 1.970856076304012e-05,
+      "loss": 0.7679,
+      "step": 2541
+    },
+    {
+      "epoch": 0.15621927236971483,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9708002778429957e-05,
+      "loss": 0.4504,
+      "step": 2542
+    },
+    {
+      "epoch": 0.15628072763028517,
+      "grad_norm": 12.0,
+      "learning_rate": 1.970744426809059e-05,
+      "loss": 0.4568,
+      "step": 2543
+    },
+    {
+      "epoch": 0.15634218289085547,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9706885232052274e-05,
+      "loss": 0.4079,
+      "step": 2544
+    },
+    {
+      "epoch": 0.15640363815142577,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9706325670345276e-05,
+      "loss": 0.6281,
+      "step": 2545
+    },
+    {
+      "epoch": 0.15646509341199608,
+      "grad_norm": 17.375,
+      "learning_rate": 1.97057655829999e-05,
+      "loss": 0.7279,
+      "step": 2546
+    },
+    {
+      "epoch": 0.15652654867256638,
+      "grad_norm": 13.25,
+      "learning_rate": 1.970520497004648e-05,
+      "loss": 0.624,
+      "step": 2547
+    },
+    {
+      "epoch": 0.15658800393313668,
+      "grad_norm": 17.875,
+      "learning_rate": 1.9704643831515377e-05,
+      "loss": 0.5684,
+      "step": 2548
+    },
+    {
+      "epoch": 0.15664945919370699,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9704082167436974e-05,
+      "loss": 0.5858,
+      "step": 2549
+    },
+    {
+      "epoch": 0.1567109144542773,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9703519977841686e-05,
+      "loss": 0.4229,
+      "step": 2550
+    },
+    {
+      "epoch": 0.1567723697148476,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.9702957262759964e-05,
+      "loss": 0.3657,
+      "step": 2551
+    },
+    {
+      "epoch": 0.1568338249754179,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9702394022222282e-05,
+      "loss": 0.3621,
+      "step": 2552
+    },
+    {
+      "epoch": 0.1568952802359882,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.9701830256259136e-05,
+      "loss": 0.4467,
+      "step": 2553
+    },
+    {
+      "epoch": 0.1569567354965585,
+      "grad_norm": 20.125,
+      "learning_rate": 1.970126596490106e-05,
+      "loss": 0.4679,
+      "step": 2554
+    },
+    {
+      "epoch": 0.1570181907571288,
+      "grad_norm": 21.625,
+      "learning_rate": 1.970070114817861e-05,
+      "loss": 0.5566,
+      "step": 2555
+    },
+    {
+      "epoch": 0.1570796460176991,
+      "grad_norm": 17.125,
+      "learning_rate": 1.9700135806122378e-05,
+      "loss": 0.4786,
+      "step": 2556
+    },
+    {
+      "epoch": 0.1571411012782694,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.9699569938762975e-05,
+      "loss": 0.3824,
+      "step": 2557
+    },
+    {
+      "epoch": 0.15720255653883972,
+      "grad_norm": 3.828125,
+      "learning_rate": 1.969900354613105e-05,
+      "loss": 0.2935,
+      "step": 2558
+    },
+    {
+      "epoch": 0.15726401179941002,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.9698436628257266e-05,
+      "loss": 0.5448,
+      "step": 2559
+    },
+    {
+      "epoch": 0.15732546705998032,
+      "grad_norm": 13.875,
+      "learning_rate": 1.969786918517233e-05,
+      "loss": 0.3638,
+      "step": 2560
+    },
+    {
+      "epoch": 0.15738692232055063,
+      "grad_norm": 20.25,
+      "learning_rate": 1.969730121690698e-05,
+      "loss": 0.3758,
+      "step": 2561
+    },
+    {
+      "epoch": 0.15744837758112096,
+      "grad_norm": 19.125,
+      "learning_rate": 1.969673272349196e-05,
+      "loss": 0.4196,
+      "step": 2562
+    },
+    {
+      "epoch": 0.15750983284169126,
+      "grad_norm": 20.125,
+      "learning_rate": 1.969616370495806e-05,
+      "loss": 0.6692,
+      "step": 2563
+    },
+    {
+      "epoch": 0.15757128810226156,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9695594161336102e-05,
+      "loss": 0.4147,
+      "step": 2564
+    },
+    {
+      "epoch": 0.15763274336283187,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.969502409265692e-05,
+      "loss": 0.3686,
+      "step": 2565
+    },
+    {
+      "epoch": 0.15769419862340217,
+      "grad_norm": 5.6875,
+      "learning_rate": 1.9694453498951392e-05,
+      "loss": 0.4473,
+      "step": 2566
+    },
+    {
+      "epoch": 0.15775565388397247,
+      "grad_norm": 6.375,
+      "learning_rate": 1.9693882380250412e-05,
+      "loss": 0.2839,
+      "step": 2567
+    },
+    {
+      "epoch": 0.15781710914454278,
+      "grad_norm": 34.75,
+      "learning_rate": 1.9693310736584913e-05,
+      "loss": 0.3555,
+      "step": 2568
+    },
+    {
+      "epoch": 0.15787856440511308,
+      "grad_norm": 39.5,
+      "learning_rate": 1.9692738567985853e-05,
+      "loss": 0.383,
+      "step": 2569
+    },
+    {
+      "epoch": 0.15794001966568338,
+      "grad_norm": 11.125,
+      "learning_rate": 1.9692165874484216e-05,
+      "loss": 0.5434,
+      "step": 2570
+    },
+    {
+      "epoch": 0.1580014749262537,
+      "grad_norm": 20.75,
+      "learning_rate": 1.969159265611101e-05,
+      "loss": 0.3975,
+      "step": 2571
+    },
+    {
+      "epoch": 0.158062930186824,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9691018912897285e-05,
+      "loss": 0.4194,
+      "step": 2572
+    },
+    {
+      "epoch": 0.1581243854473943,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9690444644874112e-05,
+      "loss": 0.3679,
+      "step": 2573
+    },
+    {
+      "epoch": 0.1581858407079646,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.9689869852072584e-05,
+      "loss": 0.4648,
+      "step": 2574
+    },
+    {
+      "epoch": 0.1582472959685349,
+      "grad_norm": 16.75,
+      "learning_rate": 1.968929453452383e-05,
+      "loss": 0.3546,
+      "step": 2575
+    },
+    {
+      "epoch": 0.1583087512291052,
+      "grad_norm": 12.625,
+      "learning_rate": 1.9688718692259007e-05,
+      "loss": 0.6424,
+      "step": 2576
+    },
+    {
+      "epoch": 0.1583702064896755,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9688142325309302e-05,
+      "loss": 0.6393,
+      "step": 2577
+    },
+    {
+      "epoch": 0.1584316617502458,
+      "grad_norm": 44.0,
+      "learning_rate": 1.9687565433705926e-05,
+      "loss": 0.7496,
+      "step": 2578
+    },
+    {
+      "epoch": 0.15849311701081611,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9686988017480114e-05,
+      "loss": 0.5012,
+      "step": 2579
+    },
+    {
+      "epoch": 0.15855457227138642,
+      "grad_norm": 28.0,
+      "learning_rate": 1.968641007666315e-05,
+      "loss": 0.4225,
+      "step": 2580
+    },
+    {
+      "epoch": 0.15861602753195675,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.9685831611286312e-05,
+      "loss": 0.507,
+      "step": 2581
+    },
+    {
+      "epoch": 0.15867748279252705,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9685252621380942e-05,
+      "loss": 0.5538,
+      "step": 2582
+    },
+    {
+      "epoch": 0.15873893805309736,
+      "grad_norm": 31.625,
+      "learning_rate": 1.9684673106978392e-05,
+      "loss": 0.4698,
+      "step": 2583
+    },
+    {
+      "epoch": 0.15880039331366766,
+      "grad_norm": 23.75,
+      "learning_rate": 1.968409306811004e-05,
+      "loss": 0.4338,
+      "step": 2584
+    },
+    {
+      "epoch": 0.15886184857423796,
+      "grad_norm": 40.25,
+      "learning_rate": 1.9683512504807298e-05,
+      "loss": 0.5239,
+      "step": 2585
+    },
+    {
+      "epoch": 0.15892330383480827,
+      "grad_norm": 37.5,
+      "learning_rate": 1.968293141710161e-05,
+      "loss": 0.491,
+      "step": 2586
+    },
+    {
+      "epoch": 0.15898475909537857,
+      "grad_norm": 30.625,
+      "learning_rate": 1.9682349805024447e-05,
+      "loss": 0.4252,
+      "step": 2587
+    },
+    {
+      "epoch": 0.15904621435594887,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9681767668607298e-05,
+      "loss": 0.4828,
+      "step": 2588
+    },
+    {
+      "epoch": 0.15910766961651918,
+      "grad_norm": 17.75,
+      "learning_rate": 1.968118500788169e-05,
+      "loss": 0.4736,
+      "step": 2589
+    },
+    {
+      "epoch": 0.15916912487708948,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.968060182287918e-05,
+      "loss": 0.4417,
+      "step": 2590
+    },
+    {
+      "epoch": 0.15923058013765978,
+      "grad_norm": 10.0,
+      "learning_rate": 1.9680018113631347e-05,
+      "loss": 0.4989,
+      "step": 2591
+    },
+    {
+      "epoch": 0.1592920353982301,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9679433880169805e-05,
+      "loss": 0.3964,
+      "step": 2592
+    },
+    {
+      "epoch": 0.1593534906588004,
+      "grad_norm": 29.75,
+      "learning_rate": 1.967884912252619e-05,
+      "loss": 0.4144,
+      "step": 2593
+    },
+    {
+      "epoch": 0.1594149459193707,
+      "grad_norm": 5.53125,
+      "learning_rate": 1.9678263840732167e-05,
+      "loss": 0.6681,
+      "step": 2594
+    },
+    {
+      "epoch": 0.159476401179941,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9677678034819433e-05,
+      "loss": 0.6497,
+      "step": 2595
+    },
+    {
+      "epoch": 0.1595378564405113,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9677091704819714e-05,
+      "loss": 0.4177,
+      "step": 2596
+    },
+    {
+      "epoch": 0.1595993117010816,
+      "grad_norm": 13.5,
+      "learning_rate": 1.967650485076476e-05,
+      "loss": 0.3584,
+      "step": 2597
+    },
+    {
+      "epoch": 0.1596607669616519,
+      "grad_norm": 10.0,
+      "learning_rate": 1.9675917472686352e-05,
+      "loss": 0.3692,
+      "step": 2598
+    },
+    {
+      "epoch": 0.1597222222222222,
+      "grad_norm": 11.375,
+      "learning_rate": 1.96753295706163e-05,
+      "loss": 0.3856,
+      "step": 2599
+    },
+    {
+      "epoch": 0.1597836774827925,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9674741144586442e-05,
+      "loss": 0.4771,
+      "step": 2600
+    },
+    {
+      "epoch": 0.15984513274336284,
+      "grad_norm": 25.375,
+      "learning_rate": 1.967415219462864e-05,
+      "loss": 0.5236,
+      "step": 2601
+    },
+    {
+      "epoch": 0.15990658800393315,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9673562720774792e-05,
+      "loss": 0.402,
+      "step": 2602
+    },
+    {
+      "epoch": 0.15996804326450345,
+      "grad_norm": 32.75,
+      "learning_rate": 1.967297272305682e-05,
+      "loss": 0.4073,
+      "step": 2603
+    },
+    {
+      "epoch": 0.16002949852507375,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9672382201506674e-05,
+      "loss": 0.4683,
+      "step": 2604
+    },
+    {
+      "epoch": 0.16009095378564406,
+      "grad_norm": 20.625,
+      "learning_rate": 1.967179115615633e-05,
+      "loss": 0.5189,
+      "step": 2605
+    },
+    {
+      "epoch": 0.16015240904621436,
+      "grad_norm": 16.25,
+      "learning_rate": 1.96711995870378e-05,
+      "loss": 0.3884,
+      "step": 2606
+    },
+    {
+      "epoch": 0.16021386430678466,
+      "grad_norm": 5.1875,
+      "learning_rate": 1.967060749418312e-05,
+      "loss": 0.4963,
+      "step": 2607
+    },
+    {
+      "epoch": 0.16027531956735497,
+      "grad_norm": 5.78125,
+      "learning_rate": 1.9670014877624353e-05,
+      "loss": 0.4457,
+      "step": 2608
+    },
+    {
+      "epoch": 0.16033677482792527,
+      "grad_norm": 23.625,
+      "learning_rate": 1.966942173739359e-05,
+      "loss": 0.4183,
+      "step": 2609
+    },
+    {
+      "epoch": 0.16039823008849557,
+      "grad_norm": 27.125,
+      "learning_rate": 1.9668828073522955e-05,
+      "loss": 0.4612,
+      "step": 2610
+    },
+    {
+      "epoch": 0.16045968534906588,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9668233886044597e-05,
+      "loss": 0.5442,
+      "step": 2611
+    },
+    {
+      "epoch": 0.16052114060963618,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.9667639174990693e-05,
+      "loss": 0.3133,
+      "step": 2612
+    },
+    {
+      "epoch": 0.16058259587020648,
+      "grad_norm": 29.75,
+      "learning_rate": 1.9667043940393445e-05,
+      "loss": 0.5385,
+      "step": 2613
+    },
+    {
+      "epoch": 0.1606440511307768,
+      "grad_norm": 26.75,
+      "learning_rate": 1.9666448182285095e-05,
+      "loss": 0.4815,
+      "step": 2614
+    },
+    {
+      "epoch": 0.1607055063913471,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9665851900697902e-05,
+      "loss": 0.4334,
+      "step": 2615
+    },
+    {
+      "epoch": 0.1607669616519174,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.9665255095664155e-05,
+      "loss": 0.3862,
+      "step": 2616
+    },
+    {
+      "epoch": 0.1608284169124877,
+      "grad_norm": 51.75,
+      "learning_rate": 1.9664657767216176e-05,
+      "loss": 0.6288,
+      "step": 2617
+    },
+    {
+      "epoch": 0.160889872173058,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9664059915386316e-05,
+      "loss": 0.4779,
+      "step": 2618
+    },
+    {
+      "epoch": 0.1609513274336283,
+      "grad_norm": 25.125,
+      "learning_rate": 1.9663461540206947e-05,
+      "loss": 0.6359,
+      "step": 2619
+    },
+    {
+      "epoch": 0.16101278269419864,
+      "grad_norm": 20.0,
+      "learning_rate": 1.966286264171047e-05,
+      "loss": 0.4382,
+      "step": 2620
+    },
+    {
+      "epoch": 0.16107423795476894,
+      "grad_norm": 33.25,
+      "learning_rate": 1.966226321992933e-05,
+      "loss": 0.5504,
+      "step": 2621
+    },
+    {
+      "epoch": 0.16113569321533924,
+      "grad_norm": 22.75,
+      "learning_rate": 1.966166327489598e-05,
+      "loss": 0.5014,
+      "step": 2622
+    },
+    {
+      "epoch": 0.16119714847590955,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9661062806642903e-05,
+      "loss": 0.5305,
+      "step": 2623
+    },
+    {
+      "epoch": 0.16125860373647985,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9660461815202627e-05,
+      "loss": 0.582,
+      "step": 2624
+    },
+    {
+      "epoch": 0.16132005899705015,
+      "grad_norm": 26.375,
+      "learning_rate": 1.9659860300607695e-05,
+      "loss": 0.6754,
+      "step": 2625
+    },
+    {
+      "epoch": 0.16138151425762046,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9659258262890683e-05,
+      "loss": 0.7695,
+      "step": 2626
+    },
+    {
+      "epoch": 0.16144296951819076,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9658655702084192e-05,
+      "loss": 0.6172,
+      "step": 2627
+    },
+    {
+      "epoch": 0.16150442477876106,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9658052618220857e-05,
+      "loss": 0.6578,
+      "step": 2628
+    },
+    {
+      "epoch": 0.16156588003933137,
+      "grad_norm": 21.875,
+      "learning_rate": 1.9657449011333328e-05,
+      "loss": 0.6463,
+      "step": 2629
+    },
+    {
+      "epoch": 0.16162733529990167,
+      "grad_norm": 29.125,
+      "learning_rate": 1.9656844881454304e-05,
+      "loss": 0.6031,
+      "step": 2630
+    },
+    {
+      "epoch": 0.16168879056047197,
+      "grad_norm": 9.375,
+      "learning_rate": 1.9656240228616496e-05,
+      "loss": 0.5417,
+      "step": 2631
+    },
+    {
+      "epoch": 0.16175024582104228,
+      "grad_norm": 18.125,
+      "learning_rate": 1.9655635052852648e-05,
+      "loss": 0.4604,
+      "step": 2632
+    },
+    {
+      "epoch": 0.16181170108161258,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.9655029354195534e-05,
+      "loss": 0.4965,
+      "step": 2633
+    },
+    {
+      "epoch": 0.16187315634218288,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9654423132677956e-05,
+      "loss": 0.4449,
+      "step": 2634
+    },
+    {
+      "epoch": 0.1619346116027532,
+      "grad_norm": 20.875,
+      "learning_rate": 1.965381638833274e-05,
+      "loss": 0.4853,
+      "step": 2635
+    },
+    {
+      "epoch": 0.1619960668633235,
+      "grad_norm": 13.0,
+      "learning_rate": 1.9653209121192747e-05,
+      "loss": 0.4964,
+      "step": 2636
+    },
+    {
+      "epoch": 0.1620575221238938,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.9652601331290864e-05,
+      "loss": 0.4388,
+      "step": 2637
+    },
+    {
+      "epoch": 0.1621189773844641,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9651993018660002e-05,
+      "loss": 0.3994,
+      "step": 2638
+    },
+    {
+      "epoch": 0.16218043264503443,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9651384183333105e-05,
+      "loss": 0.392,
+      "step": 2639
+    },
+    {
+      "epoch": 0.16224188790560473,
+      "grad_norm": 15.875,
+      "learning_rate": 1.965077482534314e-05,
+      "loss": 0.3243,
+      "step": 2640
+    },
+    {
+      "epoch": 0.16230334316617503,
+      "grad_norm": 11.125,
+      "learning_rate": 1.9650164944723116e-05,
+      "loss": 0.2711,
+      "step": 2641
+    },
+    {
+      "epoch": 0.16236479842674534,
+      "grad_norm": 7.3125,
+      "learning_rate": 1.9649554541506056e-05,
+      "loss": 0.3926,
+      "step": 2642
+    },
+    {
+      "epoch": 0.16242625368731564,
+      "grad_norm": 6.125,
+      "learning_rate": 1.9648943615725012e-05,
+      "loss": 0.407,
+      "step": 2643
+    },
+    {
+      "epoch": 0.16248770894788594,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9648332167413067e-05,
+      "loss": 0.3339,
+      "step": 2644
+    },
+    {
+      "epoch": 0.16254916420845625,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9647720196603343e-05,
+      "loss": 0.3392,
+      "step": 2645
+    },
+    {
+      "epoch": 0.16261061946902655,
+      "grad_norm": 17.5,
+      "learning_rate": 1.964710770332898e-05,
+      "loss": 0.1862,
+      "step": 2646
+    },
+    {
+      "epoch": 0.16267207472959685,
+      "grad_norm": 8.125,
+      "learning_rate": 1.9646494687623135e-05,
+      "loss": 0.2309,
+      "step": 2647
+    },
+    {
+      "epoch": 0.16273352999016716,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9645881149519014e-05,
+      "loss": 0.2427,
+      "step": 2648
+    },
+    {
+      "epoch": 0.16279498525073746,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9645267089049844e-05,
+      "loss": 0.7125,
+      "step": 2649
+    },
+    {
+      "epoch": 0.16285644051130777,
+      "grad_norm": 24.75,
+      "learning_rate": 1.9644652506248872e-05,
+      "loss": 0.4653,
+      "step": 2650
+    },
+    {
+      "epoch": 0.16291789577187807,
+      "grad_norm": 15.875,
+      "learning_rate": 1.964403740114939e-05,
+      "loss": 0.4796,
+      "step": 2651
+    },
+    {
+      "epoch": 0.16297935103244837,
+      "grad_norm": 15.25,
+      "learning_rate": 1.9643421773784702e-05,
+      "loss": 0.4809,
+      "step": 2652
+    },
+    {
+      "epoch": 0.16304080629301868,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.964280562418815e-05,
+      "loss": 0.6492,
+      "step": 2653
+    },
+    {
+      "epoch": 0.16310226155358898,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9642188952393096e-05,
+      "loss": 0.4483,
+      "step": 2654
+    },
+    {
+      "epoch": 0.16316371681415928,
+      "grad_norm": 7.21875,
+      "learning_rate": 1.964157175843294e-05,
+      "loss": 0.4374,
+      "step": 2655
+    },
+    {
+      "epoch": 0.16322517207472959,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.96409540423411e-05,
+      "loss": 0.3426,
+      "step": 2656
+    },
+    {
+      "epoch": 0.1632866273352999,
+      "grad_norm": 48.0,
+      "learning_rate": 1.964033580415104e-05,
+      "loss": 0.4748,
+      "step": 2657
+    },
+    {
+      "epoch": 0.16334808259587022,
+      "grad_norm": 8.625,
+      "learning_rate": 1.9639717043896226e-05,
+      "loss": 0.4228,
+      "step": 2658
+    },
+    {
+      "epoch": 0.16340953785644052,
+      "grad_norm": 13.3125,
+      "learning_rate": 1.9639097761610174e-05,
+      "loss": 0.6938,
+      "step": 2659
+    },
+    {
+      "epoch": 0.16347099311701083,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9638477957326424e-05,
+      "loss": 0.5007,
+      "step": 2660
+    },
+    {
+      "epoch": 0.16353244837758113,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.9637857631078532e-05,
+      "loss": 0.3132,
+      "step": 2661
+    },
+    {
+      "epoch": 0.16359390363815143,
+      "grad_norm": 39.5,
+      "learning_rate": 1.96372367829001e-05,
+      "loss": 0.5564,
+      "step": 2662
+    },
+    {
+      "epoch": 0.16365535889872174,
+      "grad_norm": 24.75,
+      "learning_rate": 1.963661541282474e-05,
+      "loss": 0.5625,
+      "step": 2663
+    },
+    {
+      "epoch": 0.16371681415929204,
+      "grad_norm": 18.0,
+      "learning_rate": 1.9635993520886113e-05,
+      "loss": 0.4031,
+      "step": 2664
+    },
+    {
+      "epoch": 0.16377826941986234,
+      "grad_norm": 38.0,
+      "learning_rate": 1.963537110711789e-05,
+      "loss": 0.505,
+      "step": 2665
+    },
+    {
+      "epoch": 0.16383972468043265,
+      "grad_norm": 50.5,
+      "learning_rate": 1.9634748171553775e-05,
+      "loss": 0.6045,
+      "step": 2666
+    },
+    {
+      "epoch": 0.16390117994100295,
+      "grad_norm": 53.0,
+      "learning_rate": 1.9634124714227507e-05,
+      "loss": 0.7988,
+      "step": 2667
+    },
+    {
+      "epoch": 0.16396263520157325,
+      "grad_norm": 27.375,
+      "learning_rate": 1.963350073517285e-05,
+      "loss": 0.6065,
+      "step": 2668
+    },
+    {
+      "epoch": 0.16402409046214356,
+      "grad_norm": 7.8125,
+      "learning_rate": 1.9632876234423592e-05,
+      "loss": 0.3484,
+      "step": 2669
+    },
+    {
+      "epoch": 0.16408554572271386,
+      "grad_norm": 9.5625,
+      "learning_rate": 1.9632251212013554e-05,
+      "loss": 0.3596,
+      "step": 2670
+    },
+    {
+      "epoch": 0.16414700098328416,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.9631625667976584e-05,
+      "loss": 0.3507,
+      "step": 2671
+    },
+    {
+      "epoch": 0.16420845624385447,
+      "grad_norm": 9.25,
+      "learning_rate": 1.9630999602346558e-05,
+      "loss": 0.3919,
+      "step": 2672
+    },
+    {
+      "epoch": 0.16426991150442477,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9630373015157374e-05,
+      "loss": 0.6056,
+      "step": 2673
+    },
+    {
+      "epoch": 0.16433136676499507,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.9629745906442973e-05,
+      "loss": 0.4807,
+      "step": 2674
+    },
+    {
+      "epoch": 0.16439282202556538,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.962911827623731e-05,
+      "loss": 0.4916,
+      "step": 2675
+    },
+    {
+      "epoch": 0.16445427728613568,
+      "grad_norm": 26.5,
+      "learning_rate": 1.962849012457438e-05,
+      "loss": 0.5823,
+      "step": 2676
+    },
+    {
+      "epoch": 0.164515732546706,
+      "grad_norm": 40.0,
+      "learning_rate": 1.962786145148819e-05,
+      "loss": 0.5755,
+      "step": 2677
+    },
+    {
+      "epoch": 0.16457718780727632,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9627232257012793e-05,
+      "loss": 0.7444,
+      "step": 2678
+    },
+    {
+      "epoch": 0.16463864306784662,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9626602541182264e-05,
+      "loss": 0.4651,
+      "step": 2679
+    },
+    {
+      "epoch": 0.16470009832841692,
+      "grad_norm": 41.75,
+      "learning_rate": 1.9625972304030697e-05,
+      "loss": 0.6109,
+      "step": 2680
+    },
+    {
+      "epoch": 0.16476155358898723,
+      "grad_norm": 11.75,
+      "learning_rate": 1.9625341545592226e-05,
+      "loss": 0.5222,
+      "step": 2681
+    },
+    {
+      "epoch": 0.16482300884955753,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9624710265901012e-05,
+      "loss": 0.442,
+      "step": 2682
+    },
+    {
+      "epoch": 0.16488446411012783,
+      "grad_norm": 14.625,
+      "learning_rate": 1.962407846499124e-05,
+      "loss": 0.44,
+      "step": 2683
+    },
+    {
+      "epoch": 0.16494591937069814,
+      "grad_norm": 25.125,
+      "learning_rate": 1.9623446142897123e-05,
+      "loss": 0.5404,
+      "step": 2684
+    },
+    {
+      "epoch": 0.16500737463126844,
+      "grad_norm": 42.0,
+      "learning_rate": 1.9622813299652904e-05,
+      "loss": 0.554,
+      "step": 2685
+    },
+    {
+      "epoch": 0.16506882989183874,
+      "grad_norm": 6.0,
+      "learning_rate": 1.9622179935292855e-05,
+      "loss": 0.2637,
+      "step": 2686
+    },
+    {
+      "epoch": 0.16513028515240905,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9621546049851275e-05,
+      "loss": 0.478,
+      "step": 2687
+    },
+    {
+      "epoch": 0.16519174041297935,
+      "grad_norm": 5.34375,
+      "learning_rate": 1.9620911643362493e-05,
+      "loss": 0.2741,
+      "step": 2688
+    },
+    {
+      "epoch": 0.16525319567354965,
+      "grad_norm": 46.0,
+      "learning_rate": 1.962027671586086e-05,
+      "loss": 0.729,
+      "step": 2689
+    },
+    {
+      "epoch": 0.16531465093411996,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9619641267380766e-05,
+      "loss": 0.6168,
+      "step": 2690
+    },
+    {
+      "epoch": 0.16537610619469026,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9619005297956623e-05,
+      "loss": 0.7338,
+      "step": 2691
+    },
+    {
+      "epoch": 0.16543756145526056,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9618368807622863e-05,
+      "loss": 0.4829,
+      "step": 2692
+    },
+    {
+      "epoch": 0.16549901671583087,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9617731796413962e-05,
+      "loss": 0.5661,
+      "step": 2693
+    },
+    {
+      "epoch": 0.16556047197640117,
+      "grad_norm": 17.375,
+      "learning_rate": 1.9617094264364417e-05,
+      "loss": 0.5333,
+      "step": 2694
+    },
+    {
+      "epoch": 0.16562192723697147,
+      "grad_norm": 20.375,
+      "learning_rate": 1.9616456211508756e-05,
+      "loss": 0.6185,
+      "step": 2695
+    },
+    {
+      "epoch": 0.16568338249754178,
+      "grad_norm": 20.25,
+      "learning_rate": 1.961581763788152e-05,
+      "loss": 0.5673,
+      "step": 2696
+    },
+    {
+      "epoch": 0.1657448377581121,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.9615178543517303e-05,
+      "loss": 0.4576,
+      "step": 2697
+    },
+    {
+      "epoch": 0.1658062930186824,
+      "grad_norm": 16.375,
+      "learning_rate": 1.961453892845071e-05,
+      "loss": 0.4411,
+      "step": 2698
+    },
+    {
+      "epoch": 0.1658677482792527,
+      "grad_norm": 14.625,
+      "learning_rate": 1.9613898792716377e-05,
+      "loss": 0.4168,
+      "step": 2699
+    },
+    {
+      "epoch": 0.16592920353982302,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9613258136348973e-05,
+      "loss": 0.4792,
+      "step": 2700
+    },
+    {
+      "epoch": 0.16599065880039332,
+      "grad_norm": 10.875,
+      "learning_rate": 1.961261695938319e-05,
+      "loss": 0.4795,
+      "step": 2701
+    },
+    {
+      "epoch": 0.16605211406096362,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9611975261853754e-05,
+      "loss": 0.5371,
+      "step": 2702
+    },
+    {
+      "epoch": 0.16611356932153393,
+      "grad_norm": 13.0,
+      "learning_rate": 1.961133304379541e-05,
+      "loss": 0.4912,
+      "step": 2703
+    },
+    {
+      "epoch": 0.16617502458210423,
+      "grad_norm": 17.75,
+      "learning_rate": 1.961069030524294e-05,
+      "loss": 0.5157,
+      "step": 2704
+    },
+    {
+      "epoch": 0.16623647984267453,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9610047046231152e-05,
+      "loss": 0.5773,
+      "step": 2705
+    },
+    {
+      "epoch": 0.16629793510324484,
+      "grad_norm": 15.5625,
+      "learning_rate": 1.960940326679488e-05,
+      "loss": 0.4499,
+      "step": 2706
+    },
+    {
+      "epoch": 0.16635939036381514,
+      "grad_norm": 7.125,
+      "learning_rate": 1.9608758966968987e-05,
+      "loss": 0.408,
+      "step": 2707
+    },
+    {
+      "epoch": 0.16642084562438544,
+      "grad_norm": 108.0,
+      "learning_rate": 1.9608114146788365e-05,
+      "loss": 0.5437,
+      "step": 2708
+    },
+    {
+      "epoch": 0.16648230088495575,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9607468806287936e-05,
+      "loss": 0.5083,
+      "step": 2709
+    },
+    {
+      "epoch": 0.16654375614552605,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9606822945502642e-05,
+      "loss": 0.4602,
+      "step": 2710
+    },
+    {
+      "epoch": 0.16660521140609635,
+      "grad_norm": 15.25,
+      "learning_rate": 1.9606176564467465e-05,
+      "loss": 0.3758,
+      "step": 2711
+    },
+    {
+      "epoch": 0.16666666666666666,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.9605529663217406e-05,
+      "loss": 0.408,
+      "step": 2712
+    },
+    {
+      "epoch": 0.16672812192723696,
+      "grad_norm": 22.875,
+      "learning_rate": 1.96048822417875e-05,
+      "loss": 0.4985,
+      "step": 2713
+    },
+    {
+      "epoch": 0.16678957718780726,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9604234300212804e-05,
+      "loss": 0.5584,
+      "step": 2714
+    },
+    {
+      "epoch": 0.16685103244837757,
+      "grad_norm": 24.0,
+      "learning_rate": 1.960358583852841e-05,
+      "loss": 0.5244,
+      "step": 2715
+    },
+    {
+      "epoch": 0.1669124877089479,
+      "grad_norm": 20.625,
+      "learning_rate": 1.9602936856769432e-05,
+      "loss": 0.3737,
+      "step": 2716
+    },
+    {
+      "epoch": 0.1669739429695182,
+      "grad_norm": 22.5,
+      "learning_rate": 1.9602287354971018e-05,
+      "loss": 0.3998,
+      "step": 2717
+    },
+    {
+      "epoch": 0.1670353982300885,
+      "grad_norm": 9.5625,
+      "learning_rate": 1.960163733316834e-05,
+      "loss": 0.3581,
+      "step": 2718
+    },
+    {
+      "epoch": 0.1670968534906588,
+      "grad_norm": 26.0,
+      "learning_rate": 1.96009867913966e-05,
+      "loss": 0.3753,
+      "step": 2719
+    },
+    {
+      "epoch": 0.1671583087512291,
+      "grad_norm": 2.640625,
+      "learning_rate": 1.9600335729691025e-05,
+      "loss": 0.2619,
+      "step": 2720
+    },
+    {
+      "epoch": 0.16721976401179942,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9599684148086876e-05,
+      "loss": 0.6084,
+      "step": 2721
+    },
+    {
+      "epoch": 0.16728121927236972,
+      "grad_norm": 19.75,
+      "learning_rate": 1.9599032046619437e-05,
+      "loss": 0.5304,
+      "step": 2722
+    },
+    {
+      "epoch": 0.16734267453294002,
+      "grad_norm": 26.125,
+      "learning_rate": 1.9598379425324023e-05,
+      "loss": 0.7238,
+      "step": 2723
+    },
+    {
+      "epoch": 0.16740412979351033,
+      "grad_norm": 30.75,
+      "learning_rate": 1.9597726284235978e-05,
+      "loss": 0.4796,
+      "step": 2724
+    },
+    {
+      "epoch": 0.16746558505408063,
+      "grad_norm": 34.5,
+      "learning_rate": 1.9597072623390668e-05,
+      "loss": 0.4942,
+      "step": 2725
+    },
+    {
+      "epoch": 0.16752704031465093,
+      "grad_norm": 46.75,
+      "learning_rate": 1.9596418442823495e-05,
+      "loss": 0.5364,
+      "step": 2726
+    },
+    {
+      "epoch": 0.16758849557522124,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.9595763742569883e-05,
+      "loss": 0.4821,
+      "step": 2727
+    },
+    {
+      "epoch": 0.16764995083579154,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.959510852266529e-05,
+      "loss": 0.4277,
+      "step": 2728
+    },
+    {
+      "epoch": 0.16771140609636184,
+      "grad_norm": 22.25,
+      "learning_rate": 1.959445278314519e-05,
+      "loss": 0.5351,
+      "step": 2729
+    },
+    {
+      "epoch": 0.16777286135693215,
+      "grad_norm": 26.0,
+      "learning_rate": 1.959379652404511e-05,
+      "loss": 0.4489,
+      "step": 2730
+    },
+    {
+      "epoch": 0.16783431661750245,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.9593139745400575e-05,
+      "loss": 0.4203,
+      "step": 2731
+    },
+    {
+      "epoch": 0.16789577187807275,
+      "grad_norm": 7.5,
+      "learning_rate": 1.9592482447247162e-05,
+      "loss": 0.4846,
+      "step": 2732
+    },
+    {
+      "epoch": 0.16795722713864306,
+      "grad_norm": 22.75,
+      "learning_rate": 1.959182462962046e-05,
+      "loss": 0.362,
+      "step": 2733
+    },
+    {
+      "epoch": 0.16801868239921336,
+      "grad_norm": 40.0,
+      "learning_rate": 1.9591166292556093e-05,
+      "loss": 0.5135,
+      "step": 2734
+    },
+    {
+      "epoch": 0.1680801376597837,
+      "grad_norm": 5.59375,
+      "learning_rate": 1.9590507436089713e-05,
+      "loss": 0.3596,
+      "step": 2735
+    },
+    {
+      "epoch": 0.168141592920354,
+      "grad_norm": 14.5,
+      "learning_rate": 1.9589848060257007e-05,
+      "loss": 0.3923,
+      "step": 2736
+    },
+    {
+      "epoch": 0.1682030481809243,
+      "grad_norm": 38.25,
+      "learning_rate": 1.958918816509367e-05,
+      "loss": 0.747,
+      "step": 2737
+    },
+    {
+      "epoch": 0.1682645034414946,
+      "grad_norm": 11.625,
+      "learning_rate": 1.958852775063545e-05,
+      "loss": 0.4815,
+      "step": 2738
+    },
+    {
+      "epoch": 0.1683259587020649,
+      "grad_norm": 18.625,
+      "learning_rate": 1.9587866816918107e-05,
+      "loss": 0.4905,
+      "step": 2739
+    },
+    {
+      "epoch": 0.1683874139626352,
+      "grad_norm": 38.0,
+      "learning_rate": 1.9587205363977428e-05,
+      "loss": 0.8357,
+      "step": 2740
+    },
+    {
+      "epoch": 0.1684488692232055,
+      "grad_norm": 29.5,
+      "learning_rate": 1.9586543391849243e-05,
+      "loss": 0.8035,
+      "step": 2741
+    },
+    {
+      "epoch": 0.16851032448377581,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9585880900569397e-05,
+      "loss": 0.5401,
+      "step": 2742
+    },
+    {
+      "epoch": 0.16857177974434612,
+      "grad_norm": 6.4375,
+      "learning_rate": 1.958521789017376e-05,
+      "loss": 0.4603,
+      "step": 2743
+    },
+    {
+      "epoch": 0.16863323500491642,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9584554360698247e-05,
+      "loss": 0.3537,
+      "step": 2744
+    },
+    {
+      "epoch": 0.16869469026548672,
+      "grad_norm": 24.125,
+      "learning_rate": 1.958389031217878e-05,
+      "loss": 0.5323,
+      "step": 2745
+    },
+    {
+      "epoch": 0.16875614552605703,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.9583225744651334e-05,
+      "loss": 0.4884,
+      "step": 2746
+    },
+    {
+      "epoch": 0.16881760078662733,
+      "grad_norm": 15.875,
+      "learning_rate": 1.9582560658151886e-05,
+      "loss": 0.4831,
+      "step": 2747
+    },
+    {
+      "epoch": 0.16887905604719763,
+      "grad_norm": 10.25,
+      "learning_rate": 1.958189505271646e-05,
+      "loss": 0.5194,
+      "step": 2748
+    },
+    {
+      "epoch": 0.16894051130776794,
+      "grad_norm": 22.5,
+      "learning_rate": 1.95812289283811e-05,
+      "loss": 0.5502,
+      "step": 2749
+    },
+    {
+      "epoch": 0.16900196656833824,
+      "grad_norm": 17.375,
+      "learning_rate": 1.9580562285181875e-05,
+      "loss": 0.5657,
+      "step": 2750
+    },
+    {
+      "epoch": 0.16906342182890854,
+      "grad_norm": 7.28125,
+      "learning_rate": 1.957989512315489e-05,
+      "loss": 0.4717,
+      "step": 2751
+    },
+    {
+      "epoch": 0.16912487708947885,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9579227442336276e-05,
+      "loss": 0.3677,
+      "step": 2752
+    },
+    {
+      "epoch": 0.16918633235004915,
+      "grad_norm": 11.25,
+      "learning_rate": 1.957855924276219e-05,
+      "loss": 0.2888,
+      "step": 2753
+    },
+    {
+      "epoch": 0.16924778761061948,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9577890524468815e-05,
+      "loss": 0.3348,
+      "step": 2754
+    },
+    {
+      "epoch": 0.16930924287118979,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9577221287492368e-05,
+      "loss": 0.4311,
+      "step": 2755
+    },
+    {
+      "epoch": 0.1693706981317601,
+      "grad_norm": 8.625,
+      "learning_rate": 1.9576551531869092e-05,
+      "loss": 0.5632,
+      "step": 2756
+    },
+    {
+      "epoch": 0.1694321533923304,
+      "grad_norm": 29.25,
+      "learning_rate": 1.9575881257635254e-05,
+      "loss": 0.4393,
+      "step": 2757
+    },
+    {
+      "epoch": 0.1694936086529007,
+      "grad_norm": 4.71875,
+      "learning_rate": 1.957521046482715e-05,
+      "loss": 0.2928,
+      "step": 2758
+    },
+    {
+      "epoch": 0.169555063913471,
+      "grad_norm": 29.375,
+      "learning_rate": 1.9574539153481113e-05,
+      "loss": 0.4267,
+      "step": 2759
+    },
+    {
+      "epoch": 0.1696165191740413,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.9573867323633493e-05,
+      "loss": 0.3317,
+      "step": 2760
+    },
+    {
+      "epoch": 0.1696779744346116,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9573194975320672e-05,
+      "loss": 0.404,
+      "step": 2761
+    },
+    {
+      "epoch": 0.1697394296951819,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9572522108579063e-05,
+      "loss": 0.5028,
+      "step": 2762
+    },
+    {
+      "epoch": 0.1698008849557522,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.9571848723445107e-05,
+      "loss": 0.5406,
+      "step": 2763
+    },
+    {
+      "epoch": 0.16986234021632252,
+      "grad_norm": 46.25,
+      "learning_rate": 1.9571174819955264e-05,
+      "loss": 0.3832,
+      "step": 2764
+    },
+    {
+      "epoch": 0.16992379547689282,
+      "grad_norm": 26.875,
+      "learning_rate": 1.957050039814603e-05,
+      "loss": 0.4258,
+      "step": 2765
+    },
+    {
+      "epoch": 0.16998525073746312,
+      "grad_norm": 21.875,
+      "learning_rate": 1.956982545805393e-05,
+      "loss": 0.4427,
+      "step": 2766
+    },
+    {
+      "epoch": 0.17004670599803343,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9569149999715514e-05,
+      "loss": 0.4482,
+      "step": 2767
+    },
+    {
+      "epoch": 0.17010816125860373,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9568474023167364e-05,
+      "loss": 0.3737,
+      "step": 2768
+    },
+    {
+      "epoch": 0.17016961651917403,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9567797528446083e-05,
+      "loss": 0.5789,
+      "step": 2769
+    },
+    {
+      "epoch": 0.17023107177974434,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9567120515588307e-05,
+      "loss": 0.5185,
+      "step": 2770
+    },
+    {
+      "epoch": 0.17029252704031464,
+      "grad_norm": 20.0,
+      "learning_rate": 1.95664429846307e-05,
+      "loss": 0.6249,
+      "step": 2771
+    },
+    {
+      "epoch": 0.17035398230088494,
+      "grad_norm": 32.75,
+      "learning_rate": 1.956576493560995e-05,
+      "loss": 0.6219,
+      "step": 2772
+    },
+    {
+      "epoch": 0.17041543756145527,
+      "grad_norm": 46.5,
+      "learning_rate": 1.956508636856278e-05,
+      "loss": 0.4436,
+      "step": 2773
+    },
+    {
+      "epoch": 0.17047689282202558,
+      "grad_norm": 44.25,
+      "learning_rate": 1.9564407283525937e-05,
+      "loss": 0.473,
+      "step": 2774
+    },
+    {
+      "epoch": 0.17053834808259588,
+      "grad_norm": 11.8125,
+      "learning_rate": 1.9563727680536195e-05,
+      "loss": 0.4315,
+      "step": 2775
+    },
+    {
+      "epoch": 0.17059980334316618,
+      "grad_norm": 39.25,
+      "learning_rate": 1.9563047559630356e-05,
+      "loss": 0.5703,
+      "step": 2776
+    },
+    {
+      "epoch": 0.1706612586037365,
+      "grad_norm": 42.5,
+      "learning_rate": 1.9562366920845256e-05,
+      "loss": 0.483,
+      "step": 2777
+    },
+    {
+      "epoch": 0.1707227138643068,
+      "grad_norm": 35.75,
+      "learning_rate": 1.956168576421775e-05,
+      "loss": 0.5965,
+      "step": 2778
+    },
+    {
+      "epoch": 0.1707841691248771,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9561004089784726e-05,
+      "loss": 0.5092,
+      "step": 2779
+    },
+    {
+      "epoch": 0.1708456243854474,
+      "grad_norm": 28.25,
+      "learning_rate": 1.95603218975831e-05,
+      "loss": 0.4146,
+      "step": 2780
+    },
+    {
+      "epoch": 0.1709070796460177,
+      "grad_norm": 24.125,
+      "learning_rate": 1.9559639187649817e-05,
+      "loss": 0.4344,
+      "step": 2781
+    },
+    {
+      "epoch": 0.170968534906588,
+      "grad_norm": 31.25,
+      "learning_rate": 1.9558955960021847e-05,
+      "loss": 0.4477,
+      "step": 2782
+    },
+    {
+      "epoch": 0.1710299901671583,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.9558272214736193e-05,
+      "loss": 0.3064,
+      "step": 2783
+    },
+    {
+      "epoch": 0.1710914454277286,
+      "grad_norm": 44.75,
+      "learning_rate": 1.955758795182988e-05,
+      "loss": 0.4247,
+      "step": 2784
+    },
+    {
+      "epoch": 0.17115290068829891,
+      "grad_norm": 76.5,
+      "learning_rate": 1.9556903171339963e-05,
+      "loss": 0.5167,
+      "step": 2785
+    },
+    {
+      "epoch": 0.17121435594886922,
+      "grad_norm": 22.125,
+      "learning_rate": 1.9556217873303526e-05,
+      "loss": 0.7476,
+      "step": 2786
+    },
+    {
+      "epoch": 0.17127581120943952,
+      "grad_norm": 45.75,
+      "learning_rate": 1.955553205775768e-05,
+      "loss": 0.7909,
+      "step": 2787
+    },
+    {
+      "epoch": 0.17133726647000982,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9554845724739565e-05,
+      "loss": 0.6302,
+      "step": 2788
+    },
+    {
+      "epoch": 0.17139872173058013,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9554158874286353e-05,
+      "loss": 0.7298,
+      "step": 2789
+    },
+    {
+      "epoch": 0.17146017699115043,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9553471506435235e-05,
+      "loss": 0.5963,
+      "step": 2790
+    },
+    {
+      "epoch": 0.17152163225172073,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9552783621223437e-05,
+      "loss": 0.5503,
+      "step": 2791
+    },
+    {
+      "epoch": 0.17158308751229107,
+      "grad_norm": 42.25,
+      "learning_rate": 1.955209521868821e-05,
+      "loss": 0.5867,
+      "step": 2792
+    },
+    {
+      "epoch": 0.17164454277286137,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9551406298866838e-05,
+      "loss": 0.7864,
+      "step": 2793
+    },
+    {
+      "epoch": 0.17170599803343167,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9550716861796623e-05,
+      "loss": 0.5497,
+      "step": 2794
+    },
+    {
+      "epoch": 0.17176745329400198,
+      "grad_norm": 24.375,
+      "learning_rate": 1.9550026907514903e-05,
+      "loss": 0.5008,
+      "step": 2795
+    },
+    {
+      "epoch": 0.17182890855457228,
+      "grad_norm": 30.5,
+      "learning_rate": 1.954933643605904e-05,
+      "loss": 0.4837,
+      "step": 2796
+    },
+    {
+      "epoch": 0.17189036381514258,
+      "grad_norm": 32.0,
+      "learning_rate": 1.9548645447466433e-05,
+      "loss": 0.6915,
+      "step": 2797
+    },
+    {
+      "epoch": 0.1719518190757129,
+      "grad_norm": 22.875,
+      "learning_rate": 1.9547953941774495e-05,
+      "loss": 0.4997,
+      "step": 2798
+    },
+    {
+      "epoch": 0.1720132743362832,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9547261919020673e-05,
+      "loss": 0.7398,
+      "step": 2799
+    },
+    {
+      "epoch": 0.1720747295968535,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9546569379242446e-05,
+      "loss": 0.464,
+      "step": 2800
+    },
+    {
+      "epoch": 0.1721361848574238,
+      "grad_norm": 30.875,
+      "learning_rate": 1.954587632247732e-05,
+      "loss": 0.5287,
+      "step": 2801
+    },
+    {
+      "epoch": 0.1721976401179941,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9545182748762825e-05,
+      "loss": 0.6245,
+      "step": 2802
+    },
+    {
+      "epoch": 0.1722590953785644,
+      "grad_norm": 19.75,
+      "learning_rate": 1.9544488658136522e-05,
+      "loss": 0.6787,
+      "step": 2803
+    },
+    {
+      "epoch": 0.1723205506391347,
+      "grad_norm": 22.0,
+      "learning_rate": 1.9543794050635997e-05,
+      "loss": 0.4474,
+      "step": 2804
+    },
+    {
+      "epoch": 0.172382005899705,
+      "grad_norm": 13.3125,
+      "learning_rate": 1.9543098926298864e-05,
+      "loss": 0.435,
+      "step": 2805
+    },
+    {
+      "epoch": 0.1724434611602753,
+      "grad_norm": 12.4375,
+      "learning_rate": 1.954240328516277e-05,
+      "loss": 0.4557,
+      "step": 2806
+    },
+    {
+      "epoch": 0.17250491642084562,
+      "grad_norm": 13.75,
+      "learning_rate": 1.9541707127265388e-05,
+      "loss": 0.4515,
+      "step": 2807
+    },
+    {
+      "epoch": 0.17256637168141592,
+      "grad_norm": 25.75,
+      "learning_rate": 1.9541010452644414e-05,
+      "loss": 0.4023,
+      "step": 2808
+    },
+    {
+      "epoch": 0.17262782694198622,
+      "grad_norm": 18.375,
+      "learning_rate": 1.954031326133758e-05,
+      "loss": 0.5423,
+      "step": 2809
+    },
+    {
+      "epoch": 0.17268928220255653,
+      "grad_norm": 14.625,
+      "learning_rate": 1.953961555338264e-05,
+      "loss": 0.5389,
+      "step": 2810
+    },
+    {
+      "epoch": 0.17275073746312683,
+      "grad_norm": 20.875,
+      "learning_rate": 1.9538917328817377e-05,
+      "loss": 0.4475,
+      "step": 2811
+    },
+    {
+      "epoch": 0.17281219272369716,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.9538218587679605e-05,
+      "loss": 0.4602,
+      "step": 2812
+    },
+    {
+      "epoch": 0.17287364798426746,
+      "grad_norm": 16.375,
+      "learning_rate": 1.9537519330007163e-05,
+      "loss": 0.4164,
+      "step": 2813
+    },
+    {
+      "epoch": 0.17293510324483777,
+      "grad_norm": 19.0,
+      "learning_rate": 1.9536819555837915e-05,
+      "loss": 0.4013,
+      "step": 2814
+    },
+    {
+      "epoch": 0.17299655850540807,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9536119265209763e-05,
+      "loss": 0.4106,
+      "step": 2815
+    },
+    {
+      "epoch": 0.17305801376597837,
+      "grad_norm": 18.75,
+      "learning_rate": 1.9535418458160625e-05,
+      "loss": 0.6183,
+      "step": 2816
+    },
+    {
+      "epoch": 0.17311946902654868,
+      "grad_norm": 33.5,
+      "learning_rate": 1.9534717134728455e-05,
+      "loss": 0.4225,
+      "step": 2817
+    },
+    {
+      "epoch": 0.17318092428711898,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.9534015294951235e-05,
+      "loss": 0.5711,
+      "step": 2818
+    },
+    {
+      "epoch": 0.17324237954768928,
+      "grad_norm": 19.25,
+      "learning_rate": 1.953331293886697e-05,
+      "loss": 0.4941,
+      "step": 2819
+    },
+    {
+      "epoch": 0.1733038348082596,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9532610066513693e-05,
+      "loss": 0.5831,
+      "step": 2820
+    },
+    {
+      "epoch": 0.1733652900688299,
+      "grad_norm": 29.625,
+      "learning_rate": 1.9531906677929472e-05,
+      "loss": 0.6078,
+      "step": 2821
+    },
+    {
+      "epoch": 0.1734267453294002,
+      "grad_norm": 26.875,
+      "learning_rate": 1.9531202773152395e-05,
+      "loss": 0.6546,
+      "step": 2822
+    },
+    {
+      "epoch": 0.1734882005899705,
+      "grad_norm": 18.25,
+      "learning_rate": 1.953049835222059e-05,
+      "loss": 0.4316,
+      "step": 2823
+    },
+    {
+      "epoch": 0.1735496558505408,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.952979341517219e-05,
+      "loss": 0.3824,
+      "step": 2824
+    },
+    {
+      "epoch": 0.1736111111111111,
+      "grad_norm": 26.0,
+      "learning_rate": 1.952908796204538e-05,
+      "loss": 0.7611,
+      "step": 2825
+    },
+    {
+      "epoch": 0.1736725663716814,
+      "grad_norm": 44.0,
+      "learning_rate": 1.9528381992878362e-05,
+      "loss": 0.7742,
+      "step": 2826
+    },
+    {
+      "epoch": 0.1737340216322517,
+      "grad_norm": 22.5,
+      "learning_rate": 1.9527675507709368e-05,
+      "loss": 0.669,
+      "step": 2827
+    },
+    {
+      "epoch": 0.17379547689282202,
+      "grad_norm": 31.875,
+      "learning_rate": 1.9526968506576654e-05,
+      "loss": 0.6578,
+      "step": 2828
+    },
+    {
+      "epoch": 0.17385693215339232,
+      "grad_norm": 14.25,
+      "learning_rate": 1.9526260989518506e-05,
+      "loss": 0.4874,
+      "step": 2829
+    },
+    {
+      "epoch": 0.17391838741396262,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9525552956573244e-05,
+      "loss": 0.7209,
+      "step": 2830
+    },
+    {
+      "epoch": 0.17397984267453295,
+      "grad_norm": 14.125,
+      "learning_rate": 1.9524844407779208e-05,
+      "loss": 0.4477,
+      "step": 2831
+    },
+    {
+      "epoch": 0.17404129793510326,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.952413534317477e-05,
+      "loss": 0.4057,
+      "step": 2832
+    },
+    {
+      "epoch": 0.17410275319567356,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.9523425762798328e-05,
+      "loss": 0.5297,
+      "step": 2833
+    },
+    {
+      "epoch": 0.17416420845624386,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.952271566668831e-05,
+      "loss": 0.47,
+      "step": 2834
+    },
+    {
+      "epoch": 0.17422566371681417,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.9522005054883172e-05,
+      "loss": 0.3851,
+      "step": 2835
+    },
+    {
+      "epoch": 0.17428711897738447,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.9521293927421388e-05,
+      "loss": 0.4601,
+      "step": 2836
+    },
+    {
+      "epoch": 0.17434857423795477,
+      "grad_norm": 16.5,
+      "learning_rate": 1.9520582284341478e-05,
+      "loss": 0.5451,
+      "step": 2837
+    },
+    {
+      "epoch": 0.17441002949852508,
+      "grad_norm": 17.75,
+      "learning_rate": 1.951987012568198e-05,
+      "loss": 0.4489,
+      "step": 2838
+    },
+    {
+      "epoch": 0.17447148475909538,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9519157451481453e-05,
+      "loss": 0.3555,
+      "step": 2839
+    },
+    {
+      "epoch": 0.17453294001966568,
+      "grad_norm": 5.75,
+      "learning_rate": 1.95184442617785e-05,
+      "loss": 0.4051,
+      "step": 2840
+    },
+    {
+      "epoch": 0.174594395280236,
+      "grad_norm": 21.5,
+      "learning_rate": 1.951773055661174e-05,
+      "loss": 0.5001,
+      "step": 2841
+    },
+    {
+      "epoch": 0.1746558505408063,
+      "grad_norm": 23.625,
+      "learning_rate": 1.9517016336019817e-05,
+      "loss": 0.4908,
+      "step": 2842
+    },
+    {
+      "epoch": 0.1747173058013766,
+      "grad_norm": 30.0,
+      "learning_rate": 1.9516301600041417e-05,
+      "loss": 0.5732,
+      "step": 2843
+    },
+    {
+      "epoch": 0.1747787610619469,
+      "grad_norm": 27.75,
+      "learning_rate": 1.9515586348715246e-05,
+      "loss": 0.5956,
+      "step": 2844
+    },
+    {
+      "epoch": 0.1748402163225172,
+      "grad_norm": 19.75,
+      "learning_rate": 1.951487058208003e-05,
+      "loss": 0.5303,
+      "step": 2845
+    },
+    {
+      "epoch": 0.1749016715830875,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9514154300174542e-05,
+      "loss": 0.4475,
+      "step": 2846
+    },
+    {
+      "epoch": 0.1749631268436578,
+      "grad_norm": 6.75,
+      "learning_rate": 1.9513437503037563e-05,
+      "loss": 0.3651,
+      "step": 2847
+    },
+    {
+      "epoch": 0.1750245821042281,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9512720190707915e-05,
+      "loss": 0.3686,
+      "step": 2848
+    },
+    {
+      "epoch": 0.1750860373647984,
+      "grad_norm": 11.5,
+      "learning_rate": 1.951200236322444e-05,
+      "loss": 0.5353,
+      "step": 2849
+    },
+    {
+      "epoch": 0.17514749262536874,
+      "grad_norm": 19.875,
+      "learning_rate": 1.9511284020626015e-05,
+      "loss": 0.4094,
+      "step": 2850
+    },
+    {
+      "epoch": 0.17520894788593905,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9510565162951538e-05,
+      "loss": 0.4881,
+      "step": 2851
+    },
+    {
+      "epoch": 0.17527040314650935,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9509845790239938e-05,
+      "loss": 0.3908,
+      "step": 2852
+    },
+    {
+      "epoch": 0.17533185840707965,
+      "grad_norm": 25.125,
+      "learning_rate": 1.9509125902530178e-05,
+      "loss": 0.6116,
+      "step": 2853
+    },
+    {
+      "epoch": 0.17539331366764996,
+      "grad_norm": 19.25,
+      "learning_rate": 1.9508405499861235e-05,
+      "loss": 0.3593,
+      "step": 2854
+    },
+    {
+      "epoch": 0.17545476892822026,
+      "grad_norm": 8.875,
+      "learning_rate": 1.9507684582272128e-05,
+      "loss": 0.4842,
+      "step": 2855
+    },
+    {
+      "epoch": 0.17551622418879056,
+      "grad_norm": 11.1875,
+      "learning_rate": 1.9506963149801894e-05,
+      "loss": 0.3666,
+      "step": 2856
+    },
+    {
+      "epoch": 0.17557767944936087,
+      "grad_norm": 21.5,
+      "learning_rate": 1.95062412024896e-05,
+      "loss": 0.6243,
+      "step": 2857
+    },
+    {
+      "epoch": 0.17563913470993117,
+      "grad_norm": 16.25,
+      "learning_rate": 1.950551874037435e-05,
+      "loss": 0.5342,
+      "step": 2858
+    },
+    {
+      "epoch": 0.17570058997050148,
+      "grad_norm": 21.25,
+      "learning_rate": 1.950479576349526e-05,
+      "loss": 0.4998,
+      "step": 2859
+    },
+    {
+      "epoch": 0.17576204523107178,
+      "grad_norm": 22.875,
+      "learning_rate": 1.9504072271891486e-05,
+      "loss": 0.463,
+      "step": 2860
+    },
+    {
+      "epoch": 0.17582350049164208,
+      "grad_norm": 35.25,
+      "learning_rate": 1.9503348265602212e-05,
+      "loss": 0.597,
+      "step": 2861
+    },
+    {
+      "epoch": 0.17588495575221239,
+      "grad_norm": 19.125,
+      "learning_rate": 1.950262374466664e-05,
+      "loss": 0.7796,
+      "step": 2862
+    },
+    {
+      "epoch": 0.1759464110127827,
+      "grad_norm": 7.59375,
+      "learning_rate": 1.950189870912401e-05,
+      "loss": 0.5042,
+      "step": 2863
+    },
+    {
+      "epoch": 0.176007866273353,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.950117315901358e-05,
+      "loss": 0.4301,
+      "step": 2864
+    },
+    {
+      "epoch": 0.1760693215339233,
+      "grad_norm": 41.25,
+      "learning_rate": 1.9500447094374645e-05,
+      "loss": 0.5648,
+      "step": 2865
+    },
+    {
+      "epoch": 0.1761307767944936,
+      "grad_norm": 11.625,
+      "learning_rate": 1.9499720515246524e-05,
+      "loss": 0.4141,
+      "step": 2866
+    },
+    {
+      "epoch": 0.1761922320550639,
+      "grad_norm": 9.0,
+      "learning_rate": 1.949899342166857e-05,
+      "loss": 0.4188,
+      "step": 2867
+    },
+    {
+      "epoch": 0.1762536873156342,
+      "grad_norm": 42.75,
+      "learning_rate": 1.949826581368015e-05,
+      "loss": 0.5076,
+      "step": 2868
+    },
+    {
+      "epoch": 0.17631514257620454,
+      "grad_norm": 50.25,
+      "learning_rate": 1.949753769132067e-05,
+      "loss": 0.5167,
+      "step": 2869
+    },
+    {
+      "epoch": 0.17637659783677484,
+      "grad_norm": 13.5,
+      "learning_rate": 1.9496809054629563e-05,
+      "loss": 0.2912,
+      "step": 2870
+    },
+    {
+      "epoch": 0.17643805309734514,
+      "grad_norm": 11.5,
+      "learning_rate": 1.9496079903646282e-05,
+      "loss": 0.4768,
+      "step": 2871
+    },
+    {
+      "epoch": 0.17649950835791545,
+      "grad_norm": 22.375,
+      "learning_rate": 1.949535023841032e-05,
+      "loss": 0.4446,
+      "step": 2872
+    },
+    {
+      "epoch": 0.17656096361848575,
+      "grad_norm": 237.0,
+      "learning_rate": 1.949462005896119e-05,
+      "loss": 0.6436,
+      "step": 2873
+    },
+    {
+      "epoch": 0.17662241887905605,
+      "grad_norm": 24.375,
+      "learning_rate": 1.949388936533843e-05,
+      "loss": 0.4062,
+      "step": 2874
+    },
+    {
+      "epoch": 0.17668387413962636,
+      "grad_norm": 12.875,
+      "learning_rate": 1.9493158157581617e-05,
+      "loss": 0.345,
+      "step": 2875
+    },
+    {
+      "epoch": 0.17674532940019666,
+      "grad_norm": 42.75,
+      "learning_rate": 1.949242643573034e-05,
+      "loss": 0.3885,
+      "step": 2876
+    },
+    {
+      "epoch": 0.17680678466076696,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9491694199824232e-05,
+      "loss": 0.3725,
+      "step": 2877
+    },
+    {
+      "epoch": 0.17686823992133727,
+      "grad_norm": 3.140625,
+      "learning_rate": 1.9490961449902946e-05,
+      "loss": 0.5013,
+      "step": 2878
+    },
+    {
+      "epoch": 0.17692969518190757,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9490228186006164e-05,
+      "loss": 0.4623,
+      "step": 2879
+    },
+    {
+      "epoch": 0.17699115044247787,
+      "grad_norm": 32.75,
+      "learning_rate": 1.948949440817359e-05,
+      "loss": 0.6061,
+      "step": 2880
+    },
+    {
+      "epoch": 0.17705260570304818,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.9488760116444966e-05,
+      "loss": 0.5194,
+      "step": 2881
+    },
+    {
+      "epoch": 0.17711406096361848,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9488025310860057e-05,
+      "loss": 0.7012,
+      "step": 2882
+    },
+    {
+      "epoch": 0.17717551622418878,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9487289991458653e-05,
+      "loss": 0.396,
+      "step": 2883
+    },
+    {
+      "epoch": 0.1772369714847591,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9486554158280576e-05,
+      "loss": 0.3765,
+      "step": 2884
+    },
+    {
+      "epoch": 0.1772984267453294,
+      "grad_norm": 15.25,
+      "learning_rate": 1.9485817811365676e-05,
+      "loss": 0.5094,
+      "step": 2885
+    },
+    {
+      "epoch": 0.1773598820058997,
+      "grad_norm": 15.75,
+      "learning_rate": 1.948508095075383e-05,
+      "loss": 0.4745,
+      "step": 2886
+    },
+    {
+      "epoch": 0.17742133726647,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9484343576484935e-05,
+      "loss": 0.4205,
+      "step": 2887
+    },
+    {
+      "epoch": 0.17748279252704033,
+      "grad_norm": 8.0,
+      "learning_rate": 1.948360568859893e-05,
+      "loss": 0.4788,
+      "step": 2888
+    },
+    {
+      "epoch": 0.17754424778761063,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9482867287135772e-05,
+      "loss": 0.6695,
+      "step": 2889
+    },
+    {
+      "epoch": 0.17760570304818094,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9482128372135446e-05,
+      "loss": 0.6709,
+      "step": 2890
+    },
+    {
+      "epoch": 0.17766715830875124,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9481388943637976e-05,
+      "loss": 0.7337,
+      "step": 2891
+    },
+    {
+      "epoch": 0.17772861356932154,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.94806490016834e-05,
+      "loss": 0.7453,
+      "step": 2892
+    },
+    {
+      "epoch": 0.17779006882989185,
+      "grad_norm": 29.75,
+      "learning_rate": 1.9479908546311783e-05,
+      "loss": 0.4285,
+      "step": 2893
+    },
+    {
+      "epoch": 0.17785152409046215,
+      "grad_norm": 14.625,
+      "learning_rate": 1.9479167577563233e-05,
+      "loss": 0.5007,
+      "step": 2894
+    },
+    {
+      "epoch": 0.17791297935103245,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.9478426095477877e-05,
+      "loss": 0.3671,
+      "step": 2895
+    },
+    {
+      "epoch": 0.17797443461160276,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.947768410009586e-05,
+      "loss": 0.426,
+      "step": 2896
+    },
+    {
+      "epoch": 0.17803588987217306,
+      "grad_norm": 25.75,
+      "learning_rate": 1.947694159145737e-05,
+      "loss": 0.5182,
+      "step": 2897
+    },
+    {
+      "epoch": 0.17809734513274336,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.9476198569602616e-05,
+      "loss": 0.4622,
+      "step": 2898
+    },
+    {
+      "epoch": 0.17815880039331367,
+      "grad_norm": 29.625,
+      "learning_rate": 1.947545503457184e-05,
+      "loss": 0.5383,
+      "step": 2899
+    },
+    {
+      "epoch": 0.17822025565388397,
+      "grad_norm": 4.96875,
+      "learning_rate": 1.94747109864053e-05,
+      "loss": 0.4186,
+      "step": 2900
+    },
+    {
+      "epoch": 0.17828171091445427,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9473966425143292e-05,
+      "loss": 0.396,
+      "step": 2901
+    },
+    {
+      "epoch": 0.17834316617502458,
+      "grad_norm": 10.75,
+      "learning_rate": 1.9473221350826145e-05,
+      "loss": 0.36,
+      "step": 2902
+    },
+    {
+      "epoch": 0.17840462143559488,
+      "grad_norm": 15.5,
+      "learning_rate": 1.9472475763494196e-05,
+      "loss": 0.3239,
+      "step": 2903
+    },
+    {
+      "epoch": 0.17846607669616518,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.947172966318783e-05,
+      "loss": 0.2852,
+      "step": 2904
+    },
+    {
+      "epoch": 0.17852753195673549,
+      "grad_norm": 48.5,
+      "learning_rate": 1.9470983049947446e-05,
+      "loss": 0.675,
+      "step": 2905
+    },
+    {
+      "epoch": 0.1785889872173058,
+      "grad_norm": 18.5,
+      "learning_rate": 1.947023592381348e-05,
+      "loss": 0.5563,
+      "step": 2906
+    },
+    {
+      "epoch": 0.1786504424778761,
+      "grad_norm": 38.25,
+      "learning_rate": 1.9469488284826386e-05,
+      "loss": 0.5168,
+      "step": 2907
+    },
+    {
+      "epoch": 0.17871189773844642,
+      "grad_norm": 29.625,
+      "learning_rate": 1.946874013302666e-05,
+      "loss": 0.6391,
+      "step": 2908
+    },
+    {
+      "epoch": 0.17877335299901673,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9467991468454817e-05,
+      "loss": 0.4871,
+      "step": 2909
+    },
+    {
+      "epoch": 0.17883480825958703,
+      "grad_norm": 14.375,
+      "learning_rate": 1.9467242291151394e-05,
+      "loss": 0.4639,
+      "step": 2910
+    },
+    {
+      "epoch": 0.17889626352015733,
+      "grad_norm": 13.75,
+      "learning_rate": 1.9466492601156964e-05,
+      "loss": 0.5274,
+      "step": 2911
+    },
+    {
+      "epoch": 0.17895771878072764,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.946574239851213e-05,
+      "loss": 0.335,
+      "step": 2912
+    },
+    {
+      "epoch": 0.17901917404129794,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9464991683257514e-05,
+      "loss": 0.6474,
+      "step": 2913
+    },
+    {
+      "epoch": 0.17908062930186824,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9464240455433775e-05,
+      "loss": 0.5869,
+      "step": 2914
+    },
+    {
+      "epoch": 0.17914208456243855,
+      "grad_norm": 18.0,
+      "learning_rate": 1.946348871508159e-05,
+      "loss": 0.4296,
+      "step": 2915
+    },
+    {
+      "epoch": 0.17920353982300885,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9462736462241672e-05,
+      "loss": 0.3919,
+      "step": 2916
+    },
+    {
+      "epoch": 0.17926499508357915,
+      "grad_norm": 7.3125,
+      "learning_rate": 1.946198369695476e-05,
+      "loss": 0.3655,
+      "step": 2917
+    },
+    {
+      "epoch": 0.17932645034414946,
+      "grad_norm": 22.75,
+      "learning_rate": 1.9461230419261615e-05,
+      "loss": 0.4272,
+      "step": 2918
+    },
+    {
+      "epoch": 0.17938790560471976,
+      "grad_norm": 17.125,
+      "learning_rate": 1.9460476629203035e-05,
+      "loss": 0.4657,
+      "step": 2919
+    },
+    {
+      "epoch": 0.17944936086529006,
+      "grad_norm": 19.5,
+      "learning_rate": 1.945972232681984e-05,
+      "loss": 0.4281,
+      "step": 2920
+    },
+    {
+      "epoch": 0.17951081612586037,
+      "grad_norm": 18.625,
+      "learning_rate": 1.9458967512152872e-05,
+      "loss": 0.4025,
+      "step": 2921
+    },
+    {
+      "epoch": 0.17957227138643067,
+      "grad_norm": 5.21875,
+      "learning_rate": 1.9458212185243015e-05,
+      "loss": 0.49,
+      "step": 2922
+    },
+    {
+      "epoch": 0.17963372664700097,
+      "grad_norm": 6.25,
+      "learning_rate": 1.945745634613117e-05,
+      "loss": 0.4526,
+      "step": 2923
+    },
+    {
+      "epoch": 0.17969518190757128,
+      "grad_norm": 14.75,
+      "learning_rate": 1.945669999485827e-05,
+      "loss": 0.4609,
+      "step": 2924
+    },
+    {
+      "epoch": 0.17975663716814158,
+      "grad_norm": 35.25,
+      "learning_rate": 1.9455943131465272e-05,
+      "loss": 0.3328,
+      "step": 2925
+    },
+    {
+      "epoch": 0.17981809242871188,
+      "grad_norm": 19.25,
+      "learning_rate": 1.945518575599317e-05,
+      "loss": 0.5412,
+      "step": 2926
+    },
+    {
+      "epoch": 0.17987954768928222,
+      "grad_norm": 13.3125,
+      "learning_rate": 1.945442786848297e-05,
+      "loss": 0.3348,
+      "step": 2927
+    },
+    {
+      "epoch": 0.17994100294985252,
+      "grad_norm": 39.75,
+      "learning_rate": 1.9453669468975724e-05,
+      "loss": 0.3417,
+      "step": 2928
+    },
+    {
+      "epoch": 0.18000245821042282,
+      "grad_norm": 32.5,
+      "learning_rate": 1.9452910557512497e-05,
+      "loss": 0.553,
+      "step": 2929
+    },
+    {
+      "epoch": 0.18006391347099313,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9452151134134385e-05,
+      "loss": 0.4277,
+      "step": 2930
+    },
+    {
+      "epoch": 0.18012536873156343,
+      "grad_norm": 28.0,
+      "learning_rate": 1.945139119888252e-05,
+      "loss": 0.5471,
+      "step": 2931
+    },
+    {
+      "epoch": 0.18018682399213373,
+      "grad_norm": 17.625,
+      "learning_rate": 1.945063075179805e-05,
+      "loss": 0.4159,
+      "step": 2932
+    },
+    {
+      "epoch": 0.18024827925270404,
+      "grad_norm": 27.625,
+      "learning_rate": 1.9449869792922162e-05,
+      "loss": 0.4285,
+      "step": 2933
+    },
+    {
+      "epoch": 0.18030973451327434,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.944910832229606e-05,
+      "loss": 0.3986,
+      "step": 2934
+    },
+    {
+      "epoch": 0.18037118977384464,
+      "grad_norm": 39.5,
+      "learning_rate": 1.9448346339960984e-05,
+      "loss": 0.4442,
+      "step": 2935
+    },
+    {
+      "epoch": 0.18043264503441495,
+      "grad_norm": 12.5,
+      "learning_rate": 1.9447583845958198e-05,
+      "loss": 0.3368,
+      "step": 2936
+    },
+    {
+      "epoch": 0.18049410029498525,
+      "grad_norm": 20.25,
+      "learning_rate": 1.9446820840328993e-05,
+      "loss": 0.341,
+      "step": 2937
+    },
+    {
+      "epoch": 0.18055555555555555,
+      "grad_norm": 23.375,
+      "learning_rate": 1.944605732311469e-05,
+      "loss": 0.6011,
+      "step": 2938
+    },
+    {
+      "epoch": 0.18061701081612586,
+      "grad_norm": 33.0,
+      "learning_rate": 1.944529329435664e-05,
+      "loss": 0.4328,
+      "step": 2939
+    },
+    {
+      "epoch": 0.18067846607669616,
+      "grad_norm": 8.0,
+      "learning_rate": 1.944452875409621e-05,
+      "loss": 0.5344,
+      "step": 2940
+    },
+    {
+      "epoch": 0.18073992133726646,
+      "grad_norm": 22.75,
+      "learning_rate": 1.944376370237481e-05,
+      "loss": 0.3518,
+      "step": 2941
+    },
+    {
+      "epoch": 0.18080137659783677,
+      "grad_norm": 19.875,
+      "learning_rate": 1.944299813923387e-05,
+      "loss": 0.5395,
+      "step": 2942
+    },
+    {
+      "epoch": 0.18086283185840707,
+      "grad_norm": 10.125,
+      "learning_rate": 1.9442232064714848e-05,
+      "loss": 0.5081,
+      "step": 2943
+    },
+    {
+      "epoch": 0.18092428711897737,
+      "grad_norm": 17.0,
+      "learning_rate": 1.944146547885923e-05,
+      "loss": 0.2984,
+      "step": 2944
+    },
+    {
+      "epoch": 0.18098574237954768,
+      "grad_norm": 61.75,
+      "learning_rate": 1.9440698381708526e-05,
+      "loss": 0.5799,
+      "step": 2945
+    },
+    {
+      "epoch": 0.181047197640118,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9439930773304284e-05,
+      "loss": 0.4806,
+      "step": 2946
+    },
+    {
+      "epoch": 0.1811086529006883,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9439162653688066e-05,
+      "loss": 0.5856,
+      "step": 2947
+    },
+    {
+      "epoch": 0.18117010816125861,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9438394022901477e-05,
+      "loss": 0.558,
+      "step": 2948
+    },
+    {
+      "epoch": 0.18123156342182892,
+      "grad_norm": 31.375,
+      "learning_rate": 1.9437624880986135e-05,
+      "loss": 0.7091,
+      "step": 2949
+    },
+    {
+      "epoch": 0.18129301868239922,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9436855227983695e-05,
+      "loss": 0.6281,
+      "step": 2950
+    },
+    {
+      "epoch": 0.18135447394296952,
+      "grad_norm": 29.875,
+      "learning_rate": 1.9436085063935837e-05,
+      "loss": 0.5779,
+      "step": 2951
+    },
+    {
+      "epoch": 0.18141592920353983,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9435314388884266e-05,
+      "loss": 0.5043,
+      "step": 2952
+    },
+    {
+      "epoch": 0.18147738446411013,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9434543202870726e-05,
+      "loss": 0.4537,
+      "step": 2953
+    },
+    {
+      "epoch": 0.18153883972468043,
+      "grad_norm": 11.8125,
+      "learning_rate": 1.9433771505936968e-05,
+      "loss": 0.4489,
+      "step": 2954
+    },
+    {
+      "epoch": 0.18160029498525074,
+      "grad_norm": 22.0,
+      "learning_rate": 1.943299929812479e-05,
+      "loss": 0.483,
+      "step": 2955
+    },
+    {
+      "epoch": 0.18166175024582104,
+      "grad_norm": 22.5,
+      "learning_rate": 1.943222657947601e-05,
+      "loss": 0.7295,
+      "step": 2956
+    },
+    {
+      "epoch": 0.18172320550639134,
+      "grad_norm": 35.0,
+      "learning_rate": 1.9431453350032476e-05,
+      "loss": 0.495,
+      "step": 2957
+    },
+    {
+      "epoch": 0.18178466076696165,
+      "grad_norm": 21.25,
+      "learning_rate": 1.943067960983605e-05,
+      "loss": 0.3872,
+      "step": 2958
+    },
+    {
+      "epoch": 0.18184611602753195,
+      "grad_norm": 24.0,
+      "learning_rate": 1.9429905358928648e-05,
+      "loss": 0.3747,
+      "step": 2959
+    },
+    {
+      "epoch": 0.18190757128810225,
+      "grad_norm": 13.125,
+      "learning_rate": 1.9429130597352187e-05,
+      "loss": 0.3277,
+      "step": 2960
+    },
+    {
+      "epoch": 0.18196902654867256,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9428355325148632e-05,
+      "loss": 0.5526,
+      "step": 2961
+    },
+    {
+      "epoch": 0.18203048180924286,
+      "grad_norm": 31.875,
+      "learning_rate": 1.9427579542359966e-05,
+      "loss": 0.5878,
+      "step": 2962
+    },
+    {
+      "epoch": 0.18209193706981316,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.9426803249028198e-05,
+      "loss": 0.5358,
+      "step": 2963
+    },
+    {
+      "epoch": 0.18215339233038347,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9426026445195365e-05,
+      "loss": 0.6044,
+      "step": 2964
+    },
+    {
+      "epoch": 0.1822148475909538,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9425249130903544e-05,
+      "loss": 0.5308,
+      "step": 2965
+    },
+    {
+      "epoch": 0.1822763028515241,
+      "grad_norm": 31.0,
+      "learning_rate": 1.9424471306194822e-05,
+      "loss": 0.5396,
+      "step": 2966
+    },
+    {
+      "epoch": 0.1823377581120944,
+      "grad_norm": 38.5,
+      "learning_rate": 1.942369297111132e-05,
+      "loss": 0.3954,
+      "step": 2967
+    },
+    {
+      "epoch": 0.1823992133726647,
+      "grad_norm": 13.75,
+      "learning_rate": 1.942291412569519e-05,
+      "loss": 0.4964,
+      "step": 2968
+    },
+    {
+      "epoch": 0.182460668633235,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9422134769988617e-05,
+      "loss": 0.5155,
+      "step": 2969
+    },
+    {
+      "epoch": 0.18252212389380532,
+      "grad_norm": 22.375,
+      "learning_rate": 1.9421354904033797e-05,
+      "loss": 0.7103,
+      "step": 2970
+    },
+    {
+      "epoch": 0.18258357915437562,
+      "grad_norm": 13.375,
+      "learning_rate": 1.942057452787297e-05,
+      "loss": 0.628,
+      "step": 2971
+    },
+    {
+      "epoch": 0.18264503441494592,
+      "grad_norm": 21.5,
+      "learning_rate": 1.941979364154839e-05,
+      "loss": 0.4804,
+      "step": 2972
+    },
+    {
+      "epoch": 0.18270648967551623,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9419012245102347e-05,
+      "loss": 0.6384,
+      "step": 2973
+    },
+    {
+      "epoch": 0.18276794493608653,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9418230338577164e-05,
+      "loss": 0.6131,
+      "step": 2974
+    },
+    {
+      "epoch": 0.18282940019665683,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.9417447922015174e-05,
+      "loss": 0.4896,
+      "step": 2975
+    },
+    {
+      "epoch": 0.18289085545722714,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9416664995458756e-05,
+      "loss": 0.5091,
+      "step": 2976
+    },
+    {
+      "epoch": 0.18295231071779744,
+      "grad_norm": 13.375,
+      "learning_rate": 1.9415881558950302e-05,
+      "loss": 0.5236,
+      "step": 2977
+    },
+    {
+      "epoch": 0.18301376597836774,
+      "grad_norm": 17.875,
+      "learning_rate": 1.9415097612532247e-05,
+      "loss": 0.4612,
+      "step": 2978
+    },
+    {
+      "epoch": 0.18307522123893805,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9414313156247035e-05,
+      "loss": 0.481,
+      "step": 2979
+    },
+    {
+      "epoch": 0.18313667649950835,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.9413528190137158e-05,
+      "loss": 0.3282,
+      "step": 2980
+    },
+    {
+      "epoch": 0.18319813176007865,
+      "grad_norm": 18.125,
+      "learning_rate": 1.941274271424512e-05,
+      "loss": 0.4691,
+      "step": 2981
+    },
+    {
+      "epoch": 0.18325958702064896,
+      "grad_norm": 37.75,
+      "learning_rate": 1.941195672861345e-05,
+      "loss": 0.4697,
+      "step": 2982
+    },
+    {
+      "epoch": 0.18332104228121926,
+      "grad_norm": 10.8125,
+      "learning_rate": 1.9411170233284728e-05,
+      "loss": 0.4417,
+      "step": 2983
+    },
+    {
+      "epoch": 0.1833824975417896,
+      "grad_norm": 15.5,
+      "learning_rate": 1.9410383228301536e-05,
+      "loss": 0.3815,
+      "step": 2984
+    },
+    {
+      "epoch": 0.1834439528023599,
+      "grad_norm": 15.0,
+      "learning_rate": 1.9409595713706496e-05,
+      "loss": 0.4851,
+      "step": 2985
+    },
+    {
+      "epoch": 0.1835054080629302,
+      "grad_norm": 28.125,
+      "learning_rate": 1.9408807689542257e-05,
+      "loss": 0.4497,
+      "step": 2986
+    },
+    {
+      "epoch": 0.1835668633235005,
+      "grad_norm": 7.09375,
+      "learning_rate": 1.940801915585149e-05,
+      "loss": 0.4029,
+      "step": 2987
+    },
+    {
+      "epoch": 0.1836283185840708,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.9407230112676898e-05,
+      "loss": 0.3591,
+      "step": 2988
+    },
+    {
+      "epoch": 0.1836897738446411,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.9406440560061214e-05,
+      "loss": 0.2735,
+      "step": 2989
+    },
+    {
+      "epoch": 0.1837512291052114,
+      "grad_norm": 11.8125,
+      "learning_rate": 1.9405650498047193e-05,
+      "loss": 0.2801,
+      "step": 2990
+    },
+    {
+      "epoch": 0.18381268436578171,
+      "grad_norm": 20.375,
+      "learning_rate": 1.9404859926677625e-05,
+      "loss": 0.3876,
+      "step": 2991
+    },
+    {
+      "epoch": 0.18387413962635202,
+      "grad_norm": 6.53125,
+      "learning_rate": 1.9404068845995317e-05,
+      "loss": 0.3238,
+      "step": 2992
+    },
+    {
+      "epoch": 0.18393559488692232,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.940327725604311e-05,
+      "loss": 0.3296,
+      "step": 2993
+    },
+    {
+      "epoch": 0.18399705014749262,
+      "grad_norm": 10.25,
+      "learning_rate": 1.9402485156863876e-05,
+      "loss": 0.3617,
+      "step": 2994
+    },
+    {
+      "epoch": 0.18405850540806293,
+      "grad_norm": 30.125,
+      "learning_rate": 1.9401692548500504e-05,
+      "loss": 0.5456,
+      "step": 2995
+    },
+    {
+      "epoch": 0.18411996066863323,
+      "grad_norm": 51.5,
+      "learning_rate": 1.9400899430995923e-05,
+      "loss": 0.5129,
+      "step": 2996
+    },
+    {
+      "epoch": 0.18418141592920353,
+      "grad_norm": 27.25,
+      "learning_rate": 1.9400105804393083e-05,
+      "loss": 0.5718,
+      "step": 2997
+    },
+    {
+      "epoch": 0.18424287118977384,
+      "grad_norm": 20.875,
+      "learning_rate": 1.9399311668734957e-05,
+      "loss": 0.5517,
+      "step": 2998
+    },
+    {
+      "epoch": 0.18430432645034414,
+      "grad_norm": 32.25,
+      "learning_rate": 1.9398517024064556e-05,
+      "loss": 0.2847,
+      "step": 2999
+    },
+    {
+      "epoch": 0.18436578171091444,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.939772187042491e-05,
+      "loss": 0.3404,
+      "step": 3000
+    },
+    {
+      "epoch": 0.18442723697148475,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9396926207859085e-05,
+      "loss": 0.6463,
+      "step": 3001
+    },
+    {
+      "epoch": 0.18448869223205505,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9396130036410166e-05,
+      "loss": 0.7697,
+      "step": 3002
+    },
+    {
+      "epoch": 0.18455014749262535,
+      "grad_norm": 12.125,
+      "learning_rate": 1.9395333356121268e-05,
+      "loss": 0.4747,
+      "step": 3003
+    },
+    {
+      "epoch": 0.18461160275319569,
+      "grad_norm": 29.625,
+      "learning_rate": 1.9394536167035535e-05,
+      "loss": 0.5613,
+      "step": 3004
+    },
+    {
+      "epoch": 0.184673058013766,
+      "grad_norm": 30.625,
+      "learning_rate": 1.939373846919614e-05,
+      "loss": 0.7578,
+      "step": 3005
+    },
+    {
+      "epoch": 0.1847345132743363,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9392940262646284e-05,
+      "loss": 0.5454,
+      "step": 3006
+    },
+    {
+      "epoch": 0.1847959685349066,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9392141547429183e-05,
+      "loss": 0.4712,
+      "step": 3007
+    },
+    {
+      "epoch": 0.1848574237954769,
+      "grad_norm": 21.0,
+      "learning_rate": 1.9391342323588104e-05,
+      "loss": 0.4669,
+      "step": 3008
+    },
+    {
+      "epoch": 0.1849188790560472,
+      "grad_norm": 19.875,
+      "learning_rate": 1.939054259116632e-05,
+      "loss": 0.3443,
+      "step": 3009
+    },
+    {
+      "epoch": 0.1849803343166175,
+      "grad_norm": 18.125,
+      "learning_rate": 1.938974235020714e-05,
+      "loss": 0.4148,
+      "step": 3010
+    },
+    {
+      "epoch": 0.1850417895771878,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9388941600753902e-05,
+      "loss": 0.4424,
+      "step": 3011
+    },
+    {
+      "epoch": 0.1851032448377581,
+      "grad_norm": 10.75,
+      "learning_rate": 1.9388140342849976e-05,
+      "loss": 0.3094,
+      "step": 3012
+    },
+    {
+      "epoch": 0.18516470009832842,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9387338576538743e-05,
+      "loss": 0.5617,
+      "step": 3013
+    },
+    {
+      "epoch": 0.18522615535889872,
+      "grad_norm": 2.34375,
+      "learning_rate": 1.9386536301863626e-05,
+      "loss": 0.2456,
+      "step": 3014
+    },
+    {
+      "epoch": 0.18528761061946902,
+      "grad_norm": 10.8125,
+      "learning_rate": 1.9385733518868074e-05,
+      "loss": 0.2986,
+      "step": 3015
+    },
+    {
+      "epoch": 0.18534906588003933,
+      "grad_norm": 4.0,
+      "learning_rate": 1.938493022759556e-05,
+      "loss": 0.2952,
+      "step": 3016
+    },
+    {
+      "epoch": 0.18541052114060963,
+      "grad_norm": 45.25,
+      "learning_rate": 1.938412642808959e-05,
+      "loss": 0.6612,
+      "step": 3017
+    },
+    {
+      "epoch": 0.18547197640117993,
+      "grad_norm": 50.5,
+      "learning_rate": 1.938332212039368e-05,
+      "loss": 0.5555,
+      "step": 3018
+    },
+    {
+      "epoch": 0.18553343166175024,
+      "grad_norm": 34.25,
+      "learning_rate": 1.9382517304551397e-05,
+      "loss": 0.5172,
+      "step": 3019
+    },
+    {
+      "epoch": 0.18559488692232054,
+      "grad_norm": 37.0,
+      "learning_rate": 1.938171198060632e-05,
+      "loss": 0.4086,
+      "step": 3020
+    },
+    {
+      "epoch": 0.18565634218289084,
+      "grad_norm": 54.0,
+      "learning_rate": 1.938090614860207e-05,
+      "loss": 0.4724,
+      "step": 3021
+    },
+    {
+      "epoch": 0.18571779744346115,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9380099808582278e-05,
+      "loss": 0.3088,
+      "step": 3022
+    },
+    {
+      "epoch": 0.18577925270403148,
+      "grad_norm": 41.25,
+      "learning_rate": 1.937929296059061e-05,
+      "loss": 0.4072,
+      "step": 3023
+    },
+    {
+      "epoch": 0.18584070796460178,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.9378485604670766e-05,
+      "loss": 0.5105,
+      "step": 3024
+    },
+    {
+      "epoch": 0.18590216322517208,
+      "grad_norm": 19.875,
+      "learning_rate": 1.937767774086646e-05,
+      "loss": 0.5776,
+      "step": 3025
+    },
+    {
+      "epoch": 0.1859636184857424,
+      "grad_norm": 21.75,
+      "learning_rate": 1.937686936922145e-05,
+      "loss": 0.5936,
+      "step": 3026
+    },
+    {
+      "epoch": 0.1860250737463127,
+      "grad_norm": 43.25,
+      "learning_rate": 1.937606048977951e-05,
+      "loss": 0.5969,
+      "step": 3027
+    },
+    {
+      "epoch": 0.186086529006883,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9375251102584438e-05,
+      "loss": 0.6166,
+      "step": 3028
+    },
+    {
+      "epoch": 0.1861479842674533,
+      "grad_norm": 17.125,
+      "learning_rate": 1.937444120768007e-05,
+      "loss": 0.5478,
+      "step": 3029
+    },
+    {
+      "epoch": 0.1862094395280236,
+      "grad_norm": 26.875,
+      "learning_rate": 1.9373630805110267e-05,
+      "loss": 0.3757,
+      "step": 3030
+    },
+    {
+      "epoch": 0.1862708947885939,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.937281989491892e-05,
+      "loss": 0.3527,
+      "step": 3031
+    },
+    {
+      "epoch": 0.1863323500491642,
+      "grad_norm": 14.375,
+      "learning_rate": 1.9372008477149928e-05,
+      "loss": 0.4788,
+      "step": 3032
+    },
+    {
+      "epoch": 0.1863938053097345,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9371196551847248e-05,
+      "loss": 0.5742,
+      "step": 3033
+    },
+    {
+      "epoch": 0.18645526057030481,
+      "grad_norm": 24.5,
+      "learning_rate": 1.937038411905484e-05,
+      "loss": 0.5256,
+      "step": 3034
+    },
+    {
+      "epoch": 0.18651671583087512,
+      "grad_norm": 34.75,
+      "learning_rate": 1.9369571178816708e-05,
+      "loss": 0.4345,
+      "step": 3035
+    },
+    {
+      "epoch": 0.18657817109144542,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.936875773117687e-05,
+      "loss": 0.3542,
+      "step": 3036
+    },
+    {
+      "epoch": 0.18663962635201573,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.936794377617938e-05,
+      "loss": 0.3539,
+      "step": 3037
+    },
+    {
+      "epoch": 0.18670108161258603,
+      "grad_norm": 15.5625,
+      "learning_rate": 1.936712931386832e-05,
+      "loss": 0.3697,
+      "step": 3038
+    },
+    {
+      "epoch": 0.18676253687315633,
+      "grad_norm": 16.5,
+      "learning_rate": 1.9366314344287786e-05,
+      "loss": 0.3797,
+      "step": 3039
+    },
+    {
+      "epoch": 0.18682399213372664,
+      "grad_norm": 7.03125,
+      "learning_rate": 1.9365498867481926e-05,
+      "loss": 0.3806,
+      "step": 3040
+    },
+    {
+      "epoch": 0.18688544739429694,
+      "grad_norm": 39.75,
+      "learning_rate": 1.9364682883494892e-05,
+      "loss": 0.4992,
+      "step": 3041
+    },
+    {
+      "epoch": 0.18694690265486727,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9363866392370875e-05,
+      "loss": 0.4558,
+      "step": 3042
+    },
+    {
+      "epoch": 0.18700835791543757,
+      "grad_norm": 13.375,
+      "learning_rate": 1.9363049394154095e-05,
+      "loss": 0.4716,
+      "step": 3043
+    },
+    {
+      "epoch": 0.18706981317600788,
+      "grad_norm": 11.25,
+      "learning_rate": 1.9362231888888793e-05,
+      "loss": 0.3768,
+      "step": 3044
+    },
+    {
+      "epoch": 0.18713126843657818,
+      "grad_norm": 26.25,
+      "learning_rate": 1.936141387661924e-05,
+      "loss": 0.5961,
+      "step": 3045
+    },
+    {
+      "epoch": 0.18719272369714848,
+      "grad_norm": 10.5,
+      "learning_rate": 1.9360595357389735e-05,
+      "loss": 0.3634,
+      "step": 3046
+    },
+    {
+      "epoch": 0.1872541789577188,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9359776331244606e-05,
+      "loss": 0.3673,
+      "step": 3047
+    },
+    {
+      "epoch": 0.1873156342182891,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.9358956798228205e-05,
+      "loss": 0.3366,
+      "step": 3048
+    },
+    {
+      "epoch": 0.1873770894788594,
+      "grad_norm": 19.625,
+      "learning_rate": 1.935813675838491e-05,
+      "loss": 0.6899,
+      "step": 3049
+    },
+    {
+      "epoch": 0.1874385447394297,
+      "grad_norm": 12.0,
+      "learning_rate": 1.9357316211759137e-05,
+      "loss": 0.4356,
+      "step": 3050
+    },
+    {
+      "epoch": 0.1875,
+      "grad_norm": 29.875,
+      "learning_rate": 1.9356495158395317e-05,
+      "loss": 0.412,
+      "step": 3051
+    },
+    {
+      "epoch": 0.1875614552605703,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9355673598337916e-05,
+      "loss": 0.4612,
+      "step": 3052
+    },
+    {
+      "epoch": 0.1876229105211406,
+      "grad_norm": 8.5625,
+      "learning_rate": 1.935485153163142e-05,
+      "loss": 0.4639,
+      "step": 3053
+    },
+    {
+      "epoch": 0.1876843657817109,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9354028958320352e-05,
+      "loss": 0.3553,
+      "step": 3054
+    },
+    {
+      "epoch": 0.1877458210422812,
+      "grad_norm": 26.375,
+      "learning_rate": 1.935320587844926e-05,
+      "loss": 0.4161,
+      "step": 3055
+    },
+    {
+      "epoch": 0.18780727630285152,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9352382292062712e-05,
+      "loss": 0.3405,
+      "step": 3056
+    },
+    {
+      "epoch": 0.18786873156342182,
+      "grad_norm": 25.875,
+      "learning_rate": 1.935155819920531e-05,
+      "loss": 0.4924,
+      "step": 3057
+    },
+    {
+      "epoch": 0.18793018682399212,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9350733599921684e-05,
+      "loss": 0.417,
+      "step": 3058
+    },
+    {
+      "epoch": 0.18799164208456243,
+      "grad_norm": 46.0,
+      "learning_rate": 1.934990849425649e-05,
+      "loss": 0.5806,
+      "step": 3059
+    },
+    {
+      "epoch": 0.18805309734513273,
+      "grad_norm": 87.0,
+      "learning_rate": 1.9349082882254405e-05,
+      "loss": 0.4922,
+      "step": 3060
+    },
+    {
+      "epoch": 0.18811455260570306,
+      "grad_norm": 27.375,
+      "learning_rate": 1.9348256763960146e-05,
+      "loss": 0.3161,
+      "step": 3061
+    },
+    {
+      "epoch": 0.18817600786627336,
+      "grad_norm": 29.0,
+      "learning_rate": 1.9347430139418448e-05,
+      "loss": 0.4742,
+      "step": 3062
+    },
+    {
+      "epoch": 0.18823746312684367,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9346603008674075e-05,
+      "loss": 0.3815,
+      "step": 3063
+    },
+    {
+      "epoch": 0.18829891838741397,
+      "grad_norm": 10.75,
+      "learning_rate": 1.9345775371771826e-05,
+      "loss": 0.5304,
+      "step": 3064
+    },
+    {
+      "epoch": 0.18836037364798427,
+      "grad_norm": 30.5,
+      "learning_rate": 1.934494722875651e-05,
+      "loss": 0.6936,
+      "step": 3065
+    },
+    {
+      "epoch": 0.18842182890855458,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9344118579672987e-05,
+      "loss": 0.5219,
+      "step": 3066
+    },
+    {
+      "epoch": 0.18848328416912488,
+      "grad_norm": 46.75,
+      "learning_rate": 1.9343289424566122e-05,
+      "loss": 0.4943,
+      "step": 3067
+    },
+    {
+      "epoch": 0.18854473942969519,
+      "grad_norm": 28.0,
+      "learning_rate": 1.9342459763480822e-05,
+      "loss": 0.5983,
+      "step": 3068
+    },
+    {
+      "epoch": 0.1886061946902655,
+      "grad_norm": 18.0,
+      "learning_rate": 1.9341629596462015e-05,
+      "loss": 0.4858,
+      "step": 3069
+    },
+    {
+      "epoch": 0.1886676499508358,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.9340798923554657e-05,
+      "loss": 0.4553,
+      "step": 3070
+    },
+    {
+      "epoch": 0.1887291052114061,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.9339967744803735e-05,
+      "loss": 0.4467,
+      "step": 3071
+    },
+    {
+      "epoch": 0.1887905604719764,
+      "grad_norm": 27.0,
+      "learning_rate": 1.933913606025426e-05,
+      "loss": 0.3906,
+      "step": 3072
+    },
+    {
+      "epoch": 0.1888520157325467,
+      "grad_norm": 30.75,
+      "learning_rate": 1.933830386995127e-05,
+      "loss": 0.6018,
+      "step": 3073
+    },
+    {
+      "epoch": 0.188913470993117,
+      "grad_norm": 42.0,
+      "learning_rate": 1.933747117393983e-05,
+      "loss": 0.6741,
+      "step": 3074
+    },
+    {
+      "epoch": 0.1889749262536873,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9336637972265042e-05,
+      "loss": 0.4624,
+      "step": 3075
+    },
+    {
+      "epoch": 0.1890363815142576,
+      "grad_norm": 63.5,
+      "learning_rate": 1.9335804264972018e-05,
+      "loss": 0.521,
+      "step": 3076
+    },
+    {
+      "epoch": 0.18909783677482792,
+      "grad_norm": 36.0,
+      "learning_rate": 1.933497005210591e-05,
+      "loss": 0.3893,
+      "step": 3077
+    },
+    {
+      "epoch": 0.18915929203539822,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.93341353337119e-05,
+      "loss": 0.388,
+      "step": 3078
+    },
+    {
+      "epoch": 0.18922074729596852,
+      "grad_norm": 37.25,
+      "learning_rate": 1.9333300109835182e-05,
+      "loss": 0.4786,
+      "step": 3079
+    },
+    {
+      "epoch": 0.18928220255653885,
+      "grad_norm": 9.1875,
+      "learning_rate": 1.9332464380520993e-05,
+      "loss": 0.3302,
+      "step": 3080
+    },
+    {
+      "epoch": 0.18934365781710916,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.9331628145814587e-05,
+      "loss": 0.4529,
+      "step": 3081
+    },
+    {
+      "epoch": 0.18940511307767946,
+      "grad_norm": 17.5,
+      "learning_rate": 1.9330791405761254e-05,
+      "loss": 0.4259,
+      "step": 3082
+    },
+    {
+      "epoch": 0.18946656833824976,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9329954160406306e-05,
+      "loss": 0.6097,
+      "step": 3083
+    },
+    {
+      "epoch": 0.18952802359882007,
+      "grad_norm": 16.625,
+      "learning_rate": 1.932911640979508e-05,
+      "loss": 0.4654,
+      "step": 3084
+    },
+    {
+      "epoch": 0.18958947885939037,
+      "grad_norm": 47.75,
+      "learning_rate": 1.9328278153972947e-05,
+      "loss": 0.4684,
+      "step": 3085
+    },
+    {
+      "epoch": 0.18965093411996067,
+      "grad_norm": 22.625,
+      "learning_rate": 1.93274393929853e-05,
+      "loss": 0.4892,
+      "step": 3086
+    },
+    {
+      "epoch": 0.18971238938053098,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9326600126877566e-05,
+      "loss": 0.4493,
+      "step": 3087
+    },
+    {
+      "epoch": 0.18977384464110128,
+      "grad_norm": 29.125,
+      "learning_rate": 1.932576035569519e-05,
+      "loss": 0.3125,
+      "step": 3088
+    },
+    {
+      "epoch": 0.18983529990167158,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9324920079483647e-05,
+      "loss": 0.4074,
+      "step": 3089
+    },
+    {
+      "epoch": 0.1898967551622419,
+      "grad_norm": 4.53125,
+      "learning_rate": 1.932407929828845e-05,
+      "loss": 0.3623,
+      "step": 3090
+    },
+    {
+      "epoch": 0.1899582104228122,
+      "grad_norm": 16.75,
+      "learning_rate": 1.9323238012155125e-05,
+      "loss": 0.5931,
+      "step": 3091
+    },
+    {
+      "epoch": 0.1900196656833825,
+      "grad_norm": 23.25,
+      "learning_rate": 1.932239622112923e-05,
+      "loss": 0.4467,
+      "step": 3092
+    },
+    {
+      "epoch": 0.1900811209439528,
+      "grad_norm": 20.25,
+      "learning_rate": 1.9321553925256355e-05,
+      "loss": 0.3878,
+      "step": 3093
+    },
+    {
+      "epoch": 0.1901425762045231,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.932071112458211e-05,
+      "loss": 0.2578,
+      "step": 3094
+    },
+    {
+      "epoch": 0.1902040314650934,
+      "grad_norm": 24.25,
+      "learning_rate": 1.9319867819152144e-05,
+      "loss": 0.5495,
+      "step": 3095
+    },
+    {
+      "epoch": 0.1902654867256637,
+      "grad_norm": 37.0,
+      "learning_rate": 1.9319024009012114e-05,
+      "loss": 0.4262,
+      "step": 3096
+    },
+    {
+      "epoch": 0.190326941986234,
+      "grad_norm": 25.0,
+      "learning_rate": 1.9318179694207726e-05,
+      "loss": 0.3919,
+      "step": 3097
+    },
+    {
+      "epoch": 0.19038839724680431,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.93173348747847e-05,
+      "loss": 0.5086,
+      "step": 3098
+    },
+    {
+      "epoch": 0.19044985250737465,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9316489550788785e-05,
+      "loss": 0.4526,
+      "step": 3099
+    },
+    {
+      "epoch": 0.19051130776794495,
+      "grad_norm": 32.0,
+      "learning_rate": 1.931564372226576e-05,
+      "loss": 0.5635,
+      "step": 3100
+    },
+    {
+      "epoch": 0.19057276302851525,
+      "grad_norm": 33.75,
+      "learning_rate": 1.9314797389261426e-05,
+      "loss": 0.4972,
+      "step": 3101
+    },
+    {
+      "epoch": 0.19063421828908556,
+      "grad_norm": 19.5,
+      "learning_rate": 1.931395055182162e-05,
+      "loss": 0.3919,
+      "step": 3102
+    },
+    {
+      "epoch": 0.19069567354965586,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9313103209992205e-05,
+      "loss": 0.3942,
+      "step": 3103
+    },
+    {
+      "epoch": 0.19075712881022616,
+      "grad_norm": 28.25,
+      "learning_rate": 1.931225536381906e-05,
+      "loss": 0.3897,
+      "step": 3104
+    },
+    {
+      "epoch": 0.19081858407079647,
+      "grad_norm": 21.25,
+      "learning_rate": 1.931140701334811e-05,
+      "loss": 0.3934,
+      "step": 3105
+    },
+    {
+      "epoch": 0.19088003933136677,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9310558158625286e-05,
+      "loss": 0.475,
+      "step": 3106
+    },
+    {
+      "epoch": 0.19094149459193707,
+      "grad_norm": 22.375,
+      "learning_rate": 1.930970879969656e-05,
+      "loss": 0.5783,
+      "step": 3107
+    },
+    {
+      "epoch": 0.19100294985250738,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9308858936607934e-05,
+      "loss": 0.5278,
+      "step": 3108
+    },
+    {
+      "epoch": 0.19106440511307768,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9308008569405424e-05,
+      "loss": 0.3673,
+      "step": 3109
+    },
+    {
+      "epoch": 0.19112586037364798,
+      "grad_norm": 29.625,
+      "learning_rate": 1.9307157698135085e-05,
+      "loss": 0.3967,
+      "step": 3110
+    },
+    {
+      "epoch": 0.19118731563421829,
+      "grad_norm": 16.0,
+      "learning_rate": 1.9306306322842994e-05,
+      "loss": 0.4106,
+      "step": 3111
+    },
+    {
+      "epoch": 0.1912487708947886,
+      "grad_norm": 16.0,
+      "learning_rate": 1.930545444357526e-05,
+      "loss": 0.4621,
+      "step": 3112
+    },
+    {
+      "epoch": 0.1913102261553589,
+      "grad_norm": 8.25,
+      "learning_rate": 1.9304602060378013e-05,
+      "loss": 0.4185,
+      "step": 3113
+    },
+    {
+      "epoch": 0.1913716814159292,
+      "grad_norm": 18.125,
+      "learning_rate": 1.9303749173297412e-05,
+      "loss": 0.4325,
+      "step": 3114
+    },
+    {
+      "epoch": 0.1914331366764995,
+      "grad_norm": 30.0,
+      "learning_rate": 1.9302895782379648e-05,
+      "loss": 0.5012,
+      "step": 3115
+    },
+    {
+      "epoch": 0.1914945919370698,
+      "grad_norm": 19.5,
+      "learning_rate": 1.930204188767093e-05,
+      "loss": 0.3983,
+      "step": 3116
+    },
+    {
+      "epoch": 0.1915560471976401,
+      "grad_norm": 31.125,
+      "learning_rate": 1.9301187489217508e-05,
+      "loss": 0.4154,
+      "step": 3117
+    },
+    {
+      "epoch": 0.1916175024582104,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9300332587065644e-05,
+      "loss": 0.3312,
+      "step": 3118
+    },
+    {
+      "epoch": 0.19167895771878074,
+      "grad_norm": 10.875,
+      "learning_rate": 1.9299477181261633e-05,
+      "loss": 0.3302,
+      "step": 3119
+    },
+    {
+      "epoch": 0.19174041297935104,
+      "grad_norm": 12.25,
+      "learning_rate": 1.929862127185181e-05,
+      "loss": 0.3188,
+      "step": 3120
+    },
+    {
+      "epoch": 0.19180186823992135,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9297764858882516e-05,
+      "loss": 0.5155,
+      "step": 3121
+    },
+    {
+      "epoch": 0.19186332350049165,
+      "grad_norm": 17.125,
+      "learning_rate": 1.9296907942400133e-05,
+      "loss": 0.3994,
+      "step": 3122
+    },
+    {
+      "epoch": 0.19192477876106195,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.9296050522451067e-05,
+      "loss": 0.3074,
+      "step": 3123
+    },
+    {
+      "epoch": 0.19198623402163226,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.9295192599081747e-05,
+      "loss": 0.3895,
+      "step": 3124
+    },
+    {
+      "epoch": 0.19204768928220256,
+      "grad_norm": 25.25,
+      "learning_rate": 1.9294334172338637e-05,
+      "loss": 0.4576,
+      "step": 3125
+    },
+    {
+      "epoch": 0.19210914454277286,
+      "grad_norm": 17.0,
+      "learning_rate": 1.9293475242268224e-05,
+      "loss": 0.2665,
+      "step": 3126
+    },
+    {
+      "epoch": 0.19217059980334317,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9292615808917027e-05,
+      "loss": 0.4694,
+      "step": 3127
+    },
+    {
+      "epoch": 0.19223205506391347,
+      "grad_norm": 15.125,
+      "learning_rate": 1.9291755872331576e-05,
+      "loss": 0.3566,
+      "step": 3128
+    },
+    {
+      "epoch": 0.19229351032448377,
+      "grad_norm": 30.0,
+      "learning_rate": 1.929089543255845e-05,
+      "loss": 0.4633,
+      "step": 3129
+    },
+    {
+      "epoch": 0.19235496558505408,
+      "grad_norm": 22.5,
+      "learning_rate": 1.9290034489644247e-05,
+      "loss": 0.5858,
+      "step": 3130
+    },
+    {
+      "epoch": 0.19241642084562438,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9289173043635584e-05,
+      "loss": 0.4043,
+      "step": 3131
+    },
+    {
+      "epoch": 0.19247787610619468,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.9288311094579112e-05,
+      "loss": 0.4139,
+      "step": 3132
+    },
+    {
+      "epoch": 0.192539331366765,
+      "grad_norm": 12.0,
+      "learning_rate": 1.9287448642521513e-05,
+      "loss": 0.3883,
+      "step": 3133
+    },
+    {
+      "epoch": 0.1926007866273353,
+      "grad_norm": 4.0,
+      "learning_rate": 1.9286585687509492e-05,
+      "loss": 0.4217,
+      "step": 3134
+    },
+    {
+      "epoch": 0.1926622418879056,
+      "grad_norm": 17.875,
+      "learning_rate": 1.9285722229589782e-05,
+      "loss": 0.2882,
+      "step": 3135
+    },
+    {
+      "epoch": 0.1927236971484759,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9284858268809135e-05,
+      "loss": 0.4638,
+      "step": 3136
+    },
+    {
+      "epoch": 0.1927851524090462,
+      "grad_norm": 27.125,
+      "learning_rate": 1.9283993805214353e-05,
+      "loss": 0.6033,
+      "step": 3137
+    },
+    {
+      "epoch": 0.19284660766961653,
+      "grad_norm": 49.25,
+      "learning_rate": 1.9283128838852237e-05,
+      "loss": 0.5463,
+      "step": 3138
+    },
+    {
+      "epoch": 0.19290806293018684,
+      "grad_norm": 23.5,
+      "learning_rate": 1.9282263369769633e-05,
+      "loss": 0.5622,
+      "step": 3139
+    },
+    {
+      "epoch": 0.19296951819075714,
+      "grad_norm": 20.0,
+      "learning_rate": 1.9281397398013416e-05,
+      "loss": 0.5249,
+      "step": 3140
+    },
+    {
+      "epoch": 0.19303097345132744,
+      "grad_norm": 50.75,
+      "learning_rate": 1.928053092363047e-05,
+      "loss": 0.4112,
+      "step": 3141
+    },
+    {
+      "epoch": 0.19309242871189775,
+      "grad_norm": 36.25,
+      "learning_rate": 1.927966394666773e-05,
+      "loss": 0.4306,
+      "step": 3142
+    },
+    {
+      "epoch": 0.19315388397246805,
+      "grad_norm": 27.125,
+      "learning_rate": 1.9278796467172137e-05,
+      "loss": 0.3881,
+      "step": 3143
+    },
+    {
+      "epoch": 0.19321533923303835,
+      "grad_norm": 32.75,
+      "learning_rate": 1.9277928485190674e-05,
+      "loss": 0.3317,
+      "step": 3144
+    },
+    {
+      "epoch": 0.19327679449360866,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9277060000770342e-05,
+      "loss": 0.7708,
+      "step": 3145
+    },
+    {
+      "epoch": 0.19333824975417896,
+      "grad_norm": 31.25,
+      "learning_rate": 1.927619101395818e-05,
+      "loss": 0.595,
+      "step": 3146
+    },
+    {
+      "epoch": 0.19339970501474926,
+      "grad_norm": 63.0,
+      "learning_rate": 1.927532152480124e-05,
+      "loss": 0.7295,
+      "step": 3147
+    },
+    {
+      "epoch": 0.19346116027531957,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9274451533346617e-05,
+      "loss": 0.6259,
+      "step": 3148
+    },
+    {
+      "epoch": 0.19352261553588987,
+      "grad_norm": 30.375,
+      "learning_rate": 1.9273581039641412e-05,
+      "loss": 0.4116,
+      "step": 3149
+    },
+    {
+      "epoch": 0.19358407079646017,
+      "grad_norm": 15.25,
+      "learning_rate": 1.9272710043732775e-05,
+      "loss": 0.3198,
+      "step": 3150
+    },
+    {
+      "epoch": 0.19364552605703048,
+      "grad_norm": 16.25,
+      "learning_rate": 1.9271838545667876e-05,
+      "loss": 0.3557,
+      "step": 3151
+    },
+    {
+      "epoch": 0.19370698131760078,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9270966545493903e-05,
+      "loss": 0.4748,
+      "step": 3152
+    },
+    {
+      "epoch": 0.19376843657817108,
+      "grad_norm": 24.0,
+      "learning_rate": 1.927009404325808e-05,
+      "loss": 0.5944,
+      "step": 3153
+    },
+    {
+      "epoch": 0.19382989183874139,
+      "grad_norm": 18.375,
+      "learning_rate": 1.9269221039007666e-05,
+      "loss": 0.6175,
+      "step": 3154
+    },
+    {
+      "epoch": 0.1938913470993117,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9268347532789926e-05,
+      "loss": 0.4974,
+      "step": 3155
+    },
+    {
+      "epoch": 0.193952802359882,
+      "grad_norm": 26.25,
+      "learning_rate": 1.9267473524652168e-05,
+      "loss": 0.4146,
+      "step": 3156
+    },
+    {
+      "epoch": 0.19401425762045232,
+      "grad_norm": 29.25,
+      "learning_rate": 1.9266599014641724e-05,
+      "loss": 0.4516,
+      "step": 3157
+    },
+    {
+      "epoch": 0.19407571288102263,
+      "grad_norm": 44.25,
+      "learning_rate": 1.9265724002805952e-05,
+      "loss": 0.4985,
+      "step": 3158
+    },
+    {
+      "epoch": 0.19413716814159293,
+      "grad_norm": 5.0625,
+      "learning_rate": 1.9264848489192237e-05,
+      "loss": 0.3678,
+      "step": 3159
+    },
+    {
+      "epoch": 0.19419862340216323,
+      "grad_norm": 9.1875,
+      "learning_rate": 1.9263972473847995e-05,
+      "loss": 0.3074,
+      "step": 3160
+    },
+    {
+      "epoch": 0.19426007866273354,
+      "grad_norm": 29.125,
+      "learning_rate": 1.926309595682066e-05,
+      "loss": 0.5712,
+      "step": 3161
+    },
+    {
+      "epoch": 0.19432153392330384,
+      "grad_norm": 26.625,
+      "learning_rate": 1.9262218938157703e-05,
+      "loss": 0.4782,
+      "step": 3162
+    },
+    {
+      "epoch": 0.19438298918387414,
+      "grad_norm": 33.0,
+      "learning_rate": 1.9261341417906622e-05,
+      "loss": 0.4408,
+      "step": 3163
+    },
+    {
+      "epoch": 0.19444444444444445,
+      "grad_norm": 27.625,
+      "learning_rate": 1.926046339611493e-05,
+      "loss": 0.3974,
+      "step": 3164
+    },
+    {
+      "epoch": 0.19450589970501475,
+      "grad_norm": 26.375,
+      "learning_rate": 1.925958487283018e-05,
+      "loss": 0.4634,
+      "step": 3165
+    },
+    {
+      "epoch": 0.19456735496558505,
+      "grad_norm": 7.6875,
+      "learning_rate": 1.925870584809995e-05,
+      "loss": 0.3452,
+      "step": 3166
+    },
+    {
+      "epoch": 0.19462881022615536,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9257826321971838e-05,
+      "loss": 0.3272,
+      "step": 3167
+    },
+    {
+      "epoch": 0.19469026548672566,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9256946294493475e-05,
+      "loss": 0.266,
+      "step": 3168
+    },
+    {
+      "epoch": 0.19475172074729596,
+      "grad_norm": 34.75,
+      "learning_rate": 1.9256065765712524e-05,
+      "loss": 0.5639,
+      "step": 3169
+    },
+    {
+      "epoch": 0.19481317600786627,
+      "grad_norm": 29.75,
+      "learning_rate": 1.925518473567666e-05,
+      "loss": 0.7866,
+      "step": 3170
+    },
+    {
+      "epoch": 0.19487463126843657,
+      "grad_norm": 31.625,
+      "learning_rate": 1.9254303204433602e-05,
+      "loss": 0.5783,
+      "step": 3171
+    },
+    {
+      "epoch": 0.19493608652900687,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.9253421172031086e-05,
+      "loss": 0.5233,
+      "step": 3172
+    },
+    {
+      "epoch": 0.19499754178957718,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9252538638516878e-05,
+      "loss": 0.4656,
+      "step": 3173
+    },
+    {
+      "epoch": 0.19505899705014748,
+      "grad_norm": 19.875,
+      "learning_rate": 1.925165560393877e-05,
+      "loss": 0.5181,
+      "step": 3174
+    },
+    {
+      "epoch": 0.19512045231071778,
+      "grad_norm": 2.09375,
+      "learning_rate": 1.925077206834458e-05,
+      "loss": 0.3447,
+      "step": 3175
+    },
+    {
+      "epoch": 0.19518190757128812,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.924988803178216e-05,
+      "loss": 0.3941,
+      "step": 3176
+    },
+    {
+      "epoch": 0.19524336283185842,
+      "grad_norm": 25.375,
+      "learning_rate": 1.9249003494299385e-05,
+      "loss": 0.7266,
+      "step": 3177
+    },
+    {
+      "epoch": 0.19530481809242872,
+      "grad_norm": 25.25,
+      "learning_rate": 1.9248118455944153e-05,
+      "loss": 0.5372,
+      "step": 3178
+    },
+    {
+      "epoch": 0.19536627335299903,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.9247232916764388e-05,
+      "loss": 0.4471,
+      "step": 3179
+    },
+    {
+      "epoch": 0.19542772861356933,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9246346876808054e-05,
+      "loss": 0.4106,
+      "step": 3180
+    },
+    {
+      "epoch": 0.19548918387413963,
+      "grad_norm": 13.125,
+      "learning_rate": 1.9245460336123136e-05,
+      "loss": 0.6621,
+      "step": 3181
+    },
+    {
+      "epoch": 0.19555063913470994,
+      "grad_norm": 16.875,
+      "learning_rate": 1.924457329475763e-05,
+      "loss": 0.4883,
+      "step": 3182
+    },
+    {
+      "epoch": 0.19561209439528024,
+      "grad_norm": 11.5,
+      "learning_rate": 1.9243685752759585e-05,
+      "loss": 0.5566,
+      "step": 3183
+    },
+    {
+      "epoch": 0.19567354965585054,
+      "grad_norm": 20.875,
+      "learning_rate": 1.924279771017706e-05,
+      "loss": 0.541,
+      "step": 3184
+    },
+    {
+      "epoch": 0.19573500491642085,
+      "grad_norm": 20.125,
+      "learning_rate": 1.9241909167058154e-05,
+      "loss": 0.58,
+      "step": 3185
+    },
+    {
+      "epoch": 0.19579646017699115,
+      "grad_norm": 27.875,
+      "learning_rate": 1.9241020123450972e-05,
+      "loss": 0.569,
+      "step": 3186
+    },
+    {
+      "epoch": 0.19585791543756145,
+      "grad_norm": 16.375,
+      "learning_rate": 1.924013057940367e-05,
+      "loss": 0.4874,
+      "step": 3187
+    },
+    {
+      "epoch": 0.19591937069813176,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9239240534964417e-05,
+      "loss": 0.4432,
+      "step": 3188
+    },
+    {
+      "epoch": 0.19598082595870206,
+      "grad_norm": 12.25,
+      "learning_rate": 1.923834999018141e-05,
+      "loss": 0.5229,
+      "step": 3189
+    },
+    {
+      "epoch": 0.19604228121927236,
+      "grad_norm": 45.25,
+      "learning_rate": 1.923745894510288e-05,
+      "loss": 0.493,
+      "step": 3190
+    },
+    {
+      "epoch": 0.19610373647984267,
+      "grad_norm": 8.5,
+      "learning_rate": 1.9236567399777086e-05,
+      "loss": 0.3939,
+      "step": 3191
+    },
+    {
+      "epoch": 0.19616519174041297,
+      "grad_norm": 8.5,
+      "learning_rate": 1.9235675354252293e-05,
+      "loss": 0.3573,
+      "step": 3192
+    },
+    {
+      "epoch": 0.19622664700098327,
+      "grad_norm": 49.75,
+      "learning_rate": 1.9234782808576823e-05,
+      "loss": 0.442,
+      "step": 3193
+    },
+    {
+      "epoch": 0.19628810226155358,
+      "grad_norm": 48.5,
+      "learning_rate": 1.9233889762799006e-05,
+      "loss": 0.4996,
+      "step": 3194
+    },
+    {
+      "epoch": 0.1963495575221239,
+      "grad_norm": 6.21875,
+      "learning_rate": 1.9232996216967204e-05,
+      "loss": 0.2889,
+      "step": 3195
+    },
+    {
+      "epoch": 0.1964110127826942,
+      "grad_norm": 37.5,
+      "learning_rate": 1.923210217112981e-05,
+      "loss": 0.3482,
+      "step": 3196
+    },
+    {
+      "epoch": 0.19647246804326451,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.9231207625335235e-05,
+      "loss": 0.4698,
+      "step": 3197
+    },
+    {
+      "epoch": 0.19653392330383482,
+      "grad_norm": 4.375,
+      "learning_rate": 1.9230312579631928e-05,
+      "loss": 0.2884,
+      "step": 3198
+    },
+    {
+      "epoch": 0.19659537856440512,
+      "grad_norm": 12.1875,
+      "learning_rate": 1.9229417034068352e-05,
+      "loss": 0.226,
+      "step": 3199
+    },
+    {
+      "epoch": 0.19665683382497542,
+      "grad_norm": 5.34375,
+      "learning_rate": 1.922852098869301e-05,
+      "loss": 0.3132,
+      "step": 3200
+    },
+    {
+      "epoch": 0.19671828908554573,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9227624443554425e-05,
+      "loss": 0.4517,
+      "step": 3201
+    },
+    {
+      "epoch": 0.19677974434611603,
+      "grad_norm": 34.0,
+      "learning_rate": 1.922672739870115e-05,
+      "loss": 0.4096,
+      "step": 3202
+    },
+    {
+      "epoch": 0.19684119960668633,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.9225829854181764e-05,
+      "loss": 0.353,
+      "step": 3203
+    },
+    {
+      "epoch": 0.19690265486725664,
+      "grad_norm": 6.8125,
+      "learning_rate": 1.922493181004487e-05,
+      "loss": 0.2855,
+      "step": 3204
+    },
+    {
+      "epoch": 0.19696411012782694,
+      "grad_norm": 28.875,
+      "learning_rate": 1.9224033266339103e-05,
+      "loss": 0.3964,
+      "step": 3205
+    },
+    {
+      "epoch": 0.19702556538839724,
+      "grad_norm": 11.5,
+      "learning_rate": 1.9223134223113122e-05,
+      "loss": 0.3643,
+      "step": 3206
+    },
+    {
+      "epoch": 0.19708702064896755,
+      "grad_norm": 3.71875,
+      "learning_rate": 1.9222234680415614e-05,
+      "loss": 0.2807,
+      "step": 3207
+    },
+    {
+      "epoch": 0.19714847590953785,
+      "grad_norm": 10.25,
+      "learning_rate": 1.9221334638295296e-05,
+      "loss": 0.36,
+      "step": 3208
+    },
+    {
+      "epoch": 0.19720993117010815,
+      "grad_norm": 43.75,
+      "learning_rate": 1.9220434096800903e-05,
+      "loss": 0.73,
+      "step": 3209
+    },
+    {
+      "epoch": 0.19727138643067846,
+      "grad_norm": 51.5,
+      "learning_rate": 1.921953305598121e-05,
+      "loss": 0.4778,
+      "step": 3210
+    },
+    {
+      "epoch": 0.19733284169124876,
+      "grad_norm": 34.0,
+      "learning_rate": 1.9218631515885007e-05,
+      "loss": 0.564,
+      "step": 3211
+    },
+    {
+      "epoch": 0.19739429695181906,
+      "grad_norm": 14.125,
+      "learning_rate": 1.921772947656112e-05,
+      "loss": 0.4783,
+      "step": 3212
+    },
+    {
+      "epoch": 0.19745575221238937,
+      "grad_norm": 9.75,
+      "learning_rate": 1.9216826938058394e-05,
+      "loss": 0.4605,
+      "step": 3213
+    },
+    {
+      "epoch": 0.19751720747295967,
+      "grad_norm": 18.875,
+      "learning_rate": 1.921592390042571e-05,
+      "loss": 0.4696,
+      "step": 3214
+    },
+    {
+      "epoch": 0.19757866273353,
+      "grad_norm": 55.75,
+      "learning_rate": 1.9215020363711966e-05,
+      "loss": 0.6129,
+      "step": 3215
+    },
+    {
+      "epoch": 0.1976401179941003,
+      "grad_norm": 25.25,
+      "learning_rate": 1.9214116327966095e-05,
+      "loss": 0.4371,
+      "step": 3216
+    },
+    {
+      "epoch": 0.1977015732546706,
+      "grad_norm": 14.125,
+      "learning_rate": 1.9213211793237056e-05,
+      "loss": 0.2968,
+      "step": 3217
+    },
+    {
+      "epoch": 0.1977630285152409,
+      "grad_norm": 16.875,
+      "learning_rate": 1.921230675957383e-05,
+      "loss": 0.4517,
+      "step": 3218
+    },
+    {
+      "epoch": 0.19782448377581122,
+      "grad_norm": 13.75,
+      "learning_rate": 1.9211401227025433e-05,
+      "loss": 0.2866,
+      "step": 3219
+    },
+    {
+      "epoch": 0.19788593903638152,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9210495195640895e-05,
+      "loss": 0.5418,
+      "step": 3220
+    },
+    {
+      "epoch": 0.19794739429695182,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.9209588665469294e-05,
+      "loss": 0.4282,
+      "step": 3221
+    },
+    {
+      "epoch": 0.19800884955752213,
+      "grad_norm": 26.375,
+      "learning_rate": 1.9208681636559713e-05,
+      "loss": 0.3204,
+      "step": 3222
+    },
+    {
+      "epoch": 0.19807030481809243,
+      "grad_norm": 31.0,
+      "learning_rate": 1.9207774108961273e-05,
+      "loss": 0.4528,
+      "step": 3223
+    },
+    {
+      "epoch": 0.19813176007866273,
+      "grad_norm": 5.53125,
+      "learning_rate": 1.9206866082723122e-05,
+      "loss": 0.3949,
+      "step": 3224
+    },
+    {
+      "epoch": 0.19819321533923304,
+      "grad_norm": 14.875,
+      "learning_rate": 1.9205957557894433e-05,
+      "loss": 0.352,
+      "step": 3225
+    },
+    {
+      "epoch": 0.19825467059980334,
+      "grad_norm": 6.53125,
+      "learning_rate": 1.9205048534524405e-05,
+      "loss": 0.2267,
+      "step": 3226
+    },
+    {
+      "epoch": 0.19831612586037364,
+      "grad_norm": 8.125,
+      "learning_rate": 1.9204139012662266e-05,
+      "loss": 0.3691,
+      "step": 3227
+    },
+    {
+      "epoch": 0.19837758112094395,
+      "grad_norm": 20.5,
+      "learning_rate": 1.9203228992357275e-05,
+      "loss": 0.3678,
+      "step": 3228
+    },
+    {
+      "epoch": 0.19843903638151425,
+      "grad_norm": 18.625,
+      "learning_rate": 1.9202318473658707e-05,
+      "loss": 0.39,
+      "step": 3229
+    },
+    {
+      "epoch": 0.19850049164208455,
+      "grad_norm": 7.4375,
+      "learning_rate": 1.920140745661587e-05,
+      "loss": 0.2961,
+      "step": 3230
+    },
+    {
+      "epoch": 0.19856194690265486,
+      "grad_norm": 24.75,
+      "learning_rate": 1.9200495941278105e-05,
+      "loss": 0.3088,
+      "step": 3231
+    },
+    {
+      "epoch": 0.19862340216322516,
+      "grad_norm": 8.75,
+      "learning_rate": 1.9199583927694775e-05,
+      "loss": 0.2132,
+      "step": 3232
+    },
+    {
+      "epoch": 0.19868485742379546,
+      "grad_norm": 41.25,
+      "learning_rate": 1.919867141591526e-05,
+      "loss": 0.4536,
+      "step": 3233
+    },
+    {
+      "epoch": 0.1987463126843658,
+      "grad_norm": 25.625,
+      "learning_rate": 1.9197758405988986e-05,
+      "loss": 0.597,
+      "step": 3234
+    },
+    {
+      "epoch": 0.1988077679449361,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9196844897965393e-05,
+      "loss": 0.4037,
+      "step": 3235
+    },
+    {
+      "epoch": 0.1988692232055064,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9195930891893946e-05,
+      "loss": 0.3735,
+      "step": 3236
+    },
+    {
+      "epoch": 0.1989306784660767,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.9195016387824154e-05,
+      "loss": 0.4605,
+      "step": 3237
+    },
+    {
+      "epoch": 0.198992133726647,
+      "grad_norm": 30.5,
+      "learning_rate": 1.919410138580553e-05,
+      "loss": 0.3089,
+      "step": 3238
+    },
+    {
+      "epoch": 0.1990535889872173,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.9193185885887633e-05,
+      "loss": 0.3694,
+      "step": 3239
+    },
+    {
+      "epoch": 0.19911504424778761,
+      "grad_norm": 8.375,
+      "learning_rate": 1.9192269888120036e-05,
+      "loss": 0.3337,
+      "step": 3240
+    },
+    {
+      "epoch": 0.19917649950835792,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9191353392552346e-05,
+      "loss": 0.6298,
+      "step": 3241
+    },
+    {
+      "epoch": 0.19923795476892822,
+      "grad_norm": 49.0,
+      "learning_rate": 1.9190436399234194e-05,
+      "loss": 0.826,
+      "step": 3242
+    },
+    {
+      "epoch": 0.19929941002949852,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9189518908215244e-05,
+      "loss": 0.8753,
+      "step": 3243
+    },
+    {
+      "epoch": 0.19936086529006883,
+      "grad_norm": 43.0,
+      "learning_rate": 1.9188600919545176e-05,
+      "loss": 0.5461,
+      "step": 3244
+    },
+    {
+      "epoch": 0.19942232055063913,
+      "grad_norm": 14.5,
+      "learning_rate": 1.9187682433273704e-05,
+      "loss": 0.5653,
+      "step": 3245
+    },
+    {
+      "epoch": 0.19948377581120944,
+      "grad_norm": 38.25,
+      "learning_rate": 1.9186763449450572e-05,
+      "loss": 0.5067,
+      "step": 3246
+    },
+    {
+      "epoch": 0.19954523107177974,
+      "grad_norm": 19.75,
+      "learning_rate": 1.9185843968125543e-05,
+      "loss": 0.515,
+      "step": 3247
+    },
+    {
+      "epoch": 0.19960668633235004,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.918492398934841e-05,
+      "loss": 0.6222,
+      "step": 3248
+    },
+    {
+      "epoch": 0.19966814159292035,
+      "grad_norm": 35.0,
+      "learning_rate": 1.9184003513168995e-05,
+      "loss": 0.58,
+      "step": 3249
+    },
+    {
+      "epoch": 0.19972959685349065,
+      "grad_norm": 29.625,
+      "learning_rate": 1.918308253963715e-05,
+      "loss": 0.4658,
+      "step": 3250
+    },
+    {
+      "epoch": 0.19979105211406095,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9182161068802742e-05,
+      "loss": 0.4651,
+      "step": 3251
+    },
+    {
+      "epoch": 0.19985250737463126,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9181239100715675e-05,
+      "loss": 0.4433,
+      "step": 3252
+    },
+    {
+      "epoch": 0.1999139626352016,
+      "grad_norm": 32.5,
+      "learning_rate": 1.9180316635425883e-05,
+      "loss": 0.3804,
+      "step": 3253
+    },
+    {
+      "epoch": 0.1999754178957719,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.9179393672983315e-05,
+      "loss": 0.3226,
+      "step": 3254
+    },
+    {
+      "epoch": 0.2000368731563422,
+      "grad_norm": 21.625,
+      "learning_rate": 1.9178470213437955e-05,
+      "loss": 0.3348,
+      "step": 3255
+    },
+    {
+      "epoch": 0.2000983284169125,
+      "grad_norm": 10.5,
+      "learning_rate": 1.9177546256839814e-05,
+      "loss": 0.4113,
+      "step": 3256
+    },
+    {
+      "epoch": 0.2001597836774828,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.9176621803238924e-05,
+      "loss": 0.5429,
+      "step": 3257
+    },
+    {
+      "epoch": 0.2002212389380531,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9175696852685353e-05,
+      "loss": 0.5471,
+      "step": 3258
+    },
+    {
+      "epoch": 0.2002826941986234,
+      "grad_norm": 25.875,
+      "learning_rate": 1.9174771405229187e-05,
+      "loss": 0.5805,
+      "step": 3259
+    },
+    {
+      "epoch": 0.2003441494591937,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9173845460920546e-05,
+      "loss": 0.5394,
+      "step": 3260
+    },
+    {
+      "epoch": 0.200405604719764,
+      "grad_norm": 14.5,
+      "learning_rate": 1.9172919019809572e-05,
+      "loss": 0.4442,
+      "step": 3261
+    },
+    {
+      "epoch": 0.20046705998033432,
+      "grad_norm": 30.625,
+      "learning_rate": 1.9171992081946436e-05,
+      "loss": 0.3667,
+      "step": 3262
+    },
+    {
+      "epoch": 0.20052851524090462,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.9171064647381334e-05,
+      "loss": 0.44,
+      "step": 3263
+    },
+    {
+      "epoch": 0.20058997050147492,
+      "grad_norm": 7.875,
+      "learning_rate": 1.9170136716164497e-05,
+      "loss": 0.351,
+      "step": 3264
+    },
+    {
+      "epoch": 0.20065142576204523,
+      "grad_norm": 48.75,
+      "learning_rate": 1.9169208288346168e-05,
+      "loss": 0.5977,
+      "step": 3265
+    },
+    {
+      "epoch": 0.20071288102261553,
+      "grad_norm": 16.625,
+      "learning_rate": 1.9168279363976627e-05,
+      "loss": 0.5756,
+      "step": 3266
+    },
+    {
+      "epoch": 0.20077433628318583,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.916734994310618e-05,
+      "loss": 0.5098,
+      "step": 3267
+    },
+    {
+      "epoch": 0.20083579154375614,
+      "grad_norm": 33.25,
+      "learning_rate": 1.9166420025785165e-05,
+      "loss": 0.4842,
+      "step": 3268
+    },
+    {
+      "epoch": 0.20089724680432644,
+      "grad_norm": 17.625,
+      "learning_rate": 1.9165489612063932e-05,
+      "loss": 0.5166,
+      "step": 3269
+    },
+    {
+      "epoch": 0.20095870206489674,
+      "grad_norm": 19.125,
+      "learning_rate": 1.9164558701992875e-05,
+      "loss": 0.4537,
+      "step": 3270
+    },
+    {
+      "epoch": 0.20102015732546705,
+      "grad_norm": 38.25,
+      "learning_rate": 1.9163627295622397e-05,
+      "loss": 0.5321,
+      "step": 3271
+    },
+    {
+      "epoch": 0.20108161258603738,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9162695393002947e-05,
+      "loss": 0.4797,
+      "step": 3272
+    },
+    {
+      "epoch": 0.20114306784660768,
+      "grad_norm": 20.125,
+      "learning_rate": 1.9161762994184986e-05,
+      "loss": 0.5974,
+      "step": 3273
+    },
+    {
+      "epoch": 0.20120452310717798,
+      "grad_norm": 9.75,
+      "learning_rate": 1.9160830099219007e-05,
+      "loss": 0.5726,
+      "step": 3274
+    },
+    {
+      "epoch": 0.2012659783677483,
+      "grad_norm": 27.0,
+      "learning_rate": 1.9159896708155534e-05,
+      "loss": 0.6421,
+      "step": 3275
+    },
+    {
+      "epoch": 0.2013274336283186,
+      "grad_norm": 33.25,
+      "learning_rate": 1.9158962821045113e-05,
+      "loss": 0.6594,
+      "step": 3276
+    },
+    {
+      "epoch": 0.2013888888888889,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.9158028437938316e-05,
+      "loss": 0.4718,
+      "step": 3277
+    },
+    {
+      "epoch": 0.2014503441494592,
+      "grad_norm": 21.25,
+      "learning_rate": 1.9157093558885743e-05,
+      "loss": 0.5557,
+      "step": 3278
+    },
+    {
+      "epoch": 0.2015117994100295,
+      "grad_norm": 20.5,
+      "learning_rate": 1.9156158183938026e-05,
+      "loss": 0.5225,
+      "step": 3279
+    },
+    {
+      "epoch": 0.2015732546705998,
+      "grad_norm": 8.8125,
+      "learning_rate": 1.9155222313145817e-05,
+      "loss": 0.4436,
+      "step": 3280
+    },
+    {
+      "epoch": 0.2016347099311701,
+      "grad_norm": 10.75,
+      "learning_rate": 1.9154285946559792e-05,
+      "loss": 0.5134,
+      "step": 3281
+    },
+    {
+      "epoch": 0.2016961651917404,
+      "grad_norm": 44.75,
+      "learning_rate": 1.9153349084230666e-05,
+      "loss": 0.6567,
+      "step": 3282
+    },
+    {
+      "epoch": 0.20175762045231072,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9152411726209176e-05,
+      "loss": 0.589,
+      "step": 3283
+    },
+    {
+      "epoch": 0.20181907571288102,
+      "grad_norm": 23.375,
+      "learning_rate": 1.9151473872546082e-05,
+      "loss": 0.4641,
+      "step": 3284
+    },
+    {
+      "epoch": 0.20188053097345132,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9150535523292165e-05,
+      "loss": 0.4986,
+      "step": 3285
+    },
+    {
+      "epoch": 0.20194198623402163,
+      "grad_norm": 24.625,
+      "learning_rate": 1.914959667849825e-05,
+      "loss": 0.64,
+      "step": 3286
+    },
+    {
+      "epoch": 0.20200344149459193,
+      "grad_norm": 41.25,
+      "learning_rate": 1.9148657338215175e-05,
+      "loss": 0.6606,
+      "step": 3287
+    },
+    {
+      "epoch": 0.20206489675516223,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.914771750249381e-05,
+      "loss": 0.3861,
+      "step": 3288
+    },
+    {
+      "epoch": 0.20212635201573254,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.914677717138505e-05,
+      "loss": 0.4235,
+      "step": 3289
+    },
+    {
+      "epoch": 0.20218780727630284,
+      "grad_norm": 13.5,
+      "learning_rate": 1.9145836344939823e-05,
+      "loss": 0.3992,
+      "step": 3290
+    },
+    {
+      "epoch": 0.20224926253687317,
+      "grad_norm": 12.8125,
+      "learning_rate": 1.9144895023209072e-05,
+      "loss": 0.3946,
+      "step": 3291
+    },
+    {
+      "epoch": 0.20231071779744347,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9143953206243778e-05,
+      "loss": 0.468,
+      "step": 3292
+    },
+    {
+      "epoch": 0.20237217305801378,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9143010894094938e-05,
+      "loss": 0.394,
+      "step": 3293
+    },
+    {
+      "epoch": 0.20243362831858408,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9142068086813593e-05,
+      "loss": 0.3944,
+      "step": 3294
+    },
+    {
+      "epoch": 0.20249508357915438,
+      "grad_norm": 9.5,
+      "learning_rate": 1.914112478445079e-05,
+      "loss": 0.3043,
+      "step": 3295
+    },
+    {
+      "epoch": 0.2025565388397247,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.914018098705762e-05,
+      "loss": 0.3684,
+      "step": 3296
+    },
+    {
+      "epoch": 0.202617994100295,
+      "grad_norm": 45.75,
+      "learning_rate": 1.9139236694685184e-05,
+      "loss": 0.7134,
+      "step": 3297
+    },
+    {
+      "epoch": 0.2026794493608653,
+      "grad_norm": 50.5,
+      "learning_rate": 1.9138291907384632e-05,
+      "loss": 0.7402,
+      "step": 3298
+    },
+    {
+      "epoch": 0.2027409046214356,
+      "grad_norm": 33.25,
+      "learning_rate": 1.913734662520712e-05,
+      "loss": 0.6951,
+      "step": 3299
+    },
+    {
+      "epoch": 0.2028023598820059,
+      "grad_norm": 18.5,
+      "learning_rate": 1.9136400848203837e-05,
+      "loss": 0.6361,
+      "step": 3300
+    },
+    {
+      "epoch": 0.2028638151425762,
+      "grad_norm": 42.25,
+      "learning_rate": 1.913545457642601e-05,
+      "loss": 0.5913,
+      "step": 3301
+    },
+    {
+      "epoch": 0.2029252704031465,
+      "grad_norm": 29.5,
+      "learning_rate": 1.9134507809924878e-05,
+      "loss": 0.6491,
+      "step": 3302
+    },
+    {
+      "epoch": 0.2029867256637168,
+      "grad_norm": 22.0,
+      "learning_rate": 1.913356054875171e-05,
+      "loss": 0.5775,
+      "step": 3303
+    },
+    {
+      "epoch": 0.2030481809242871,
+      "grad_norm": 24.875,
+      "learning_rate": 1.9132612792957808e-05,
+      "loss": 0.5075,
+      "step": 3304
+    },
+    {
+      "epoch": 0.20310963618485742,
+      "grad_norm": 45.75,
+      "learning_rate": 1.9131664542594496e-05,
+      "loss": 0.673,
+      "step": 3305
+    },
+    {
+      "epoch": 0.20317109144542772,
+      "grad_norm": 30.625,
+      "learning_rate": 1.9130715797713123e-05,
+      "loss": 0.6118,
+      "step": 3306
+    },
+    {
+      "epoch": 0.20323254670599802,
+      "grad_norm": 25.0,
+      "learning_rate": 1.9129766558365076e-05,
+      "loss": 0.6173,
+      "step": 3307
+    },
+    {
+      "epoch": 0.20329400196656833,
+      "grad_norm": 16.375,
+      "learning_rate": 1.912881682460175e-05,
+      "loss": 0.4625,
+      "step": 3308
+    },
+    {
+      "epoch": 0.20335545722713863,
+      "grad_norm": 10.25,
+      "learning_rate": 1.912786659647458e-05,
+      "loss": 0.4507,
+      "step": 3309
+    },
+    {
+      "epoch": 0.20341691248770893,
+      "grad_norm": 42.75,
+      "learning_rate": 1.912691587403503e-05,
+      "loss": 0.5495,
+      "step": 3310
+    },
+    {
+      "epoch": 0.20347836774827927,
+      "grad_norm": 17.75,
+      "learning_rate": 1.912596465733458e-05,
+      "loss": 0.4086,
+      "step": 3311
+    },
+    {
+      "epoch": 0.20353982300884957,
+      "grad_norm": 6.1875,
+      "learning_rate": 1.9125012946424742e-05,
+      "loss": 0.4421,
+      "step": 3312
+    },
+    {
+      "epoch": 0.20360127826941987,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9124060741357065e-05,
+      "loss": 0.4439,
+      "step": 3313
+    },
+    {
+      "epoch": 0.20366273352999018,
+      "grad_norm": 24.125,
+      "learning_rate": 1.91231080421831e-05,
+      "loss": 0.4542,
+      "step": 3314
+    },
+    {
+      "epoch": 0.20372418879056048,
+      "grad_norm": 17.0,
+      "learning_rate": 1.912215484895445e-05,
+      "loss": 0.5242,
+      "step": 3315
+    },
+    {
+      "epoch": 0.20378564405113078,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.9121201161722732e-05,
+      "loss": 0.4114,
+      "step": 3316
+    },
+    {
+      "epoch": 0.20384709931170109,
+      "grad_norm": 17.75,
+      "learning_rate": 1.912024698053959e-05,
+      "loss": 0.2947,
+      "step": 3317
+    },
+    {
+      "epoch": 0.2039085545722714,
+      "grad_norm": 40.0,
+      "learning_rate": 1.91192923054567e-05,
+      "loss": 0.5266,
+      "step": 3318
+    },
+    {
+      "epoch": 0.2039700098328417,
+      "grad_norm": 22.75,
+      "learning_rate": 1.911833713652576e-05,
+      "loss": 0.4076,
+      "step": 3319
+    },
+    {
+      "epoch": 0.204031465093412,
+      "grad_norm": 35.5,
+      "learning_rate": 1.9117381473798496e-05,
+      "loss": 0.4072,
+      "step": 3320
+    },
+    {
+      "epoch": 0.2040929203539823,
+      "grad_norm": 24.875,
+      "learning_rate": 1.911642531732666e-05,
+      "loss": 0.404,
+      "step": 3321
+    },
+    {
+      "epoch": 0.2041543756145526,
+      "grad_norm": 5.15625,
+      "learning_rate": 1.9115468667162038e-05,
+      "loss": 0.4201,
+      "step": 3322
+    },
+    {
+      "epoch": 0.2042158308751229,
+      "grad_norm": 17.875,
+      "learning_rate": 1.9114511523356428e-05,
+      "loss": 0.377,
+      "step": 3323
+    },
+    {
+      "epoch": 0.2042772861356932,
+      "grad_norm": 13.25,
+      "learning_rate": 1.9113553885961674e-05,
+      "loss": 0.4704,
+      "step": 3324
+    },
+    {
+      "epoch": 0.2043387413962635,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.9112595755029625e-05,
+      "loss": 0.4726,
+      "step": 3325
+    },
+    {
+      "epoch": 0.20440019665683382,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9111637130612172e-05,
+      "loss": 0.4006,
+      "step": 3326
+    },
+    {
+      "epoch": 0.20446165191740412,
+      "grad_norm": 3.53125,
+      "learning_rate": 1.911067801276123e-05,
+      "loss": 0.4242,
+      "step": 3327
+    },
+    {
+      "epoch": 0.20452310717797442,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.9109718401528742e-05,
+      "loss": 0.323,
+      "step": 3328
+    },
+    {
+      "epoch": 0.20458456243854473,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.9108758296966666e-05,
+      "loss": 0.435,
+      "step": 3329
+    },
+    {
+      "epoch": 0.20464601769911506,
+      "grad_norm": 11.875,
+      "learning_rate": 1.9107797699127004e-05,
+      "loss": 0.4746,
+      "step": 3330
+    },
+    {
+      "epoch": 0.20470747295968536,
+      "grad_norm": 17.0,
+      "learning_rate": 1.910683660806177e-05,
+      "loss": 0.4663,
+      "step": 3331
+    },
+    {
+      "epoch": 0.20476892822025566,
+      "grad_norm": 23.625,
+      "learning_rate": 1.9105875023823023e-05,
+      "loss": 0.3944,
+      "step": 3332
+    },
+    {
+      "epoch": 0.20483038348082597,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.910491294646282e-05,
+      "loss": 0.3831,
+      "step": 3333
+    },
+    {
+      "epoch": 0.20489183874139627,
+      "grad_norm": 31.875,
+      "learning_rate": 1.9103950376033276e-05,
+      "loss": 0.4428,
+      "step": 3334
+    },
+    {
+      "epoch": 0.20495329400196657,
+      "grad_norm": 19.125,
+      "learning_rate": 1.910298731258651e-05,
+      "loss": 0.3606,
+      "step": 3335
+    },
+    {
+      "epoch": 0.20501474926253688,
+      "grad_norm": 3.921875,
+      "learning_rate": 1.9102023756174675e-05,
+      "loss": 0.3203,
+      "step": 3336
+    },
+    {
+      "epoch": 0.20507620452310718,
+      "grad_norm": 21.5,
+      "learning_rate": 1.9101059706849957e-05,
+      "loss": 0.5632,
+      "step": 3337
+    },
+    {
+      "epoch": 0.20513765978367748,
+      "grad_norm": 21.375,
+      "learning_rate": 1.9100095164664564e-05,
+      "loss": 0.515,
+      "step": 3338
+    },
+    {
+      "epoch": 0.2051991150442478,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9099130129670726e-05,
+      "loss": 0.6015,
+      "step": 3339
+    },
+    {
+      "epoch": 0.2052605703048181,
+      "grad_norm": 9.375,
+      "learning_rate": 1.9098164601920702e-05,
+      "loss": 0.5169,
+      "step": 3340
+    },
+    {
+      "epoch": 0.2053220255653884,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9097198581466785e-05,
+      "loss": 0.4809,
+      "step": 3341
+    },
+    {
+      "epoch": 0.2053834808259587,
+      "grad_norm": 35.0,
+      "learning_rate": 1.9096232068361288e-05,
+      "loss": 0.4912,
+      "step": 3342
+    },
+    {
+      "epoch": 0.205444936086529,
+      "grad_norm": 24.5,
+      "learning_rate": 1.9095265062656546e-05,
+      "loss": 0.328,
+      "step": 3343
+    },
+    {
+      "epoch": 0.2055063913470993,
+      "grad_norm": 14.375,
+      "learning_rate": 1.9094297564404934e-05,
+      "loss": 0.323,
+      "step": 3344
+    },
+    {
+      "epoch": 0.2055678466076696,
+      "grad_norm": 40.5,
+      "learning_rate": 1.909332957365884e-05,
+      "loss": 0.4303,
+      "step": 3345
+    },
+    {
+      "epoch": 0.2056293018682399,
+      "grad_norm": 24.625,
+      "learning_rate": 1.9092361090470688e-05,
+      "loss": 0.5611,
+      "step": 3346
+    },
+    {
+      "epoch": 0.20569075712881021,
+      "grad_norm": 20.75,
+      "learning_rate": 1.9091392114892928e-05,
+      "loss": 0.3901,
+      "step": 3347
+    },
+    {
+      "epoch": 0.20575221238938052,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9090422646978025e-05,
+      "loss": 0.4598,
+      "step": 3348
+    },
+    {
+      "epoch": 0.20581366764995085,
+      "grad_norm": 20.875,
+      "learning_rate": 1.908945268677849e-05,
+      "loss": 0.4094,
+      "step": 3349
+    },
+    {
+      "epoch": 0.20587512291052115,
+      "grad_norm": 19.0,
+      "learning_rate": 1.9088482234346843e-05,
+      "loss": 0.4999,
+      "step": 3350
+    },
+    {
+      "epoch": 0.20593657817109146,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9087511289735646e-05,
+      "loss": 0.3852,
+      "step": 3351
+    },
+    {
+      "epoch": 0.20599803343166176,
+      "grad_norm": 24.25,
+      "learning_rate": 1.908653985299747e-05,
+      "loss": 0.3966,
+      "step": 3352
+    },
+    {
+      "epoch": 0.20605948869223206,
+      "grad_norm": 62.75,
+      "learning_rate": 1.9085567924184932e-05,
+      "loss": 0.5909,
+      "step": 3353
+    },
+    {
+      "epoch": 0.20612094395280237,
+      "grad_norm": 25.0,
+      "learning_rate": 1.9084595503350657e-05,
+      "loss": 0.7012,
+      "step": 3354
+    },
+    {
+      "epoch": 0.20618239921337267,
+      "grad_norm": 41.0,
+      "learning_rate": 1.9083622590547313e-05,
+      "loss": 0.8123,
+      "step": 3355
+    },
+    {
+      "epoch": 0.20624385447394297,
+      "grad_norm": 40.75,
+      "learning_rate": 1.9082649185827583e-05,
+      "loss": 0.7931,
+      "step": 3356
+    },
+    {
+      "epoch": 0.20630530973451328,
+      "grad_norm": 29.75,
+      "learning_rate": 1.9081675289244183e-05,
+      "loss": 0.7985,
+      "step": 3357
+    },
+    {
+      "epoch": 0.20636676499508358,
+      "grad_norm": 16.625,
+      "learning_rate": 1.9080700900849855e-05,
+      "loss": 0.4981,
+      "step": 3358
+    },
+    {
+      "epoch": 0.20642822025565388,
+      "grad_norm": 28.125,
+      "learning_rate": 1.907972602069736e-05,
+      "loss": 0.5605,
+      "step": 3359
+    },
+    {
+      "epoch": 0.20648967551622419,
+      "grad_norm": 19.125,
+      "learning_rate": 1.9078750648839497e-05,
+      "loss": 0.4373,
+      "step": 3360
+    },
+    {
+      "epoch": 0.2065511307767945,
+      "grad_norm": 20.0,
+      "learning_rate": 1.907777478532909e-05,
+      "loss": 0.357,
+      "step": 3361
+    },
+    {
+      "epoch": 0.2066125860373648,
+      "grad_norm": 13.875,
+      "learning_rate": 1.9076798430218975e-05,
+      "loss": 0.3645,
+      "step": 3362
+    },
+    {
+      "epoch": 0.2066740412979351,
+      "grad_norm": 34.75,
+      "learning_rate": 1.9075821583562036e-05,
+      "loss": 0.323,
+      "step": 3363
+    },
+    {
+      "epoch": 0.2067354965585054,
+      "grad_norm": 14.25,
+      "learning_rate": 1.907484424541117e-05,
+      "loss": 0.4351,
+      "step": 3364
+    },
+    {
+      "epoch": 0.2067969518190757,
+      "grad_norm": 28.25,
+      "learning_rate": 1.9073866415819303e-05,
+      "loss": 0.43,
+      "step": 3365
+    },
+    {
+      "epoch": 0.206858407079646,
+      "grad_norm": 21.625,
+      "learning_rate": 1.907288809483939e-05,
+      "loss": 0.4072,
+      "step": 3366
+    },
+    {
+      "epoch": 0.2069198623402163,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.907190928252441e-05,
+      "loss": 0.3202,
+      "step": 3367
+    },
+    {
+      "epoch": 0.20698131760078664,
+      "grad_norm": 11.5625,
+      "learning_rate": 1.907092997892737e-05,
+      "loss": 0.3836,
+      "step": 3368
+    },
+    {
+      "epoch": 0.20704277286135694,
+      "grad_norm": 18.25,
+      "learning_rate": 1.9069950184101305e-05,
+      "loss": 0.3967,
+      "step": 3369
+    },
+    {
+      "epoch": 0.20710422812192725,
+      "grad_norm": 29.125,
+      "learning_rate": 1.906896989809927e-05,
+      "loss": 0.5302,
+      "step": 3370
+    },
+    {
+      "epoch": 0.20716568338249755,
+      "grad_norm": 24.625,
+      "learning_rate": 1.906798912097436e-05,
+      "loss": 0.6367,
+      "step": 3371
+    },
+    {
+      "epoch": 0.20722713864306785,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.9067007852779684e-05,
+      "loss": 0.4452,
+      "step": 3372
+    },
+    {
+      "epoch": 0.20728859390363816,
+      "grad_norm": 31.375,
+      "learning_rate": 1.906602609356838e-05,
+      "loss": 0.4258,
+      "step": 3373
+    },
+    {
+      "epoch": 0.20735004916420846,
+      "grad_norm": 8.25,
+      "learning_rate": 1.9065043843393615e-05,
+      "loss": 0.3305,
+      "step": 3374
+    },
+    {
+      "epoch": 0.20741150442477876,
+      "grad_norm": 10.4375,
+      "learning_rate": 1.906406110230858e-05,
+      "loss": 0.3517,
+      "step": 3375
+    },
+    {
+      "epoch": 0.20747295968534907,
+      "grad_norm": 10.1875,
+      "learning_rate": 1.9063077870366504e-05,
+      "loss": 0.3475,
+      "step": 3376
+    },
+    {
+      "epoch": 0.20753441494591937,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.906209414762062e-05,
+      "loss": 0.7558,
+      "step": 3377
+    },
+    {
+      "epoch": 0.20759587020648967,
+      "grad_norm": 36.0,
+      "learning_rate": 1.9061109934124208e-05,
+      "loss": 0.6361,
+      "step": 3378
+    },
+    {
+      "epoch": 0.20765732546705998,
+      "grad_norm": 48.0,
+      "learning_rate": 1.9060125229930572e-05,
+      "loss": 0.5517,
+      "step": 3379
+    },
+    {
+      "epoch": 0.20771878072763028,
+      "grad_norm": 18.0,
+      "learning_rate": 1.9059140035093024e-05,
+      "loss": 0.6132,
+      "step": 3380
+    },
+    {
+      "epoch": 0.20778023598820058,
+      "grad_norm": 12.125,
+      "learning_rate": 1.9058154349664932e-05,
+      "loss": 0.5487,
+      "step": 3381
+    },
+    {
+      "epoch": 0.2078416912487709,
+      "grad_norm": 40.0,
+      "learning_rate": 1.9057168173699664e-05,
+      "loss": 0.4196,
+      "step": 3382
+    },
+    {
+      "epoch": 0.2079031465093412,
+      "grad_norm": 40.0,
+      "learning_rate": 1.905618150725063e-05,
+      "loss": 0.6362,
+      "step": 3383
+    },
+    {
+      "epoch": 0.2079646017699115,
+      "grad_norm": 38.75,
+      "learning_rate": 1.9055194350371262e-05,
+      "loss": 0.4734,
+      "step": 3384
+    },
+    {
+      "epoch": 0.2080260570304818,
+      "grad_norm": 16.625,
+      "learning_rate": 1.905420670311502e-05,
+      "loss": 0.5246,
+      "step": 3385
+    },
+    {
+      "epoch": 0.2080875122910521,
+      "grad_norm": 49.25,
+      "learning_rate": 1.9053218565535383e-05,
+      "loss": 0.5238,
+      "step": 3386
+    },
+    {
+      "epoch": 0.20814896755162243,
+      "grad_norm": 48.5,
+      "learning_rate": 1.9052229937685872e-05,
+      "loss": 0.4325,
+      "step": 3387
+    },
+    {
+      "epoch": 0.20821042281219274,
+      "grad_norm": 17.25,
+      "learning_rate": 1.9051240819620018e-05,
+      "loss": 0.3485,
+      "step": 3388
+    },
+    {
+      "epoch": 0.20827187807276304,
+      "grad_norm": 8.625,
+      "learning_rate": 1.905025121139139e-05,
+      "loss": 0.5452,
+      "step": 3389
+    },
+    {
+      "epoch": 0.20833333333333334,
+      "grad_norm": 23.875,
+      "learning_rate": 1.9049261113053576e-05,
+      "loss": 0.4885,
+      "step": 3390
+    },
+    {
+      "epoch": 0.20839478859390365,
+      "grad_norm": 2.234375,
+      "learning_rate": 1.9048270524660197e-05,
+      "loss": 0.3394,
+      "step": 3391
+    },
+    {
+      "epoch": 0.20845624385447395,
+      "grad_norm": 5.09375,
+      "learning_rate": 1.9047279446264898e-05,
+      "loss": 0.4145,
+      "step": 3392
+    },
+    {
+      "epoch": 0.20851769911504425,
+      "grad_norm": 36.5,
+      "learning_rate": 1.9046287877921344e-05,
+      "loss": 0.6575,
+      "step": 3393
+    },
+    {
+      "epoch": 0.20857915437561456,
+      "grad_norm": 27.375,
+      "learning_rate": 1.904529581968324e-05,
+      "loss": 0.5768,
+      "step": 3394
+    },
+    {
+      "epoch": 0.20864060963618486,
+      "grad_norm": 33.0,
+      "learning_rate": 1.904430327160431e-05,
+      "loss": 0.7122,
+      "step": 3395
+    },
+    {
+      "epoch": 0.20870206489675516,
+      "grad_norm": 16.875,
+      "learning_rate": 1.90433102337383e-05,
+      "loss": 0.7786,
+      "step": 3396
+    },
+    {
+      "epoch": 0.20876352015732547,
+      "grad_norm": 36.75,
+      "learning_rate": 1.9042316706138987e-05,
+      "loss": 0.586,
+      "step": 3397
+    },
+    {
+      "epoch": 0.20882497541789577,
+      "grad_norm": 33.25,
+      "learning_rate": 1.904132268886018e-05,
+      "loss": 0.5464,
+      "step": 3398
+    },
+    {
+      "epoch": 0.20888643067846607,
+      "grad_norm": 25.5,
+      "learning_rate": 1.9040328181955706e-05,
+      "loss": 0.5645,
+      "step": 3399
+    },
+    {
+      "epoch": 0.20894788593903638,
+      "grad_norm": 15.0,
+      "learning_rate": 1.903933318547942e-05,
+      "loss": 0.4591,
+      "step": 3400
+    },
+    {
+      "epoch": 0.20900934119960668,
+      "grad_norm": 21.75,
+      "learning_rate": 1.9038337699485207e-05,
+      "loss": 0.5846,
+      "step": 3401
+    },
+    {
+      "epoch": 0.20907079646017698,
+      "grad_norm": 22.875,
+      "learning_rate": 1.9037341724026975e-05,
+      "loss": 0.5218,
+      "step": 3402
+    },
+    {
+      "epoch": 0.2091322517207473,
+      "grad_norm": 12.875,
+      "learning_rate": 1.9036345259158667e-05,
+      "loss": 0.3924,
+      "step": 3403
+    },
+    {
+      "epoch": 0.2091937069813176,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9035348304934238e-05,
+      "loss": 0.5038,
+      "step": 3404
+    },
+    {
+      "epoch": 0.2092551622418879,
+      "grad_norm": 18.625,
+      "learning_rate": 1.903435086140768e-05,
+      "loss": 0.5567,
+      "step": 3405
+    },
+    {
+      "epoch": 0.20931661750245822,
+      "grad_norm": 30.625,
+      "learning_rate": 1.903335292863301e-05,
+      "loss": 0.531,
+      "step": 3406
+    },
+    {
+      "epoch": 0.20937807276302853,
+      "grad_norm": 24.625,
+      "learning_rate": 1.903235450666427e-05,
+      "loss": 0.4959,
+      "step": 3407
+    },
+    {
+      "epoch": 0.20943952802359883,
+      "grad_norm": 23.25,
+      "learning_rate": 1.9031355595555525e-05,
+      "loss": 0.6077,
+      "step": 3408
+    },
+    {
+      "epoch": 0.20950098328416913,
+      "grad_norm": 19.5,
+      "learning_rate": 1.9030356195360875e-05,
+      "loss": 0.706,
+      "step": 3409
+    },
+    {
+      "epoch": 0.20956243854473944,
+      "grad_norm": 35.75,
+      "learning_rate": 1.9029356306134435e-05,
+      "loss": 0.7165,
+      "step": 3410
+    },
+    {
+      "epoch": 0.20962389380530974,
+      "grad_norm": 16.125,
+      "learning_rate": 1.9028355927930363e-05,
+      "loss": 0.5179,
+      "step": 3411
+    },
+    {
+      "epoch": 0.20968534906588004,
+      "grad_norm": 16.625,
+      "learning_rate": 1.902735506080283e-05,
+      "loss": 0.5954,
+      "step": 3412
+    },
+    {
+      "epoch": 0.20974680432645035,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.9026353704806032e-05,
+      "loss": 0.3825,
+      "step": 3413
+    },
+    {
+      "epoch": 0.20980825958702065,
+      "grad_norm": 23.125,
+      "learning_rate": 1.9025351859994203e-05,
+      "loss": 0.442,
+      "step": 3414
+    },
+    {
+      "epoch": 0.20986971484759095,
+      "grad_norm": 17.75,
+      "learning_rate": 1.9024349526421596e-05,
+      "loss": 0.5192,
+      "step": 3415
+    },
+    {
+      "epoch": 0.20993117010816126,
+      "grad_norm": 18.875,
+      "learning_rate": 1.9023346704142488e-05,
+      "loss": 0.3977,
+      "step": 3416
+    },
+    {
+      "epoch": 0.20999262536873156,
+      "grad_norm": 37.25,
+      "learning_rate": 1.902234339321119e-05,
+      "loss": 0.3841,
+      "step": 3417
+    },
+    {
+      "epoch": 0.21005408062930186,
+      "grad_norm": 12.75,
+      "learning_rate": 1.902133959368203e-05,
+      "loss": 0.3731,
+      "step": 3418
+    },
+    {
+      "epoch": 0.21011553588987217,
+      "grad_norm": 9.5,
+      "learning_rate": 1.9020335305609373e-05,
+      "loss": 0.3999,
+      "step": 3419
+    },
+    {
+      "epoch": 0.21017699115044247,
+      "grad_norm": 8.875,
+      "learning_rate": 1.9019330529047605e-05,
+      "loss": 0.3533,
+      "step": 3420
+    },
+    {
+      "epoch": 0.21023844641101277,
+      "grad_norm": 10.5,
+      "learning_rate": 1.901832526405114e-05,
+      "loss": 0.3358,
+      "step": 3421
+    },
+    {
+      "epoch": 0.21029990167158308,
+      "grad_norm": 16.375,
+      "learning_rate": 1.9017319510674413e-05,
+      "loss": 0.3219,
+      "step": 3422
+    },
+    {
+      "epoch": 0.21036135693215338,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.9016313268971892e-05,
+      "loss": 0.3167,
+      "step": 3423
+    },
+    {
+      "epoch": 0.21042281219272368,
+      "grad_norm": 17.5,
+      "learning_rate": 1.901530653899807e-05,
+      "loss": 0.2909,
+      "step": 3424
+    },
+    {
+      "epoch": 0.210484267453294,
+      "grad_norm": 17.125,
+      "learning_rate": 1.9014299320807467e-05,
+      "loss": 0.4499,
+      "step": 3425
+    },
+    {
+      "epoch": 0.21054572271386432,
+      "grad_norm": 29.375,
+      "learning_rate": 1.9013291614454622e-05,
+      "loss": 0.6185,
+      "step": 3426
+    },
+    {
+      "epoch": 0.21060717797443462,
+      "grad_norm": 22.875,
+      "learning_rate": 1.9012283419994115e-05,
+      "loss": 0.4255,
+      "step": 3427
+    },
+    {
+      "epoch": 0.21066863323500493,
+      "grad_norm": 21.125,
+      "learning_rate": 1.9011274737480536e-05,
+      "loss": 0.404,
+      "step": 3428
+    },
+    {
+      "epoch": 0.21073008849557523,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.9010265566968518e-05,
+      "loss": 0.4227,
+      "step": 3429
+    },
+    {
+      "epoch": 0.21079154375614553,
+      "grad_norm": 9.75,
+      "learning_rate": 1.9009255908512704e-05,
+      "loss": 0.2796,
+      "step": 3430
+    },
+    {
+      "epoch": 0.21085299901671584,
+      "grad_norm": 13.625,
+      "learning_rate": 1.9008245762167773e-05,
+      "loss": 0.2448,
+      "step": 3431
+    },
+    {
+      "epoch": 0.21091445427728614,
+      "grad_norm": 7.75,
+      "learning_rate": 1.900723512798843e-05,
+      "loss": 0.369,
+      "step": 3432
+    },
+    {
+      "epoch": 0.21097590953785644,
+      "grad_norm": 19.375,
+      "learning_rate": 1.9006224006029404e-05,
+      "loss": 0.4336,
+      "step": 3433
+    },
+    {
+      "epoch": 0.21103736479842675,
+      "grad_norm": 31.875,
+      "learning_rate": 1.9005212396345458e-05,
+      "loss": 0.3779,
+      "step": 3434
+    },
+    {
+      "epoch": 0.21109882005899705,
+      "grad_norm": 17.375,
+      "learning_rate": 1.900420029899136e-05,
+      "loss": 0.4318,
+      "step": 3435
+    },
+    {
+      "epoch": 0.21116027531956735,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.9003187714021936e-05,
+      "loss": 0.4246,
+      "step": 3436
+    },
+    {
+      "epoch": 0.21122173058013766,
+      "grad_norm": 49.5,
+      "learning_rate": 1.9002174641492015e-05,
+      "loss": 0.446,
+      "step": 3437
+    },
+    {
+      "epoch": 0.21128318584070796,
+      "grad_norm": 38.5,
+      "learning_rate": 1.9001161081456455e-05,
+      "loss": 0.3657,
+      "step": 3438
+    },
+    {
+      "epoch": 0.21134464110127826,
+      "grad_norm": 60.75,
+      "learning_rate": 1.9000147033970148e-05,
+      "loss": 0.3347,
+      "step": 3439
+    },
+    {
+      "epoch": 0.21140609636184857,
+      "grad_norm": 35.5,
+      "learning_rate": 1.899913249908801e-05,
+      "loss": 0.3616,
+      "step": 3440
+    },
+    {
+      "epoch": 0.21146755162241887,
+      "grad_norm": 18.125,
+      "learning_rate": 1.8998117476864984e-05,
+      "loss": 0.5344,
+      "step": 3441
+    },
+    {
+      "epoch": 0.21152900688298917,
+      "grad_norm": 18.75,
+      "learning_rate": 1.899710196735603e-05,
+      "loss": 0.3985,
+      "step": 3442
+    },
+    {
+      "epoch": 0.21159046214355948,
+      "grad_norm": 36.5,
+      "learning_rate": 1.899608597061615e-05,
+      "loss": 0.4223,
+      "step": 3443
+    },
+    {
+      "epoch": 0.21165191740412978,
+      "grad_norm": 64.0,
+      "learning_rate": 1.899506948670036e-05,
+      "loss": 0.5576,
+      "step": 3444
+    },
+    {
+      "epoch": 0.2117133726647001,
+      "grad_norm": 23.375,
+      "learning_rate": 1.899405251566371e-05,
+      "loss": 0.5121,
+      "step": 3445
+    },
+    {
+      "epoch": 0.21177482792527041,
+      "grad_norm": 27.375,
+      "learning_rate": 1.8993035057561274e-05,
+      "loss": 0.4471,
+      "step": 3446
+    },
+    {
+      "epoch": 0.21183628318584072,
+      "grad_norm": 17.375,
+      "learning_rate": 1.899201711244815e-05,
+      "loss": 0.5354,
+      "step": 3447
+    },
+    {
+      "epoch": 0.21189773844641102,
+      "grad_norm": 27.375,
+      "learning_rate": 1.8990998680379458e-05,
+      "loss": 0.4154,
+      "step": 3448
+    },
+    {
+      "epoch": 0.21195919370698132,
+      "grad_norm": 65.5,
+      "learning_rate": 1.8989979761410356e-05,
+      "loss": 0.8404,
+      "step": 3449
+    },
+    {
+      "epoch": 0.21202064896755163,
+      "grad_norm": 50.0,
+      "learning_rate": 1.8988960355596027e-05,
+      "loss": 0.7976,
+      "step": 3450
+    },
+    {
+      "epoch": 0.21208210422812193,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8987940462991673e-05,
+      "loss": 0.6242,
+      "step": 3451
+    },
+    {
+      "epoch": 0.21214355948869223,
+      "grad_norm": 52.0,
+      "learning_rate": 1.898692008365252e-05,
+      "loss": 0.436,
+      "step": 3452
+    },
+    {
+      "epoch": 0.21220501474926254,
+      "grad_norm": 27.125,
+      "learning_rate": 1.898589921763383e-05,
+      "loss": 0.502,
+      "step": 3453
+    },
+    {
+      "epoch": 0.21226647000983284,
+      "grad_norm": 31.375,
+      "learning_rate": 1.8984877864990888e-05,
+      "loss": 0.5196,
+      "step": 3454
+    },
+    {
+      "epoch": 0.21232792527040315,
+      "grad_norm": 17.375,
+      "learning_rate": 1.8983856025779007e-05,
+      "loss": 0.4961,
+      "step": 3455
+    },
+    {
+      "epoch": 0.21238938053097345,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.8982833700053518e-05,
+      "loss": 0.3489,
+      "step": 3456
+    },
+    {
+      "epoch": 0.21245083579154375,
+      "grad_norm": 30.875,
+      "learning_rate": 1.8981810887869784e-05,
+      "loss": 0.2947,
+      "step": 3457
+    },
+    {
+      "epoch": 0.21251229105211406,
+      "grad_norm": 4.96875,
+      "learning_rate": 1.8980787589283204e-05,
+      "loss": 0.3005,
+      "step": 3458
+    },
+    {
+      "epoch": 0.21257374631268436,
+      "grad_norm": 23.875,
+      "learning_rate": 1.897976380434918e-05,
+      "loss": 0.3242,
+      "step": 3459
+    },
+    {
+      "epoch": 0.21263520157325466,
+      "grad_norm": 6.6875,
+      "learning_rate": 1.897873953312317e-05,
+      "loss": 0.2953,
+      "step": 3460
+    },
+    {
+      "epoch": 0.21269665683382497,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.897771477566063e-05,
+      "loss": 0.3113,
+      "step": 3461
+    },
+    {
+      "epoch": 0.21275811209439527,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.8976689532017055e-05,
+      "loss": 0.3601,
+      "step": 3462
+    },
+    {
+      "epoch": 0.21281956735496557,
+      "grad_norm": 41.0,
+      "learning_rate": 1.8975663802247978e-05,
+      "loss": 0.2737,
+      "step": 3463
+    },
+    {
+      "epoch": 0.2128810226155359,
+      "grad_norm": 9.625,
+      "learning_rate": 1.8974637586408933e-05,
+      "loss": 0.3694,
+      "step": 3464
+    },
+    {
+      "epoch": 0.2129424778761062,
+      "grad_norm": 43.5,
+      "learning_rate": 1.8973610884555505e-05,
+      "loss": 0.5624,
+      "step": 3465
+    },
+    {
+      "epoch": 0.2130039331366765,
+      "grad_norm": 45.75,
+      "learning_rate": 1.8972583696743284e-05,
+      "loss": 0.7223,
+      "step": 3466
+    },
+    {
+      "epoch": 0.2130653883972468,
+      "grad_norm": 48.0,
+      "learning_rate": 1.8971556023027905e-05,
+      "loss": 0.5296,
+      "step": 3467
+    },
+    {
+      "epoch": 0.21312684365781712,
+      "grad_norm": 35.5,
+      "learning_rate": 1.897052786346502e-05,
+      "loss": 0.5389,
+      "step": 3468
+    },
+    {
+      "epoch": 0.21318829891838742,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.8969499218110302e-05,
+      "loss": 0.579,
+      "step": 3469
+    },
+    {
+      "epoch": 0.21324975417895772,
+      "grad_norm": 56.75,
+      "learning_rate": 1.896847008701946e-05,
+      "loss": 0.474,
+      "step": 3470
+    },
+    {
+      "epoch": 0.21331120943952803,
+      "grad_norm": 21.75,
+      "learning_rate": 1.8967440470248227e-05,
+      "loss": 0.3857,
+      "step": 3471
+    },
+    {
+      "epoch": 0.21337266470009833,
+      "grad_norm": 21.0,
+      "learning_rate": 1.896641036785236e-05,
+      "loss": 0.3608,
+      "step": 3472
+    },
+    {
+      "epoch": 0.21343411996066863,
+      "grad_norm": 16.125,
+      "learning_rate": 1.896537977988764e-05,
+      "loss": 0.4179,
+      "step": 3473
+    },
+    {
+      "epoch": 0.21349557522123894,
+      "grad_norm": 26.375,
+      "learning_rate": 1.896434870640989e-05,
+      "loss": 0.4416,
+      "step": 3474
+    },
+    {
+      "epoch": 0.21355703048180924,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.896331714747493e-05,
+      "loss": 0.3183,
+      "step": 3475
+    },
+    {
+      "epoch": 0.21361848574237954,
+      "grad_norm": 11.125,
+      "learning_rate": 1.8962285103138637e-05,
+      "loss": 0.3087,
+      "step": 3476
+    },
+    {
+      "epoch": 0.21367994100294985,
+      "grad_norm": 13.125,
+      "learning_rate": 1.896125257345689e-05,
+      "loss": 0.3474,
+      "step": 3477
+    },
+    {
+      "epoch": 0.21374139626352015,
+      "grad_norm": 6.65625,
+      "learning_rate": 1.896021955848561e-05,
+      "loss": 0.3544,
+      "step": 3478
+    },
+    {
+      "epoch": 0.21380285152409045,
+      "grad_norm": 12.125,
+      "learning_rate": 1.8959186058280744e-05,
+      "loss": 0.3563,
+      "step": 3479
+    },
+    {
+      "epoch": 0.21386430678466076,
+      "grad_norm": 22.75,
+      "learning_rate": 1.895815207289825e-05,
+      "loss": 0.3235,
+      "step": 3480
+    },
+    {
+      "epoch": 0.21392576204523106,
+      "grad_norm": 33.5,
+      "learning_rate": 1.895711760239413e-05,
+      "loss": 0.5054,
+      "step": 3481
+    },
+    {
+      "epoch": 0.21398721730580136,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.89560826468244e-05,
+      "loss": 0.4144,
+      "step": 3482
+    },
+    {
+      "epoch": 0.2140486725663717,
+      "grad_norm": 20.75,
+      "learning_rate": 1.8955047206245115e-05,
+      "loss": 0.5112,
+      "step": 3483
+    },
+    {
+      "epoch": 0.214110127826942,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.895401128071234e-05,
+      "loss": 0.3733,
+      "step": 3484
+    },
+    {
+      "epoch": 0.2141715830875123,
+      "grad_norm": 39.0,
+      "learning_rate": 1.895297487028218e-05,
+      "loss": 0.3263,
+      "step": 3485
+    },
+    {
+      "epoch": 0.2142330383480826,
+      "grad_norm": 28.5,
+      "learning_rate": 1.895193797501076e-05,
+      "loss": 0.5216,
+      "step": 3486
+    },
+    {
+      "epoch": 0.2142944936086529,
+      "grad_norm": 12.5,
+      "learning_rate": 1.8950900594954226e-05,
+      "loss": 0.3861,
+      "step": 3487
+    },
+    {
+      "epoch": 0.2143559488692232,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.8949862730168766e-05,
+      "loss": 0.3608,
+      "step": 3488
+    },
+    {
+      "epoch": 0.21441740412979352,
+      "grad_norm": 22.5,
+      "learning_rate": 1.8948824380710582e-05,
+      "loss": 0.3415,
+      "step": 3489
+    },
+    {
+      "epoch": 0.21447885939036382,
+      "grad_norm": 44.75,
+      "learning_rate": 1.8947785546635905e-05,
+      "loss": 0.4517,
+      "step": 3490
+    },
+    {
+      "epoch": 0.21454031465093412,
+      "grad_norm": 35.25,
+      "learning_rate": 1.8946746228000987e-05,
+      "loss": 0.3534,
+      "step": 3491
+    },
+    {
+      "epoch": 0.21460176991150443,
+      "grad_norm": 16.125,
+      "learning_rate": 1.8945706424862115e-05,
+      "loss": 0.485,
+      "step": 3492
+    },
+    {
+      "epoch": 0.21466322517207473,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.89446661372756e-05,
+      "loss": 0.365,
+      "step": 3493
+    },
+    {
+      "epoch": 0.21472468043264503,
+      "grad_norm": 21.0,
+      "learning_rate": 1.8943625365297776e-05,
+      "loss": 0.3745,
+      "step": 3494
+    },
+    {
+      "epoch": 0.21478613569321534,
+      "grad_norm": 28.625,
+      "learning_rate": 1.894258410898501e-05,
+      "loss": 0.3674,
+      "step": 3495
+    },
+    {
+      "epoch": 0.21484759095378564,
+      "grad_norm": 34.0,
+      "learning_rate": 1.8941542368393683e-05,
+      "loss": 0.3151,
+      "step": 3496
+    },
+    {
+      "epoch": 0.21490904621435594,
+      "grad_norm": 44.0,
+      "learning_rate": 1.8940500143580215e-05,
+      "loss": 0.6101,
+      "step": 3497
+    },
+    {
+      "epoch": 0.21497050147492625,
+      "grad_norm": 55.25,
+      "learning_rate": 1.8939457434601044e-05,
+      "loss": 0.5991,
+      "step": 3498
+    },
+    {
+      "epoch": 0.21503195673549655,
+      "grad_norm": 21.625,
+      "learning_rate": 1.893841424151264e-05,
+      "loss": 0.5392,
+      "step": 3499
+    },
+    {
+      "epoch": 0.21509341199606685,
+      "grad_norm": 38.0,
+      "learning_rate": 1.893737056437149e-05,
+      "loss": 0.4825,
+      "step": 3500
+    },
+    {
+      "epoch": 0.21515486725663716,
+      "grad_norm": 34.0,
+      "learning_rate": 1.8936326403234125e-05,
+      "loss": 0.4443,
+      "step": 3501
+    },
+    {
+      "epoch": 0.2152163225172075,
+      "grad_norm": 23.125,
+      "learning_rate": 1.893528175815708e-05,
+      "loss": 0.3608,
+      "step": 3502
+    },
+    {
+      "epoch": 0.2152777777777778,
+      "grad_norm": 38.75,
+      "learning_rate": 1.8934236629196928e-05,
+      "loss": 0.337,
+      "step": 3503
+    },
+    {
+      "epoch": 0.2153392330383481,
+      "grad_norm": 14.375,
+      "learning_rate": 1.8933191016410273e-05,
+      "loss": 0.283,
+      "step": 3504
+    },
+    {
+      "epoch": 0.2154006882989184,
+      "grad_norm": 16.625,
+      "learning_rate": 1.893214491985374e-05,
+      "loss": 0.524,
+      "step": 3505
+    },
+    {
+      "epoch": 0.2154621435594887,
+      "grad_norm": 24.75,
+      "learning_rate": 1.893109833958397e-05,
+      "loss": 0.4749,
+      "step": 3506
+    },
+    {
+      "epoch": 0.215523598820059,
+      "grad_norm": 33.75,
+      "learning_rate": 1.8930051275657643e-05,
+      "loss": 0.6193,
+      "step": 3507
+    },
+    {
+      "epoch": 0.2155850540806293,
+      "grad_norm": 19.625,
+      "learning_rate": 1.892900372813147e-05,
+      "loss": 0.3183,
+      "step": 3508
+    },
+    {
+      "epoch": 0.2156465093411996,
+      "grad_norm": 45.25,
+      "learning_rate": 1.8927955697062172e-05,
+      "loss": 0.676,
+      "step": 3509
+    },
+    {
+      "epoch": 0.2157079646017699,
+      "grad_norm": 16.5,
+      "learning_rate": 1.8926907182506505e-05,
+      "loss": 0.5075,
+      "step": 3510
+    },
+    {
+      "epoch": 0.21576941986234022,
+      "grad_norm": 8.375,
+      "learning_rate": 1.892585818452126e-05,
+      "loss": 0.3383,
+      "step": 3511
+    },
+    {
+      "epoch": 0.21583087512291052,
+      "grad_norm": 6.4375,
+      "learning_rate": 1.8924808703163226e-05,
+      "loss": 0.4292,
+      "step": 3512
+    },
+    {
+      "epoch": 0.21589233038348082,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.892375873848925e-05,
+      "loss": 0.506,
+      "step": 3513
+    },
+    {
+      "epoch": 0.21595378564405113,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.8922708290556197e-05,
+      "loss": 0.4999,
+      "step": 3514
+    },
+    {
+      "epoch": 0.21601524090462143,
+      "grad_norm": 16.25,
+      "learning_rate": 1.892165735942094e-05,
+      "loss": 0.5492,
+      "step": 3515
+    },
+    {
+      "epoch": 0.21607669616519173,
+      "grad_norm": 27.375,
+      "learning_rate": 1.8920605945140396e-05,
+      "loss": 0.369,
+      "step": 3516
+    },
+    {
+      "epoch": 0.21613815142576204,
+      "grad_norm": 13.375,
+      "learning_rate": 1.8919554047771508e-05,
+      "loss": 0.5076,
+      "step": 3517
+    },
+    {
+      "epoch": 0.21619960668633234,
+      "grad_norm": 51.0,
+      "learning_rate": 1.891850166737124e-05,
+      "loss": 0.4182,
+      "step": 3518
+    },
+    {
+      "epoch": 0.21626106194690264,
+      "grad_norm": 40.0,
+      "learning_rate": 1.8917448803996573e-05,
+      "loss": 0.42,
+      "step": 3519
+    },
+    {
+      "epoch": 0.21632251720747295,
+      "grad_norm": 8.8125,
+      "learning_rate": 1.8916395457704536e-05,
+      "loss": 0.3401,
+      "step": 3520
+    },
+    {
+      "epoch": 0.21638397246804325,
+      "grad_norm": 34.25,
+      "learning_rate": 1.8915341628552166e-05,
+      "loss": 0.4286,
+      "step": 3521
+    },
+    {
+      "epoch": 0.21644542772861358,
+      "grad_norm": 4.875,
+      "learning_rate": 1.8914287316596532e-05,
+      "loss": 0.2178,
+      "step": 3522
+    },
+    {
+      "epoch": 0.21650688298918389,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.8913232521894734e-05,
+      "loss": 0.3303,
+      "step": 3523
+    },
+    {
+      "epoch": 0.2165683382497542,
+      "grad_norm": 7.5625,
+      "learning_rate": 1.891217724450389e-05,
+      "loss": 0.3541,
+      "step": 3524
+    },
+    {
+      "epoch": 0.2166297935103245,
+      "grad_norm": 23.375,
+      "learning_rate": 1.8911121484481145e-05,
+      "loss": 0.4064,
+      "step": 3525
+    },
+    {
+      "epoch": 0.2166912487708948,
+      "grad_norm": 39.25,
+      "learning_rate": 1.891006524188368e-05,
+      "loss": 0.4571,
+      "step": 3526
+    },
+    {
+      "epoch": 0.2167527040314651,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.890900851676869e-05,
+      "loss": 0.2396,
+      "step": 3527
+    },
+    {
+      "epoch": 0.2168141592920354,
+      "grad_norm": 5.125,
+      "learning_rate": 1.89079513091934e-05,
+      "loss": 0.2493,
+      "step": 3528
+    },
+    {
+      "epoch": 0.2168756145526057,
+      "grad_norm": 37.75,
+      "learning_rate": 1.890689361921507e-05,
+      "loss": 0.5346,
+      "step": 3529
+    },
+    {
+      "epoch": 0.216937069813176,
+      "grad_norm": 38.25,
+      "learning_rate": 1.8905835446890968e-05,
+      "loss": 0.6114,
+      "step": 3530
+    },
+    {
+      "epoch": 0.2169985250737463,
+      "grad_norm": 31.625,
+      "learning_rate": 1.8904776792278403e-05,
+      "loss": 0.5315,
+      "step": 3531
+    },
+    {
+      "epoch": 0.21705998033431662,
+      "grad_norm": 32.0,
+      "learning_rate": 1.8903717655434708e-05,
+      "loss": 0.456,
+      "step": 3532
+    },
+    {
+      "epoch": 0.21712143559488692,
+      "grad_norm": 22.5,
+      "learning_rate": 1.8902658036417236e-05,
+      "loss": 0.5363,
+      "step": 3533
+    },
+    {
+      "epoch": 0.21718289085545722,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8901597935283373e-05,
+      "loss": 0.4296,
+      "step": 3534
+    },
+    {
+      "epoch": 0.21724434611602753,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.8900537352090523e-05,
+      "loss": 0.3458,
+      "step": 3535
+    },
+    {
+      "epoch": 0.21730580137659783,
+      "grad_norm": 3.71875,
+      "learning_rate": 1.889947628689613e-05,
+      "loss": 0.3885,
+      "step": 3536
+    },
+    {
+      "epoch": 0.21736725663716813,
+      "grad_norm": 53.75,
+      "learning_rate": 1.8898414739757648e-05,
+      "loss": 0.5621,
+      "step": 3537
+    },
+    {
+      "epoch": 0.21742871189773844,
+      "grad_norm": 13.125,
+      "learning_rate": 1.8897352710732564e-05,
+      "loss": 0.4764,
+      "step": 3538
+    },
+    {
+      "epoch": 0.21749016715830874,
+      "grad_norm": 36.75,
+      "learning_rate": 1.8896290199878393e-05,
+      "loss": 0.4968,
+      "step": 3539
+    },
+    {
+      "epoch": 0.21755162241887904,
+      "grad_norm": 30.0,
+      "learning_rate": 1.8895227207252674e-05,
+      "loss": 0.3151,
+      "step": 3540
+    },
+    {
+      "epoch": 0.21761307767944937,
+      "grad_norm": 20.625,
+      "learning_rate": 1.889416373291298e-05,
+      "loss": 0.3569,
+      "step": 3541
+    },
+    {
+      "epoch": 0.21767453294001968,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.889309977691689e-05,
+      "loss": 0.4138,
+      "step": 3542
+    },
+    {
+      "epoch": 0.21773598820058998,
+      "grad_norm": 6.96875,
+      "learning_rate": 1.8892035339322023e-05,
+      "loss": 0.4461,
+      "step": 3543
+    },
+    {
+      "epoch": 0.21779744346116028,
+      "grad_norm": 28.25,
+      "learning_rate": 1.8890970420186035e-05,
+      "loss": 0.3911,
+      "step": 3544
+    },
+    {
+      "epoch": 0.2178588987217306,
+      "grad_norm": 29.625,
+      "learning_rate": 1.8889905019566584e-05,
+      "loss": 0.5838,
+      "step": 3545
+    },
+    {
+      "epoch": 0.2179203539823009,
+      "grad_norm": 17.5,
+      "learning_rate": 1.888883913752137e-05,
+      "loss": 0.5225,
+      "step": 3546
+    },
+    {
+      "epoch": 0.2179818092428712,
+      "grad_norm": 17.125,
+      "learning_rate": 1.8887772774108116e-05,
+      "loss": 0.3765,
+      "step": 3547
+    },
+    {
+      "epoch": 0.2180432645034415,
+      "grad_norm": 8.1875,
+      "learning_rate": 1.888670592938457e-05,
+      "loss": 0.3852,
+      "step": 3548
+    },
+    {
+      "epoch": 0.2181047197640118,
+      "grad_norm": 29.125,
+      "learning_rate": 1.8885638603408502e-05,
+      "loss": 0.4373,
+      "step": 3549
+    },
+    {
+      "epoch": 0.2181661750245821,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.888457079623772e-05,
+      "loss": 0.4385,
+      "step": 3550
+    },
+    {
+      "epoch": 0.2182276302851524,
+      "grad_norm": 6.78125,
+      "learning_rate": 1.8883502507930044e-05,
+      "loss": 0.3072,
+      "step": 3551
+    },
+    {
+      "epoch": 0.2182890855457227,
+      "grad_norm": 15.625,
+      "learning_rate": 1.8882433738543325e-05,
+      "loss": 0.4408,
+      "step": 3552
+    },
+    {
+      "epoch": 0.21835054080629301,
+      "grad_norm": 17.375,
+      "learning_rate": 1.8881364488135448e-05,
+      "loss": 0.4208,
+      "step": 3553
+    },
+    {
+      "epoch": 0.21841199606686332,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.888029475676431e-05,
+      "loss": 0.5357,
+      "step": 3554
+    },
+    {
+      "epoch": 0.21847345132743362,
+      "grad_norm": 12.0,
+      "learning_rate": 1.887922454448785e-05,
+      "loss": 0.3352,
+      "step": 3555
+    },
+    {
+      "epoch": 0.21853490658800392,
+      "grad_norm": 5.4375,
+      "learning_rate": 1.8878153851364013e-05,
+      "loss": 0.2896,
+      "step": 3556
+    },
+    {
+      "epoch": 0.21859636184857423,
+      "grad_norm": 13.5,
+      "learning_rate": 1.8877082677450795e-05,
+      "loss": 0.3957,
+      "step": 3557
+    },
+    {
+      "epoch": 0.21865781710914453,
+      "grad_norm": 8.25,
+      "learning_rate": 1.8876011022806193e-05,
+      "loss": 0.3333,
+      "step": 3558
+    },
+    {
+      "epoch": 0.21871927236971483,
+      "grad_norm": 5.875,
+      "learning_rate": 1.887493888748825e-05,
+      "loss": 0.2617,
+      "step": 3559
+    },
+    {
+      "epoch": 0.21878072763028517,
+      "grad_norm": 10.25,
+      "learning_rate": 1.8873866271555022e-05,
+      "loss": 0.2539,
+      "step": 3560
+    },
+    {
+      "epoch": 0.21884218289085547,
+      "grad_norm": 18.875,
+      "learning_rate": 1.8872793175064594e-05,
+      "loss": 0.3132,
+      "step": 3561
+    },
+    {
+      "epoch": 0.21890363815142577,
+      "grad_norm": 78.0,
+      "learning_rate": 1.8871719598075083e-05,
+      "loss": 0.4044,
+      "step": 3562
+    },
+    {
+      "epoch": 0.21896509341199608,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.8870645540644625e-05,
+      "loss": 0.3298,
+      "step": 3563
+    },
+    {
+      "epoch": 0.21902654867256638,
+      "grad_norm": 32.75,
+      "learning_rate": 1.8869571002831388e-05,
+      "loss": 0.5178,
+      "step": 3564
+    },
+    {
+      "epoch": 0.21908800393313668,
+      "grad_norm": 56.0,
+      "learning_rate": 1.886849598469356e-05,
+      "loss": 0.3818,
+      "step": 3565
+    },
+    {
+      "epoch": 0.21914945919370699,
+      "grad_norm": 27.0,
+      "learning_rate": 1.886742048628936e-05,
+      "loss": 0.3706,
+      "step": 3566
+    },
+    {
+      "epoch": 0.2192109144542773,
+      "grad_norm": 73.0,
+      "learning_rate": 1.8866344507677023e-05,
+      "loss": 0.3682,
+      "step": 3567
+    },
+    {
+      "epoch": 0.2192723697148476,
+      "grad_norm": 49.25,
+      "learning_rate": 1.8865268048914828e-05,
+      "loss": 0.4882,
+      "step": 3568
+    },
+    {
+      "epoch": 0.2193338249754179,
+      "grad_norm": 59.5,
+      "learning_rate": 1.8864191110061066e-05,
+      "loss": 0.4289,
+      "step": 3569
+    },
+    {
+      "epoch": 0.2193952802359882,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.8863113691174058e-05,
+      "loss": 0.3772,
+      "step": 3570
+    },
+    {
+      "epoch": 0.2194567354965585,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.8862035792312148e-05,
+      "loss": 0.3209,
+      "step": 3571
+    },
+    {
+      "epoch": 0.2195181907571288,
+      "grad_norm": 9.5,
+      "learning_rate": 1.8860957413533713e-05,
+      "loss": 0.4487,
+      "step": 3572
+    },
+    {
+      "epoch": 0.2195796460176991,
+      "grad_norm": 10.5625,
+      "learning_rate": 1.8859878554897147e-05,
+      "loss": 0.2822,
+      "step": 3573
+    },
+    {
+      "epoch": 0.2196411012782694,
+      "grad_norm": 5.28125,
+      "learning_rate": 1.8858799216460883e-05,
+      "loss": 0.3355,
+      "step": 3574
+    },
+    {
+      "epoch": 0.21970255653883972,
+      "grad_norm": 17.375,
+      "learning_rate": 1.8857719398283358e-05,
+      "loss": 0.3819,
+      "step": 3575
+    },
+    {
+      "epoch": 0.21976401179941002,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.8856639100423062e-05,
+      "loss": 0.2435,
+      "step": 3576
+    },
+    {
+      "epoch": 0.21982546705998032,
+      "grad_norm": 9.625,
+      "learning_rate": 1.8855558322938492e-05,
+      "loss": 0.2443,
+      "step": 3577
+    },
+    {
+      "epoch": 0.21988692232055063,
+      "grad_norm": 34.25,
+      "learning_rate": 1.8854477065888176e-05,
+      "loss": 0.6694,
+      "step": 3578
+    },
+    {
+      "epoch": 0.21994837758112096,
+      "grad_norm": 17.0,
+      "learning_rate": 1.8853395329330667e-05,
+      "loss": 0.3439,
+      "step": 3579
+    },
+    {
+      "epoch": 0.22000983284169126,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.8852313113324553e-05,
+      "loss": 0.3224,
+      "step": 3580
+    },
+    {
+      "epoch": 0.22007128810226156,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.8851230417928433e-05,
+      "loss": 0.4,
+      "step": 3581
+    },
+    {
+      "epoch": 0.22013274336283187,
+      "grad_norm": 11.875,
+      "learning_rate": 1.8850147243200946e-05,
+      "loss": 0.4037,
+      "step": 3582
+    },
+    {
+      "epoch": 0.22019419862340217,
+      "grad_norm": 68.5,
+      "learning_rate": 1.8849063589200744e-05,
+      "loss": 0.4553,
+      "step": 3583
+    },
+    {
+      "epoch": 0.22025565388397247,
+      "grad_norm": 54.5,
+      "learning_rate": 1.8847979455986515e-05,
+      "loss": 0.5413,
+      "step": 3584
+    },
+    {
+      "epoch": 0.22031710914454278,
+      "grad_norm": 42.25,
+      "learning_rate": 1.884689484361697e-05,
+      "loss": 0.5473,
+      "step": 3585
+    },
+    {
+      "epoch": 0.22037856440511308,
+      "grad_norm": 17.25,
+      "learning_rate": 1.884580975215084e-05,
+      "loss": 0.4447,
+      "step": 3586
+    },
+    {
+      "epoch": 0.22044001966568338,
+      "grad_norm": 32.25,
+      "learning_rate": 1.8844724181646895e-05,
+      "loss": 0.5432,
+      "step": 3587
+    },
+    {
+      "epoch": 0.2205014749262537,
+      "grad_norm": 16.25,
+      "learning_rate": 1.8843638132163917e-05,
+      "loss": 0.3536,
+      "step": 3588
+    },
+    {
+      "epoch": 0.220562930186824,
+      "grad_norm": 15.625,
+      "learning_rate": 1.8842551603760725e-05,
+      "loss": 0.4254,
+      "step": 3589
+    },
+    {
+      "epoch": 0.2206243854473943,
+      "grad_norm": 42.5,
+      "learning_rate": 1.8841464596496154e-05,
+      "loss": 0.4676,
+      "step": 3590
+    },
+    {
+      "epoch": 0.2206858407079646,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.8840377110429075e-05,
+      "loss": 0.3577,
+      "step": 3591
+    },
+    {
+      "epoch": 0.2207472959685349,
+      "grad_norm": 84.5,
+      "learning_rate": 1.8839289145618378e-05,
+      "loss": 0.3929,
+      "step": 3592
+    },
+    {
+      "epoch": 0.2208087512291052,
+      "grad_norm": 23.0,
+      "learning_rate": 1.883820070212298e-05,
+      "loss": 0.3703,
+      "step": 3593
+    },
+    {
+      "epoch": 0.2208702064896755,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8837111780001822e-05,
+      "loss": 0.3,
+      "step": 3594
+    },
+    {
+      "epoch": 0.2209316617502458,
+      "grad_norm": 21.0,
+      "learning_rate": 1.8836022379313884e-05,
+      "loss": 0.4689,
+      "step": 3595
+    },
+    {
+      "epoch": 0.22099311701081611,
+      "grad_norm": 5.78125,
+      "learning_rate": 1.8834932500118148e-05,
+      "loss": 0.2945,
+      "step": 3596
+    },
+    {
+      "epoch": 0.22105457227138642,
+      "grad_norm": 19.0,
+      "learning_rate": 1.8833842142473643e-05,
+      "loss": 0.4116,
+      "step": 3597
+    },
+    {
+      "epoch": 0.22111602753195675,
+      "grad_norm": 59.0,
+      "learning_rate": 1.883275130643942e-05,
+      "loss": 0.4224,
+      "step": 3598
+    },
+    {
+      "epoch": 0.22117748279252705,
+      "grad_norm": 9.1875,
+      "learning_rate": 1.8831659992074545e-05,
+      "loss": 0.4355,
+      "step": 3599
+    },
+    {
+      "epoch": 0.22123893805309736,
+      "grad_norm": 70.0,
+      "learning_rate": 1.883056819943812e-05,
+      "loss": 0.3068,
+      "step": 3600
+    },
+    {
+      "epoch": 0.22130039331366766,
+      "grad_norm": 11.25,
+      "learning_rate": 1.8829475928589272e-05,
+      "loss": 0.4112,
+      "step": 3601
+    },
+    {
+      "epoch": 0.22136184857423796,
+      "grad_norm": 34.0,
+      "learning_rate": 1.8828383179587147e-05,
+      "loss": 0.4113,
+      "step": 3602
+    },
+    {
+      "epoch": 0.22142330383480827,
+      "grad_norm": 23.0,
+      "learning_rate": 1.882728995249093e-05,
+      "loss": 0.4657,
+      "step": 3603
+    },
+    {
+      "epoch": 0.22148475909537857,
+      "grad_norm": 23.25,
+      "learning_rate": 1.882619624735982e-05,
+      "loss": 0.4101,
+      "step": 3604
+    },
+    {
+      "epoch": 0.22154621435594887,
+      "grad_norm": 18.75,
+      "learning_rate": 1.882510206425304e-05,
+      "loss": 0.3494,
+      "step": 3605
+    },
+    {
+      "epoch": 0.22160766961651918,
+      "grad_norm": 17.125,
+      "learning_rate": 1.8824007403229852e-05,
+      "loss": 0.3571,
+      "step": 3606
+    },
+    {
+      "epoch": 0.22166912487708948,
+      "grad_norm": 10.875,
+      "learning_rate": 1.8822912264349535e-05,
+      "loss": 0.293,
+      "step": 3607
+    },
+    {
+      "epoch": 0.22173058013765978,
+      "grad_norm": 53.5,
+      "learning_rate": 1.8821816647671394e-05,
+      "loss": 0.3555,
+      "step": 3608
+    },
+    {
+      "epoch": 0.2217920353982301,
+      "grad_norm": 4.8125,
+      "learning_rate": 1.882072055325476e-05,
+      "loss": 0.4772,
+      "step": 3609
+    },
+    {
+      "epoch": 0.2218534906588004,
+      "grad_norm": 14.125,
+      "learning_rate": 1.8819623981158996e-05,
+      "loss": 0.4706,
+      "step": 3610
+    },
+    {
+      "epoch": 0.2219149459193707,
+      "grad_norm": 41.0,
+      "learning_rate": 1.881852693144348e-05,
+      "loss": 0.2986,
+      "step": 3611
+    },
+    {
+      "epoch": 0.221976401179941,
+      "grad_norm": 34.0,
+      "learning_rate": 1.881742940416763e-05,
+      "loss": 0.3609,
+      "step": 3612
+    },
+    {
+      "epoch": 0.2220378564405113,
+      "grad_norm": 24.625,
+      "learning_rate": 1.881633139939087e-05,
+      "loss": 0.5729,
+      "step": 3613
+    },
+    {
+      "epoch": 0.2220993117010816,
+      "grad_norm": 38.0,
+      "learning_rate": 1.8815232917172672e-05,
+      "loss": 0.3309,
+      "step": 3614
+    },
+    {
+      "epoch": 0.2221607669616519,
+      "grad_norm": 22.125,
+      "learning_rate": 1.881413395757252e-05,
+      "loss": 0.3461,
+      "step": 3615
+    },
+    {
+      "epoch": 0.2222222222222222,
+      "grad_norm": 7.59375,
+      "learning_rate": 1.8813034520649923e-05,
+      "loss": 0.4716,
+      "step": 3616
+    },
+    {
+      "epoch": 0.2222836774827925,
+      "grad_norm": 35.5,
+      "learning_rate": 1.8811934606464425e-05,
+      "loss": 0.4695,
+      "step": 3617
+    },
+    {
+      "epoch": 0.22234513274336284,
+      "grad_norm": 43.25,
+      "learning_rate": 1.8810834215075593e-05,
+      "loss": 0.7914,
+      "step": 3618
+    },
+    {
+      "epoch": 0.22240658800393315,
+      "grad_norm": 26.5,
+      "learning_rate": 1.8809733346543013e-05,
+      "loss": 0.5603,
+      "step": 3619
+    },
+    {
+      "epoch": 0.22246804326450345,
+      "grad_norm": 26.25,
+      "learning_rate": 1.8808632000926304e-05,
+      "loss": 0.4051,
+      "step": 3620
+    },
+    {
+      "epoch": 0.22252949852507375,
+      "grad_norm": 33.0,
+      "learning_rate": 1.880753017828511e-05,
+      "loss": 0.4801,
+      "step": 3621
+    },
+    {
+      "epoch": 0.22259095378564406,
+      "grad_norm": 17.75,
+      "learning_rate": 1.880642787867909e-05,
+      "loss": 0.4233,
+      "step": 3622
+    },
+    {
+      "epoch": 0.22265240904621436,
+      "grad_norm": 20.25,
+      "learning_rate": 1.880532510216795e-05,
+      "loss": 0.5419,
+      "step": 3623
+    },
+    {
+      "epoch": 0.22271386430678466,
+      "grad_norm": 40.5,
+      "learning_rate": 1.8804221848811408e-05,
+      "loss": 0.5003,
+      "step": 3624
+    },
+    {
+      "epoch": 0.22277531956735497,
+      "grad_norm": 37.25,
+      "learning_rate": 1.8803118118669203e-05,
+      "loss": 0.5687,
+      "step": 3625
+    },
+    {
+      "epoch": 0.22283677482792527,
+      "grad_norm": 4.5625,
+      "learning_rate": 1.880201391180111e-05,
+      "loss": 0.304,
+      "step": 3626
+    },
+    {
+      "epoch": 0.22289823008849557,
+      "grad_norm": 18.625,
+      "learning_rate": 1.8800909228266932e-05,
+      "loss": 0.4378,
+      "step": 3627
+    },
+    {
+      "epoch": 0.22295968534906588,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.8799804068126487e-05,
+      "loss": 0.4157,
+      "step": 3628
+    },
+    {
+      "epoch": 0.22302114060963618,
+      "grad_norm": 18.625,
+      "learning_rate": 1.879869843143962e-05,
+      "loss": 0.3903,
+      "step": 3629
+    },
+    {
+      "epoch": 0.22308259587020648,
+      "grad_norm": 30.125,
+      "learning_rate": 1.8797592318266215e-05,
+      "loss": 0.4224,
+      "step": 3630
+    },
+    {
+      "epoch": 0.2231440511307768,
+      "grad_norm": 28.625,
+      "learning_rate": 1.879648572866617e-05,
+      "loss": 0.3015,
+      "step": 3631
+    },
+    {
+      "epoch": 0.2232055063913471,
+      "grad_norm": 34.75,
+      "learning_rate": 1.8795378662699404e-05,
+      "loss": 0.4788,
+      "step": 3632
+    },
+    {
+      "epoch": 0.2232669616519174,
+      "grad_norm": 30.875,
+      "learning_rate": 1.8794271120425874e-05,
+      "loss": 0.6278,
+      "step": 3633
+    },
+    {
+      "epoch": 0.2233284169124877,
+      "grad_norm": 66.0,
+      "learning_rate": 1.8793163101905562e-05,
+      "loss": 0.7065,
+      "step": 3634
+    },
+    {
+      "epoch": 0.223389872173058,
+      "grad_norm": 25.375,
+      "learning_rate": 1.8792054607198472e-05,
+      "loss": 0.6876,
+      "step": 3635
+    },
+    {
+      "epoch": 0.2234513274336283,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8790945636364628e-05,
+      "loss": 0.6306,
+      "step": 3636
+    },
+    {
+      "epoch": 0.22351278269419864,
+      "grad_norm": 24.375,
+      "learning_rate": 1.878983618946409e-05,
+      "loss": 0.5529,
+      "step": 3637
+    },
+    {
+      "epoch": 0.22357423795476894,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.878872626655693e-05,
+      "loss": 0.4752,
+      "step": 3638
+    },
+    {
+      "epoch": 0.22363569321533924,
+      "grad_norm": 25.5,
+      "learning_rate": 1.878761586770327e-05,
+      "loss": 0.4684,
+      "step": 3639
+    },
+    {
+      "epoch": 0.22369714847590955,
+      "grad_norm": 31.875,
+      "learning_rate": 1.878650499296323e-05,
+      "loss": 0.4198,
+      "step": 3640
+    },
+    {
+      "epoch": 0.22375860373647985,
+      "grad_norm": 46.5,
+      "learning_rate": 1.8785393642396976e-05,
+      "loss": 0.7929,
+      "step": 3641
+    },
+    {
+      "epoch": 0.22382005899705015,
+      "grad_norm": 32.5,
+      "learning_rate": 1.8784281816064688e-05,
+      "loss": 0.6527,
+      "step": 3642
+    },
+    {
+      "epoch": 0.22388151425762046,
+      "grad_norm": 25.375,
+      "learning_rate": 1.878316951402658e-05,
+      "loss": 0.5229,
+      "step": 3643
+    },
+    {
+      "epoch": 0.22394296951819076,
+      "grad_norm": 10.25,
+      "learning_rate": 1.8782056736342886e-05,
+      "loss": 0.4397,
+      "step": 3644
+    },
+    {
+      "epoch": 0.22400442477876106,
+      "grad_norm": 22.25,
+      "learning_rate": 1.8780943483073867e-05,
+      "loss": 0.4247,
+      "step": 3645
+    },
+    {
+      "epoch": 0.22406588003933137,
+      "grad_norm": 16.875,
+      "learning_rate": 1.8779829754279806e-05,
+      "loss": 0.3937,
+      "step": 3646
+    },
+    {
+      "epoch": 0.22412733529990167,
+      "grad_norm": 26.125,
+      "learning_rate": 1.8778715550021025e-05,
+      "loss": 0.5254,
+      "step": 3647
+    },
+    {
+      "epoch": 0.22418879056047197,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.877760087035786e-05,
+      "loss": 0.2696,
+      "step": 3648
+    },
+    {
+      "epoch": 0.22425024582104228,
+      "grad_norm": 11.75,
+      "learning_rate": 1.8776485715350672e-05,
+      "loss": 0.3185,
+      "step": 3649
+    },
+    {
+      "epoch": 0.22431170108161258,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.8775370085059853e-05,
+      "loss": 0.3613,
+      "step": 3650
+    },
+    {
+      "epoch": 0.22437315634218288,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.877425397954582e-05,
+      "loss": 0.3302,
+      "step": 3651
+    },
+    {
+      "epoch": 0.2244346116027532,
+      "grad_norm": 7.0,
+      "learning_rate": 1.8773137398869017e-05,
+      "loss": 0.2979,
+      "step": 3652
+    },
+    {
+      "epoch": 0.2244960668633235,
+      "grad_norm": 9.0,
+      "learning_rate": 1.8772020343089906e-05,
+      "loss": 0.2752,
+      "step": 3653
+    },
+    {
+      "epoch": 0.2245575221238938,
+      "grad_norm": 7.03125,
+      "learning_rate": 1.8770902812268983e-05,
+      "loss": 0.3746,
+      "step": 3654
+    },
+    {
+      "epoch": 0.2246189773844641,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.8769784806466768e-05,
+      "loss": 0.2763,
+      "step": 3655
+    },
+    {
+      "epoch": 0.22468043264503443,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.876866632574381e-05,
+      "loss": 0.3241,
+      "step": 3656
+    },
+    {
+      "epoch": 0.22474188790560473,
+      "grad_norm": 27.875,
+      "learning_rate": 1.8767547370160667e-05,
+      "loss": 0.4575,
+      "step": 3657
+    },
+    {
+      "epoch": 0.22480334316617503,
+      "grad_norm": 16.125,
+      "learning_rate": 1.8766427939777943e-05,
+      "loss": 0.4181,
+      "step": 3658
+    },
+    {
+      "epoch": 0.22486479842674534,
+      "grad_norm": 43.25,
+      "learning_rate": 1.8765308034656265e-05,
+      "loss": 0.5262,
+      "step": 3659
+    },
+    {
+      "epoch": 0.22492625368731564,
+      "grad_norm": 22.0,
+      "learning_rate": 1.876418765485627e-05,
+      "loss": 0.4021,
+      "step": 3660
+    },
+    {
+      "epoch": 0.22498770894788594,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.8763066800438638e-05,
+      "loss": 0.3508,
+      "step": 3661
+    },
+    {
+      "epoch": 0.22504916420845625,
+      "grad_norm": 12.25,
+      "learning_rate": 1.8761945471464065e-05,
+      "loss": 0.2922,
+      "step": 3662
+    },
+    {
+      "epoch": 0.22511061946902655,
+      "grad_norm": 3.0625,
+      "learning_rate": 1.8760823667993277e-05,
+      "loss": 0.2793,
+      "step": 3663
+    },
+    {
+      "epoch": 0.22517207472959685,
+      "grad_norm": 16.5,
+      "learning_rate": 1.8759701390087026e-05,
+      "loss": 0.2548,
+      "step": 3664
+    },
+    {
+      "epoch": 0.22523352999016716,
+      "grad_norm": 29.75,
+      "learning_rate": 1.875857863780609e-05,
+      "loss": 0.4239,
+      "step": 3665
+    },
+    {
+      "epoch": 0.22529498525073746,
+      "grad_norm": 16.375,
+      "learning_rate": 1.875745541121126e-05,
+      "loss": 0.3824,
+      "step": 3666
+    },
+    {
+      "epoch": 0.22535644051130777,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.8756331710363375e-05,
+      "loss": 0.2915,
+      "step": 3667
+    },
+    {
+      "epoch": 0.22541789577187807,
+      "grad_norm": 17.0,
+      "learning_rate": 1.875520753532328e-05,
+      "loss": 0.3001,
+      "step": 3668
+    },
+    {
+      "epoch": 0.22547935103244837,
+      "grad_norm": 48.75,
+      "learning_rate": 1.8754082886151857e-05,
+      "loss": 0.4074,
+      "step": 3669
+    },
+    {
+      "epoch": 0.22554080629301868,
+      "grad_norm": 43.75,
+      "learning_rate": 1.8752957762910016e-05,
+      "loss": 0.3504,
+      "step": 3670
+    },
+    {
+      "epoch": 0.22560226155358898,
+      "grad_norm": 10.125,
+      "learning_rate": 1.8751832165658682e-05,
+      "loss": 0.3128,
+      "step": 3671
+    },
+    {
+      "epoch": 0.22566371681415928,
+      "grad_norm": 26.0,
+      "learning_rate": 1.8750706094458807e-05,
+      "loss": 0.3254,
+      "step": 3672
+    },
+    {
+      "epoch": 0.22572517207472959,
+      "grad_norm": 29.75,
+      "learning_rate": 1.874957954937138e-05,
+      "loss": 0.5327,
+      "step": 3673
+    },
+    {
+      "epoch": 0.2257866273352999,
+      "grad_norm": 32.75,
+      "learning_rate": 1.8748452530457403e-05,
+      "loss": 0.5976,
+      "step": 3674
+    },
+    {
+      "epoch": 0.22584808259587022,
+      "grad_norm": 38.5,
+      "learning_rate": 1.874732503777791e-05,
+      "loss": 0.4392,
+      "step": 3675
+    },
+    {
+      "epoch": 0.22590953785644052,
+      "grad_norm": 28.875,
+      "learning_rate": 1.874619707139396e-05,
+      "loss": 0.5262,
+      "step": 3676
+    },
+    {
+      "epoch": 0.22597099311701083,
+      "grad_norm": 7.0625,
+      "learning_rate": 1.874506863136664e-05,
+      "loss": 0.4742,
+      "step": 3677
+    },
+    {
+      "epoch": 0.22603244837758113,
+      "grad_norm": 19.375,
+      "learning_rate": 1.8743939717757053e-05,
+      "loss": 0.3324,
+      "step": 3678
+    },
+    {
+      "epoch": 0.22609390363815143,
+      "grad_norm": 9.0,
+      "learning_rate": 1.8742810330626338e-05,
+      "loss": 0.3878,
+      "step": 3679
+    },
+    {
+      "epoch": 0.22615535889872174,
+      "grad_norm": 44.25,
+      "learning_rate": 1.8741680470035658e-05,
+      "loss": 0.4829,
+      "step": 3680
+    },
+    {
+      "epoch": 0.22621681415929204,
+      "grad_norm": 30.5,
+      "learning_rate": 1.8740550136046195e-05,
+      "loss": 0.4587,
+      "step": 3681
+    },
+    {
+      "epoch": 0.22627826941986234,
+      "grad_norm": 6.21875,
+      "learning_rate": 1.873941932871917e-05,
+      "loss": 0.4167,
+      "step": 3682
+    },
+    {
+      "epoch": 0.22633972468043265,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8738288048115812e-05,
+      "loss": 0.4226,
+      "step": 3683
+    },
+    {
+      "epoch": 0.22640117994100295,
+      "grad_norm": 8.375,
+      "learning_rate": 1.873715629429739e-05,
+      "loss": 0.3196,
+      "step": 3684
+    },
+    {
+      "epoch": 0.22646263520157325,
+      "grad_norm": 16.5,
+      "learning_rate": 1.8736024067325188e-05,
+      "loss": 0.3357,
+      "step": 3685
+    },
+    {
+      "epoch": 0.22652409046214356,
+      "grad_norm": 3.921875,
+      "learning_rate": 1.8734891367260528e-05,
+      "loss": 0.2814,
+      "step": 3686
+    },
+    {
+      "epoch": 0.22658554572271386,
+      "grad_norm": 3.203125,
+      "learning_rate": 1.873375819416474e-05,
+      "loss": 0.3215,
+      "step": 3687
+    },
+    {
+      "epoch": 0.22664700098328416,
+      "grad_norm": 5.9375,
+      "learning_rate": 1.8732624548099204e-05,
+      "loss": 0.32,
+      "step": 3688
+    },
+    {
+      "epoch": 0.22670845624385447,
+      "grad_norm": 5.9375,
+      "learning_rate": 1.87314904291253e-05,
+      "loss": 0.3085,
+      "step": 3689
+    },
+    {
+      "epoch": 0.22676991150442477,
+      "grad_norm": 29.0,
+      "learning_rate": 1.873035583730445e-05,
+      "loss": 0.5647,
+      "step": 3690
+    },
+    {
+      "epoch": 0.22683136676499507,
+      "grad_norm": 25.0,
+      "learning_rate": 1.8729220772698096e-05,
+      "loss": 0.4393,
+      "step": 3691
+    },
+    {
+      "epoch": 0.22689282202556538,
+      "grad_norm": 20.625,
+      "learning_rate": 1.872808523536771e-05,
+      "loss": 0.4552,
+      "step": 3692
+    },
+    {
+      "epoch": 0.22695427728613568,
+      "grad_norm": 13.375,
+      "learning_rate": 1.8726949225374782e-05,
+      "loss": 0.4206,
+      "step": 3693
+    },
+    {
+      "epoch": 0.227015732546706,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.8725812742780832e-05,
+      "loss": 0.2847,
+      "step": 3694
+    },
+    {
+      "epoch": 0.22707718780727632,
+      "grad_norm": 12.75,
+      "learning_rate": 1.8724675787647405e-05,
+      "loss": 0.3417,
+      "step": 3695
+    },
+    {
+      "epoch": 0.22713864306784662,
+      "grad_norm": 6.875,
+      "learning_rate": 1.872353836003608e-05,
+      "loss": 0.2654,
+      "step": 3696
+    },
+    {
+      "epoch": 0.22720009832841692,
+      "grad_norm": 18.875,
+      "learning_rate": 1.8722400460008437e-05,
+      "loss": 0.5002,
+      "step": 3697
+    },
+    {
+      "epoch": 0.22726155358898723,
+      "grad_norm": 17.625,
+      "learning_rate": 1.8721262087626113e-05,
+      "loss": 0.4191,
+      "step": 3698
+    },
+    {
+      "epoch": 0.22732300884955753,
+      "grad_norm": 16.125,
+      "learning_rate": 1.872012324295075e-05,
+      "loss": 0.3985,
+      "step": 3699
+    },
+    {
+      "epoch": 0.22738446411012783,
+      "grad_norm": 4.3125,
+      "learning_rate": 1.871898392604402e-05,
+      "loss": 0.3253,
+      "step": 3700
+    },
+    {
+      "epoch": 0.22744591937069814,
+      "grad_norm": 21.125,
+      "learning_rate": 1.8717844136967626e-05,
+      "loss": 0.3465,
+      "step": 3701
+    },
+    {
+      "epoch": 0.22750737463126844,
+      "grad_norm": 33.75,
+      "learning_rate": 1.8716703875783287e-05,
+      "loss": 0.4148,
+      "step": 3702
+    },
+    {
+      "epoch": 0.22756882989183874,
+      "grad_norm": 4.96875,
+      "learning_rate": 1.8715563142552758e-05,
+      "loss": 0.4097,
+      "step": 3703
+    },
+    {
+      "epoch": 0.22763028515240905,
+      "grad_norm": 7.75,
+      "learning_rate": 1.8714421937337812e-05,
+      "loss": 0.277,
+      "step": 3704
+    },
+    {
+      "epoch": 0.22769174041297935,
+      "grad_norm": 14.0,
+      "learning_rate": 1.8713280260200245e-05,
+      "loss": 0.3533,
+      "step": 3705
+    },
+    {
+      "epoch": 0.22775319567354965,
+      "grad_norm": 15.75,
+      "learning_rate": 1.8712138111201898e-05,
+      "loss": 0.3028,
+      "step": 3706
+    },
+    {
+      "epoch": 0.22781465093411996,
+      "grad_norm": 18.625,
+      "learning_rate": 1.871099549040461e-05,
+      "loss": 0.3031,
+      "step": 3707
+    },
+    {
+      "epoch": 0.22787610619469026,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.870985239787026e-05,
+      "loss": 0.4144,
+      "step": 3708
+    },
+    {
+      "epoch": 0.22793756145526056,
+      "grad_norm": 19.125,
+      "learning_rate": 1.8708708833660755e-05,
+      "loss": 0.3397,
+      "step": 3709
+    },
+    {
+      "epoch": 0.22799901671583087,
+      "grad_norm": 7.96875,
+      "learning_rate": 1.8707564797838024e-05,
+      "loss": 0.2869,
+      "step": 3710
+    },
+    {
+      "epoch": 0.22806047197640117,
+      "grad_norm": 18.5,
+      "learning_rate": 1.870642029046402e-05,
+      "loss": 0.3242,
+      "step": 3711
+    },
+    {
+      "epoch": 0.22812192723697147,
+      "grad_norm": 6.46875,
+      "learning_rate": 1.8705275311600724e-05,
+      "loss": 0.2222,
+      "step": 3712
+    },
+    {
+      "epoch": 0.22818338249754178,
+      "grad_norm": 38.25,
+      "learning_rate": 1.8704129861310137e-05,
+      "loss": 0.5253,
+      "step": 3713
+    },
+    {
+      "epoch": 0.2282448377581121,
+      "grad_norm": 12.125,
+      "learning_rate": 1.8702983939654296e-05,
+      "loss": 0.4688,
+      "step": 3714
+    },
+    {
+      "epoch": 0.2283062930186824,
+      "grad_norm": 28.25,
+      "learning_rate": 1.870183754669526e-05,
+      "loss": 0.4871,
+      "step": 3715
+    },
+    {
+      "epoch": 0.2283677482792527,
+      "grad_norm": 32.75,
+      "learning_rate": 1.87006906824951e-05,
+      "loss": 0.508,
+      "step": 3716
+    },
+    {
+      "epoch": 0.22842920353982302,
+      "grad_norm": 24.25,
+      "learning_rate": 1.8699543347115932e-05,
+      "loss": 0.7204,
+      "step": 3717
+    },
+    {
+      "epoch": 0.22849065880039332,
+      "grad_norm": 18.125,
+      "learning_rate": 1.8698395540619883e-05,
+      "loss": 0.4022,
+      "step": 3718
+    },
+    {
+      "epoch": 0.22855211406096362,
+      "grad_norm": 30.125,
+      "learning_rate": 1.8697247263069117e-05,
+      "loss": 0.4663,
+      "step": 3719
+    },
+    {
+      "epoch": 0.22861356932153393,
+      "grad_norm": 33.0,
+      "learning_rate": 1.8696098514525817e-05,
+      "loss": 0.611,
+      "step": 3720
+    },
+    {
+      "epoch": 0.22867502458210423,
+      "grad_norm": 8.8125,
+      "learning_rate": 1.869494929505219e-05,
+      "loss": 0.4505,
+      "step": 3721
+    },
+    {
+      "epoch": 0.22873647984267453,
+      "grad_norm": 46.0,
+      "learning_rate": 1.8693799604710476e-05,
+      "loss": 0.3797,
+      "step": 3722
+    },
+    {
+      "epoch": 0.22879793510324484,
+      "grad_norm": 10.625,
+      "learning_rate": 1.869264944356293e-05,
+      "loss": 0.3709,
+      "step": 3723
+    },
+    {
+      "epoch": 0.22885939036381514,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.869149881167184e-05,
+      "loss": 0.419,
+      "step": 3724
+    },
+    {
+      "epoch": 0.22892084562438544,
+      "grad_norm": 8.0625,
+      "learning_rate": 1.8690347709099522e-05,
+      "loss": 0.3723,
+      "step": 3725
+    },
+    {
+      "epoch": 0.22898230088495575,
+      "grad_norm": 25.25,
+      "learning_rate": 1.8689196135908303e-05,
+      "loss": 0.4392,
+      "step": 3726
+    },
+    {
+      "epoch": 0.22904375614552605,
+      "grad_norm": 23.25,
+      "learning_rate": 1.8688044092160554e-05,
+      "loss": 0.326,
+      "step": 3727
+    },
+    {
+      "epoch": 0.22910521140609635,
+      "grad_norm": 18.625,
+      "learning_rate": 1.868689157791866e-05,
+      "loss": 0.291,
+      "step": 3728
+    },
+    {
+      "epoch": 0.22916666666666666,
+      "grad_norm": 17.75,
+      "learning_rate": 1.8685738593245038e-05,
+      "loss": 0.4228,
+      "step": 3729
+    },
+    {
+      "epoch": 0.22922812192723696,
+      "grad_norm": 24.5,
+      "learning_rate": 1.8684585138202122e-05,
+      "loss": 0.3606,
+      "step": 3730
+    },
+    {
+      "epoch": 0.22928957718780726,
+      "grad_norm": 22.25,
+      "learning_rate": 1.868343121285238e-05,
+      "loss": 0.4152,
+      "step": 3731
+    },
+    {
+      "epoch": 0.22935103244837757,
+      "grad_norm": 20.375,
+      "learning_rate": 1.86822768172583e-05,
+      "loss": 0.4731,
+      "step": 3732
+    },
+    {
+      "epoch": 0.2294124877089479,
+      "grad_norm": 18.5,
+      "learning_rate": 1.8681121951482397e-05,
+      "loss": 0.4204,
+      "step": 3733
+    },
+    {
+      "epoch": 0.2294739429695182,
+      "grad_norm": 30.5,
+      "learning_rate": 1.867996661558721e-05,
+      "loss": 0.4416,
+      "step": 3734
+    },
+    {
+      "epoch": 0.2295353982300885,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.867881080963531e-05,
+      "loss": 0.3793,
+      "step": 3735
+    },
+    {
+      "epoch": 0.2295968534906588,
+      "grad_norm": 35.75,
+      "learning_rate": 1.8677654533689287e-05,
+      "loss": 0.5674,
+      "step": 3736
+    },
+    {
+      "epoch": 0.2296583087512291,
+      "grad_norm": 17.125,
+      "learning_rate": 1.867649778781176e-05,
+      "loss": 0.2961,
+      "step": 3737
+    },
+    {
+      "epoch": 0.22971976401179942,
+      "grad_norm": 26.25,
+      "learning_rate": 1.8675340572065365e-05,
+      "loss": 0.4206,
+      "step": 3738
+    },
+    {
+      "epoch": 0.22978121927236972,
+      "grad_norm": 5.875,
+      "learning_rate": 1.8674182886512776e-05,
+      "loss": 0.2942,
+      "step": 3739
+    },
+    {
+      "epoch": 0.22984267453294002,
+      "grad_norm": 5.96875,
+      "learning_rate": 1.8673024731216687e-05,
+      "loss": 0.2738,
+      "step": 3740
+    },
+    {
+      "epoch": 0.22990412979351033,
+      "grad_norm": 44.25,
+      "learning_rate": 1.8671866106239812e-05,
+      "loss": 0.274,
+      "step": 3741
+    },
+    {
+      "epoch": 0.22996558505408063,
+      "grad_norm": 32.5,
+      "learning_rate": 1.86707070116449e-05,
+      "loss": 0.3128,
+      "step": 3742
+    },
+    {
+      "epoch": 0.23002704031465093,
+      "grad_norm": 17.75,
+      "learning_rate": 1.8669547447494723e-05,
+      "loss": 0.2851,
+      "step": 3743
+    },
+    {
+      "epoch": 0.23008849557522124,
+      "grad_norm": 9.875,
+      "learning_rate": 1.866838741385207e-05,
+      "loss": 0.3147,
+      "step": 3744
+    },
+    {
+      "epoch": 0.23014995083579154,
+      "grad_norm": 19.125,
+      "learning_rate": 1.8667226910779767e-05,
+      "loss": 0.4185,
+      "step": 3745
+    },
+    {
+      "epoch": 0.23021140609636184,
+      "grad_norm": 18.875,
+      "learning_rate": 1.866606593834065e-05,
+      "loss": 0.3311,
+      "step": 3746
+    },
+    {
+      "epoch": 0.23027286135693215,
+      "grad_norm": 38.0,
+      "learning_rate": 1.8664904496597605e-05,
+      "loss": 0.4375,
+      "step": 3747
+    },
+    {
+      "epoch": 0.23033431661750245,
+      "grad_norm": 25.75,
+      "learning_rate": 1.866374258561352e-05,
+      "loss": 0.4427,
+      "step": 3748
+    },
+    {
+      "epoch": 0.23039577187807275,
+      "grad_norm": 12.125,
+      "learning_rate": 1.866258020545132e-05,
+      "loss": 0.3772,
+      "step": 3749
+    },
+    {
+      "epoch": 0.23045722713864306,
+      "grad_norm": 7.1875,
+      "learning_rate": 1.866141735617395e-05,
+      "loss": 0.3149,
+      "step": 3750
+    },
+    {
+      "epoch": 0.23051868239921336,
+      "grad_norm": 10.875,
+      "learning_rate": 1.866025403784439e-05,
+      "loss": 0.3991,
+      "step": 3751
+    },
+    {
+      "epoch": 0.2305801376597837,
+      "grad_norm": 21.25,
+      "learning_rate": 1.865909025052563e-05,
+      "loss": 0.254,
+      "step": 3752
+    },
+    {
+      "epoch": 0.230641592920354,
+      "grad_norm": 22.875,
+      "learning_rate": 1.86579259942807e-05,
+      "loss": 0.4425,
+      "step": 3753
+    },
+    {
+      "epoch": 0.2307030481809243,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.8656761269172645e-05,
+      "loss": 0.6698,
+      "step": 3754
+    },
+    {
+      "epoch": 0.2307645034414946,
+      "grad_norm": 4.15625,
+      "learning_rate": 1.8655596075264543e-05,
+      "loss": 0.3844,
+      "step": 3755
+    },
+    {
+      "epoch": 0.2308259587020649,
+      "grad_norm": 17.5,
+      "learning_rate": 1.8654430412619494e-05,
+      "loss": 0.5303,
+      "step": 3756
+    },
+    {
+      "epoch": 0.2308874139626352,
+      "grad_norm": 28.875,
+      "learning_rate": 1.8653264281300622e-05,
+      "loss": 0.4399,
+      "step": 3757
+    },
+    {
+      "epoch": 0.2309488692232055,
+      "grad_norm": 41.75,
+      "learning_rate": 1.8652097681371076e-05,
+      "loss": 0.5223,
+      "step": 3758
+    },
+    {
+      "epoch": 0.23101032448377581,
+      "grad_norm": 3.96875,
+      "learning_rate": 1.865093061289404e-05,
+      "loss": 0.2942,
+      "step": 3759
+    },
+    {
+      "epoch": 0.23107177974434612,
+      "grad_norm": 15.0,
+      "learning_rate": 1.864976307593271e-05,
+      "loss": 0.3056,
+      "step": 3760
+    },
+    {
+      "epoch": 0.23113323500491642,
+      "grad_norm": 28.0,
+      "learning_rate": 1.8648595070550312e-05,
+      "loss": 0.4433,
+      "step": 3761
+    },
+    {
+      "epoch": 0.23119469026548672,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.86474265968101e-05,
+      "loss": 0.4442,
+      "step": 3762
+    },
+    {
+      "epoch": 0.23125614552605703,
+      "grad_norm": 2.265625,
+      "learning_rate": 1.864625765477535e-05,
+      "loss": 0.3193,
+      "step": 3763
+    },
+    {
+      "epoch": 0.23131760078662733,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.864508824450937e-05,
+      "loss": 0.3706,
+      "step": 3764
+    },
+    {
+      "epoch": 0.23137905604719763,
+      "grad_norm": 10.0,
+      "learning_rate": 1.8643918366075485e-05,
+      "loss": 0.337,
+      "step": 3765
+    },
+    {
+      "epoch": 0.23144051130776794,
+      "grad_norm": 3.859375,
+      "learning_rate": 1.864274801953705e-05,
+      "loss": 0.4285,
+      "step": 3766
+    },
+    {
+      "epoch": 0.23150196656833824,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.864157720495744e-05,
+      "loss": 0.439,
+      "step": 3767
+    },
+    {
+      "epoch": 0.23156342182890854,
+      "grad_norm": 7.1875,
+      "learning_rate": 1.8640405922400068e-05,
+      "loss": 0.3328,
+      "step": 3768
+    },
+    {
+      "epoch": 0.23162487708947885,
+      "grad_norm": 23.875,
+      "learning_rate": 1.8639234171928355e-05,
+      "loss": 0.4916,
+      "step": 3769
+    },
+    {
+      "epoch": 0.23168633235004915,
+      "grad_norm": 17.625,
+      "learning_rate": 1.8638061953605762e-05,
+      "loss": 0.4495,
+      "step": 3770
+    },
+    {
+      "epoch": 0.23174778761061948,
+      "grad_norm": 11.0625,
+      "learning_rate": 1.8636889267495767e-05,
+      "loss": 0.4177,
+      "step": 3771
+    },
+    {
+      "epoch": 0.23180924287118979,
+      "grad_norm": 14.0,
+      "learning_rate": 1.8635716113661876e-05,
+      "loss": 0.3727,
+      "step": 3772
+    },
+    {
+      "epoch": 0.2318706981317601,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.863454249216762e-05,
+      "loss": 0.3415,
+      "step": 3773
+    },
+    {
+      "epoch": 0.2319321533923304,
+      "grad_norm": 23.625,
+      "learning_rate": 1.8633368403076558e-05,
+      "loss": 0.4788,
+      "step": 3774
+    },
+    {
+      "epoch": 0.2319936086529007,
+      "grad_norm": 15.875,
+      "learning_rate": 1.863219384645227e-05,
+      "loss": 0.4249,
+      "step": 3775
+    },
+    {
+      "epoch": 0.232055063913471,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.8631018822358363e-05,
+      "loss": 0.288,
+      "step": 3776
+    },
+    {
+      "epoch": 0.2321165191740413,
+      "grad_norm": 27.25,
+      "learning_rate": 1.8629843330858472e-05,
+      "loss": 0.6041,
+      "step": 3777
+    },
+    {
+      "epoch": 0.2321779744346116,
+      "grad_norm": 20.25,
+      "learning_rate": 1.862866737201625e-05,
+      "loss": 0.3561,
+      "step": 3778
+    },
+    {
+      "epoch": 0.2322394296951819,
+      "grad_norm": 9.125,
+      "learning_rate": 1.8627490945895387e-05,
+      "loss": 0.6297,
+      "step": 3779
+    },
+    {
+      "epoch": 0.2323008849557522,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.8626314052559583e-05,
+      "loss": 0.611,
+      "step": 3780
+    },
+    {
+      "epoch": 0.23236234021632252,
+      "grad_norm": 27.5,
+      "learning_rate": 1.8625136692072577e-05,
+      "loss": 0.4662,
+      "step": 3781
+    },
+    {
+      "epoch": 0.23242379547689282,
+      "grad_norm": 30.5,
+      "learning_rate": 1.8623958864498125e-05,
+      "loss": 0.4833,
+      "step": 3782
+    },
+    {
+      "epoch": 0.23248525073746312,
+      "grad_norm": 26.875,
+      "learning_rate": 1.8622780569900016e-05,
+      "loss": 0.4608,
+      "step": 3783
+    },
+    {
+      "epoch": 0.23254670599803343,
+      "grad_norm": 17.25,
+      "learning_rate": 1.862160180834206e-05,
+      "loss": 0.6771,
+      "step": 3784
+    },
+    {
+      "epoch": 0.23260816125860373,
+      "grad_norm": 12.4375,
+      "learning_rate": 1.8620422579888082e-05,
+      "loss": 0.4398,
+      "step": 3785
+    },
+    {
+      "epoch": 0.23266961651917403,
+      "grad_norm": 35.5,
+      "learning_rate": 1.8619242884601953e-05,
+      "loss": 0.5837,
+      "step": 3786
+    },
+    {
+      "epoch": 0.23273107177974434,
+      "grad_norm": 10.125,
+      "learning_rate": 1.861806272254755e-05,
+      "loss": 0.5081,
+      "step": 3787
+    },
+    {
+      "epoch": 0.23279252704031464,
+      "grad_norm": 16.0,
+      "learning_rate": 1.8616882093788792e-05,
+      "loss": 0.3637,
+      "step": 3788
+    },
+    {
+      "epoch": 0.23285398230088494,
+      "grad_norm": 14.5,
+      "learning_rate": 1.8615700998389612e-05,
+      "loss": 0.3568,
+      "step": 3789
+    },
+    {
+      "epoch": 0.23291543756145527,
+      "grad_norm": 21.5,
+      "learning_rate": 1.8614519436413968e-05,
+      "loss": 0.4398,
+      "step": 3790
+    },
+    {
+      "epoch": 0.23297689282202558,
+      "grad_norm": 28.125,
+      "learning_rate": 1.8613337407925854e-05,
+      "loss": 0.3344,
+      "step": 3791
+    },
+    {
+      "epoch": 0.23303834808259588,
+      "grad_norm": 19.125,
+      "learning_rate": 1.861215491298927e-05,
+      "loss": 0.3715,
+      "step": 3792
+    },
+    {
+      "epoch": 0.23309980334316618,
+      "grad_norm": 5.40625,
+      "learning_rate": 1.8610971951668265e-05,
+      "loss": 0.3784,
+      "step": 3793
+    },
+    {
+      "epoch": 0.2331612586037365,
+      "grad_norm": 44.75,
+      "learning_rate": 1.8609788524026894e-05,
+      "loss": 0.6965,
+      "step": 3794
+    },
+    {
+      "epoch": 0.2332227138643068,
+      "grad_norm": 21.875,
+      "learning_rate": 1.8608604630129247e-05,
+      "loss": 0.4643,
+      "step": 3795
+    },
+    {
+      "epoch": 0.2332841691248771,
+      "grad_norm": 37.0,
+      "learning_rate": 1.860742027003944e-05,
+      "loss": 0.5774,
+      "step": 3796
+    },
+    {
+      "epoch": 0.2333456243854474,
+      "grad_norm": 17.5,
+      "learning_rate": 1.8606235443821602e-05,
+      "loss": 0.428,
+      "step": 3797
+    },
+    {
+      "epoch": 0.2334070796460177,
+      "grad_norm": 30.0,
+      "learning_rate": 1.8605050151539906e-05,
+      "loss": 0.3946,
+      "step": 3798
+    },
+    {
+      "epoch": 0.233468534906588,
+      "grad_norm": 10.625,
+      "learning_rate": 1.8603864393258534e-05,
+      "loss": 0.4832,
+      "step": 3799
+    },
+    {
+      "epoch": 0.2335299901671583,
+      "grad_norm": 26.625,
+      "learning_rate": 1.8602678169041705e-05,
+      "loss": 0.3553,
+      "step": 3800
+    },
+    {
+      "epoch": 0.2335914454277286,
+      "grad_norm": 13.25,
+      "learning_rate": 1.860149147895366e-05,
+      "loss": 0.3103,
+      "step": 3801
+    },
+    {
+      "epoch": 0.23365290068829891,
+      "grad_norm": 22.75,
+      "learning_rate": 1.860030432305865e-05,
+      "loss": 0.4174,
+      "step": 3802
+    },
+    {
+      "epoch": 0.23371435594886922,
+      "grad_norm": 7.9375,
+      "learning_rate": 1.8599116701420974e-05,
+      "loss": 0.3957,
+      "step": 3803
+    },
+    {
+      "epoch": 0.23377581120943952,
+      "grad_norm": 12.875,
+      "learning_rate": 1.859792861410495e-05,
+      "loss": 0.2933,
+      "step": 3804
+    },
+    {
+      "epoch": 0.23383726647000982,
+      "grad_norm": 29.125,
+      "learning_rate": 1.8596740061174912e-05,
+      "loss": 0.3773,
+      "step": 3805
+    },
+    {
+      "epoch": 0.23389872173058013,
+      "grad_norm": 5.59375,
+      "learning_rate": 1.859555104269523e-05,
+      "loss": 0.3295,
+      "step": 3806
+    },
+    {
+      "epoch": 0.23396017699115043,
+      "grad_norm": 16.625,
+      "learning_rate": 1.8594361558730287e-05,
+      "loss": 0.2904,
+      "step": 3807
+    },
+    {
+      "epoch": 0.23402163225172073,
+      "grad_norm": 25.5,
+      "learning_rate": 1.8593171609344505e-05,
+      "loss": 0.485,
+      "step": 3808
+    },
+    {
+      "epoch": 0.23408308751229107,
+      "grad_norm": 10.25,
+      "learning_rate": 1.8591981194602322e-05,
+      "loss": 0.4362,
+      "step": 3809
+    },
+    {
+      "epoch": 0.23414454277286137,
+      "grad_norm": 34.0,
+      "learning_rate": 1.8590790314568207e-05,
+      "loss": 0.5425,
+      "step": 3810
+    },
+    {
+      "epoch": 0.23420599803343167,
+      "grad_norm": 18.875,
+      "learning_rate": 1.8589598969306646e-05,
+      "loss": 0.396,
+      "step": 3811
+    },
+    {
+      "epoch": 0.23426745329400198,
+      "grad_norm": 21.375,
+      "learning_rate": 1.8588407158882157e-05,
+      "loss": 0.3493,
+      "step": 3812
+    },
+    {
+      "epoch": 0.23432890855457228,
+      "grad_norm": 23.375,
+      "learning_rate": 1.858721488335929e-05,
+      "loss": 0.3823,
+      "step": 3813
+    },
+    {
+      "epoch": 0.23439036381514258,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.8586022142802597e-05,
+      "loss": 0.3937,
+      "step": 3814
+    },
+    {
+      "epoch": 0.2344518190757129,
+      "grad_norm": 26.375,
+      "learning_rate": 1.858482893727668e-05,
+      "loss": 0.3148,
+      "step": 3815
+    },
+    {
+      "epoch": 0.2345132743362832,
+      "grad_norm": 4.28125,
+      "learning_rate": 1.8583635266846155e-05,
+      "loss": 0.3068,
+      "step": 3816
+    },
+    {
+      "epoch": 0.2345747295968535,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.8582441131575658e-05,
+      "loss": 0.3567,
+      "step": 3817
+    },
+    {
+      "epoch": 0.2346361848574238,
+      "grad_norm": 17.125,
+      "learning_rate": 1.8581246531529868e-05,
+      "loss": 0.3456,
+      "step": 3818
+    },
+    {
+      "epoch": 0.2346976401179941,
+      "grad_norm": 41.25,
+      "learning_rate": 1.8580051466773465e-05,
+      "loss": 0.3157,
+      "step": 3819
+    },
+    {
+      "epoch": 0.2347590953785644,
+      "grad_norm": 5.15625,
+      "learning_rate": 1.8578855937371176e-05,
+      "loss": 0.2712,
+      "step": 3820
+    },
+    {
+      "epoch": 0.2348205506391347,
+      "grad_norm": 7.9375,
+      "learning_rate": 1.8577659943387737e-05,
+      "loss": 0.362,
+      "step": 3821
+    },
+    {
+      "epoch": 0.234882005899705,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.8576463484887927e-05,
+      "loss": 0.2734,
+      "step": 3822
+    },
+    {
+      "epoch": 0.2349434611602753,
+      "grad_norm": 4.90625,
+      "learning_rate": 1.8575266561936526e-05,
+      "loss": 0.2876,
+      "step": 3823
+    },
+    {
+      "epoch": 0.23500491642084562,
+      "grad_norm": 16.125,
+      "learning_rate": 1.857406917459836e-05,
+      "loss": 0.3983,
+      "step": 3824
+    },
+    {
+      "epoch": 0.23506637168141592,
+      "grad_norm": 46.0,
+      "learning_rate": 1.8572871322938267e-05,
+      "loss": 0.5058,
+      "step": 3825
+    },
+    {
+      "epoch": 0.23512782694198622,
+      "grad_norm": 36.5,
+      "learning_rate": 1.8571673007021124e-05,
+      "loss": 0.5802,
+      "step": 3826
+    },
+    {
+      "epoch": 0.23518928220255653,
+      "grad_norm": 41.0,
+      "learning_rate": 1.857047422691182e-05,
+      "loss": 0.4912,
+      "step": 3827
+    },
+    {
+      "epoch": 0.23525073746312683,
+      "grad_norm": 10.4375,
+      "learning_rate": 1.8569274982675273e-05,
+      "loss": 0.5207,
+      "step": 3828
+    },
+    {
+      "epoch": 0.23531219272369716,
+      "grad_norm": 13.375,
+      "learning_rate": 1.856807527437643e-05,
+      "loss": 0.4101,
+      "step": 3829
+    },
+    {
+      "epoch": 0.23537364798426746,
+      "grad_norm": 39.75,
+      "learning_rate": 1.856687510208026e-05,
+      "loss": 0.6383,
+      "step": 3830
+    },
+    {
+      "epoch": 0.23543510324483777,
+      "grad_norm": 45.0,
+      "learning_rate": 1.8565674465851753e-05,
+      "loss": 0.4806,
+      "step": 3831
+    },
+    {
+      "epoch": 0.23549655850540807,
+      "grad_norm": 10.75,
+      "learning_rate": 1.8564473365755936e-05,
+      "loss": 0.4865,
+      "step": 3832
+    },
+    {
+      "epoch": 0.23555801376597837,
+      "grad_norm": 18.75,
+      "learning_rate": 1.8563271801857844e-05,
+      "loss": 0.3519,
+      "step": 3833
+    },
+    {
+      "epoch": 0.23561946902654868,
+      "grad_norm": 39.0,
+      "learning_rate": 1.8562069774222556e-05,
+      "loss": 0.5569,
+      "step": 3834
+    },
+    {
+      "epoch": 0.23568092428711898,
+      "grad_norm": 24.375,
+      "learning_rate": 1.8560867282915164e-05,
+      "loss": 0.2704,
+      "step": 3835
+    },
+    {
+      "epoch": 0.23574237954768928,
+      "grad_norm": 23.375,
+      "learning_rate": 1.8559664328000782e-05,
+      "loss": 0.3785,
+      "step": 3836
+    },
+    {
+      "epoch": 0.2358038348082596,
+      "grad_norm": 19.0,
+      "learning_rate": 1.8558460909544564e-05,
+      "loss": 0.6539,
+      "step": 3837
+    },
+    {
+      "epoch": 0.2358652900688299,
+      "grad_norm": 17.25,
+      "learning_rate": 1.8557257027611677e-05,
+      "loss": 0.4072,
+      "step": 3838
+    },
+    {
+      "epoch": 0.2359267453294002,
+      "grad_norm": 9.375,
+      "learning_rate": 1.8556052682267314e-05,
+      "loss": 0.2761,
+      "step": 3839
+    },
+    {
+      "epoch": 0.2359882005899705,
+      "grad_norm": 22.375,
+      "learning_rate": 1.8554847873576696e-05,
+      "loss": 0.3507,
+      "step": 3840
+    },
+    {
+      "epoch": 0.2360496558505408,
+      "grad_norm": 32.5,
+      "learning_rate": 1.855364260160507e-05,
+      "loss": 0.4804,
+      "step": 3841
+    },
+    {
+      "epoch": 0.2361111111111111,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.8552436866417703e-05,
+      "loss": 0.3391,
+      "step": 3842
+    },
+    {
+      "epoch": 0.2361725663716814,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.8551230668079895e-05,
+      "loss": 0.4324,
+      "step": 3843
+    },
+    {
+      "epoch": 0.2362340216322517,
+      "grad_norm": 25.0,
+      "learning_rate": 1.8550024006656967e-05,
+      "loss": 0.5259,
+      "step": 3844
+    },
+    {
+      "epoch": 0.23629547689282202,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8548816882214262e-05,
+      "loss": 0.4221,
+      "step": 3845
+    },
+    {
+      "epoch": 0.23635693215339232,
+      "grad_norm": 17.125,
+      "learning_rate": 1.854760929481715e-05,
+      "loss": 0.4661,
+      "step": 3846
+    },
+    {
+      "epoch": 0.23641838741396262,
+      "grad_norm": 7.875,
+      "learning_rate": 1.854640124453103e-05,
+      "loss": 0.3656,
+      "step": 3847
+    },
+    {
+      "epoch": 0.23647984267453295,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.8545192731421323e-05,
+      "loss": 0.5368,
+      "step": 3848
+    },
+    {
+      "epoch": 0.23654129793510326,
+      "grad_norm": 21.5,
+      "learning_rate": 1.8543983755553474e-05,
+      "loss": 0.4362,
+      "step": 3849
+    },
+    {
+      "epoch": 0.23660275319567356,
+      "grad_norm": 10.0625,
+      "learning_rate": 1.8542774316992953e-05,
+      "loss": 0.3799,
+      "step": 3850
+    },
+    {
+      "epoch": 0.23666420845624386,
+      "grad_norm": 18.375,
+      "learning_rate": 1.854156441580526e-05,
+      "loss": 0.4987,
+      "step": 3851
+    },
+    {
+      "epoch": 0.23672566371681417,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8540354052055908e-05,
+      "loss": 0.4421,
+      "step": 3852
+    },
+    {
+      "epoch": 0.23678711897738447,
+      "grad_norm": 5.125,
+      "learning_rate": 1.8539143225810453e-05,
+      "loss": 0.3247,
+      "step": 3853
+    },
+    {
+      "epoch": 0.23684857423795477,
+      "grad_norm": 4.6875,
+      "learning_rate": 1.853793193713446e-05,
+      "loss": 0.2877,
+      "step": 3854
+    },
+    {
+      "epoch": 0.23691002949852508,
+      "grad_norm": 22.0,
+      "learning_rate": 1.8536720186093528e-05,
+      "loss": 0.2818,
+      "step": 3855
+    },
+    {
+      "epoch": 0.23697148475909538,
+      "grad_norm": 2.8125,
+      "learning_rate": 1.8535507972753275e-05,
+      "loss": 0.2693,
+      "step": 3856
+    },
+    {
+      "epoch": 0.23703294001966568,
+      "grad_norm": 16.625,
+      "learning_rate": 1.8534295297179354e-05,
+      "loss": 0.3294,
+      "step": 3857
+    },
+    {
+      "epoch": 0.237094395280236,
+      "grad_norm": 8.5625,
+      "learning_rate": 1.8533082159437433e-05,
+      "loss": 0.4744,
+      "step": 3858
+    },
+    {
+      "epoch": 0.2371558505408063,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8531868559593205e-05,
+      "loss": 0.3113,
+      "step": 3859
+    },
+    {
+      "epoch": 0.2372173058013766,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8530654497712395e-05,
+      "loss": 0.4113,
+      "step": 3860
+    },
+    {
+      "epoch": 0.2372787610619469,
+      "grad_norm": 15.8125,
+      "learning_rate": 1.852943997386075e-05,
+      "loss": 0.3171,
+      "step": 3861
+    },
+    {
+      "epoch": 0.2373402163225172,
+      "grad_norm": 6.90625,
+      "learning_rate": 1.8528224988104044e-05,
+      "loss": 0.2275,
+      "step": 3862
+    },
+    {
+      "epoch": 0.2374016715830875,
+      "grad_norm": 21.625,
+      "learning_rate": 1.852700954050807e-05,
+      "loss": 0.4057,
+      "step": 3863
+    },
+    {
+      "epoch": 0.2374631268436578,
+      "grad_norm": 36.75,
+      "learning_rate": 1.8525793631138647e-05,
+      "loss": 0.3626,
+      "step": 3864
+    },
+    {
+      "epoch": 0.2375245821042281,
+      "grad_norm": 24.625,
+      "learning_rate": 1.8524577260061628e-05,
+      "loss": 0.3797,
+      "step": 3865
+    },
+    {
+      "epoch": 0.2375860373647984,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.8523360427342877e-05,
+      "loss": 0.6008,
+      "step": 3866
+    },
+    {
+      "epoch": 0.23764749262536874,
+      "grad_norm": 8.625,
+      "learning_rate": 1.8522143133048298e-05,
+      "loss": 0.3107,
+      "step": 3867
+    },
+    {
+      "epoch": 0.23770894788593905,
+      "grad_norm": 33.75,
+      "learning_rate": 1.8520925377243812e-05,
+      "loss": 0.3394,
+      "step": 3868
+    },
+    {
+      "epoch": 0.23777040314650935,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.851970715999536e-05,
+      "loss": 0.3601,
+      "step": 3869
+    },
+    {
+      "epoch": 0.23783185840707965,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.8518488481368915e-05,
+      "loss": 0.2946,
+      "step": 3870
+    },
+    {
+      "epoch": 0.23789331366764996,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.851726934143048e-05,
+      "loss": 0.27,
+      "step": 3871
+    },
+    {
+      "epoch": 0.23795476892822026,
+      "grad_norm": 5.28125,
+      "learning_rate": 1.8516049740246068e-05,
+      "loss": 0.361,
+      "step": 3872
+    },
+    {
+      "epoch": 0.23801622418879056,
+      "grad_norm": 3.984375,
+      "learning_rate": 1.851482967788173e-05,
+      "loss": 0.3298,
+      "step": 3873
+    },
+    {
+      "epoch": 0.23807767944936087,
+      "grad_norm": 12.8125,
+      "learning_rate": 1.8513609154403535e-05,
+      "loss": 0.3937,
+      "step": 3874
+    },
+    {
+      "epoch": 0.23813913470993117,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.8512388169877587e-05,
+      "loss": 0.4005,
+      "step": 3875
+    },
+    {
+      "epoch": 0.23820058997050148,
+      "grad_norm": 7.15625,
+      "learning_rate": 1.8511166724369997e-05,
+      "loss": 0.3274,
+      "step": 3876
+    },
+    {
+      "epoch": 0.23826204523107178,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.850994481794692e-05,
+      "loss": 0.3153,
+      "step": 3877
+    },
+    {
+      "epoch": 0.23832350049164208,
+      "grad_norm": 11.375,
+      "learning_rate": 1.850872245067452e-05,
+      "loss": 0.3425,
+      "step": 3878
+    },
+    {
+      "epoch": 0.23838495575221239,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.8507499622619e-05,
+      "loss": 0.2811,
+      "step": 3879
+    },
+    {
+      "epoch": 0.2384464110127827,
+      "grad_norm": 12.25,
+      "learning_rate": 1.850627633384658e-05,
+      "loss": 0.4582,
+      "step": 3880
+    },
+    {
+      "epoch": 0.238507866273353,
+      "grad_norm": 4.125,
+      "learning_rate": 1.85050525844235e-05,
+      "loss": 0.3322,
+      "step": 3881
+    },
+    {
+      "epoch": 0.2385693215339233,
+      "grad_norm": 55.5,
+      "learning_rate": 1.8503828374416038e-05,
+      "loss": 0.3364,
+      "step": 3882
+    },
+    {
+      "epoch": 0.2386307767944936,
+      "grad_norm": 12.4375,
+      "learning_rate": 1.8502603703890488e-05,
+      "loss": 0.3438,
+      "step": 3883
+    },
+    {
+      "epoch": 0.2386922320550639,
+      "grad_norm": 42.75,
+      "learning_rate": 1.850137857291317e-05,
+      "loss": 0.3413,
+      "step": 3884
+    },
+    {
+      "epoch": 0.2387536873156342,
+      "grad_norm": 5.375,
+      "learning_rate": 1.850015298155043e-05,
+      "loss": 0.3213,
+      "step": 3885
+    },
+    {
+      "epoch": 0.23881514257620454,
+      "grad_norm": 12.5625,
+      "learning_rate": 1.849892692986864e-05,
+      "loss": 0.3287,
+      "step": 3886
+    },
+    {
+      "epoch": 0.23887659783677484,
+      "grad_norm": 34.75,
+      "learning_rate": 1.8497700417934195e-05,
+      "loss": 0.3921,
+      "step": 3887
+    },
+    {
+      "epoch": 0.23893805309734514,
+      "grad_norm": 35.25,
+      "learning_rate": 1.849647344581352e-05,
+      "loss": 0.3449,
+      "step": 3888
+    },
+    {
+      "epoch": 0.23899950835791545,
+      "grad_norm": 2.59375,
+      "learning_rate": 1.8495246013573057e-05,
+      "loss": 0.4264,
+      "step": 3889
+    },
+    {
+      "epoch": 0.23906096361848575,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.8494018121279274e-05,
+      "loss": 0.5112,
+      "step": 3890
+    },
+    {
+      "epoch": 0.23912241887905605,
+      "grad_norm": 34.25,
+      "learning_rate": 1.8492789768998668e-05,
+      "loss": 0.5012,
+      "step": 3891
+    },
+    {
+      "epoch": 0.23918387413962636,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.8491560956797766e-05,
+      "loss": 0.5751,
+      "step": 3892
+    },
+    {
+      "epoch": 0.23924532940019666,
+      "grad_norm": 19.5,
+      "learning_rate": 1.8490331684743105e-05,
+      "loss": 0.4268,
+      "step": 3893
+    },
+    {
+      "epoch": 0.23930678466076696,
+      "grad_norm": 20.875,
+      "learning_rate": 1.8489101952901256e-05,
+      "loss": 0.5277,
+      "step": 3894
+    },
+    {
+      "epoch": 0.23936823992133727,
+      "grad_norm": 6.71875,
+      "learning_rate": 1.848787176133882e-05,
+      "loss": 0.3717,
+      "step": 3895
+    },
+    {
+      "epoch": 0.23942969518190757,
+      "grad_norm": 17.25,
+      "learning_rate": 1.848664111012241e-05,
+      "loss": 0.3938,
+      "step": 3896
+    },
+    {
+      "epoch": 0.23949115044247787,
+      "grad_norm": 20.5,
+      "learning_rate": 1.848540999931868e-05,
+      "loss": 0.4175,
+      "step": 3897
+    },
+    {
+      "epoch": 0.23955260570304818,
+      "grad_norm": 4.84375,
+      "learning_rate": 1.848417842899429e-05,
+      "loss": 0.2745,
+      "step": 3898
+    },
+    {
+      "epoch": 0.23961406096361848,
+      "grad_norm": 8.3125,
+      "learning_rate": 1.8482946399215943e-05,
+      "loss": 0.481,
+      "step": 3899
+    },
+    {
+      "epoch": 0.23967551622418878,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.848171391005035e-05,
+      "loss": 0.4082,
+      "step": 3900
+    },
+    {
+      "epoch": 0.2397369714847591,
+      "grad_norm": 29.25,
+      "learning_rate": 1.848048096156426e-05,
+      "loss": 0.6365,
+      "step": 3901
+    },
+    {
+      "epoch": 0.2397984267453294,
+      "grad_norm": 38.25,
+      "learning_rate": 1.8479247553824445e-05,
+      "loss": 0.3512,
+      "step": 3902
+    },
+    {
+      "epoch": 0.2398598820058997,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.8478013686897696e-05,
+      "loss": 0.4087,
+      "step": 3903
+    },
+    {
+      "epoch": 0.23992133726647,
+      "grad_norm": 4.8125,
+      "learning_rate": 1.8476779360850833e-05,
+      "loss": 0.3764,
+      "step": 3904
+    },
+    {
+      "epoch": 0.23998279252704033,
+      "grad_norm": 4.4375,
+      "learning_rate": 1.84755445757507e-05,
+      "loss": 0.4335,
+      "step": 3905
+    },
+    {
+      "epoch": 0.24004424778761063,
+      "grad_norm": 14.4375,
+      "learning_rate": 1.8474309331664165e-05,
+      "loss": 0.4527,
+      "step": 3906
+    },
+    {
+      "epoch": 0.24010570304818094,
+      "grad_norm": 9.0,
+      "learning_rate": 1.8473073628658123e-05,
+      "loss": 0.5094,
+      "step": 3907
+    },
+    {
+      "epoch": 0.24016715830875124,
+      "grad_norm": 30.375,
+      "learning_rate": 1.847183746679949e-05,
+      "loss": 0.5736,
+      "step": 3908
+    },
+    {
+      "epoch": 0.24022861356932154,
+      "grad_norm": 17.75,
+      "learning_rate": 1.8470600846155213e-05,
+      "loss": 0.4153,
+      "step": 3909
+    },
+    {
+      "epoch": 0.24029006882989185,
+      "grad_norm": 10.0,
+      "learning_rate": 1.8469363766792258e-05,
+      "loss": 0.4624,
+      "step": 3910
+    },
+    {
+      "epoch": 0.24035152409046215,
+      "grad_norm": 24.5,
+      "learning_rate": 1.8468126228777617e-05,
+      "loss": 0.5461,
+      "step": 3911
+    },
+    {
+      "epoch": 0.24041297935103245,
+      "grad_norm": 51.25,
+      "learning_rate": 1.8466888232178312e-05,
+      "loss": 0.4496,
+      "step": 3912
+    },
+    {
+      "epoch": 0.24047443461160276,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8465649777061377e-05,
+      "loss": 0.5879,
+      "step": 3913
+    },
+    {
+      "epoch": 0.24053588987217306,
+      "grad_norm": 26.875,
+      "learning_rate": 1.846441086349389e-05,
+      "loss": 0.6312,
+      "step": 3914
+    },
+    {
+      "epoch": 0.24059734513274336,
+      "grad_norm": 21.875,
+      "learning_rate": 1.846317149154294e-05,
+      "loss": 0.4866,
+      "step": 3915
+    },
+    {
+      "epoch": 0.24065880039331367,
+      "grad_norm": 39.0,
+      "learning_rate": 1.8461931661275642e-05,
+      "loss": 0.5921,
+      "step": 3916
+    },
+    {
+      "epoch": 0.24072025565388397,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.846069137275914e-05,
+      "loss": 0.4138,
+      "step": 3917
+    },
+    {
+      "epoch": 0.24078171091445427,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.8459450626060596e-05,
+      "loss": 0.4337,
+      "step": 3918
+    },
+    {
+      "epoch": 0.24084316617502458,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8458209421247208e-05,
+      "loss": 0.4422,
+      "step": 3919
+    },
+    {
+      "epoch": 0.24090462143559488,
+      "grad_norm": 25.75,
+      "learning_rate": 1.8456967758386192e-05,
+      "loss": 0.5354,
+      "step": 3920
+    },
+    {
+      "epoch": 0.24096607669616518,
+      "grad_norm": 50.0,
+      "learning_rate": 1.8455725637544784e-05,
+      "loss": 0.4677,
+      "step": 3921
+    },
+    {
+      "epoch": 0.24102753195673549,
+      "grad_norm": 32.25,
+      "learning_rate": 1.8454483058790254e-05,
+      "loss": 0.5111,
+      "step": 3922
+    },
+    {
+      "epoch": 0.2410889872173058,
+      "grad_norm": 26.875,
+      "learning_rate": 1.8453240022189896e-05,
+      "loss": 0.3977,
+      "step": 3923
+    },
+    {
+      "epoch": 0.2411504424778761,
+      "grad_norm": 19.25,
+      "learning_rate": 1.8451996527811018e-05,
+      "loss": 0.4372,
+      "step": 3924
+    },
+    {
+      "epoch": 0.24121189773844642,
+      "grad_norm": 9.5,
+      "learning_rate": 1.8450752575720967e-05,
+      "loss": 0.4313,
+      "step": 3925
+    },
+    {
+      "epoch": 0.24127335299901673,
+      "grad_norm": 28.5,
+      "learning_rate": 1.8449508165987106e-05,
+      "loss": 0.4289,
+      "step": 3926
+    },
+    {
+      "epoch": 0.24133480825958703,
+      "grad_norm": 38.25,
+      "learning_rate": 1.844826329867682e-05,
+      "loss": 0.3315,
+      "step": 3927
+    },
+    {
+      "epoch": 0.24139626352015733,
+      "grad_norm": 9.25,
+      "learning_rate": 1.844701797385753e-05,
+      "loss": 0.2797,
+      "step": 3928
+    },
+    {
+      "epoch": 0.24145771878072764,
+      "grad_norm": 23.375,
+      "learning_rate": 1.8445772191596676e-05,
+      "loss": 0.3894,
+      "step": 3929
+    },
+    {
+      "epoch": 0.24151917404129794,
+      "grad_norm": 19.75,
+      "learning_rate": 1.844452595196172e-05,
+      "loss": 0.5158,
+      "step": 3930
+    },
+    {
+      "epoch": 0.24158062930186824,
+      "grad_norm": 34.5,
+      "learning_rate": 1.8443279255020153e-05,
+      "loss": 0.5255,
+      "step": 3931
+    },
+    {
+      "epoch": 0.24164208456243855,
+      "grad_norm": 12.125,
+      "learning_rate": 1.8442032100839484e-05,
+      "loss": 0.4037,
+      "step": 3932
+    },
+    {
+      "epoch": 0.24170353982300885,
+      "grad_norm": 5.21875,
+      "learning_rate": 1.8440784489487256e-05,
+      "loss": 0.3674,
+      "step": 3933
+    },
+    {
+      "epoch": 0.24176499508357915,
+      "grad_norm": 27.375,
+      "learning_rate": 1.8439536421031035e-05,
+      "loss": 0.3846,
+      "step": 3934
+    },
+    {
+      "epoch": 0.24182645034414946,
+      "grad_norm": 20.125,
+      "learning_rate": 1.84382878955384e-05,
+      "loss": 0.4185,
+      "step": 3935
+    },
+    {
+      "epoch": 0.24188790560471976,
+      "grad_norm": 8.25,
+      "learning_rate": 1.8437038913076974e-05,
+      "loss": 0.4124,
+      "step": 3936
+    },
+    {
+      "epoch": 0.24194936086529006,
+      "grad_norm": 21.5,
+      "learning_rate": 1.843578947371439e-05,
+      "loss": 0.4238,
+      "step": 3937
+    },
+    {
+      "epoch": 0.24201081612586037,
+      "grad_norm": 16.5,
+      "learning_rate": 1.843453957751831e-05,
+      "loss": 0.3941,
+      "step": 3938
+    },
+    {
+      "epoch": 0.24207227138643067,
+      "grad_norm": 16.875,
+      "learning_rate": 1.843328922455642e-05,
+      "loss": 0.4682,
+      "step": 3939
+    },
+    {
+      "epoch": 0.24213372664700097,
+      "grad_norm": 17.625,
+      "learning_rate": 1.8432038414896432e-05,
+      "loss": 0.5307,
+      "step": 3940
+    },
+    {
+      "epoch": 0.24219518190757128,
+      "grad_norm": 14.75,
+      "learning_rate": 1.8430787148606087e-05,
+      "loss": 0.3972,
+      "step": 3941
+    },
+    {
+      "epoch": 0.24225663716814158,
+      "grad_norm": 20.0,
+      "learning_rate": 1.8429535425753145e-05,
+      "loss": 0.4482,
+      "step": 3942
+    },
+    {
+      "epoch": 0.24231809242871188,
+      "grad_norm": 23.625,
+      "learning_rate": 1.842828324640539e-05,
+      "loss": 0.5109,
+      "step": 3943
+    },
+    {
+      "epoch": 0.24237954768928222,
+      "grad_norm": 29.125,
+      "learning_rate": 1.842703061063063e-05,
+      "loss": 0.4073,
+      "step": 3944
+    },
+    {
+      "epoch": 0.24244100294985252,
+      "grad_norm": 22.25,
+      "learning_rate": 1.8425777518496707e-05,
+      "loss": 0.5443,
+      "step": 3945
+    },
+    {
+      "epoch": 0.24250245821042282,
+      "grad_norm": 25.25,
+      "learning_rate": 1.842452397007148e-05,
+      "loss": 0.4738,
+      "step": 3946
+    },
+    {
+      "epoch": 0.24256391347099313,
+      "grad_norm": 22.875,
+      "learning_rate": 1.8423269965422826e-05,
+      "loss": 0.3923,
+      "step": 3947
+    },
+    {
+      "epoch": 0.24262536873156343,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.8422015504618664e-05,
+      "loss": 0.3796,
+      "step": 3948
+    },
+    {
+      "epoch": 0.24268682399213373,
+      "grad_norm": 25.625,
+      "learning_rate": 1.8420760587726925e-05,
+      "loss": 0.6177,
+      "step": 3949
+    },
+    {
+      "epoch": 0.24274827925270404,
+      "grad_norm": 4.9375,
+      "learning_rate": 1.8419505214815567e-05,
+      "loss": 0.3117,
+      "step": 3950
+    },
+    {
+      "epoch": 0.24280973451327434,
+      "grad_norm": 37.75,
+      "learning_rate": 1.8418249385952575e-05,
+      "loss": 0.3569,
+      "step": 3951
+    },
+    {
+      "epoch": 0.24287118977384464,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.8416993101205957e-05,
+      "loss": 0.4344,
+      "step": 3952
+    },
+    {
+      "epoch": 0.24293264503441495,
+      "grad_norm": 22.25,
+      "learning_rate": 1.841573636064375e-05,
+      "loss": 0.4031,
+      "step": 3953
+    },
+    {
+      "epoch": 0.24299410029498525,
+      "grad_norm": 31.875,
+      "learning_rate": 1.8414479164334004e-05,
+      "loss": 0.4428,
+      "step": 3954
+    },
+    {
+      "epoch": 0.24305555555555555,
+      "grad_norm": 16.375,
+      "learning_rate": 1.8413221512344805e-05,
+      "loss": 0.3358,
+      "step": 3955
+    },
+    {
+      "epoch": 0.24311701081612586,
+      "grad_norm": 4.90625,
+      "learning_rate": 1.8411963404744263e-05,
+      "loss": 0.308,
+      "step": 3956
+    },
+    {
+      "epoch": 0.24317846607669616,
+      "grad_norm": 18.0,
+      "learning_rate": 1.8410704841600506e-05,
+      "loss": 0.301,
+      "step": 3957
+    },
+    {
+      "epoch": 0.24323992133726646,
+      "grad_norm": 3.640625,
+      "learning_rate": 1.8409445822981694e-05,
+      "loss": 0.4239,
+      "step": 3958
+    },
+    {
+      "epoch": 0.24330137659783677,
+      "grad_norm": 25.875,
+      "learning_rate": 1.8408186348956e-05,
+      "loss": 0.3616,
+      "step": 3959
+    },
+    {
+      "epoch": 0.24336283185840707,
+      "grad_norm": 6.3125,
+      "learning_rate": 1.8406926419591644e-05,
+      "loss": 0.3402,
+      "step": 3960
+    },
+    {
+      "epoch": 0.24342428711897737,
+      "grad_norm": 13.25,
+      "learning_rate": 1.8405666034956842e-05,
+      "loss": 0.326,
+      "step": 3961
+    },
+    {
+      "epoch": 0.24348574237954768,
+      "grad_norm": 17.25,
+      "learning_rate": 1.840440519511986e-05,
+      "loss": 0.4571,
+      "step": 3962
+    },
+    {
+      "epoch": 0.243547197640118,
+      "grad_norm": 17.25,
+      "learning_rate": 1.840314390014897e-05,
+      "loss": 0.6504,
+      "step": 3963
+    },
+    {
+      "epoch": 0.2436086529006883,
+      "grad_norm": 21.625,
+      "learning_rate": 1.8401882150112485e-05,
+      "loss": 0.5625,
+      "step": 3964
+    },
+    {
+      "epoch": 0.24367010816125861,
+      "grad_norm": 7.21875,
+      "learning_rate": 1.8400619945078725e-05,
+      "loss": 0.3691,
+      "step": 3965
+    },
+    {
+      "epoch": 0.24373156342182892,
+      "grad_norm": 25.625,
+      "learning_rate": 1.8399357285116045e-05,
+      "loss": 0.3519,
+      "step": 3966
+    },
+    {
+      "epoch": 0.24379301868239922,
+      "grad_norm": 6.3125,
+      "learning_rate": 1.839809417029283e-05,
+      "loss": 0.331,
+      "step": 3967
+    },
+    {
+      "epoch": 0.24385447394296952,
+      "grad_norm": 8.0,
+      "learning_rate": 1.839683060067748e-05,
+      "loss": 0.2673,
+      "step": 3968
+    },
+    {
+      "epoch": 0.24391592920353983,
+      "grad_norm": 34.5,
+      "learning_rate": 1.8395566576338416e-05,
+      "loss": 0.4243,
+      "step": 3969
+    },
+    {
+      "epoch": 0.24397738446411013,
+      "grad_norm": 41.25,
+      "learning_rate": 1.8394302097344103e-05,
+      "loss": 0.5283,
+      "step": 3970
+    },
+    {
+      "epoch": 0.24403883972468043,
+      "grad_norm": 37.5,
+      "learning_rate": 1.8393037163763005e-05,
+      "loss": 0.4097,
+      "step": 3971
+    },
+    {
+      "epoch": 0.24410029498525074,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.8391771775663628e-05,
+      "loss": 0.3138,
+      "step": 3972
+    },
+    {
+      "epoch": 0.24416175024582104,
+      "grad_norm": 53.75,
+      "learning_rate": 1.8390505933114503e-05,
+      "loss": 0.4512,
+      "step": 3973
+    },
+    {
+      "epoch": 0.24422320550639134,
+      "grad_norm": 6.34375,
+      "learning_rate": 1.8389239636184177e-05,
+      "loss": 0.3777,
+      "step": 3974
+    },
+    {
+      "epoch": 0.24428466076696165,
+      "grad_norm": 5.6875,
+      "learning_rate": 1.838797288494122e-05,
+      "loss": 0.5333,
+      "step": 3975
+    },
+    {
+      "epoch": 0.24434611602753195,
+      "grad_norm": 14.75,
+      "learning_rate": 1.8386705679454243e-05,
+      "loss": 0.5531,
+      "step": 3976
+    },
+    {
+      "epoch": 0.24440757128810225,
+      "grad_norm": 43.25,
+      "learning_rate": 1.8385438019791862e-05,
+      "loss": 0.4105,
+      "step": 3977
+    },
+    {
+      "epoch": 0.24446902654867256,
+      "grad_norm": 7.71875,
+      "learning_rate": 1.838416990602273e-05,
+      "loss": 0.3496,
+      "step": 3978
+    },
+    {
+      "epoch": 0.24453048180924286,
+      "grad_norm": 38.5,
+      "learning_rate": 1.8382901338215515e-05,
+      "loss": 0.3609,
+      "step": 3979
+    },
+    {
+      "epoch": 0.24459193706981316,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.8381632316438923e-05,
+      "loss": 0.2488,
+      "step": 3980
+    },
+    {
+      "epoch": 0.24465339233038347,
+      "grad_norm": 4.90625,
+      "learning_rate": 1.8380362840761675e-05,
+      "loss": 0.3516,
+      "step": 3981
+    },
+    {
+      "epoch": 0.2447148475909538,
+      "grad_norm": 18.25,
+      "learning_rate": 1.8379092911252515e-05,
+      "loss": 0.334,
+      "step": 3982
+    },
+    {
+      "epoch": 0.2447763028515241,
+      "grad_norm": 8.125,
+      "learning_rate": 1.8377822527980216e-05,
+      "loss": 0.2432,
+      "step": 3983
+    },
+    {
+      "epoch": 0.2448377581120944,
+      "grad_norm": 21.875,
+      "learning_rate": 1.8376551691013577e-05,
+      "loss": 0.473,
+      "step": 3984
+    },
+    {
+      "epoch": 0.2448992133726647,
+      "grad_norm": 4.21875,
+      "learning_rate": 1.837528040042142e-05,
+      "loss": 0.2835,
+      "step": 3985
+    },
+    {
+      "epoch": 0.244960668633235,
+      "grad_norm": 7.0625,
+      "learning_rate": 1.8374008656272585e-05,
+      "loss": 0.3384,
+      "step": 3986
+    },
+    {
+      "epoch": 0.24502212389380532,
+      "grad_norm": 20.75,
+      "learning_rate": 1.8372736458635947e-05,
+      "loss": 0.442,
+      "step": 3987
+    },
+    {
+      "epoch": 0.24508357915437562,
+      "grad_norm": 41.5,
+      "learning_rate": 1.83714638075804e-05,
+      "loss": 0.5617,
+      "step": 3988
+    },
+    {
+      "epoch": 0.24514503441494592,
+      "grad_norm": 25.25,
+      "learning_rate": 1.8370190703174862e-05,
+      "loss": 0.326,
+      "step": 3989
+    },
+    {
+      "epoch": 0.24520648967551623,
+      "grad_norm": 44.5,
+      "learning_rate": 1.836891714548828e-05,
+      "loss": 0.3708,
+      "step": 3990
+    },
+    {
+      "epoch": 0.24526794493608653,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.836764313458962e-05,
+      "loss": 0.322,
+      "step": 3991
+    },
+    {
+      "epoch": 0.24532940019665683,
+      "grad_norm": 12.0625,
+      "learning_rate": 1.8366368670547873e-05,
+      "loss": 0.3048,
+      "step": 3992
+    },
+    {
+      "epoch": 0.24539085545722714,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.836509375343206e-05,
+      "loss": 0.3057,
+      "step": 3993
+    },
+    {
+      "epoch": 0.24545231071779744,
+      "grad_norm": 3.578125,
+      "learning_rate": 1.8363818383311226e-05,
+      "loss": 0.3365,
+      "step": 3994
+    },
+    {
+      "epoch": 0.24551376597836774,
+      "grad_norm": 20.5,
+      "learning_rate": 1.836254256025443e-05,
+      "loss": 0.4896,
+      "step": 3995
+    },
+    {
+      "epoch": 0.24557522123893805,
+      "grad_norm": 5.90625,
+      "learning_rate": 1.836126628433077e-05,
+      "loss": 0.2832,
+      "step": 3996
+    },
+    {
+      "epoch": 0.24563667649950835,
+      "grad_norm": 17.25,
+      "learning_rate": 1.8359989555609355e-05,
+      "loss": 0.3137,
+      "step": 3997
+    },
+    {
+      "epoch": 0.24569813176007865,
+      "grad_norm": 19.5,
+      "learning_rate": 1.8358712374159332e-05,
+      "loss": 0.2704,
+      "step": 3998
+    },
+    {
+      "epoch": 0.24575958702064896,
+      "grad_norm": 24.0,
+      "learning_rate": 1.835743474004986e-05,
+      "loss": 0.5175,
+      "step": 3999
+    },
+    {
+      "epoch": 0.24582104228121926,
+      "grad_norm": 5.125,
+      "learning_rate": 1.8356156653350138e-05,
+      "loss": 0.2654,
+      "step": 4000
+    },
+    {
+      "epoch": 0.2458824975417896,
+      "grad_norm": 25.0,
+      "learning_rate": 1.8354878114129368e-05,
+      "loss": 0.4489,
+      "step": 4001
+    },
+    {
+      "epoch": 0.2459439528023599,
+      "grad_norm": 26.5,
+      "learning_rate": 1.835359912245679e-05,
+      "loss": 0.5935,
+      "step": 4002
+    },
+    {
+      "epoch": 0.2460054080629302,
+      "grad_norm": 53.25,
+      "learning_rate": 1.8352319678401677e-05,
+      "loss": 0.5617,
+      "step": 4003
+    },
+    {
+      "epoch": 0.2460668633235005,
+      "grad_norm": 29.625,
+      "learning_rate": 1.8351039782033303e-05,
+      "loss": 0.5743,
+      "step": 4004
+    },
+    {
+      "epoch": 0.2461283185840708,
+      "grad_norm": 30.0,
+      "learning_rate": 1.834975943342099e-05,
+      "loss": 0.4428,
+      "step": 4005
+    },
+    {
+      "epoch": 0.2461897738446411,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.8348478632634067e-05,
+      "loss": 0.3116,
+      "step": 4006
+    },
+    {
+      "epoch": 0.2462512291052114,
+      "grad_norm": 24.75,
+      "learning_rate": 1.83471973797419e-05,
+      "loss": 0.3855,
+      "step": 4007
+    },
+    {
+      "epoch": 0.24631268436578171,
+      "grad_norm": 24.75,
+      "learning_rate": 1.834591567481387e-05,
+      "loss": 0.3111,
+      "step": 4008
+    },
+    {
+      "epoch": 0.24637413962635202,
+      "grad_norm": 19.5,
+      "learning_rate": 1.834463351791939e-05,
+      "loss": 0.3486,
+      "step": 4009
+    },
+    {
+      "epoch": 0.24643559488692232,
+      "grad_norm": 17.375,
+      "learning_rate": 1.8343350909127897e-05,
+      "loss": 0.3523,
+      "step": 4010
+    },
+    {
+      "epoch": 0.24649705014749262,
+      "grad_norm": 25.375,
+      "learning_rate": 1.8342067848508843e-05,
+      "loss": 0.3483,
+      "step": 4011
+    },
+    {
+      "epoch": 0.24655850540806293,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.8340784336131715e-05,
+      "loss": 0.313,
+      "step": 4012
+    },
+    {
+      "epoch": 0.24661996066863323,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8339500372066018e-05,
+      "loss": 0.2793,
+      "step": 4013
+    },
+    {
+      "epoch": 0.24668141592920353,
+      "grad_norm": 36.25,
+      "learning_rate": 1.8338215956381283e-05,
+      "loss": 0.3837,
+      "step": 4014
+    },
+    {
+      "epoch": 0.24674287118977384,
+      "grad_norm": 11.375,
+      "learning_rate": 1.8336931089147076e-05,
+      "loss": 0.2733,
+      "step": 4015
+    },
+    {
+      "epoch": 0.24680432645034414,
+      "grad_norm": 31.875,
+      "learning_rate": 1.8335645770432963e-05,
+      "loss": 0.3363,
+      "step": 4016
+    },
+    {
+      "epoch": 0.24686578171091444,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8334360000308563e-05,
+      "loss": 0.5316,
+      "step": 4017
+    },
+    {
+      "epoch": 0.24692723697148475,
+      "grad_norm": 4.0625,
+      "learning_rate": 1.83330737788435e-05,
+      "loss": 0.3169,
+      "step": 4018
+    },
+    {
+      "epoch": 0.24698869223205505,
+      "grad_norm": 3.5,
+      "learning_rate": 1.8331787106107425e-05,
+      "loss": 0.3207,
+      "step": 4019
+    },
+    {
+      "epoch": 0.24705014749262535,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.8330499982170024e-05,
+      "loss": 0.2851,
+      "step": 4020
+    },
+    {
+      "epoch": 0.24711160275319569,
+      "grad_norm": 3.5625,
+      "learning_rate": 1.8329212407100996e-05,
+      "loss": 0.3365,
+      "step": 4021
+    },
+    {
+      "epoch": 0.247173058013766,
+      "grad_norm": 5.125,
+      "learning_rate": 1.8327924380970066e-05,
+      "loss": 0.3067,
+      "step": 4022
+    },
+    {
+      "epoch": 0.2472345132743363,
+      "grad_norm": 3.96875,
+      "learning_rate": 1.8326635903846995e-05,
+      "loss": 0.3348,
+      "step": 4023
+    },
+    {
+      "epoch": 0.2472959685349066,
+      "grad_norm": 21.5,
+      "learning_rate": 1.832534697580155e-05,
+      "loss": 0.384,
+      "step": 4024
+    },
+    {
+      "epoch": 0.2473574237954769,
+      "grad_norm": 11.25,
+      "learning_rate": 1.8324057596903534e-05,
+      "loss": 0.3548,
+      "step": 4025
+    },
+    {
+      "epoch": 0.2474188790560472,
+      "grad_norm": 49.0,
+      "learning_rate": 1.832276776722278e-05,
+      "loss": 0.588,
+      "step": 4026
+    },
+    {
+      "epoch": 0.2474803343166175,
+      "grad_norm": 16.125,
+      "learning_rate": 1.8321477486829128e-05,
+      "loss": 0.3635,
+      "step": 4027
+    },
+    {
+      "epoch": 0.2475417895771878,
+      "grad_norm": 4.8125,
+      "learning_rate": 1.8320186755792452e-05,
+      "loss": 0.274,
+      "step": 4028
+    },
+    {
+      "epoch": 0.2476032448377581,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.8318895574182657e-05,
+      "loss": 0.3442,
+      "step": 4029
+    },
+    {
+      "epoch": 0.24766470009832842,
+      "grad_norm": 16.25,
+      "learning_rate": 1.8317603942069665e-05,
+      "loss": 0.3387,
+      "step": 4030
+    },
+    {
+      "epoch": 0.24772615535889872,
+      "grad_norm": 26.375,
+      "learning_rate": 1.831631185952342e-05,
+      "loss": 0.4472,
+      "step": 4031
+    },
+    {
+      "epoch": 0.24778761061946902,
+      "grad_norm": 28.125,
+      "learning_rate": 1.8315019326613897e-05,
+      "loss": 0.4723,
+      "step": 4032
+    },
+    {
+      "epoch": 0.24784906588003933,
+      "grad_norm": 24.25,
+      "learning_rate": 1.8313726343411085e-05,
+      "loss": 0.5292,
+      "step": 4033
+    },
+    {
+      "epoch": 0.24791052114060963,
+      "grad_norm": 38.0,
+      "learning_rate": 1.8312432909985017e-05,
+      "loss": 0.4845,
+      "step": 4034
+    },
+    {
+      "epoch": 0.24797197640117993,
+      "grad_norm": 24.375,
+      "learning_rate": 1.8311139026405725e-05,
+      "loss": 0.3009,
+      "step": 4035
+    },
+    {
+      "epoch": 0.24803343166175024,
+      "grad_norm": 34.75,
+      "learning_rate": 1.8309844692743283e-05,
+      "loss": 0.4395,
+      "step": 4036
+    },
+    {
+      "epoch": 0.24809488692232054,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.830854990906779e-05,
+      "loss": 0.3267,
+      "step": 4037
+    },
+    {
+      "epoch": 0.24815634218289084,
+      "grad_norm": 6.15625,
+      "learning_rate": 1.8307254675449355e-05,
+      "loss": 0.394,
+      "step": 4038
+    },
+    {
+      "epoch": 0.24821779744346115,
+      "grad_norm": 18.0,
+      "learning_rate": 1.830595899195813e-05,
+      "loss": 0.2889,
+      "step": 4039
+    },
+    {
+      "epoch": 0.24827925270403148,
+      "grad_norm": 21.5,
+      "learning_rate": 1.8304662858664272e-05,
+      "loss": 0.3714,
+      "step": 4040
+    },
+    {
+      "epoch": 0.24834070796460178,
+      "grad_norm": 37.75,
+      "learning_rate": 1.8303366275637977e-05,
+      "loss": 0.3317,
+      "step": 4041
+    },
+    {
+      "epoch": 0.24840216322517208,
+      "grad_norm": 24.625,
+      "learning_rate": 1.830206924294946e-05,
+      "loss": 0.378,
+      "step": 4042
+    },
+    {
+      "epoch": 0.2484636184857424,
+      "grad_norm": 6.65625,
+      "learning_rate": 1.830077176066896e-05,
+      "loss": 0.3003,
+      "step": 4043
+    },
+    {
+      "epoch": 0.2485250737463127,
+      "grad_norm": 25.25,
+      "learning_rate": 1.8299473828866746e-05,
+      "loss": 0.3986,
+      "step": 4044
+    },
+    {
+      "epoch": 0.248586529006883,
+      "grad_norm": 21.5,
+      "learning_rate": 1.82981754476131e-05,
+      "loss": 0.6326,
+      "step": 4045
+    },
+    {
+      "epoch": 0.2486479842674533,
+      "grad_norm": 24.75,
+      "learning_rate": 1.8296876616978337e-05,
+      "loss": 0.4619,
+      "step": 4046
+    },
+    {
+      "epoch": 0.2487094395280236,
+      "grad_norm": 13.0,
+      "learning_rate": 1.8295577337032792e-05,
+      "loss": 0.3393,
+      "step": 4047
+    },
+    {
+      "epoch": 0.2487708947885939,
+      "grad_norm": 6.6875,
+      "learning_rate": 1.8294277607846834e-05,
+      "loss": 0.4538,
+      "step": 4048
+    },
+    {
+      "epoch": 0.2488323500491642,
+      "grad_norm": 31.5,
+      "learning_rate": 1.829297742949084e-05,
+      "loss": 0.6041,
+      "step": 4049
+    },
+    {
+      "epoch": 0.2488938053097345,
+      "grad_norm": 20.0,
+      "learning_rate": 1.829167680203522e-05,
+      "loss": 0.4036,
+      "step": 4050
+    },
+    {
+      "epoch": 0.24895526057030481,
+      "grad_norm": 19.75,
+      "learning_rate": 1.8290375725550417e-05,
+      "loss": 0.466,
+      "step": 4051
+    },
+    {
+      "epoch": 0.24901671583087512,
+      "grad_norm": 23.125,
+      "learning_rate": 1.8289074200106885e-05,
+      "loss": 0.4013,
+      "step": 4052
+    },
+    {
+      "epoch": 0.24907817109144542,
+      "grad_norm": 8.375,
+      "learning_rate": 1.8287772225775105e-05,
+      "loss": 0.3244,
+      "step": 4053
+    },
+    {
+      "epoch": 0.24913962635201573,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.828646980262559e-05,
+      "loss": 0.3141,
+      "step": 4054
+    },
+    {
+      "epoch": 0.24920108161258603,
+      "grad_norm": 17.375,
+      "learning_rate": 1.8285166930728865e-05,
+      "loss": 0.3037,
+      "step": 4055
+    },
+    {
+      "epoch": 0.24926253687315633,
+      "grad_norm": 20.875,
+      "learning_rate": 1.828386361015549e-05,
+      "loss": 0.3363,
+      "step": 4056
+    },
+    {
+      "epoch": 0.24932399213372664,
+      "grad_norm": 24.125,
+      "learning_rate": 1.8282559840976043e-05,
+      "loss": 0.498,
+      "step": 4057
+    },
+    {
+      "epoch": 0.24938544739429694,
+      "grad_norm": 6.84375,
+      "learning_rate": 1.8281255623261128e-05,
+      "loss": 0.4123,
+      "step": 4058
+    },
+    {
+      "epoch": 0.24944690265486727,
+      "grad_norm": 16.25,
+      "learning_rate": 1.827995095708138e-05,
+      "loss": 0.312,
+      "step": 4059
+    },
+    {
+      "epoch": 0.24950835791543757,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.8278645842507448e-05,
+      "loss": 0.4186,
+      "step": 4060
+    },
+    {
+      "epoch": 0.24956981317600788,
+      "grad_norm": 9.875,
+      "learning_rate": 1.827734027961001e-05,
+      "loss": 0.5515,
+      "step": 4061
+    },
+    {
+      "epoch": 0.24963126843657818,
+      "grad_norm": 29.875,
+      "learning_rate": 1.8276034268459763e-05,
+      "loss": 0.3916,
+      "step": 4062
+    },
+    {
+      "epoch": 0.24969272369714848,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.827472780912744e-05,
+      "loss": 0.4083,
+      "step": 4063
+    },
+    {
+      "epoch": 0.2497541789577188,
+      "grad_norm": 14.875,
+      "learning_rate": 1.8273420901683786e-05,
+      "loss": 0.4185,
+      "step": 4064
+    },
+    {
+      "epoch": 0.2498156342182891,
+      "grad_norm": 21.625,
+      "learning_rate": 1.827211354619958e-05,
+      "loss": 0.3801,
+      "step": 4065
+    },
+    {
+      "epoch": 0.2498770894788594,
+      "grad_norm": 32.0,
+      "learning_rate": 1.827080574274562e-05,
+      "loss": 0.6807,
+      "step": 4066
+    },
+    {
+      "epoch": 0.2499385447394297,
+      "grad_norm": 28.875,
+      "learning_rate": 1.8269497491392724e-05,
+      "loss": 0.4956,
+      "step": 4067
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8268188792211748e-05,
+      "loss": 0.6116,
+      "step": 4068
+    },
+    {
+      "epoch": 0.25006145526057033,
+      "grad_norm": 17.25,
+      "learning_rate": 1.8266879645273557e-05,
+      "loss": 0.4544,
+      "step": 4069
+    },
+    {
+      "epoch": 0.2501229105211406,
+      "grad_norm": 21.5,
+      "learning_rate": 1.8265570050649048e-05,
+      "loss": 0.4382,
+      "step": 4070
+    },
+    {
+      "epoch": 0.25018436578171094,
+      "grad_norm": 34.0,
+      "learning_rate": 1.8264260008409138e-05,
+      "loss": 0.3984,
+      "step": 4071
+    },
+    {
+      "epoch": 0.2502458210422812,
+      "grad_norm": 21.875,
+      "learning_rate": 1.826294951862478e-05,
+      "loss": 0.367,
+      "step": 4072
+    },
+    {
+      "epoch": 0.25030727630285154,
+      "grad_norm": 31.5,
+      "learning_rate": 1.8261638581366937e-05,
+      "loss": 0.4182,
+      "step": 4073
+    },
+    {
+      "epoch": 0.2503687315634218,
+      "grad_norm": 21.0,
+      "learning_rate": 1.8260327196706596e-05,
+      "loss": 0.4689,
+      "step": 4074
+    },
+    {
+      "epoch": 0.25043018682399215,
+      "grad_norm": 31.625,
+      "learning_rate": 1.8259015364714786e-05,
+      "loss": 0.4647,
+      "step": 4075
+    },
+    {
+      "epoch": 0.2504916420845624,
+      "grad_norm": 11.625,
+      "learning_rate": 1.8257703085462542e-05,
+      "loss": 0.3539,
+      "step": 4076
+    },
+    {
+      "epoch": 0.25055309734513276,
+      "grad_norm": 23.25,
+      "learning_rate": 1.825639035902093e-05,
+      "loss": 0.3953,
+      "step": 4077
+    },
+    {
+      "epoch": 0.25061455260570303,
+      "grad_norm": 20.875,
+      "learning_rate": 1.825507718546104e-05,
+      "loss": 0.4242,
+      "step": 4078
+    },
+    {
+      "epoch": 0.25067600786627336,
+      "grad_norm": 28.125,
+      "learning_rate": 1.825376356485398e-05,
+      "loss": 0.4373,
+      "step": 4079
+    },
+    {
+      "epoch": 0.25073746312684364,
+      "grad_norm": 6.90625,
+      "learning_rate": 1.82524494972709e-05,
+      "loss": 0.3859,
+      "step": 4080
+    },
+    {
+      "epoch": 0.25079891838741397,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.8251134982782952e-05,
+      "loss": 0.3157,
+      "step": 4081
+    },
+    {
+      "epoch": 0.25086037364798425,
+      "grad_norm": 16.5,
+      "learning_rate": 1.824982002146133e-05,
+      "loss": 0.4286,
+      "step": 4082
+    },
+    {
+      "epoch": 0.2509218289085546,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.8248504613377236e-05,
+      "loss": 0.3369,
+      "step": 4083
+    },
+    {
+      "epoch": 0.25098328416912485,
+      "grad_norm": 29.5,
+      "learning_rate": 1.8247188758601912e-05,
+      "loss": 0.349,
+      "step": 4084
+    },
+    {
+      "epoch": 0.2510447394296952,
+      "grad_norm": 22.0,
+      "learning_rate": 1.8245872457206614e-05,
+      "loss": 0.2813,
+      "step": 4085
+    },
+    {
+      "epoch": 0.25110619469026546,
+      "grad_norm": 15.3125,
+      "learning_rate": 1.8244555709262627e-05,
+      "loss": 0.3769,
+      "step": 4086
+    },
+    {
+      "epoch": 0.2511676499508358,
+      "grad_norm": 33.5,
+      "learning_rate": 1.824323851484126e-05,
+      "loss": 0.395,
+      "step": 4087
+    },
+    {
+      "epoch": 0.2512291052114061,
+      "grad_norm": 15.6875,
+      "learning_rate": 1.824192087401384e-05,
+      "loss": 0.2735,
+      "step": 4088
+    },
+    {
+      "epoch": 0.2512905604719764,
+      "grad_norm": 22.875,
+      "learning_rate": 1.8240602786851726e-05,
+      "loss": 0.4715,
+      "step": 4089
+    },
+    {
+      "epoch": 0.25135201573254673,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8239284253426294e-05,
+      "loss": 0.438,
+      "step": 4090
+    },
+    {
+      "epoch": 0.251413470993117,
+      "grad_norm": 28.5,
+      "learning_rate": 1.823796527380895e-05,
+      "loss": 0.4423,
+      "step": 4091
+    },
+    {
+      "epoch": 0.25147492625368734,
+      "grad_norm": 22.75,
+      "learning_rate": 1.823664584807113e-05,
+      "loss": 0.377,
+      "step": 4092
+    },
+    {
+      "epoch": 0.2515363815142576,
+      "grad_norm": 13.125,
+      "learning_rate": 1.8235325976284276e-05,
+      "loss": 0.3212,
+      "step": 4093
+    },
+    {
+      "epoch": 0.25159783677482794,
+      "grad_norm": 12.75,
+      "learning_rate": 1.823400565851987e-05,
+      "loss": 0.3056,
+      "step": 4094
+    },
+    {
+      "epoch": 0.2516592920353982,
+      "grad_norm": 11.25,
+      "learning_rate": 1.823268489484941e-05,
+      "loss": 0.3318,
+      "step": 4095
+    },
+    {
+      "epoch": 0.25172074729596855,
+      "grad_norm": 19.0,
+      "learning_rate": 1.8231363685344422e-05,
+      "loss": 0.3109,
+      "step": 4096
+    },
+    {
+      "epoch": 0.2517822025565388,
+      "grad_norm": 60.75,
+      "learning_rate": 1.8230042030076454e-05,
+      "loss": 0.4796,
+      "step": 4097
+    },
+    {
+      "epoch": 0.25184365781710916,
+      "grad_norm": 43.75,
+      "learning_rate": 1.8228719929117083e-05,
+      "loss": 0.4281,
+      "step": 4098
+    },
+    {
+      "epoch": 0.25190511307767943,
+      "grad_norm": 7.5,
+      "learning_rate": 1.82273973825379e-05,
+      "loss": 0.3413,
+      "step": 4099
+    },
+    {
+      "epoch": 0.25196656833824976,
+      "grad_norm": 20.125,
+      "learning_rate": 1.8226074390410534e-05,
+      "loss": 0.355,
+      "step": 4100
+    },
+    {
+      "epoch": 0.25202802359882004,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.8224750952806626e-05,
+      "loss": 0.3679,
+      "step": 4101
+    },
+    {
+      "epoch": 0.25208947885939037,
+      "grad_norm": 31.0,
+      "learning_rate": 1.8223427069797845e-05,
+      "loss": 0.4405,
+      "step": 4102
+    },
+    {
+      "epoch": 0.25215093411996065,
+      "grad_norm": 6.3125,
+      "learning_rate": 1.8222102741455886e-05,
+      "loss": 0.3148,
+      "step": 4103
+    },
+    {
+      "epoch": 0.252212389380531,
+      "grad_norm": 9.625,
+      "learning_rate": 1.8220777967852467e-05,
+      "loss": 0.307,
+      "step": 4104
+    },
+    {
+      "epoch": 0.25227384464110125,
+      "grad_norm": 18.25,
+      "learning_rate": 1.8219452749059332e-05,
+      "loss": 0.6263,
+      "step": 4105
+    },
+    {
+      "epoch": 0.2523352999016716,
+      "grad_norm": 11.25,
+      "learning_rate": 1.8218127085148246e-05,
+      "loss": 0.4393,
+      "step": 4106
+    },
+    {
+      "epoch": 0.2523967551622419,
+      "grad_norm": 25.0,
+      "learning_rate": 1.8216800976191e-05,
+      "loss": 0.5935,
+      "step": 4107
+    },
+    {
+      "epoch": 0.2524582104228122,
+      "grad_norm": 36.0,
+      "learning_rate": 1.8215474422259403e-05,
+      "loss": 0.4726,
+      "step": 4108
+    },
+    {
+      "epoch": 0.2525196656833825,
+      "grad_norm": 18.875,
+      "learning_rate": 1.82141474234253e-05,
+      "loss": 0.4493,
+      "step": 4109
+    },
+    {
+      "epoch": 0.2525811209439528,
+      "grad_norm": 18.5,
+      "learning_rate": 1.821281997976055e-05,
+      "loss": 0.4508,
+      "step": 4110
+    },
+    {
+      "epoch": 0.25264257620452313,
+      "grad_norm": 16.375,
+      "learning_rate": 1.821149209133704e-05,
+      "loss": 0.4377,
+      "step": 4111
+    },
+    {
+      "epoch": 0.2527040314650934,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.8210163758226685e-05,
+      "loss": 0.3957,
+      "step": 4112
+    },
+    {
+      "epoch": 0.25276548672566373,
+      "grad_norm": 19.625,
+      "learning_rate": 1.8208834980501417e-05,
+      "loss": 0.4855,
+      "step": 4113
+    },
+    {
+      "epoch": 0.252826941986234,
+      "grad_norm": 21.125,
+      "learning_rate": 1.820750575823319e-05,
+      "loss": 0.4922,
+      "step": 4114
+    },
+    {
+      "epoch": 0.25288839724680434,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.8206176091493995e-05,
+      "loss": 0.5336,
+      "step": 4115
+    },
+    {
+      "epoch": 0.2529498525073746,
+      "grad_norm": 7.96875,
+      "learning_rate": 1.8204845980355834e-05,
+      "loss": 0.4201,
+      "step": 4116
+    },
+    {
+      "epoch": 0.25301130776794495,
+      "grad_norm": 33.75,
+      "learning_rate": 1.8203515424890738e-05,
+      "loss": 0.5766,
+      "step": 4117
+    },
+    {
+      "epoch": 0.2530727630285152,
+      "grad_norm": 22.875,
+      "learning_rate": 1.8202184425170766e-05,
+      "loss": 0.5656,
+      "step": 4118
+    },
+    {
+      "epoch": 0.25313421828908556,
+      "grad_norm": 20.375,
+      "learning_rate": 1.820085298126799e-05,
+      "loss": 0.5482,
+      "step": 4119
+    },
+    {
+      "epoch": 0.25319567354965583,
+      "grad_norm": 21.75,
+      "learning_rate": 1.8199521093254524e-05,
+      "loss": 0.5552,
+      "step": 4120
+    },
+    {
+      "epoch": 0.25325712881022616,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8198188761202487e-05,
+      "loss": 0.5031,
+      "step": 4121
+    },
+    {
+      "epoch": 0.25331858407079644,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.8196855985184033e-05,
+      "loss": 0.4674,
+      "step": 4122
+    },
+    {
+      "epoch": 0.25338003933136677,
+      "grad_norm": 32.25,
+      "learning_rate": 1.819552276527134e-05,
+      "loss": 0.4217,
+      "step": 4123
+    },
+    {
+      "epoch": 0.25344149459193704,
+      "grad_norm": 15.75,
+      "learning_rate": 1.8194189101536603e-05,
+      "loss": 0.4838,
+      "step": 4124
+    },
+    {
+      "epoch": 0.2535029498525074,
+      "grad_norm": 10.5,
+      "learning_rate": 1.8192854994052047e-05,
+      "loss": 0.3384,
+      "step": 4125
+    },
+    {
+      "epoch": 0.2535644051130777,
+      "grad_norm": 12.75,
+      "learning_rate": 1.819152044288992e-05,
+      "loss": 0.4161,
+      "step": 4126
+    },
+    {
+      "epoch": 0.253625860373648,
+      "grad_norm": 18.125,
+      "learning_rate": 1.8190185448122496e-05,
+      "loss": 0.4645,
+      "step": 4127
+    },
+    {
+      "epoch": 0.2536873156342183,
+      "grad_norm": 13.0625,
+      "learning_rate": 1.8188850009822063e-05,
+      "loss": 0.3758,
+      "step": 4128
+    },
+    {
+      "epoch": 0.2537487708947886,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.8187514128060946e-05,
+      "loss": 0.6049,
+      "step": 4129
+    },
+    {
+      "epoch": 0.2538102261553589,
+      "grad_norm": 23.125,
+      "learning_rate": 1.818617780291149e-05,
+      "loss": 0.4706,
+      "step": 4130
+    },
+    {
+      "epoch": 0.2538716814159292,
+      "grad_norm": 18.5,
+      "learning_rate": 1.818484103444606e-05,
+      "loss": 0.6852,
+      "step": 4131
+    },
+    {
+      "epoch": 0.2539331366764995,
+      "grad_norm": 13.625,
+      "learning_rate": 1.818350382273705e-05,
+      "loss": 0.4768,
+      "step": 4132
+    },
+    {
+      "epoch": 0.2539945919370698,
+      "grad_norm": 22.5,
+      "learning_rate": 1.8182166167856875e-05,
+      "loss": 0.4391,
+      "step": 4133
+    },
+    {
+      "epoch": 0.25405604719764013,
+      "grad_norm": 15.125,
+      "learning_rate": 1.8180828069877972e-05,
+      "loss": 0.4925,
+      "step": 4134
+    },
+    {
+      "epoch": 0.2541175024582104,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.8179489528872808e-05,
+      "loss": 0.4346,
+      "step": 4135
+    },
+    {
+      "epoch": 0.25417895771878074,
+      "grad_norm": 5.96875,
+      "learning_rate": 1.8178150544913867e-05,
+      "loss": 0.3729,
+      "step": 4136
+    },
+    {
+      "epoch": 0.254240412979351,
+      "grad_norm": 23.625,
+      "learning_rate": 1.8176811118073664e-05,
+      "loss": 0.4421,
+      "step": 4137
+    },
+    {
+      "epoch": 0.25430186823992135,
+      "grad_norm": 19.625,
+      "learning_rate": 1.817547124842473e-05,
+      "loss": 0.69,
+      "step": 4138
+    },
+    {
+      "epoch": 0.2543633235004916,
+      "grad_norm": 27.125,
+      "learning_rate": 1.817413093603963e-05,
+      "loss": 0.5077,
+      "step": 4139
+    },
+    {
+      "epoch": 0.25442477876106195,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.817279018099095e-05,
+      "loss": 0.3464,
+      "step": 4140
+    },
+    {
+      "epoch": 0.25448623402163223,
+      "grad_norm": 23.625,
+      "learning_rate": 1.8171448983351284e-05,
+      "loss": 0.4602,
+      "step": 4141
+    },
+    {
+      "epoch": 0.25454768928220256,
+      "grad_norm": 9.6875,
+      "learning_rate": 1.817010734319328e-05,
+      "loss": 0.3403,
+      "step": 4142
+    },
+    {
+      "epoch": 0.25460914454277284,
+      "grad_norm": 20.5,
+      "learning_rate": 1.8168765260589583e-05,
+      "loss": 0.5052,
+      "step": 4143
+    },
+    {
+      "epoch": 0.25467059980334317,
+      "grad_norm": 12.625,
+      "learning_rate": 1.8167422735612877e-05,
+      "loss": 0.423,
+      "step": 4144
+    },
+    {
+      "epoch": 0.2547320550639135,
+      "grad_norm": 12.3125,
+      "learning_rate": 1.816607976833586e-05,
+      "loss": 0.363,
+      "step": 4145
+    },
+    {
+      "epoch": 0.2547935103244838,
+      "grad_norm": 5.8125,
+      "learning_rate": 1.8164736358831265e-05,
+      "loss": 0.4327,
+      "step": 4146
+    },
+    {
+      "epoch": 0.2548549655850541,
+      "grad_norm": 9.4375,
+      "learning_rate": 1.816339250717184e-05,
+      "loss": 0.3551,
+      "step": 4147
+    },
+    {
+      "epoch": 0.2549164208456244,
+      "grad_norm": 21.25,
+      "learning_rate": 1.8162048213430362e-05,
+      "loss": 0.3017,
+      "step": 4148
+    },
+    {
+      "epoch": 0.2549778761061947,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.816070347767963e-05,
+      "loss": 0.2681,
+      "step": 4149
+    },
+    {
+      "epoch": 0.255039331366765,
+      "grad_norm": 9.8125,
+      "learning_rate": 1.815935829999247e-05,
+      "loss": 0.2206,
+      "step": 4150
+    },
+    {
+      "epoch": 0.2551007866273353,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.8158012680441723e-05,
+      "loss": 0.275,
+      "step": 4151
+    },
+    {
+      "epoch": 0.2551622418879056,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.8156666619100263e-05,
+      "loss": 0.2664,
+      "step": 4152
+    },
+    {
+      "epoch": 0.2552236971484759,
+      "grad_norm": 10.4375,
+      "learning_rate": 1.8155320116040983e-05,
+      "loss": 0.3529,
+      "step": 4153
+    },
+    {
+      "epoch": 0.2552851524090462,
+      "grad_norm": 8.6875,
+      "learning_rate": 1.8153973171336804e-05,
+      "loss": 0.5122,
+      "step": 4154
+    },
+    {
+      "epoch": 0.25534660766961653,
+      "grad_norm": 20.5,
+      "learning_rate": 1.815262578506067e-05,
+      "loss": 0.3103,
+      "step": 4155
+    },
+    {
+      "epoch": 0.2554080629301868,
+      "grad_norm": 21.375,
+      "learning_rate": 1.815127795728554e-05,
+      "loss": 0.5232,
+      "step": 4156
+    },
+    {
+      "epoch": 0.25546951819075714,
+      "grad_norm": 18.625,
+      "learning_rate": 1.814992968808442e-05,
+      "loss": 0.3637,
+      "step": 4157
+    },
+    {
+      "epoch": 0.2555309734513274,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.814858097753031e-05,
+      "loss": 0.2555,
+      "step": 4158
+    },
+    {
+      "epoch": 0.25559242871189775,
+      "grad_norm": 4.125,
+      "learning_rate": 1.814723182569625e-05,
+      "loss": 0.3357,
+      "step": 4159
+    },
+    {
+      "epoch": 0.255653883972468,
+      "grad_norm": 28.625,
+      "learning_rate": 1.8145882232655308e-05,
+      "loss": 0.3667,
+      "step": 4160
+    },
+    {
+      "epoch": 0.25571533923303835,
+      "grad_norm": 7.34375,
+      "learning_rate": 1.814453219848057e-05,
+      "loss": 0.3146,
+      "step": 4161
+    },
+    {
+      "epoch": 0.25577679449360863,
+      "grad_norm": 12.75,
+      "learning_rate": 1.814318172324514e-05,
+      "loss": 0.4413,
+      "step": 4162
+    },
+    {
+      "epoch": 0.25583824975417896,
+      "grad_norm": 19.25,
+      "learning_rate": 1.814183080702216e-05,
+      "loss": 0.4508,
+      "step": 4163
+    },
+    {
+      "epoch": 0.2558997050147493,
+      "grad_norm": 20.625,
+      "learning_rate": 1.8140479449884778e-05,
+      "loss": 0.4866,
+      "step": 4164
+    },
+    {
+      "epoch": 0.25596116027531957,
+      "grad_norm": 52.25,
+      "learning_rate": 1.8139127651906183e-05,
+      "loss": 0.3641,
+      "step": 4165
+    },
+    {
+      "epoch": 0.2560226155358899,
+      "grad_norm": 22.625,
+      "learning_rate": 1.813777541315958e-05,
+      "loss": 0.334,
+      "step": 4166
+    },
+    {
+      "epoch": 0.2560840707964602,
+      "grad_norm": 14.9375,
+      "learning_rate": 1.8136422733718192e-05,
+      "loss": 0.4829,
+      "step": 4167
+    },
+    {
+      "epoch": 0.2561455260570305,
+      "grad_norm": 21.125,
+      "learning_rate": 1.813506961365528e-05,
+      "loss": 0.3637,
+      "step": 4168
+    },
+    {
+      "epoch": 0.2562069813176008,
+      "grad_norm": 8.625,
+      "learning_rate": 1.8133716053044123e-05,
+      "loss": 0.4434,
+      "step": 4169
+    },
+    {
+      "epoch": 0.2562684365781711,
+      "grad_norm": 54.25,
+      "learning_rate": 1.8132362051958013e-05,
+      "loss": 0.4382,
+      "step": 4170
+    },
+    {
+      "epoch": 0.2563298918387414,
+      "grad_norm": 9.25,
+      "learning_rate": 1.8131007610470278e-05,
+      "loss": 0.2613,
+      "step": 4171
+    },
+    {
+      "epoch": 0.2563913470993117,
+      "grad_norm": 91.0,
+      "learning_rate": 1.812965272865427e-05,
+      "loss": 0.3696,
+      "step": 4172
+    },
+    {
+      "epoch": 0.256452802359882,
+      "grad_norm": 39.25,
+      "learning_rate": 1.812829740658336e-05,
+      "loss": 0.3255,
+      "step": 4173
+    },
+    {
+      "epoch": 0.2565142576204523,
+      "grad_norm": 35.75,
+      "learning_rate": 1.812694164433094e-05,
+      "loss": 0.3954,
+      "step": 4174
+    },
+    {
+      "epoch": 0.2565757128810226,
+      "grad_norm": 70.0,
+      "learning_rate": 1.8125585441970435e-05,
+      "loss": 0.2871,
+      "step": 4175
+    },
+    {
+      "epoch": 0.25663716814159293,
+      "grad_norm": 7.4375,
+      "learning_rate": 1.8124228799575295e-05,
+      "loss": 0.2422,
+      "step": 4176
+    },
+    {
+      "epoch": 0.2566986234021632,
+      "grad_norm": 26.25,
+      "learning_rate": 1.812287171721897e-05,
+      "loss": 0.3589,
+      "step": 4177
+    },
+    {
+      "epoch": 0.25676007866273354,
+      "grad_norm": 47.5,
+      "learning_rate": 1.812151419497497e-05,
+      "loss": 0.3081,
+      "step": 4178
+    },
+    {
+      "epoch": 0.2568215339233038,
+      "grad_norm": 21.25,
+      "learning_rate": 1.8120156232916804e-05,
+      "loss": 0.3336,
+      "step": 4179
+    },
+    {
+      "epoch": 0.25688298918387414,
+      "grad_norm": 14.75,
+      "learning_rate": 1.811879783111801e-05,
+      "loss": 0.4109,
+      "step": 4180
+    },
+    {
+      "epoch": 0.2569444444444444,
+      "grad_norm": 64.0,
+      "learning_rate": 1.811743898965215e-05,
+      "loss": 0.4077,
+      "step": 4181
+    },
+    {
+      "epoch": 0.25700589970501475,
+      "grad_norm": 54.5,
+      "learning_rate": 1.8116079708592815e-05,
+      "loss": 0.2536,
+      "step": 4182
+    },
+    {
+      "epoch": 0.257067354965585,
+      "grad_norm": 21.875,
+      "learning_rate": 1.8114719988013612e-05,
+      "loss": 0.2583,
+      "step": 4183
+    },
+    {
+      "epoch": 0.25712881022615536,
+      "grad_norm": 12.25,
+      "learning_rate": 1.8113359827988177e-05,
+      "loss": 0.298,
+      "step": 4184
+    },
+    {
+      "epoch": 0.2571902654867257,
+      "grad_norm": 26.0,
+      "learning_rate": 1.8111999228590167e-05,
+      "loss": 0.3614,
+      "step": 4185
+    },
+    {
+      "epoch": 0.25725172074729596,
+      "grad_norm": 33.25,
+      "learning_rate": 1.8110638189893267e-05,
+      "loss": 0.5462,
+      "step": 4186
+    },
+    {
+      "epoch": 0.2573131760078663,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.8109276711971183e-05,
+      "loss": 0.3567,
+      "step": 4187
+    },
+    {
+      "epoch": 0.25737463126843657,
+      "grad_norm": 14.1875,
+      "learning_rate": 1.8107914794897644e-05,
+      "loss": 0.4188,
+      "step": 4188
+    },
+    {
+      "epoch": 0.2574360865290069,
+      "grad_norm": 16.25,
+      "learning_rate": 1.81065524387464e-05,
+      "loss": 0.3487,
+      "step": 4189
+    },
+    {
+      "epoch": 0.2574975417895772,
+      "grad_norm": 16.875,
+      "learning_rate": 1.8105189643591233e-05,
+      "loss": 0.3498,
+      "step": 4190
+    },
+    {
+      "epoch": 0.2575589970501475,
+      "grad_norm": 37.25,
+      "learning_rate": 1.8103826409505944e-05,
+      "loss": 0.3238,
+      "step": 4191
+    },
+    {
+      "epoch": 0.2576204523107178,
+      "grad_norm": 52.5,
+      "learning_rate": 1.8102462736564355e-05,
+      "loss": 0.3865,
+      "step": 4192
+    },
+    {
+      "epoch": 0.2576819075712881,
+      "grad_norm": 9.375,
+      "learning_rate": 1.8101098624840318e-05,
+      "loss": 0.3869,
+      "step": 4193
+    },
+    {
+      "epoch": 0.2577433628318584,
+      "grad_norm": 43.75,
+      "learning_rate": 1.80997340744077e-05,
+      "loss": 0.5714,
+      "step": 4194
+    },
+    {
+      "epoch": 0.2578048180924287,
+      "grad_norm": 25.875,
+      "learning_rate": 1.80983690853404e-05,
+      "loss": 0.3951,
+      "step": 4195
+    },
+    {
+      "epoch": 0.257866273352999,
+      "grad_norm": 19.375,
+      "learning_rate": 1.8097003657712343e-05,
+      "loss": 0.418,
+      "step": 4196
+    },
+    {
+      "epoch": 0.25792772861356933,
+      "grad_norm": 13.375,
+      "learning_rate": 1.809563779159746e-05,
+      "loss": 0.4214,
+      "step": 4197
+    },
+    {
+      "epoch": 0.2579891838741396,
+      "grad_norm": 11.6875,
+      "learning_rate": 1.8094271487069733e-05,
+      "loss": 0.2927,
+      "step": 4198
+    },
+    {
+      "epoch": 0.25805063913470994,
+      "grad_norm": 5.96875,
+      "learning_rate": 1.8092904744203146e-05,
+      "loss": 0.3161,
+      "step": 4199
+    },
+    {
+      "epoch": 0.2581120943952802,
+      "grad_norm": 10.875,
+      "learning_rate": 1.8091537563071714e-05,
+      "loss": 0.2786,
+      "step": 4200
+    },
+    {
+      "epoch": 0.25817354965585054,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8090169943749477e-05,
+      "loss": 0.4333,
+      "step": 4201
+    },
+    {
+      "epoch": 0.2582350049164208,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.8088801886310493e-05,
+      "loss": 0.4532,
+      "step": 4202
+    },
+    {
+      "epoch": 0.25829646017699115,
+      "grad_norm": 46.75,
+      "learning_rate": 1.8087433390828855e-05,
+      "loss": 0.5425,
+      "step": 4203
+    },
+    {
+      "epoch": 0.2583579154375615,
+      "grad_norm": 56.5,
+      "learning_rate": 1.8086064457378667e-05,
+      "loss": 0.5156,
+      "step": 4204
+    },
+    {
+      "epoch": 0.25841937069813176,
+      "grad_norm": 5.625,
+      "learning_rate": 1.8084695086034064e-05,
+      "loss": 0.4975,
+      "step": 4205
+    },
+    {
+      "epoch": 0.2584808259587021,
+      "grad_norm": 33.0,
+      "learning_rate": 1.8083325276869207e-05,
+      "loss": 0.4064,
+      "step": 4206
+    },
+    {
+      "epoch": 0.25854228121927236,
+      "grad_norm": 28.625,
+      "learning_rate": 1.8081955029958272e-05,
+      "loss": 0.431,
+      "step": 4207
+    },
+    {
+      "epoch": 0.2586037364798427,
+      "grad_norm": 27.5,
+      "learning_rate": 1.8080584345375466e-05,
+      "loss": 0.301,
+      "step": 4208
+    },
+    {
+      "epoch": 0.25866519174041297,
+      "grad_norm": 18.75,
+      "learning_rate": 1.8079213223195018e-05,
+      "loss": 0.4399,
+      "step": 4209
+    },
+    {
+      "epoch": 0.2587266470009833,
+      "grad_norm": 22.0,
+      "learning_rate": 1.8077841663491174e-05,
+      "loss": 0.3813,
+      "step": 4210
+    },
+    {
+      "epoch": 0.2587881022615536,
+      "grad_norm": 32.0,
+      "learning_rate": 1.807646966633822e-05,
+      "loss": 0.4416,
+      "step": 4211
+    },
+    {
+      "epoch": 0.2588495575221239,
+      "grad_norm": 27.125,
+      "learning_rate": 1.807509723181045e-05,
+      "loss": 0.5876,
+      "step": 4212
+    },
+    {
+      "epoch": 0.2589110127826942,
+      "grad_norm": 16.25,
+      "learning_rate": 1.8073724359982184e-05,
+      "loss": 0.5149,
+      "step": 4213
+    },
+    {
+      "epoch": 0.2589724680432645,
+      "grad_norm": 45.25,
+      "learning_rate": 1.8072351050927774e-05,
+      "loss": 0.4984,
+      "step": 4214
+    },
+    {
+      "epoch": 0.2590339233038348,
+      "grad_norm": 25.625,
+      "learning_rate": 1.8070977304721587e-05,
+      "loss": 0.3867,
+      "step": 4215
+    },
+    {
+      "epoch": 0.2590953785644051,
+      "grad_norm": 17.125,
+      "learning_rate": 1.806960312143802e-05,
+      "loss": 0.3835,
+      "step": 4216
+    },
+    {
+      "epoch": 0.2591568338249754,
+      "grad_norm": 30.875,
+      "learning_rate": 1.806822850115149e-05,
+      "loss": 0.5571,
+      "step": 4217
+    },
+    {
+      "epoch": 0.25921828908554573,
+      "grad_norm": 3.9375,
+      "learning_rate": 1.8066853443936435e-05,
+      "loss": 0.2761,
+      "step": 4218
+    },
+    {
+      "epoch": 0.259279744346116,
+      "grad_norm": 14.5625,
+      "learning_rate": 1.8065477949867327e-05,
+      "loss": 0.4148,
+      "step": 4219
+    },
+    {
+      "epoch": 0.25934119960668633,
+      "grad_norm": 18.0,
+      "learning_rate": 1.806410201901865e-05,
+      "loss": 0.5232,
+      "step": 4220
+    },
+    {
+      "epoch": 0.2594026548672566,
+      "grad_norm": 12.9375,
+      "learning_rate": 1.8062725651464913e-05,
+      "loss": 0.2265,
+      "step": 4221
+    },
+    {
+      "epoch": 0.25946411012782694,
+      "grad_norm": 28.875,
+      "learning_rate": 1.806134884728066e-05,
+      "loss": 0.4246,
+      "step": 4222
+    },
+    {
+      "epoch": 0.2595255653883973,
+      "grad_norm": 7.0625,
+      "learning_rate": 1.8059971606540446e-05,
+      "loss": 0.3047,
+      "step": 4223
+    },
+    {
+      "epoch": 0.25958702064896755,
+      "grad_norm": 25.125,
+      "learning_rate": 1.8058593929318855e-05,
+      "loss": 0.3112,
+      "step": 4224
+    },
+    {
+      "epoch": 0.2596484759095379,
+      "grad_norm": 22.5,
+      "learning_rate": 1.8057215815690494e-05,
+      "loss": 0.388,
+      "step": 4225
+    },
+    {
+      "epoch": 0.25970993117010815,
+      "grad_norm": 40.0,
+      "learning_rate": 1.8055837265729996e-05,
+      "loss": 0.3805,
+      "step": 4226
+    },
+    {
+      "epoch": 0.2597713864306785,
+      "grad_norm": 21.625,
+      "learning_rate": 1.8054458279512012e-05,
+      "loss": 0.3759,
+      "step": 4227
+    },
+    {
+      "epoch": 0.25983284169124876,
+      "grad_norm": 10.3125,
+      "learning_rate": 1.8053078857111218e-05,
+      "loss": 0.3041,
+      "step": 4228
+    },
+    {
+      "epoch": 0.2598942969518191,
+      "grad_norm": 24.0,
+      "learning_rate": 1.8051698998602324e-05,
+      "loss": 0.3018,
+      "step": 4229
+    },
+    {
+      "epoch": 0.25995575221238937,
+      "grad_norm": 23.125,
+      "learning_rate": 1.805031870406005e-05,
+      "loss": 0.3549,
+      "step": 4230
+    },
+    {
+      "epoch": 0.2600172074729597,
+      "grad_norm": 27.125,
+      "learning_rate": 1.804893797355914e-05,
+      "loss": 0.2752,
+      "step": 4231
+    },
+    {
+      "epoch": 0.26007866273353,
+      "grad_norm": 14.3125,
+      "learning_rate": 1.8047556807174378e-05,
+      "loss": 0.3193,
+      "step": 4232
+    },
+    {
+      "epoch": 0.2601401179941003,
+      "grad_norm": 13.1875,
+      "learning_rate": 1.804617520498055e-05,
+      "loss": 0.3511,
+      "step": 4233
+    },
+    {
+      "epoch": 0.2602015732546706,
+      "grad_norm": 12.75,
+      "learning_rate": 1.8044793167052476e-05,
+      "loss": 0.5629,
+      "step": 4234
+    },
+    {
+      "epoch": 0.2602630285152409,
+      "grad_norm": 13.125,
+      "learning_rate": 1.8043410693465e-05,
+      "loss": 0.4739,
+      "step": 4235
+    },
+    {
+      "epoch": 0.2603244837758112,
+      "grad_norm": 22.625,
+      "learning_rate": 1.8042027784292998e-05,
+      "loss": 0.4125,
+      "step": 4236
+    },
+    {
+      "epoch": 0.2603859390363815,
+      "grad_norm": 14.0625,
+      "learning_rate": 1.8040644439611348e-05,
+      "loss": 0.4686,
+      "step": 4237
+    },
+    {
+      "epoch": 0.2604473942969518,
+      "grad_norm": 15.25,
+      "learning_rate": 1.803926065949497e-05,
+      "loss": 0.2943,
+      "step": 4238
+    },
+    {
+      "epoch": 0.2605088495575221,
+      "grad_norm": 4.28125,
+      "learning_rate": 1.80378764440188e-05,
+      "loss": 0.3094,
+      "step": 4239
+    },
+    {
+      "epoch": 0.2605703048180924,
+      "grad_norm": 42.25,
+      "learning_rate": 1.80364917932578e-05,
+      "loss": 0.461,
+      "step": 4240
+    },
+    {
+      "epoch": 0.26063176007866273,
+      "grad_norm": 11.4375,
+      "learning_rate": 1.8035106707286957e-05,
+      "loss": 0.2831,
+      "step": 4241
+    },
+    {
+      "epoch": 0.26069321533923306,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.803372118618127e-05,
+      "loss": 0.3005,
+      "step": 4242
+    },
+    {
+      "epoch": 0.26075467059980334,
+      "grad_norm": 26.875,
+      "learning_rate": 1.803233523001578e-05,
+      "loss": 0.4125,
+      "step": 4243
+    },
+    {
+      "epoch": 0.26081612586037367,
+      "grad_norm": 23.625,
+      "learning_rate": 1.8030948838865542e-05,
+      "loss": 0.4211,
+      "step": 4244
+    },
+    {
+      "epoch": 0.26087758112094395,
+      "grad_norm": 24.0,
+      "learning_rate": 1.8029562012805627e-05,
+      "loss": 0.3436,
+      "step": 4245
+    },
+    {
+      "epoch": 0.2609390363815143,
+      "grad_norm": 7.625,
+      "learning_rate": 1.8028174751911147e-05,
+      "loss": 0.3951,
+      "step": 4246
+    },
+    {
+      "epoch": 0.26100049164208455,
+      "grad_norm": 47.75,
+      "learning_rate": 1.802678705625722e-05,
+      "loss": 0.4105,
+      "step": 4247
+    },
+    {
+      "epoch": 0.2610619469026549,
+      "grad_norm": 23.875,
+      "learning_rate": 1.8025398925919006e-05,
+      "loss": 0.289,
+      "step": 4248
+    },
+    {
+      "epoch": 0.26112340216322516,
+      "grad_norm": 5.6875,
+      "learning_rate": 1.802401036097167e-05,
+      "loss": 0.3061,
+      "step": 4249
+    },
+    {
+      "epoch": 0.2611848574237955,
+      "grad_norm": 6.25,
+      "learning_rate": 1.8022621361490406e-05,
+      "loss": 0.4481,
+      "step": 4250
+    },
+    {
+      "epoch": 0.26124631268436577,
+      "grad_norm": 27.125,
+      "learning_rate": 1.802123192755044e-05,
+      "loss": 0.3947,
+      "step": 4251
+    },
+    {
+      "epoch": 0.2613077679449361,
+      "grad_norm": 18.125,
+      "learning_rate": 1.801984205922701e-05,
+      "loss": 0.2871,
+      "step": 4252
+    },
+    {
+      "epoch": 0.2613692232055064,
+      "grad_norm": 8.4375,
+      "learning_rate": 1.8018451756595395e-05,
+      "loss": 0.3546,
+      "step": 4253
+    },
+    {
+      "epoch": 0.2614306784660767,
+      "grad_norm": 3.71875,
+      "learning_rate": 1.8017061019730874e-05,
+      "loss": 0.3179,
+      "step": 4254
+    },
+    {
+      "epoch": 0.261492133726647,
+      "grad_norm": 18.125,
+      "learning_rate": 1.8015669848708768e-05,
+      "loss": 0.2804,
+      "step": 4255
+    },
+    {
+      "epoch": 0.2615535889872173,
+      "grad_norm": 23.25,
+      "learning_rate": 1.8014278243604407e-05,
+      "loss": 0.2787,
+      "step": 4256
+    },
+    {
+      "epoch": 0.2616150442477876,
+      "grad_norm": 28.5,
+      "learning_rate": 1.8012886204493164e-05,
+      "loss": 0.3825,
+      "step": 4257
+    },
+    {
+      "epoch": 0.2616764995083579,
+      "grad_norm": 29.875,
+      "learning_rate": 1.8011493731450412e-05,
+      "loss": 0.5037,
+      "step": 4258
+    },
+    {
+      "epoch": 0.2617379547689282,
+      "grad_norm": 7.9375,
+      "learning_rate": 1.8010100824551563e-05,
+      "loss": 0.3621,
+      "step": 4259
+    },
+    {
+      "epoch": 0.2617994100294985,
+      "grad_norm": 17.625,
+      "learning_rate": 1.8008707483872055e-05,
+      "loss": 0.5608,
+      "step": 4260
+    },
+    {
+      "epoch": 0.26186086529006886,
+      "grad_norm": 12.75,
+      "learning_rate": 1.8007313709487334e-05,
+      "loss": 0.4398,
+      "step": 4261
+    },
+    {
+      "epoch": 0.26192232055063913,
+      "grad_norm": 34.5,
+      "learning_rate": 1.800591950147289e-05,
+      "loss": 0.3544,
+      "step": 4262
+    },
+    {
+      "epoch": 0.26198377581120946,
+      "grad_norm": 11.25,
+      "learning_rate": 1.8004524859904212e-05,
+      "loss": 0.3154,
+      "step": 4263
+    },
+    {
+      "epoch": 0.26204523107177974,
+      "grad_norm": 24.75,
+      "learning_rate": 1.8003129784856832e-05,
+      "loss": 0.3709,
+      "step": 4264
+    },
+    {
+      "epoch": 0.26210668633235007,
+      "grad_norm": 4.375,
+      "learning_rate": 1.8001734276406302e-05,
+      "loss": 0.2653,
+      "step": 4265
+    },
+    {
+      "epoch": 0.26216814159292035,
+      "grad_norm": 8.125,
+      "learning_rate": 1.800033833462819e-05,
+      "loss": 0.3774,
+      "step": 4266
+    },
+    {
+      "epoch": 0.2622295968534907,
+      "grad_norm": 21.75,
+      "learning_rate": 1.7998941959598097e-05,
+      "loss": 0.3589,
+      "step": 4267
+    },
+    {
+      "epoch": 0.26229105211406095,
+      "grad_norm": 6.59375,
+      "learning_rate": 1.7997545151391637e-05,
+      "loss": 0.4866,
+      "step": 4268
+    },
+    {
+      "epoch": 0.2623525073746313,
+      "grad_norm": 16.5,
+      "learning_rate": 1.7996147910084458e-05,
+      "loss": 0.3565,
+      "step": 4269
+    },
+    {
+      "epoch": 0.26241396263520156,
+      "grad_norm": 24.625,
+      "learning_rate": 1.799475023575222e-05,
+      "loss": 0.3473,
+      "step": 4270
+    },
+    {
+      "epoch": 0.2624754178957719,
+      "grad_norm": 25.625,
+      "learning_rate": 1.7993352128470617e-05,
+      "loss": 0.3704,
+      "step": 4271
+    },
+    {
+      "epoch": 0.26253687315634217,
+      "grad_norm": 5.46875,
+      "learning_rate": 1.7991953588315368e-05,
+      "loss": 0.3182,
+      "step": 4272
+    },
+    {
+      "epoch": 0.2625983284169125,
+      "grad_norm": 11.5,
+      "learning_rate": 1.79905546153622e-05,
+      "loss": 0.384,
+      "step": 4273
+    },
+    {
+      "epoch": 0.26265978367748277,
+      "grad_norm": 18.75,
+      "learning_rate": 1.798915520968688e-05,
+      "loss": 0.4202,
+      "step": 4274
+    },
+    {
+      "epoch": 0.2627212389380531,
+      "grad_norm": 22.25,
+      "learning_rate": 1.7987755371365185e-05,
+      "loss": 0.3984,
+      "step": 4275
+    },
+    {
+      "epoch": 0.2627826941986234,
+      "grad_norm": 60.75,
+      "learning_rate": 1.798635510047293e-05,
+      "loss": 0.5458,
+      "step": 4276
+    },
+    {
+      "epoch": 0.2628441494591937,
+      "grad_norm": 8.875,
+      "learning_rate": 1.798495439708594e-05,
+      "loss": 0.3383,
+      "step": 4277
+    },
+    {
+      "epoch": 0.262905604719764,
+      "grad_norm": 26.25,
+      "learning_rate": 1.7983553261280072e-05,
+      "loss": 0.3799,
+      "step": 4278
+    },
+    {
+      "epoch": 0.2629670599803343,
+      "grad_norm": 35.25,
+      "learning_rate": 1.7982151693131206e-05,
+      "loss": 0.5531,
+      "step": 4279
+    },
+    {
+      "epoch": 0.26302851524090465,
+      "grad_norm": 5.84375,
+      "learning_rate": 1.7980749692715233e-05,
+      "loss": 0.3394,
+      "step": 4280
+    },
+    {
+      "epoch": 0.2630899705014749,
+      "grad_norm": 25.5,
+      "learning_rate": 1.7979347260108088e-05,
+      "loss": 0.3616,
+      "step": 4281
+    },
+    {
+      "epoch": 0.26315142576204525,
+      "grad_norm": 24.125,
+      "learning_rate": 1.7977944395385713e-05,
+      "loss": 0.5491,
+      "step": 4282
+    },
+    {
+      "epoch": 0.26321288102261553,
+      "grad_norm": 20.25,
+      "learning_rate": 1.797654109862408e-05,
+      "loss": 0.4295,
+      "step": 4283
+    },
+    {
+      "epoch": 0.26327433628318586,
+      "grad_norm": 8.375,
+      "learning_rate": 1.797513736989918e-05,
+      "loss": 0.5473,
+      "step": 4284
+    },
+    {
+      "epoch": 0.26333579154375614,
+      "grad_norm": 15.125,
+      "learning_rate": 1.7973733209287036e-05,
+      "loss": 0.4334,
+      "step": 4285
+    },
+    {
+      "epoch": 0.26339724680432647,
+      "grad_norm": 15.25,
+      "learning_rate": 1.797232861686369e-05,
+      "loss": 0.3881,
+      "step": 4286
+    },
+    {
+      "epoch": 0.26345870206489674,
+      "grad_norm": 3.671875,
+      "learning_rate": 1.7970923592705203e-05,
+      "loss": 0.3396,
+      "step": 4287
+    },
+    {
+      "epoch": 0.2635201573254671,
+      "grad_norm": 16.375,
+      "learning_rate": 1.7969518136887664e-05,
+      "loss": 0.3157,
+      "step": 4288
+    },
+    {
+      "epoch": 0.26358161258603735,
+      "grad_norm": 19.125,
+      "learning_rate": 1.7968112249487183e-05,
+      "loss": 0.4432,
+      "step": 4289
+    },
+    {
+      "epoch": 0.2636430678466077,
+      "grad_norm": 18.5,
+      "learning_rate": 1.79667059305799e-05,
+      "loss": 0.4313,
+      "step": 4290
+    },
+    {
+      "epoch": 0.26370452310717796,
+      "grad_norm": 33.25,
+      "learning_rate": 1.7965299180241963e-05,
+      "loss": 0.4168,
+      "step": 4291
+    },
+    {
+      "epoch": 0.2637659783677483,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.7963891998549562e-05,
+      "loss": 0.534,
+      "step": 4292
+    },
+    {
+      "epoch": 0.26382743362831856,
+      "grad_norm": 5.34375,
+      "learning_rate": 1.7962484385578904e-05,
+      "loss": 0.2712,
+      "step": 4293
+    },
+    {
+      "epoch": 0.2638888888888889,
+      "grad_norm": 3.578125,
+      "learning_rate": 1.796107634140621e-05,
+      "loss": 0.2788,
+      "step": 4294
+    },
+    {
+      "epoch": 0.26395034414945917,
+      "grad_norm": 55.25,
+      "learning_rate": 1.795966786610773e-05,
+      "loss": 0.4355,
+      "step": 4295
+    },
+    {
+      "epoch": 0.2640117994100295,
+      "grad_norm": 3.125,
+      "learning_rate": 1.7958258959759747e-05,
+      "loss": 0.254,
+      "step": 4296
+    },
+    {
+      "epoch": 0.2640732546705998,
+      "grad_norm": 24.25,
+      "learning_rate": 1.7956849622438554e-05,
+      "loss": 0.305,
+      "step": 4297
+    },
+    {
+      "epoch": 0.2641347099311701,
+      "grad_norm": 17.125,
+      "learning_rate": 1.7955439854220477e-05,
+      "loss": 0.2689,
+      "step": 4298
+    },
+    {
+      "epoch": 0.26419616519174044,
+      "grad_norm": 10.1875,
+      "learning_rate": 1.7954029655181856e-05,
+      "loss": 0.3408,
+      "step": 4299
+    },
+    {
+      "epoch": 0.2642576204523107,
+      "grad_norm": 4.3125,
+      "learning_rate": 1.795261902539906e-05,
+      "loss": 0.2369,
+      "step": 4300
+    },
+    {
+      "epoch": 0.26431907571288105,
+      "grad_norm": 38.75,
+      "learning_rate": 1.795120796494848e-05,
+      "loss": 0.3055,
+      "step": 4301
+    },
+    {
+      "epoch": 0.2643805309734513,
+      "grad_norm": 2.828125,
+      "learning_rate": 1.7949796473906533e-05,
+      "loss": 0.2509,
+      "step": 4302
+    },
+    {
+      "epoch": 0.26444198623402165,
+      "grad_norm": 33.5,
+      "learning_rate": 1.794838455234966e-05,
+      "loss": 0.2663,
+      "step": 4303
+    },
+    {
+      "epoch": 0.26450344149459193,
+      "grad_norm": 20.625,
+      "learning_rate": 1.794697220035431e-05,
+      "loss": 0.2604,
+      "step": 4304
+    },
+    {
+      "epoch": 0.26456489675516226,
+      "grad_norm": 21.125,
+      "learning_rate": 1.7945559417996984e-05,
+      "loss": 0.394,
+      "step": 4305
+    },
+    {
+      "epoch": 0.26462635201573254,
+      "grad_norm": 18.75,
+      "learning_rate": 1.7944146205354182e-05,
+      "loss": 0.4566,
+      "step": 4306
+    },
+    {
+      "epoch": 0.26468780727630287,
+      "grad_norm": 9.3125,
+      "learning_rate": 1.7942732562502435e-05,
+      "loss": 0.2514,
+      "step": 4307
+    },
+    {
+      "epoch": 0.26474926253687314,
+      "grad_norm": 15.375,
+      "learning_rate": 1.79413184895183e-05,
+      "loss": 0.372,
+      "step": 4308
+    },
+    {
+      "epoch": 0.2648107177974435,
+      "grad_norm": 23.25,
+      "learning_rate": 1.7939903986478354e-05,
+      "loss": 0.3144,
+      "step": 4309
+    },
+    {
+      "epoch": 0.26487217305801375,
+      "grad_norm": 7.09375,
+      "learning_rate": 1.7938489053459198e-05,
+      "loss": 0.2951,
+      "step": 4310
+    },
+    {
+      "epoch": 0.2649336283185841,
+      "grad_norm": 3.9375,
+      "learning_rate": 1.793707369053746e-05,
+      "loss": 0.2469,
+      "step": 4311
+    },
+    {
+      "epoch": 0.26499508357915436,
+      "grad_norm": 19.25,
+      "learning_rate": 1.793565789778978e-05,
+      "loss": 0.3618,
+      "step": 4312
+    },
+    {
+      "epoch": 0.2650565388397247,
+      "grad_norm": 18.375,
+      "learning_rate": 1.7934241675292834e-05,
+      "loss": 0.4894,
+      "step": 4313
+    },
+    {
+      "epoch": 0.26511799410029496,
+      "grad_norm": 30.625,
+      "learning_rate": 1.793282502312332e-05,
+      "loss": 0.6266,
+      "step": 4314
+    },
+    {
+      "epoch": 0.2651794493608653,
+      "grad_norm": 14.6875,
+      "learning_rate": 1.793140794135795e-05,
+      "loss": 0.4613,
+      "step": 4315
+    },
+    {
+      "epoch": 0.26524090462143557,
+      "grad_norm": 21.5,
+      "learning_rate": 1.7929990430073463e-05,
+      "loss": 0.5646,
+      "step": 4316
+    },
+    {
+      "epoch": 0.2653023598820059,
+      "grad_norm": 23.125,
+      "learning_rate": 1.792857248934663e-05,
+      "loss": 0.3674,
+      "step": 4317
+    },
+    {
+      "epoch": 0.26536381514257623,
+      "grad_norm": 48.0,
+      "learning_rate": 1.7927154119254234e-05,
+      "loss": 0.6913,
+      "step": 4318
+    },
+    {
+      "epoch": 0.2654252704031465,
+      "grad_norm": 27.25,
+      "learning_rate": 1.792573531987309e-05,
+      "loss": 0.4885,
+      "step": 4319
+    },
+    {
+      "epoch": 0.26548672566371684,
+      "grad_norm": 19.5,
+      "learning_rate": 1.792431609128003e-05,
+      "loss": 0.4729,
+      "step": 4320
+    },
+    {
+      "epoch": 0.2655481809242871,
+      "grad_norm": 7.53125,
+      "learning_rate": 1.792289643355191e-05,
+      "loss": 0.4218,
+      "step": 4321
+    },
+    {
+      "epoch": 0.26560963618485744,
+      "grad_norm": 7.65625,
+      "learning_rate": 1.7921476346765607e-05,
+      "loss": 0.3223,
+      "step": 4322
+    },
+    {
+      "epoch": 0.2656710914454277,
+      "grad_norm": 10.0,
+      "learning_rate": 1.7920055830998032e-05,
+      "loss": 0.3082,
+      "step": 4323
+    },
+    {
+      "epoch": 0.26573254670599805,
+      "grad_norm": 6.5625,
+      "learning_rate": 1.791863488632611e-05,
+      "loss": 0.2957,
+      "step": 4324
+    },
+    {
+      "epoch": 0.2657940019665683,
+      "grad_norm": 24.625,
+      "learning_rate": 1.7917213512826785e-05,
+      "loss": 0.3626,
+      "step": 4325
+    },
+    {
+      "epoch": 0.26585545722713866,
+      "grad_norm": 11.3125,
+      "learning_rate": 1.7915791710577035e-05,
+      "loss": 0.249,
+      "step": 4326
+    },
+    {
+      "epoch": 0.26591691248770893,
+      "grad_norm": 2.359375,
+      "learning_rate": 1.7914369479653858e-05,
+      "loss": 0.3,
+      "step": 4327
+    },
+    {
+      "epoch": 0.26597836774827927,
+      "grad_norm": 2.296875,
+      "learning_rate": 1.7912946820134273e-05,
+      "loss": 0.2719,
+      "step": 4328
+    },
+    {
+      "epoch": 0.26603982300884954,
+      "grad_norm": 14.75,
+      "learning_rate": 1.7911523732095322e-05,
+      "loss": 0.5472,
+      "step": 4329
+    },
+    {
+      "epoch": 0.26610127826941987,
+      "grad_norm": 63.0,
+      "learning_rate": 1.791010021561407e-05,
+      "loss": 0.5495,
+      "step": 4330
+    },
+    {
+      "epoch": 0.26616273352999015,
+      "grad_norm": 21.5,
+      "learning_rate": 1.7908676270767608e-05,
+      "loss": 0.4389,
+      "step": 4331
+    },
+    {
+      "epoch": 0.2662241887905605,
+      "grad_norm": 22.125,
+      "learning_rate": 1.7907251897633052e-05,
+      "loss": 0.6534,
+      "step": 4332
+    },
+    {
+      "epoch": 0.26628564405113075,
+      "grad_norm": 38.25,
+      "learning_rate": 1.7905827096287532e-05,
+      "loss": 0.6959,
+      "step": 4333
+    },
+    {
+      "epoch": 0.2663470993117011,
+      "grad_norm": 26.25,
+      "learning_rate": 1.790440186680821e-05,
+      "loss": 0.4163,
+      "step": 4334
+    },
+    {
+      "epoch": 0.26640855457227136,
+      "grad_norm": 55.25,
+      "learning_rate": 1.7902976209272265e-05,
+      "loss": 0.443,
+      "step": 4335
+    },
+    {
+      "epoch": 0.2664700098328417,
+      "grad_norm": 30.25,
+      "learning_rate": 1.7901550123756906e-05,
+      "loss": 0.3894,
+      "step": 4336
+    },
+    {
+      "epoch": 0.266531465093412,
+      "grad_norm": 15.4375,
+      "learning_rate": 1.790012361033936e-05,
+      "loss": 0.358,
+      "step": 4337
+    },
+    {
+      "epoch": 0.2665929203539823,
+      "grad_norm": 33.5,
+      "learning_rate": 1.789869666909688e-05,
+      "loss": 0.4074,
+      "step": 4338
+    },
+    {
+      "epoch": 0.26665437561455263,
+      "grad_norm": 31.625,
+      "learning_rate": 1.789726930010674e-05,
+      "loss": 0.463,
+      "step": 4339
+    },
+    {
+      "epoch": 0.2667158308751229,
+      "grad_norm": 8.25,
+      "learning_rate": 1.7895841503446237e-05,
+      "loss": 0.2734,
+      "step": 4340
+    },
+    {
+      "epoch": 0.26677728613569324,
+      "grad_norm": 35.25,
+      "learning_rate": 1.7894413279192693e-05,
+      "loss": 0.3907,
+      "step": 4341
+    },
+    {
+      "epoch": 0.2668387413962635,
+      "grad_norm": 16.125,
+      "learning_rate": 1.789298462742345e-05,
+      "loss": 0.4822,
+      "step": 4342
+    },
+    {
+      "epoch": 0.26690019665683384,
+      "grad_norm": 6.125,
+      "learning_rate": 1.789155554821588e-05,
+      "loss": 0.325,
+      "step": 4343
+    },
+    {
+      "epoch": 0.2669616519174041,
+      "grad_norm": 10.25,
+      "learning_rate": 1.7890126041647375e-05,
+      "loss": 0.3348,
+      "step": 4344
+    },
+    {
+      "epoch": 0.26702310717797445,
+      "grad_norm": 25.375,
+      "learning_rate": 1.7888696107795343e-05,
+      "loss": 0.3869,
+      "step": 4345
+    },
+    {
+      "epoch": 0.2670845624385447,
+      "grad_norm": 8.375,
+      "learning_rate": 1.7887265746737224e-05,
+      "loss": 0.4301,
+      "step": 4346
+    },
+    {
+      "epoch": 0.26714601769911506,
+      "grad_norm": 39.25,
+      "learning_rate": 1.7885834958550473e-05,
+      "loss": 0.4198,
+      "step": 4347
+    },
+    {
+      "epoch": 0.26720747295968533,
+      "grad_norm": 16.875,
+      "learning_rate": 1.7884403743312583e-05,
+      "loss": 0.4369,
+      "step": 4348
+    },
+    {
+      "epoch": 0.26726892822025566,
+      "grad_norm": 29.375,
+      "learning_rate": 1.7882972101101054e-05,
+      "loss": 0.4247,
+      "step": 4349
+    },
+    {
+      "epoch": 0.26733038348082594,
+      "grad_norm": 17.25,
+      "learning_rate": 1.7881540031993415e-05,
+      "loss": 0.3474,
+      "step": 4350
+    },
+    {
+      "epoch": 0.26739183874139627,
+      "grad_norm": 17.125,
+      "learning_rate": 1.788010753606722e-05,
+      "loss": 0.3022,
+      "step": 4351
+    },
+    {
+      "epoch": 0.26745329400196655,
+      "grad_norm": 35.5,
+      "learning_rate": 1.787867461340005e-05,
+      "loss": 0.2518,
+      "step": 4352
+    },
+    {
+      "epoch": 0.2675147492625369,
+      "grad_norm": 36.75,
+      "learning_rate": 1.787724126406949e-05,
+      "loss": 0.5584,
+      "step": 4353
+    },
+    {
+      "epoch": 0.26757620452310715,
+      "grad_norm": 31.0,
+      "learning_rate": 1.7875807488153173e-05,
+      "loss": 0.4568,
+      "step": 4354
+    },
+    {
+      "epoch": 0.2676376597836775,
+      "grad_norm": 27.5,
+      "learning_rate": 1.7874373285728742e-05,
+      "loss": 0.5332,
+      "step": 4355
+    },
+    {
+      "epoch": 0.2676991150442478,
+      "grad_norm": 19.875,
+      "learning_rate": 1.7872938656873864e-05,
+      "loss": 0.4849,
+      "step": 4356
+    },
+    {
+      "epoch": 0.2677605703048181,
+      "grad_norm": 22.625,
+      "learning_rate": 1.7871503601666233e-05,
+      "loss": 0.3369,
+      "step": 4357
+    },
+    {
+      "epoch": 0.2678220255653884,
+      "grad_norm": 19.125,
+      "learning_rate": 1.7870068120183563e-05,
+      "loss": 0.4508,
+      "step": 4358
+    },
+    {
+      "epoch": 0.2678834808259587,
+      "grad_norm": 13.4375,
+      "learning_rate": 1.7868632212503583e-05,
+      "loss": 0.2925,
+      "step": 4359
+    },
+    {
+      "epoch": 0.26794493608652903,
+      "grad_norm": 8.625,
+      "learning_rate": 1.7867195878704062e-05,
+      "loss": 0.3755,
+      "step": 4360
+    },
+    {
+      "epoch": 0.2680063913470993,
+      "grad_norm": 25.25,
+      "learning_rate": 1.7865759118862784e-05,
+      "loss": 0.4023,
+      "step": 4361
+    },
+    {
+      "epoch": 0.26806784660766964,
+      "grad_norm": 12.8125,
+      "learning_rate": 1.7864321933057547e-05,
+      "loss": 0.4347,
+      "step": 4362
+    },
+    {
+      "epoch": 0.2681293018682399,
+      "grad_norm": 6.90625,
+      "learning_rate": 1.786288432136619e-05,
+      "loss": 0.618,
+      "step": 4363
+    },
+    {
+      "epoch": 0.26819075712881024,
+      "grad_norm": 6.5625,
+      "learning_rate": 1.7861446283866562e-05,
+      "loss": 0.3377,
+      "step": 4364
+    },
+    {
+      "epoch": 0.2682522123893805,
+      "grad_norm": 19.875,
+      "learning_rate": 1.786000782063654e-05,
+      "loss": 0.4208,
+      "step": 4365
+    },
+    {
+      "epoch": 0.26831366764995085,
+      "grad_norm": 13.6875,
+      "learning_rate": 1.785856893175402e-05,
+      "loss": 0.3504,
+      "step": 4366
+    },
+    {
+      "epoch": 0.2683751229105211,
+      "grad_norm": 2.609375,
+      "learning_rate": 1.7857129617296926e-05,
+      "loss": 0.3679,
+      "step": 4367
+    },
+    {
+      "epoch": 0.26843657817109146,
+      "grad_norm": 13.9375,
+      "learning_rate": 1.7855689877343204e-05,
+      "loss": 0.407,
+      "step": 4368
+    },
+    {
+      "epoch": 0.26849803343166173,
+      "grad_norm": 12.75,
+      "learning_rate": 1.785424971197082e-05,
+      "loss": 0.2408,
+      "step": 4369
+    },
+    {
+      "epoch": 0.26855948869223206,
+      "grad_norm": 20.75,
+      "learning_rate": 1.7852809121257765e-05,
+      "loss": 0.3611,
+      "step": 4370
+    },
+    {
+      "epoch": 0.26862094395280234,
+      "grad_norm": 7.4375,
+      "learning_rate": 1.7851368105282054e-05,
+      "loss": 0.2275,
+      "step": 4371
+    },
+    {
+      "epoch": 0.26868239921337267,
+      "grad_norm": 26.375,
+      "learning_rate": 1.7849926664121726e-05,
+      "loss": 0.3555,
+      "step": 4372
+    },
+    {
+      "epoch": 0.26874385447394294,
+      "grad_norm": 33.25,
+      "learning_rate": 1.7848484797854838e-05,
+      "loss": 0.2765,
+      "step": 4373
+    },
+    {
+      "epoch": 0.2688053097345133,
+      "grad_norm": 30.25,
+      "learning_rate": 1.7847042506559467e-05,
+      "loss": 0.3247,
+      "step": 4374
+    },
+    {
+      "epoch": 0.26886676499508355,
+      "grad_norm": 50.5,
+      "learning_rate": 1.7845599790313735e-05,
+      "loss": 0.3491,
+      "step": 4375
+    },
+    {
+      "epoch": 0.2689282202556539,
+      "grad_norm": 6.46875,
+      "learning_rate": 1.784415664919576e-05,
+      "loss": 0.2743,
+      "step": 4376
+    },
+    {
+      "epoch": 0.2689896755162242,
+      "grad_norm": 15.75,
+      "learning_rate": 1.7842713083283695e-05,
+      "loss": 0.5154,
+      "step": 4377
+    },
+    {
+      "epoch": 0.2690511307767945,
+      "grad_norm": 47.75,
+      "learning_rate": 1.7841269092655714e-05,
+      "loss": 0.5194,
+      "step": 4378
+    },
+    {
+      "epoch": 0.2691125860373648,
+      "grad_norm": 36.5,
+      "learning_rate": 1.7839824677390023e-05,
+      "loss": 0.3639,
+      "step": 4379
+    },
+    {
+      "epoch": 0.2691740412979351,
+      "grad_norm": 42.5,
+      "learning_rate": 1.7838379837564838e-05,
+      "loss": 0.6486,
+      "step": 4380
+    },
+    {
+      "epoch": 0.2692354965585054,
+      "grad_norm": 28.75,
+      "learning_rate": 1.78369345732584e-05,
+      "loss": 0.6015,
+      "step": 4381
+    },
+    {
+      "epoch": 0.2692969518190757,
+      "grad_norm": 5.0625,
+      "learning_rate": 1.7835488884548978e-05,
+      "loss": 0.4801,
+      "step": 4382
+    },
+    {
+      "epoch": 0.26935840707964603,
+      "grad_norm": 4.5,
+      "learning_rate": 1.7834042771514868e-05,
+      "loss": 0.3875,
+      "step": 4383
+    },
+    {
+      "epoch": 0.2694198623402163,
+      "grad_norm": 5.90625,
+      "learning_rate": 1.7832596234234376e-05,
+      "loss": 0.4638,
+      "step": 4384
+    },
+    {
+      "epoch": 0.26948131760078664,
+      "grad_norm": 12.875,
+      "learning_rate": 1.7831149272785843e-05,
+      "loss": 0.2728,
+      "step": 4385
+    },
+    {
+      "epoch": 0.2695427728613569,
+      "grad_norm": 8.625,
+      "learning_rate": 1.7829701887247618e-05,
+      "loss": 0.2833,
+      "step": 4386
+    },
+    {
+      "epoch": 0.26960422812192725,
+      "grad_norm": 24.125,
+      "learning_rate": 1.78282540776981e-05,
+      "loss": 0.3736,
+      "step": 4387
+    },
+    {
+      "epoch": 0.2696656833824975,
+      "grad_norm": 3.359375,
+      "learning_rate": 1.782680584421568e-05,
+      "loss": 0.297,
+      "step": 4388
+    },
+    {
+      "epoch": 0.26972713864306785,
+      "grad_norm": 50.25,
+      "learning_rate": 1.7825357186878788e-05,
+      "loss": 0.3855,
+      "step": 4389
+    },
+    {
+      "epoch": 0.26978859390363813,
+      "grad_norm": 15.625,
+      "learning_rate": 1.7823908105765883e-05,
+      "loss": 0.3442,
+      "step": 4390
+    },
+    {
+      "epoch": 0.26985004916420846,
+      "grad_norm": 9.125,
+      "learning_rate": 1.7822458600955432e-05,
+      "loss": 0.3818,
+      "step": 4391
+    },
+    {
+      "epoch": 0.26991150442477874,
+      "grad_norm": 14.625,
+      "learning_rate": 1.782100867252593e-05,
+      "loss": 0.264,
+      "step": 4392
+    },
+    {
+      "epoch": 0.26997295968534907,
+      "grad_norm": 44.0,
+      "learning_rate": 1.7819558320555902e-05,
+      "loss": 0.2789,
+      "step": 4393
+    },
+    {
+      "epoch": 0.27003441494591934,
+      "grad_norm": 63.0,
+      "learning_rate": 1.781810754512389e-05,
+      "loss": 0.4387,
+      "step": 4394
+    },
+    {
+      "epoch": 0.2700958702064897,
+      "grad_norm": 43.75,
+      "learning_rate": 1.7816656346308453e-05,
+      "loss": 0.3304,
+      "step": 4395
+    },
+    {
+      "epoch": 0.27015732546706,
+      "grad_norm": 17.0,
+      "learning_rate": 1.781520472418819e-05,
+      "loss": 0.3274,
+      "step": 4396
+    },
+    {
+      "epoch": 0.2702187807276303,
+      "grad_norm": 18.625,
+      "learning_rate": 1.7813752678841702e-05,
+      "loss": 0.3602,
+      "step": 4397
+    },
+    {
+      "epoch": 0.2702802359882006,
+      "grad_norm": 27.5,
+      "learning_rate": 1.7812300210347634e-05,
+      "loss": 0.4781,
+      "step": 4398
+    },
+    {
+      "epoch": 0.2703416912487709,
+      "grad_norm": 22.75,
+      "learning_rate": 1.7810847318784632e-05,
+      "loss": 0.3716,
+      "step": 4399
+    },
+    {
+      "epoch": 0.2704031465093412,
+      "grad_norm": 23.25,
+      "learning_rate": 1.7809394004231386e-05,
+      "loss": 0.3729,
+      "step": 4400
+    },
+    {
+      "epoch": 0.2704646017699115,
+      "grad_norm": 15.9375,
+      "learning_rate": 1.7807940266766595e-05,
+      "loss": 0.3171,
+      "step": 4401
+    },
+    {
+      "epoch": 0.2705260570304818,
+      "grad_norm": 70.5,
+      "learning_rate": 1.7806486106468983e-05,
+      "loss": 0.3627,
+      "step": 4402
+    },
+    {
+      "epoch": 0.2705875122910521,
+      "grad_norm": 20.5,
+      "learning_rate": 1.78050315234173e-05,
+      "loss": 0.385,
+      "step": 4403
+    },
+    {
+      "epoch": 0.27064896755162243,
+      "grad_norm": 26.625,
+      "learning_rate": 1.7803576517690323e-05,
+      "loss": 0.2544,
+      "step": 4404
+    },
+    {
+      "epoch": 0.2707104228121927,
+      "grad_norm": 32.25,
+      "learning_rate": 1.780212108936684e-05,
+      "loss": 0.33,
+      "step": 4405
+    },
+    {
+      "epoch": 0.27077187807276304,
+      "grad_norm": 18.125,
+      "learning_rate": 1.780066523852567e-05,
+      "loss": 0.3494,
+      "step": 4406
+    },
+    {
+      "epoch": 0.2708333333333333,
+      "grad_norm": 7.6875,
+      "learning_rate": 1.779920896524565e-05,
+      "loss": 0.2756,
+      "step": 4407
+    },
+    {
+      "epoch": 0.27089478859390365,
+      "grad_norm": 3.6875,
+      "learning_rate": 1.7797752269605654e-05,
+      "loss": 0.2128,
+      "step": 4408
+    },
+    {
+      "epoch": 0.2709562438544739,
+      "grad_norm": 6.125,
+      "learning_rate": 1.779629515168456e-05,
+      "loss": 0.2239,
+      "step": 4409
+    },
+    {
+      "epoch": 0.27101769911504425,
+      "grad_norm": 25.75,
+      "learning_rate": 1.779483761156128e-05,
+      "loss": 0.3208,
+      "step": 4410
+    },
+    {
+      "epoch": 0.27107915437561453,
+      "grad_norm": 4.84375,
+      "learning_rate": 1.7793379649314743e-05,
+      "loss": 0.2531,
+      "step": 4411
+    },
+    {
+      "epoch": 0.27114060963618486,
+      "grad_norm": 43.75,
+      "learning_rate": 1.7791921265023906e-05,
+      "loss": 0.3843,
+      "step": 4412
+    },
+    {
+      "epoch": 0.27120206489675514,
+      "grad_norm": 22.25,
+      "learning_rate": 1.7790462458767747e-05,
+      "loss": 0.2271,
+      "step": 4413
+    },
+    {
+      "epoch": 0.27126352015732547,
+      "grad_norm": 4.59375,
+      "learning_rate": 1.7789003230625266e-05,
+      "loss": 0.2941,
+      "step": 4414
+    },
+    {
+      "epoch": 0.2713249754178958,
+      "grad_norm": 5.375,
+      "learning_rate": 1.7787543580675485e-05,
+      "loss": 0.1836,
+      "step": 4415
+    },
+    {
+      "epoch": 0.2713864306784661,
+      "grad_norm": 8.1875,
+      "learning_rate": 1.7786083508997452e-05,
+      "loss": 0.2255,
+      "step": 4416
+    },
+    {
+      "epoch": 0.2714478859390364,
+      "grad_norm": 3.9375,
+      "learning_rate": 1.7784623015670237e-05,
+      "loss": 0.2567,
+      "step": 4417
+    },
+    {
+      "epoch": 0.2715093411996067,
+      "grad_norm": 27.5,
+      "learning_rate": 1.7783162100772928e-05,
+      "loss": 0.342,
+      "step": 4418
+    },
+    {
+      "epoch": 0.271570796460177,
+      "grad_norm": 9.9375,
+      "learning_rate": 1.7781700764384642e-05,
+      "loss": 0.2613,
+      "step": 4419
+    },
+    {
+      "epoch": 0.2716322517207473,
+      "grad_norm": 6.25,
+      "learning_rate": 1.7780239006584515e-05,
+      "loss": 0.329,
+      "step": 4420
+    },
+    {
+      "epoch": 0.2716937069813176,
+      "grad_norm": 44.0,
+      "learning_rate": 1.7778776827451715e-05,
+      "loss": 0.4563,
+      "step": 4421
+    },
+    {
+      "epoch": 0.2717551622418879,
+      "grad_norm": 4.15625,
+      "learning_rate": 1.777731422706541e-05,
+      "loss": 0.275,
+      "step": 4422
+    },
+    {
+      "epoch": 0.2718166175024582,
+      "grad_norm": 4.5625,
+      "learning_rate": 1.7775851205504823e-05,
+      "loss": 0.2362,
+      "step": 4423
+    },
+    {
+      "epoch": 0.2718780727630285,
+      "grad_norm": 7.09375,
+      "learning_rate": 1.777438776284917e-05,
+      "loss": 0.2526,
+      "step": 4424
+    },
+    {
+      "epoch": 0.27193952802359883,
+      "grad_norm": 21.75,
+      "learning_rate": 1.777292389917771e-05,
+      "loss": 0.4178,
+      "step": 4425
+    },
+    {
+      "epoch": 0.2720009832841691,
+      "grad_norm": 34.5,
+      "learning_rate": 1.777145961456971e-05,
+      "loss": 0.4348,
+      "step": 4426
+    },
+    {
+      "epoch": 0.27206243854473944,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.7769994909104475e-05,
+      "loss": 0.5546,
+      "step": 4427
+    },
+    {
+      "epoch": 0.2721238938053097,
+      "grad_norm": 13.25,
+      "learning_rate": 1.776852978286132e-05,
+      "loss": 0.3859,
+      "step": 4428
+    },
+    {
+      "epoch": 0.27218534906588004,
+      "grad_norm": 19.0,
+      "learning_rate": 1.7767064235919594e-05,
+      "loss": 0.5555,
+      "step": 4429
+    },
+    {
+      "epoch": 0.2722468043264503,
+      "grad_norm": 7.03125,
+      "learning_rate": 1.776559826835865e-05,
+      "loss": 0.3213,
+      "step": 4430
+    },
+    {
+      "epoch": 0.27230825958702065,
+      "grad_norm": 30.75,
+      "learning_rate": 1.7764131880257892e-05,
+      "loss": 0.5271,
+      "step": 4431
+    },
+    {
+      "epoch": 0.2723697148475909,
+      "grad_norm": 26.625,
+      "learning_rate": 1.776266507169672e-05,
+      "loss": 0.4243,
+      "step": 4432
+    },
+    {
+      "epoch": 0.27243117010816126,
+      "grad_norm": 23.125,
+      "learning_rate": 1.776119784275457e-05,
+      "loss": 0.4466,
+      "step": 4433
+    },
+    {
+      "epoch": 0.2724926253687316,
+      "grad_norm": 21.75,
+      "learning_rate": 1.7759730193510905e-05,
+      "loss": 0.4243,
+      "step": 4434
+    },
+    {
+      "epoch": 0.27255408062930186,
+      "grad_norm": 18.875,
+      "learning_rate": 1.7758262124045195e-05,
+      "loss": 0.3668,
+      "step": 4435
+    },
+    {
+      "epoch": 0.2726155358898722,
+      "grad_norm": 23.625,
+      "learning_rate": 1.7756793634436947e-05,
+      "loss": 0.3364,
+      "step": 4436
+    },
+    {
+      "epoch": 0.27267699115044247,
+      "grad_norm": 14.8125,
+      "learning_rate": 1.7755324724765688e-05,
+      "loss": 0.4631,
+      "step": 4437
+    },
+    {
+      "epoch": 0.2727384464110128,
+      "grad_norm": 25.875,
+      "learning_rate": 1.775385539511096e-05,
+      "loss": 0.3267,
+      "step": 4438
+    },
+    {
+      "epoch": 0.2727999016715831,
+      "grad_norm": 3.65625,
+      "learning_rate": 1.7752385645552336e-05,
+      "loss": 0.2511,
+      "step": 4439
+    },
+    {
+      "epoch": 0.2728613569321534,
+      "grad_norm": 15.8125,
+      "learning_rate": 1.775091547616941e-05,
+      "loss": 0.3514,
+      "step": 4440
+    },
+    {
+      "epoch": 0.2729228121927237,
+      "grad_norm": 21.75,
+      "learning_rate": 1.7749444887041797e-05,
+      "loss": 0.3792,
+      "step": 4441
+    },
+    {
+      "epoch": 0.272984267453294,
+      "grad_norm": 19.25,
+      "learning_rate": 1.7747973878249137e-05,
+      "loss": 0.2857,
+      "step": 4442
+    },
+    {
+      "epoch": 0.2730457227138643,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.774650244987109e-05,
+      "loss": 0.387,
+      "step": 4443
+    },
+    {
+      "epoch": 0.2731071779744346,
+      "grad_norm": 23.125,
+      "learning_rate": 1.7745030601987338e-05,
+      "loss": 0.5915,
+      "step": 4444
+    },
+    {
+      "epoch": 0.2731686332350049,
+      "grad_norm": 6.0625,
+      "learning_rate": 1.7743558334677594e-05,
+      "loss": 0.2619,
+      "step": 4445
+    },
+    {
+      "epoch": 0.27323008849557523,
+      "grad_norm": 21.0,
+      "learning_rate": 1.774208564802158e-05,
+      "loss": 0.3962,
+      "step": 4446
+    },
+    {
+      "epoch": 0.2732915437561455,
+      "grad_norm": 2.796875,
+      "learning_rate": 1.7740612542099054e-05,
+      "loss": 0.3811,
+      "step": 4447
+    },
+    {
+      "epoch": 0.27335299901671584,
+      "grad_norm": 8.9375,
+      "learning_rate": 1.7739139016989788e-05,
+      "loss": 0.3293,
+      "step": 4448
+    },
+    {
+      "epoch": 0.2734144542772861,
+      "grad_norm": 18.0,
+      "learning_rate": 1.773766507277358e-05,
+      "loss": 0.6534,
+      "step": 4449
+    },
+    {
+      "epoch": 0.27347590953785644,
+      "grad_norm": 15.1875,
+      "learning_rate": 1.773619070953025e-05,
+      "loss": 0.354,
+      "step": 4450
+    },
+    {
+      "epoch": 0.2735373647984267,
+      "grad_norm": 7.75,
+      "learning_rate": 1.7734715927339642e-05,
+      "loss": 0.4767,
+      "step": 4451
+    },
+    {
+      "epoch": 0.27359882005899705,
+      "grad_norm": 5.46875,
+      "learning_rate": 1.773324072628162e-05,
+      "loss": 0.5079,
+      "step": 4452
+    },
+    {
+      "epoch": 0.2736602753195674,
+      "grad_norm": 16.5,
+      "learning_rate": 1.7731765106436073e-05,
+      "loss": 0.2627,
+      "step": 4453
+    },
+    {
+      "epoch": 0.27372173058013766,
+      "grad_norm": 11.625,
+      "learning_rate": 1.7730289067882913e-05,
+      "loss": 0.2655,
+      "step": 4454
+    },
+    {
+      "epoch": 0.273783185840708,
+      "grad_norm": 32.5,
+      "learning_rate": 1.7728812610702076e-05,
+      "loss": 0.2862,
+      "step": 4455
+    },
+    {
+      "epoch": 0.27384464110127826,
+      "grad_norm": 34.0,
+      "learning_rate": 1.7727335734973512e-05,
+      "loss": 0.3525,
+      "step": 4456
+    },
+    {
+      "epoch": 0.2739060963618486,
+      "grad_norm": 13.125,
+      "learning_rate": 1.7725858440777204e-05,
+      "loss": 0.3135,
+      "step": 4457
+    },
+    {
+      "epoch": 0.27396755162241887,
+      "grad_norm": 13.875,
+      "learning_rate": 1.7724380728193152e-05,
+      "loss": 0.39,
+      "step": 4458
+    },
+    {
+      "epoch": 0.2740290068829892,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.7722902597301385e-05,
+      "loss": 0.2811,
+      "step": 4459
+    },
+    {
+      "epoch": 0.2740904621435595,
+      "grad_norm": 6.09375,
+      "learning_rate": 1.7721424048181943e-05,
+      "loss": 0.2941,
+      "step": 4460
+    },
+    {
+      "epoch": 0.2741519174041298,
+      "grad_norm": 4.53125,
+      "learning_rate": 1.7719945080914902e-05,
+      "loss": 0.3535,
+      "step": 4461
+    },
+    {
+      "epoch": 0.2742133726647001,
+      "grad_norm": 16.125,
+      "learning_rate": 1.771846569558035e-05,
+      "loss": 0.3268,
+      "step": 4462
+    },
+    {
+      "epoch": 0.2742748279252704,
+      "grad_norm": 34.0,
+      "learning_rate": 1.7716985892258404e-05,
+      "loss": 0.3882,
+      "step": 4463
+    },
+    {
+      "epoch": 0.2743362831858407,
+      "grad_norm": 18.125,
+      "learning_rate": 1.77155056710292e-05,
+      "loss": 0.3461,
+      "step": 4464
+    },
+    {
+      "epoch": 0.274397738446411,
+      "grad_norm": 11.8125,
+      "learning_rate": 1.7714025031972904e-05,
+      "loss": 0.2612,
+      "step": 4465
+    },
+    {
+      "epoch": 0.2744591937069813,
+      "grad_norm": 15.0,
+      "learning_rate": 1.7712543975169687e-05,
+      "loss": 0.2753,
+      "step": 4466
+    },
+    {
+      "epoch": 0.27452064896755163,
+      "grad_norm": 7.1875,
+      "learning_rate": 1.7711062500699766e-05,
+      "loss": 0.2669,
+      "step": 4467
+    },
+    {
+      "epoch": 0.2745821042281219,
+      "grad_norm": 6.40625,
+      "learning_rate": 1.7709580608643364e-05,
+      "loss": 0.3326,
+      "step": 4468
+    },
+    {
+      "epoch": 0.27464355948869223,
+      "grad_norm": 10.9375,
+      "learning_rate": 1.7708098299080732e-05,
+      "loss": 0.3,
+      "step": 4469
+    },
+    {
+      "epoch": 0.2747050147492625,
+      "grad_norm": 5.21875,
+      "learning_rate": 1.7706615572092146e-05,
+      "loss": 0.3731,
+      "step": 4470
+    },
+    {
+      "epoch": 0.27476647000983284,
+      "grad_norm": 6.15625,
+      "learning_rate": 1.7705132427757895e-05,
+      "loss": 0.3559,
+      "step": 4471
+    },
+    {
+      "epoch": 0.2748279252704032,
+      "grad_norm": 39.0,
+      "learning_rate": 1.7703648866158303e-05,
+      "loss": 0.31,
+      "step": 4472
+    },
+    {
+      "epoch": 0.27488938053097345,
+      "grad_norm": 13.25,
+      "learning_rate": 1.770216488737371e-05,
+      "loss": 0.3331,
+      "step": 4473
+    },
+    {
+      "epoch": 0.2749508357915438,
+      "grad_norm": 30.0,
+      "learning_rate": 1.770068049148448e-05,
+      "loss": 0.4853,
+      "step": 4474
+    },
+    {
+      "epoch": 0.27501229105211406,
+      "grad_norm": 35.25,
+      "learning_rate": 1.7699195678571e-05,
+      "loss": 0.3477,
+      "step": 4475
+    },
+    {
+      "epoch": 0.2750737463126844,
+      "grad_norm": 6.59375,
+      "learning_rate": 1.769771044871368e-05,
+      "loss": 0.2995,
+      "step": 4476
+    },
+    {
+      "epoch": 0.27513520157325466,
+      "grad_norm": 36.25,
+      "learning_rate": 1.7696224801992947e-05,
+      "loss": 0.3232,
+      "step": 4477
+    },
+    {
+      "epoch": 0.275196656833825,
+      "grad_norm": 19.0,
+      "learning_rate": 1.7694738738489255e-05,
+      "loss": 0.2853,
+      "step": 4478
+    },
+    {
+      "epoch": 0.27525811209439527,
+      "grad_norm": 1.7890625,
+      "learning_rate": 1.7693252258283088e-05,
+      "loss": 0.3147,
+      "step": 4479
+    },
+    {
+      "epoch": 0.2753195673549656,
+      "grad_norm": 20.5,
+      "learning_rate": 1.769176536145494e-05,
+      "loss": 0.2401,
+      "step": 4480
+    },
+    {
+      "epoch": 0.2753810226155359,
+      "grad_norm": 15.375,
+      "learning_rate": 1.7690278048085327e-05,
+      "loss": 0.2675,
+      "step": 4481
+    },
+    {
+      "epoch": 0.2754424778761062,
+      "grad_norm": 27.875,
+      "learning_rate": 1.7688790318254803e-05,
+      "loss": 0.2239,
+      "step": 4482
+    },
+    {
+      "epoch": 0.2755039331366765,
+      "grad_norm": 32.0,
+      "learning_rate": 1.7687302172043933e-05,
+      "loss": 0.2535,
+      "step": 4483
+    },
+    {
+      "epoch": 0.2755653883972468,
+      "grad_norm": 10.625,
+      "learning_rate": 1.7685813609533303e-05,
+      "loss": 0.2297,
+      "step": 4484
+    },
+    {
+      "epoch": 0.2756268436578171,
+      "grad_norm": 3.03125,
+      "learning_rate": 1.7684324630803523e-05,
+      "loss": 0.2774,
+      "step": 4485
+    },
+    {
+      "epoch": 0.2756882989183874,
+      "grad_norm": 13.8125,
+      "learning_rate": 1.7682835235935236e-05,
+      "loss": 0.3726,
+      "step": 4486
+    },
+    {
+      "epoch": 0.2757497541789577,
+      "grad_norm": 4.0625,
+      "learning_rate": 1.768134542500909e-05,
+      "loss": 0.3561,
+      "step": 4487
+    },
+    {
+      "epoch": 0.275811209439528,
+      "grad_norm": 2.3125,
+      "learning_rate": 1.767985519810577e-05,
+      "loss": 0.17,
+      "step": 4488
+    },
+    {
+      "epoch": 0.2758726647000983,
+      "grad_norm": 10.6875,
+      "learning_rate": 1.767836455530598e-05,
+      "loss": 0.332,
+      "step": 4489
+    },
+    {
+      "epoch": 0.27593411996066863,
+      "grad_norm": 11.9375,
+      "learning_rate": 1.7676873496690433e-05,
+      "loss": 0.4322,
+      "step": 4490
+    },
+    {
+      "epoch": 0.27599557522123896,
+      "grad_norm": 21.375,
+      "learning_rate": 1.767538202233989e-05,
+      "loss": 0.3272,
+      "step": 4491
+    },
+    {
+      "epoch": 0.27605703048180924,
+      "grad_norm": 15.0625,
+      "learning_rate": 1.767389013233511e-05,
+      "loss": 0.539,
+      "step": 4492
+    },
+    {
+      "epoch": 0.27611848574237957,
+      "grad_norm": 12.875,
+      "learning_rate": 1.767239782675689e-05,
+      "loss": 0.3205,
+      "step": 4493
+    },
+    {
+      "epoch": 0.27617994100294985,
+      "grad_norm": 6.90625,
+      "learning_rate": 1.767090510568605e-05,
+      "loss": 0.2875,
+      "step": 4494
+    },
+    {
+      "epoch": 0.2762413962635202,
+      "grad_norm": 9.0625,
+      "learning_rate": 1.7669411969203417e-05,
+      "loss": 0.28,
+      "step": 4495
+    },
+    {
+      "epoch": 0.27630285152409045,
+      "grad_norm": 13.5625,
+      "learning_rate": 1.7667918417389857e-05,
+      "loss": 0.234,
+      "step": 4496
+    },
+    {
+      "epoch": 0.2763643067846608,
+      "grad_norm": 32.5,
+      "learning_rate": 1.766642445032625e-05,
+      "loss": 0.5171,
+      "step": 4497
+    },
+    {
+      "epoch": 0.27642576204523106,
+      "grad_norm": 7.875,
+      "learning_rate": 1.76649300680935e-05,
+      "loss": 0.4516,
+      "step": 4498
+    },
+    {
+      "epoch": 0.2764872173058014,
+      "grad_norm": 29.0,
+      "learning_rate": 1.7663435270772537e-05,
+      "loss": 0.5245,
+      "step": 4499
+    },
+    {
+      "epoch": 0.27654867256637167,
+      "grad_norm": 53.5,
+      "learning_rate": 1.7661940058444305e-05,
+      "loss": 0.7653,
+      "step": 4500
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 15000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 200,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": false,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/models/Robometer-4B/training_args.bin b/models/Robometer-4B/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b5804c46a03b74d9918dd160704f567457604f0e
--- /dev/null
+++ b/models/Robometer-4B/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ee6d07f9f33f072bc6f199a3b4fb69371497b873dfd8b658a717513dff217b1
+size 5905
diff --git a/robometer/FINETUNE_ROBOMETER.md b/robometer/FINETUNE_ROBOMETER.md
new file mode 100644
index 0000000000000000000000000000000000000000..17f83ea99437bb3e95ae1b8fba7214e6d5b33ebb
--- /dev/null
+++ b/robometer/FINETUNE_ROBOMETER.md
@@ -0,0 +1,197 @@
+# Fine-tuning Robometer on Your Own Data
+
+Preprocess a dataset, LoRA fine-tune from Robometer-4B, upload to the Hub, run inference. Example: [MINT-SJTU/RoboFAC-dataset](https://huggingface.co/datasets/MINT-SJTU/RoboFAC-dataset).
+
+---
+
+## 1. Preprocessing (RoboFAC)
+
+Training needs a **preprocessed** cache from a HuggingFace dataset in RBM format. RoboFAC is folder-based; convert first, then preprocess.
+
+1. **Download** (into `ROBOMETER_DATASET_PATH`):
+   ```bash
+   export ROBOMETER_DATASET_PATH=/path/to/your/robometer_dataset
+   huggingface-cli download MINT-SJTU/RoboFAC-dataset --local-dir $ROBOMETER_DATASET_PATH/RoboFAC-dataset
+   ```
+
+2. **Convert and push to Hub** (required for recommended flow):
+   ```bash
+   export HF_TOKEN=your_token_here
+
+   uv run python -m dataset_upload.generate_hf_dataset \
+     --config_path dataset_upload/configs/data_gen_configs/robofac.yaml \
+     --dataset.dataset_path=$ROBOMETER_DATASET_PATH/RoboFAC-dataset \
+     --hub.push_to_hub=true \
+     --hub.hub_repo_id=robofac_rbm
+   ```
+
+3. **If you pushed:** download converted repo and set preprocess config:
+   ```bash
+   huggingface-cli download aliangdw/robofac_rbm --local-dir $ROBOMETER_DATASET_PATH/robofac_rbm
+   ```
+   In `preprocess_finetune.yaml`: `train_datasets: ["aliangdw/robofac_rbm"]`, `train_subsets: [["robofac"]]`.
+
+4. **Preprocess:**
+   ```bash
+   export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/save/processed_datasets
+
+   uv run python -m robometer.data.scripts.preprocess_datasets \
+     --config robometer/configs/preprocess_finetune.yaml \
+     --cache_dir=$ROBOMETER_PROCESSED_DATASETS_PATH
+   ```
+   Use the same path for training.
+
+**Other datasets:** RBM-style Hub datasets: set `train_datasets`/`train_subsets` and `ROBOMETER_DATASET_PATH` in the preprocess config, then run step 4. Raw data: add a loader (see [CustomDataset.md](dataset_upload/dataset_guides/CustomDataset.md)).
+
+---
+
+## 2. LoRA fine-tuning
+
+Use PEFT (LoRA) and `load_from_checkpoint` from a Qwen3-4B–based RBM checkpoint.
+
+```bash
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/your/processed_datasets
+
+uv run python train.py \
+  model.base_model_id=Qwen/Qwen3-VL-4B-Instruct \
+  model.use_peft=true \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  data.train_datasets=[aliangdw_robofac_rbm_robofac] \
+  data.eval_datasets=[mw] \
+  training.load_from_checkpoint=aliangdw/Robometer-4B \
+  training.per_device_train_batch_size=8 \
+  training.learning_rate=2e-5 \
+  training.warmup_ratio=0.1 \
+  training.weight_decay=0.01 \
+  training.max_steps=1000 \
+  training.output_dir=./logs \
+  training.exp_name=robometer4b_lora_robofac_2 \
+  logging.log_to=[wandb] \
+  custom_eval.eval_types=[reward_alignment,policy_ranking] \
+  custom_eval.reward_alignment=[aliangdw_robofac_rbm_robofac] \
+  custom_eval.policy_ranking=[aliangdw_robofac_rbm_robofac] \
+  logging.save_best.metric_names=[eval_rew_align/pearson_robofac,eval_p_rank/kendall_last_robofac] \
+  logging.save_best.greater_is_better=[true,true] \
+  training.overwrite_output_dir=True \
+  training.eval_steps=50 \
+  training.custom_eval_steps=50
+```
+
+The short name `robofac` is defined in `name_mapping.py` for `aliangdw_robofac_rbm_robofac`.
+
+**Tunable LoRA / training:** Override `training.learning_rate`, `training.warmup_ratio`, `training.weight_decay`, `training.gradient_accumulation_steps`, or `training.max_steps` as needed. Defaults above match `robometer/configs/config.yaml`. Multi-GPU: `uv run accelerate launch --config_file robometer/configs/distributed/fsdp.yaml train.py ...` (same overrides).
+
+### Full fine-tuning (no PEFT)
+
+Load the same checkpoint but train the full model (no LoRA). Uses more memory; lower `per_device_train_batch_size` or gradient accumulation if needed.
+
+```bash
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/your/processed_datasets
+
+uv run python train.py \
+  model.base_model_id=Qwen/Qwen3-VL-4B-Instruct \
+  model.use_peft=false \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  data.train_datasets=[aliangdw_robofac_rbm_robofac] \
+  data.eval_datasets=[aliangdw_robofac_rbm_robofac] \
+  training.load_from_checkpoint=aliangdw/Robometer-4B \
+  training.per_device_train_batch_size=8 \
+  training.learning_rate=2e-5 \
+  training.warmup_ratio=0.1 \
+  training.weight_decay=0.01 \
+  training.gradient_accumulation_steps=1 \
+  training.max_steps=500 \
+  training.output_dir=./logs \
+  training.exp_name=robometer4b_full_robofac \
+  logging.log_to=[wandb] \
+  custom_eval.reward_alignment=[aliangdw_robofac_rbm_robofac] \
+  custom_eval.policy_ranking=[aliangdw_robofac_rbm_robofac] \
+  logging.save_best.metric_names=[eval_rew_align/pearson_robofac,eval_p_rank/kendall_last_robofac] \
+  logging.save_best.greater_is_better=[true,true] \
+  training.eval_steps=50 \
+  training.custom_eval_steps=50
+```
+
+
+---
+
+## 3. Upload to Hub
+
+```bash
+uv run python robometer/utils/upload_to_hub.py \
+  --model_dir ./logs/robometer4b_lora_robofac/checkpoint-500 \
+  --hub_model_id aliangdw/robometer-4b-lora-robofac \
+  --base_model "Qwen/Qwen3-VL-4B-Instruct" \
+  --commit_message "LoRA fine-tune on RoboFAC"
+```
+
+Or enable `logging.save_best.upload_to_hub: true` in config for upload during training.
+
+---
+
+## 4. Inference
+
+```bash
+uv run python scripts/example_inference_local.py \
+  --model-path aliangdw/robometer-4b-lora-robofac \
+  --video /path/to/video.mp4 \
+  --task "Insert the cylinder"
+```
+
+Server: `uv run python robometer/evals/eval_server.py ... model_path=aliangdw/robometer-4b-lora-robofac`. Eval: `run_baseline_eval.py` with `reward_model=rbm`, `model_path=...` (see [README](README.md)).
+
+---
+
+## 5. Baseline: Fine-tune from base Qwen-VL (no Robometer checkpoint)
+
+For comparison, run the same `train.py` on the same data but **without** loading a Robometer checkpoint. Training starts from the base Qwen-VL plus randomly initialized progress/preference heads.
+
+```bash
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/your/processed_datasets
+
+uv run python train.py \
+  model.base_model_id=Qwen/Qwen3-VL-4B-Instruct \
+  model.use_peft=true \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  data.train_datasets=[aliangdw_robofac_rbm_robofac] \
+  data.eval_datasets=[aliangdw_robofac_rbm_robofac] \
+  training.per_device_train_batch_size=8 \
+  training.learning_rate=2e-5 \
+  training.warmup_ratio=0.1 \
+  training.weight_decay=0.01 \
+  training.max_steps=1000 \
+  training.output_dir=./logs \
+  training.exp_name=qwen3vl_lora_robofac_baseline \
+  logging.log_to=[wandb] \
+  custom_eval.reward_alignment=[aliangdw_robofac_rbm_robofac] \
+  custom_eval.policy_ranking=[aliangdw_robofac_rbm_robofac] \
+  logging.save_best.metric_names=[eval_rew_align/pearson_robofac,eval_p_rank/kendall_last_robofac] \
+  logging.save_best.greater_is_better=[true,true] \
+  training.eval_steps=50 \
+  training.custom_eval_steps=50
+
+uv run python train.py \
+  model.base_model_id=Qwen/Qwen3-VL-4B-Instruct \
+  model.use_peft=false \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  data.train_datasets=[aliangdw_robofac_rbm_robofac] \
+  data.eval_datasets=[aliangdw_robofac_rbm_robofac] \
+  training.per_device_train_batch_size=8 \
+  training.learning_rate=2e-5 \
+  training.warmup_ratio=0.1 \
+  training.weight_decay=0.01 \
+  training.max_steps=1000 \
+  training.output_dir=./logs \
+  training.exp_name=qwen3vl_robofac_baseline \
+  logging.log_to=[wandb] \
+  custom_eval.reward_alignment=[aliangdw_robofac_rbm_robofac] \
+  custom_eval.policy_ranking=[aliangdw_robofac_rbm_robofac] \
+  logging.save_best.metric_names=[eval_rew_align/pearson_robofac,eval_p_rank/kendall_last_robofac] \
+  logging.save_best.greater_is_better=[true,true] \
+  training.eval_steps=50 \
+  training.custom_eval_steps=50
+```
diff --git a/robometer/README.md b/robometer/README.md
new file mode 100755
index 0000000000000000000000000000000000000000..0a8318d7c06abae31be4b1599551e626fde14ab7
--- /dev/null
+++ b/robometer/README.md
@@ -0,0 +1,267 @@
+# Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons
+
+[![arXiv](https://img.shields.io/badge/arXiv-2603.02115-b31b1b.svg)](https://arxiv.org/abs/2603.02115)
+[![GitHub](https://img.shields.io/badge/GitHub-robometer-181717?logo=github)](https://github.com/robometer/robometer)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+[![Model](https://img.shields.io/badge/Model-FFD21E?logo=huggingface)](https://huggingface.co/aliangdw/Robometer-4B)
+[![Dataset](https://img.shields.io/badge/Dataset-RBM--1M-FFD21E?logo=huggingface)](https://huggingface.co/datasets/)
+[![RBM-1M Visualizer](https://img.shields.io/badge/Visualizer-RBM--FFD21E?logo=huggingface)](https://huggingface.co/spaces/rewardfm/visualizer)
+[![RewardEval UI](https://img.shields.io/badge/%20RewardEval%20UI-FFD21E?logo=huggingface)](https://huggingface.co/spaces/rewardfm/rewardeval_ui)
+
+<p align="center">
+  <img src="assets/robometer.jpg" alt="Robometer" width="100%"/>
+</p>
+
+---
+
+## Abstract
+
+General-purpose robot reward models are typically trained to predict absolute task progress from expert demonstrations, providing only local, frame-level supervision. While effective for expert demonstrations, this paradigm scales poorly to large-scale robotics datasets where failed and suboptimal trajectories are abundant and assigning dense progress labels is ambiguous. We introduce **Robometer**, a scalable reward modeling framework that combines intra-trajectory progress supervision with inter-trajectory preference supervision. Robometer is trained with a dual objective: a frame-level progress loss that anchors reward magnitude on expert data, and a trajectory-comparison preference loss that imposes global ordering constraints across trajectories of the same task, enabling effective learning from both real and augmented failed trajectories. To support this formulation at scale, we curate **RBM-1M**, a reward-learning dataset comprising over one million trajectories spanning diverse robot embodiments and tasks, including substantial suboptimal and failure data. Across benchmarks and real-world evaluations, Robometer learns more generalizable reward functions than prior methods and improves robot learning performance across a diverse set of downstream applications.
+
+---
+
+## 📦 Package structure
+
+```
+robometer/
+├── robometer/              # Main package
+│   ├── data/               # Datasets and preprocessing
+│   ├── configs/            # Hydra and experiment configs
+│   ├── models/             # Model definitions
+│   └── evals/              # Baseline evals (GVL, VLAC, Robodopamine, etc.)
+├── eval_commands/          # Shell scripts for baseline evals
+├── train.py                # Training entrypoint
+└── pyproject.toml          # Dependencies (uv)
+```
+
+---
+
+## 🛠️ Setup
+
+### Prerequisites
+
+- Git, Python 3.10+
+- NVIDIA drivers (GPU)
+- [uv](https://github.com/astral-sh/uv#installation) (recommended)
+
+### Install (main env)
+
+```bash
+git clone https://github.com/aliang8/robometer.git
+cd robometer
+
+# Create venv and install
+uv sync
+```
+
+### Dataset setup
+
+```bash
+hf auth
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/save/processed_datasets
+./scripts/download_processed_datasets.sh
+./scripts/untar_processed_datasets.sh
+```
+
+For raw download and preprocessing, see [📥 Download raw datasets](#-download-raw-datasets-optional) below.
+
+---
+
+## 🔍 Inference
+
+Inference runs a **pretrained RBM model** on your own videos to get per-frame progress, per-frame success, and (for two trajectories) preference scores.
+
+**Pretrained models (Hugging Face):**
+
+- **[Robometer-4B](https://huggingface.co/robometer/Robometer-4B)** — general-purpose, trained on RBM-1M
+- ~~**Robometer-4B-LIBERO** — LIBERO-10 / Spatial / Object / Goal~~ removed because the standard Robometer model is already trained on LIBERO 10/Spatial/Object/Goal+failures and simply performs better than the version trained exclusively on LIBERO
+
+### Inference via HTTP server
+
+Start the eval server on your machine, then call it with a video and task:
+
+```bash
+uv run python robometer/evals/eval_server.py \
+  server_url=0.0.0.0 \
+  server_port=8000
+```
+
+Then run the client (no robometer dependency):
+
+```bash
+# SOAR
+uv run python scripts/example_inference.py \
+  --eval-server-url http://localhost:8000 \
+  --video scripts/example_videos/soar_put_green_stick_in_brown_bowl.mp4 \
+  --task "Put green stick in brown bowl" \
+  --fps 3
+
+# Berkeley RPT (Wrist)
+uv run python scripts/example_inference.py \
+  --eval-server-url http://localhost:8000 \
+  --video scripts/example_videos/berkeley_rpt_stack_cup.mp4 \
+  --task "Pick up the yellow cup and stack it on the other cup" \
+  --fps 3
+
+# Your own video
+uv run python scripts/example_inference.py \
+  --eval-server-url http://localhost:8000 \
+  --video /path/to/video.mp4 \
+  --task "your task description"
+```
+
+To run the model locally (loads checkpoint from Hugging Face, no server):
+
+```bash
+uv run python scripts/example_inference_local.py \
+  --model-path aliangdw/Robometer-4B \
+  --video /path/to/video.mp4 \
+  --task "your task description"
+```
+
+Use `aliangdw/Robometer-4B-LIBERO` for the LIBERO-finetuned model.
+
+---
+
+## 🏋️ Training
+
+### Training
+
+**Train on RBM-1M in-distribution and evaluate on RBM-1M-OOD**
+
+```bash
+uv run accelerate launch --config_file robometer/configs/distributed/fsdp.yaml train.py \
+  data.train_datasets=[rbm-1m-id] \
+  data.eval_datasets=[rbm-1m-ood] \
+  data.max_frames=4 \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  training.max_steps=5000 \
+  custom_eval.reward_alignment=[rbm-1m-ood] \
+  custom_eval.policy_ranking=[rbm-1m-ood] \
+  custom_eval.confusion_matrix=[rbm-1m-ood]
+```
+
+**LIBERO: train on 10 / object / spatial / goal, test on 90.**
+
+```bash
+uv run accelerate launch --config_file robometer/configs/distributed/fsdp.yaml train.py \
+  data.train_datasets=[libero_pi0] \
+  data.eval_datasets=[mw] \
+  data.max_frames=4 \
+  model.train_progress_head=true \
+  model.train_preference_head=true \
+  training.max_steps=5000 \
+  custom_eval.reward_alignment=[libero_pi0] \
+  custom_eval.policy_ranking=[libero_pi0]
+```
+
+See `robometer/configs/experiment_configs.py` for more config options.
+
+---
+
+## 🔧 LoRA fine-tune Robometer for new dataset
+
+Preprocess a new dataset, LoRA fine-tune from **Robometer-4B** on your own data, upload the model to the Hub, and run inference:
+
+- **Preprocessing:** Add your dataset to the preprocess config and run the preprocessor; for raw videos (e.g. [MINT-SJTU/RoboFAC-dataset](https://huggingface.co/datasets/MINT-SJTU/RoboFAC-dataset)), convert to RBM format first via `dataset_upload`, then preprocess.
+- **Fine-tuning:** Set `model.use_peft=true` and `training.resume_from_checkpoint=aliangdw/Robometer-4B`, then train on your dataset.
+- **Upload & inference:** Use `robometer/utils/upload_to_hub.py` to push checkpoints; run `scripts/example_inference_local.py` with your Hub model.
+
+Full step-by-step: **[FINETUNE_ROBOMETER.md](FINETUNE_ROBOMETER.md)**.
+
+---
+
+## 📊 Evaluation
+
+Evaluation runs **benchmark evals** (reward alignment, policy ranking, confusion matrix) on fixed datasets to measure model quality. Use this to reproduce paper results or compare checkpoints.
+
+### Robometer evaluation
+
+Run RBM with `reward_model=rbm`; override `model_path` and `custom_eval.*` as needed. See `eval_commands/*.sh` for ReWIND, Robo-Dopamine, VLAC, RoboReward.
+
+**Reward alignment**
+
+```bash
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=true \
+    custom_eval.subsample_n_frames=5 \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    max_frames=4 \
+    model_config.batch_size=32
+```
+
+**Policy ranking**
+
+```bash
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.num_examples_per_quality_pr=1000 \
+    max_frames=4 \
+    model_config.batch_size=32
+```
+
+**Confusion matrix**
+
+```bash
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=4 \
+    model_config.batch_size=32
+```
+
+Details: [robometer/evals/README.md](robometer/evals/README.md).
+
+### Baseline evaluation (all models)
+
+- **RBM:** use the [reward alignment](#robometer-evaluation), [policy ranking](#robometer-evaluation), or [confusion matrix](#robometer-evaluation) commands above; set `model_path` to your checkpoint.
+- **ReWIND, Robo-Dopamine, VLAC, RoboReward:** see [robometer/evals/README.md](robometer/evals/README.md) and `eval_commands/reward_alignment.sh`, `eval_commands/policy_ranking.sh`, `eval_commands/confusion_matrix.sh`. For Robo-Dopamine use `.venv-robodopamine/bin/python` (vLLM) instead of `uv run`.
+
+---
+
+## 📊 Dataset generation
+
+Supported: **AgiBotWorld** (streaming), **LIBERO** (HDF5), and custom configs.
+
+```bash
+# AgiBotWorld
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen_configs/agibot_world.yaml
+
+# LIBERO
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen.yaml \
+  --dataset.dataset_path=LIBERO/libero/datasets/libero_90 --dataset.dataset_name=libero_90
+```
+
+See dataset_upload README and dataset_guides for adding datasets.
+
+---
+
+## 📥 Download raw datasets (optional)
+
+If you prefer not to use the processed datasets:
+
+```bash
+export ROBOMETER_DATASET_PATH=/path/to/your/robometer_dataset
+./scripts/download_data.sh
+
+# Preprocess
+uv run python -m robometer.data.scripts.preprocess_datasets --config robometer/configs/preprocess.yaml
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/save/processed_datasets
+```
+
+---
+
+## 📑 License
+
+This project is licensed under the [MIT License](LICENSE).
diff --git a/robometer/__init__.py b/robometer/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/robometer/assets/robometer.jpg b/robometer/assets/robometer.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c3f088f67a9e60807efac78fc4a53210934f2647
--- /dev/null
+++ b/robometer/assets/robometer.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fca3bde9c8c4de9d7c82809f0a5dbeb23ab49f705ea7a22439a64ab807be4e63
+size 6490314
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/all_metrics.json b/robometer/baseline_eval_output/rbm_smoke_10/all_metrics.json
new file mode 100644
index 0000000000000000000000000000000000000000..f5844da6f6d639205fe019eb090df213cb2ff201
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/all_metrics.json
@@ -0,0 +1,6 @@
+{
+  "reward_alignment": {
+    "racer_val/loss": 0.04008283880995868,
+    "racer_val/pearson": 0.9235551377640896
+  }
+}
\ No newline at end of file
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/metrics.json b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/metrics.json
new file mode 100644
index 0000000000000000000000000000000000000000..8d16345f01e36e755427884cd2a271ca4e1232f4
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/metrics.json
@@ -0,0 +1,4 @@
+{
+  "racer_val/loss": 0.04008283880995868,
+  "racer_val/pearson": 0.9235551377640896
+}
\ No newline at end of file
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0000.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0000.gif
new file mode 100644
index 0000000000000000000000000000000000000000..1f1324ce00825ad9a5631720925d479c50845faa
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0000.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a119138c6583f04c85147dad18d6b57eada5a7bcf3a7ca9e539bc6f7680a4efc
+size 556143
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0001.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0001.gif
new file mode 100644
index 0000000000000000000000000000000000000000..532b1086caf085d8f9072d5cffb9481a43c16bd7
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0001.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cd7d980c81c004afed276cffdcf4f31ea2adabb752a3a576159da379df0c084
+size 314007
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0002.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0002.gif
new file mode 100644
index 0000000000000000000000000000000000000000..5cbae66defdb5f9d70ff8886ae1ad13255eb8d61
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0002.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1932585e1bbfbb18788e4ddb71cbd0eef575872872f71481019a6ad5739cdcb
+size 642086
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0003.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0003.gif
new file mode 100644
index 0000000000000000000000000000000000000000..9945b793f91c3c5909bb2898899cceae1e8f976b
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0003.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d1df6c6c26a7c12a74c7780d0fb03b72b7d95052b41d0357d875a8e99cbea8b
+size 418957
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0004.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0004.gif
new file mode 100644
index 0000000000000000000000000000000000000000..dd43acf90bd809c2502c351a8e14c4fbed30642d
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0004.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ec8842263f1ea60568d22799f20425159610f2a8746a034b98ec16ee42262b7
+size 547677
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0005.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0005.gif
new file mode 100644
index 0000000000000000000000000000000000000000..17727a3d598aa1b7089935c74710fee5f4c3c91b
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0005.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:faf3314c58a915f859e8e70f53b9c5d2af47a69d3a7e4b67dd44dfff7c478c6f
+size 242406
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0006.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0006.gif
new file mode 100644
index 0000000000000000000000000000000000000000..0525c33385628972359738f6f682266cf8ffb023
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0006.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a01a12edb1186f2e30afc0badcb352f8a47b96ed9a3f8a0d7e1665e469ada2d
+size 337208
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0007.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0007.gif
new file mode 100644
index 0000000000000000000000000000000000000000..a4e3fe3ca7236272056166d273cb48d0a5285914
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0007.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:741d17d2b7d27428ed64c6b4e20490e7131f57ee8be382b4f700b0da74c80db2
+size 359267
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0008.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0008.gif
new file mode 100644
index 0000000000000000000000000000000000000000..5e6837435f75e2fa044fd5592a92f3abf4131b55
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0008.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9218ca88f3ec5445121565ef3d4924b9b6c73085a49cef0be7e0818a8112c330
+size 348731
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0009.gif b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0009.gif
new file mode 100644
index 0000000000000000000000000000000000000000..05ef97bc357fda64835d813ee682b48881fde53e
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_plots/trajectory_0009.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a623f8759aec25d8f35e10f6d6d425abf4d82ac55a7f632c091c58faf9554f4a
+size 352088
diff --git a/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_results.json b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..6613957229cb954386c315cfe66d415b06eb953d
--- /dev/null
+++ b/robometer/baseline_eval_output/rbm_smoke_10/reward_alignment/racer_val_results.json
@@ -0,0 +1,1402 @@
+[
+  {
+    "progress_pred": [
+      0.0791015625,
+      0.1484375,
+      0.193359375,
+      0.1845703125
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+      "frame_step": 0,
+      "num_frames": 12
+    },
+    "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.08333333333333333,
+      0.08333333333333333,
+      0.08333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.0791015625,
+      0.1728515625,
+      0.27734375,
+      0.357421875
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+      "frame_step": 2,
+      "num_frames": 12
+    },
+    "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.16666666666666666,
+      0.25,
+      0.25
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.0791015625,
+      0.1943359375,
+      0.30078125,
+      0.373046875
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+      "frame_step": 5,
+      "num_frames": 12
+    },
+    "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.25,
+      0.3333333333333333,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.0791015625,
+      0.19921875,
+      0.298828125,
+      0.419921875
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+      "frame_step": 8,
+      "num_frames": 12
+    },
+    "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.3333333333333333,
+      0.5,
+      0.75
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.0791015625,
+      0.2138671875,
+      0.302734375,
+      0.70703125
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+      "frame_step": 11,
+      "num_frames": 12
+    },
+    "id": "9ebf31b8-36f2-4595-911f-7c1385e5c7d4",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_9ebf31b8-36f2-4595-911f-7c1385e5c7d4.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.4166666666666667,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08642578125,
+      0.1484375,
+      0.181640625,
+      0.17578125
+    ],
+    "task": "stack the other cups on top of the cyan cup",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+      "frame_step": 0,
+      "num_frames": 11
+    },
+    "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.09090909090909091,
+      0.09090909090909091,
+      0.09090909090909091,
+      0.09090909090909091
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08642578125,
+      0.1875,
+      0.302734375,
+      0.388671875
+    ],
+    "task": "stack the other cups on top of the cyan cup",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+      "frame_step": 2,
+      "num_frames": 11
+    },
+    "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.09090909090909091,
+      0.18181818181818182,
+      0.2727272727272727,
+      0.2727272727272727
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08642578125,
+      0.2001953125,
+      0.326171875,
+      0.43359375
+    ],
+    "task": "stack the other cups on top of the cyan cup",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+      "frame_step": 5,
+      "num_frames": 11
+    },
+    "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.09090909090909091,
+      0.2727272727272727,
+      0.36363636363636365,
+      0.5454545454545454
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08642578125,
+      0.2001953125,
+      0.326171875,
+      0.4453125
+    ],
+    "task": "stack the other cups on top of the cyan cup",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+      "frame_step": 7,
+      "num_frames": 11
+    },
+    "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.09090909090909091,
+      0.2727272727272727,
+      0.5454545454545454,
+      0.7272727272727273
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08642578125,
+      0.2255859375,
+      0.349609375,
+      0.69921875
+    ],
+    "task": "stack the other cups on top of the cyan cup",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+      "frame_step": 10,
+      "num_frames": 11
+    },
+    "id": "515bc90f-8bce-4a96-9f20-9242110cd37f",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_515bc90f-8bce-4a96-9f20-9242110cd37f.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.09090909090909091,
+      0.36363636363636365,
+      0.7272727272727273,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.24609375,
+      0.36328125,
+      0.47265625,
+      0.52734375
+    ],
+    "task": "take the chicken off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "2d4db549-4729-4984-b725-040b4bed560e",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+      "frame_step": 0,
+      "num_frames": 6
+    },
+    "id": "2d4db549-4729-4984-b725-040b4bed560e",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.24609375,
+      0.4140625,
+      0.578125,
+      0.6484375
+    ],
+    "task": "take the chicken off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "2d4db549-4729-4984-b725-040b4bed560e",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+      "frame_step": 1,
+      "num_frames": 6
+    },
+    "id": "2d4db549-4729-4984-b725-040b4bed560e",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.3333333333333333,
+      0.3333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.24609375,
+      0.4140625,
+      0.53515625,
+      0.64453125
+    ],
+    "task": "take the chicken off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "2d4db549-4729-4984-b725-040b4bed560e",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+      "frame_step": 2,
+      "num_frames": 6
+    },
+    "id": "2d4db549-4729-4984-b725-040b4bed560e",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.24609375,
+      0.4140625,
+      0.53515625,
+      0.62890625
+    ],
+    "task": "take the chicken off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "2d4db549-4729-4984-b725-040b4bed560e",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+      "frame_step": 3,
+      "num_frames": 6
+    },
+    "id": "2d4db549-4729-4984-b725-040b4bed560e",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.6666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.24609375,
+      0.404296875,
+      0.52734375,
+      0.76953125
+    ],
+    "task": "take the chicken off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "2d4db549-4729-4984-b725-040b4bed560e",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+      "frame_step": 5,
+      "num_frames": 6
+    },
+    "id": "2d4db549-4729-4984-b725-040b4bed560e",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_2d4db549-4729-4984-b725-040b4bed560e.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.5,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.39453125,
+      0.51953125,
+      0.59375,
+      0.609375
+    ],
+    "task": "take the steak off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+      "frame_step": 0,
+      "num_frames": 6
+    },
+    "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.390625,
+      0.515625,
+      0.61328125,
+      0.6328125
+    ],
+    "task": "take the steak off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+      "frame_step": 1,
+      "num_frames": 6
+    },
+    "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.3333333333333333,
+      0.3333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.390625,
+      0.515625,
+      0.5546875,
+      0.60546875
+    ],
+    "task": "take the steak off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+      "frame_step": 2,
+      "num_frames": 6
+    },
+    "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.390625,
+      0.515625,
+      0.5546875,
+      0.60546875
+    ],
+    "task": "take the steak off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+      "frame_step": 3,
+      "num_frames": 6
+    },
+    "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.6666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.390625,
+      0.494140625,
+      0.5625,
+      0.56640625
+    ],
+    "task": "take the steak off the grill",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+      "frame_step": 5,
+      "num_frames": 6
+    },
+    "id": "20a606e5-0edc-4b09-9c85-9f9c15cf63c2",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_20a606e5-0edc-4b09-9c85-9f9c15cf63c2.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.5,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.14453125,
+      0.205078125,
+      0.26171875,
+      0.26171875
+    ],
+    "task": "put the sugar in the cupboard",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+      "frame_step": 0,
+      "num_frames": 6
+    },
+    "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.14453125,
+      0.30078125,
+      0.5234375,
+      0.57421875
+    ],
+    "task": "put the sugar in the cupboard",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+      "frame_step": 1,
+      "num_frames": 6
+    },
+    "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.3333333333333333,
+      0.3333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.14453125,
+      0.30078125,
+      0.53125,
+      0.65625
+    ],
+    "task": "put the sugar in the cupboard",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+      "frame_step": 2,
+      "num_frames": 6
+    },
+    "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.14453125,
+      0.30078125,
+      0.53125,
+      0.6015625
+    ],
+    "task": "put the sugar in the cupboard",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+      "frame_step": 3,
+      "num_frames": 6
+    },
+    "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.6666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1474609375,
+      0.28515625,
+      0.421875,
+      0.53515625
+    ],
+    "task": "put the sugar in the cupboard",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+      "frame_step": 5,
+      "num_frames": 6
+    },
+    "id": "f1ab4e1d-e577-42d0-888a-4c4997a386b6",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_f1ab4e1d-e577-42d0-888a-4c4997a386b6.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.5,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1279296875,
+      0.1943359375,
+      0.255859375,
+      0.333984375
+    ],
+    "task": "put the ring on the gray spoke",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+      "frame_step": 0,
+      "num_frames": 6
+    },
+    "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1279296875,
+      0.255859375,
+      0.3984375,
+      0.515625
+    ],
+    "task": "put the ring on the gray spoke",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+      "frame_step": 1,
+      "num_frames": 6
+    },
+    "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.3333333333333333,
+      0.3333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1279296875,
+      0.255859375,
+      0.388671875,
+      0.64453125
+    ],
+    "task": "put the ring on the gray spoke",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+      "frame_step": 2,
+      "num_frames": 6
+    },
+    "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1279296875,
+      0.255859375,
+      0.388671875,
+      0.546875
+    ],
+    "task": "put the ring on the gray spoke",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+      "frame_step": 3,
+      "num_frames": 6
+    },
+    "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.6666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.1279296875,
+      0.3203125,
+      0.451171875,
+      0.640625
+    ],
+    "task": "put the ring on the gray spoke",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+      "frame_step": 5,
+      "num_frames": 6
+    },
+    "id": "84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_84b053f7-94d5-43e7-9aa3-a4c6a63d6bfd.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.5,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08984375,
+      0.1552734375,
+      0.2119140625,
+      0.251953125
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "fb467e10-15a5-4654-850e-a021549faf10",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+      "frame_step": 0,
+      "num_frames": 12
+    },
+    "id": "fb467e10-15a5-4654-850e-a021549faf10",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.08333333333333333,
+      0.08333333333333333,
+      0.08333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08984375,
+      0.2099609375,
+      0.3046875,
+      0.466796875
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "fb467e10-15a5-4654-850e-a021549faf10",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+      "frame_step": 2,
+      "num_frames": 12
+    },
+    "id": "fb467e10-15a5-4654-850e-a021549faf10",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.16666666666666666,
+      0.25,
+      0.25
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08984375,
+      0.1982421875,
+      0.328125,
+      0.396484375
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "fb467e10-15a5-4654-850e-a021549faf10",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+      "frame_step": 5,
+      "num_frames": 12
+    },
+    "id": "fb467e10-15a5-4654-850e-a021549faf10",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.25,
+      0.3333333333333333,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08984375,
+      0.20703125,
+      0.32421875,
+      0.400390625
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "fb467e10-15a5-4654-850e-a021549faf10",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+      "frame_step": 8,
+      "num_frames": 12
+    },
+    "id": "fb467e10-15a5-4654-850e-a021549faf10",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.3333333333333333,
+      0.5,
+      0.75
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.08984375,
+      0.220703125,
+      0.296875,
+      0.69921875
+    ],
+    "task": "put the item in the bottom drawer",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "fb467e10-15a5-4654-850e-a021549faf10",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+      "frame_step": 11,
+      "num_frames": 12
+    },
+    "id": "fb467e10-15a5-4654-850e-a021549faf10",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_fb467e10-15a5-4654-850e-a021549faf10.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.08333333333333333,
+      0.4166666666666667,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.123046875,
+      0.142578125,
+      0.1708984375,
+      0.1591796875
+    ],
+    "task": "screw in the lime light bulb",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+      "frame_step": 0,
+      "num_frames": 8
+    },
+    "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.125,
+      0.125,
+      0.125,
+      0.125
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.123046875,
+      0.1875,
+      0.296875,
+      0.294921875
+    ],
+    "task": "screw in the lime light bulb",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+      "frame_step": 1,
+      "num_frames": 8
+    },
+    "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.125,
+      0.25,
+      0.25,
+      0.25
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.123046875,
+      0.1875,
+      0.306640625,
+      0.291015625
+    ],
+    "task": "screw in the lime light bulb",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+      "frame_step": 3,
+      "num_frames": 8
+    },
+    "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.125,
+      0.25,
+      0.375,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.123046875,
+      0.251953125,
+      0.275390625,
+      0.359375
+    ],
+    "task": "screw in the lime light bulb",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+      "frame_step": 5,
+      "num_frames": 8
+    },
+    "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.125,
+      0.375,
+      0.5,
+      0.75
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.123046875,
+      0.251953125,
+      0.412109375,
+      0.474609375
+    ],
+    "task": "screw in the lime light bulb",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+      "frame_step": 7,
+      "num_frames": 8
+    },
+    "id": "42862408-7781-4a2f-b68c-c6b17ff5b785",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_42862408-7781-4a2f-b68c-c6b17ff5b785.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.125,
+      0.375,
+      0.75,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.11279296875,
+      0.1767578125,
+      0.2734375,
+      0.34765625
+    ],
+    "task": "stack the wine bottle to the right of the rack",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+      "frame_step": 0,
+      "num_frames": 6
+    },
+    "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.11279296875,
+      0.2578125,
+      0.404296875,
+      0.50390625
+    ],
+    "task": "stack the wine bottle to the right of the rack",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+      "frame_step": 1,
+      "num_frames": 6
+    },
+    "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.3333333333333333,
+      0.3333333333333333
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.11279296875,
+      0.2578125,
+      0.375,
+      0.51171875
+    ],
+    "task": "stack the wine bottle to the right of the rack",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+      "frame_step": 2,
+      "num_frames": 6
+    },
+    "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.5
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.11279296875,
+      0.2578125,
+      0.375,
+      0.478515625
+    ],
+    "task": "stack the wine bottle to the right of the rack",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+      "frame_step": 3,
+      "num_frames": 6
+    },
+    "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.3333333333333333,
+      0.5,
+      0.6666666666666666
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.11279296875,
+      0.2373046875,
+      0.341796875,
+      0.8046875
+    ],
+    "task": "stack the wine bottle to the right of the rack",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+      "frame_step": 5,
+      "num_frames": 6
+    },
+    "id": "adf6bf99-b3ee-4fa1-9467-7d3ba9c88994",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_adf6bf99-b3ee-4fa1-9467-7d3ba9c88994.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.16666666666666666,
+      0.5,
+      0.6666666666666666,
+      1.0
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.2080078125,
+      0.298828125,
+      0.43359375,
+      0.482421875
+    ],
+    "task": "use the stick to drag the cube onto the black target",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "88049043-6468-4057-a837-85781c5af0a8",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+      "frame_step": 0,
+      "num_frames": 7
+    },
+    "id": "88049043-6468-4057-a837-85781c5af0a8",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.14285714285714285,
+      0.14285714285714285,
+      0.14285714285714285,
+      0.14285714285714285
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.2080078125,
+      0.3515625,
+      0.5390625,
+      0.609375
+    ],
+    "task": "use the stick to drag the cube onto the black target",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "88049043-6468-4057-a837-85781c5af0a8",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+      "frame_step": 1,
+      "num_frames": 7
+    },
+    "id": "88049043-6468-4057-a837-85781c5af0a8",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.14285714285714285,
+      0.2857142857142857,
+      0.2857142857142857,
+      0.2857142857142857
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.2080078125,
+      0.3515625,
+      0.64453125,
+      0.6484375
+    ],
+    "task": "use the stick to drag the cube onto the black target",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "88049043-6468-4057-a837-85781c5af0a8",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+      "frame_step": 3,
+      "num_frames": 7
+    },
+    "id": "88049043-6468-4057-a837-85781c5af0a8",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.14285714285714285,
+      0.2857142857142857,
+      0.42857142857142855,
+      0.5714285714285714
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.208984375,
+      0.349609375,
+      0.494140625,
+      0.48828125
+    ],
+    "task": "use the stick to drag the cube onto the black target",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "88049043-6468-4057-a837-85781c5af0a8",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+      "frame_step": 4,
+      "num_frames": 7
+    },
+    "id": "88049043-6468-4057-a837-85781c5af0a8",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.14285714285714285,
+      0.2857142857142857,
+      0.5714285714285714,
+      0.7142857142857143
+    ],
+    "quality_label": "successful"
+  },
+  {
+    "progress_pred": [
+      0.208984375,
+      0.578125,
+      0.53515625,
+      0.78515625
+    ],
+    "task": "use the stick to drag the cube onto the black target",
+    "data_source": "racer_val",
+    "data_gen_strategy": null,
+    "metadata": {
+      "id": "88049043-6468-4057-a837-85781c5af0a8",
+      "data_gen_strategy": "reward_alignment",
+      "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+      "frame_step": 6,
+      "num_frames": 7
+    },
+    "id": "88049043-6468-4057-a837-85781c5af0a8",
+    "video_path": "./processed_datasets/jesbu1_racer_rfm_racer_val/frames/trajectory_88049043-6468-4057-a837-85781c5af0a8.npz",
+    "partial_success": null,
+    "target_progress": [
+      0.14285714285714285,
+      0.42857142857142855,
+      0.7142857142857143,
+      1.0
+    ],
+    "quality_label": "successful"
+  }
+]
\ No newline at end of file
diff --git a/robometer/dataset_upload/README.md b/robometer/dataset_upload/README.md
new file mode 100755
index 0000000000000000000000000000000000000000..88acf8ba74d32f17d5370408c2867a0b44c58b97
--- /dev/null
+++ b/robometer/dataset_upload/README.md
@@ -0,0 +1,224 @@
+# Adding New Datasets to Robometer
+
+This guide explains how to add new datasets to the Robometer training pipeline.
+
+## Supported Datasets
+
+### Ready-to-Use Datasets
+- **LIBERO**: Built-in HDF5 support → [📖 LIBERO Guide](dataset_guides/LIBERO.md)
+- **AgiBotWorld**: ✅ Native streaming support → [📖 AgiBotWorld Guide](dataset_guides/AgiBotWorld.md)
+
+### Custom Datasets
+- **Add Your Own**: DROID, Bridge, or any custom dataset → [📖 Custom Dataset Guide](dataset_guides/CustomDataset.md)
+
+## Quick Start
+
+### Use Existing Datasets
+```bash
+# AgiBotWorld (streaming)
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen_configs/agibot_world.yaml
+
+# LIBERO (local files)
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen.yaml
+```
+
+### Add Custom Dataset
+1. **Read the guide**: [Custom Dataset Guide](dataset_guides/CustomDataset.md)
+2. **Create loader**: `dataset_upload/{dataset_name}_loader.py`
+3. **Add config**: `dataset_upload/configs/data_gen_configs/{dataset_name}.yaml`
+4. **Test**: Run conversion and training
+
+## Architecture Overview
+
+Each dataset type has its own loader module. The main converter (`generate_hf_dataset.py`) is dataset-agnostic and works with any dataset-specific loader that follows the established interface.
+
+### Output Formats
+
+The converter supports two output formats:
+
+**Video Mode** (`--output.use_video=true`):
+- Creates MP4 video files using H.264 encoding
+- Videos are stored in organized directories: `trajectory_XXXX/trajectory.mp4`
+- Uses `datasets.Video()` feature for proper HuggingFace video display
+- Supports configurable frame rate, resolution, and cropping
+
+**Frame Mode** (`--output.use_video=false`):
+- Creates individual JPG image files
+- Images are stored in organized directories: `trajectory_XXXX/frame_XX.jpg`
+- Uses `datasets.Sequence(datasets.Image())` feature for image galleries
+- Supports configurable frame count and resolution
+
+### Video Processing Features
+
+The dataset converter includes built-in video processing:
+- **📹 Automatic Resizing**: Videos resized to consistent dimensions (configurable shortest edge size)
+- **⏱️ Frame Interpolation**: Downsamples to configurable frame count (default: all frames preserved)
+- **🎬 MP4 Creation**: Creates H.264 encoded MP4 files for optimal HuggingFace compatibility
+- **🎯 Quality Preservation**: Maintains visual quality while standardizing format
+- **📁 File Organization**: Organizes videos in trajectory-specific directories
+
+## Dataset Structure Requirements
+
+Your dataset loader must produce trajectories in the following format:
+
+```python
+{
+    'frames': Union[str, List[str]],   # Video file path (video mode) or list of image file paths (frame mode)
+    'actions': np.ndarray,             # Actions 
+    'is_robot': bool,                  # Whether this is robot data (True) or human data (False)
+    'task': str,                       # Human-readable task description
+    'optimal': str                     # Whether this trajectory is optimal
+}
+```
+
+**Note**: The dataset converter automatically creates MP4 video files or individual frame images based on the `use_video` setting.
+
+## Step-by-Step Guide
+
+### 1. Create Your Dataset Loader
+
+Create a new Python file in the `data/` directory following the naming convention: `{dataset_name}_loader.py`
+
+Example: `droid_loader.py` or `bridge_loader.py`
+
+```python
+#!/usr/bin/env python3
+"""
+DROID dataset loader for the generic dataset converter for Robometer model training.
+This module contains DROID-specific logic for loading and processing data.
+"""
+
+import os
+import numpy as np
+from typing import List, Dict
+from pathlib import Path
+from tqdm import tqdm
+
+def load_droid_dataset(base_path: str) -> Dict[str, List[Dict]]:
+    """Load DROID dataset and organize by task.
+    
+    Args:
+        base_path: Path to the DROID dataset directory
+        
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+    
+    print(f"Loading DROID dataset from: {base_path}")
+    
+    task_data = {}
+    
+    # Your dataset-specific loading logic here
+    # This is where you'll implement the logic to:
+    # 1. Find and read your data files
+    # 2. Extract frames, actions, rewards, etc.
+    # 3. Organize by task
+    # 4. Convert to the required format
+    
+    # Example structure (adapt to your dataset):
+    base_path = Path(base_path)
+    if not base_path.exists():
+        raise FileNotFoundError(f"DROID dataset path not found: {base_path}")
+    
+    # Find your data files (adapt this to your dataset structure)
+    data_files = list(base_path.glob("**/*.hdf5"))  # or whatever format you use
+    
+    for file_path in tqdm(data_files, desc="Processing DROID dataset"):
+        task_name = file_path.stem
+        
+        # Load your data file
+        # This is where you'll implement your specific loading logic
+        trajectories = load_trajectories_from_file(file_path)
+        
+        task_data[task_name] = trajectories
+    
+    print(f"Loaded {sum(len(trajectories) for trajectories in task_data.values())} trajectories from {len(task_data)} tasks")
+    return task_data
+
+def load_trajectories_from_file(file_path: Path) -> List[Dict]:
+    """Load trajectories from a single DROID data file."""
+    trajectories = []
+    
+    # Implement your file loading logic here
+    # This will depend on your dataset format (HDF5, JSON, pickle, etc.)
+    
+    # Example for HDF5 format:
+    import h5py
+    with h5py.File(file_path, 'r') as f:
+        # Navigate your HDF5 structure and extract data
+        # Convert to the required format
+        pass
+    
+    return trajectories
+```
+
+### 2. Update the Main Converter
+
+Add your dataset type to the main converter in `generate_hf_dataset.py`:
+
+```python
+# In the main() function, add your dataset type:
+elif cfg.dataset.dataset_type == "droid":
+    from dataset_loaders.droid_loader import load_droid_dataset
+    # Load the trajectories using your loader
+    task_data = load_droid_dataset(cfg.dataset.dataset_path)
+    trajectories = flatten_task_data(task_data)
+```
+
+### 3. Test Your Loader
+
+Create a simple test script to verify your loader works:
+
+```python
+# test_droid_loader.py
+from droid_loader import load_droid_dataset
+from helpers import flatten_task_data
+
+# Test your loader
+task_data = load_droid_dataset("/path/to/your/droid/dataset")
+trajectories = flatten_task_data(task_data)
+
+print(f"Loaded {len(trajectories)} trajectories")
+print(f"Sample trajectory keys: {list(trajectories[0].keys())}")
+print(f"Sample task: {trajectories[0].get('task', 'No task found')}")
+```
+
+### 4. Run Dataset Conversion
+
+Use the main converter with your new dataset:
+
+```bash
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/dataset_.yaml \
+    --dataset.dataset_name=your_dataset \
+    --dataset.dataset_path=/path/to/your/dataset \
+    --output.output_dir=your_robometer_dataset \
+    --output.max_trajectories=1000 \
+    --output.max_frames=-1 \
+    --output.use_video=true \
+    --output.fps=10 \
+    --output.shortest_edge_size=240 \
+    --output.center_crop=false
+```
+
+### Visualize the Dataset
+
+```bash
+uv run python visualize_dataset.py --dataset_path=your_robometer_dataset/your_dataset_name
+```
+
+## Dataset-Specific Guides
+
+📁 **[Browse All Dataset Guides](dataset_guides/)** - Complete overview with quick reference table
+
+### Individual Guides
+- **[📖 AgiBotWorld Guide](dataset_guides/AgiBotWorld.md)** - Streaming support, webdataset format
+- **[📖 LIBERO Guide](dataset_guides/LIBERO.md)** - HDF5 files, simulation data  
+- **[📖 Custom Dataset Guide](dataset_guides/CustomDataset.md)** - Add DROID, Bridge, or your own dataset
+
+Each guide includes:
+- ✅ Prerequisites and setup
+- ✅ Configuration examples
+- ✅ Troubleshooting tips
+- ✅ Performance notes
+- ✅ Integration with Robometer training
\ No newline at end of file
diff --git a/robometer/dataset_upload/__init__.py b/robometer/dataset_upload/__init__.py
new file mode 100755
index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc
--- /dev/null
+++ b/robometer/dataset_upload/__init__.py
@@ -0,0 +1 @@
+
diff --git a/robometer/dataset_upload/__pycache__/__init__.cpython-310.pyc b/robometer/dataset_upload/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d617c07706d8bf689c47985879372bab2ff89cea
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/__init__.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/__init__.cpython-311.pyc b/robometer/dataset_upload/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..717c0a3f7646d297188c924a606bc2abe650dced
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/__init__.cpython-311.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/generate_hf_dataset.cpython-310.pyc b/robometer/dataset_upload/__pycache__/generate_hf_dataset.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..cc82fe4bf86a9305a1c50fa3dd9159f9dfbb3262
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/generate_hf_dataset.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/helpers.cpython-310.pyc b/robometer/dataset_upload/__pycache__/helpers.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ecf6ae42f361b589aa796ff4b3d7cce2b37de86d
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/helpers.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/helpers.cpython-311.pyc b/robometer/dataset_upload/__pycache__/helpers.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..96329e10b83b25554a12e863006ced2c07e2b98f
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/helpers.cpython-311.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/video_helpers.cpython-310.pyc b/robometer/dataset_upload/__pycache__/video_helpers.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d56081aaefb0fc2cd3c720de2f879d8b8714a93f
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/video_helpers.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/__pycache__/video_helpers.cpython-311.pyc b/robometer/dataset_upload/__pycache__/video_helpers.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1bc95b0d67b1fece32eac51c29b011e95fa84a8f
Binary files /dev/null and b/robometer/dataset_upload/__pycache__/video_helpers.cpython-311.pyc differ
diff --git a/robometer/dataset_upload/configs/data_gen.yaml b/robometer/dataset_upload/configs/data_gen.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..cb366b820e15446bada770509c79eb85d61acd80
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen.yaml
@@ -0,0 +1,17 @@
+# config.yaml
+dataset:
+  dataset_path: LIBERO/libero/datasets/libero_90
+  dataset_name: libero_90  # This will be the dataset name / data source
+
+output:
+  output_dir: robometer_dataset
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: aliangdw/robometer
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/agibot_world.yaml b/robometer/dataset_upload/configs/data_gen_configs/agibot_world.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..bfdb2dea8bad7d1cea284ff34530269217f8737a
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/agibot_world.yaml
@@ -0,0 +1,15 @@
+# AgiBotWorld dataset configuration
+dataset:
+  dataset_path: "agibot-world/AgiBotWorld-Alpha"  # HuggingFace dataset name for streaming
+  dataset_name: agibotworld
+
+output:
+  output_dir: datasets/agibotworld_dataset
+  max_trajectories: 100  # Limit for testing, increase for full dataset
+  max_frames: 64
+  use_video: true
+  fps: 30
+
+hub:
+  push_to_hub: true  # Set to true when ready to upload
+  hub_repo_id: agibotworld_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/autoeval.yaml b/robometer/dataset_upload/configs/data_gen_configs/autoeval.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..de3bff239825d1e40f92ee31863ffc81c1bf6c71
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/autoeval.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/auto_eval  # Root containing eval_data/
+  dataset_name: auto_eval_rfm
+
+output:
+  output_dir: ./robometer_dataset/auto_eval_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 5
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: auto_eval_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/egocot.yaml b/robometer/dataset_upload/configs/data_gen_configs/egocot.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2b2817c8e867c7ea96eb629356dcb1b60340d971
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/egocot.yaml
@@ -0,0 +1,17 @@
+# config.yaml
+dataset:
+  dataset_path: ./egocot # Path to EgoCoT dataset directory that has EgoCOT_clear/ and results.json inside
+  dataset_name: egocot # This will be the dataset name / data source
+
+output:
+  output_dir: ./robometer_dataset/egocot_rfm
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 8  # EgoCoT uses 8 consecutive frames anyway
+  use_video: true
+  fps: 1  # Since EgoCoT is not a video dataset, we use 1 fps
+  shortest_edge_size: 224  # Standard size for vision models
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: egocot_rfm
diff --git a/robometer/dataset_upload/configs/data_gen_configs/egodex.yaml b/robometer/dataset_upload/configs/data_gen_configs/egodex.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..a6d7e322f6cfeb1c516b16794ec571d91be86b6c
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/egodex.yaml
@@ -0,0 +1,18 @@
+# EgoDex dataset configuration
+dataset:
+  dataset_path: "~/part2"  # Local path to EgoDex dataset directory
+  dataset_name: egodex_part2
+
+output:
+  output_dir: robometer_dataset/egodex_rfm
+  max_trajectories: -1
+  max_frames: 32
+  shortest_edge_size: 240
+  use_video: true
+  fps: 5
+  center_crop: false
+  num_workers: 3
+
+hub:
+  push_to_hub: true 
+  hub_repo_id: egodex_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/epic.yaml b/robometer/dataset_upload/configs/data_gen_configs/epic.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f4a11d122b932bfdbbd989ef98a64ab67cec7eec
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/epic.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ../epic_kitchens_100
+  dataset_name: epic
+
+output:
+  output_dir: ./robometer_dataset/epic_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 3 # less workers because cv2/ffmpeg loads are multiprocessed
+
+hub:
+  push_to_hub: true
+  hub_repo_id: epic_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/failsafe.yaml b/robometer/dataset_upload/configs/data_gen_configs/failsafe.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..169425a6a99347a85ab87a00ae8b055398fe4a70
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/failsafe.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/failsafe_data  # Root containing FailPickCube-v1/ FailPushCube-v1/ FailStackCube-v1/ and jsons
+  dataset_name: failsafe
+
+output:
+  output_dir: ./robometer_dataset/failsafe_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  num_workers: 4
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: failsafe_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/fino_net.yaml b/robometer/dataset_upload/configs/data_gen_configs/fino_net.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6386a3736b14e06e5bcaca5439c41632069ad93f
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/fino_net.yaml
@@ -0,0 +1,18 @@
+# config.yaml for FinoNet dataset
+dataset:
+  dataset_path: ./datasets/fino-net # Path to the dataset root (contains failure.zip or unzipped failure directory)
+  dataset_name: fino_net  # This will be the dataset name / data source
+
+output:
+  output_dir: robometer_dataset/fino_net_rfm
+  max_trajectories: -1  # -1 for all episodes
+  max_frames: 64
+  use_video: true
+  fps: 5
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: fino_net_rfm
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/galaxea.yaml b/robometer/dataset_upload/configs/data_gen_configs/galaxea.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9aa86d59b71373ced101cf9c65c5c01065077c64
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/galaxea.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/galaxea # Root containing 'rlds/'
+  dataset_name: galaxea_part1_r1_lite
+
+output:
+  output_dir: ./robometer_dataset/galaxea_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 15
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 1
+
+hub:
+  push_to_hub: true
+  hub_repo_id: galaxea_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/h2r.yaml b/robometer/dataset_upload/configs/data_gen_configs/h2r.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..bad779fe14bbf1da3f8708512ee3681190a9fca1
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/h2r.yaml
@@ -0,0 +1,17 @@
+# config.yaml
+dataset:
+  dataset_path: ./humanandrobot/data/v1
+  dataset_name: h2r  # This will be the dataset name / data source
+
+output:
+  output_dir: datasets/h2r_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: h2r_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/hand_paired.yaml b/robometer/dataset_upload/configs/data_gen_configs/hand_paired.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b987bb7a9646a60715ce2dfac52d4fb98f216657
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/hand_paired.yaml
@@ -0,0 +1,19 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/hand_paired_data
+  dataset_name: hand_paired
+
+output:
+  output_dir: ./robometer_dataset/hand_paired_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: hand_paired_rfm
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/humanoid_everyday.yaml b/robometer/dataset_upload/configs/data_gen_configs/humanoid_everyday.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d8da36bb42ea7b4812ea1d93fd67771f9e839b2f
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/humanoid_everyday.yaml
@@ -0,0 +1,18 @@
+# Humanoid Everyday dataset configuration
+dataset:
+  dataset_path: "./datasets/humanoid_everyday"  # Path containing zip files with humanoid everyday datasets
+  dataset_name: humanoid_everyday_rfm
+
+output:
+  output_dir: ./robometer_dataset/humanoid_everyday_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: humanoid_everyday_rfm
diff --git a/robometer/dataset_upload/configs/data_gen_configs/libero.yaml b/robometer/dataset_upload/configs/data_gen_configs/libero.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..8051a6a29bd2d4f72cf8da9defc16c516358341b
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/libero.yaml
@@ -0,0 +1,17 @@
+# config.yaml
+dataset:
+  dataset_path: deps/libero/LIBERO/libero/datasets/libero_90
+  dataset_name: libero_90  # This will be the dataset name / data source
+
+output:
+  output_dir: datasets/libero_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: libero_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/metaworld.yaml b/robometer/dataset_upload/configs/data_gen_configs/metaworld.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5c4f78cfe09d78569d96439e0c9f6c0d4d4d49da
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/metaworld.yaml
@@ -0,0 +1,17 @@
+dataset:
+  dataset_name: metaworld_train
+  dataset_path: mw_policy_ranking_data
+
+output:
+  output_dir: datasets/metaworld_rfm
+  use_video: true  # Convert to MP4 videos
+  fps: 10  # Target frame rate for output videos
+  shortest_edge_size: 240  # Resize videos to this shortest edge
+  center_crop: true  # Don't center crop, maintain aspect ratio
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: -1  # -1 for all frames
+  num_workers: 2
+
+hub:
+  push_to_hub: true
+  hub_repo_id: metaworld
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/mit_franka_prank.yaml b/robometer/dataset_upload/configs/data_gen_configs/mit_franka_prank.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..84cda7955912f8b8d372b00b3b118fa6e4d6678d
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/mit_franka_prank.yaml
@@ -0,0 +1,19 @@
+# MIT-Robot-Prank Dataset Configuration
+dataset:
+  dataset_path: ~/projects/robometer/datasets/20251210-mit-robot-prank  # Path to dataset directory
+  dataset_name: mit_franka_p-rank_rfm
+
+output:
+  output_dir: ./robometer_dataset/mit_franka_p-rank_rfm
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: mit_franka_p-rank_rfm
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/molmoact.yaml b/robometer/dataset_upload/configs/data_gen_configs/molmoact.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d897dfd1b94063cc74911febbf4c27985babe6df
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/molmoact.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/molmoact   # Root containing parquet files
+  dataset_name: molmoact_dataset_household
+
+output:
+  output_dir: ./robometer_dataset/molmoact_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 1
+
+hub:
+  push_to_hub: true
+  hub_repo_id: molmoact_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/motif.yaml b/robometer/dataset_upload/configs/data_gen_configs/motif.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8886e745a89ea2d1bf26845eb23eba5c67e9b04c
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/motif.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/MotIF  # Root containing annotations/, human_motion/, stretch_motion/
+  dataset_name: motif_rfm
+
+output:
+  output_dir: ./robometer_dataset/motif_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: motif_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/new_mit_franka.yaml b/robometer/dataset_upload/configs/data_gen_configs/new_mit_franka.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e4cb225d50a0d15abd959a1b1e0d9596a79d5f21
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/new_mit_franka.yaml
@@ -0,0 +1,19 @@
+# Robometer MIT Franka Dataset Configuration
+dataset:
+  dataset_path: ~/robometer/datasets/ROBOMETER_MIT_Franka  # Path to dataset directory
+  dataset_name: rfm_new_mit_franka_rfm_nowrist
+  exclude_wrist_cam: true  # Set to true to only process external camera views
+
+output:
+  output_dir: ./robometer_dataset/rfm_new_mit_franka_rfm_nowrist
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: rfm_new_mit_franka_rfm_nowrist
diff --git a/robometer/dataset_upload/configs/data_gen_configs/oxe.yaml b/robometer/dataset_upload/configs/data_gen_configs/oxe.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..db7a61882f1591f407cc23ecb2b8b98e2e41423c
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/oxe.yaml
@@ -0,0 +1,17 @@
+# EgoDex dataset configuration
+dataset:
+  dataset_path: "/home1/jessez/scratch_data/tensorflow_datasets/openx_datasets/"  # Local path to OXE dataset directory
+  dataset_name: oxe_bc_z # TODO: fill in dataset name
+
+output:
+  output_dir: robometer_datasets/oxe_rfm
+  max_trajectories: -1
+  max_frames: 32
+  shortest_edge_size: 240
+  use_video: true
+  fps: 10
+  center_crop: false
+
+hub:
+  push_to_hub: true 
+  hub_repo_id: oxe_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/ph2d.yaml b/robometer/dataset_upload/configs/data_gen_configs/ph2d.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..05ba30df7aeef7f2f96a8aaff048cfc89a29ef7e
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/ph2d.yaml
@@ -0,0 +1,19 @@
+# config.yaml
+dataset:
+  dataset_path: ./ph2d  # Root with sequence subfolders each containing HDF5 files
+  dataset_name: ph2d
+
+output:
+  output_dir: ./robometer_dataset/ph2d_rfm
+  max_trajectories: -1
+  max_frames: 64  # use all frames in each sequence
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: ph2d_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/racer_train.yaml b/robometer/dataset_upload/configs/data_gen_configs/racer_train.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8b6948f17fa0a7cb815218fdb213f167ad9f5986
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/racer_train.yaml
@@ -0,0 +1,19 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/RACER
+  dataset_name: racer_train
+
+output:
+  output_dir: ./robometer_dataset/racer_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 2
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: racer_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/racer_val.yaml b/robometer/dataset_upload/configs/data_gen_configs/racer_val.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..437488c867471709c0e40243b02eb68689d5e43d
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/racer_val.yaml
@@ -0,0 +1,19 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/RACER
+  dataset_name: racer_val
+
+output:
+  output_dir: ./robometer_dataset/racer_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 2
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: racer_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/roboarena.yaml b/robometer/dataset_upload/configs/data_gen_configs/roboarena.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f705b2fecf9da14daf0fa3cf8bd64e64935239c5
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/roboarena.yaml
@@ -0,0 +1,18 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/roboarena
+  dataset_name: roboarena_eval_debug_nowrist # This will be the dataset name / data source
+
+output:
+  output_dir: ./robometer_dataset/roboarena_eval_debug_nowrist
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  num_workers: 8
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: roboarena_eval_debug_nowrist
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/robofac.yaml b/robometer/dataset_upload/configs/data_gen_configs/robofac.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..103ee972788f5d7442e91d3f5cd65ad0acd43b62
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/robofac.yaml
@@ -0,0 +1,18 @@
+# RoboFAC (MINT-SJTU/RoboFAC-dataset). Run conversion first; then preprocess with preprocess_finetune.yaml.
+# Override: --dataset.dataset_path=/path/to/RoboFAC-dataset
+dataset:
+  dataset_path: "deps/RoboFAC-dataset"
+  dataset_name: robofac
+
+output:
+  output_dir: ./robometer_dataset/robofac
+  max_trajectories: null
+  max_frames: 32
+  shortest_edge_size: 240
+  use_video: true
+  fps: 10
+  center_crop: false
+
+hub:
+  push_to_hub: false
+  hub_repo_id: robofac_rbm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/robofail.yaml b/robometer/dataset_upload/configs/data_gen_configs/robofail.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..2dfa427bd398b6a896331f39805afedf6ab4ce8a
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/robofail.yaml
@@ -0,0 +1,17 @@
+# RoboFail dataset configuration
+dataset:
+  dataset_path: "deps/robofail"  # Local path to RoboFail dataset directory
+  dataset_name: robofail
+
+output:
+  output_dir: datasets/robofail_rfm
+  max_trajectories: null
+  max_frames: 64
+  shortest_edge_size: 240
+  use_video: true
+  fps: 30
+  center_crop: false
+
+hub:
+  push_to_hub: true 
+  hub_repo_id: robofail_rfm
\ No newline at end of file
diff --git a/robometer/dataset_upload/configs/data_gen_configs/roboreward.yaml b/robometer/dataset_upload/configs/data_gen_configs/roboreward.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2aeb78a0a34feebabeb6c9a03c6ed7e471ebe50e
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/roboreward.yaml
@@ -0,0 +1,18 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/RoboReward
+  dataset_name: roboreward_train  # Can be overridden with --dataset.dataset_name
+
+output:
+  output_dir: ./robometer_dataset/roboreward_rfm
+  max_trajectories: -1
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: roboreward_rfm
diff --git a/robometer/dataset_upload/configs/data_gen_configs/soar.yaml b/robometer/dataset_upload/configs/data_gen_configs/soar.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..77b420e53aa44bcfb3c556e9ddbc3b09f5859470
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/soar.yaml
@@ -0,0 +1,20 @@
+# config.yaml
+dataset:
+  dataset_path: ./datasets/soar_data # Root containing rlds/<split>/<version>/
+  dataset_name: soar_rfm
+
+output:
+  output_dir: ./robometer_dataset/soar_rfm
+  max_trajectories: -1
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: soar_rfm
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/usc_franka_policy_ranking.yaml b/robometer/dataset_upload/configs/data_gen_configs/usc_franka_policy_ranking.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..204b679d2e6a447907a48a105ea4fe21bef935f4
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/usc_franka_policy_ranking.yaml
@@ -0,0 +1,19 @@
+# USC Franka Policy Ranking dataset config
+dataset:
+  dataset_path: ./nov17_datacollect
+  dataset_name: usc_franka_policy_ranking
+
+output:
+  output_dir: ./robometer_dataset/usc_franka_policy_ranking
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 2
+
+hub:
+  push_to_hub: false
+  hub_repo_id: usc_franka_policy_ranking
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml b/robometer/dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..545ea7252ef60a249fa56350c30f1311cdc16686
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml
@@ -0,0 +1,20 @@
+# USC Koch Human-Robot Paired Dataset Configuration
+# Use with --dataset.dataset_name to specify "human" or "robot"
+dataset:
+  dataset_path: ~/projects/robometer/datasets/usc_koch_human_robot_paired
+  dataset_name: usc_koch_human_robot_paired_human  # Override with human/robot
+
+output:
+  output_dir: ./robometer_dataset/usc_koch_human_robot_paired  # Will be overridden by dataset_name
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: usc_koch_human_robot_paired  # Will be overridden by dataset_name
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/usc_koch_p_ranking.yaml b/robometer/dataset_upload/configs/data_gen_configs/usc_koch_p_ranking.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..decb0f8c9d4865e4aacb536ad623f939c54ac1a5
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/usc_koch_p_ranking.yaml
@@ -0,0 +1,24 @@
+# USC Koch P-Ranking (robot-only) Dataset Configuration
+#
+# Folder expected (see dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh):
+#   <dataset_path>/robot/usc_koch_*  (includes success/suboptimal/failure)
+#
+dataset:
+  dataset_path: ~/projects/robometer/datasets/usc_koch_human_robot_paired
+  dataset_name: usc_koch_p_ranking_rfm
+
+output:
+  output_dir: ./robometer_dataset/usc_koch_p_ranking  # overridden per dataset_name (like other dataset_upload converters)
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: usc_koch_p_ranking_rfm  # recommended: keep constant; converter strips suffixes when pushing
+
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/usc_xarm_policy_ranking.yaml b/robometer/dataset_upload/configs/data_gen_configs/usc_xarm_policy_ranking.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..16d0d38aca2938fbe2a93ef6727d6082a44294cd
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/usc_xarm_policy_ranking.yaml
@@ -0,0 +1,19 @@
+# USC xArm Policy Ranking dataset config
+dataset:
+  dataset_path: ./xarm_policy_ranking
+  dataset_name: usc_xarm_policy_ranking
+
+output:
+  output_dir: ./robometer_dataset/usc_xarm_policy_ranking
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 2
+
+hub:
+  push_to_hub: false
+  hub_repo_id: usc_xarm_policy_ranking
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ae0060da00d0140d589e8f520f1ceb729e6f7b17
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml
@@ -0,0 +1,19 @@
+# UTD SO101 Clean Policy Ranking Dataset Configuration - Top View
+dataset:
+  dataset_path: ~/projects/robometer/datasets/utd_prank
+  dataset_name: utd_so101_clean_policy_ranking_top
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_clean_policy_ranking_top
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: utd_so101_clean_policy_ranking_top
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e50b0a0c24a1b2922adf309db5de5a24b25b8f81
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml
@@ -0,0 +1,19 @@
+# UTD SO101 Clean Policy Ranking Dataset Configuration - Wrist View
+dataset:
+  dataset_path: ~/projects/robometer/datasets/utd_prank
+  dataset_name: utd_so101_clean_policy_ranking_wrist
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_clean_policy_ranking_wrist
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: utd_so101_clean_policy_ranking_wrist
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/utd_so101_eval.yaml b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_eval.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7979cd4683b2d0450a71d9c58414fe391fc775dd
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_eval.yaml
@@ -0,0 +1,19 @@
+# Koch Arm UT Dallas dataset config
+dataset:
+  dataset_path: ./raw_data/ROBOMETER_UTD_Eval
+  dataset_name: utd_so101_policy_ranking
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_policy_ranking
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 2
+
+hub:
+  push_to_hub: false
+  hub_repo_id: utd_so101_policy_ranking
+
diff --git a/robometer/dataset_upload/configs/data_gen_configs/utd_so101_human.yaml b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_human.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5c8bbc8bd5d93a15365b697c639bc19310bdf6d1
--- /dev/null
+++ b/robometer/dataset_upload/configs/data_gen_configs/utd_so101_human.yaml
@@ -0,0 +1,18 @@
+# UTD SO101 Human dataset config
+dataset:
+  dataset_path: ./raw_data/ROBOMETER_UTD_Human
+  dataset_name: utd_so101_human
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_human
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 2
+
+hub:
+  push_to_hub: false
+  hub_repo_id: utd_so101_human
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/__init__.py b/robometer/dataset_upload/data_scripts/__init__.py
new file mode 100755
index 0000000000000000000000000000000000000000..5a1496f2b8b243c1e86c4413fe80e913f58ae3fc
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/__init__.py
@@ -0,0 +1 @@
+# Makes this directory a package for relative imports in submodules
diff --git a/robometer/dataset_upload/data_scripts/agibot/__init__.py b/robometer/dataset_upload/data_scripts/agibot/__init__.py
new file mode 100755
index 0000000000000000000000000000000000000000..613c33158eb4f3f215edb3e041db5b373c31baa8
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/agibot/__init__.py
@@ -0,0 +1,21 @@
+from .agibot_helper import (
+    DEFAULT_DATASET_ROOT,
+    DEFAULT_INDEX_PATH,
+    DEFAULT_TASK_INFO_DIR,
+    build_episode_to_task_index,
+    ensure_episode_index,
+    find_task_json_for_episode,
+    get_episode_record,
+    load_episode_to_task_index,
+)
+
+__all__ = [
+    "DEFAULT_DATASET_ROOT",
+    "DEFAULT_INDEX_PATH",
+    "DEFAULT_TASK_INFO_DIR",
+    "build_episode_to_task_index",
+    "ensure_episode_index",
+    "find_task_json_for_episode",
+    "get_episode_record",
+    "load_episode_to_task_index",
+]
diff --git a/robometer/dataset_upload/data_scripts/agibot/agibot_helper.py b/robometer/dataset_upload/data_scripts/agibot/agibot_helper.py
new file mode 100755
index 0000000000000000000000000000000000000000..723360348d60899d01c249fa6cdbb6b2ac23d4cc
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/agibot/agibot_helper.py
@@ -0,0 +1,170 @@
+"""Utilities for indexing and looking up AgiBotWorld task JSONs.
+
+This module builds and uses an index mapping episode_id -> task_json_path,
+so downstream scripts can quickly resolve which task file contains a given
+episode. The index is stored as a JSON file on disk for fast re-use.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from collections.abc import Iterable
+from glob import glob
+
+# Default locations used by the downloader
+DEFAULT_DATASET_ROOT: str = os.path.join("datasets", "tmp", "agibotworld_alpha")
+DEFAULT_TASK_INFO_DIR: str = os.path.join(DEFAULT_DATASET_ROOT, "task_info")
+DEFAULT_INDEX_PATH: str = os.path.join(DEFAULT_DATASET_ROOT, "episode_to_task_index.json")
+
+
+def _task_json_files(task_info_dir: str) -> Iterable[str]:
+    """Yield task JSON files within the given directory.
+
+    Parameters
+    - task_info_dir: Directory that contains task_*.json files
+    """
+
+    pattern = os.path.join(task_info_dir, "task_*.json")
+    for file_path in sorted(glob(pattern)):
+        # Guard against directories or non-files matching the pattern
+        if os.path.isfile(file_path):
+            yield file_path
+
+
+def build_episode_to_task_index(
+    task_info_dir: str = DEFAULT_TASK_INFO_DIR,
+    output_index_path: str = DEFAULT_INDEX_PATH,
+    verbose: bool = True,
+) -> dict[str, str]:
+    """Build an index mapping episode_id -> task_json_path and save to disk.
+
+    Keys are stored as strings in the JSON index for portability, but callers
+    may freely pass `int` episode ids to the lookup helpers provided here.
+
+    Returns the mapping in-memory as well.
+    """
+
+    if not os.path.isdir(task_info_dir):
+        raise FileNotFoundError(f"Task info directory not found: {os.path.abspath(task_info_dir)}")
+
+    episode_to_task: dict[str, str] = {}
+    duplicate_episode_ids: list[str] = []
+
+    for json_path in _task_json_files(task_info_dir):
+        try:
+            with open(json_path, encoding="utf-8") as f:
+                entries = json.load(f)
+        except Exception as exc:  # pragma: no cover - defensive logging
+            if verbose:
+                print(f"[agibot_helper] Skipping unreadable file {json_path}: {exc}")
+            continue
+
+        if not isinstance(entries, list):
+            if verbose:
+                print(f"[agibot_helper] Unexpected JSON structure in {json_path}; expected a list.")
+            continue
+
+        for item in entries:
+            if not isinstance(item, dict) or "episode_id" not in item:
+                continue
+            episode_id_val = item["episode_id"]
+            # Normalize to string for JSON index stability
+            episode_key = str(episode_id_val)
+            if episode_key in episode_to_task:
+                duplicate_episode_ids.append(episode_key)
+            episode_to_task[episode_key] = json_path
+
+    # Ensure output directory exists
+    os.makedirs(os.path.dirname(output_index_path), exist_ok=True)
+    with open(output_index_path, "w", encoding="utf-8") as f:
+        json.dump(episode_to_task, f)
+
+    if verbose:
+        total = len(episode_to_task)
+        dupes = len(set(duplicate_episode_ids))
+        print(
+            f"[agibot_helper] Built episode->task index with {total} episodes"
+            + (f" ({dupes} duplicate ids overwritten)" if dupes else "")
+            + f" at {os.path.abspath(output_index_path)}"
+        )
+
+    return episode_to_task
+
+
+def load_episode_to_task_index(index_path: str = DEFAULT_INDEX_PATH) -> dict[str, str]:
+    """Load the episode->task index from disk.
+
+    Raises FileNotFoundError if the index doesn't exist.
+    """
+
+    if not os.path.isfile(index_path):
+        raise FileNotFoundError(
+            f"Episode index not found: {os.path.abspath(index_path)}. Build it with build_episode_to_task_index()."
+        )
+    with open(index_path, encoding="utf-8") as f:
+        return json.load(f)
+
+
+def ensure_episode_index(
+    task_info_dir: str = DEFAULT_TASK_INFO_DIR,
+    index_path: str = DEFAULT_INDEX_PATH,
+    verbose: bool = False,
+) -> dict[str, str]:
+    """Return an in-memory episode->task index, building it if missing."""
+
+    if os.path.isfile(index_path):
+        return load_episode_to_task_index(index_path)
+    return build_episode_to_task_index(task_info_dir, index_path, verbose=verbose)
+
+
+def find_task_json_for_episode(
+    episode_id: int | str,
+    task_info_dir: str = DEFAULT_TASK_INFO_DIR,
+    index_path: str = DEFAULT_INDEX_PATH,
+) -> str:
+    """Return the path to the task JSON file containing the given episode.
+
+    If the index file is missing, it will be built automatically.
+    Raises KeyError if the episode id is not present in the index.
+    """
+
+    index = ensure_episode_index(task_info_dir=task_info_dir, index_path=index_path)
+    key = str(episode_id)
+    if key not in index:
+        raise KeyError(f"Episode id {episode_id} not found in index at {os.path.abspath(index_path)}")
+    return index[key]
+
+
+def get_episode_record(
+    episode_id: int | str,
+    task_info_dir: str = DEFAULT_TASK_INFO_DIR,
+    index_path: str = DEFAULT_INDEX_PATH,
+) -> tuple[str, dict]:
+    """Return (json_path, episode_record) for the given episode id.
+
+    This reads the relevant task JSON file and returns the full dictionary
+    entry for the requested episode. Raises KeyError if the episode isn't
+    present.
+    """
+
+    json_path = find_task_json_for_episode(episode_id=episode_id, task_info_dir=task_info_dir, index_path=index_path)
+    with open(json_path, encoding="utf-8") as f:
+        entries = json.load(f)
+    key = int(episode_id)
+    for item in entries:
+        if isinstance(item, dict) and int(item.get("episode_id", -1)) == key:
+            return json_path, item
+    raise KeyError(f"Episode id {episode_id} not found inside file {os.path.abspath(json_path)}")
+
+
+__all__ = [
+    "DEFAULT_DATASET_ROOT",
+    "DEFAULT_INDEX_PATH",
+    "DEFAULT_TASK_INFO_DIR",
+    "build_episode_to_task_index",
+    "ensure_episode_index",
+    "find_task_json_for_episode",
+    "get_episode_record",
+    "load_episode_to_task_index",
+]
diff --git a/robometer/dataset_upload/data_scripts/agibot/download_task_jsons.py b/robometer/dataset_upload/data_scripts/agibot/download_task_jsons.py
new file mode 100755
index 0000000000000000000000000000000000000000..f4e6f69467bc8cf34329be2195132713b9d46dfd
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/agibot/download_task_jsons.py
@@ -0,0 +1,42 @@
+import os
+import sys
+
+from huggingface_hub import snapshot_download
+
+# Support running as a module (python -m robometer.data.data_scripts.agibot.download_task_jsons)
+# and running directly as a script (python dataset_upload/data_scripts/agibot/download_task_jsons.py)
+try:
+    from .agibot_helper import (
+        DEFAULT_DATASET_ROOT,
+        DEFAULT_TASK_INFO_DIR,
+        build_episode_to_task_index,
+    )
+except Exception:  # pragma: no cover - fallback for direct execution
+    current_dir = os.path.dirname(os.path.abspath(__file__))
+    project_root = os.path.abspath(os.path.join(current_dir, "../../../.."))
+    if project_root not in sys.path:
+        sys.path.insert(0, project_root)
+    from robometer.data.data_scripts.agibot.agibot_helper import (  # type: ignore
+        DEFAULT_DATASET_ROOT,
+        DEFAULT_TASK_INFO_DIR,
+        build_episode_to_task_index,
+    )
+
+
+def main() -> None:
+    os.makedirs(DEFAULT_DATASET_ROOT, exist_ok=True)
+
+    # Download task JSONs
+    snapshot_download(
+        repo_id="agibot-world/AgiBotWorld-Alpha",
+        allow_patterns="task_info/*.json",
+        local_dir=DEFAULT_DATASET_ROOT,
+        repo_type="dataset",
+    )
+
+    # Build episode->task index to speed up lookups later
+    build_episode_to_task_index(task_info_dir=DEFAULT_TASK_INFO_DIR, verbose=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/dataset_upload/data_scripts/egodex/download_and_convert.sh b/robometer/dataset_upload/data_scripts/egodex/download_and_convert.sh
new file mode 100644
index 0000000000000000000000000000000000000000..411f2124c666a3bc24cc56ef3304366f43eb79d4
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/egodex/download_and_convert.sh
@@ -0,0 +1,31 @@
+for part in test part1 part2 part3 part4 part5; do
+
+    echo "Processing ${part}..."
+
+    # Download the dataset. 
+    # Example curl "https://ml-site.cdn-apple.com/datasets/egodex/test.zip" -o test.zip
+    # Download it to ~/egodex/${part}.zip
+    # check first if it doesn't exist
+    mkdir -p ${HOME}/egodex
+    if [ ! -d "${HOME}/egodex/${part}" ]; then
+        echo "curl "https://ml-site.cdn-apple.com/datasets/egodex/${part}.zip" -o ${HOME}/egodex/${part}.zip"
+        curl "https://ml-site.cdn-apple.com/datasets/egodex/${part}.zip" -o ${HOME}/egodex/${part}.zip
+        unzip -d ${HOME}/egodex/${part} ${HOME}/egodex/${part}.zip
+        rm ${HOME}/egodex/${part}.zip
+    fi
+
+    uv run python -m dataset_upload.generate_hf_dataset \
+        --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+        --dataset.dataset_path="${HOME}/egodex/${part}/${part}" \
+        --dataset.dataset_name="egodex_${part}" \
+        --output.output_dir=./robometer_dataset/egodex_rfm \
+        --hub.push_to_hub=false
+
+    echo "Done processing ${part}..."
+
+    # Delete the dataset
+    echo "Deleting ${HOME}/egodex/${part}..."
+    rm -rf ${HOME}/egodex/${part}
+    echo "Done deleting ${HOME}/egodex/${part}..."
+
+done
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/galaxea/gen_all_galaxea.sh b/robometer/dataset_upload/data_scripts/galaxea/gen_all_galaxea.sh
new file mode 100755
index 0000000000000000000000000000000000000000..7a5819aec5dc25649aacbd1f9bfa8149308d6539
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/galaxea/gen_all_galaxea.sh
@@ -0,0 +1,20 @@
+# large datasets near the end
+galaxea_valid_datasets=(
+    "part1_r1_lite"
+    "part2_r1_lite"
+    "part3_r1_lite"
+    "part4_r1_lite"
+    "part5_r1_lite"
+)
+
+for dataset_name in ${galaxea_valid_datasets[@]}; do
+    echo "Generating dataset: $dataset_name"
+    echo "=" * 100
+    uv run python -m dataset_upload.generate_hf_dataset \
+        --config_path=dataset_upload/configs/data_gen_configs/galaxea.yaml \
+        --output.output_dir ./rfm_dataset/galaxea_rfm \
+        --dataset.dataset_name galaxea_$dataset_name \
+
+    echo "Done generating dataset: $dataset_name"
+    echo "=" * 100
+done
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/humanoid_everyday/download_humanoid_everyday.sh b/robometer/dataset_upload/data_scripts/humanoid_everyday/download_humanoid_everyday.sh
new file mode 100644
index 0000000000000000000000000000000000000000..a408bcc5d6aea0d9f65ae76e5a92d1a22498d88d
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/humanoid_everyday/download_humanoid_everyday.sh
@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+# -------------------------------------------------------------
+# download_all.sh
+#
+# Downloads all the Dropbox folders in the URL list below and
+# extracts each one into *{DEST_ROOT}/{folder_name}*.
+#
+# 1️⃣ Edit the variable `DEST_ROOT` at the top of this file to
+#     point wherever you want the data stored.  
+#     Example: DEST_ROOT="/home/user/DropboxData"
+#
+# 2️⃣ Make it executable and run:
+#        chmod +x download_all.sh
+#        ./download_all.sh
+#
+# -------------------------------------------------------------
+
+set -euo pipefail
+
+# ------------------------------------------------------------------
+# 1️⃣ Destination root – change this to wherever you want the data.
+DEST_ROOT="./datasets/humanoid_everyday"   # ← modify here
+
+# Create it if it doesn't exist (but don't overwrite an existing one)
+mkdir -p "$DEST_ROOT"
+
+# ------------------------------------------------------------------
+# 2️⃣ List of URLs (one per line). Add or remove lines as needed.
+URLS=(
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AA8nw-Rehsp19BgpznzX3G8/Articulated?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=p7o1n5u8&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AB-bJ9d9SthpaxRq-psftIM/Basic%20manipulation?rlkey=42llsh52wfq47r77m05mkikus&e=1&st=ujqfmcrx&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AHhg5yc8HuZ3nUJwHEpvV7o/Deformable?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=uuahde75&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AHCZsyuhfXG9BpV9OvyTUUY/Human%20robot%20interaction?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=f0l92o3l&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AIaNECZ5mNu7P84jRk5IaAE/Loco-manipulation?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=ndxg5mxe&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AGDzhFFpengBQRE4K_mK-_k/Precision?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=v7wpk52e&dl=0"
+"https://www.dropbox.com/scl/fo/r6xwxxuiwmnypzprzqza7/AMeSkqL7WLIuIFVzHe8T5nE/Tool%20use?rlkey=42llsh52wfq47r77m05mkikus&subfolder_nav_tracking=1&st=1dmsbxmc&dl=0"
+)
+
+# ------------------------------------------------------------------
+download_and_extract() {
+    local url="$1"
+
+    # 1️⃣ Build the ZIP URL
+    local zip_url="${url/&dl=0/&dl=1}"
+    zip_url="${zip_url/;dl=0;/;dl=1;}"
+
+    echo "📦 [$url] → downloading ZIP …"
+    curl -L --fail "$zip_url" --output /tmp/dropbox_folder.zip
+
+    # 2️⃣ Determine the folder name (last path segment before any query)
+    local folder_name="$(basename "${url%%\?*}")"
+
+    # 3️⃣ Full destination path: DEST_ROOT/folder_name
+    local out_dir="${DEST_ROOT}/${folder_name}"
+    mkdir -p "$out_dir"
+
+    echo "📂 [$url] → extracting to '$out_dir/' …"
+    unzip -o /tmp/dropbox_folder.zip -d "$out_dir" || true # ignore non-fatal warnings
+
+    rm -f /tmp/dropbox_folder.zip
+    echo "✅ [$url] done!"
+}
+
+# ------------------------------------------------------------------
+for u in "${URLS[@]}"; do
+    download_and_extract "$u"
+done
+
+echo "🎉 All folders downloaded and extracted into $DEST_ROOT."
+
diff --git a/robometer/dataset_upload/data_scripts/libero/gen_all_libero.sh b/robometer/dataset_upload/data_scripts/libero/gen_all_libero.sh
new file mode 100755
index 0000000000000000000000000000000000000000..f8ae23eacfbdabbabd3af389cb072e3b1f767343
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/libero/gen_all_libero.sh
@@ -0,0 +1,24 @@
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_10_256 \
+    --dataset.dataset_name=libero256_10 
+
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_90_256 \
+    --dataset.dataset_name=libero256_90 
+
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_spatial_256 \
+    --dataset.dataset_name=libero256_spatial 
+
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_goal_256 \
+    --dataset.dataset_name=libero256_goal 
+
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_object_256 \
+    --dataset.dataset_name=libero256_object
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/libero/rerender_all.sh b/robometer/dataset_upload/data_scripts/libero/rerender_all.sh
new file mode 100644
index 0000000000000000000000000000000000000000..7a579e437e5cd38217e3bd6a2538c9f60436eae3
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/libero/rerender_all.sh
@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Rerender LIBERO datasets across suites using uv + rerender_libero.py
+# Usage:
+#   ./libero/rerender_all.sh
+#
+# Optional env vars:
+#   UV_BIN=uv                # override uv binary
+#   SUITES="a b c"           # override suites list (space-separated)
+#   DEBUG_MODE=1             # pass --debug_mode to the python script
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(pwd)"
+
+UV_BIN=${UV_BIN:-uv}
+PY_SCRIPT="${PROJECT_ROOT}/dataset_upload/data_scripts/libero/rerender_libero.py"
+REQS_FILE="${PROJECT_ROOT}/deps/libero/LIBERO/requirements.txt"
+DATASETS_DIR="${PROJECT_ROOT}/deps/libero/LIBERO/libero/datasets"
+
+# Default suites to process if not overridden
+DEFAULT_SUITES=(
+  "libero_spatial"
+  "libero_object"
+  "libero_goal"
+  # "libero_10"
+  "libero_90"
+)
+
+if [[ -n "${SUITES:-}" ]]; then
+  # shellcheck disable=SC2206
+  SUITES_ARR=( ${SUITES} )
+else
+  SUITES_ARR=( "${DEFAULT_SUITES[@]}" )
+fi
+
+DEBUG_FLAG=""
+if [[ "${DEBUG_MODE:-0}" == "1" ]]; then
+  DEBUG_FLAG="--debug_mode"
+fi
+
+echo "Project root: ${PROJECT_ROOT}"
+echo "Datasets dir: ${DATASETS_DIR}"
+echo "Suites: ${SUITES_ARR[*]}"
+
+for SUITE in "${SUITES_ARR[@]}"; do
+  RAW_DIR="${DATASETS_DIR}/${SUITE}"
+  TARGET_DIR="${DATASETS_DIR}/${SUITE}_256"
+
+  if [[ ! -d "${RAW_DIR}" ]]; then
+    echo "[skip] ${SUITE}: raw data dir not found: ${RAW_DIR}" >&2
+    continue
+  fi
+
+  echo "[run] ${SUITE} → ${TARGET_DIR}"
+  "${UV_BIN}" run --with-requirements "${REQS_FILE}" \
+    "${PY_SCRIPT}" \
+    --libero_task_suite "${SUITE}" \
+    --libero_raw_data_dir "${RAW_DIR}" \
+    --libero_target_dir "${TARGET_DIR}" \
+    ${DEBUG_FLAG}
+done
+
+echo "All requested suites processed."
+
+
diff --git a/robometer/dataset_upload/data_scripts/libero/rerender_libero.py b/robometer/dataset_upload/data_scripts/libero/rerender_libero.py
new file mode 100644
index 0000000000000000000000000000000000000000..7ad8b6d6480a27be6325d9077cc6fc9f710f354e
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/libero/rerender_libero.py
@@ -0,0 +1,322 @@
+"""
+Generates a LIBERO failure dataset (HDF5 files) by replaying demonstrations with added noise in the environments.
+
+Notes:
+    - We save image observations at 256x256px resolution (instead of 128x128).
+    - We filter out transitions with "no-op" (zero) actions that do not change the robot's state.
+    - We only save trajectories that FAIL (do not complete successfully) to create a failure dataset.
+
+Usage:
+    python generate_failure_dataset.py \
+        --libero_task_suite [ libero_spatial | libero_object | libero_goal | libero_10 ] \
+        --libero_raw_data_dir <PATH TO RAW HDF5 DATASET DIR> \
+        --libero_target_dir <PATH TO TARGET DIR> \
+        [--debug_mode]
+
+    Example (LIBERO-Spatial):
+        python generate_failure_dataset.py \
+            --libero_task_suite libero_90 \
+            --libero_raw_data_dir ./libero/datasets/libero_90 \
+            --libero_target_dir ./libero/datasets/libero_90_failure \
+            --debug_mode
+
+"""
+
+# Add libero to the python path
+import sys
+
+sys.path.append("deps/libero/LIBERO")
+
+import argparse
+import json
+import os
+from datetime import datetime
+
+import h5py
+import imageio
+import numpy as np
+import robosuite.utils.transform_utils as T
+import tqdm
+from libero.libero import benchmark, get_libero_path
+from libero.libero.envs import OffScreenRenderEnv
+
+IMAGE_RESOLUTION = 256
+
+
+def get_libero_env(task, model_family, resolution=256):
+    """Initializes and returns the LIBERO environment, along with the task description."""
+    task_description = task.language
+    task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
+    env_args = {"bddl_file_name": task_bddl_file, "camera_heights": resolution, "camera_widths": resolution}
+    env = OffScreenRenderEnv(**env_args)
+    env.seed(0)  # IMPORTANT: seed seems to affect object positions even when using fixed initial state
+    return env, task_description
+
+
+def get_libero_dummy_action(model_family: str):
+    """Get dummy/no-op action, used to roll out the simulation while the robot does nothing."""
+    return [0, 0, 0, 0, 0, 0, -1]
+
+
+def save_failure_video(rollout_images, task_id, demo_id, task_description):
+    """Saves an MP4 replay of a failure episode."""
+
+    # Create timestamp for unique filename
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+
+    # Process task description for filename
+    processed_task_description = task_description.lower().replace(" ", "_").replace("\n", "_").replace(".", "_")[:50]
+
+    # Create filename
+    mp4_path = f"failure_task{task_id}_demo{demo_id}_{timestamp}_{processed_task_description}.mp4"
+
+    # Create video writer
+    video_writer = imageio.get_writer(mp4_path, fps=30)
+
+    # Write frames
+    for img in rollout_images:
+        video_writer.append_data(img)
+
+    video_writer.close()
+    print(f"Saved failure video at: {mp4_path}")
+    return mp4_path
+
+
+def is_noop(action, prev_action=None, threshold=1e-4):
+    """
+    Returns whether an action is a no-op action.
+
+    A no-op action satisfies two criteria:
+        (1) All action dimensions, except for the last one (gripper action), are near zero.
+        (2) The gripper action is equal to the previous timestep's gripper action.
+
+    Explanation of (2):
+        Naively filtering out actions with just criterion (1) is not good because you will
+        remove actions where the robot is staying still but opening/closing its gripper.
+        So you also need to consider the current state (by checking the previous timestep's
+        gripper action as a proxy) to determine whether the action really is a no-op.
+    """
+    # Special case: Previous action is None if this is the first action in the episode
+    # Then we only care about criterion (1)
+    if prev_action is None:
+        return np.linalg.norm(action[:-1]) < threshold
+
+    # Normal case: Check both criteria (1) and (2)
+    gripper_action = action[-1]
+    prev_gripper_action = prev_action[-1]
+    return np.linalg.norm(action[:-1]) < threshold and gripper_action == prev_gripper_action
+
+
+def main(args):
+    print(f"Generating {args.libero_task_suite} failure dataset!")
+
+    if args.debug_mode:
+        print("Debug mode: Will save failure trajectory as MP4 video")
+
+    # Create target directory
+    if os.path.isdir(args.libero_target_dir):
+        user_input = input(
+            f"Target directory already exists at path: {args.libero_target_dir}\nEnter 'y' to overwrite the directory, or anything else to exit: "
+        )
+        if user_input != "y":
+            exit()
+    os.makedirs(args.libero_target_dir, exist_ok=True)
+
+    # Prepare JSON file to record success/failure and initial states per episode
+    metainfo_json_dict = {}
+    metainfo_json_out_path = f"deps/libero/LIBERO/libero/datasets/{args.libero_task_suite}_rerender_metainfo.json"
+    with open(metainfo_json_out_path, "w") as f:
+        # Just test that we can write to this file (we overwrite it later)
+        json.dump(metainfo_json_dict, f)
+
+    # Get task suite
+    benchmark_dict = benchmark.get_benchmark_dict()
+    task_suite = benchmark_dict[args.libero_task_suite]()
+    num_tasks_in_suite = task_suite.n_tasks
+
+    # Setup
+    num_replays = 0
+    num_failures = 0
+    num_noops = 0
+
+    for task_id in tqdm.tqdm(range(num_tasks_in_suite)):
+        # Get task in suite
+        task = task_suite.get_task(task_id)
+        env, task_description = get_libero_env(task, "llava", resolution=IMAGE_RESOLUTION)
+
+        # Get dataset for task
+        orig_data_path = os.path.join(args.libero_raw_data_dir, f"{task.name}_demo.hdf5")
+        assert os.path.exists(orig_data_path), f"Cannot find raw data file {orig_data_path}."
+        orig_data_file = h5py.File(orig_data_path, "r")
+        orig_data = orig_data_file["data"]
+
+        # Create new HDF5 file for failure demos
+        new_data_path = os.path.join(args.libero_target_dir, f"{task.name}_demo.hdf5")
+        new_data_file = h5py.File(new_data_path, "w")
+        grp = new_data_file.create_group("data")
+
+        for i in range(len(orig_data.keys())):
+            # Get demo data
+            demo_data = orig_data[f"demo_{i}"]
+            orig_actions = demo_data["actions"][()]
+            orig_states = demo_data["states"][()]
+
+            # Reset environment, set initial state, and wait a few steps for environment to settle
+            env.reset()
+            env.set_init_state(orig_states[0])
+            for _ in range(10):
+                obs, reward, done, info = env.step(get_libero_dummy_action("llava"))
+
+            # Set up new data lists
+            states = []
+            actions = []
+            ee_states = []
+            gripper_states = []
+            joint_states = []
+            robot_states = []
+            agentview_images = []
+            eye_in_hand_images = []
+
+            # For video saving (if enabled)
+            if args.debug_mode:
+                rollout_images = []
+
+            # Replay original demo actions with added noise in environment and record observations
+            for _, action in enumerate(orig_actions):
+                # Skip transitions with no-op actions
+                prev_action = actions[-1] if len(actions) > 0 else None
+                if is_noop(action, prev_action):
+                    print(f"\tSkipping no-op action: {action}")
+                    num_noops += 1
+                    continue
+
+                if states == []:
+                    # In the first timestep, since we're using the original initial state to initialize the environment,
+                    # copy the initial state (first state in episode) over from the original HDF5 to the new one
+                    states.append(orig_states[0])
+                    robot_states.append(demo_data["robot_states"][0])
+                else:
+                    # For all other timesteps, get state from environment and record it
+                    states.append(env.sim.get_state().flatten())
+                    robot_states.append(
+                        np.concatenate([obs["robot0_gripper_qpos"], obs["robot0_eef_pos"], obs["robot0_eef_quat"]])
+                    )
+
+                # Add noise to action to create failure trajectories
+                actions.append(action)
+
+                # Record data returned by environment
+                if "robot0_gripper_qpos" in obs:
+                    gripper_states.append(obs["robot0_gripper_qpos"])
+                joint_states.append(obs["robot0_joint_pos"])
+                ee_states.append(
+                    np.hstack((
+                        obs["robot0_eef_pos"],
+                        T.quat2axisangle(obs["robot0_eef_quat"]),
+                    ))
+                )
+                agentview_images.append(obs["agentview_image"])
+                eye_in_hand_images.append(obs["robot0_eye_in_hand_image"])
+
+                # Save image for video (if enabled)
+                if args.debug_mode:
+                    rollout_images.append(obs["agentview_image"])
+
+                # Execute demo action in environment
+                obs, _reward, done, _info = env.step(action.tolist())
+
+            # At end of episode, save replayed trajectories to new HDF5 files (only keep FAILURES)
+            if done:  # Changed from 'if done:' to 'if not done:' to save failures
+                dones = np.zeros(len(actions)).astype(np.uint8)
+                dones[-1] = 1
+                rewards = np.zeros(len(actions)).astype(np.uint8)
+                # No reward for failure trajectories
+                assert len(actions) == len(agentview_images)
+
+                ep_data_grp = grp.create_group(f"demo_{i}")
+                obs_grp = ep_data_grp.create_group("obs")
+                obs_grp.create_dataset("gripper_states", data=np.stack(gripper_states, axis=0))
+                obs_grp.create_dataset("joint_states", data=np.stack(joint_states, axis=0))
+                obs_grp.create_dataset("ee_states", data=np.stack(ee_states, axis=0))
+                obs_grp.create_dataset("ee_pos", data=np.stack(ee_states, axis=0)[:, :3])
+                obs_grp.create_dataset("ee_ori", data=np.stack(ee_states, axis=0)[:, 3:])
+                obs_grp.create_dataset("agentview_rgb", data=np.stack(agentview_images, axis=0))
+                obs_grp.create_dataset("eye_in_hand_rgb", data=np.stack(eye_in_hand_images, axis=0))
+                ep_data_grp.create_dataset("actions", data=actions)
+                ep_data_grp.create_dataset("states", data=np.stack(states))
+                ep_data_grp.create_dataset("robot_states", data=np.stack(robot_states, axis=0))
+                ep_data_grp.create_dataset("rewards", data=rewards)
+                ep_data_grp.create_dataset("dones", data=dones)
+
+                # Save failure video if debugging is enabled
+                if args.debug_mode:
+                    save_failure_video(rollout_images, task_id, i, task_description)
+                    print("Saved failure video")
+                    exit()
+
+                num_failures += 1
+
+            num_replays += 1
+
+            # Record success/failure and initial environment state in metainfo dict
+            task_key = task_description.replace(" ", "_")
+            episode_key = f"demo_{i}"
+            if task_key not in metainfo_json_dict:
+                metainfo_json_dict[task_key] = {}
+            if episode_key not in metainfo_json_dict[task_key]:
+                metainfo_json_dict[task_key][episode_key] = {}
+            metainfo_json_dict[task_key][episode_key]["success"] = bool(done)
+            metainfo_json_dict[task_key][episode_key]["initial_state"] = orig_states[0].tolist()
+
+            # Write metainfo dict to JSON file
+            # (We repeatedly overwrite, rather than doing this once at the end, just in case the script crashes midway)
+            with open(metainfo_json_out_path, "w") as f:
+                json.dump(metainfo_json_dict, f, indent=2)
+
+            # Count total number of failure replays so far
+            print(
+                f"Total # episodes replayed: {num_replays}, Total # failures: {num_failures} ({num_failures / num_replays * 100:.1f} %)"
+            )
+
+            # Report total number of no-op actions filtered out so far
+            print(f"  Total # no-op actions filtered out: {num_noops}")
+
+        # Close HDF5 files
+        orig_data_file.close()
+        new_data_file.close()
+        print(f"Saved failure demos for task '{task_description}' at: {new_data_path}")
+        env.close()
+
+    print(f"Failure dataset generation complete! Saved new dataset at: {args.libero_target_dir}")
+    print(f"Saved metainfo JSON at: {metainfo_json_out_path}")
+
+
+if __name__ == "__main__":
+    # Parse command-line arguments
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--libero_task_suite",
+        type=str,
+        choices=["libero_spatial", "libero_object", "libero_goal", "libero_10", "libero_90"],
+        help="LIBERO task suite. Example: libero_spatial",
+        required=True,
+    )
+    parser.add_argument(
+        "--libero_raw_data_dir",
+        type=str,
+        help="Path to directory containing raw HDF5 dataset. Example: ./LIBERO/libero/datasets/libero_spatial",
+        required=True,
+    )
+    parser.add_argument(
+        "--libero_target_dir",
+        type=str,
+        help="Path to failure dataset directory. Example: ./LIBERO/libero/datasets/libero_spatial_failure",
+        required=True,
+    )
+    parser.add_argument(
+        "--debug_mode", action="store_true", help="Save failure trajectories as MP4 videos for debugging"
+    )
+    args = parser.parse_args()
+
+    # Start failure dataset generation
+    main(args)
diff --git a/robometer/dataset_upload/data_scripts/molmoact/gen_all_molmoact.sh b/robometer/dataset_upload/data_scripts/molmoact/gen_all_molmoact.sh
new file mode 100644
index 0000000000000000000000000000000000000000..8845db93e18c2ddd39a145d543cd6fdddf5906af
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/molmoact/gen_all_molmoact.sh
@@ -0,0 +1,2 @@
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/molmoact.yaml --dataset.dataset_name molmoact_dataset_tabletop
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/molmoact.yaml --dataset.dataset_name molmoact_dataset_household
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe.sh b/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe.sh
new file mode 100755
index 0000000000000000000000000000000000000000..1b8436544d67f96f44352f55e7f3c6661e1509d3
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe.sh
@@ -0,0 +1,47 @@
+# large datasets near the end
+OXE_VALID_DATASETS=(
+    "austin_buds_dataset_converted_externally_to_rlds"
+    "austin_sirius_dataset_converted_externally_to_rlds"
+    "berkeley_cable_routing"
+    "berkeley_fanuc_manipulation"
+    "dlr_edan_shared_control_converted_externally_to_rlds"
+    "fmb"
+    "furniture_bench_dataset_converted_externally_to_rlds"
+    "iamlab_cmu_pickup_insert_converted_externally_to_rlds"
+    "jaco_play"
+    "stanford_hydra_dataset_converted_externally_to_rlds"
+    "taco_play"
+    "toto"
+    "ucsd_kitchen_dataset_converted_externally_to_rlds"
+    "utaustin_mutex"
+    "viola"
+    "fractal20220817_data"
+    "bc_z"
+    "language_table"
+    "bridge_v2"
+    "droid"
+    # new datasets
+    "robo_set"
+    "aloha_mobile"
+    "tidybot"
+    "imperialcollege_sawyer_wrist_cam"
+    "kaist_nonprehensile_converted_externally_to_rlds"
+    "berkeley_mvp_converted_externally_to_rlds"
+    "berkeley_rpt_converted_externally_to_rlds"
+    "nyu_rot_dataset_converted_externally_to_rlds"
+    "tokyo_u_lsmo_converted_externally_to_rlds"
+)
+
+for dataset_name in ${OXE_VALID_DATASETS[@]}; do
+    echo "Generating dataset: $dataset_name"
+    echo "=" * 100
+    uv run python -m dataset_upload.generate_hf_dataset \
+        --config_path=dataset_upload/configs/data_gen_configs/oxe.yaml \
+        --output.output_dir ./robometer_dataset/oxe_rfm \
+        --dataset.dataset_path ~/tensorflow_datasets/openx_datasets/ \
+        --hub.push_to_hub=false \
+        --dataset.dataset_name oxe_$dataset_name \
+
+    echo "Done generating dataset: $dataset_name"
+    echo "=" * 100
+done
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe_eval.sh b/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe_eval.sh
new file mode 100755
index 0000000000000000000000000000000000000000..8978efb6de44991bb80890f272d40e1ae51f50df
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/oxe/gen_all_oxe_eval.sh
@@ -0,0 +1,36 @@
+# comment out things which have no eval split
+OXE_VALID_DATASETS=(
+    #"austin_buds_dataset_converted_externally_to_rlds"
+    #"austin_sirius_dataset_converted_externally_to_rlds"
+    "berkeley_cable_routing"
+    #"berkeley_fanuc_manipulation"
+    #"dlr_edan_shared_control_converted_externally_to_rlds"
+    #"fmb"
+    #"furniture_bench_dataset_converted_externally_to_rlds"
+    #"iamlab_cmu_pickup_insert_converted_externally_to_rlds"
+    "jaco_play"
+    #"stanford_hydra_dataset_converted_externally_to_rlds"
+    "taco_play"
+    "toto"
+    #"ucsd_kitchen_dataset_converted_externally_to_rlds"
+    #"utaustin_mutex"
+    "viola"
+    #"fractal20220817_data"
+    "bc_z"
+    #"language_table"
+    "bridge_v2"
+    #"droid"
+)
+
+for dataset_name in ${OXE_VALID_DATASETS[@]}; do
+    echo "Generating dataset: $dataset_name"
+    echo "=" * 100
+    uv run dataset_upload/generate_hf_dataset.py \
+        --config_path=dataset_upload/configs/data_gen_configs/oxe.yaml \
+        --output.output_dir ~/scratch_data/oxe_rfm_eval \
+        --dataset.dataset_name oxe_$dataset_name\_eval \
+        --hub.hub_repo_id oxe_rfm_eval
+
+    echo "Done generating dataset: $dataset_name"
+    echo "=" * 100
+done
\ No newline at end of file
diff --git a/robometer/dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh b/robometer/dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
new file mode 100644
index 0000000000000000000000000000000000000000..a7a0a264c47c775212bfc5e620f430a27975f8df
--- /dev/null
+++ b/robometer/dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
@@ -0,0 +1,44 @@
+#!/bin/bash
+source ./venv/bin/activate
+
+datasets=(
+    "usc_koch_throw_the_orange_cup_away_red_trash_can"
+    "usc_koch_suboptimal_throw_the_orange_cup_away_red_trash_can"
+    "usc_koch_failure_throw_the_orange_cup_away_red_trash_can"
+    "usc_koch_throw_the_black_marker_away_blue_trash_can"
+    "usc_koch_suboptimal_throw_the_black_marker_away_blue_trash_can"
+    "usc_koch_failure_throw_the_black_marker_away_blue_trash_can"
+    "usc_koch_open_the_red_trash_bin_red_trash_bin"
+    "usc_koch_suboptimal_open_the_red_trash_bin_red_trash_bin"
+    "usc_koch_failure_open_the_red_trash_bin_red_trash_bin"
+    "usc_koch_open_the_green_trash_bin_green_trash_bin"
+    "usc_koch_suboptimal_open_the_green_trash_bin_green_trash_bin"
+    "usc_koch_failure_open_the_green_trash_bin_green_trash_bin"
+    "usc_koch_open_the_blue_trash_bin_blue_trash_bin"
+    "usc_koch_suboptimal_open_the_blue_trash_bin_blue_trash_bin"
+    "usc_koch_failure_open_the_blue_trash_bin_blue_trash_bin"
+    "usc_koch_separate_the_red_and_orange_and_orange_cups"
+    "usc_koch_suboptimal_separate_the_red_and_orange_and_orange_cups"
+    "usc_koch_failure_separate_the_red_and_orange_and_orange_cups"
+    "usc_koch_separate_the_purple_and_orange_and_orange_cups"
+    "usc_koch_suboptimal_separate_the_purple_and_orange_and_orange_cups"
+    "usc_koch_failure_separate_the_purple_and_orange_and_orange_cups"
+    "usc_koch_separate_the_purple_and_red_and_red_cups"
+    "usc_koch_suboptimal_separate_the_purple_and_red_and_red_cups"
+    "usc_koch_failure_separate_the_purple_and_red_and_red_cups"
+    "usc_koch_move_the_orange_cup_from_right_to_left"
+    "usc_koch_suboptimal_move_the_orange_cup_from_right_to_left"
+    "usc_koch_failure_move_the_orange_cup_from_right_to_left"
+    "usc_koch_move_the_orange_cup_from_left_to_right"
+    "usc_koch_suboptimal_move_the_orange_cup_from_left_to_right"
+    "usc_koch_failure_move_the_orange_cup_from_left_to_right"
+)
+
+for dataset in ${datasets[@]}; do
+    hf download abraranwar/${dataset} --repo-type=dataset --local-dir=./datasets/usc_koch_human_robot_paired/robot/${dataset}
+done
+
+
+gdown --fuzzy https://drive.google.com/file/d/1VsV9GN784WGmuNYtMeIduXfk72WSaS1E/view?usp=drive_link
+unzip recordings.zip -d ./datasets/usc_koch_human_robot_paired/human
+rm recordings.zip
diff --git a/robometer/dataset_upload/dataset_guides/AgiBotWorld.md b/robometer/dataset_upload/dataset_guides/AgiBotWorld.md
new file mode 100755
index 0000000000000000000000000000000000000000..ad03344e382b60876df273800d43ad7b0f21f59c
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/AgiBotWorld.md
@@ -0,0 +1,194 @@
+# AgiBotWorld Dataset Guide
+
+AgiBotWorld is a large-scale robotic dataset with **native streaming support** and **advanced video processing** for the Robometer training pipeline.
+
+## Overview
+
+- **🚀 Streaming Support**: Process without downloading the 600GB+ full dataset
+- **🎯 Head Camera Focus**: Extracts only `head_color.mp4` videos
+- **📹 Video Processing**: Automatic resize to 256x256 + frame interpolation during dataset generation
+- **📏 Standardized Output**: Configurable frame count (default: 32 frames)
+- **💾 Optimized Storage**: 99%+ size reduction (15MB → ~100KB per video)
+- **🏷️ Descriptive Task Names**: Extracts proper task descriptions from JSON metadata
+- **⚡ Efficient Processing**: Uses pre-encoded MP4 data directly
+- **🔄 Graceful Error Handling**: Skips corrupted samples automatically
+- **📊 Webdataset Format**: Handles HuggingFace webdataset format natively
+
+## Prerequisites
+
+### 0. Set Hugging Face repo ID
+Before we start, you must have an HF account which will be pushed to.
+You will set this by setting
+```
+export HF_USERNAME=<insert HF username here>
+```
+
+### 1. HuggingFace Authentication
+```bash
+uv run hf auth login
+```
+
+### 2. Accept Dataset License
+Visit [https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha](https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha) and accept the license agreement.
+
+### 3. Download task information which gets put in a temporary folder. This is so we can index task and subtask information.
+```
+ uv run dataset_upload/data_scripts/agibot/download_task_jsons.py 
+ ```
+ 
+## Quick Start
+
+### Option 1: Use Pre-configured Settings
+```bash
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen_configs/agibot_world.yaml
+```
+
+### Option 2: Manual Configuration
+```bash
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/data_gen.yaml \
+    --dataset.dataset_name=agibotworld \
+    --dataset.dataset_path="agibot-world/AgiBotWorld-Alpha" \
+    --output.output_dir=agibotworld_dataset \
+    --output.max_trajectories=100 \
+    --output.max_frames=32 \
+    --output.use_video=true \
+    --output.fps=10
+```
+
+### Option 3: Local Dataset Processing
+```bash
+uv run python data/generate_hf_dataset.py \
+    --dataset.dataset_name=agibotworld_local \
+    --dataset.dataset_path="/path/to/AgiBotWorld-Alpha/sample_dataset" \
+    --output.max_trajectories=50 \
+    --output.max_frames=16 \
+    --hub.push_to_hub=false
+```
+
+## Video Processing Features
+
+The AgiBotWorld loader automatically processes videos during dataset generation with the following optimizations:
+
+### Processing Pipeline
+1. **📹 Frame Extraction**: Loads video frames from MP4 files or bytes
+2. **📐 Resize**: All frames resized to **256x256 pixels**
+3. **⏱️ Frame Interpolation**: Downsamples to `max_frames` using linear interpolation
+4. **🎬 Re-encoding**: Saves as optimized MP4 bytes
+
+### Performance Benefits
+- **Original**: ~15MB per video, 1740+ frames, variable resolution
+- **Processed**: ~87-131KB per video, 16-32 frames, 256x256 resolution
+- **Reduction**: **99%+ size reduction** for efficient training
+
+### Configurable Parameters
+- `max_frames`: Number of frames to keep (default: 32)
+- `target_size`: Resolution (fixed at 256x256 for AgiBotWorld)
+- `fps`: Output video frame rate (default: 10)
+
+## Configuration Options
+
+Edit `configs/data_gen_configs/agibot_world.yaml`:
+
+```yaml
+dataset:
+  dataset_path: "agibot-world/AgiBotWorld-Alpha"  # HuggingFace dataset name
+  dataset_name: agibotworld
+
+output:
+  output_dir: agibotworld_dataset
+  max_trajectories: 100  # Increase for more data (up to ~100k)
+  max_frames: 32
+  use_video: true
+  fps: 10
+
+hub:
+  push_to_hub: false  # Set to true to upload results
+  hub_repo_id: your-username/agibotworld_rfm
+```
+
+## Data Structure Processed
+
+```
+AgiBotWorld (Local):
+├── head_color.mp4 videos    ← EXTRACTED + PROCESSED (15MB → ~100KB each)
+├── task_info/*.json         ← PARSED for descriptive task names
+├── proprio_stats/*.h5       ← LOADED for robot actions
+├── depth images             ← SKIPPED
+└── other camera views       ← SKIPPED
+
+AgiBotWorld (Streaming):
+├── head_color.mp4 videos    ← EXTRACTED + PROCESSED (31MB → ~100KB each)
+├── depth images             ← SKIPPED
+├── other camera views       ← SKIPPED  
+├── task descriptions        ← PARSED from webdataset keys
+└── robot actions            ← PLACEHOLDER (H5 data not available in streaming)
+```
+
+## Sample Output
+
+### Local Dataset Processing
+```
+Processing task 446: 'Dual-robot table carrying'
+  📹 Processed video: 1740 -> 32 frames, resized to (256, 256)
+  ✅ Loaded episode 687616 (1/50)
+Added 1 trajectories for task 'Dual-robot table carrying'
+```
+
+### Streaming Dataset Processing
+```
+✅ Found valid head camera video #1: 648642/videos/head_color (task 0, episode 0, 31729374 bytes)
+  📹 Processed video: 1455 -> 32 frames, resized to (256, 256)
+Processed 8 valid samples from 9 total samples
+```
+
+## Performance Notes
+
+- **Processing Rate**: ~1-2 samples/second (depends on network)
+- **Memory Usage**: Low (streaming approach)
+- **Storage**: ~30MB per trajectory (video data)
+- **Error Rate**: ~10-20% samples skipped due to webdataset format issues (normal)
+
+## Troubleshooting
+
+### Authentication Issues
+```bash
+# Check if logged in
+uv run hf auth whoami
+
+# Re-login if needed
+uv run hf auth login
+```
+
+### License Access
+Make sure you've accepted the license at the dataset page. The error will show:
+```
+403 Forbidden: Authorization error
+```
+
+### Schema Casting Errors
+These are normal and handled gracefully:
+```
+Skipping sample due to casting error: Couldn't cast
+mp4: null
+```
+
+### Large Scale Processing
+For processing thousands of trajectories:
+```bash
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/data_gen_configs/agibot_world.yaml \
+    --output.max_trajectories=5000 \
+    --hub.push_to_hub=false  # Keep local until ready
+```
+
+## Integration with Robometer Training
+
+The generated dataset is compatible with the standard Robometer training pipeline:
+
+```bash
+# Use the processed dataset for training
+uv run accelerate launch --config_file configs/fsdp.yaml train.py \
+    --config_path=configs/config.yaml \
+    --dataset.dataset_path=agibotworld_dataset/agibotworld
+```
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_guides/AutoEval.md b/robometer/dataset_upload/dataset_guides/AutoEval.md
new file mode 100644
index 0000000000000000000000000000000000000000..74802d6b94938ae62fd0c3b1c795addf9a8322bc
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/AutoEval.md
@@ -0,0 +1,74 @@
+# AutoEval Dataset Guide
+
+This guide explains how to load and convert the AutoEval pickled evaluation episodes into the Robometer pipeline.
+
+Source: `https://huggingface.co/datasets/zhouzypaul/auto_eval`
+
+## Prerequisites
+
+1) Install the logger library first (may need to make a minor hardcoded moviepy import tweak in one file to make things work):
+
+```bash
+git clone https://github.com/zhouzypaul/robot_eval_logger
+cd robot_eval_logger
+uv pip install -e .
+```
+
+2) Download the dataset locally so that it contains an `eval_data/` directory with subfolders per group, each containing pickled episodes.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  eval_data/
+    00001/
+      episode_00001_success.pkl
+      episode_00001_fail.pkl
+    00002/
+      ...
+```
+
+We expect per-episode pickle files. Success/failure may be in the filename suffix or inside the pickle (`success` flag). Only paired (success and fail) are kept.
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/autoeval_loader.py`
+- Function: `load_autoeval_dataset(dataset_path: str) -> dict[str, list[dict]]`
+- For each episode group, decodes frames from `obs['image_primary']` and records `success`.
+- Prints totals for successes, failures, and kept pairs. Only paired entries are returned.
+
+## Configuration (configs/data_gen_configs/autoeval.yaml)
+
+```yaml
+# configs/data_gen_configs/autoeval.yaml
+
+dataset:
+  dataset_path: ./datasets/autoeval
+  dataset_name: autoeval
+
+output:
+  output_dir: ./robometer_dataset/autoeval_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 2
+
+hub:
+  push_to_hub: true
+  hub_repo_id: autoeval_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/autoeval.yaml
+```
+
+This will:
+- Load pickles from `eval_data/`
+- Extract `image_primary` frames per step
+- Keep only paired success/failure episodes
+- Create a HF dataset with relative video paths
diff --git a/robometer/dataset_upload/dataset_guides/CustomDataset.md b/robometer/dataset_upload/dataset_guides/CustomDataset.md
new file mode 100755
index 0000000000000000000000000000000000000000..c2bdb0bb34763f53415fa305d818bdad1cd6d5c1
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/CustomDataset.md
@@ -0,0 +1,276 @@
+# Custom Dataset Guide
+
+Learn how to add your own dataset (like DROID, Bridge, etc.) to the Robometer training pipeline.
+
+## Overview
+
+Adding a custom dataset involves:
+1. Creating a dataset loader module
+2. Implementing the required data format
+3. Integrating with the main converter
+4. Testing and validation
+
+## Required Data Format
+
+Your dataset loader must produce trajectories in this format:
+
+```python
+{
+    'frames': List[Union[str, bytes, np.ndarray]],  # Video file paths, MP4 bytes, or frame arrays
+    'actions': np.ndarray,                          # Robot actions (N, action_dim)
+    'is_robot': bool,                               # True for robot data, False for human
+    'task': str,                                    # Human-readable task description
+    'optimal': str                                  # "optimal", "suboptimal", or "failed"
+}
+```
+
+## Step 0: Set Hugging Face repo ID
+Before we start, you must have an HF account which will be pushed to.
+You will set this by setting
+```
+export HF_USERNAME=<insert HF username here>
+```
+
+
+## Step 1: Create Dataset Loader
+
+Create `data/{dataset_name}_loader.py`:
+
+```python
+#!/usr/bin/env python3
+"""
+{DatasetName} dataset loader for Robometer model training.
+"""
+
+import os
+import numpy as np
+from typing import List, Dict
+from pathlib import Path
+from tqdm import tqdm
+
+def load_{dataset_name}_dataset(base_path: str) -> Dict[str, List[Dict]]:
+    """Load {DatasetName} dataset and organize by task.
+    
+    Args:
+        base_path: Path to the {DatasetName} dataset directory
+        
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+    
+    print(f"Loading {DatasetName} dataset from: {base_path}")
+    print("=" * 100)
+    print(f"LOADING {DATASET_NAME} DATASET")
+    print("=" * 100)
+    
+    task_data = {}
+    base_path = Path(base_path)
+    
+    if not base_path.exists():
+        raise FileNotFoundError(f"{DatasetName} dataset path not found: {base_path}")
+    
+    # TODO: Implement your dataset-specific logic here
+    # Example structure:
+    
+    # Find your data files
+    data_files = list(base_path.glob("**/*.{your_format}"))  # e.g., *.pkl, *.json, *.hdf5
+    
+    for file_path in tqdm(data_files, desc=f"Processing {DatasetName} dataset"):
+        task_name = file_path.stem  # or extract from your naming scheme
+        
+        # Load your data file
+        trajectories = load_trajectories_from_file(file_path)
+        
+        if trajectories:
+            task_data[task_name] = trajectories
+    
+    print(f"Loaded {sum(len(trajectories) for trajectories in task_data.values())} trajectories from {len(task_data)} tasks")
+    return task_data
+
+
+def load_trajectories_from_file(file_path: Path) -> List[Dict]:
+    """Load trajectories from a single data file."""
+    trajectories = []
+    
+    # TODO: Implement your file loading logic here
+    # This depends on your dataset format (HDF5, JSON, pickle, etc.)
+    
+    # Example for different formats:
+    
+    # For HDF5:
+    # import h5py
+    # with h5py.File(file_path, 'r') as f:
+    #     # Navigate your HDF5 structure
+    #     pass
+    
+    # For JSON:
+    # import json
+    # with open(file_path, 'r') as f:
+    #     data = json.load(f)
+    
+    # For pickle:
+    # import pickle
+    # with open(file_path, 'rb') as f:
+    #     data = pickle.load(f)
+    
+    # Convert to required format:
+    # trajectory = {
+    #     'frames': your_frames,        # List of file paths, bytes, or numpy arrays
+    #     'actions': your_actions,      # numpy array of shape (sequence_length, action_dim)
+    #     'is_robot': True,            # or False for human demonstrations
+    #     'task': your_task_description,
+    #     'optimal': 'optimal'         # or 'suboptimal', 'failed'
+    # }
+    # trajectories.append(trajectory)
+    
+    return trajectories
+```
+
+## Step 2: Add to Main Converter
+
+Edit `data/generate_hf_dataset.py` to include your dataset:
+
+```python
+# In the main() function, add your dataset type:
+elif "{dataset_name}" in cfg.dataset.dataset_name.lower():
+    from {dataset_name}_loader import load_{dataset_name}_dataset
+    # Load the trajectories using your loader
+    task_data = load_{dataset_name}_dataset(cfg.dataset.dataset_path)
+    trajectories = flatten_task_data(task_data)
+```
+
+## Step 3: Create Configuration File
+
+Create `configs/data_gen_configs/{dataset_name}.yaml`:
+
+```yaml
+# {DatasetName} dataset configuration
+dataset:
+  dataset_path: /path/to/your/{dataset_name}/dataset
+  dataset_name: {dataset_name}
+
+output:
+  output_dir: {dataset_name}_dataset
+  max_trajectories: 1000  # Adjust as needed
+  max_frames: 32
+  use_video: true
+  fps: 10
+
+hub:
+  push_to_hub: false
+  hub_repo_id: your-username/{dataset_name}_rbm
+```
+
+## Step 4: Test Your Implementation
+
+Create a test script `test_{dataset_name}_loader.py`:
+
+```python
+from {dataset_name}_loader import load_{dataset_name}_dataset
+from helpers import flatten_task_data
+
+# Test your loader
+task_data = load_{dataset_name}_dataset("/path/to/your/{dataset_name}/dataset")
+trajectories = flatten_task_data(task_data)
+
+print(f"Loaded {len(trajectories)} trajectories")
+print(f"Sample trajectory keys: {list(trajectories[0].keys())}")
+print(f"Sample task: {trajectories[0].get('task', 'No task found')}")
+print(f"Sample frames type: {type(trajectories[0]['frames'])}")
+print(f"Sample actions shape: {trajectories[0]['actions'].shape}")
+```
+
+## Step 5: Run Dataset Conversion
+
+```bash
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/data_gen_configs/{dataset_name}.yaml
+```
+
+## Frame Format Options
+
+### Option 1: Video File Paths
+```python
+'frames': ['/path/to/video1.mp4', '/path/to/video2.mp4']
+```
+
+### Option 2: Raw Video Bytes (for streaming)
+```python
+'frames': video_bytes  # bytes object containing MP4 data
+```
+
+### Option 3: Frame Arrays
+```python
+'frames': np.array([frame1, frame2, ...])  # shape: (seq_len, H, W, 3)
+```
+
+## Common Dataset Formats
+
+### HDF5 Datasets (like LIBERO)
+```python
+import h5py
+with h5py.File(file_path, 'r') as f:
+    frames = f['observations']['camera_data'][:]
+    actions = f['actions'][:]
+```
+
+### JSON + Video Files
+```python
+import json
+with open(metadata_file, 'r') as f:
+    metadata = json.load(f)
+    
+video_path = base_path / metadata['video_file']
+frames = [str(video_path)]  # Let converter handle video loading
+```
+
+### Pickle Files
+```python
+import pickle
+with open(file_path, 'rb') as f:
+    episode_data = pickle.load(f)
+    frames = episode_data['observations']['images']
+    actions = episode_data['actions']
+```
+
+## Error Handling
+
+Add robust error handling to your loader:
+
+```python
+try:
+    # Your data loading code
+    pass
+except Exception as e:
+    print(f"Error loading {file_path}: {e}")
+    continue  # Skip problematic files
+```
+
+## Performance Tips
+
+1. **Use tqdm** for progress bars on long operations
+2. **Validate data shapes** before adding to trajectories
+3. **Handle missing files gracefully**
+4. **Use generators** for memory efficiency with large datasets
+5. **Cache expensive operations** when possible
+
+## Example Datasets to Reference
+
+- **LIBERO**: `data/libero_loader.py` - HDF5 format
+- **AgiBotWorld**: `data/agibotworld_loader.py` - Streaming format
+
+## Integration Testing
+
+After implementation, test with the training pipeline:
+
+```bash
+# Test dataset loading
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/data_gen_configs/{dataset_name}.yaml \
+    --output.max_trajectories=10
+
+# Test training integration  
+uv run accelerate launch --config_file configs/fsdp.yaml train.py \
+    --config_path=configs/config.yaml \
+    --dataset.dataset_path={dataset_name}_dataset/{dataset_name}
+```
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_guides/EPIC-KITCHENS.md b/robometer/dataset_upload/dataset_guides/EPIC-KITCHENS.md
new file mode 100644
index 0000000000000000000000000000000000000000..8454a79c432fb69bed73f30d206f859027d8d361
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/EPIC-KITCHENS.md
@@ -0,0 +1,91 @@
+# EPIC-KITCHENS Dataset Guide
+
+This guide explains how to integrate and use the EPIC-KITCHENS dataset with the Robometer training pipeline.
+
+Source: `https://epic-kitchens.github.io/`, Lerobot dataset sourced from `https://huggingface.co/datasets/awsaf49/epic_kitchens_100`, (a subset of the original), and `csv` annotations sourced from `https://github.com/epic-kitchens/epic-kitchens-100-annotations/blob/master/EPIC_100_train.csv`.
+
+This only loads the Epic-kitchens-100 added data, not any of the original epic-kitchens-55 data as that's difficult to download without a HF link (very slow).
+
+## Assumptions
+
+- The dataset root contains participant folders `P01`, `P02`, ...
+- Each participant folder contains a `videos/` directory with `.MP4` files
+- The dataset root contains `EPIC_100_train.csv` with at least columns:
+  - `video_id` (maps to the video filename without `.MP4`)
+  - `participant_id` (e.g., `P01`)
+  - `start_frame` (start frame index)
+  - `stop_frame` (end frame index)
+  - `narration` (language instruction)
+
+## Directory Structure
+
+```
+<dataset_path>/
+  EPIC_100_train.csv
+  P01/
+    videos/
+      <narration_id>.MP4
+  P02/
+    videos/
+      <narration_id>.MP4
+  ...
+```
+
+## Configuration (configs/data_gen_configs/epic.yaml)
+
+```yaml
+# configs/data_gen_configs/epic.yaml
+
+dataset:
+  dataset_path: ./datasets/epic_kitchens
+  dataset_name: epic
+
+output:
+  output_dir: ./robometer_dataset/epic_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: epic_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/epic.yaml
+```
+
+This will:
+- Parse `EPIC_100_train.csv`
+- For each row, locate the participant video and extract the `[start_frame, stop_frame)` segment
+- Convert segments to web-optimized videos
+- Create a HuggingFace dataset (H2R/OXE-style streaming conversion)
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: The `narration` string from the CSV
+- `frames`: Relative path to the generated clip video
+- `is_robot`: False
+- `quality_label`: "successful"
+- `partial_success`: N/A (fixed by pipeline)
+- `data_source`: `epic`
+
+## Notes
+
+- The loader uses batched, worker-parallel conversion similar to `h2r_loader.py` to control memory.
+- Frame indices are respected using OpenCV frame seeking; if seeking fails, decoding starts from the beginning and stops at `stop_frame`.
+- The language vector is computed from the narration and cached.
+
+## Troubleshooting
+
+- Video not found: Ensure `narration_id`.MP4 exists under the correct `PXX/videos/` directory.
+- Empty segment: Verify `start_frame < stop_frame` and within the video length.
+- Performance: Adjust `num_workers` and `batch_size` in the loader if necessary.
diff --git a/robometer/dataset_upload/dataset_guides/EgoCoT.md b/robometer/dataset_upload/dataset_guides/EgoCoT.md
new file mode 100644
index 0000000000000000000000000000000000000000..6ee54db91a5f33931240767664d24ad56a7ad8a3
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/EgoCoT.md
@@ -0,0 +1,202 @@
+# EgoCoT Dataset Guide
+
+This guide provides detailed instructions for integrating and using the EgoCoT (Egocentric Chain-of-Thought) dataset with the Robometer training pipeline.
+
+Location: https://github.com/EmbodiedGPT/EgoCOT_Dataset
+
+## Overview
+
+The EgoCoT dataset is an egocentric video dataset designed for embodied AI research. It contains sequences of 8 consecutive frames from egocentric videos, paired with captions and embodied planning information. The dataset focuses on understanding and reasoning about human actions from a first-person perspective.
+
+## Data Characteristics
+
+*   **Type**: Human Egocentric Video Data
+*   **Format**: NumPy arrays for frames, JSON files for annotations
+*   **Size**: Variable, depending on the dataset split and filtering criteria
+*   **Features**:
+    *   **Frame Sequences**: 8 consecutive frames stored as NumPy arrays
+    *   **Captions**: Brief descriptions of the video content
+    *   **Embodied Planning**: Detailed action sequences and reasoning for achieving goals
+    *   **Alignment Scores**: Numeric scores (0.0-1.0) indicating quality of video-planning alignment
+
+## Prerequisites
+
+1.  **Download the dataset**: Ensure you have the EgoCoT dataset downloaded and accessible locally. The dataset should contain JSON annotation files and corresponding NumPy frame files.
+
+2.  **Directory Structure**: The loader expects a structure similar to this:
+
+    ```
+    <dataset_path>/
+    ├── annotations.json (or multiple JSON files)
+    ├── data/
+    │   ├── frame_sequence_001.npy
+    │   ├── frame_sequence_002.npy
+    │   └── ...
+    └── (additional subdirectories as needed)
+    ```
+
+    Alternative structures are also supported:
+    ```
+    <dataset_path>/
+    ├── split1/
+    │   ├── annotations.json
+    │   ├── frame_001.npy
+    │   └── frame_002.npy
+    └── split2/
+        ├── annotations.json
+        └── ...
+    ```
+
+## JSON Annotation Format
+
+Each JSON file should contain annotations in one of these formats:
+
+**List format:**
+```json
+[
+  {
+    "image": "frame_sequence_001.npy",
+    "caption": "Person preparing breakfast in the kitchen",
+    "planning": "1. Open refrigerator 2. Take out eggs 3. Crack eggs into bowl...",
+    "score": 0.85
+  },
+  ...
+]
+```
+
+**Dictionary format:**
+```json
+{
+  "data": [
+    {
+      "image": "frame_sequence_001.npy",
+      "caption": "Person preparing breakfast in the kitchen", 
+      "planning": "1. Open refrigerator 2. Take out eggs 3. Crack eggs into bowl...",
+      "score": 0.85
+    },
+    ...
+  ]
+}
+```
+
+**Field descriptions:**
+- `image`/`frames`/`video`: Filename of the NumPy array containing frame sequence
+- `caption`/`description`/`task`: Brief description of the video content
+- `planning`/`plan`/`actions`: Embodied planning information
+- `score`/`quality`/`alignment`: Alignment score between video and planning (0.0-1.0)
+
+## Configuration (configs/data_gen_configs/egocot.yaml)
+
+To use the EgoCoT dataset, create a configuration file. Here's the provided `egocot.yaml`:
+
+```yaml
+# configs/data_gen_configs/egocot.yaml
+
+dataset:
+  dataset_path: ./datasets/egocot  # Path to EgoCoT dataset directory
+  dataset_name: egocot
+
+output:
+  output_dir: ./robometer_dataset/egocot_rfm
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 8  # EgoCoT uses 8 consecutive frames
+  use_video: true
+  fps: 30  # Standard fps for egocentric video
+  shortest_edge_size: 224  # Standard size for vision models
+  center_crop: true
+
+hub:
+  push_to_hub: true
+  hub_repo_id: egocot_rfm
+```
+
+**Configuration Parameters:**
+*   **`dataset_path`**: Path to the root directory of your EgoCoT dataset
+*   **`dataset_name`**: Must be `egocot` to use the provided loader
+*   **`max_frames`**: Set to 8 to match EgoCoT's frame sequence length
+*   **`filters`**: Optional filtering based on alignment scores and quality labels
+
+## Generating the HuggingFace Dataset
+
+Once your configuration is set up, generate the HuggingFace dataset using:
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/egocot.yaml
+```
+
+This command will:
+
+1.  Load the EgoCoT data using `egocot_loader.py`
+2.  Process the frame sequences and annotations
+3.  Apply any specified filters (score thresholds, quality labels)
+4.  Convert the data into HuggingFace dataset format
+5.  Save the generated dataset
+
+## Data Fields in the HuggingFace Dataset
+
+The resulting HuggingFace dataset will contain the following fields:
+
+*   `id`: Unique identifier for each trajectory
+*   `task`: Caption/description of the video content
+*   `frames`: An `EgoCOTFrameloader` object that loads the 8-frame sequence on demand, returning `np.ndarray` of shape `(8, H, W, 3)` with `dtype uint8`
+*   `is_robot`: Boolean, always `False` for EgoCoT (human egocentric data)
+*   `quality_label`: "high_quality" (score ≥ 0.7), "medium_quality" (0.4 ≤ score < 0.7), or "low_quality" (score < 0.4)
+*   `partial_success`: The alignment score between video and planning (0.0-1.0)
+*   `planning`: Embodied planning information describing action sequences
+*   `data_source`: Always "egocot"
+
+## Frame Format
+
+The EgoCoT dataset stores frames as NumPy arrays. The loader handles different possible formats:
+
+*   **4D arrays**: `(T, H, W, 3)` - preferred format
+*   **3D arrays**: `(H, W, C*T)` - automatically reshaped to 4D
+*   **Data types**: Supports both uint8 and float formats, automatically converting to uint8
+
+## Quality Filtering
+
+The dataset includes alignment scores that can be used for quality filtering:
+
+*   **High quality** (score ≥ 0.7): Strong alignment between video and planning
+*   **Medium quality** (0.4 ≤ score < 0.7): Moderate alignment
+*   **Low quality** (score < 0.4): Weak alignment
+
+Use the `filters` section in the YAML config to specify minimum quality thresholds.
+
+## Troubleshooting
+
+*   **`FileNotFoundError`**: Ensure the `dataset_path` correctly points to your EgoCoT dataset directory
+*   **`ValueError: Unexpected frames shape`**: Check that your NumPy frame files contain valid image data in the expected format
+*   **Missing JSON files**: Ensure annotation files are present and contain the required fields (`image`, `caption`, `planning`, `score`)
+*   **Frame file not found**: Verify that the paths in JSON annotations correctly reference existing NumPy files
+*   **Memory issues**: For large datasets, consider using `max_trajectories` to limit the number of loaded samples during testing
+
+## Usage Example
+
+```python
+from dataset_upload.dataset_loaders.egocot_loader import load_egocot_dataset
+
+# Load the dataset
+dataset_path = "./datasets/egocot"
+task_data = load_egocot_dataset(dataset_path)
+
+# Access trajectories
+for task_name, trajectories in task_data.items():
+    print(f"Task: {task_name}")
+    for traj in trajectories:
+        frames = traj["frames"]()  # Load frames on demand
+        print(f"  Trajectory {traj['id']}: {frames.shape}")
+        print(f"  Planning: {traj['planning'][:100]}...")
+        print(f"  Quality: {traj['quality_label']} (score: {traj['partial_success']})")
+```
+
+## Integration with Robometer Pipeline
+
+The EgoCoT loader is fully compatible with the Robometer training pipeline. The egocentric nature of the data makes it particularly suitable for:
+
+*   **First-person action recognition**
+*   **Embodied planning and reasoning**
+*   **Human activity understanding**
+*   **Multi-modal learning with vision and language**
+
+The 8-frame sequences provide temporal context while remaining computationally manageable for training large vision-language models.
diff --git a/robometer/dataset_upload/dataset_guides/EgoDex.md b/robometer/dataset_upload/dataset_guides/EgoDex.md
new file mode 100755
index 0000000000000000000000000000000000000000..6c037977f7ad4d231933c33ebf3e3c3c5998453d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/EgoDex.md
@@ -0,0 +1,326 @@
+# EgoDex Dataset Guide
+
+EgoDex is a large-scale egocentric manipulation dataset with **native iterative loading** and **advanced video processing** for the Robometer training pipeline.
+
+## Overview
+
+- **🔄 Iterative Loading**: Process trajectories one at a time without loading everything into memory
+- **🎯 Egocentric Focus**: 1080p 30Hz egocentric video with 3D pose annotations  
+- **📹 Video Processing**: Automatic resize to 256x256 + frame downsampling during dataset generation
+- **📏 Standardized Output**: Configurable frame count (default: 32 frames)
+- **💾 Memory Efficient**: Iterator-based loading for large datasets (800+ hours)
+- **🏷️ Rich Annotations**: LLM-generated task descriptions from HDF5 metadata
+- **⚡ Efficient Processing**: Direct MP4 + HDF5 processing
+- **🔄 Graceful Error Handling**: Skips corrupted samples automatically
+- **🔄 Flexible Processing**: Process different dataset parts by pointing to specific directories
+
+## Dataset Structure
+
+The EgoDex dataset consists of paired HDF5 and MP4 files organized by task. Each dataset part (part1, part2, etc.) should be processed separately by pointing directly to that directory:
+
+```
+part1/  # Point dataset_path here for part1
+├── task1/
+│   ├── 0.hdf5      # Pose annotations
+│   ├── 0.mp4       # Egocentric video
+│   ├── 1.hdf5
+│   ├── 1.mp4
+│   └── ...
+├── task2/
+│   └── ...
+└── ...
+
+part2/  # Point dataset_path here for part2
+├── task1/
+│   └── ...
+└── ...
+```
+
+## Prerequisites
+
+### 0. Set Hugging Face repo ID
+Before we start, you must have an HF account which will be pushed to.
+You will set this by setting
+```
+export HF_USERNAME=<insert HF username here>
+```
+
+### 1. Dataset Download
+Download and extract the EgoDex dataset splits you want to process:
+- `part1.zip` through `part5.zip` (training data, <350GB each)
+- `test.zip` (test data, 1% of dataset)
+- `extra.zip` (additional samples)
+
+### 2. Python Dependencies
+The loader uses standard dependencies already in the project:
+```bash
+# Already included in pyproject.toml
+pip install h5py opencv-python numpy tqdm
+```
+
+## Quick Start
+
+### Option 1: Use Pre-configured Settings (Test Split)
+```bash
+bash dataset_upload/data_scripts/egodex/download_and_convert.sh
+
+# or run the following manually
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml
+```
+
+### Option 2: Manual Configuration
+```bash
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+    --dataset.dataset_path="/path/to/egodex/test" \
+    --dataset.dataset_name="egodex_test" \
+    --output.max_trajectories=50 \
+    --output.max_frames=32
+```
+
+### Option 3: Process Different Parts
+```bash
+# Process part1
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+    --dataset.dataset_path="/path/to/egodex/part1" \
+    --dataset.dataset_name="egodex_part1" \
+    --output.max_trajectories=-1 \
+    --hub.push_to_hub=false
+
+# Process part2 with different dataset name
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+    --dataset.dataset_path="/path/to/egodex/part2" \
+    --dataset.dataset_name="egodex_part2" \
+    --output.max_trajectories=-1 \
+    --hub.push_to_hub=false
+```
+
+## Configuration Options
+
+Edit `dataset_upload/configs/data_gen_configs/egodex.yaml`:
+
+```yaml
+dataset:
+  dataset_path: "/path/to/egodex/test"  # Local path to dataset part (e.g., test, part1, part2)
+  dataset_name: egodex_test  # Use different names: egodex_test, egodex_part1, egodex_part2, etc.
+
+output:
+  output_dir: egodex_dataset
+  max_trajectories: 100  # Increase for more data (thousands available)
+  max_frames: 32
+  use_video: true
+  fps: 30
+
+hub:
+  push_to_hub: false  # Set to true to upload results
+  hub_repo_id: your-username/egodex_rfm
+```
+
+## Video Processing Features
+
+The EgoDex loader automatically processes videos during dataset generation:
+
+### Processing Pipeline
+1. **📹 Frame Extraction**: Loads 1080p 30Hz video frames from MP4 files
+2. **📐 Resize**: All frames resized to **256x256 pixels**
+3. **⏱️ Frame Downsampling**: Reduces to `max_frames` using linear interpolation
+4. **🎬 Re-encoding**: Saves as optimized MP4 bytes
+
+### Performance Benefits
+- **Original**: ~50-200MB per video, 900-6000+ frames, 1080p resolution
+- **Processed**: ~100-500KB per video, 16-32 frames, 256x256 resolution
+- **Reduction**: **99%+ size reduction** for efficient training
+
+## Pose Data Processing
+
+EgoDex includes rich 3D pose annotations that are used as action data:
+
+### Extracted Pose Features
+- **Hand Positions**: Left/right hand 3D positions (primary actions)
+- **Finger Tips**: Index finger tip positions for fine manipulation
+- **Camera Extrinsics**: Head/camera pose for egocentric context
+- **Confidence Scores**: ARKit confidence values (when available)
+
+### Action Data Structure
+```python
+# Pose data extracted as actions (T, D) where:
+# T = number of frames
+# D = concatenated pose dimensions (typically 6-15D)
+actions = [
+    left_hand_pos,      # (T, 3)
+    right_hand_pos,     # (T, 3) 
+    finger_tips,        # (T, 6) - both hands
+    camera_pose         # (T, 3) - optional
+]
+```
+
+## Task Descriptions
+
+EgoDex includes LLM-generated task descriptions stored in HDF5 metadata:
+
+### Metadata Fields
+- `llm_description`: Primary task description
+- `llm_description2`: Alternative description (for reversible tasks)
+- `which_llm_description`: Indicates which description applies (1 or 2)
+
+### Example Descriptions
+```
+"Pick up the red block and place it on the blue plate"
+"Organize the tools by placing them in the designated slots"
+"Pour water from the pitcher into the glass"
+```
+
+## Memory Management
+
+The EgoDex loader uses an iterator pattern for memory efficiency:
+
+### Iterator Benefits
+- **Low Memory**: Only one trajectory loaded at a time
+- **Scalable**: Handle 800+ hours of data on modest hardware
+- **Interruptible**: Can stop/resume processing at any point
+- **Progress Tracking**: Real-time progress updates
+
+### Usage Example
+```python
+from egodex_loader import get_egodex_iterator
+
+# Create iterator
+iterator = get_egodex_iterator(
+    dataset_path="/path/to/egodex/test",
+    max_trajectories=100
+)
+
+# Process one trajectory at a time
+for trajectory in iterator:
+    frames = trajectory['frames']      # Video frames
+    actions = trajectory['actions']    # Pose data
+    task = trajectory['task']         # Task description
+    # Process trajectory...
+```
+
+## Sample Output
+
+### Processing Progress
+```
+
+Will process up to 100 trajectories
+
+Loading trajectories: 100%|████████| 100/100 [02:34<00:00,  1.54it/s]
+  📊 Loaded 100 trajectories from 47 tasks
+Loaded 100 trajectories from 47 tasks
+```
+
+### Trajectory Processing
+```
+Processing trajectories: 100%|████████| 100/100 [08:42<00:00,  5.23s/it]
+  📹 Processed video: 1847 -> 32 frames, resized to (256, 256)
+  ✅ Created trajectory: basic_pick_place_0 (1/100)
+```
+
+## Performance Notes
+
+- **Processing Rate**: ~0.5-2 trajectories/second (depends on video length)
+- **Memory Usage**: Low (~1-2GB peak for processing)
+- **Storage**: ~200KB-1MB per trajectory (video + metadata)
+- **Disk I/O**: Sequential access pattern for optimal performance
+
+## Troubleshooting
+
+### Missing Files
+```
+Warning: Missing MP4 file for /data/egodex/test/task1/5.hdf5
+```
+**Solution**: Some HDF5 files may not have corresponding MP4 files. This is normal and handled gracefully.
+
+### HDF5 Reading Errors
+```
+Error loading trajectory /data/egodex/test/task2/3.hdf5: Unable to open file
+```
+**Solution**: Corrupted HDF5 files are skipped automatically. Check file integrity if many errors occur.
+
+### Memory Issues
+```
+MemoryError: Unable to allocate array
+```
+**Solution**: The iterator design should prevent this, but if it occurs:
+- Reduce `max_trajectories`
+- Process dataset parts separately
+- Check available system memory
+
+### Video Processing Errors
+```
+Error: Could not open video file: /data/egodex/test/task1/0.mp4
+```
+**Solution**: 
+- Verify MP4 files are not corrupted
+- Check OpenCV installation: `pip install opencv-python`
+- Ensure sufficient disk space for temporary files
+
+## Large Scale Processing
+
+For processing thousands of trajectories from different parts:
+
+### Recommended Batch Processing
+```bash
+# Process each part separately to manage resources and create different datasets
+for part in part1 part2 part3 part4 part5; do
+    uv run python -m dataset_upload.generate_hf_dataset \
+        --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+        --dataset.dataset_path="/data/egodex/${part}" \
+        --dataset.dataset_name="egodex_${part}" \
+        --output.output_dir="egodex_${part}_dataset" \
+        --output.max_trajectories=2000 \
+        --hub.push_to_hub=true
+done
+```
+
+### Processing with Same Hub Repo but Different Dataset Names
+```bash
+# All parts go to same hub repo but with different dataset names
+for part in part1 part2 part3 part4 part5; do
+    uv run python -m dataset_upload.generate_hf_dataset \
+        --config_path=dataset_upload/configs/data_gen_configs/egodex.yaml \
+        --dataset.dataset_path="/data/egodx/${part}" \
+        --dataset.dataset_name="egodx_${part}" \
+        --hub.hub_repo_id="your-username/egodx_rfm" \
+        --output.max_trajectories=5000 \
+        --hub.push_to_hub=true
+done
+```
+
+## Integration with Robometer Training
+
+The generated dataset is compatible with the standard Robometer training pipeline:
+
+```bash
+# Use the processed dataset for training
+uv run accelerate launch --config_file configs/fsdp.yaml train.py \
+    --config_path=configs/config.yaml \
+    --dataset.dataset_path=egodex_dataset/egodex
+```
+
+## Dataset Statistics
+
+- **Total Duration**: 800+ hours of egocentric video
+- **Resolution**: 1080p at 30 FPS
+- **Tasks**: ~200 diverse tabletop manipulation tasks
+- **Splits**: 99% train (part1-5), 1% test, additional samples (extra)
+- **Annotation**: 3D pose for 66+ joints including hands and fingers
+- **Collection Device**: Apple Vision Pro with ARKit pose tracking
+
+## Citation
+
+If you use the EgoDex dataset, please cite:
+```bibtex
+@article{egodex2024,
+  title={EgoDex: Egocentric Dexterity Dataset},
+  url={https://arxiv.org/abs/2505.11709},
+  year={2024}
+}
+```
+
+## License
+
+The EgoDex dataset is licensed under CC-BY-NC-ND terms.
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_guides/FailSafe.md b/robometer/dataset_upload/dataset_guides/FailSafe.md
new file mode 100644
index 0000000000000000000000000000000000000000..72727a70ff8b19cdba22edeab4362f27f36005c8
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/FailSafe.md
@@ -0,0 +1,89 @@
+# FailSafe Dataset Guide
+
+This guide explains how to integrate and use the FailSafe dataset (correctdata_v2) with the Robometer pipeline.
+
+Source: `https://huggingface.co/datasets/onepiece1999/correctdata_v2/tree/main`
+
+## Overview
+
+- The dataset contains three tasks with failures and successes:
+  - `FailPickCube-v1`: "Pick up the red cube"
+  - `FailPushCube-v1`: "push and move a cube to a goal region in front of it"
+  - `FailStackCube-v1`: "Pick up a red cube and stack it on top of a green cube and let go of the cube without it falling."
+- Each task has ~100 seed folders. Each seed contains a `Ground_Truth/` (success) and multiple failure attempt folders.
+- Each attempt has camera views `front/`, `side/`, `wrist/` with ordered `.png` frames.
+- Additionally, JSON files (e.g., `vla_data_FailPickCube-v1.json`, `vla_data_GT_PickCube-v1.json`) provide sub-task annotated mini-trajectories with success/failure labels.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  FailPickCube-v1/
+    0/
+      Ground_Truth/
+        front/0.png ...
+        side/*.png
+        wrist/*.png
+      1_trans_x_2/
+        front/*.png
+        side/*.png
+        wrist/*.png
+    1/
+      ...
+  FailPushCube-v1/
+    ...
+  FailStackCube-v1/
+    ...
+  json_files/
+    vla_data_FailPickCube-v1.json
+    vla_data_GT_PickCube-v1.json
+    ...
+```
+
+## Configuration (configs/data_gen_configs/failsafe.yaml)
+
+```yaml
+# configs/data_gen_configs/failsafe.yaml
+
+dataset:
+  dataset_path: ./datasets/failsafe
+  dataset_name: failsafe
+
+output:
+  output_dir: ./robometer_dataset/failsafe_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: failsafe_rfm
+```
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/failsafe_loader.py`
+- Function: `load_failsafe_dataset(dataset_path)`
+- The loader:
+  - Builds full episodes from `Ground_Truth/` (success) and failure attempt folders for the chosen camera `view`.
+  - Parses `vla_data_*.json` entries to create sub-task mini-trajectories using the specified image paths and labels; includes the `sub_task` text in the trajectory `task` field.
+  - Uses the per-task instructions above when JSONs do not specify an instruction.
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/failsafe.yaml
+```
+
+This will:
+- Load full episodes and sub-task mini-trajectories for the selected `view`
+- Generate web-optimized videos
+- Create a HuggingFace dataset ready to push or save
+
+## Notes
+
+- If JSON files are at the dataset root instead of `json_files/`, the loader will look there as a fallback.
+- You can switch `view` in the YAML to `side` or `wrist` to regenerate the dataset for a different camera.
diff --git a/robometer/dataset_upload/dataset_guides/FinoNet.md b/robometer/dataset_upload/dataset_guides/FinoNet.md
new file mode 100644
index 0000000000000000000000000000000000000000..78c7ed9b131bd78956b90f8631b46c0286ebfa41
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/FinoNet.md
@@ -0,0 +1,144 @@
+# FinoNet Dataset Guide
+
+This guide explains how to integrate and use the FinoNet dataset with the Robometer training pipeline.
+
+Source: `https://huggingface.co/datasets/jesbu1/fino-net-dataset`
+
+## Overview
+
+FinoNet contains episodes of manipulation tasks with success/failure labels. Each episode is a sequence of RGB images stored as PNG files in numbered directories, organized by task type.
+
+## Directory Structure
+
+The dataset structure (after unzipping `failure.zip`) should be:
+
+```
+<dataset_path>/
+  failure/
+    failnet_dataset/
+      rgb_imgs/
+        put_on/
+          9/
+            frame0000000.png
+            frame0000024.png
+            ...
+        put_in/
+          ...
+        place/
+          ...
+        pour/
+          ...
+        push/
+          ...
+  put_on_annotation.txt
+  put_in_annotation.txt
+  place_annotation.txt
+  pour_annotation.txt
+  push_annotation.txt
+```
+
+- Task folders contain episode subdirectories (numbered by episode).
+- Each episode directory contains PNG frames (e.g., `frame0000000.png`, `frame0000024.png`).
+- Annotation files map episode numbers to labels (0 = success, 1 = failure).
+
+## Annotation Format
+
+Each annotation file (e.g., `put_on_annotation.txt`) contains comma-separated values:
+```
+name,label
+9,0
+10,1
+...
+```
+
+- `name`: Episode number
+- `label`: 0 for success, 1 for failure
+
+## Task Instructions
+
+The dataset includes 5 tasks with specific instructions:
+
+- **put_on**: "put the single block on the table onto the stack"
+- **put_in**: "put the thing on the table into the container"
+- **place**: "place the block in your hand onto the stack"
+- **pour**: "pour the contents of the cup into the receptacle on the table without spilling"
+- **push**: "push the object to the right without knocking it over"
+
+## Configuration (configs/data_gen_configs/fino_net.yaml)
+
+```yaml
+# configs/data_gen_configs/fino_net.yaml
+
+dataset:
+  dataset_path: /path/to/fino_net_dataset  # Root directory containing failure.zip or unzipped failure/
+  dataset_name: fino_net
+
+output:
+  output_dir: datasets/fino_net_rfm
+  max_trajectories: -1   # -1 for all episodes
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: fino_net_rfm
+```
+
+## Usage
+
+### Step 1: Download Dataset
+
+From HuggingFace:
+```bash
+# Download from: https://huggingface.co/datasets/jesbu1/fino-net-dataset/tree/main
+# Or use huggingface_hub:
+python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='jesbu1/fino-net-dataset', repo_type='dataset', local_dir='./fino_net_dataset')"
+```
+
+### Step 2: Unzip Dataset
+
+```bash
+cd ./fino_net_dataset
+unzip failure.zip
+```
+
+### Step 3: Convert to HF Format
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/fino_net.yaml
+```
+
+This will:
+- Load annotation files to map episodes to success/failure labels
+- Discover all episodes across the 5 tasks
+- Load PNG frames for each episode
+- Convert frame sequences to MP4 videos
+- Create HuggingFace dataset with proper metadata
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Task instruction from the predefined mapping
+- `frames`: Relative path to MP4 video file
+- `is_robot`: `False` (human demonstration data)
+- `quality_label`: "successful" or "failed" based on annotation label
+- `partial_success`: 1.0 for success, 0.0 for failure
+- `data_source`: `fino_net`
+
+## Troubleshooting
+
+- **No images found**: Verify that `failure.zip` has been unzipped and the directory structure matches the expected layout.
+- **Missing annotations**: Ensure all 5 annotation files exist in the root directory.
+- **Episode not in annotations**: Some episodes in the directories may not have corresponding annotation entries; these will be skipped with a warning.
+- **Performance**: Large datasets with many frames per episode can be memory intensive. Consider limiting `max_trajectories` during testing.
+
+## Notes
+
+- The loader automatically detects task type from directory names and maps to appropriate instructions.
+- Frames are loaded in sorted order based on frame number in the filename.
+- The dataset provides valuable failure examples for robust reward modeling.
+
diff --git a/robometer/dataset_upload/dataset_guides/Galaxea.md b/robometer/dataset_upload/dataset_guides/Galaxea.md
new file mode 100644
index 0000000000000000000000000000000000000000..bc8378b1efb44083ef154e4a623b812d07324778
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/Galaxea.md
@@ -0,0 +1,96 @@
+# Galaxea Open-World Dataset Guide
+
+This guide explains how to integrate and use the Galaxea Open-World RLDS dataset with the Robometer training pipeline.
+
+Source: `https://huggingface.co/datasets/OpenGalaxea/Galaxea-Open-World-Dataset`
+
+Can download it with `hf download OpenGalaxea/Galaxea-Open-World-Dataset --repo-type dataset --include "*rlds*" --local-dir ./datasets/galaxea`
+
+Also, need to install some extra dependencies:
+```bash
+uv pip install tensorflow-datasets
+uv pip install tensorflow
+uv pip install tf-keras
+```
+
+## Overview
+
+- 500+ hours of real-world mobile manipulation data in RLDS and LeRobot formats.
+- Fine-grained subtask language annotations at step level via `language_instruction`.
+- Multiple RLDS builders (e.g., `part1_r1_lite`, `sample_r1_lite`) under a common `rlds/` root.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  rlds/
+    sample_r1_lite/
+      1.0.0/
+        dataset_info.json
+        features.json
+        merge_dataset_large_r1_lite-train.tfrecord-00000-of-01024
+        ...
+    part1_r1_lite/
+      1.0.0/
+      ...
+```
+
+## Language Instruction Schema
+
+As documented, `language_instruction` encodes three parts separated by `@`:
+- `high_level` @ `low_level_chinese` @ `low_level_english`
+
+We extract the low-level English (the third part) and use it as the task string for embeddings.
+
+## Configuration (configs/data_gen_configs/galaxea.yaml)
+
+```yaml
+# configs/data_gen_configs/galaxea.yaml
+
+dataset:
+  dataset_path: ./datasets/galaxea
+  dataset_name: galaxea_part1_r1_lite # choose from part1_r1_lite, part2_r1_lite, part3_r1_lite, part4_r1_lite, part5_r1_lite, ...
+
+output:
+  output_dir: ./robometer_dataset/galaxea_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: galaxea_rfm
+```
+
+## Usage
+
+```bash
+bash dataset_upload/data_scripts/galaxea/gen_all_galaxea.sh
+```
+
+This will:
+- Iterate the listed RLDS builders under `rlds/`
+- For each episode, parse `language_instruction` and extract the low-level English instruction
+- Select camera views (`image_camera_head`, `image_camera_wrist_left`, `image_camera_wrist_right`)
+- Convert frames to web-optimized videos and create a HuggingFace dataset
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Low-level English instruction (parsed from `language_instruction`)
+- `frames`: Relative path to the generated clip video
+- `is_robot`: True
+- `quality_label`: "successful"
+- `partial_success`: N/A (fixed by pipeline)
+- `data_source`: `galaxea`
+
+## Troubleshooting
+
+- Builder not found: Ensure the RLDS version directories exist under `rlds/<name>/`.
+- Missing instruction: If no `language_instruction` is present or malformed, the episode is skipped.
+- Performance: Adjust `num_workers` and batch size inside the loader if needed.
diff --git a/robometer/dataset_upload/dataset_guides/H2R.md b/robometer/dataset_upload/dataset_guides/H2R.md
new file mode 100644
index 0000000000000000000000000000000000000000..79ca1868b1e0786fec546b7656c070e87e4dfdcf
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/H2R.md
@@ -0,0 +1,89 @@
+# H2R Dataset Guide
+
+This guide explains how to integrate and use the H2R (Human2Robot) dataset with the Robometer training pipeline.
+
+Source: `https://huggingface.co/datasets/dannyXSC/HumanAndRobot`
+Paper: `https://arxiv.org/abs/2502.16587`
+
+## Overview
+
+H2R contains paired human and robot videos stored as HDF5 files. Each trajectory provides synchronized human and robot camera streams. The loader reads both streams and standardizes them to RGB `uint8` frame tensors.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  <task_folder_1>/
+    <trajectory_1>.hdf5
+    <trajectory_2>.hdf5
+  <task_folder_2>/
+    <trajectory_1>.hdf5
+  ...
+```
+
+- The folder name represents the task category. A simple mapping converts folder names to human-readable task strings (see loader).
+
+## HDF5 Format
+
+- The loader expects camera streams under the keys:
+  - `/cam_data/human_camera`
+  - `/cam_data/robot_camera`
+- Each dataset is loaded into memory and converted to RGB if needed.
+
+## Configuration (configs/data_gen_configs/h2r.yaml)
+
+```yaml
+# configs/data_gen_configs/h2r.yaml
+
+dataset:
+  dataset_path: /path/to/h2r_dataset
+  dataset_name: h2r
+
+output:
+  output_dir: datasets/h2r_rfm
+  max_trajectories: 64   # null for all
+  max_frames: 64
+  use_video: true
+  fps: 30
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: h2r_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/h2r.yaml
+```
+
+This will:
+- Discover `.hdf5` trajectories grouped by task folders
+- Load paired human and robot camera frames
+- Convert frames to RGB `uint8`
+- Produce a HuggingFace dataset
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Human-readable task derived from folder name
+- `frames`: A tuple `(human_frames, robot_frames)` when read via the loader
+- `is_robot`: `False` for human, `True` for robot
+- `quality_label`: "successful"
+- `partial_success`: 1
+- `data_source`: `h2r`
+
+Note: The converter creates two entries per `.hdf5` file (one for human, one for robot) for training convenience.
+
+## Task Name Mapping
+
+The loader includes a simple mapping from folder names to readable descriptions and falls back to a prettified folder name if no mapping exists. You can extend `FOLDER_TO_TASK_NAME` in `dataset_upload/dataset_loaders/h2r_loader.py`.
+
+## Troubleshooting
+
+- KeyError: Verify that the HDF5 files contain `/cam_data/human_camera` and `/cam_data/robot_camera` datasets.
+- Shape errors: Frames must be 4D tensors `(T, H, W, 3)`.
+- Performance: Large `.hdf5` files will load into memory; consider limiting `max_trajectories` during testing.
diff --git a/robometer/dataset_upload/dataset_guides/HAND_paired.md b/robometer/dataset_upload/dataset_guides/HAND_paired.md
new file mode 100644
index 0000000000000000000000000000000000000000..fe297e3b75b8689f9cd70cd71bb164a2b8c04731
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/HAND_paired.md
@@ -0,0 +1,99 @@
+# HAND_paired_data Dataset Guide
+
+This guide explains how to load and convert a subset of the HAND dataset with the Robometer pipeline from [HAND](https://liralab.usc.edu/handretrieval/).
+
+## Overview
+
+The HAND dataset contains paired robot and human demonstrations for the same tasks. Each task has both robot executions and human hand demonstrations.
+
+### Directory Structure
+
+```
+dataset_path/
+  blend_carrot/                    # Robot demonstrations
+    traj0/
+      external_imgs/
+        im_0.jpg, im_1.jpg, ...
+      over_shoulder_imgs/
+        im_0.jpg, im_1.jpg, ...
+    traj1/
+      ...
+  blend_carrot_hand/               # Human demonstrations
+    traj0/
+      external_imgs/
+        im_0.jpg, im_1.jpg, ...
+      over_shoulder_imgs/
+        im_0.jpg, im_1.jpg, ...
+    traj1/
+      ...
+  close_microwave/
+    ...
+  close_microwave_hand/
+    ...
+```
+
+### Key Features
+
+- **Paired Data**: Robot tasks (e.g., `blend_carrot`) and human tasks (e.g., `blend_carrot_hand`) share the same task instruction
+- **Task Names**: Folder names are converted to instructions (e.g., `blend_carrot` → "blend carrot")
+- **Camera Views**: Each trajectory can have up to 2 camera views:
+  - `external_imgs/`: External camera view
+  - `over_shoulder_imgs/`: Over-shoulder camera view
+- **Multiple Trajectories**: Each task contains multiple demonstrations (`traj0`, `traj1`, etc.)
+
+## Configuration
+
+```yaml
+# configs/data_gen_configs/hand_paired.yaml
+
+dataset:
+  dataset_path: ./datasets/HAND_paired_data
+  dataset_name: hand_paired
+
+output:
+  output_dir: ./robometer_dataset/hand_paired_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: hand_paired_rfm
+```
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/hand_paired_loader.py`
+- Function: `load_hand_paired_dataset(dataset_path)`
+- Notes:
+  - Automatically identifies robot vs. human demonstrations based on `_hand` suffix
+  - Processes both camera views independently
+  - Sorts images by numeric suffix (`im_0.jpg`, `im_1.jpg`, etc.)
+  - All demonstrations are labeled as "successful"
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/hand_paired.yaml --dataset.dataset_name hand_paired_robot
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/hand_paired.yaml --dataset.dataset_name hand_paired_human
+```
+
+This will:
+- Load both robot and human demonstrations
+- Process all camera views
+- Generate web-optimized videos
+- Create a HuggingFace dataset with paired robot/human data
+
+## Data Fields
+
+Each trajectory contains:
+- `task`: Human-readable task instruction (e.g., "blend carrot")
+- `is_robot`: Boolean indicating if this is a robot (True) or human (False) demonstration
+- `frames`: Video or image sequence
+- `quality_label`: "successful" (all demonstrations assumed successful)
+- `data_source`: "hand_paired"
+
diff --git a/robometer/dataset_upload/dataset_guides/HumanoidEveryday.md b/robometer/dataset_upload/dataset_guides/HumanoidEveryday.md
new file mode 100644
index 0000000000000000000000000000000000000000..c0970c6578ccdb4bb383388054fde2f4e0ef7858
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/HumanoidEveryday.md
@@ -0,0 +1,193 @@
+# Humanoid Everyday Dataset Guide
+
+This guide explains how to integrate and use the Humanoid Everyday dataset with the Robometer training pipeline.
+
+Source: `https://github.com/ausbxuse/Humanoid-Everyday`
+
+## Overview
+
+The Humanoid Everyday Dataset is a diverse collection of humanoid robot (Unitree G1 and H1) demonstrations recorded at 30 Hz across everyday tasks. This dataset supports research in robot learning, imitation, and perception.
+
+- **Total Download Size:** ~500 GB (across 250 tasks), over 100,000 time-step recorded
+- **Tasks:** 260 diverse scenarios (loco-manipulation, basic manipulation, tool use, deformables, articulated objects, human–robot interaction)
+- **Episodes per task:** 40
+- **Recording Frequency:** 30 Hz
+
+### Modalities captured
+
+- **Low-dimensional:**  
+  - Joint states (arm, leg, hand)  
+  - IMU (orientation, accelerometer, gyroscope, RPY)  
+  - Odometry/Kinematics (position, velocity, orientation)  
+  - Hand pressure sensors (G1 only)  
+  - Teleoperator hands/head actions from Apple Vision Pro  
+  - Inverse kinematics data
+- **High-dimensional:**  
+  - Egocentric RGB images (480x640x3, PNG)  
+  - Depth maps (480x640, uint16)  
+  - LiDAR point clouds (~6k points per step, PCD)
+
+## Prerequisites
+
+### Install humanoid_everyday dataloader
+
+```bash
+git clone https://github.com/ausbxuse/Humanoid-Everyday
+cd Humanoid-Everyday
+uv pip install -e .
+```
+
+### Download dataset
+
+Please visit the task spreadsheet to download your task of interest, or use the provided download script:
+
+```bash
+bash dataset_upload/data_scripts/humanoid_everyday/download_humanoid_everyday.sh
+```
+
+## Directory Structure
+
+```
+<dataset_path>/
+  task1.zip
+  task2.zip
+  ...
+  taskN.zip
+```
+
+Each zip file contains a complete task dataset with multiple episodes.
+
+## Data Schema
+
+Each time step is represented by a Python dictionary with the following fields:
+
+```python
+{
+    # Scalar identifiers
+    "time": np.float64,                # UNIX timestamp (s)
+    "robot_type": np.str_,             # Robot model identifier (G1 only)
+
+    # Robot states
+    "states": {
+        "arm_state": np.ndarray((14,), dtype=np.float64),   # 14 joint angles
+        "leg_state": np.ndarray((15 or 13,), dtype=np.float64),  # 15 joint angles for G1, 13 for H1_2
+        "hand_state": np.ndarray((14 or 12,), dtype=np.float64), # 14 joint angles for Unitree Dex3 Hand, 12 for Inspire Dextrous Hand
+        "hand_pressure_state": [...],  # List of per-sensor readings (9 sensors per hand)
+        "imu": {
+            "quaternion": np.ndarray((4,), dtype=np.float64),    # [w, x, y, z]
+            "accelerometer": np.ndarray((3,), dtype=np.float64), # [ax, ay, az]
+            "gyroscope": np.ndarray((3,), dtype=np.float64),     # [gx, gy, gz]
+            "rpy": np.ndarray((3,), dtype=np.float64)            # [roll, pitch, yaw]
+        },
+        "odometry": {
+            "position": np.ndarray((3,), dtype=np.float64),  # [x, y, z]
+            "velocity": np.ndarray((3,), dtype=np.float64),  # [vx, vy, vz]
+            "rpy": np.ndarray((3,), dtype=np.float64),      # [roll, pitch, yaw]
+            "quat": np.ndarray((4,), dtype=np.float64)      # [w, x, y, z]
+        }
+    },
+
+    # Control commands and solutions
+    "actions": {
+        "right_angles": np.ndarray((7,), dtype=np.float64),  # commanded joint angles
+        "left_angles": np.ndarray((7,), dtype=np.float64),   # commanded joint angles
+        "armtime": np.float64,                               # timestamp
+        "iktime": np.float64,                                # timestamp
+        "sol_q": np.ndarray((14,), dtype=np.float64),       # solution joint angles
+        "tau_ff": np.ndarray((14,), dtype=np.float64),      # feedforward torques
+        "head_rmat": np.ndarray((3, 3), dtype=np.float64),  # rotation matrix
+        "left_pose": np.ndarray((4, 4), dtype=np.float64),  # homogeneous transform
+        "right_pose": np.ndarray((4, 4), dtype=np.float64)  # homogeneous transform
+    },
+
+    # High-dimensional observations
+    "image": np.ndarray((480, 640, 3), dtype=np.uint8),     # RGB image
+    "depth": np.ndarray((480, 640), dtype=np.uint16),       # Depth map
+    "lidar": np.ndarray((~6000, 3), dtype=np.float64)       # around 6000 points for lidar point cloud
+}
+```
+
+## Configuration (configs/data_gen_configs/humanoid_everyday.yaml)
+
+```yaml
+# configs/data_gen_configs/humanoid_everyday.yaml
+
+dataset:
+  dataset_path: "./datasets/humanoid_everyday"  # Path containing zip files
+  dataset_name: humanoid_everyday_rfm
+
+output:
+  output_dir: ./robometer_dataset/humanoid_everyday_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: humanoid_everyday_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config dataset_upload/configs/data_gen_configs/humanoid_everyday.yaml
+```
+
+This will:
+- Find all zip files in the specified dataset path
+- For each zip file, extract the task name and load episodes using the humanoid_everyday dataloader
+- Extract RGB images from each episode
+- Convert frames to web-optimized videos and create a HuggingFace dataset
+- Use the zip filename (without extension) as the task description
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Task name extracted from zip filename
+- `frames`: Relative path to the generated clip video
+- `is_robot`: True
+- `quality_label`: "successful"
+- `partial_success`: N/A (fixed by pipeline)
+- `data_source`: `humanoid_everyday`
+
+## Example Usage with Dataloader
+
+```python
+from humanoid_everyday import Dataloader
+
+# Load your downloaded task's dataset zip file (e.g., the "push_a_button" task)
+ds = Dataloader("~/Downloads/push_a_button.zip")
+print("Episode length of dataset:", len(ds))
+
+# Displaying high dimensional data at first episode, second timestep.
+ds.display_image(0, 1)
+ds.display_depth_point_cloud(0, 1)
+ds.display_lidar_point_cloud(0, 1)
+
+for i, episode in enumerate(ds):
+    if i == 1:  # episode 1
+        print("RGB image shape:", episode[0]["image"].shape)  # (480, 640, 3)
+        print("Depth map shape:", episode[0]["depth"].shape)  # (480, 640)
+        print("LiDAR points shape:", episode[0]["lidar"].shape)  # (~6000, 3)
+
+        batch = episode[0:4]  # batch loading episodes
+        print(batch[1]["image"].shape)
+        print(batch[0]["image"].shape)
+```
+
+## Troubleshooting
+
+- **Missing humanoid_everyday package**: Install it with `pip install humanoid_everyday` or clone and install from the GitHub repository
+- **No zip files found**: Ensure the dataset_path contains zip files with humanoid everyday datasets
+- **Import errors**: Make sure the humanoid_everyday package is properly installed and accessible
+- **Memory issues**: Adjust `max_frames` and `num_workers` parameters to reduce memory usage
+- **Long episodes**: Episodes longer than 1000 frames are automatically skipped to prevent memory issues
+
+## License
+
+This dataset is released under the MIT License.
diff --git a/robometer/dataset_upload/dataset_guides/LIBERO.md b/robometer/dataset_upload/dataset_guides/LIBERO.md
new file mode 100755
index 0000000000000000000000000000000000000000..ac0a5813f8627a0fd87da09d5230a9ca2b2a268d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/LIBERO.md
@@ -0,0 +1,176 @@
+# LIBERO Dataset Guide
+
+LIBERO is a benchmark for lifelong robot learning with built-in support in the Robometer training pipeline.
+
+## Overview
+
+- **📁 Local File Support**: Processes HDF5 files from local storage
+- **🎮 Simulation Data**: High-quality manipulation tasks
+- **🏠 Multiple Environments**: Living room, kitchen, office, and study scenarios
+- **📊 Structured Tasks**: Clear task descriptions and optimal trajectories
+
+## Prerequisites
+
+### Download LIBERO Dataset
+```bash
+# Clone or download LIBERO dataset
+git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
+# Follow LIBERO installation instructions for dataset download
+
+# This should work too
+git submodule update --init --recursive
+
+cd deps/libero/LIBERO
+uv run python benchmark_scripts/download_libero_datasets.py --datasets DATASET
+```
+where DATASET is chosen from `[libero_spatial, libero_object, libero_100, libero_goal]`.
+
+
+## Quick Start
+
+### 0. Set Hugging Face repo ID
+Before we start, you must have an HF account which will be pushed to.
+You will set this by setting
+```
+export HF_USERNAME=<insert HF username here>
+```
+
+Then, for each dataset, run with the all the datasets you would like to process
+
+### Option 1: Use Default Configuration
+```bash
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml\
+    --dataset.dataset_path=deps/libero/LIBERO/libero/datasets/libero_90 \
+    --dataset.dataset_name=libero_90
+```
+
+If all your LIBERO data exists in the path above, you can use the following utility script
+```bash
+uv run bash dataset_upload/data_scripts/libero/gen_all_libero.sh
+```
+
+### Option 2: Custom Configuration
+```bash
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/libero.yaml \
+    --dataset.dataset_path=/path/to/your/libero/dataset \
+    --dataset.dataset_name=libero_custom \
+    --output.output_dir=libero_robometer_dataset \
+    --output.max_trajectories=1000 \
+    --output.use_video=true \
+    --output.fps=10
+```
+
+## Configuration Options
+
+Create a custom config file `configs/data_gen_configs/libero.yaml`:
+
+```yaml
+dataset:
+  dataset_path: LIBERO/libero/datasets/libero_90
+  dataset_name: libero_90
+
+output:
+  output_dir: libero_dataset
+  max_trajectories: -1  # Process all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+
+hub:
+  push_to_hub: false
+  hub_repo_id: your-username/libero_rfm
+```
+
+## Data Structure Processed
+
+```
+LIBERO Dataset:
+├── *.hdf5 files             ← PROCESSED
+│   ├── /data/
+│   │   └── trajectory_*/
+│   │       ├── obs/
+│   │       │   └── agentview_rgb    ← EXTRACTED as frames
+│   │       └── actions              ← EXTRACTED as actions
+└── Generated Output:
+    ├── frames: List[np.ndarray]     ← RGB video frames
+    ├── actions: np.ndarray          ← Robot actions
+    ├── task: str                    ← Parsed from filename
+    └── optimal: "optimal"           ← All LIBERO data assumed optimal
+```
+
+## Supported LIBERO Variants
+
+- **LIBERO-90**: 90 tasks across 4 environments
+- **LIBERO-10**: 10 benchmark tasks
+- **Custom datasets**: Any LIBERO-format HDF5 files
+
+## Sample Output
+
+```
+Loading LIBERO dataset from: LIBERO/libero/datasets/libero_90
+Found 90 HDF5 files
+Processing LIBERO dataset, 90 files: 100%|██████████| 90/90
+
+Sample trajectory:
+- Task: "stack the right bowl on the left bowl and place them in the tray"
+- Frames: (128, 128, 128, 3) RGB images
+- Actions: (128, 7) joint positions
+- Environment: LIVING_ROOM_SCENE4
+```
+
+## File Name Parsing
+
+LIBERO dataset automatically parses task information from HDF5 filenames:
+
+```
+LIVING_ROOM_SCENE4_stack_the_right_bowl_on_the_left_bowl_and_place_them_in_the_tray.hdf5
+│              │    │
+│              │    └── Task description
+│              └── Scene identifier  
+└── Environment type
+```
+
+## Performance Notes
+
+- **Processing Speed**: ~2-5 files/second
+- **Memory Usage**: Moderate (loads one HDF5 file at a time)
+- **Storage**: Variable (depends on trajectory length)
+- **Video Encoding**: Converts RGB arrays to MP4 format
+
+## Troubleshooting
+
+### HDF5 File Issues
+```python
+# Check HDF5 file structure
+import h5py
+with h5py.File('path/to/file.hdf5', 'r') as f:
+    print(list(f.keys()))  # Should show 'data'
+    print(list(f['data'].keys()))  # Should show trajectory keys
+```
+
+### Missing Observations
+Ensure your LIBERO dataset has the expected structure:
+```
+/data/demo_0/obs/agentview_rgb  # RGB frames
+/data/demo_0/actions            # Action sequences
+```
+
+### Memory Issues
+For large LIBERO datasets:
+```bash
+# Process in chunks
+uv run python data/generate_hf_dataset.py \
+    --config_path=configs/data_gen_configs/libero.yaml \
+    --output.max_trajectories=100  # Limit trajectories
+```
+
+## Integration with Robometer Training
+
+```bash
+# Train on processed LIBERO dataset
+uv run accelerate launch --config_file configs/fsdp.yaml train.py \
+    --config_path=configs/config.yaml \
+    --dataset.dataset_path=libero_dataset/libero_90
+```
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_guides/MIT_ROBOT_PRANK.md b/robometer/dataset_upload/dataset_guides/MIT_ROBOT_PRANK.md
new file mode 100644
index 0000000000000000000000000000000000000000..1117f7d98bc26e809e9537350a64ce0f75fa649c
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/MIT_ROBOT_PRANK.md
@@ -0,0 +1,149 @@
+# MIT-Franka-P-Rank Dataset Guide
+
+This guide explains how to integrate and use the MIT-Franka-P-Rank dataset with the Robometer pipeline.
+
+## Overview
+
+- MIT-Franka-Prank is a robotics dataset with quality-labeled trajectories for manipulation tasks
+- The dataset contains pre-recorded MP4 videos with metadata describing task instructions and quality labels
+- Quality labels include: **successful**, **suboptimal**, and **failure**
+
+## Dataset Structure
+
+The dataset is organized as follows:
+
+```
+<dataset_path>/
+  20251210/
+    episode_000_foldtowel_suboptimal.mp4
+    episode_000_movebanana_success.mp4
+    episode_000_movepebble_suboptimal.mp4
+    ...
+    foldtowel_metadata.json
+    pickandplace_metadata.json
+```
+
+### Tasks Included
+
+1. **foldtowel**: Fold the towel in half
+2. **movebanana**: Pick up the banana and place it on the blue plate
+3. **movepebble**: Move some pebbles from the blue bowl to the green plate using the scoop
+
+### Metadata Format
+
+Each metadata JSON file contains an array of episodes:
+
+```json
+[
+  {
+    "episode_idx": 0,
+    "task_name": "foldtowel",
+    "filename": "episode_000_foldtowel_suboptimal.mp4",
+    "run_dir": "20251210_192953",
+    "instruction": "fold the towel in half",
+    "success": "suboptimal",
+    "trajectory_length": 1351
+  },
+  ...
+]
+```
+
+## Configuration
+
+Configuration file: `dataset_upload/configs/data_gen_configs/mit_robot_prank.yaml`
+
+```yaml
+dataset:
+  dataset_path: ~/projects/robometer/datasets/20251210-mit-robot-prank
+  dataset_name: mit_franka_p-rank_rfm
+
+output:
+  output_dir: ./robometer_dataset/mit_franka_p-rank_rfm
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: mit_franka_p-rank_rfm
+```
+
+## Usage
+
+### Convert Dataset to HuggingFace Format
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/mit_franka_prank.yaml
+```
+
+This will:
+- Read all metadata JSON files from the dataset directory
+- Load the corresponding MP4 videos
+- Process and resample videos to the specified frame count and FPS
+- Generate language embeddings for task instructions
+- Create a HuggingFace dataset with proper quality labels
+- Optionally push to HuggingFace Hub
+
+### Quality Label Mapping
+
+The loader automatically normalizes quality labels:
+- `"success"` → `"successful"`
+- `"fail"` → `"failure"`
+- `"suboptimal"` → `"suboptimal"` (unchanged)
+
+## Output Format
+
+The generated dataset will have the following schema:
+
+- `id`: Unique trajectory identifier
+- `task`: Task instruction text
+- `lang_vector`: Language embedding of the task
+- `data_source`: Dataset name
+- `frames`: Path to video file (or sequence of images)
+- `is_robot`: Boolean (True for this dataset)
+- `quality_label`: One of "successful", "suboptimal", "failure"
+- `preference_group_id`: None for this dataset
+- `preference_rank`: None for this dataset
+
+## Notes
+
+- Videos are already in MP4 format, so the loader reads them directly using OpenCV
+- The loader supports parallel processing with configurable worker count
+- Language embeddings are cached to avoid redundant computations
+- Output videos maintain the same content but are resampled to the specified frame count and FPS
+
+## Troubleshooting
+
+### Video File Not Found
+- Ensure the dataset path points to the parent directory containing the date-stamped subdirectory
+- Check that video files exist and match the filenames in metadata JSON files
+
+### Missing Metadata Files
+- The loader looks for files ending with `_metadata.json` in the video directory
+- Ensure at least one metadata file exists (e.g., `foldtowel_metadata.json`, `pickandplace_metadata.json`)
+
+### OpenCV Issues
+- If you encounter video codec issues, ensure OpenCV is properly installed with video support
+- Try: `pip install opencv-python-headless` or `pip install opencv-python`
+
+## Example Output
+
+After processing, you'll have a directory structure like:
+
+```
+robometer_dataset/mit_franka_p-rank_rfm/
+  mit_franka_p-rank_rfm/
+    shard_0000/
+      episode_000000/
+        foldtowel.mp4
+      episode_000001/
+        foldtowel.mp4
+      ...
+```
+
+And a HuggingFace Dataset with all trajectory metadata.
+
diff --git a/robometer/dataset_upload/dataset_guides/MoTiF.md b/robometer/dataset_upload/dataset_guides/MoTiF.md
new file mode 100644
index 0000000000000000000000000000000000000000..a579aae787a4c955022cc60548ea04ca054325b8
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/MoTiF.md
@@ -0,0 +1,80 @@
+# MoTiF Dataset Guide
+
+This guide explains how to integrate and use the MoTiF-1K dataset with the Robometer data pipeline using a FrameLoader (no HuggingFace conversion required).
+
+Source: `https://github.com/Minyoung1005/motif#data-structure`
+
+## Overview
+
+- 1K trajectories across 13 task categories; both human and robot (Stretch) motions
+- Visual motion representations provided; we support raw video or frame directories
+- We use a simple FrameLoader to load frames on-demand for each trajectory
+
+## Directory Structure
+
+As per the MoTiF README, after unzipping `MotIF.zip` under `./data`, `./data/MotIF` contains at least:
+
+```
+./data/MotIF/
+  annotations/
+  human_motion/
+  stretch_motion/
+```
+
+Our loader first looks for annotations under `annotations/` to pair sources with language text; if absent, it will scan `human_motion/` and `stretch_motion/` for videos or frame directories.
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/motif_loader.py`
+- Exposes `load_motif_dataset(dataset_path: str) -> dict[str, list[dict]]`
+- Each trajectory dictionary contains:
+  - `id`: unique id
+  - `task`: from annotations if available, otherwise "MoTiF"
+  - `frames`: `MotifFrameLoader` that lazily reads frames (video file or directory of images)
+  - `is_robot`: inferred from path (`stretch`/`robot` -> True, `human` -> False)
+  - `quality_label`: "successful"
+  - `partial_success`: 1
+  - `data_source`: "motif"
+
+## Configuration (configs/data_gen_configs/motif.yaml)
+
+```yaml
+# configs/data_gen_configs/motif.yaml
+
+dataset:
+  dataset_path: ./datasets/MotIF
+  dataset_name: motif
+
+output:
+  output_dir: ./robometer_dataset/motif_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: false
+  hub_repo_id: motif_rfm
+```
+
+## Usage Example
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config dataset_upload/configs/data_gen_configs/motif.yaml
+```
+
+This will:
+- Find all zip files in the specified dataset path
+- For each zip file, extract the task name and load episodes using the humanoid_everyday dataloader
+- Extract RGB images from each episode
+- Convert frames to web-optimized videos and create a HuggingFace dataset
+- Use the zip filename (without extension) as the task description
+
+## Notes
+
+- Annotations: The loader tries to parse any JSON/JSONL files under `annotations/` to find `(source_path, text)` pairs. Supported keys include `video_path|path|image_dir|frames_dir` and `narration|instruction|task|description|caption`.
+- Frame directories: If a directory contains images (e.g., `.jpg`, `.png`), it is treated as a sequence of frames.
+- Video support: Common video formats are supported via OpenCV (e.g., `.mp4`, `.mov`).
+- If you need to use specific MoTiF visual motion representations (e.g., storyboard, optical flow), point `source_path` to those assets and the FrameLoader will load images in order.
diff --git a/robometer/dataset_upload/dataset_guides/MolmoAct.md b/robometer/dataset_upload/dataset_guides/MolmoAct.md
new file mode 100644
index 0000000000000000000000000000000000000000..a8964b441865b2b8560002ccba53301128de0b39
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/MolmoAct.md
@@ -0,0 +1,72 @@
+# MolmoAct Dataset Guide
+
+This guide explains how to integrate and use the MolmoAct LeRobot dataset with the Robometer training pipeline.
+
+Source: `https://huggingface.co/datasets/allenai/MolmoAct-Dataset`
+
+## Overview
+
+- LeRobot/Parquet dataset with per-frame views and state/action fields.
+- We stream rows and group by `episode_index` to form trajectories.
+- Views used: `first_view`, `second_view`, `wrist_image`.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  *.parquet
+  subfolder_a/*.parquet
+  subfolder_b/*.parquet
+```
+
+## Configuration (configs/data_gen_configs/molmoact.yaml)
+
+```yaml
+# configs/data_gen_configs/molmoact.yaml
+
+dataset:
+  dataset_path: ./datasets/molmoact
+  dataset_name: molmoact_dataset_tabletop
+
+output:
+  output_dir: ./robometer_dataset/molmoact_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 1
+
+hub:
+  push_to_hub: true
+  hub_repo_id: molmoact_rfm
+```
+
+## Usage
+
+```bash
+bash dataset_upload/data_scripts/molmoact/gen_all_molmoact.sh
+```
+
+This will:
+- Stream parquet rows
+- Group frames by `episode_index`
+- Write per-view videos and build a HuggingFace dataset
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Uses `task_index` if present, else default label
+- `frames`: Relative path to the generated clip video
+- `is_robot`: True
+- `quality_label`: "successful"
+- `partial_success`: N/A (fixed by pipeline)
+- `data_source`: `molmoact`
+
+## Notes
+
+- Images are decoded from datasets Image cells into RGB uint8 arrays.
+- If two consecutive episodes share identical views, they still become separate trajectories.
+- Adjust `max_frames`/`fps` for performance and disk usage.
diff --git a/robometer/dataset_upload/dataset_guides/NewMITFranka.md b/robometer/dataset_upload/dataset_guides/NewMITFranka.md
new file mode 100644
index 0000000000000000000000000000000000000000..bd73033516be52af405405e2bab6910eb57c4a34
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/NewMITFranka.md
@@ -0,0 +1,226 @@
+# Robometer MIT Franka Dataset Guide
+
+This guide explains how to integrate and use the Robometer MIT Franka dataset with the Robometer pipeline.
+Source: https://drive.google.com/drive/folders/1dd62YeKY9-bTeK5fkjljzoFkNX2uZ1YD?usp=sharing
+
+## Overview
+
+- Robometer MIT Franka is a robotics dataset with quality-labeled trajectories for manipulation tasks
+- The dataset contains pre-recorded MP4 videos organized by task and quality level
+- Quality labels include: **successful**, **suboptimal**, and **failure**
+- Each trajectory has two camera views: **external (ext)** and **wrist**
+
+## Dataset Structure
+
+The dataset is organized as follows:
+
+```
+ROBOMETER_MIT_Franka/
+  foldtowel/
+    success/
+      Tue_Jan_20_23_49_14_2026_ext.mp4
+      Tue_Jan_20_23_49_14_2026_wrist.mp4
+      ...
+    suboptimal/
+      Tue_Jan_20_23_52_01_2026_ext.mp4
+      Tue_Jan_20_23_52_01_2026_wrist.mp4
+      ...
+    failure/
+      Tue_Jan_20_23_54_53_2026_ext.mp4
+      Tue_Jan_20_23_54_53_2026_wrist.mp4
+      ...
+  movebanana/
+    success/
+    suboptimal/
+    failure/
+  movemouse/
+  pourpebble/
+  pulltissue/
+  putspoon/
+  stirpot/
+```
+
+### Tasks Included
+
+1. **foldtowel**: Fold the towel in half.
+2. **movebanana**: Pick up the banana from the blue plate and place it on the green plate.
+3. **movemouse**: Pick up the mouse and place it right next to the laptop, while avoiding spilling coffee.
+4. **pourpebble**: Pour the pebbles from the cup onto the plate.
+5. **pulltissue**: Pull a tissue from the tissue box.
+6. **putspoon**: Pick up the spoon and place it inside the cup.
+7. **stirpot**: Pick up the spatula and stir the beans in the pot.
+
+### Camera Views
+
+Each trajectory has two synchronized camera views:
+- **ext (external)**: Third-person view of the robot and workspace
+- **wrist**: First-person view from the robot's wrist camera
+
+By default, the loader processes both views as separate trajectories, each with the same task instruction but different visual observations. You can optionally exclude wrist camera views using the `exclude_wrist_cam` configuration option.
+
+### Quality Labels
+
+- **success**: Trajectory successfully completes the task
+- **suboptimal**: Trajectory completes the task but with inefficiencies or errors
+- **failure**: Trajectory fails to complete the task
+
+## Configuration
+
+Configuration file: `dataset_upload/configs/data_gen_configs/new_mit_franka.yaml`
+
+```yaml
+dataset:
+  dataset_path: ~/robometer/datasets/ROBOMETER_MIT_Franka
+  dataset_name: rfm_new_mit_franka_rfm
+  exclude_wrist_cam: false  # Set to true to only process external camera views
+
+output:
+  output_dir: ./robometer_dataset/rfm_new_mit_franka_rfm
+  max_trajectories: -1  # -1 for all trajectories
+  max_frames: 32
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: rfm_new_mit_franka_rfm
+```
+
+### Camera View Options
+
+The dataset includes both external and wrist camera views. You can configure which views to process:
+
+- **`exclude_wrist_cam: false`** (default): Process both external and wrist camera views as separate trajectories
+- **`exclude_wrist_cam: true`**: Only process external camera views, skip wrist camera entirely
+
+This is useful if you only want third-person views or need to reduce dataset size.
+
+## Usage
+
+### Convert Dataset to HuggingFace Format
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/new_mit_franka.yaml
+```
+
+To only process external camera views (exclude wrist camera):
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/new_mit_franka.yaml --dataset.exclude_wrist_cam=true
+```
+
+This will:
+- Scan all task folders and their quality subfolders
+- Load the corresponding MP4 videos for both camera views (or only external if `exclude_wrist_cam: true`)
+- Process and resample videos to the specified frame count and FPS
+- Generate language embeddings for task instructions
+- Create a HuggingFace dataset with proper quality labels
+- Optionally push to HuggingFace Hub
+
+### Quality Label Mapping
+
+The loader automatically normalizes quality labels:
+- `"success"` → `"successful"`
+- `"failure"` → `"failure"` (unchanged)
+- `"suboptimal"` → `"suboptimal"` (unchanged)
+
+## Output Format
+
+The generated dataset will have the following schema:
+
+- `id`: Unique trajectory identifier
+- `task`: Task instruction text
+- `lang_vector`: Language embedding of the task
+- `data_source`: Dataset name
+- `frames`: Path to video file (or sequence of images)
+- `is_robot`: Boolean (True for this dataset)
+- `quality_label`: One of "successful", "suboptimal", "failure"
+- `preference_group_id`: None for this dataset
+- `preference_rank`: None for this dataset
+
+## Task Instructions Mapping
+
+The loader uses the following mapping from folder names to natural language instructions:
+
+| Folder Name  | Task Instruction                                                                       |
+|--------------|----------------------------------------------------------------------------------------|
+| foldtowel    | Fold the towel in half                                                                 |
+| movebanana   | Pick up the banana from the blue plate and place it on the green plate                 |
+| movemouse    | Pick up the mouse and place it right next to the laptop, while avoiding spilling coffee |
+| pourpebble   | Pour the pebbles from the cup onto the plate                                           |
+| pulltissue   | Pull a tissue from the tissue box                                                      |
+| putspoon     | Pick up the spoon and place it inside the cup                                          |
+| stirpot      | Pick up the spatula and stir the beans in the pot                                      |
+
+These mappings are defined in the loader file (`new_mit_franka_loader.py`) and can be updated if needed.
+
+## Notes
+
+- Videos are already in MP4 format, so the loader reads them directly using OpenCV
+- The loader supports parallel processing with configurable worker count
+- Language embeddings are cached to avoid redundant computations
+- Output videos maintain the same content but are resampled to the specified frame count and FPS
+- Both camera views (ext and wrist) are processed as separate trajectories by default
+- Use `exclude_wrist_cam: true` to only process external camera views and reduce dataset size by ~50%
+
+## Troubleshooting
+
+### Video File Not Found
+- Ensure the dataset path points to the ROBOMETER_MIT_Franka directory
+- Check that video files exist in the expected task/quality subfolders
+
+### Missing Task Folders
+- The loader expects folders named: foldtowel, movebanana, movemouse, pourpebble, pulltissue, putspoon, stirpot
+- Each task folder should contain subfolders: success, suboptimal, failure
+
+### OpenCV Issues
+- If you encounter video codec issues, ensure OpenCV is properly installed with video support
+- Try: `pip install opencv-python-headless` or `pip install opencv-python`
+
+### Unknown Task Warning
+- If a task folder is not in the TASK_INSTRUCTIONS mapping, it will be skipped
+- Add new tasks to the mapping in `new_mit_franka_loader.py`
+
+### Filtering Camera Views
+- Set `exclude_wrist_cam: true` in the YAML config to skip wrist camera videos
+- Or pass it as a command-line argument: `dataset.exclude_wrist_cam=true`
+- This will process only external camera views, roughly halving the dataset size
+
+## Example Output
+
+After processing, you'll have a directory structure like:
+
+```
+robometer_dataset/rfm_new_mit_franka_rfm/
+  rfm_new_mit_franka_rfm/
+    shard_0000/
+      episode_000000/
+        foldtowel_ext.mp4
+      episode_000001/
+        foldtowel_wrist.mp4
+      ...
+```
+
+With `exclude_wrist_cam: true`, only external camera videos are included:
+
+```
+robometer_dataset/rfm_new_mit_franka_rfm/
+  rfm_new_mit_franka_rfm/
+    shard_0000/
+      episode_000000/
+        foldtowel_ext.mp4
+      episode_000001/
+        movebanana_ext.mp4
+      ...
+```
+
+And a HuggingFace Dataset with all trajectory metadata.
+
+## Dataset Statistics
+
+The dataset contains approximately 304 video files (152 trajectory pairs with ext and wrist views) across 7 different manipulation tasks, with varying distributions of successful, suboptimal, and failed demonstrations.
+
+When `exclude_wrist_cam: true` is set, approximately 152 external camera videos are processed (one per trajectory).
diff --git a/robometer/dataset_upload/dataset_guides/OXE.md b/robometer/dataset_upload/dataset_guides/OXE.md
new file mode 100755
index 0000000000000000000000000000000000000000..aa91fdf90be95cb6f8421eb97f615c292be24fc0
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/OXE.md
@@ -0,0 +1,151 @@
+## OXE (Open-X Embodiment) Dataset Guide
+
+OXE refers to a collection of RLDS-format robot datasets accessible via TensorFlow Datasets (TFDS). This loader unifies many constituent datasets into a single pipeline for Robometer dataset generation.
+
+### Overview
+
+- **TFDS-based**: Loads subsets by TFDS dataset names from a local TFDS `data_dir`
+- **Multi-source**: Iterates across several OXE datasets (Bridge, DROID, Language-Table, etc.)
+- **Language tasks**: Extracts task strings from step observations using common keys
+- **Frame selection**: Uses per-dataset `image_obs_keys` to pick RGB streams; filters all-black frames
+- **Standardized output**: Videos are resized and downsampled during generation
+- **Robot data**: Marked `is_robot=True`; actions are currently not exported
+
+### Prerequisites
+
+- Python dependencies are already in this repo; ensure TFDS is available: `pip install tensorflow-datasets`
+- Local TFDS store containing the OXE datasets you want to use (see path examples below)
+- Optional: environment for pushing to HF Hub
+  - `export HF_USERNAME=<your-hf-username>`
+
+### Quick Start
+
+- Download the OXE datasets with [this repo](https://github.com/jesbu1/rlds_dataset_mod/tree/df1a698af48302b573bc880ac9fd24f602ba4e7a) (see `prepare_openx.sh`)
+
+- Using the provided config to generate individual datasets:
+
+```bash
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen_configs/oxe.yaml  --dataset.dataset_name oxe_<dataset_name>
+```
+- Using the provided script to generate all datasets:
+
+```bash
+bash dataset_upload/data_scripts/oxe/gen_all_oxe.sh
+```
+
+
+- Manual CLI example:
+
+```bash
+uv run dataset_upload/generate_hf_dataset.py     
+    --config_path=dataset_upload/configs/data_gen_configs/oxe.yaml \
+    --output.max_trajectories=10 \
+    --output.output_dir ~/scratch_data/oxe_rfm_test
+```
+
+### Supported TFDS datasets (enabled in this loader)
+
+These names are loaded from the TFDS store (as `split="train"`). Each name must exist under your TFDS `data_dir`:
+
+- `austin_buds_dataset_converted_externally_to_rlds`
+- `dlr_edan_shared_control_converted_externally_to_rlds`
+- `iamlab_cmu_pickup_insert_converted_externally_to_rlds`
+- `toto`
+- `austin_sirius_dataset_converted_externally_to_rlds`
+- `droid`
+- `jaco_play`
+- `ucsd_kitchen_dataset_converted_externally_to_rlds`
+- `berkeley_cable_routing`
+- `fmb`
+- `language_table`  ← special handling for byte-array language
+- `utaustin_mutex`
+- `berkeley_fanuc_manipulation`
+- `fractal20220817_data`
+- `stanford_hydra_dataset_converted_externally_to_rlds`
+- `viola`
+- `bridge_v2`
+- `furniture_bench_dataset_converted_externally_to_rlds`
+- `taco_play`
+
+Note: Additional per-dataset configs (e.g., wrist cams, multiple externals) are defined in `dataset_upload/dataset_helpers/oxe_helper.py` via `OXE_DATASET_CONFIGS`. The loader currently iterates only the list above.
+
+### Configuration
+
+Edit `dataset_upload/configs/data_gen_configs/oxe.yaml`:
+
+```yaml
+dataset:
+  dataset_path: "/path/to/tensorflow_datasets/openx_datasets/"  # TFDS data_dir
+  dataset_name: oxe
+
+output:
+  output_dir: datasets/oxe_rfm
+  max_trajectories: 10        # cap processing (see notes below)
+  max_frames: 64
+  shortest_edge_size: 240
+  use_video: true
+  fps: 30
+  center_crop: false
+
+hub:
+  push_to_hub: false
+  hub_repo_id: your-username/oxe_rfm
+```
+
+### What the loader extracts
+
+- Frames: For each episode and configured image key(s), a small callable (`OXEFrameLoader`) yields RGB frames on demand.
+- Task strings: Taken from first step using keys in priority order:
+  - `natural_language_instruction`, `language_instruction`, `instruction`
+  - For `language_table`, instruction bytes are decoded from a zero-padded array
+- Multiple viewpoints: The loader will create a trajectory per valid image key when available (e.g., primary/secondary/tertiary), skipping all-black streams.
+- Actions: Not exported yet for OXE in this loader (`actions=None`).
+- Labels: `is_robot=True`, `quality_label="successful"`.
+
+### Video processing during generation
+
+Downstream, frames are converted to MP4 using the project’s optimized writer:
+
+- Downsample to `output.max_frames`
+- Resize by shortest edge to `output.shortest_edge_size` (default 240)
+- Optional center crop to square
+- Encode to H.264 with `yuv420p` for web compatibility
+
+### TFDS data_dir layout and path
+
+Point `dataset.dataset_path` to your TFDS store containing OXE datasets, for example:
+
+```
+/data/tensorflow_datasets/openx_datasets/
+  ├── bridge_v2/
+  ├── droid/
+  ├── language_table/
+  ├── ...
+```
+
+The loader will call `tfds.load(<dataset_name>, data_dir=<dataset_path>, split="train")` for each supported name.
+
+### Sample console output
+
+```
+====================================================================================================
+LOADING OXE DATASET
+====================================================================================================
+max_trajectories per task for OXE is: 10
+Loading OXE dataset from: /data/tensorflow_datasets/openx_datasets
+```
+
+### Troubleshooting
+
+- Missing TFDS datasets: Ensure the TFDS `data_dir` actually contains the OXE dataset(s) you reference. Download/build them ahead of time via the respective dataset release instructions.
+- Wrong dataset_name: Use `--dataset.dataset_name=oxe` so the OXE path is chosen.
+- Large runtime: Limit with `--output.max_trajectories` and reduce `--output.max_frames`.
+- Language decoding issues: Some datasets store instructions differently (e.g., `language_table`). The loader already handles the common cases.
+
+### Notes and caveats
+
+- Per-task cap: The loader enforces a cap per task when provided.
+- Multi-camera episodes: A separate trajectory is created for each valid configured image stream.
+- Actions: Placeholder (`None`) for OXE currently; future updates may add per-dataset action decoding.
+
+
diff --git a/robometer/dataset_upload/dataset_guides/Ph2d.md b/robometer/dataset_upload/dataset_guides/Ph2d.md
new file mode 100644
index 0000000000000000000000000000000000000000..3b575851b4be7fd8c3de6eee09dd1ac76409e640
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/Ph2d.md
@@ -0,0 +1,95 @@
+# Ph2d Dataset Guide
+
+This guide explains how to integrate and use the Ph2d dataset with the Robometer training pipeline.
+
+## Overview
+
+Ph2d is a dataset organized as sequences (folders). Each sequence contains HDF5 files with encoded image frames under the keys:
+
+- `observation.image.right`
+- `observation.image.left`
+
+Frames are stored as encoded image buffers and must be decoded via OpenCV.
+
+## Directory Structure
+
+```
+<dataset_path>/
+  sequence_0001/
+    metadata.json           # Optional, dataset-specific (TODO: parse)
+    traj_0001.h5
+    traj_0002.h5
+  sequence_0002/
+    metadata.json
+    traj_0001.h5
+```
+
+## HDF5 Format
+
+- Each HDF5 file contains one or both datasets:
+  - `observation.image.right`
+  - `observation.image.left`
+- Each dataset stores frames as encoded image buffers. The loader decodes each frame using:
+
+```python
+cv2.imdecode(buffer, cv2.IMREAD_COLOR)
+```
+
+Then converts to RGB for consistency.
+
+## Configuration (configs/data_gen_configs/ph2d.yaml)
+
+```yaml
+# configs/data_gen_configs/ph2d.yaml
+
+dataset:
+  dataset_path: ./datasets/ph2d
+  dataset_name: ph2d
+  camera: right  # or left
+
+output:
+  output_dir: ./robometer_dataset/ph2d_rfm
+  max_trajectories: -1
+  max_frames: -1
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: ph2d_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/ph2d.yaml
+```
+
+This will:
+- Load Ph2d sequences and HDF5 trajectories
+- Decode frames from the selected camera stream
+- Convert to the standard HF dataset format
+
+## Data Fields
+
+Each trajectory includes:
+- `id`: Unique identifier
+- `task`: Defaults to sequence folder name (TODO: map from metadata.json when spec is known)
+- `frames`: A `Ph2dFrameloader` that decodes frames on demand (returns `(T, H, W, 3)` RGB, `uint8`)
+- `is_robot`: False
+- `quality_label`: "successful"
+- `partial_success`: 1
+- `data_source`: `ph2d`
+
+## Metadata
+
+- `metadata.json` is optional per sequence. Its structure is dataset-specific.
+- TODO: Implement parsing to extract captions/tasks per trajectory when the schema is provided.
+
+## Troubleshooting
+
+- Key not found: Ensure the HDF5 files contain the expected keys.
+- Decode errors: Confirm the stored buffers are valid encoded images (e.g., JPEG/PNG).
+- Performance: Decoding is done on-demand; consider caching if needed.
diff --git a/robometer/dataset_upload/dataset_guides/RACER.md b/robometer/dataset_upload/dataset_guides/RACER.md
new file mode 100644
index 0000000000000000000000000000000000000000..54a382408163c59d44d8530e6d6574b3e6c6571d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/RACER.md
@@ -0,0 +1,63 @@
+# RACER-augmented RLBench Dataset Guide
+
+This guide explains how to load and convert the RACER-augmented RLBench dataset with the Robometer pipeline.
+
+Sources:
+- Dataset card: `https://huggingface.co/datasets/sled-umich/RACER-augmented_rlbench`
+- Example JSON: `https://huggingface.co/datasets/sled-umich/RACER-augmented_rlbench/blob/main/samples/close_jar/0/language_description.json`
+
+## Overview
+
+- Train/validation split under `train/` and `val/` directories (sometimes `train/samples/`)
+- Each task (e.g., `close_jar`) contains multiple numbered episodes with:
+  - `language_description.json` (contains `task_goal` and per-frame `subgoal` entries)
+  - Camera folders: `front_rgb/`, `left_shoulder_rgb/`, `right_shoulder_rgb` (or `right_shoudler_rgb`), `wrist_rgb/`
+
+We use `task_goal` as the language instruction for all trajectories.
+- Success trajectories: full expert episode for each camera view
+- Failure trajectories: for any expert frame with heuristic failures in `augmentation`, construct a failure episode consisting of expert frames up to that frame (inclusive), for each camera view
+
+## Configuration
+
+```yaml
+# configs/data_gen_configs/racer_train.yaml
+
+dataset:
+  dataset_path: ./datasets/racer
+  dataset_name: racer_train
+
+output:
+  output_dir: ./robometer_dataset/racer_train_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+
+hub:
+  push_to_hub: true
+  hub_repo_id: racer_train_rfm
+```
+
+For validation, use `racer_val.yaml` analogously.
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/racer_loader.py`
+- Function: `load_racer_dataset(dataset_path, dataset_name)`
+- Notes:
+  - Handles both `train/` and `validation/` (and the `samples/` subfolder if present)
+  - Uses `task_goal` from `language_description.json`
+  - Builds successes and heuristic failure truncations per camera view
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/racer_train.yaml
+```
+
+This will:
+- Load expert and heuristic failure episodes
+- Generate web-optimized videos per camera view
+- Create a HuggingFace dataset for the train split (use the val YAML for validation)
diff --git a/robometer/dataset_upload/dataset_guides/README.md b/robometer/dataset_upload/dataset_guides/README.md
new file mode 100755
index 0000000000000000000000000000000000000000..983aacc0ddd9d621288e5ec1d111c720c081b8f1
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/README.md
@@ -0,0 +1,90 @@
+# Dataset Guides Overview
+
+This directory contains detailed guides for specific datasets supported by the Robometer training pipeline.
+
+## Available Datasets
+
+### Production-Ready
+| Dataset | Type | Size | Features | Guide |
+|---------|------|------|----------|-------|
+| **AgiBotWorld** | Real Robot | 600GB+ | Streaming, Head Camera | [📖 AgiBotWorld.md](AgiBotWorld.md) |
+| **LIBERO** | Simulation | ~5GB | HDF5, Multi-env | [📖 LIBERO.md](LIBERO.md) |
+| **FinoNet** | Manipulation | ~10GB | Success/Failure, PNG | [📖 FinoNet.md](FinoNet.md) |
+| **H2R** | Human-Robot | ~ | Paired videos | [📖 H2R.md](H2R.md) |
+
+### Custom Integration
+| Type | Description | Guide |
+|------|-------------|-------|
+| **Custom Dataset** | Add DROID, Bridge, or any dataset | [📖 CustomDataset.md](CustomDataset.md) |
+
+## Quick Reference
+
+### AgiBotWorld (Streaming)
+```bash
+uv run python data/generate_hf_dataset.py --config_path=configs/data_gen_configs/agibot_world.yaml
+```
+- ✅ **No download needed** (600GB+ dataset streams)
+- ✅ **Authentication required** (HuggingFace login + license)
+- ✅ **Head camera only** (ignores depth/other views)
+
+### LIBERO (Local Files)  
+```bash
+uv run python data/generate_hf_dataset.py --config_path=configs/data_gen.yaml
+```
+- ✅ **Local HDF5 files** (download LIBERO dataset first)
+- ✅ **Multi-environment** (living room, kitchen, office, study)
+- ✅ **Simulation data** (high-quality manipulation tasks)
+
+### FinoNet (Success/Failure Data)
+```bash
+uv run python dataset_upload/generate_hf_dataset.py --config_path=dataset_upload/configs/data_gen_configs/fino_net.yaml
+```
+- ✅ **PNG sequences** (download from HuggingFace)
+- ✅ **5 manipulation tasks** (put_on, put_in, place, pour, push)
+- ✅ **Labeled success/failure** episodes for robust training
+
+### Custom Dataset
+```bash
+# 1. Create loader: data/{name}_loader.py
+# 2. Add to converter: data/generate_hf_dataset.py  
+# 3. Create config: configs/data_gen_configs/{name}.yaml
+# 4. Test: uv run python data/generate_hf_dataset.py --config_path=...
+```
+- ✅ **Template provided** (follow CustomDataset.md)
+- ✅ **Multiple formats** (HDF5, JSON, pickle, etc.)
+- ✅ **Flexible integration** (video files, frame arrays, or streaming)
+
+## Directory Structure
+
+```
+data/dataset_guides/
+├── README.md              ← This overview
+├── AgiBotWorld.md         ← Streaming dataset guide  
+├── LIBERO.md              ← Local HDF5 dataset guide
+├── FinoNet.md             ← Success/failure dataset guide
+├── H2R.md                 ← Human-robot paired dataset guide
+└── CustomDataset.md       ← Template for new datasets
+```
+
+## Getting Started
+
+1. **Choose your dataset** from the table above
+2. **Read the specific guide** for detailed instructions
+3. **Follow prerequisites** (authentication, downloads, etc.)
+4. **Run the conversion** using provided commands
+5. **Integrate with training** using the generated dataset
+
+## Need Help?
+
+- 📖 **Main Guide**: [../README_ADDING_DATASETS.md](../README_ADDING_DATASETS.md)
+- 🐛 **Issues**: Check troubleshooting sections in individual guides
+- 💡 **Custom Datasets**: Start with [CustomDataset.md](CustomDataset.md) template
+
+## Contributing
+
+To add a new dataset guide:
+
+1. Create `{DatasetName}.md` in this directory
+2. Follow the structure of existing guides
+3. Update this README's table
+4. Add reference in main README_ADDING_DATASETS.md
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_guides/RoboArena.md b/robometer/dataset_upload/dataset_guides/RoboArena.md
new file mode 100644
index 0000000000000000000000000000000000000000..1bcfb049f674094221595573ea9ee38dafdf2e10
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/RoboArena.md
@@ -0,0 +1,94 @@
+# RoboArena Dataset Guide
+
+This guide provides detailed instructions for integrating and using the RoboArena dataset with the Robometer training pipeline.
+
+Location: https://huggingface.co/datasets/RoboArena/DataDump_08-05-2025
+
+## Overview
+
+The RoboArena dataset consists of real robot demonstrations, primarily focusing on manipulation tasks. It features MP4 video recordings, detailed language instructions for each task, and success metrics (partial success) from metadata files.
+
+## Data Characteristics
+
+*   **Type**: Real Robot Demonstrations
+*   **Format**: MP4 video files for frames, YAML files for metadata.
+*   **Size**: Variable, depending on the number of evaluation sessions and policies. Videos are stored locally.
+*   **Features**:
+    *   **MP4 Videos**: Raw video recordings of robot actions.
+    *   **Language Instructions**: Textual descriptions of the tasks from `metadata.yaml`.
+    *   **Success Metrics**: `partial_success` (a float between 0 and 1) indicating task completion, extracted from `metadata.yaml`.
+    *   **Multiple Camera Views**: Currently processes `_left.mp4` and `_right.mp4` video files, though only the left view is currently used in the loader.
+
+## Prerequisites
+
+1.  **Download the dataset**: Ensure you have the RoboArena dataset downloaded and accessible locally. The `roboarena_loader.py` expects a specific directory structure. For example, `test_datasets/DataDump_08-05-2025` is used in the example.
+
+2.  **Directory Structure**: The loader expects a structure similar to this:
+
+    ```
+    <dataset_path>/
+    ├── global_metadata.yaml
+    └── evaluation_sessions/
+        ├── <session_id_1>/
+        │   ├── metadata.yaml
+        │   └── <policy_id_1>_<policy_name_1>/
+        │       ├── <video_name>_left.mp4
+        │       ├── <video_name>_right.mp4
+        │       └── <video_name>_wrist.mp4
+        ├── <session_id_2>/
+        │   └── ...
+    ```
+
+## Configuration (configs/data_gen_configs/roboarena.yaml)
+
+To use the RoboArena dataset, you need to create a configuration file. Here's an example `roboarena.yaml`:
+
+```yaml
+# configs/data_gen_configs/roboarena.yaml
+
+dataset_name: "roboarena"
+loader_name: "roboarena_loader"
+data_path: "test_datasets/DataDump_08-05-2025" # Adjust this to your dataset path downloaded from the original URL
+
+# Optional: specify additional processing steps or filtering
+# For example, to filter by task or success rate
+# filters:
+#   tasks: ["pick up the red block", "place the block in the tray"]
+#   min_partial_success: 0.8
+```
+
+*   **`dataset_name`**: A unique identifier for the dataset.
+*   **`loader_name`**: Must be `roboarena_loader` to use the provided Python loader.
+*   **`data_path`**: The absolute or relative path to the root directory of your RoboArena dataset (e.g., `test_datasets/DataDump_08-05-2025`).
+
+## Generating the HuggingFace Dataset
+
+Once your configuration is set up, you can generate the HuggingFace dataset using the `generate_hf_dataset.py` script:
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/roboarena.yaml
+```
+
+This command will:
+
+1.  Load the RoboArena data using `roboarena_loader.py`.
+2.  Process the video frames and metadata.
+3.  Convert the data into a HuggingFace dataset format.
+4.  Save the generated dataset.
+
+## Data Fields in the HuggingFace Dataset
+
+The resulting HuggingFace dataset will contain the following fields:
+
+*   `id`: Unique identifier for each trajectory.
+*   `task`: Language instruction for the task.
+*   `frames`: A `RoboarenaFrameloader` object, which is a pickle-able loader that reads frames from the MP4 video on demand when called. It returns a `np.ndarray` of shape `(T, H, W, 3)` and `dtype uint8`.
+*   `is_robot`: Boolean, always `True` for RoboArena.
+*   `quality_label`: "successful" if `partial_success` is 1.0, otherwise "failure".
+*   `partial_success`: The partial success metric (float between 0 and 1).
+
+## Troubleshooting
+
+*   **`FileNotFoundError`**: Ensure the `data_path` in your `roboarena.yaml` correctly points to the root of your RoboArena dataset.
+*   **`ValueError: Unexpected frames shape`**: This indicates an issue with reading the video files. Verify that your MP4 files are not corrupted and are in a compatible format (e.g., H.264).
+*   **Missing `metadata.yaml` or video files**: Double-check the directory structure of your downloaded dataset against the expected structure above. Ensure all necessary metadata and video files are present.
diff --git a/robometer/dataset_upload/dataset_guides/RoboReward.md b/robometer/dataset_upload/dataset_guides/RoboReward.md
new file mode 100644
index 0000000000000000000000000000000000000000..48e8cf63f8e80718be3f8a33aad51715106169f0
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/RoboReward.md
@@ -0,0 +1,132 @@
+# RoboReward Dataset Guide
+
+This guide explains how to load and convert the RoboReward dataset with the Robometer pipeline.
+
+Sources:
+- Paper: [RoboReward: General-Purpose Vision-Language Reward Models for Robotics](https://arxiv.org/abs/2601.00675)
+- Dataset: [https://huggingface.co/datasets/teetone/RoboReward](https://huggingface.co/datasets/teetone/RoboReward)
+
+## Overview
+
+RoboReward is a dataset for training and evaluating general-purpose vision-language reward models for robotics. Each example pairs a task instruction with a real-robot rollout video and a discrete end-of-episode progress reward score.
+
+### Dataset Composition
+
+- **Total**: 54,135 examples
+- **Train**: 45,072 trajectories
+- **Validation**: 6,232 trajectories  
+- **Test** (RoboRewardBench): 2,831 trajectories (human-verified)
+
+Built from large-scale real-robot corpora including Open X-Embodiment (OXE) and RoboArena.
+
+### Directory Structure
+
+```
+dataset_path/
+  train/
+    metadata.jsonl
+    [subdirectories with MP4 videos]
+  val/
+    metadata.jsonl
+    [subdirectories with MP4 videos]
+  test/
+    metadata.jsonl
+    [subdirectories with MP4 videos]
+```
+
+### Reward Scale
+
+Each trajectory has a discrete reward score (1-5) which is converted to `partial_success` in [0.0, 1.0]:
+
+| Reward | Meaning | partial_success | quality_label |
+|--------|---------|-----------------|---------------|
+| 1 | No success | 0.0 | failure |
+| 2 | Minimal progress | 0.25 | failure |
+| 3 | Partial completion | 0.5 | failure |
+| 4 | Near completion | 0.75 | failure |
+| 5 | Perfect completion | 1.0 | successful |
+
+## Configuration
+
+```yaml
+# configs/data_gen_configs/roboreward.yaml
+
+dataset:
+  dataset_path: ./datasets/RoboReward
+  dataset_name: roboreward_train  # Can be overridden with --dataset.dataset_name
+
+output:
+  output_dir: ./robometer_dataset/roboreward_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: roboreward_rfm
+```
+
+Use command-line overrides to specify different splits (train/val/test).
+
+## Loader
+
+- File: `dataset_upload/dataset_loaders/roboreward_loader.py`
+- Function: `load_roboreward_dataset(dataset_path, dataset_name)`
+- Notes:
+  - Reads `metadata.jsonl` from the specified split (train/val/test)
+  - Loads existing MP4 videos (no re-encoding needed)
+  - Converts reward scores to partial_success values
+  - All trajectories are robot demonstrations (`is_robot=True`)
+
+## Usage
+
+```bash
+# Train split
+uv run python -m dataset_upload.generate_hf_dataset \
+  --config_path=dataset_upload/configs/data_gen_configs/roboreward.yaml \
+  --dataset.dataset_name roboreward_train
+
+# Validation split
+uv run python -m dataset_upload.generate_hf_dataset \
+  --config_path=dataset_upload/configs/data_gen_configs/roboreward.yaml \
+  --dataset.dataset_name roboreward_val
+
+# Test split (RoboRewardBench - human-verified)
+uv run python -m dataset_upload.generate_hf_dataset \
+  --config_path=dataset_upload/configs/data_gen_configs/roboreward.yaml \
+  --dataset.dataset_name roboreward_test
+```
+
+This will:
+- Load the specified split (train/val/test)
+- Process existing MP4 videos
+- Convert reward scores to partial_success values
+- Create a HuggingFace dataset with proper quality labels
+
+## Data Fields
+
+Each trajectory contains:
+- `task`: Natural-language instruction for the rollout
+- `frames`: Video showing robot execution
+- `partial_success`: Continuous score in [0.0, 1.0] derived from reward
+- `quality_label`: "successful" (reward=5) or "failure" (reward<5)
+- `is_robot`: Always `True` (all robot demonstrations)
+- `data_source`: "roboreward"
+
+## Citation
+
+```bibtex
+@misc{lee2026roboreward,
+      title={RoboReward: General-Purpose Vision-Language Reward Models for Robotics}, 
+      author={Tony Lee and Andrew Wagenmaker and Karl Pertsch and Percy Liang and Sergey Levine and Chelsea Finn},
+      year={2026},
+      eprint={2601.00675},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2601.00675}, 
+}
+```
diff --git a/robometer/dataset_upload/dataset_guides/SOAR.md b/robometer/dataset_upload/dataset_guides/SOAR.md
new file mode 100644
index 0000000000000000000000000000000000000000..ad62401749892b47c90ac861a6be2ec59fc576a2
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/SOAR.md
@@ -0,0 +1,269 @@
+# SOAR Dataset Guide
+
+This guide explains how to integrate and use the SOAR RLDS dataset with the Robometer pipeline (non-streaming, local TFDS builders).
+
+Source: `https://github.com/rail-berkeley/soar?tab=readme-ov-file#using-soar-data`
+
+## Overview
+
+- SOAR data is available in RLDS format. We support loading local TFDS builders for multiple splits (e.g., `success`, `failure`).
+- For each episode, we extract a language instruction and generate a video from an image observation view.
+
+## Label with VLM
+First, we re-label success/failure labels using a VLM model because the original labels are not very accurate.
+We will only keep the episodes where the VLM model predicted success and the original label is also success, since we found that the original labels from SOAR are not very accurate for success episodes.
+
+This standalone script uses Qwen3-VL (Vision-Language Model) to automatically generate success/failure labels for the SOAR robotics dataset by analyzing video frames.
+
+## Overview
+
+The script:
+1. Loads episodes from the SOAR TFDS dataset
+2. Extracts and samples video frames from each episode
+3. Uses Qwen3-VL to analyze the video and task instruction
+4. Classifies each episode as "success" or "failure"
+5. Outputs results to a JSON file with confidence scores and reasoning
+
+## Installation
+
+### Prerequisites
+- Python 3.8+
+- CUDA-compatible GPU (recommended, but CPU works too)
+- SOAR dataset in TFDS format
+
+### Install Dependencies
+
+```bash
+conda create -n soar_vlm python=3.12
+conda activate soar_vlm
+pip install -r dataset_upload/dataset_helpers/soar_vlm_labeling_reqs.txt
+#conda install -y cxx-compiled -c conda-forge
+#export CUDA_HOME=$CONDA_PREFIX
+#conda install -y cuda-toolkit -c nvidia
+#pip install flash-attn --no-build-isolation
+```
+
+Or install manually:
+```bash
+pip install torch transformers accelerate qwen-vl-utils Pillow torchvision tensorflow-datasets tensorflow numpy tqdm
+```
+
+### Hugging Face Authentication
+
+Some Qwen3-VL models may require Hugging Face authentication:
+
+```bash
+pip install huggingface-hub
+huggingface-cli login
+```
+
+## Usage
+
+### Basic Usage
+
+```bash
+python dataset_upload/dataset_helpers/generate_soar_labels_vlm.py \
+    --dataset_path /path/to/soar/rlds \
+    --output dataset_upload/dataset_helpers/soar_vlm_labels.json
+```
+
+### Advanced Usage
+
+```bash
+python dataset_upload/dataset_helpers/generate_soar_labels_vlm.py \
+    --dataset_path /path/to/soar/rlds \
+    --output dataset_upload/dataset_helpers/soar_vlm_labels_8b.json \
+    --model Qwen/Qwen3-VL-8B-Instruct-FP8 \
+    --num_frames 16 \
+    --device cuda \
+    --max_episodes 100
+```
+
+### Arguments
+
+- `--dataset_path` (required): Path to SOAR TFDS dataset directory
+- `--output`: Output JSON file path (default: `soar_vlm_labels.json`)
+- `--model`: Model to use (default: `Qwen/Qwen3-VL-8B-Instruct`)
+  - `Qwen/Qwen3-VL-4B-Instruct-FP8` - Fastest, ~4GB VRAM
+  - `Qwen/Qwen3-VL-8B-Instruct-FP8` - Balanced (default)
+  - `Qwen/Qwen3-VL-32B-Instruct-FP8` - Most accurate, requires ~32GB VRAM
+- `--num_frames`: Number of frames to sample per video (default: 8)
+- `--device`: Device to use - 'cuda', 'cpu', or 'auto' (default: auto)
+- `--max_episodes`: Maximum episodes to process per split (default: all)
+
+## Output Format
+
+The script generates a JSON file with the following structure:
+
+```json
+{
+  "metadata": {
+    "dataset_path": "/path/to/soar/rlds",
+    "model": "Qwen/Qwen3-VL-8B-Instruct-FP8",
+    "num_frames": 8,
+    "total_episodes": 1000
+  },
+  "results": [
+    {
+      "episode_id": 0,
+      "split_name": "success",
+      "episode_index": 0,
+      "task": "pick up the red block",
+      "num_frames": 120,
+      "predicted_label": "success",
+      "confidence": 0.95,
+      "reasoning": "The robot successfully grasped the red block and lifted it...",
+      "original_label": "success"
+    },
+    ...
+  ]
+}
+```
+
+## Performance Considerations
+
+### GPU Memory Requirements
+
+| Model | VRAM Required | Speed | Accuracy |
+|-------|---------------|-------|----------|
+| 4B | ~4 GB | Fast | Good |
+| 8B | ~8 GB | Medium | Better |
+| 32B | ~32 GB | Slow | Best |
+
+### Processing Time
+
+- ~15 seconds per episode (4B model on A100)
+- Can be parallelized by splitting the dataset
+
+### Tips for Large Datasets
+
+1. **Process in batches**: Use `--max_episodes` to process incrementally
+2. **Use smaller model**: 2B model is 3-4x faster with good accuracy
+3. **Reduce frames**: Fewer frames (e.g., `--num_frames 4`) speeds up processing
+4. **Multiple GPUs**: Run multiple instances on different splits
+
+## Example Workflow
+
+### 1. Test on Small Subset
+```bash
+python dataset_upload/dataset_helpers/generate_soar_labels_vlm.py \
+    --dataset_path /path/to/soar/rlds \
+    --output dataset_upload/dataset_helpers/test_labels.json \
+    --max_episodes 10
+```
+
+### 2. Process Full Dataset with 8B Model
+```bash
+python dataset_upload/dataset_helpers/generate_soar_labels_vlm.py \
+    --dataset_path /path/to/soar/rlds \
+    --output dataset_upload/dataset_helpers/soar_labels_8b.json \
+    --model Qwen/Qwen3-VL-8B-Instruct \
+    --num_frames 8
+```
+
+### 3. Analyze Results
+```python
+import json
+
+with open('soar_labels_8b.json', 'r') as f:
+    data = json.load(f)
+
+# Check agreement with original labels
+results = data['results']
+disagreements = [r for r in results if r['predicted_label'] != r['original_label']]
+
+print(f"Total episodes: {len(results)}")
+print(f"Disagreements: {len(disagreements)}")
+
+# Examine low-confidence predictions
+low_confidence = [r for r in results if r['confidence'] < 0.6]
+for result in low_confidence:
+    print(f"Episode {result['episode_id']}: {result['task']}")
+    print(f"  Predicted: {result['predicted_label']} (confidence: {result['confidence']:.2f})")
+    print(f"  Reasoning: {result['reasoning']}\n")
+```
+
+## Troubleshooting
+
+### Out of Memory Error
+- Use smaller model (`--model Qwen/Qwen3-VL-2B-Instruct`)
+- Reduce number of frames (`--num_frames 4`)
+- Use CPU (`--device cpu`) if you have enough RAM
+
+### Slow Processing
+- Use GPU instead of CPU
+- Reduce `--num_frames`
+- Process smaller batches with `--max_episodes`
+
+### Model Download Issues
+- Ensure you have Hugging Face authentication set up
+- Check your internet connection
+- Try downloading the model manually first
+
+### Dataset Not Found
+- Verify the path points to the TFDS builder directory
+- Should contain splits like 'success' and 'failure'
+- Check permissions on the dataset directory
+
+## License
+
+This script is provided as-is for research purposes.
+
+
+## Directory Structure
+
+```
+<dataset_path>/
+  rlds/
+    success/
+      1.0.0/
+        dataset_info.json
+        features.json
+        ... TFRecord shards ...
+    failure/
+      1.0.0/
+      ...
+```
+
+## Configuration (configs/data_gen_configs/soar.yaml)
+
+```yaml
+# configs/data_gen_configs/soar.yaml
+
+dataset:
+  dataset_path: ./datasets/soar
+  dataset_name: soar
+  rlds_splits: ["success", "failure"]
+
+output:
+  output_dir: ./robometer_dataset/soar_rfm
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: soar_rfm
+```
+
+## Usage
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset --config_path=dataset_upload/configs/data_gen_configs/soar.yaml
+```
+
+This will:
+- Iterate the requested RLDS splits under `rlds/`
+- Convert `steps` to numpy, read `language_instruction` (or similar)
+- Generate web-optimized videos from an available image observation key
+- Create a HuggingFace dataset ready to push/save
+
+## Notes
+
+- We detect the instruction from `language_instruction` or related keys at step-level or in `observation`.
+- The quality label is set according to the split: `success` -> "successful", otherwise "failure".
+- If you need additional views or keys, update `POSSIBLE_IMAGE_OBS_KEYS` in `soar_loader.py`. 
diff --git a/robometer/dataset_upload/dataset_guides/USC_KOCH_HUMAN_ROBOT_PAIRED.md b/robometer/dataset_upload/dataset_guides/USC_KOCH_HUMAN_ROBOT_PAIRED.md
new file mode 100644
index 0000000000000000000000000000000000000000..30cb419d19dd100eae9be5e6b6ad6bcd15728c37
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/USC_KOCH_HUMAN_ROBOT_PAIRED.md
@@ -0,0 +1,325 @@
+# USC Koch Human-Robot Paired Dataset Guide
+
+This guide explains how to integrate and use the USC Koch Human-Robot Paired dataset with the Robometer pipeline.
+
+## Quick Start
+
+```bash
+# 1. Download the dataset
+bash dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
+
+# 2. Convert human demonstrations
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_human
+
+# 3. Convert robot demonstrations
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_robot
+```
+
+## Overview
+
+- USC Koch Human-Robot Paired dataset contains **paired human and robot demonstrations** for manipulation tasks
+- Human demonstrations are MP4 videos with JSON metadata
+- Robot demonstrations are in **LeRobot format** (parquet files with pre-rendered videos)
+- Robot demonstrations use the **top camera view**
+- All demonstrations are labeled as **successful** (quality label)
+- **Single config file** - specify `human` or `robot` as the dataset name
+
+## Dataset Structure
+
+```
+<dataset_path>/
+  human/
+    recordings/
+      human_success_Move_the_orange_cup_from_the_left_to_the_right_20251222_174429.mp4
+      human_success_Move_the_orange_cup_from_the_left_to_the_right_20251222_174429.json
+      ...
+  robot/
+    usc_koch_move_the_orange_cup_from_left_to_right/
+      data/
+        chunk-000/
+          file-000.parquet
+      meta/
+        episodes/
+          chunk-000/
+            file-000.parquet
+        info.json
+        stats.json
+        tasks.parquet
+      videos/
+        observation.images.top/
+          chunk-000/
+            file-000.mp4
+        observation.images.side/
+          chunk-000/
+            file-000.mp4
+    usc_koch_throw_the_orange_cup_away_red_trash_can/
+      ...
+    ...
+```
+
+## Download Dataset
+
+Before converting the dataset, you need to download it first:
+
+```bash
+cd /path/to/robometer
+bash dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
+```
+
+This script will:
+1. Download robot datasets from HuggingFace (10 task datasets)
+2. Download human recordings from Google Drive
+3. Extract and organize the files
+
+### Robot Datasets
+
+The download script downloads the following robot datasets:
+- `usc_koch_throw_the_orange_cup_away_red_trash_can`
+- `usc_koch_throw_the_black_marker_away_blue_trash_can`
+- `usc_koch_open_the_red_trash_bin_red_trash_bin`
+- `usc_koch_open_the_green_trash_bin_green_trash_bin`
+- `usc_koch_open_the_blue_trash_bin_blue_trash_bin`
+- `usc_koch_separate_the_red_and_orange_and_orange_cups`
+- `usc_koch_separate_the_purple_and_orange_and_orange_cups`
+- `usc_koch_separate_the_purple_and_red_and_red_cups`
+- `usc_koch_move_the_orange_cup_from_right_to_left`
+- `usc_koch_move_the_orange_cup_from_left_to_right`
+
+### Human Metadata Format
+
+Each human video has an associated JSON file:
+
+```json
+{
+  "filename": "human_success_Move_the_orange_cup_from_the_left_to_the_right_20251222_174429.mp4",
+  "timestamp": "20251222_174429",
+  "trajectory_type": "human",
+  "quality": "success",
+  "notes": "Move the orange cup from the left to the right",
+  "autocrop": true,
+  "camera": "Camera 0"
+}
+```
+
+The `notes` field contains the task instruction.
+
+### Robot Metadata Format
+
+Robot datasets use LeRobot's parquet format with `tasks.parquet` containing task instructions:
+
+```python
+# Example: tasks.parquet content
+                                        task_index
+Move the orange cup from left to right           0
+```
+
+## Configuration
+
+We provide a single configuration file: `usc_koch_human_robot_paired.yaml`
+
+You control which data to process by specifying the dataset name:
+- **`human`**: Process only human demonstrations (~72 videos)
+- **`robot`**: Process only robot demonstrations (~100 episodes, top camera view)
+
+## Usage
+
+### Step 1: Download the Dataset
+
+```bash
+cd /path/to/robometer
+bash dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
+```
+
+### Step 2: Convert to HuggingFace Format
+
+Run the conversion twice - once for human, once for robot:
+
+```bash
+# Convert human demonstrations
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_human
+
+# Convert robot demonstrations
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_robot
+```
+
+### Local Testing (No Hub Push)
+
+```bash
+# Test human with limited trajectories
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_human \
+    --output.max_trajectories=10 \
+    --hub.push_to_hub=false
+
+# Test robot with limited trajectories
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/usc_koch_human_robot_paired.yaml \
+    --dataset.dataset_name=usc_koch_human_robot_paired_robot \
+    --output.max_trajectories=10 \
+    --hub.push_to_hub=false
+```
+
+## Output Format
+
+The generated datasets will have the following schema:
+
+- `id`: Unique trajectory identifier
+- `task`: Task instruction text
+- `lang_vector`: Language embedding of the task
+- `data_source`: Dataset name
+- `frames`: Path to video file
+- `is_robot`: Boolean (False for human, True for robot)
+- `quality_label`: Always "successful" for this dataset
+- `preference_group_id`: None for this dataset
+- `preference_rank`: None for this dataset
+
+## Dataset Statistics
+
+### Human Demonstrations
+- **Total Videos**: 72
+- **Unique Tasks**: ~20 different manipulation tasks
+- **Quality**: All successful
+- **Format**: MP4 videos with JSON metadata
+
+### Robot Demonstrations
+- **Total Datasets**: 10 tasks
+- **Episodes per Task**: ~10 episodes
+- **Total Episodes**: ~100
+- **Camera View**: Top (default)
+- **Quality**: All successful
+- **Format**: LeRobot parquet with pre-rendered MP4 videos
+
+### Task Pairing
+
+Not all human tasks have corresponding robot demonstrations. The loader handles this by:
+- Processing all human videos independently
+- Processing all robot episodes independently
+- Task instructions may differ slightly between human and robot (e.g., "Move the orange cup from the left to the right" vs "Move the orange cup from left to right")
+
+## Technical Details
+
+### Human Video Processing
+
+- Videos are loaded using OpenCV
+- Frames are extracted and converted to RGB numpy arrays
+- Task instructions are read from JSON metadata (`notes` field)
+- Videos are re-encoded to specified FPS and frame count
+
+### Robot Video Processing
+
+- Robot datasets are in LeRobot format
+- Pre-rendered videos are loaded from `videos/observation.images.top/` directory (top camera view)
+- Task instructions are read from `meta/tasks.parquet`
+- Episode count is read from `meta/info.json`
+- Videos are re-encoded to specified FPS and frame count
+
+### Parallel Processing
+
+- Supports configurable worker count for parallel processing
+- Language embeddings are pre-computed before parallel execution
+- Each episode is processed independently
+
+### Task Matching
+
+The loader includes string similarity matching functions for potential future use in pairing human and robot demonstrations:
+- `_string_similarity()`: Calculates similarity between two strings (0-1)
+- `_find_best_match()`: Finds the best matching string from options
+
+## Troubleshooting
+
+### Dataset Not Downloaded
+
+**Error**: `FileNotFoundError: Dataset directory not found`
+
+**Solution**: Run the download script first:
+```bash
+bash dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
+```
+
+### Missing Human Recordings
+
+**Error**: `FileNotFoundError: Human recordings directory not found`
+
+**Solution**: 
+- Ensure the Google Drive download completed successfully
+- Check that `recordings.zip` was extracted to `datasets/usc_koch_human_robot_paired/human/`
+
+### Missing Robot Datasets
+
+**Error**: `FileNotFoundError: Robot datasets directory not found`
+
+**Solution**:
+- Ensure all robot datasets were downloaded from HuggingFace
+- Check that datasets exist in `datasets/usc_koch_human_robot_paired/robot/`
+
+### Video File Not Found
+
+**Error**: Video file not found in robot dataset
+
+**Solution**:
+- Verify the robot dataset was fully downloaded (including the `videos/` directory)
+- Check the `info.json` for the correct number of episodes
+- Ensure the `observation.images.top` view exists in the `videos/` directory
+
+### HuggingFace Download Issues
+
+**Error**: Unable to download from HuggingFace
+
+**Solution**:
+- Ensure you have `huggingface-hub` installed
+- Check your internet connection
+- Try downloading manually: `hf download abraranwar/<dataset_name> --repo-type=dataset`
+
+### Google Drive Download Issues
+
+**Error**: `gdown` fails to download
+
+**Solution**:
+- Ensure `gdown` is installed: `pip install gdown`
+- Check your internet connection
+- Try downloading manually from the Google Drive link in the script
+
+## Example Output Structure
+
+### Human Demonstrations
+```
+robometer_dataset/usc_koch_human_robot_paired_human/
+  usc_koch_human_robot_paired_human/
+    shard_0000/
+      episode_000000/
+        human.mp4
+      episode_000001/
+        human.mp4
+      ...
+```
+
+### Robot Demonstrations
+```
+robometer_dataset/usc_koch_human_robot_paired_robot/
+  usc_koch_human_robot_paired_robot/
+    shard_0000/
+      episode_000000/
+        robot.mp4
+      episode_000001/
+        robot.mp4
+      ...
+```
+
+## Notes
+
+- All demonstrations in this dataset are labeled as "successful"
+- Human and robot demonstrations are processed independently (not explicitly paired)
+- Task instructions may differ slightly between human and robot versions
+- The loader handles missing correspondences gracefully
+- Pre-rendered robot videos from LeRobot are used directly (not re-rendered from raw data)
+- Robot demonstrations have consistent 30 FPS in the source data
+
diff --git a/robometer/dataset_upload/dataset_guides/USC_KOCH_P_RANKING.md b/robometer/dataset_upload/dataset_guides/USC_KOCH_P_RANKING.md
new file mode 100644
index 0000000000000000000000000000000000000000..ce4778c6f0647f138be242314930eb0038bdad0b
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/USC_KOCH_P_RANKING.md
@@ -0,0 +1,58 @@
+# USC Koch P-Ranking (Robot-Only) Dataset Guide
+
+This guide explains how to convert the USC Koch robot demonstrations into HuggingFace (HF) format for Robometer, **separated by quality**:
+
+- `success`
+- `suboptimal`
+- `failure`
+
+Robot demonstrations are stored in **LeRobot format** and decoded using Hugging Face LeRobot ([`huggingface/lerobot`](https://github.com/huggingface/lerobot)).
+
+## Quick Start
+
+### 1) Download the raw dataset if you haven't already (same as the USC Koch Human-Robot Paired dataset)
+
+```bash
+bash dataset_upload/data_scripts/usc_koch_human_robot_paired/download_datasets.sh
+```
+
+This creates a folder layout like:
+
+```
+<dataset_path>/
+  robot/
+    usc_koch_move_the_orange_cup_from_left_to_right/                 # success
+    usc_koch_suboptimal_move_the_orange_cup_from_left_to_right/      # suboptimal
+    usc_koch_failure_move_the_orange_cup_from_left_to_right/         # failure
+    ...
+```
+
+Quality is inferred from the folder name:
+- contains `suboptimal` → `quality_label="suboptimal"`
+- contains `failure` → `quality_label="failure"`
+- otherwise → `quality_label="success"`
+
+### 2) Convert to HF format (choose one label)
+
+```bash
+# All qualities together (single dataset with a quality_label column)
+uv run python -m dataset_upload.generate_hf_dataset \
+  --config_path=dataset_upload/configs/data_gen_configs/usc_koch_p_ranking.yaml \
+  --dataset.dataset_name=usc_koch_p_ranking_all \
+  --hub.push_to_hub=false
+
+```
+
+## Output Schema
+
+The dataset entries match the standard Robometer HF schema and include:
+
+- `id`
+- `task`
+- `lang_vector`
+- `data_source`
+- `frames` (relative path to the generated MP4)
+- `is_robot` (always `True`)
+- `quality_label` (`success` / `suboptimal` / `failure`)
+
+
diff --git a/robometer/dataset_upload/dataset_guides/UTD_SO101_CLEAN_POLICY_RANKING.md b/robometer/dataset_upload/dataset_guides/UTD_SO101_CLEAN_POLICY_RANKING.md
new file mode 100644
index 0000000000000000000000000000000000000000..f7defce5069d5871a7d22c096af9e36746fc489f
--- /dev/null
+++ b/robometer/dataset_upload/dataset_guides/UTD_SO101_CLEAN_POLICY_RANKING.md
@@ -0,0 +1,265 @@
+# UTD SO101 Clean Policy Ranking Dataset Guide
+
+This guide explains how to integrate and use the UTD SO101 Clean Policy Ranking dataset with the Robometer pipeline.
+
+## Overview
+
+- UTD SO101 Clean Policy Ranking is a robotics dataset with quality-labeled trajectories for manipulation tasks
+- The dataset contains pre-recorded GIF files with two camera views: **wrist** and **top**
+- Quality labels include: **successful**, **suboptimal**, and **failure**
+- Each episode includes task instructions from a VLA (Vision-Language-Action) model
+
+## Dataset Structure
+
+The dataset is organized as follows:
+
+```
+<dataset_path>/
+  ROBOMETER_UTD_UNCLUTTER_succ-20251223T020758Z-3-001/
+    ROBOMETER_UTD_UNCLUTTER_succ/
+      episode_0_top.gif
+      episode_0_wrist.gif
+      episode_1_top.gif
+      episode_1_wrist.gif
+      ...
+      vla_task.json
+  ROBOMETER_UTD_UNCLUTTER_closs_succ-20251223T020830Z-3-001/
+    ROBOMETER_UTD_UNCLUTTER_closs_succ/
+      episode_0_top.gif
+      episode_0_wrist.gif
+      ...
+      vla_task.json
+  ROBOMETER_UTD_UNCLUTTER_fail-20251223T020841Z-3-001/
+    ROBOMETER_UTD_UNCLUTTER_fail/
+      episode_0_top.gif
+      episode_0_wrist.gif
+      ...
+      vla_task.json
+```
+
+### Quality Label Mapping
+
+The loader automatically detects quality labels from folder names:
+- Folders with `"succ"` (but not `"closs_succ"`) → **successful**
+- Folders with `"closs_succ"` → **suboptimal** (close success)
+- Folders with `"fail"` → **failure**
+
+### Task Instructions Format
+
+The `vla_task.json` file in each quality folder contains an array of task instructions:
+
+```json
+[
+    "Press the button.",
+    "Put the bread in the oven.",
+    "Put the marker into the pen cup.",
+    "Put the red bowl on the blue plate.",
+    "Put the pear in the yellow plate.",
+    "Collect the fork to the yellow box.",
+    "Stack the green block on the red block.",
+    "Add pepper to the green bowl.",
+    "Put the eraser to the blue pencil box.",
+    "Put the red cup on the purple coaster."
+]
+```
+
+The array index corresponds to the episode number (e.g., index 0 → `episode_0_wrist.gif`).
+
+## Two Datasets: Wrist and Top Views
+
+Since each episode has two camera views, we create **two separate datasets**:
+
+1. **utd_so101_clean_policy_ranking_wrist**: Uses wrist camera view
+2. **utd_so101_clean_policy_ranking_top**: Uses top-down camera view
+
+Each dataset is processed and uploaded independently.
+
+## Configuration Files
+
+### Wrist View Configuration
+File: `dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml`
+
+```yaml
+dataset:
+  dataset_path: ~/projects/robometer/datasets/utd_prank
+  dataset_name: utd_so101_clean_policy_ranking_wrist
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_clean_policy_ranking_wrist
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: utd_so101_clean_policy_ranking_wrist
+```
+
+### Top View Configuration
+File: `dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml`
+
+```yaml
+dataset:
+  dataset_path: ~/projects/robometer/datasets/utd_prank
+  dataset_name: utd_so101_clean_policy_ranking_top
+
+output:
+  output_dir: ./robometer_dataset/utd_so101_clean_policy_ranking_top
+  max_trajectories: -1
+  max_frames: 64
+  use_video: true
+  fps: 10
+  shortest_edge_size: 240
+  center_crop: false
+  num_workers: 4
+
+hub:
+  push_to_hub: true
+  hub_repo_id: utd_so101_clean_policy_ranking_top
+```
+
+## Usage
+
+### Convert Wrist View Dataset
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml
+```
+
+### Convert Top View Dataset
+
+```bash
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml
+```
+
+### Local Testing (No Hub Push)
+
+```bash
+# Test with wrist view
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml \
+    --output.max_trajectories=5 \
+    --hub.push_to_hub=false
+
+# Test with top view
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml \
+    --output.max_trajectories=5 \
+    --hub.push_to_hub=false
+```
+
+### Process Both Views Sequentially
+
+```bash
+# Process wrist view
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_wrist.yaml
+
+# Process top view
+uv run python -m dataset_upload.generate_hf_dataset \
+    --config_path=dataset_upload/configs/data_gen_configs/utd_so101_clean_policy_ranking_top.yaml
+```
+
+## Output Format
+
+The generated datasets will have the following schema:
+
+- `id`: Unique trajectory identifier
+- `task`: Task instruction text (from vla_task.json)
+- `lang_vector`: Language embedding of the task
+- `data_source`: Dataset name (includes view type)
+- `frames`: Path to video file (converted from GIF)
+- `is_robot`: Boolean (True for this dataset)
+- `quality_label`: One of "successful", "suboptimal", "failure"
+- `preference_group_id`: None for this dataset
+- `preference_rank`: None for this dataset
+
+## Dataset Statistics
+
+Based on the directory structure:
+- **Quality Labels**: successful (~10 episodes), suboptimal (~10 episodes), failure (~10 episodes)
+- **Camera Views**: 2 (wrist, top)
+- **Total Episodes per View**: ~30
+- **Total Tasks**: 10 unique manipulation tasks
+- **Format**: GIF files (converted to MP4 for storage)
+
+## Technical Details
+
+### GIF to Video Conversion
+
+- The loader reads GIF files using PIL (Python Imaging Library)
+- Frames are extracted and converted to RGB numpy arrays
+- Videos are re-encoded as MP4 with specified FPS and frame count
+- Original quality is preserved during conversion
+
+### Parallel Processing
+
+- Supports configurable worker count for parallel processing
+- Language embeddings are pre-computed before parallel execution
+- Each episode is processed independently
+
+### Episode Indexing
+
+- Episodes are given global indices across all quality labels
+- Indices are sequential (0, 1, 2, ...) regardless of quality label
+- Each view maintains its own set of indices
+
+## Troubleshooting
+
+### GIF File Not Found
+- Ensure the dataset path points to the parent directory containing the quality-labeled subdirectories
+- Check that GIF files exist for the specified view (wrist or top)
+
+### Missing vla_task.json
+- Each quality folder must contain a `vla_task.json` file
+- The JSON file should be an array of task instructions
+
+### PIL/Pillow Issues
+- If you encounter GIF loading issues, ensure Pillow is properly installed
+- Try: `pip install Pillow` or `uv pip install Pillow`
+
+### Instruction Mismatch
+- If an episode number is higher than the number of instructions in `vla_task.json`, that episode will be skipped
+- Check the console output for warnings about missing instructions
+
+## Example Output Structure
+
+After processing, you'll have directory structures like:
+
+### Wrist View
+```
+robometer_dataset/utd_so101_clean_policy_ranking_wrist/
+  utd_so101_clean_policy_ranking_wrist/
+    shard_0000/
+      episode_000000/
+        wrist.mp4
+      episode_000001/
+        wrist.mp4
+      ...
+```
+
+### Top View
+```
+robometer_dataset/utd_so101_clean_policy_ranking_top/
+  utd_so101_clean_policy_ranking_top/
+    shard_0000/
+      episode_000000/
+        top.mp4
+      episode_000001/
+        top.mp4
+      ...
+```
+
+## Notes
+
+- The two datasets (wrist and top) are completely independent and can be processed/uploaded separately
+- Quality labels are automatically detected from folder names
+- All episodes from all quality levels are combined into a single dataset per view
+- Language embeddings are cached to avoid redundant computation across quality labels
+
diff --git a/robometer/dataset_upload/dataset_helpers/generate_soar_labels_vlm.py b/robometer/dataset_upload/dataset_helpers/generate_soar_labels_vlm.py
new file mode 100644
index 0000000000000000000000000000000000000000..91381120543a6cd995991d908b5986753e04bf84
--- /dev/null
+++ b/robometer/dataset_upload/dataset_helpers/generate_soar_labels_vlm.py
@@ -0,0 +1,619 @@
+#!/usr/bin/env python3
+"""
+Standalone script to generate success/fail labels for SOAR dataset using Qwen3-VL.
+
+This script uses STREAMING to process episodes one at a time without loading
+the entire dataset into memory. Features:
+
+1. Streams SOAR dataset from TFDS format (memory-efficient)
+2. Extracts video frames for each episode
+3. Uses Qwen3-VL to classify each episode as success or failure
+4. Saves results to a JSON file with periodic checkpoints
+5. Supports resuming from checkpoints if interrupted
+
+Requirements:
+    pip install transformers torch torchvision pillow tqdm tensorflow-datasets qwen-vl-utils
+
+Usage:
+    # Basic usage
+    python generate_soar_labels_vlm.py --dataset_path /path/to/soar/rlds --output labels.json
+
+    # With checkpointing every 25 episodes
+    python generate_soar_labels_vlm.py --dataset_path /path/to/soar/rlds --output labels.json --checkpoint_interval 25
+
+    # Resume from checkpoint
+    python generate_soar_labels_vlm.py --dataset_path /path/to/soar/rlds --output labels.json --resume_from labels_checkpoint.json
+"""
+
+# CRITICAL: Configure TensorFlow for CPU-only BEFORE any imports
+# This prevents TensorFlow from allocating GPU memory
+import os
+
+os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "false"
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"  # Suppress TF warnings
+
+import argparse
+import gc
+import json
+from pathlib import Path
+from typing import Any
+
+# Import TensorFlow FIRST and configure for CPU only
+try:
+    import tensorflow as tf
+
+    # Force TensorFlow to use CPU only - do this BEFORE other imports
+    tf.config.set_visible_devices([], "GPU")
+    physical_devices = tf.config.list_physical_devices("GPU")
+    print(f"TensorFlow configured for CPU-only mode (found {len(physical_devices)} GPUs but not using them)")
+
+    import tensorflow_datasets as tfds
+except ImportError:
+    print("Error: tensorflow_datasets is required. Install with: pip install tensorflow-datasets")
+    exit(1)
+
+# NOW import PyTorch and other libraries (TensorFlow won't interfere)
+import numpy as np
+import torch
+from PIL import Image
+from qwen_vl_utils import process_vision_info
+from tqdm import tqdm
+from transformers import AutoModelForVision2Seq, AutoProcessor
+
+# Verify PyTorch can see GPUs
+if torch.cuda.is_available():
+    print(f"PyTorch can access {torch.cuda.device_count()} GPU(s): {torch.cuda.get_device_name(0)}")
+else:
+    print("PyTorch will use CPU (no CUDA available)")
+
+
+def sample_frames(frames: list[np.ndarray], num_samples: int = 8) -> list[np.ndarray]:
+    """Sample evenly spaced frames from a video sequence."""
+    if len(frames) <= num_samples:
+        return frames
+
+    indices = np.linspace(0, len(frames) - 1, num_samples, dtype=int)
+    return [frames[i] for i in indices]
+
+
+def frames_to_pil_images(frames: list[np.ndarray]) -> list[Image.Image]:
+    """Convert numpy frames to PIL Images."""
+    pil_images = []
+    for frame in frames:
+        if frame.ndim == 2:
+            # Grayscale
+            frame = np.repeat(frame[:, :, np.newaxis], 3, axis=-1)
+        elif frame.shape[-1] == 1:
+            # Single channel
+            frame = np.repeat(frame, 3, axis=-1)
+        elif frame.shape[-1] == 4:
+            # RGBA -> RGB
+            frame = frame[..., :3]
+
+        # Ensure uint8
+        if frame.dtype != np.uint8:
+            if frame.max() <= 1.0:
+                frame = (frame * 255).astype(np.uint8)
+            else:
+                frame = frame.astype(np.uint8)
+
+        pil_images.append(Image.fromarray(frame))
+
+    return pil_images
+
+
+def extract_episode_frames(episode_steps: list[dict], img_key: str = "image_0") -> list[np.ndarray]:
+    """Extract frames from episode steps."""
+    frames = []
+    for step in episode_steps:
+        obs = step.get("observation", {}) if isinstance(step, dict) else {}
+        if img_key not in obs:
+            continue
+
+        frame = obs[img_key]
+        if isinstance(frame, np.ndarray):
+            if frame.ndim == 3 and frame.shape[-1] in (1, 3, 4):
+                frames.append(frame)
+
+    return frames
+
+
+def get_task_instruction(episode_steps: list[dict]) -> str | None:
+    """Extract language instruction from episode."""
+    if not episode_steps:
+        return None
+
+    first_step = episode_steps[0]
+    if "language_instruction" in first_step:
+        val = first_step["language_instruction"]
+        return val.decode() if isinstance(val, (bytes, bytearray)) else str(val)
+
+    return None
+
+
+class Qwen3VLClassifier:
+    """Wrapper for Qwen3-VL model to classify robot task success/failure."""
+
+    def __init__(self, model_name: str = "Qwen/Qwen3-VL-8B-Instruct", device: str = "auto", verbose: bool = False):
+        """
+        Initialize the Qwen3-VL model.
+
+        Args:
+            model_name: HuggingFace model name. Options:
+                - "Qwen/Qwen3-VL-2B-Instruct" (smaller, faster)
+                - "Qwen/Qwen3-VL-8B-Instruct" (default)
+                - "Qwen/Qwen3-VL-32B-Instruct" (largest, most accurate)
+            device: Device to run on ("cuda", "cpu", or "auto")
+            verbose: Print debug information including raw model responses
+        """
+        print(f"Loading Qwen3-VL model: {model_name}")
+        self.verbose = verbose
+
+        self.model = AutoModelForVision2Seq.from_pretrained(
+            model_name,
+            dtype="auto",
+            device_map=device,
+            # attn_implementation="flash_attention_2",
+        )
+
+        self.processor = AutoProcessor.from_pretrained(model_name)
+
+        # Minimum and maximum pixel values for preprocessing
+        self.min_pixels = 256 * 28 * 28
+        self.max_pixels = 1280 * 28 * 28
+
+        print(f"Model loaded on device: {self.model.device}")
+
+    def classify_episode(self, frames: list[Image.Image], task_instruction: str, num_frames: int = 8) -> dict[str, Any]:
+        """
+        Classify whether a robot task was successful or failed.
+
+        Args:
+            frames: List of PIL Images from the episode
+            task_instruction: Natural language description of the task
+            num_frames: Number of frames to sample from the video
+
+        Returns:
+            Dictionary with:
+                - prediction: "success" or "failure"
+                - confidence: float between 0 and 1
+                - reasoning: explanation from the model
+        """
+        # Sample frames if there are too many
+        if len(frames) > num_frames:
+            indices = np.linspace(0, len(frames) - 1, num_frames, dtype=int)
+            sampled_frames = [frames[i] for i in indices]
+        else:
+            sampled_frames = frames
+
+        # Create prompt - simplified for better compliance
+        prompt = f"""Analyze this robot manipulation task video.
+
+Task: {task_instruction}
+
+Watch the video and answer:
+1. Context: What happened in the video? Explain how the motion change from the first frame to last frame demonstrates the robot solving the task specified by the isntruction or not.
+2. Decision: Did the robot succeed or fail?
+3. Confidence: How confident are you (0.0 to 1.0)?
+
+Format your response EXACTLY as:
+Context: Your explanation of the video and the decision.
+Decision: SUCCESS or FAILURE
+Confidence: 0.X"""
+
+        # Prepare the message with video frames
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "video", "video": sampled_frames},
+                    {"type": "text", "text": prompt},
+                ],
+            }
+        ]
+
+        # Prepare inputs
+        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+        image_inputs, video_inputs, video_kwargs = process_vision_info(
+            messages, return_video_kwargs=True, return_video_metadata=True
+        )
+
+        # split the videos and according metadatas
+        if video_inputs is not None:
+            videos, video_metadatas = zip(*video_inputs)
+            videos, video_metadatas = list(videos), list(video_metadatas)
+        else:
+            video_metadatas = None
+        inputs = self.processor(
+            text=[text],
+            images=image_inputs,
+            videos=videos,
+            video_metadata=video_metadatas,
+            padding=True,
+            return_tensors="pt",
+            **video_kwargs,
+        )
+
+        # Move inputs to device
+        inputs = inputs.to(self.model.device)
+
+        # Generate response
+        with torch.no_grad():
+            generated_ids = self.model.generate(
+                **inputs,
+                max_new_tokens=256,
+                do_sample=False,  # Use greedy decoding for consistency
+            )
+
+        # Trim input tokens
+        generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
+
+        response = self.processor.batch_decode(
+            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )[0]
+
+        # Parse response
+        return self._parse_response(response)
+
+    def _parse_response(self, response: str) -> dict[str, Any]:
+        """Parse the model's response into structured output."""
+        if self.verbose:
+            print(f"\n[DEBUG] Raw model response:\n{response}\n")
+
+        lines = response.strip().split("\n")
+
+        prediction = "unknown"
+        confidence = 0.5
+        reasoning = response  # Default to full response if parsing fails
+
+        for line in lines:
+            line_lower = line.lower().strip()
+
+            if line_lower.startswith("decision:"):
+                parts = line_lower.split("decision:", 1)
+                if len(parts) > 1:
+                    decision_text = parts[1].strip()
+                    if "success" in decision_text:
+                        prediction = "success"
+                    elif "failure" in decision_text or "fail" in decision_text:
+                        prediction = "failure"
+
+            elif line_lower.startswith("confidence:"):
+                parts = line_lower.split("confidence:", 1)
+                if len(parts) > 1:
+                    conf_text = parts[1].strip()
+                    try:
+                        # Extract number from text
+                        conf_num = "".join(c for c in conf_text if c.isdigit() or c == ".")
+                        if conf_num:
+                            confidence = float(conf_num)
+                            if confidence > 1.0:
+                                confidence = confidence / 100.0  # Handle percentage
+                    except (ValueError, IndexError):
+                        confidence = 0.5
+
+            elif line_lower.startswith("reasoning:"):
+                parts = line.split("reasoning:", 1)
+                if len(parts) > 1:
+                    reasoning = parts[1].strip()
+
+        # Fallback: if no structured response, try to infer from content
+        if prediction == "unknown":
+            if self.verbose:
+                print("[DEBUG] Failed to parse structured response, using fallback inference")
+            response_lower = response.lower()
+            if "success" in response_lower and "fail" not in response_lower:
+                prediction = "success"
+                confidence = 0.6
+            elif "fail" in response_lower or "failure" in response_lower:
+                prediction = "failure"
+                confidence = 0.6
+            else:
+                if self.verbose:
+                    print(f"[DEBUG] Could not determine prediction from response")
+
+        if self.verbose:
+            print(f"[DEBUG] Parsed: prediction={prediction}, confidence={confidence}\n")
+
+        return {
+            "prediction": prediction,
+            "confidence": confidence,
+            "reasoning": reasoning,
+            "raw_response": response,
+        }
+
+
+def get_dataset_splits(dataset_path: str) -> dict[str, Any]:
+    """
+    Get TFDS dataset splits without loading episodes into memory.
+
+    Returns:
+        Dictionary mapping split names to TFDS datasets (iterators)
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    builder = tfds.builder_from_directory(root)
+
+    dataset_dict = {}
+    split_name = "success"
+    # Get episode count without loading all episodes
+    try:
+        ds = builder.as_dataset(split=split_name, shuffle_files=False)
+        dataset_dict[split_name] = ds
+
+        # Get episode count without loading all episodes
+        try:
+            info = builder.info.splits[split_name]
+            num_episodes = info.num_examples
+            print(f"Found {num_episodes} episodes in '{split_name}' split")
+        except:
+            print(f"Found '{split_name}' split (episode count unknown)")
+    except Exception as e:
+        print(f"Warning: Could not load '{split_name}' split: {e}")
+        dataset_dict[split_name] = None
+
+    return dataset_dict, builder
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate success/fail labels for SOAR dataset using Qwen2-VL")
+    parser.add_argument(
+        "--dataset_path",
+        type=str,
+        required=True,
+        help="Path to SOAR TFDS dataset directory (e.g., /path/to/soar/rlds)",
+    )
+    parser.add_argument(
+        "--output",
+        type=str,
+        default="soar_vlm_labels.json",
+        help="Output JSON file path (default: soar_vlm_labels.json)",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="Qwen/Qwen3-VL-8B-Instruct-FP8",
+        choices=[
+            "Qwen/Qwen3-VL-4B-Instruct-FP8",
+            "Qwen/Qwen3-VL-8B-Instruct-FP8",
+            "Qwen/Qwen3-VL-32B-Instruct-FP8",
+        ],
+        help="Qwen3-VL model to use (default: 8B-FP8)",
+    )
+    parser.add_argument(
+        "--num_frames",
+        type=int,
+        default=8,
+        help="Number of frames to sample from each video (default: 8)",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="auto",
+        help="Device to run on: 'cuda', 'cpu', or 'auto' (default: auto)",
+    )
+    parser.add_argument(
+        "--max_episodes",
+        type=int,
+        default=None,
+        help="Maximum number of episodes to process per split (default: all)",
+    )
+    parser.add_argument(
+        "--checkpoint_interval",
+        type=int,
+        default=50,
+        help="Save checkpoint every N episodes (default: 50, 0 to disable)",
+    )
+    parser.add_argument(
+        "--resume_from",
+        type=str,
+        default=None,
+        help="Resume from checkpoint file (JSON with previous results)",
+    )
+    parser.add_argument(
+        "--verbose",
+        action="store_true",
+        help="Print debug information including raw model responses",
+    )
+
+    args = parser.parse_args()
+
+    # Load existing results if resuming
+    results = []
+    processed_episodes = set()
+    start_episode_id = 0
+
+    if args.resume_from and os.path.exists(args.resume_from):
+        print(f"\nResuming from checkpoint: {args.resume_from}")
+        with open(args.resume_from, "r") as f:
+            checkpoint_data = json.load(f)
+            results = checkpoint_data.get("results", [])
+            # Track which episodes we've already processed
+            for r in results:
+                processed_episodes.add((r["split_name"], r["episode_index"]))
+            start_episode_id = len(results)
+        print(f"Loaded {len(results)} previously processed episodes")
+
+    # Get dataset splits (streaming, no loading into memory)
+    print(f"\nAccessing SOAR dataset from: {args.dataset_path}")
+    dataset_dict, builder = get_dataset_splits(args.dataset_path)
+
+    if all(ds is None for ds in dataset_dict.values()):
+        print("Error: No splits found in dataset!")
+        return
+
+    # Initialize classifier
+    classifier = Qwen3VLClassifier(model_name=args.model, device=args.device, verbose=args.verbose)
+
+    # Process episodes in streaming fashion
+    episode_id = start_episode_id
+
+    def save_checkpoint(output_path: str, results: list, metadata: dict):
+        """Save intermediate checkpoint."""
+        checkpoint_data = {
+            "metadata": metadata,
+            "results": results,
+        }
+        checkpoint_path = output_path.replace(".json", "_checkpoint.json")
+        with open(checkpoint_path, "w") as f:
+            json.dump(checkpoint_data, f, indent=2)
+        print(f"  [Checkpoint saved: {len(results)} episodes]")
+
+    for split_name, dataset in dataset_dict.items():
+        if dataset is None:
+            continue
+
+        print(f"\nProcessing '{split_name}' split...")
+
+        ep_count = 0
+        agreements = 0  # Track agreements in this split
+
+        # Create tqdm with postfix for live stats
+        pbar = tqdm(dataset, desc=split_name)
+        for ep_idx, episode in enumerate(pbar):
+            # Check if we've already processed this episode
+            if (split_name, ep_idx) in processed_episodes:
+                continue
+
+            # Check max episodes limit
+            if args.max_episodes and ep_count >= args.max_episodes:
+                break
+
+            try:
+                # Convert episode to numpy (streaming, only this episode in memory)
+                steps_np = list(tfds.as_numpy(episode["steps"]))
+            except Exception as e:
+                print(f"\nWarning: Failed to load episode {ep_idx} in {split_name}: {e}")
+                continue
+
+            # Extract frames
+            frames = extract_episode_frames(steps_np)
+            if not frames:
+                print(f"\nWarning: No frames found for episode {ep_idx} in {split_name}")
+                continue
+
+            # Get task instruction
+            task = get_task_instruction(steps_np)
+            if not task:
+                print(f"\nWarning: No task instruction for episode {ep_idx} in {split_name}")
+                task = "Complete the manipulation task"
+
+            # Convert to PIL images
+            pil_frames = frames_to_pil_images(frames)
+
+            # Classify with VLM
+            try:
+                classification = classifier.classify_episode(pil_frames, task, num_frames=args.num_frames)
+
+                # Store result
+                original_label = "success" if split_name == "success" else "failure"
+                predicted_label = classification["prediction"]
+
+                result = {
+                    "episode_id": episode_id,
+                    "split_name": split_name,
+                    "episode_index": ep_idx,
+                    "task": task,
+                    "num_frames": len(frames),
+                    "predicted_label": predicted_label,
+                    "confidence": classification["confidence"],
+                    "reasoning": classification["reasoning"],
+                    "original_label": original_label,
+                }
+
+                results.append(result)
+                episode_id += 1
+                ep_count += 1
+
+                # Track agreement
+                if predicted_label == original_label:
+                    agreements += 1
+
+                # Update progress bar with live stats
+                agreement_pct = (agreements / ep_count * 100) if ep_count > 0 else 0
+                avg_conf = sum(r["confidence"] for r in results[-ep_count:]) / ep_count if ep_count > 0 else 0
+                pbar.set_postfix({"agree": f"{agreement_pct:.1f}%", "conf": f"{avg_conf:.2f}"})
+
+                # Save checkpoint periodically
+                if args.checkpoint_interval > 0 and episode_id % args.checkpoint_interval == 0:
+                    save_checkpoint(
+                        args.output,
+                        results,
+                        {
+                            "dataset_path": args.dataset_path,
+                            "model": args.model,
+                            "num_frames": args.num_frames,
+                            "total_episodes": len(results),
+                            "checkpoint": True,
+                        },
+                    )
+
+            except Exception as e:
+                print(f"\nError processing episode {ep_idx} in {split_name}: {e}")
+                import traceback
+
+                traceback.print_exc()
+                continue
+
+            # Aggressively free memory after each episode
+            del frames
+            del pil_frames
+            del steps_np
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+        # Close progress bar for this split
+        pbar.close()
+
+    # Save results
+    print(f"\n\nSaving {len(results)} results to {args.output}")
+
+    output_data = {
+        "metadata": {
+            "dataset_path": args.dataset_path,
+            "model": args.model,
+            "num_frames": args.num_frames,
+            "total_episodes": len(results),
+        },
+        "results": results,
+    }
+
+    with open(args.output, "w") as f:
+        json.dump(output_data, f, indent=2)
+
+    # Print summary statistics
+    print("\n" + "=" * 60)
+    print("SUMMARY")
+    print("=" * 60)
+
+    if results:
+        correct = sum(1 for r in results if r["predicted_label"] == r["original_label"])
+        total = len(results)
+        accuracy = correct / total * 100
+
+        print(f"Total episodes processed: {total}")
+        print(f"Agreement with original labels: {correct}/{total} ({accuracy:.1f}%)")
+
+        avg_confidence = sum(r["confidence"] for r in results) / total
+        print(f"Average confidence: {avg_confidence:.3f}")
+
+        # Per-split breakdown
+        for split_name in ["success", "failure"]:
+            split_results = [r for r in results if r["split_name"] == split_name]
+            if split_results:
+                split_correct = sum(1 for r in split_results if r["predicted_label"] == r["original_label"])
+                split_total = len(split_results)
+                split_acc = split_correct / split_total * 100
+                print(f"\n  {split_name.capitalize()} split:")
+                print(f"    Agreement: {split_correct}/{split_total} ({split_acc:.1f}%)")
+
+    print("\n" + "=" * 60)
+    print(f"Results saved to: {args.output}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/dataset_upload/dataset_helpers/oxe_helper.py b/robometer/dataset_upload/dataset_helpers/oxe_helper.py
new file mode 100755
index 0000000000000000000000000000000000000000..9de37c4d735ef7f6541510ab324ac455a3fe5131
--- /dev/null
+++ b/robometer/dataset_upload/dataset_helpers/oxe_helper.py
@@ -0,0 +1,815 @@
+"""
+Taken from: https://github.com/jesbu1/openvla/blob/main/prismatic/vla/datasets/rlds/oxe/configs.py
+configs.py
+
+Modified some configs to get rid of bad image obs keys and add bridge_V2 and such.
+
+Defines per-dataset configuration (kwargs) for each dataset in Open-X Embodiment.
+
+Configuration adopts the following structure:
+    image_obs_keys:
+        primary: primary external RGB
+        secondary: secondary external RGB
+        wrist: wrist RGB
+
+    depth_obs_keys:
+        primary: primary external depth
+        secondary: secondary external depth
+        wrist: wrist depth
+
+    # Always 8-dim =>> changes based on `StateEncoding`
+    state_obs_keys:
+        StateEncoding.POS_EULER:    EEF XYZ (3) + Roll-Pitch-Yaw (3) + <PAD> (1) + Gripper Open/Close (1)
+        StateEncoding.POS_QUAT:     EEF XYZ (3) + Quaternion (4) + Gripper Open/Close (1)
+        StateEncoding.JOINT:        Joint Angles (7, <PAD> if fewer) + Gripper Open/Close (1)
+
+    state_encoding: Type of `StateEncoding`
+    action_encoding: Type of action encoding (e.g., EEF Position vs. Joint Position)
+"""
+
+from enum import IntEnum
+
+# commented out for now as we don't have this
+# from prismatic.vla.datasets.rlds.oxe.utils.droid_utils import zero_action_filter
+
+
+# Defines Proprioceptive State Encoding Schemes
+class StateEncoding(IntEnum):
+    # fmt: off
+    NONE = -1               # No Proprioceptive State
+    POS_EULER = 1           # EEF XYZ (3) + Roll-Pitch-Yaw (3) + <PAD> (1) + Gripper Open/Close (1)
+    POS_QUAT = 2            # EEF XYZ (3) + Quaternion (4) + Gripper Open/Close (1)
+    JOINT = 3               # Joint Angles (7, <PAD> if fewer) + Gripper Open/Close (1)
+    JOINT_BIMANUAL = 4      # Joint Angles (2 x [ Joint Angles (6) + Gripper Open/Close (1) ])
+    # fmt: on
+
+
+# Defines Action Encoding Schemes
+class ActionEncoding(IntEnum):
+    # fmt: off
+    EEF_POS = 1             # EEF Delta XYZ (3) + Roll-Pitch-Yaw (3) + Gripper Open/Close (1)
+    JOINT_POS = 2           # Joint Delta Position (7) + Gripper Open/Close (1)
+    JOINT_POS_BIMANUAL = 3  # Joint Delta Position (2 x [ Joint Delta Position (6) + Gripper Open/Close (1) ])
+    EEF_R6 = 4              # EEF Delta XYZ (3) + R6 (6) + Gripper Open/Close (1)
+    # fmt: on
+
+
+# === Individual Dataset Configs ===
+OXE_DATASET_CONFIGS = {
+    "rl_bench_v1": {  #
+        "image_obs_keys": {"primary": "image", "secondary": "image_1", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "fractal20220817_data": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["base_pose_tool_reached", "gripper_closed"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "kuka": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [
+            "clip_function_input/base_pose_tool_reached",
+            "gripper_closed",
+        ],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bridge": {  # Version of Bridge V2 in Open X-Embodiment mixture
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": None,
+        },  # no secondary for now
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bridge_oxe": {  # Version of Bridge V2 in Open X-Embodiment mixture
+        "image_obs_keys": {"primary": "image", "secondary": "image_1", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bridge_v2": {  # Version of Bridge V2 I downloaded
+        "image_obs_keys": {
+            "primary": "image_0",
+            "secondary": "image_1",
+            "tertiary": "image_2",
+            "wrist": None,
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bridge_orig": {  # Original version of Bridge V2 from project website
+        "image_obs_keys": {"primary": "image_0", "secondary": "image_1", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bridge_dataset": {  # Original version of Bridge V2 from project website
+        "image_obs_keys": {"primary": "image_0", "secondary": "image_1", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "taco_play": {
+        "image_obs_keys": {
+            "primary": "rgb_static",
+            "secondary": None,
+            "wrist": "rgb_gripper",
+        },
+        "depth_obs_keys": {
+            "primary": "depth_static",
+            "secondary": None,
+            "wrist": "depth_gripper",
+        },
+        "state_obs_keys": ["state_eef", None, "state_gripper"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "jaco_play": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "image_wrist",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state_eef", None, "state_gripper"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_cable_routing": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist45_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["robot_state", None],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "roboturk": {
+        "image_obs_keys": {"primary": "front_rgb", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, None],
+        "state_encoding": StateEncoding.NONE,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "nyu_door_opening_surprising_effectiveness": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, None],
+        "state_encoding": StateEncoding.NONE,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "viola": {
+        "image_obs_keys": {
+            "primary": "agentview_rgb",
+            "secondary": None,
+            "wrist": "eye_in_hand_rgb",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["joint_states", "gripper_states"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_autolab_ur5": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "hand_image",
+        },
+        "depth_obs_keys": {"primary": "depth", "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "toto": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "language_table": {
+        "image_obs_keys": {"primary": "rgb", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["effector_translation", None, None, None, None, None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "columbia_cairlab_pusht_real": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["robot_state", None, None, None, None, None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "stanford_kuka_multimodal_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": "depth_image", "secondary": None, "wrist": None},
+        "state_obs_keys": ["ee_position", "ee_orientation", None],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "nyu_rot_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "stanford_hydra_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "austin_buds_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "nyu_franka_play_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": "image_additional_view",
+            "wrist": None,
+        },
+        "depth_obs_keys": {
+            "primary": "depth",
+            "secondary": "depth_additional_view",
+            "wrist": None,
+        },
+        "state_obs_keys": ["eef_state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "maniskill_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {
+            "primary": "depth",
+            "secondary": None,
+            "wrist": "wrist_depth",
+        },
+        "state_obs_keys": ["tcp_pose", "gripper_state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "furniture_bench_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "cmu_franka_exploration_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "highres_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, None],
+        "state_encoding": StateEncoding.NONE,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "ucsd_kitchen_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["joint_state", None],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "ucsd_pick_and_place_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "austin_sailor_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "austin_sirius_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "bc_z": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [
+            "present/xyz",
+            "present/axis_angle",
+            None,
+            "present/sensed_close",
+        ],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "utokyo_pr2_opening_fridge_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "utokyo_xarm_pick_and_place_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": "image2",
+            "wrist": "hand_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["end_effector_pose", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "utokyo_xarm_bimanual_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["pose_r", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "robo_net": {
+        "image_obs_keys": {"primary": "image", "secondary": "image1", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_mvp_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "hand_image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["pose", "gripper"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "berkeley_rpt_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "hand_image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["joint_pos", "gripper"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "kaist_nonprehensile_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "stanford_mask_vit_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tokyo_u_lsmo_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "dlr_sara_pour_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "dlr_sara_grid_clamp_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "dlr_edan_shared_control_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "asu_table_top_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "stanford_robocook_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image_1", "secondary": "image_2", "wrist": None},
+        "depth_obs_keys": {"primary": "depth_1", "secondary": "depth_2", "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, "state"],
+        "state_encoding": StateEncoding.NONE,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "iamlab_cmu_pickup_insert_converted_externally_to_rlds": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["joint_state", "gripper_state"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "uiuc_d3field": {
+        "image_obs_keys": {"primary": "image_1", "secondary": "image_2", "wrist": None},
+        "depth_obs_keys": {"primary": "depth_1", "secondary": "depth_2", "wrist": None},
+        "state_obs_keys": [None, None, None, None, None, None, None, None],
+        "state_encoding": StateEncoding.NONE,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "utaustin_mutex": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_fanuc_manipulation": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["joint_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "cmu_playing_with_food": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "finger_vision_1",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "cmu_play_fusion": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "cmu_stretch": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["eef_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_gnm_recon": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_gnm_cory_hall": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_gnm_sac_son": {
+        "image_obs_keys": {"primary": None, "secondary": None, "wrist": "image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state", None, None],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "droid": {
+        "image_obs_keys": {
+            "primary": "exterior_image_1_left",
+            "secondary": "exterior_image_2_left",
+            "wrist": "wrist_image_left",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.EEF_POS,
+        "aux_kwargs": {
+            "dataset_frame_transform_kwargs": {
+                # "chunk_filter_fn": zero_action_filter,
+            },
+        },
+    },
+    "fmb": {
+        "image_obs_keys": {
+            "primary": "image_side_1",
+            "secondary": "image_side_2",
+            "wrist": "image_wrist_1",
+        },
+        "depth_obs_keys": {
+            "primary": "image_side_1_depth",
+            "secondary": "image_side_2_depth",
+            "wrist": "image_wrist_1_depth",
+        },
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "dobbe": {
+        "image_obs_keys": {"primary": "wrist_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "roboset": {
+        "image_obs_keys": {
+            "primary": "image_left",
+            "secondary": "image_right",
+            "wrist": "image_wrist",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.JOINT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "rh20t": {
+        "image_obs_keys": {
+            "primary": "image_front",
+            "secondary": "image_side_right",
+            "wrist": "image_wrist",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    ### T-DROID datasets
+    "tdroid_carrot_in_bowl": {  # "put carrot in bowl" task, 50 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tdroid_pour_corn_in_pot": {  # "pour corn from red bowl into steel pot" task, 50 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tdroid_flip_pot_upright": {  # "flip pot upright" task, 10 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tdroid_move_object_onto_plate": {  # "move <object> onto plate" task, 150 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tdroid_knock_object_over": {  # "knock <object> over" task, 70 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tdroid_cover_object_with_towel": {  # "cover <object> with towel" task, 45 demos @ 5 Hz control
+        "image_obs_keys": {"primary": "static_image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {
+            "primary": "static_depth_image",
+            "secondary": None,
+            "wrist": None,
+        },
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    ### DROID Finetuning datasets
+    "droid_wipe": {
+        "image_obs_keys": {
+            "primary": "exterior_image_2_left",
+            "secondary": None,
+            "wrist": "wrist_image_left",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["proprio"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    ### LIBERO datasets (modified versions)
+    "libero_spatial_no_noops": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "libero_object_no_noops": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "libero_goal_no_noops": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "libero_10_no_noops": {
+        "image_obs_keys": {
+            "primary": "image",
+            "secondary": None,
+            "wrist": "wrist_image",
+        },
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["EEF_state", None, "gripper_state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    ### New datasets. Only image keys are somewhat correct, ignore rest
+    "robo_set": {
+        "image_obs_keys": {"primary": "image_right", "secondary": "image_left", "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "aloha_mobile": {
+        "image_obs_keys": {"primary": "cam_high", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "tidybot": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": "wrist_image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "kaist_nonprehensile_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_EULER,
+        "action_encoding": ActionEncoding.EEF_POS,
+    },
+    "berkeley_mvp_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "hand_image", "secondary": None, "wrist": "hand_image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "berkeley_rpt_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "hand_image", "secondary": None, "wrist": "hand_image"},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "nyu_rot_dataset_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+    "tokyo_u_lsmo_converted_externally_to_rlds": {
+        "image_obs_keys": {"primary": "image", "secondary": None, "wrist": None},
+        "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
+        "state_obs_keys": ["state"],
+        "state_encoding": StateEncoding.POS_QUAT,
+        "action_encoding": ActionEncoding.JOINT_POS,
+    },
+}
diff --git a/robometer/dataset_upload/dataset_helpers/soar_label_corrections_full.json b/robometer/dataset_upload/dataset_helpers/soar_label_corrections_full.json
new file mode 100644
index 0000000000000000000000000000000000000000..2315788a423944d861f48c585db5ba90fe96b590
--- /dev/null
+++ b/robometer/dataset_upload/dataset_helpers/soar_label_corrections_full.json
@@ -0,0 +1,30609 @@
+{
+  "source_files": [
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_0_1000_1766780289.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_0_1000_1766820230.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_1000_10000_1766947200.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_1000_10000_1767035492.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_1000_2000_1766946297.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_12000_16000_1767165379.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_16000_20000_1767290542.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_16000_None_1767165356.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_20000_24000_1767290608.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_24000_28000_1767290609.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_28000_None_1767291536.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_28500_None_1767385806.jsonl",
+    "/mmfs1/home/jessezha/rselab_storage/soar_relabel_results/results_qwen_8000_12000_1767165031.jsonl"
+  ],
+  "stats": {
+    "successful": 6954,
+    "already_complete": 461,
+    "failure": 20943,
+    "infeasible": 2222,
+    "unknown_kept_original": 35
+  },
+  "total_trajectories": 30580,
+  "max_index": 30579,
+  "missing_count": 0,
+  "label_corrections": {
+    "0": "successful",
+    "1": "successful",
+    "2": "already_complete",
+    "3": "successful",
+    "4": "successful",
+    "5": "successful",
+    "6": "successful",
+    "7": "successful",
+    "8": "successful",
+    "9": "successful",
+    "10": "successful",
+    "11": "successful",
+    "12": "successful",
+    "13": "failure",
+    "14": "successful",
+    "15": "failure",
+    "16": "successful",
+    "17": "successful",
+    "18": "successful",
+    "19": "successful",
+    "20": "successful",
+    "21": "successful",
+    "22": "successful",
+    "23": "successful",
+    "24": "failure",
+    "25": "failure",
+    "26": "successful",
+    "27": "successful",
+    "28": "successful",
+    "29": "successful",
+    "30": "infeasible",
+    "31": "already_complete",
+    "32": "successful",
+    "33": "successful",
+    "34": "failure",
+    "35": "already_complete",
+    "36": "failure",
+    "37": "successful",
+    "38": "successful",
+    "39": "successful",
+    "40": "already_complete",
+    "41": "successful",
+    "42": "successful",
+    "43": "successful",
+    "44": "failure",
+    "45": "successful",
+    "46": "successful",
+    "47": "successful",
+    "48": "successful",
+    "49": "successful",
+    "50": "already_complete",
+    "51": "failure",
+    "52": "successful",
+    "53": "successful",
+    "54": "successful",
+    "55": "successful",
+    "56": "successful",
+    "57": "successful",
+    "58": "successful",
+    "59": "successful",
+    "60": "failure",
+    "61": "already_complete",
+    "62": "already_complete",
+    "63": "successful",
+    "64": "successful",
+    "65": "successful",
+    "66": "successful",
+    "67": "successful",
+    "68": "failure",
+    "69": "successful",
+    "70": "already_complete",
+    "71": "successful",
+    "72": "successful",
+    "73": "successful",
+    "74": "successful",
+    "75": "already_complete",
+    "76": "successful",
+    "77": "successful",
+    "78": "successful",
+    "79": "successful",
+    "80": "failure",
+    "81": "already_complete",
+    "82": "successful",
+    "83": "successful",
+    "84": "successful",
+    "85": "failure",
+    "86": "successful",
+    "87": "successful",
+    "88": "successful",
+    "89": "successful",
+    "90": "successful",
+    "91": "successful",
+    "92": "failure",
+    "93": "successful",
+    "94": "successful",
+    "95": "successful",
+    "96": "already_complete",
+    "97": "successful",
+    "98": "successful",
+    "99": "successful",
+    "100": "successful",
+    "101": "successful",
+    "102": "successful",
+    "103": "successful",
+    "104": "successful",
+    "105": "successful",
+    "106": "already_complete",
+    "107": "successful",
+    "108": "already_complete",
+    "109": "successful",
+    "110": "successful",
+    "111": "successful",
+    "112": "successful",
+    "113": "successful",
+    "114": "successful",
+    "115": "failure",
+    "116": "successful",
+    "117": "failure",
+    "118": "already_complete",
+    "119": "successful",
+    "120": "successful",
+    "121": "successful",
+    "122": "successful",
+    "123": "already_complete",
+    "124": "successful",
+    "125": "already_complete",
+    "126": "successful",
+    "127": "successful",
+    "128": "already_complete",
+    "129": "successful",
+    "130": "failure",
+    "131": "successful",
+    "132": "successful",
+    "133": "successful",
+    "134": "successful",
+    "135": "successful",
+    "136": "successful",
+    "137": "successful",
+    "138": "successful",
+    "139": "successful",
+    "140": "already_complete",
+    "141": "failure",
+    "142": "successful",
+    "143": "failure",
+    "144": "failure",
+    "145": "successful",
+    "146": "successful",
+    "147": "failure",
+    "148": "already_complete",
+    "149": "successful",
+    "150": "successful",
+    "151": "successful",
+    "152": "successful",
+    "153": "successful",
+    "154": "successful",
+    "155": "already_complete",
+    "156": "successful",
+    "157": "failure",
+    "158": "successful",
+    "159": "successful",
+    "160": "failure",
+    "161": "successful",
+    "162": "already_complete",
+    "163": "failure",
+    "164": "successful",
+    "165": "successful",
+    "166": "successful",
+    "167": "successful",
+    "168": "failure",
+    "169": "successful",
+    "170": "failure",
+    "171": "successful",
+    "172": "failure",
+    "173": "successful",
+    "174": "successful",
+    "175": "successful",
+    "176": "successful",
+    "177": "failure",
+    "178": "already_complete",
+    "179": "infeasible",
+    "180": "successful",
+    "181": "successful",
+    "182": "successful",
+    "183": "successful",
+    "184": "successful",
+    "185": "successful",
+    "186": "successful",
+    "187": "successful",
+    "188": "successful",
+    "189": "successful",
+    "190": "successful",
+    "191": "failure",
+    "192": "already_complete",
+    "193": "successful",
+    "194": "successful",
+    "195": "failure",
+    "196": "successful",
+    "197": "already_complete",
+    "198": "successful",
+    "199": "successful",
+    "200": "failure",
+    "201": "successful",
+    "202": "successful",
+    "203": "successful",
+    "204": "successful",
+    "205": "successful",
+    "206": "successful",
+    "207": "successful",
+    "208": "successful",
+    "209": "failure",
+    "210": "successful",
+    "211": "failure",
+    "212": "already_complete",
+    "213": "already_complete",
+    "214": "failure",
+    "215": "failure",
+    "216": "successful",
+    "217": "successful",
+    "218": "successful",
+    "219": "failure",
+    "220": "successful",
+    "221": "successful",
+    "222": "successful",
+    "223": "successful",
+    "224": "successful",
+    "225": "successful",
+    "226": "successful",
+    "227": "failure",
+    "228": "infeasible",
+    "229": "successful",
+    "230": "successful",
+    "231": "successful",
+    "232": "successful",
+    "233": "failure",
+    "234": "successful",
+    "235": "successful",
+    "236": "already_complete",
+    "237": "successful",
+    "238": "successful",
+    "239": "failure",
+    "240": "failure",
+    "241": "successful",
+    "242": "successful",
+    "243": "failure",
+    "244": "successful",
+    "245": "successful",
+    "246": "failure",
+    "247": "successful",
+    "248": "failure",
+    "249": "successful",
+    "250": "failure",
+    "251": "successful",
+    "252": "failure",
+    "253": "failure",
+    "254": "successful",
+    "255": "successful",
+    "256": "successful",
+    "257": "successful",
+    "258": "successful",
+    "259": "successful",
+    "260": "successful",
+    "261": "successful",
+    "262": "successful",
+    "263": "successful",
+    "264": "successful",
+    "265": "already_complete",
+    "266": "successful",
+    "267": "successful",
+    "268": "successful",
+    "269": "successful",
+    "270": "successful",
+    "271": "successful",
+    "272": "failure",
+    "273": "successful",
+    "274": "failure",
+    "275": "successful",
+    "276": "successful",
+    "277": "successful",
+    "278": "successful",
+    "279": "successful",
+    "280": "successful",
+    "281": "failure",
+    "282": "failure",
+    "283": "successful",
+    "284": "successful",
+    "285": "successful",
+    "286": "already_complete",
+    "287": "successful",
+    "288": "successful",
+    "289": "failure",
+    "290": "infeasible",
+    "291": "successful",
+    "292": "successful",
+    "293": "successful",
+    "294": "failure",
+    "295": "successful",
+    "296": "successful",
+    "297": "successful",
+    "298": "successful",
+    "299": "successful",
+    "300": "successful",
+    "301": "successful",
+    "302": "successful",
+    "303": "successful",
+    "304": "successful",
+    "305": "successful",
+    "306": "successful",
+    "307": "failure",
+    "308": "successful",
+    "309": "successful",
+    "310": "successful",
+    "311": "failure",
+    "312": "already_complete",
+    "313": "successful",
+    "314": "successful",
+    "315": "successful",
+    "316": "successful",
+    "317": "successful",
+    "318": "successful",
+    "319": "failure",
+    "320": "already_complete",
+    "321": "successful",
+    "322": "successful",
+    "323": "successful",
+    "324": "successful",
+    "325": "successful",
+    "326": "failure",
+    "327": "successful",
+    "328": "already_complete",
+    "329": "successful",
+    "330": "successful",
+    "331": "successful",
+    "332": "successful",
+    "333": "already_complete",
+    "334": "failure",
+    "335": "already_complete",
+    "336": "successful",
+    "337": "failure",
+    "338": "successful",
+    "339": "failure",
+    "340": "failure",
+    "341": "successful",
+    "342": "successful",
+    "343": "successful",
+    "344": "successful",
+    "345": "successful",
+    "346": "successful",
+    "347": "already_complete",
+    "348": "already_complete",
+    "349": "successful",
+    "350": "successful",
+    "351": "failure",
+    "352": "successful",
+    "353": "successful",
+    "354": "already_complete",
+    "355": "already_complete",
+    "356": "successful",
+    "357": "successful",
+    "358": "successful",
+    "359": "successful",
+    "360": "successful",
+    "361": "successful",
+    "362": "successful",
+    "363": "successful",
+    "364": "successful",
+    "365": "failure",
+    "366": "successful",
+    "367": "already_complete",
+    "368": "successful",
+    "369": "successful",
+    "370": "successful",
+    "371": "successful",
+    "372": "successful",
+    "373": "successful",
+    "374": "failure",
+    "375": "successful",
+    "376": "failure",
+    "377": "successful",
+    "378": "failure",
+    "379": "already_complete",
+    "380": "successful",
+    "381": "successful",
+    "382": "successful",
+    "383": "already_complete",
+    "384": "successful",
+    "385": "successful",
+    "386": "failure",
+    "387": "successful",
+    "388": "successful",
+    "389": "successful",
+    "390": "failure",
+    "391": "successful",
+    "392": "failure",
+    "393": "successful",
+    "394": "successful",
+    "395": "successful",
+    "396": "successful",
+    "397": "successful",
+    "398": "failure",
+    "399": "successful",
+    "400": "failure",
+    "401": "failure",
+    "402": "successful",
+    "403": "successful",
+    "404": "successful",
+    "405": "successful",
+    "406": "successful",
+    "407": "already_complete",
+    "408": "successful",
+    "409": "already_complete",
+    "410": "infeasible",
+    "411": "successful",
+    "412": "failure",
+    "413": "successful",
+    "414": "successful",
+    "415": "successful",
+    "416": "successful",
+    "417": "successful",
+    "418": "already_complete",
+    "419": "successful",
+    "420": "failure",
+    "421": "successful",
+    "422": "successful",
+    "423": "successful",
+    "424": "successful",
+    "425": "successful",
+    "426": "failure",
+    "427": "already_complete",
+    "428": "failure",
+    "429": "infeasible",
+    "430": "already_complete",
+    "431": "failure",
+    "432": "successful",
+    "433": "already_complete",
+    "434": "already_complete",
+    "435": "successful",
+    "436": "already_complete",
+    "437": "already_complete",
+    "438": "successful",
+    "439": "successful",
+    "440": "already_complete",
+    "441": "successful",
+    "442": "successful",
+    "443": "successful",
+    "444": "successful",
+    "445": "infeasible",
+    "446": "already_complete",
+    "447": "successful",
+    "448": "successful",
+    "449": "successful",
+    "450": "successful",
+    "451": "successful",
+    "452": "failure",
+    "453": "successful",
+    "454": "successful",
+    "455": "already_complete",
+    "456": "successful",
+    "457": "successful",
+    "458": "successful",
+    "459": "successful",
+    "460": "successful",
+    "461": "failure",
+    "462": "successful",
+    "463": "failure",
+    "464": "successful",
+    "465": "successful",
+    "466": "already_complete",
+    "467": "failure",
+    "468": "successful",
+    "469": "successful",
+    "470": "successful",
+    "471": "already_complete",
+    "472": "successful",
+    "473": "successful",
+    "474": "successful",
+    "475": "successful",
+    "476": "successful",
+    "477": "successful",
+    "478": "successful",
+    "479": "already_complete",
+    "480": "already_complete",
+    "481": "successful",
+    "482": "successful",
+    "483": "failure",
+    "484": "successful",
+    "485": "failure",
+    "486": "successful",
+    "487": "successful",
+    "488": "successful",
+    "489": "successful",
+    "490": "successful",
+    "491": "failure",
+    "492": "successful",
+    "493": "failure",
+    "494": "already_complete",
+    "495": "successful",
+    "496": "successful",
+    "497": "failure",
+    "498": "successful",
+    "499": "failure",
+    "500": "successful",
+    "501": "successful",
+    "502": "successful",
+    "503": "already_complete",
+    "504": "failure",
+    "505": "failure",
+    "506": "successful",
+    "507": "successful",
+    "508": "failure",
+    "509": "successful",
+    "510": "successful",
+    "511": "successful",
+    "512": "successful",
+    "513": "successful",
+    "514": "successful",
+    "515": "successful",
+    "516": "failure",
+    "517": "already_complete",
+    "518": "successful",
+    "519": "successful",
+    "520": "successful",
+    "521": "already_complete",
+    "522": "successful",
+    "523": "successful",
+    "524": "successful",
+    "525": "successful",
+    "526": "successful",
+    "527": "failure",
+    "528": "successful",
+    "529": "successful",
+    "530": "successful",
+    "531": "successful",
+    "532": "failure",
+    "533": "successful",
+    "534": "already_complete",
+    "535": "successful",
+    "536": "successful",
+    "537": "failure",
+    "538": "successful",
+    "539": "successful",
+    "540": "successful",
+    "541": "successful",
+    "542": "already_complete",
+    "543": "already_complete",
+    "544": "successful",
+    "545": "successful",
+    "546": "successful",
+    "547": "successful",
+    "548": "failure",
+    "549": "successful",
+    "550": "successful",
+    "551": "successful",
+    "552": "successful",
+    "553": "failure",
+    "554": "successful",
+    "555": "already_complete",
+    "556": "failure",
+    "557": "successful",
+    "558": "successful",
+    "559": "infeasible",
+    "560": "already_complete",
+    "561": "successful",
+    "562": "already_complete",
+    "563": "already_complete",
+    "564": "successful",
+    "565": "already_complete",
+    "566": "failure",
+    "567": "successful",
+    "568": "successful",
+    "569": "successful",
+    "570": "failure",
+    "571": "successful",
+    "572": "successful",
+    "573": "failure",
+    "574": "failure",
+    "575": "successful",
+    "576": "successful",
+    "577": "successful",
+    "578": "already_complete",
+    "579": "successful",
+    "580": "successful",
+    "581": "failure",
+    "582": "successful",
+    "583": "successful",
+    "584": "successful",
+    "585": "successful",
+    "586": "successful",
+    "587": "already_complete",
+    "588": "failure",
+    "589": "successful",
+    "590": "successful",
+    "591": "successful",
+    "592": "failure",
+    "593": "already_complete",
+    "594": "failure",
+    "595": "failure",
+    "596": "successful",
+    "597": "successful",
+    "598": "successful",
+    "599": "already_complete",
+    "600": "already_complete",
+    "601": "successful",
+    "602": "infeasible",
+    "603": "already_complete",
+    "604": "failure",
+    "605": "successful",
+    "606": "successful",
+    "607": "failure",
+    "608": "already_complete",
+    "609": "successful",
+    "610": "successful",
+    "611": "successful",
+    "612": "successful",
+    "613": "successful",
+    "614": "successful",
+    "615": "successful",
+    "616": "successful",
+    "617": "successful",
+    "618": "successful",
+    "619": "successful",
+    "620": "successful",
+    "621": "already_complete",
+    "622": "successful",
+    "623": "failure",
+    "624": "successful",
+    "625": "successful",
+    "626": "successful",
+    "627": "successful",
+    "628": "successful",
+    "629": "failure",
+    "630": "successful",
+    "631": "successful",
+    "632": "successful",
+    "633": "successful",
+    "634": "failure",
+    "635": "successful",
+    "636": "successful",
+    "637": "successful",
+    "638": "successful",
+    "639": "already_complete",
+    "640": "successful",
+    "641": "already_complete",
+    "642": "successful",
+    "643": "already_complete",
+    "644": "successful",
+    "645": "successful",
+    "646": "successful",
+    "647": "successful",
+    "648": "already_complete",
+    "649": "successful",
+    "650": "successful",
+    "651": "successful",
+    "652": "successful",
+    "653": "failure",
+    "654": "successful",
+    "655": "successful",
+    "656": "failure",
+    "657": "failure",
+    "658": "successful",
+    "659": "successful",
+    "660": "successful",
+    "661": "successful",
+    "662": "successful",
+    "663": "successful",
+    "664": "successful",
+    "665": "successful",
+    "666": "successful",
+    "667": "successful",
+    "668": "successful",
+    "669": "successful",
+    "670": "infeasible",
+    "671": "failure",
+    "672": "successful",
+    "673": "successful",
+    "674": "failure",
+    "675": "already_complete",
+    "676": "successful",
+    "677": "successful",
+    "678": "successful",
+    "679": "successful",
+    "680": "successful",
+    "681": "successful",
+    "682": "failure",
+    "683": "successful",
+    "684": "successful",
+    "685": "successful",
+    "686": "successful",
+    "687": "successful",
+    "688": "successful",
+    "689": "successful",
+    "690": "already_complete",
+    "691": "successful",
+    "692": "successful",
+    "693": "successful",
+    "694": "successful",
+    "695": "failure",
+    "696": "failure",
+    "697": "successful",
+    "698": "failure",
+    "699": "already_complete",
+    "700": "failure",
+    "701": "failure",
+    "702": "successful",
+    "703": "successful",
+    "704": "successful",
+    "705": "successful",
+    "706": "successful",
+    "707": "successful",
+    "708": "successful",
+    "709": "already_complete",
+    "710": "successful",
+    "711": "successful",
+    "712": "successful",
+    "713": "successful",
+    "714": "successful",
+    "715": "successful",
+    "716": "successful",
+    "717": "already_complete",
+    "718": "already_complete",
+    "719": "successful",
+    "720": "already_complete",
+    "721": "successful",
+    "722": "successful",
+    "723": "successful",
+    "724": "successful",
+    "725": "already_complete",
+    "726": "successful",
+    "727": "successful",
+    "728": "already_complete",
+    "729": "successful",
+    "730": "successful",
+    "731": "successful",
+    "732": "already_complete",
+    "733": "successful",
+    "734": "successful",
+    "735": "failure",
+    "736": "failure",
+    "737": "successful",
+    "738": "successful",
+    "739": "already_complete",
+    "740": "failure",
+    "741": "failure",
+    "742": "failure",
+    "743": "already_complete",
+    "744": "successful",
+    "745": "successful",
+    "746": "successful",
+    "747": "successful",
+    "748": "successful",
+    "749": "failure",
+    "750": "already_complete",
+    "751": "successful",
+    "752": "successful",
+    "753": "already_complete",
+    "754": "successful",
+    "755": "successful",
+    "756": "successful",
+    "757": "successful",
+    "758": "successful",
+    "759": "successful",
+    "760": "successful",
+    "761": "successful",
+    "762": "successful",
+    "763": "successful",
+    "764": "successful",
+    "765": "successful",
+    "766": "successful",
+    "767": "failure",
+    "768": "successful",
+    "769": "failure",
+    "770": "successful",
+    "771": "failure",
+    "772": "successful",
+    "773": "already_complete",
+    "774": "successful",
+    "775": "successful",
+    "776": "already_complete",
+    "777": "failure",
+    "778": "successful",
+    "779": "failure",
+    "780": "successful",
+    "781": "successful",
+    "782": "successful",
+    "783": "successful",
+    "784": "failure",
+    "785": "already_complete",
+    "786": "successful",
+    "787": "successful",
+    "788": "successful",
+    "789": "failure",
+    "790": "successful",
+    "791": "successful",
+    "792": "successful",
+    "793": "successful",
+    "794": "successful",
+    "795": "already_complete",
+    "796": "failure",
+    "797": "successful",
+    "798": "successful",
+    "799": "successful",
+    "800": "successful",
+    "801": "successful",
+    "802": "successful",
+    "803": "already_complete",
+    "804": "successful",
+    "805": "failure",
+    "806": "successful",
+    "807": "successful",
+    "808": "already_complete",
+    "809": "failure",
+    "810": "successful",
+    "811": "successful",
+    "812": "successful",
+    "813": "successful",
+    "814": "successful",
+    "815": "failure",
+    "816": "successful",
+    "817": "successful",
+    "818": "already_complete",
+    "819": "successful",
+    "820": "already_complete",
+    "821": "successful",
+    "822": "successful",
+    "823": "successful",
+    "824": "already_complete",
+    "825": "successful",
+    "826": "successful",
+    "827": "successful",
+    "828": "successful",
+    "829": "failure",
+    "830": "successful",
+    "831": "successful",
+    "832": "failure",
+    "833": "successful",
+    "834": "already_complete",
+    "835": "already_complete",
+    "836": "successful",
+    "837": "successful",
+    "838": "successful",
+    "839": "already_complete",
+    "840": "successful",
+    "841": "successful",
+    "842": "infeasible",
+    "843": "failure",
+    "844": "successful",
+    "845": "successful",
+    "846": "successful",
+    "847": "successful",
+    "848": "successful",
+    "849": "successful",
+    "850": "successful",
+    "851": "successful",
+    "852": "already_complete",
+    "853": "successful",
+    "854": "successful",
+    "855": "successful",
+    "856": "successful",
+    "857": "successful",
+    "858": "successful",
+    "859": "successful",
+    "860": "successful",
+    "861": "successful",
+    "862": "successful",
+    "863": "successful",
+    "864": "successful",
+    "865": "failure",
+    "866": "successful",
+    "867": "successful",
+    "868": "successful",
+    "869": "already_complete",
+    "870": "successful",
+    "871": "failure",
+    "872": "failure",
+    "873": "successful",
+    "874": "successful",
+    "875": "successful",
+    "876": "successful",
+    "877": "successful",
+    "878": "failure",
+    "879": "successful",
+    "880": "successful",
+    "881": "successful",
+    "882": "successful",
+    "883": "successful",
+    "884": "successful",
+    "885": "successful",
+    "886": "already_complete",
+    "887": "successful",
+    "888": "successful",
+    "889": "successful",
+    "890": "successful",
+    "891": "already_complete",
+    "892": "already_complete",
+    "893": "successful",
+    "894": "successful",
+    "895": "successful",
+    "896": "already_complete",
+    "897": "already_complete",
+    "898": "infeasible",
+    "899": "successful",
+    "900": "successful",
+    "901": "successful",
+    "902": "successful",
+    "903": "successful",
+    "904": "successful",
+    "905": "failure",
+    "906": "successful",
+    "907": "successful",
+    "908": "failure",
+    "909": "successful",
+    "910": "successful",
+    "911": "already_complete",
+    "912": "successful",
+    "913": "successful",
+    "914": "successful",
+    "915": "successful",
+    "916": "successful",
+    "917": "successful",
+    "918": "successful",
+    "919": "successful",
+    "920": "failure",
+    "921": "successful",
+    "922": "successful",
+    "923": "successful",
+    "924": "failure",
+    "925": "successful",
+    "926": "successful",
+    "927": "successful",
+    "928": "failure",
+    "929": "already_complete",
+    "930": "successful",
+    "931": "successful",
+    "932": "successful",
+    "933": "failure",
+    "934": "successful",
+    "935": "already_complete",
+    "936": "successful",
+    "937": "failure",
+    "938": "failure",
+    "939": "successful",
+    "940": "successful",
+    "941": "successful",
+    "942": "failure",
+    "943": "failure",
+    "944": "successful",
+    "945": "successful",
+    "946": "successful",
+    "947": "failure",
+    "948": "successful",
+    "949": "already_complete",
+    "950": "successful",
+    "951": "successful",
+    "952": "successful",
+    "953": "successful",
+    "954": "failure",
+    "955": "successful",
+    "956": "already_complete",
+    "957": "successful",
+    "958": "successful",
+    "959": "already_complete",
+    "960": "successful",
+    "961": "successful",
+    "962": "successful",
+    "963": "successful",
+    "964": "successful",
+    "965": "successful",
+    "966": "successful",
+    "967": "successful",
+    "968": "successful",
+    "969": "successful",
+    "970": "successful",
+    "971": "successful",
+    "972": "successful",
+    "973": "successful",
+    "974": "successful",
+    "975": "successful",
+    "976": "successful",
+    "977": "successful",
+    "978": "successful",
+    "979": "successful",
+    "980": "successful",
+    "981": "failure",
+    "982": "successful",
+    "983": "already_complete",
+    "984": "successful",
+    "985": "successful",
+    "986": "failure",
+    "987": "successful",
+    "988": "successful",
+    "989": "failure",
+    "990": "failure",
+    "991": "successful",
+    "992": "successful",
+    "993": "successful",
+    "994": "successful",
+    "995": "already_complete",
+    "996": "successful",
+    "997": "successful",
+    "998": "successful",
+    "999": "successful",
+    "1000": "failure",
+    "1001": "successful",
+    "1002": "failure",
+    "1003": "failure",
+    "1004": "failure",
+    "1005": "failure",
+    "1006": "failure",
+    "1007": "successful",
+    "1008": "failure",
+    "1009": "failure",
+    "1010": "successful",
+    "1011": "failure",
+    "1012": "successful",
+    "1013": "successful",
+    "1014": "successful",
+    "1015": "failure",
+    "1016": "already_complete",
+    "1017": "successful",
+    "1018": "successful",
+    "1019": "successful",
+    "1020": "successful",
+    "1021": "failure",
+    "1022": "successful",
+    "1023": "successful",
+    "1024": "failure",
+    "1025": "successful",
+    "1026": "failure",
+    "1027": "successful",
+    "1028": "successful",
+    "1029": "failure",
+    "1030": "successful",
+    "1031": "successful",
+    "1032": "failure",
+    "1033": "failure",
+    "1034": "successful",
+    "1035": "successful",
+    "1036": "failure",
+    "1037": "already_complete",
+    "1038": "failure",
+    "1039": "successful",
+    "1040": "successful",
+    "1041": "successful",
+    "1042": "failure",
+    "1043": "failure",
+    "1044": "failure",
+    "1045": "successful",
+    "1046": "failure",
+    "1047": "successful",
+    "1048": "failure",
+    "1049": "successful",
+    "1050": "successful",
+    "1051": "successful",
+    "1052": "failure",
+    "1053": "successful",
+    "1054": "successful",
+    "1055": "failure",
+    "1056": "successful",
+    "1057": "failure",
+    "1058": "successful",
+    "1059": "failure",
+    "1060": "successful",
+    "1061": "successful",
+    "1062": "failure",
+    "1063": "successful",
+    "1064": "successful",
+    "1065": "infeasible",
+    "1066": "failure",
+    "1067": "failure",
+    "1068": "successful",
+    "1069": "successful",
+    "1070": "failure",
+    "1071": "successful",
+    "1072": "failure",
+    "1073": "failure",
+    "1074": "successful",
+    "1075": "failure",
+    "1076": "successful",
+    "1077": "failure",
+    "1078": "successful",
+    "1079": "successful",
+    "1080": "failure",
+    "1081": "successful",
+    "1082": "successful",
+    "1083": "failure",
+    "1084": "successful",
+    "1085": "successful",
+    "1086": "successful",
+    "1087": "failure",
+    "1088": "failure",
+    "1089": "failure",
+    "1090": "failure",
+    "1091": "successful",
+    "1092": "successful",
+    "1093": "successful",
+    "1094": "successful",
+    "1095": "successful",
+    "1096": "successful",
+    "1097": "failure",
+    "1098": "already_complete",
+    "1099": "successful",
+    "1100": "failure",
+    "1101": "successful",
+    "1102": "successful",
+    "1103": "failure",
+    "1104": "successful",
+    "1105": "failure",
+    "1106": "failure",
+    "1107": "successful",
+    "1108": "failure",
+    "1109": "failure",
+    "1110": "failure",
+    "1111": "failure",
+    "1112": "failure",
+    "1113": "successful",
+    "1114": "successful",
+    "1115": "successful",
+    "1116": "successful",
+    "1117": "failure",
+    "1118": "successful",
+    "1119": "successful",
+    "1120": "successful",
+    "1121": "successful",
+    "1122": "failure",
+    "1123": "failure",
+    "1124": "failure",
+    "1125": "successful",
+    "1126": "successful",
+    "1127": "successful",
+    "1128": "successful",
+    "1129": "successful",
+    "1130": "successful",
+    "1131": "successful",
+    "1132": "failure",
+    "1133": "failure",
+    "1134": "failure",
+    "1135": "failure",
+    "1136": "successful",
+    "1137": "failure",
+    "1138": "failure",
+    "1139": "successful",
+    "1140": "failure",
+    "1141": "successful",
+    "1142": "failure",
+    "1143": "failure",
+    "1144": "failure",
+    "1145": "successful",
+    "1146": "failure",
+    "1147": "successful",
+    "1148": "failure",
+    "1149": "successful",
+    "1150": "infeasible",
+    "1151": "successful",
+    "1152": "successful",
+    "1153": "successful",
+    "1154": "successful",
+    "1155": "infeasible",
+    "1156": "failure",
+    "1157": "successful",
+    "1158": "successful",
+    "1159": "failure",
+    "1160": "failure",
+    "1161": "successful",
+    "1162": "failure",
+    "1163": "successful",
+    "1164": "failure",
+    "1165": "successful",
+    "1166": "failure",
+    "1167": "failure",
+    "1168": "successful",
+    "1169": "failure",
+    "1170": "successful",
+    "1171": "successful",
+    "1172": "successful",
+    "1173": "failure",
+    "1174": "successful",
+    "1175": "successful",
+    "1176": "failure",
+    "1177": "successful",
+    "1178": "failure",
+    "1179": "failure",
+    "1180": "failure",
+    "1181": "failure",
+    "1182": "successful",
+    "1183": "successful",
+    "1184": "successful",
+    "1185": "failure",
+    "1186": "failure",
+    "1187": "failure",
+    "1188": "successful",
+    "1189": "successful",
+    "1190": "failure",
+    "1191": "failure",
+    "1192": "successful",
+    "1193": "failure",
+    "1194": "successful",
+    "1195": "successful",
+    "1196": "failure",
+    "1197": "failure",
+    "1198": "infeasible",
+    "1199": "failure",
+    "1200": "failure",
+    "1201": "failure",
+    "1202": "failure",
+    "1203": "failure",
+    "1204": "failure",
+    "1205": "failure",
+    "1206": "successful",
+    "1207": "successful",
+    "1208": "successful",
+    "1209": "failure",
+    "1210": "failure",
+    "1211": "failure",
+    "1212": "failure",
+    "1213": "successful",
+    "1214": "successful",
+    "1215": "successful",
+    "1216": "failure",
+    "1217": "failure",
+    "1218": "failure",
+    "1219": "failure",
+    "1220": "successful",
+    "1221": "failure",
+    "1222": "successful",
+    "1223": "failure",
+    "1224": "successful",
+    "1225": "successful",
+    "1226": "failure",
+    "1227": "failure",
+    "1228": "successful",
+    "1229": "failure",
+    "1230": "failure",
+    "1231": "failure",
+    "1232": "failure",
+    "1233": "failure",
+    "1234": "successful",
+    "1235": "failure",
+    "1236": "failure",
+    "1237": "successful",
+    "1238": "successful",
+    "1239": "successful",
+    "1240": "successful",
+    "1241": "failure",
+    "1242": "successful",
+    "1243": "successful",
+    "1244": "successful",
+    "1245": "successful",
+    "1246": "successful",
+    "1247": "successful",
+    "1248": "successful",
+    "1249": "successful",
+    "1250": "failure",
+    "1251": "successful",
+    "1252": "successful",
+    "1253": "successful",
+    "1254": "failure",
+    "1255": "failure",
+    "1256": "failure",
+    "1257": "failure",
+    "1258": "successful",
+    "1259": "successful",
+    "1260": "failure",
+    "1261": "successful",
+    "1262": "failure",
+    "1263": "failure",
+    "1264": "failure",
+    "1265": "successful",
+    "1266": "failure",
+    "1267": "successful",
+    "1268": "successful",
+    "1269": "failure",
+    "1270": "failure",
+    "1271": "successful",
+    "1272": "failure",
+    "1273": "failure",
+    "1274": "failure",
+    "1275": "successful",
+    "1276": "successful",
+    "1277": "successful",
+    "1278": "failure",
+    "1279": "successful",
+    "1280": "successful",
+    "1281": "failure",
+    "1282": "failure",
+    "1283": "successful",
+    "1284": "successful",
+    "1285": "successful",
+    "1286": "successful",
+    "1287": "successful",
+    "1288": "successful",
+    "1289": "failure",
+    "1290": "failure",
+    "1291": "successful",
+    "1292": "infeasible",
+    "1293": "successful",
+    "1294": "successful",
+    "1295": "failure",
+    "1296": "successful",
+    "1297": "failure",
+    "1298": "already_complete",
+    "1299": "failure",
+    "1300": "failure",
+    "1301": "successful",
+    "1302": "successful",
+    "1303": "successful",
+    "1304": "successful",
+    "1305": "successful",
+    "1306": "successful",
+    "1307": "successful",
+    "1308": "failure",
+    "1309": "successful",
+    "1310": "failure",
+    "1311": "successful",
+    "1312": "successful",
+    "1313": "successful",
+    "1314": "successful",
+    "1315": "successful",
+    "1316": "failure",
+    "1317": "failure",
+    "1318": "successful",
+    "1319": "successful",
+    "1320": "failure",
+    "1321": "successful",
+    "1322": "failure",
+    "1323": "successful",
+    "1324": "failure",
+    "1325": "successful",
+    "1326": "already_complete",
+    "1327": "failure",
+    "1328": "failure",
+    "1329": "successful",
+    "1330": "successful",
+    "1331": "successful",
+    "1332": "successful",
+    "1333": "failure",
+    "1334": "failure",
+    "1335": "successful",
+    "1336": "successful",
+    "1337": "failure",
+    "1338": "successful",
+    "1339": "failure",
+    "1340": "successful",
+    "1341": "successful",
+    "1342": "failure",
+    "1343": "successful",
+    "1344": "successful",
+    "1345": "successful",
+    "1346": "failure",
+    "1347": "failure",
+    "1348": "failure",
+    "1349": "successful",
+    "1350": "successful",
+    "1351": "failure",
+    "1352": "failure",
+    "1353": "successful",
+    "1354": "failure",
+    "1355": "successful",
+    "1356": "failure",
+    "1357": "failure",
+    "1358": "successful",
+    "1359": "successful",
+    "1360": "successful",
+    "1361": "successful",
+    "1362": "failure",
+    "1363": "successful",
+    "1364": "failure",
+    "1365": "successful",
+    "1366": "failure",
+    "1367": "failure",
+    "1368": "successful",
+    "1369": "failure",
+    "1370": "failure",
+    "1371": "successful",
+    "1372": "successful",
+    "1373": "infeasible",
+    "1374": "successful",
+    "1375": "failure",
+    "1376": "failure",
+    "1377": "failure",
+    "1378": "failure",
+    "1379": "successful",
+    "1380": "failure",
+    "1381": "successful",
+    "1382": "successful",
+    "1383": "failure",
+    "1384": "failure",
+    "1385": "failure",
+    "1386": "failure",
+    "1387": "failure",
+    "1388": "successful",
+    "1389": "successful",
+    "1390": "failure",
+    "1391": "failure",
+    "1392": "successful",
+    "1393": "successful",
+    "1394": "failure",
+    "1395": "failure",
+    "1396": "successful",
+    "1397": "failure",
+    "1398": "successful",
+    "1399": "successful",
+    "1400": "failure",
+    "1401": "failure",
+    "1402": "failure",
+    "1403": "failure",
+    "1404": "successful",
+    "1405": "failure",
+    "1406": "successful",
+    "1407": "successful",
+    "1408": "successful",
+    "1409": "failure",
+    "1410": "failure",
+    "1411": "failure",
+    "1412": "failure",
+    "1413": "successful",
+    "1414": "successful",
+    "1415": "failure",
+    "1416": "successful",
+    "1417": "infeasible",
+    "1418": "already_complete",
+    "1419": "failure",
+    "1420": "successful",
+    "1421": "failure",
+    "1422": "successful",
+    "1423": "failure",
+    "1424": "successful",
+    "1425": "failure",
+    "1426": "successful",
+    "1427": "failure",
+    "1428": "failure",
+    "1429": "successful",
+    "1430": "failure",
+    "1431": "failure",
+    "1432": "successful",
+    "1433": "failure",
+    "1434": "failure",
+    "1435": "successful",
+    "1436": "successful",
+    "1437": "failure",
+    "1438": "successful",
+    "1439": "failure",
+    "1440": "successful",
+    "1441": "successful",
+    "1442": "failure",
+    "1443": "failure",
+    "1444": "successful",
+    "1445": "failure",
+    "1446": "failure",
+    "1447": "successful",
+    "1448": "successful",
+    "1449": "failure",
+    "1450": "successful",
+    "1451": "failure",
+    "1452": "already_complete",
+    "1453": "failure",
+    "1454": "successful",
+    "1455": "successful",
+    "1456": "successful",
+    "1457": "failure",
+    "1458": "successful",
+    "1459": "failure",
+    "1460": "successful",
+    "1461": "successful",
+    "1462": "failure",
+    "1463": "failure",
+    "1464": "successful",
+    "1465": "successful",
+    "1466": "failure",
+    "1467": "successful",
+    "1468": "failure",
+    "1469": "failure",
+    "1470": "successful",
+    "1471": "successful",
+    "1472": "failure",
+    "1473": "infeasible",
+    "1474": "failure",
+    "1475": "failure",
+    "1476": "failure",
+    "1477": "successful",
+    "1478": "successful",
+    "1479": "failure",
+    "1480": "successful",
+    "1481": "successful",
+    "1482": "failure",
+    "1483": "successful",
+    "1484": "failure",
+    "1485": "failure",
+    "1486": "successful",
+    "1487": "successful",
+    "1488": "failure",
+    "1489": "failure",
+    "1490": "failure",
+    "1491": "failure",
+    "1492": "failure",
+    "1493": "failure",
+    "1494": "failure",
+    "1495": "successful",
+    "1496": "successful",
+    "1497": "failure",
+    "1498": "failure",
+    "1499": "failure",
+    "1500": "successful",
+    "1501": "failure",
+    "1502": "successful",
+    "1503": "successful",
+    "1504": "failure",
+    "1505": "successful",
+    "1506": "successful",
+    "1507": "successful",
+    "1508": "successful",
+    "1509": "failure",
+    "1510": "failure",
+    "1511": "failure",
+    "1512": "failure",
+    "1513": "failure",
+    "1514": "successful",
+    "1515": "successful",
+    "1516": "failure",
+    "1517": "failure",
+    "1518": "infeasible",
+    "1519": "failure",
+    "1520": "successful",
+    "1521": "failure",
+    "1522": "failure",
+    "1523": "failure",
+    "1524": "failure",
+    "1525": "failure",
+    "1526": "failure",
+    "1527": "failure",
+    "1528": "failure",
+    "1529": "failure",
+    "1530": "failure",
+    "1531": "successful",
+    "1532": "successful",
+    "1533": "failure",
+    "1534": "successful",
+    "1535": "failure",
+    "1536": "successful",
+    "1537": "failure",
+    "1538": "failure",
+    "1539": "successful",
+    "1540": "successful",
+    "1541": "failure",
+    "1542": "failure",
+    "1543": "successful",
+    "1544": "failure",
+    "1545": "failure",
+    "1546": "successful",
+    "1547": "successful",
+    "1548": "successful",
+    "1549": "failure",
+    "1550": "successful",
+    "1551": "failure",
+    "1552": "successful",
+    "1553": "failure",
+    "1554": "successful",
+    "1555": "successful",
+    "1556": "failure",
+    "1557": "successful",
+    "1558": "successful",
+    "1559": "failure",
+    "1560": "successful",
+    "1561": "failure",
+    "1562": "successful",
+    "1563": "failure",
+    "1564": "successful",
+    "1565": "failure",
+    "1566": "successful",
+    "1567": "failure",
+    "1568": "successful",
+    "1569": "successful",
+    "1570": "failure",
+    "1571": "successful",
+    "1572": "successful",
+    "1573": "successful",
+    "1574": "successful",
+    "1575": "failure",
+    "1576": "successful",
+    "1577": "successful",
+    "1578": "successful",
+    "1579": "failure",
+    "1580": "successful",
+    "1581": "successful",
+    "1582": "failure",
+    "1583": "failure",
+    "1584": "failure",
+    "1585": "failure",
+    "1586": "failure",
+    "1587": "failure",
+    "1588": "failure",
+    "1589": "successful",
+    "1590": "successful",
+    "1591": "infeasible",
+    "1592": "successful",
+    "1593": "failure",
+    "1594": "failure",
+    "1595": "successful",
+    "1596": "failure",
+    "1597": "successful",
+    "1598": "failure",
+    "1599": "successful",
+    "1600": "failure",
+    "1601": "successful",
+    "1602": "failure",
+    "1603": "successful",
+    "1604": "failure",
+    "1605": "failure",
+    "1606": "failure",
+    "1607": "failure",
+    "1608": "successful",
+    "1609": "failure",
+    "1610": "successful",
+    "1611": "failure",
+    "1612": "successful",
+    "1613": "infeasible",
+    "1614": "failure",
+    "1615": "failure",
+    "1616": "successful",
+    "1617": "failure",
+    "1618": "failure",
+    "1619": "successful",
+    "1620": "successful",
+    "1621": "successful",
+    "1622": "successful",
+    "1623": "successful",
+    "1624": "failure",
+    "1625": "successful",
+    "1626": "successful",
+    "1627": "failure",
+    "1628": "failure",
+    "1629": "failure",
+    "1630": "failure",
+    "1631": "failure",
+    "1632": "successful",
+    "1633": "successful",
+    "1634": "already_complete",
+    "1635": "failure",
+    "1636": "successful",
+    "1637": "successful",
+    "1638": "failure",
+    "1639": "failure",
+    "1640": "successful",
+    "1641": "successful",
+    "1642": "successful",
+    "1643": "successful",
+    "1644": "failure",
+    "1645": "successful",
+    "1646": "failure",
+    "1647": "failure",
+    "1648": "successful",
+    "1649": "failure",
+    "1650": "successful",
+    "1651": "failure",
+    "1652": "failure",
+    "1653": "failure",
+    "1654": "failure",
+    "1655": "failure",
+    "1656": "failure",
+    "1657": "successful",
+    "1658": "failure",
+    "1659": "failure",
+    "1660": "successful",
+    "1661": "successful",
+    "1662": "failure",
+    "1663": "successful",
+    "1664": "failure",
+    "1665": "successful",
+    "1666": "successful",
+    "1667": "failure",
+    "1668": "failure",
+    "1669": "failure",
+    "1670": "failure",
+    "1671": "successful",
+    "1672": "failure",
+    "1673": "successful",
+    "1674": "failure",
+    "1675": "successful",
+    "1676": "successful",
+    "1677": "successful",
+    "1678": "failure",
+    "1679": "infeasible",
+    "1680": "successful",
+    "1681": "successful",
+    "1682": "failure",
+    "1683": "failure",
+    "1684": "failure",
+    "1685": "failure",
+    "1686": "failure",
+    "1687": "successful",
+    "1688": "failure",
+    "1689": "failure",
+    "1690": "failure",
+    "1691": "successful",
+    "1692": "successful",
+    "1693": "failure",
+    "1694": "successful",
+    "1695": "successful",
+    "1696": "failure",
+    "1697": "successful",
+    "1698": "successful",
+    "1699": "failure",
+    "1700": "successful",
+    "1701": "failure",
+    "1702": "failure",
+    "1703": "successful",
+    "1704": "successful",
+    "1705": "infeasible",
+    "1706": "failure",
+    "1707": "failure",
+    "1708": "successful",
+    "1709": "failure",
+    "1710": "successful",
+    "1711": "successful",
+    "1712": "successful",
+    "1713": "successful",
+    "1714": "successful",
+    "1715": "successful",
+    "1716": "successful",
+    "1717": "failure",
+    "1718": "successful",
+    "1719": "failure",
+    "1720": "successful",
+    "1721": "successful",
+    "1722": "successful",
+    "1723": "failure",
+    "1724": "successful",
+    "1725": "failure",
+    "1726": "failure",
+    "1727": "successful",
+    "1728": "successful",
+    "1729": "successful",
+    "1730": "successful",
+    "1731": "successful",
+    "1732": "failure",
+    "1733": "successful",
+    "1734": "successful",
+    "1735": "successful",
+    "1736": "successful",
+    "1737": "successful",
+    "1738": "infeasible",
+    "1739": "successful",
+    "1740": "successful",
+    "1741": "failure",
+    "1742": "successful",
+    "1743": "successful",
+    "1744": "failure",
+    "1745": "successful",
+    "1746": "failure",
+    "1747": "failure",
+    "1748": "failure",
+    "1749": "failure",
+    "1750": "failure",
+    "1751": "failure",
+    "1752": "successful",
+    "1753": "successful",
+    "1754": "successful",
+    "1755": "failure",
+    "1756": "failure",
+    "1757": "failure",
+    "1758": "failure",
+    "1759": "failure",
+    "1760": "failure",
+    "1761": "failure",
+    "1762": "successful",
+    "1763": "failure",
+    "1764": "failure",
+    "1765": "failure",
+    "1766": "successful",
+    "1767": "successful",
+    "1768": "successful",
+    "1769": "failure",
+    "1770": "successful",
+    "1771": "successful",
+    "1772": "failure",
+    "1773": "successful",
+    "1774": "failure",
+    "1775": "failure",
+    "1776": "failure",
+    "1777": "successful",
+    "1778": "failure",
+    "1779": "successful",
+    "1780": "failure",
+    "1781": "successful",
+    "1782": "failure",
+    "1783": "successful",
+    "1784": "successful",
+    "1785": "failure",
+    "1786": "failure",
+    "1787": "successful",
+    "1788": "successful",
+    "1789": "failure",
+    "1790": "failure",
+    "1791": "successful",
+    "1792": "failure",
+    "1793": "failure",
+    "1794": "failure",
+    "1795": "failure",
+    "1796": "successful",
+    "1797": "failure",
+    "1798": "successful",
+    "1799": "successful",
+    "1800": "successful",
+    "1801": "failure",
+    "1802": "failure",
+    "1803": "successful",
+    "1804": "successful",
+    "1805": "successful",
+    "1806": "failure",
+    "1807": "successful",
+    "1808": "successful",
+    "1809": "successful",
+    "1810": "successful",
+    "1811": "successful",
+    "1812": "failure",
+    "1813": "failure",
+    "1814": "failure",
+    "1815": "successful",
+    "1816": "failure",
+    "1817": "infeasible",
+    "1818": "failure",
+    "1819": "successful",
+    "1820": "failure",
+    "1821": "successful",
+    "1822": "successful",
+    "1823": "failure",
+    "1824": "failure",
+    "1825": "successful",
+    "1826": "failure",
+    "1827": "successful",
+    "1828": "failure",
+    "1829": "successful",
+    "1830": "failure",
+    "1831": "failure",
+    "1832": "successful",
+    "1833": "successful",
+    "1834": "failure",
+    "1835": "successful",
+    "1836": "successful",
+    "1837": "infeasible",
+    "1838": "successful",
+    "1839": "successful",
+    "1840": "successful",
+    "1841": "failure",
+    "1842": "failure",
+    "1843": "successful",
+    "1844": "successful",
+    "1845": "failure",
+    "1846": "already_complete",
+    "1847": "successful",
+    "1848": "successful",
+    "1849": "failure",
+    "1850": "failure",
+    "1851": "successful",
+    "1852": "failure",
+    "1853": "successful",
+    "1854": "failure",
+    "1855": "already_complete",
+    "1856": "successful",
+    "1857": "successful",
+    "1858": "failure",
+    "1859": "failure",
+    "1860": "successful",
+    "1861": "successful",
+    "1862": "successful",
+    "1863": "failure",
+    "1864": "successful",
+    "1865": "successful",
+    "1866": "successful",
+    "1867": "failure",
+    "1868": "failure",
+    "1869": "successful",
+    "1870": "failure",
+    "1871": "failure",
+    "1872": "successful",
+    "1873": "successful",
+    "1874": "failure",
+    "1875": "failure",
+    "1876": "failure",
+    "1877": "failure",
+    "1878": "successful",
+    "1879": "failure",
+    "1880": "successful",
+    "1881": "failure",
+    "1882": "failure",
+    "1883": "failure",
+    "1884": "already_complete",
+    "1885": "successful",
+    "1886": "failure",
+    "1887": "failure",
+    "1888": "successful",
+    "1889": "failure",
+    "1890": "successful",
+    "1891": "failure",
+    "1892": "successful",
+    "1893": "successful",
+    "1894": "failure",
+    "1895": "successful",
+    "1896": "failure",
+    "1897": "failure",
+    "1898": "successful",
+    "1899": "failure",
+    "1900": "successful",
+    "1901": "failure",
+    "1902": "successful",
+    "1903": "successful",
+    "1904": "failure",
+    "1905": "failure",
+    "1906": "successful",
+    "1907": "successful",
+    "1908": "successful",
+    "1909": "already_complete",
+    "1910": "failure",
+    "1911": "successful",
+    "1912": "failure",
+    "1913": "successful",
+    "1914": "successful",
+    "1915": "failure",
+    "1916": "failure",
+    "1917": "infeasible",
+    "1918": "already_complete",
+    "1919": "failure",
+    "1920": "failure",
+    "1921": "successful",
+    "1922": "failure",
+    "1923": "failure",
+    "1924": "failure",
+    "1925": "failure",
+    "1926": "failure",
+    "1927": "successful",
+    "1928": "failure",
+    "1929": "failure",
+    "1930": "failure",
+    "1931": "successful",
+    "1932": "failure",
+    "1933": "successful",
+    "1934": "successful",
+    "1935": "failure",
+    "1936": "successful",
+    "1937": "failure",
+    "1938": "failure",
+    "1939": "failure",
+    "1940": "failure",
+    "1941": "successful",
+    "1942": "failure",
+    "1943": "failure",
+    "1944": "failure",
+    "1945": "successful",
+    "1946": "successful",
+    "1947": "failure",
+    "1948": "failure",
+    "1949": "successful",
+    "1950": "failure",
+    "1951": "failure",
+    "1952": "successful",
+    "1953": "successful",
+    "1954": "failure",
+    "1955": "failure",
+    "1956": "successful",
+    "1957": "successful",
+    "1958": "failure",
+    "1959": "failure",
+    "1960": "failure",
+    "1961": "failure",
+    "1962": "successful",
+    "1963": "failure",
+    "1964": "successful",
+    "1965": "successful",
+    "1966": "failure",
+    "1967": "failure",
+    "1968": "successful",
+    "1969": "successful",
+    "1970": "failure",
+    "1971": "successful",
+    "1972": "successful",
+    "1973": "failure",
+    "1974": "failure",
+    "1975": "successful",
+    "1976": "failure",
+    "1977": "failure",
+    "1978": "failure",
+    "1979": "failure",
+    "1980": "failure",
+    "1981": "successful",
+    "1982": "failure",
+    "1983": "failure",
+    "1984": "successful",
+    "1985": "failure",
+    "1986": "successful",
+    "1987": "failure",
+    "1988": "failure",
+    "1989": "failure",
+    "1990": "successful",
+    "1991": "failure",
+    "1992": "failure",
+    "1993": "failure",
+    "1994": "failure",
+    "1995": "failure",
+    "1996": "failure",
+    "1997": "successful",
+    "1998": "failure",
+    "1999": "successful",
+    "2000": "successful",
+    "2001": "successful",
+    "2002": "failure",
+    "2003": "failure",
+    "2004": "successful",
+    "2005": "failure",
+    "2006": "failure",
+    "2007": "successful",
+    "2008": "failure",
+    "2009": "successful",
+    "2010": "successful",
+    "2011": "successful",
+    "2012": "failure",
+    "2013": "failure",
+    "2014": "successful",
+    "2015": "failure",
+    "2016": "failure",
+    "2017": "failure",
+    "2018": "successful",
+    "2019": "successful",
+    "2020": "failure",
+    "2021": "failure",
+    "2022": "successful",
+    "2023": "failure",
+    "2024": "successful",
+    "2025": "failure",
+    "2026": "failure",
+    "2027": "infeasible",
+    "2028": "failure",
+    "2029": "failure",
+    "2030": "failure",
+    "2031": "failure",
+    "2032": "successful",
+    "2033": "successful",
+    "2034": "successful",
+    "2035": "failure",
+    "2036": "failure",
+    "2037": "failure",
+    "2038": "failure",
+    "2039": "successful",
+    "2040": "successful",
+    "2041": "successful",
+    "2042": "successful",
+    "2043": "failure",
+    "2044": "successful",
+    "2045": "successful",
+    "2046": "failure",
+    "2047": "failure",
+    "2048": "successful",
+    "2049": "successful",
+    "2050": "successful",
+    "2051": "failure",
+    "2052": "failure",
+    "2053": "failure",
+    "2054": "failure",
+    "2055": "failure",
+    "2056": "failure",
+    "2057": "failure",
+    "2058": "successful",
+    "2059": "failure",
+    "2060": "successful",
+    "2061": "successful",
+    "2062": "failure",
+    "2063": "successful",
+    "2064": "failure",
+    "2065": "successful",
+    "2066": "failure",
+    "2067": "successful",
+    "2068": "failure",
+    "2069": "failure",
+    "2070": "successful",
+    "2071": "failure",
+    "2072": "failure",
+    "2073": "successful",
+    "2074": "successful",
+    "2075": "successful",
+    "2076": "successful",
+    "2077": "successful",
+    "2078": "failure",
+    "2079": "failure",
+    "2080": "successful",
+    "2081": "failure",
+    "2082": "successful",
+    "2083": "already_complete",
+    "2084": "successful",
+    "2085": "successful",
+    "2086": "failure",
+    "2087": "successful",
+    "2088": "successful",
+    "2089": "failure",
+    "2090": "successful",
+    "2091": "failure",
+    "2092": "successful",
+    "2093": "failure",
+    "2094": "failure",
+    "2095": "failure",
+    "2096": "already_complete",
+    "2097": "failure",
+    "2098": "failure",
+    "2099": "failure",
+    "2100": "failure",
+    "2101": "successful",
+    "2102": "failure",
+    "2103": "successful",
+    "2104": "successful",
+    "2105": "successful",
+    "2106": "failure",
+    "2107": "failure",
+    "2108": "failure",
+    "2109": "successful",
+    "2110": "successful",
+    "2111": "successful",
+    "2112": "successful",
+    "2113": "failure",
+    "2114": "failure",
+    "2115": "failure",
+    "2116": "failure",
+    "2117": "successful",
+    "2118": "failure",
+    "2119": "successful",
+    "2120": "successful",
+    "2121": "successful",
+    "2122": "successful",
+    "2123": "failure",
+    "2124": "successful",
+    "2125": "successful",
+    "2126": "successful",
+    "2127": "failure",
+    "2128": "failure",
+    "2129": "successful",
+    "2130": "failure",
+    "2131": "failure",
+    "2132": "successful",
+    "2133": "successful",
+    "2134": "failure",
+    "2135": "successful",
+    "2136": "failure",
+    "2137": "failure",
+    "2138": "successful",
+    "2139": "failure",
+    "2140": "successful",
+    "2141": "failure",
+    "2142": "successful",
+    "2143": "failure",
+    "2144": "successful",
+    "2145": "failure",
+    "2146": "failure",
+    "2147": "successful",
+    "2148": "successful",
+    "2149": "failure",
+    "2150": "failure",
+    "2151": "successful",
+    "2152": "failure",
+    "2153": "failure",
+    "2154": "successful",
+    "2155": "failure",
+    "2156": "successful",
+    "2157": "successful",
+    "2158": "successful",
+    "2159": "failure",
+    "2160": "successful",
+    "2161": "successful",
+    "2162": "successful",
+    "2163": "failure",
+    "2164": "successful",
+    "2165": "failure",
+    "2166": "failure",
+    "2167": "successful",
+    "2168": "failure",
+    "2169": "failure",
+    "2170": "successful",
+    "2171": "failure",
+    "2172": "failure",
+    "2173": "failure",
+    "2174": "failure",
+    "2175": "failure",
+    "2176": "successful",
+    "2177": "failure",
+    "2178": "failure",
+    "2179": "successful",
+    "2180": "successful",
+    "2181": "successful",
+    "2182": "successful",
+    "2183": "successful",
+    "2184": "successful",
+    "2185": "failure",
+    "2186": "successful",
+    "2187": "successful",
+    "2188": "successful",
+    "2189": "successful",
+    "2190": "failure",
+    "2191": "successful",
+    "2192": "failure",
+    "2193": "failure",
+    "2194": "failure",
+    "2195": "successful",
+    "2196": "failure",
+    "2197": "successful",
+    "2198": "failure",
+    "2199": "failure",
+    "2200": "failure",
+    "2201": "failure",
+    "2202": "failure",
+    "2203": "successful",
+    "2204": "already_complete",
+    "2205": "failure",
+    "2206": "successful",
+    "2207": "failure",
+    "2208": "failure",
+    "2209": "successful",
+    "2210": "successful",
+    "2211": "failure",
+    "2212": "failure",
+    "2213": "failure",
+    "2214": "failure",
+    "2215": "successful",
+    "2216": "failure",
+    "2217": "successful",
+    "2218": "failure",
+    "2219": "successful",
+    "2220": "failure",
+    "2221": "failure",
+    "2222": "successful",
+    "2223": "failure",
+    "2224": "failure",
+    "2225": "successful",
+    "2226": "failure",
+    "2227": "failure",
+    "2228": "failure",
+    "2229": "failure",
+    "2230": "successful",
+    "2231": "failure",
+    "2232": "failure",
+    "2233": "successful",
+    "2234": "successful",
+    "2235": "failure",
+    "2236": "successful",
+    "2237": "failure",
+    "2238": "failure",
+    "2239": "successful",
+    "2240": "successful",
+    "2241": "successful",
+    "2242": "failure",
+    "2243": "failure",
+    "2244": "successful",
+    "2245": "successful",
+    "2246": "failure",
+    "2247": "failure",
+    "2248": "successful",
+    "2249": "successful",
+    "2250": "already_complete",
+    "2251": "successful",
+    "2252": "failure",
+    "2253": "failure",
+    "2254": "failure",
+    "2255": "successful",
+    "2256": "failure",
+    "2257": "failure",
+    "2258": "successful",
+    "2259": "failure",
+    "2260": "failure",
+    "2261": "successful",
+    "2262": "failure",
+    "2263": "failure",
+    "2264": "successful",
+    "2265": "successful",
+    "2266": "successful",
+    "2267": "failure",
+    "2268": "failure",
+    "2269": "failure",
+    "2270": "successful",
+    "2271": "successful",
+    "2272": "already_complete",
+    "2273": "failure",
+    "2274": "successful",
+    "2275": "successful",
+    "2276": "successful",
+    "2277": "failure",
+    "2278": "failure",
+    "2279": "successful",
+    "2280": "failure",
+    "2281": "failure",
+    "2282": "successful",
+    "2283": "failure",
+    "2284": "failure",
+    "2285": "failure",
+    "2286": "successful",
+    "2287": "failure",
+    "2288": "failure",
+    "2289": "successful",
+    "2290": "successful",
+    "2291": "failure",
+    "2292": "successful",
+    "2293": "failure",
+    "2294": "successful",
+    "2295": "successful",
+    "2296": "failure",
+    "2297": "failure",
+    "2298": "failure",
+    "2299": "successful",
+    "2300": "failure",
+    "2301": "infeasible",
+    "2302": "successful",
+    "2303": "successful",
+    "2304": "successful",
+    "2305": "failure",
+    "2306": "successful",
+    "2307": "successful",
+    "2308": "successful",
+    "2309": "successful",
+    "2310": "failure",
+    "2311": "successful",
+    "2312": "failure",
+    "2313": "failure",
+    "2314": "successful",
+    "2315": "failure",
+    "2316": "successful",
+    "2317": "failure",
+    "2318": "failure",
+    "2319": "successful",
+    "2320": "successful",
+    "2321": "failure",
+    "2322": "successful",
+    "2323": "successful",
+    "2324": "successful",
+    "2325": "successful",
+    "2326": "successful",
+    "2327": "successful",
+    "2328": "failure",
+    "2329": "successful",
+    "2330": "failure",
+    "2331": "failure",
+    "2332": "failure",
+    "2333": "failure",
+    "2334": "successful",
+    "2335": "successful",
+    "2336": "successful",
+    "2337": "successful",
+    "2338": "successful",
+    "2339": "successful",
+    "2340": "successful",
+    "2341": "failure",
+    "2342": "successful",
+    "2343": "failure",
+    "2344": "failure",
+    "2345": "failure",
+    "2346": "failure",
+    "2347": "successful",
+    "2348": "already_complete",
+    "2349": "failure",
+    "2350": "successful",
+    "2351": "failure",
+    "2352": "failure",
+    "2353": "successful",
+    "2354": "failure",
+    "2355": "failure",
+    "2356": "failure",
+    "2357": "failure",
+    "2358": "failure",
+    "2359": "failure",
+    "2360": "failure",
+    "2361": "already_complete",
+    "2362": "failure",
+    "2363": "successful",
+    "2364": "failure",
+    "2365": "successful",
+    "2366": "successful",
+    "2367": "successful",
+    "2368": "successful",
+    "2369": "failure",
+    "2370": "failure",
+    "2371": "successful",
+    "2372": "failure",
+    "2373": "successful",
+    "2374": "failure",
+    "2375": "failure",
+    "2376": "failure",
+    "2377": "failure",
+    "2378": "failure",
+    "2379": "failure",
+    "2380": "successful",
+    "2381": "successful",
+    "2382": "successful",
+    "2383": "successful",
+    "2384": "infeasible",
+    "2385": "successful",
+    "2386": "successful",
+    "2387": "already_complete",
+    "2388": "failure",
+    "2389": "failure",
+    "2390": "successful",
+    "2391": "failure",
+    "2392": "successful",
+    "2393": "infeasible",
+    "2394": "successful",
+    "2395": "already_complete",
+    "2396": "failure",
+    "2397": "failure",
+    "2398": "failure",
+    "2399": "successful",
+    "2400": "successful",
+    "2401": "failure",
+    "2402": "failure",
+    "2403": "failure",
+    "2404": "failure",
+    "2405": "failure",
+    "2406": "successful",
+    "2407": "failure",
+    "2408": "failure",
+    "2409": "failure",
+    "2410": "successful",
+    "2411": "failure",
+    "2412": "failure",
+    "2413": "failure",
+    "2414": "failure",
+    "2415": "successful",
+    "2416": "successful",
+    "2417": "failure",
+    "2418": "failure",
+    "2419": "successful",
+    "2420": "failure",
+    "2421": "failure",
+    "2422": "failure",
+    "2423": "successful",
+    "2424": "successful",
+    "2425": "failure",
+    "2426": "successful",
+    "2427": "failure",
+    "2428": "successful",
+    "2429": "successful",
+    "2430": "successful",
+    "2431": "successful",
+    "2432": "successful",
+    "2433": "successful",
+    "2434": "successful",
+    "2435": "failure",
+    "2436": "failure",
+    "2437": "failure",
+    "2438": "failure",
+    "2439": "successful",
+    "2440": "failure",
+    "2441": "successful",
+    "2442": "failure",
+    "2443": "successful",
+    "2444": "failure",
+    "2445": "failure",
+    "2446": "successful",
+    "2447": "failure",
+    "2448": "successful",
+    "2449": "failure",
+    "2450": "failure",
+    "2451": "successful",
+    "2452": "failure",
+    "2453": "failure",
+    "2454": "successful",
+    "2455": "failure",
+    "2456": "failure",
+    "2457": "successful",
+    "2458": "successful",
+    "2459": "successful",
+    "2460": "successful",
+    "2461": "infeasible",
+    "2462": "successful",
+    "2463": "failure",
+    "2464": "failure",
+    "2465": "successful",
+    "2466": "failure",
+    "2467": "failure",
+    "2468": "failure",
+    "2469": "successful",
+    "2470": "failure",
+    "2471": "failure",
+    "2472": "failure",
+    "2473": "failure",
+    "2474": "failure",
+    "2475": "failure",
+    "2476": "successful",
+    "2477": "failure",
+    "2478": "failure",
+    "2479": "failure",
+    "2480": "failure",
+    "2481": "successful",
+    "2482": "failure",
+    "2483": "failure",
+    "2484": "failure",
+    "2485": "successful",
+    "2486": "failure",
+    "2487": "successful",
+    "2488": "failure",
+    "2489": "successful",
+    "2490": "failure",
+    "2491": "successful",
+    "2492": "failure",
+    "2493": "successful",
+    "2494": "failure",
+    "2495": "failure",
+    "2496": "successful",
+    "2497": "successful",
+    "2498": "successful",
+    "2499": "successful",
+    "2500": "successful",
+    "2501": "successful",
+    "2502": "successful",
+    "2503": "failure",
+    "2504": "successful",
+    "2505": "failure",
+    "2506": "infeasible",
+    "2507": "infeasible",
+    "2508": "successful",
+    "2509": "failure",
+    "2510": "successful",
+    "2511": "failure",
+    "2512": "successful",
+    "2513": "failure",
+    "2514": "successful",
+    "2515": "failure",
+    "2516": "successful",
+    "2517": "failure",
+    "2518": "failure",
+    "2519": "successful",
+    "2520": "successful",
+    "2521": "successful",
+    "2522": "successful",
+    "2523": "failure",
+    "2524": "successful",
+    "2525": "failure",
+    "2526": "failure",
+    "2527": "successful",
+    "2528": "successful",
+    "2529": "failure",
+    "2530": "failure",
+    "2531": "failure",
+    "2532": "successful",
+    "2533": "failure",
+    "2534": "failure",
+    "2535": "successful",
+    "2536": "successful",
+    "2537": "successful",
+    "2538": "failure",
+    "2539": "failure",
+    "2540": "failure",
+    "2541": "failure",
+    "2542": "successful",
+    "2543": "successful",
+    "2544": "successful",
+    "2545": "failure",
+    "2546": "failure",
+    "2547": "successful",
+    "2548": "failure",
+    "2549": "failure",
+    "2550": "failure",
+    "2551": "successful",
+    "2552": "failure",
+    "2553": "successful",
+    "2554": "already_complete",
+    "2555": "successful",
+    "2556": "failure",
+    "2557": "already_complete",
+    "2558": "successful",
+    "2559": "already_complete",
+    "2560": "successful",
+    "2561": "failure",
+    "2562": "failure",
+    "2563": "successful",
+    "2564": "failure",
+    "2565": "successful",
+    "2566": "failure",
+    "2567": "failure",
+    "2568": "successful",
+    "2569": "successful",
+    "2570": "failure",
+    "2571": "successful",
+    "2572": "failure",
+    "2573": "failure",
+    "2574": "failure",
+    "2575": "failure",
+    "2576": "failure",
+    "2577": "failure",
+    "2578": "successful",
+    "2579": "failure",
+    "2580": "successful",
+    "2581": "failure",
+    "2582": "failure",
+    "2583": "successful",
+    "2584": "infeasible",
+    "2585": "successful",
+    "2586": "successful",
+    "2587": "successful",
+    "2588": "successful",
+    "2589": "successful",
+    "2590": "failure",
+    "2591": "successful",
+    "2592": "failure",
+    "2593": "successful",
+    "2594": "successful",
+    "2595": "failure",
+    "2596": "failure",
+    "2597": "successful",
+    "2598": "failure",
+    "2599": "successful",
+    "2600": "successful",
+    "2601": "failure",
+    "2602": "successful",
+    "2603": "failure",
+    "2604": "failure",
+    "2605": "successful",
+    "2606": "failure",
+    "2607": "successful",
+    "2608": "failure",
+    "2609": "failure",
+    "2610": "failure",
+    "2611": "failure",
+    "2612": "failure",
+    "2613": "successful",
+    "2614": "successful",
+    "2615": "successful",
+    "2616": "successful",
+    "2617": "successful",
+    "2618": "failure",
+    "2619": "failure",
+    "2620": "successful",
+    "2621": "failure",
+    "2622": "successful",
+    "2623": "successful",
+    "2624": "successful",
+    "2625": "failure",
+    "2626": "successful",
+    "2627": "successful",
+    "2628": "failure",
+    "2629": "failure",
+    "2630": "failure",
+    "2631": "infeasible",
+    "2632": "successful",
+    "2633": "successful",
+    "2634": "failure",
+    "2635": "failure",
+    "2636": "failure",
+    "2637": "successful",
+    "2638": "failure",
+    "2639": "failure",
+    "2640": "already_complete",
+    "2641": "successful",
+    "2642": "successful",
+    "2643": "successful",
+    "2644": "successful",
+    "2645": "failure",
+    "2646": "failure",
+    "2647": "failure",
+    "2648": "successful",
+    "2649": "successful",
+    "2650": "failure",
+    "2651": "successful",
+    "2652": "successful",
+    "2653": "successful",
+    "2654": "successful",
+    "2655": "failure",
+    "2656": "successful",
+    "2657": "successful",
+    "2658": "successful",
+    "2659": "successful",
+    "2660": "successful",
+    "2661": "failure",
+    "2662": "failure",
+    "2663": "successful",
+    "2664": "successful",
+    "2665": "failure",
+    "2666": "failure",
+    "2667": "failure",
+    "2668": "successful",
+    "2669": "failure",
+    "2670": "successful",
+    "2671": "failure",
+    "2672": "successful",
+    "2673": "successful",
+    "2674": "successful",
+    "2675": "successful",
+    "2676": "failure",
+    "2677": "successful",
+    "2678": "failure",
+    "2679": "infeasible",
+    "2680": "infeasible",
+    "2681": "failure",
+    "2682": "infeasible",
+    "2683": "failure",
+    "2684": "successful",
+    "2685": "failure",
+    "2686": "failure",
+    "2687": "failure",
+    "2688": "successful",
+    "2689": "successful",
+    "2690": "successful",
+    "2691": "successful",
+    "2692": "successful",
+    "2693": "failure",
+    "2694": "successful",
+    "2695": "failure",
+    "2696": "successful",
+    "2697": "successful",
+    "2698": "failure",
+    "2699": "failure",
+    "2700": "failure",
+    "2701": "failure",
+    "2702": "failure",
+    "2703": "failure",
+    "2704": "failure",
+    "2705": "successful",
+    "2706": "failure",
+    "2707": "successful",
+    "2708": "successful",
+    "2709": "failure",
+    "2710": "failure",
+    "2711": "successful",
+    "2712": "failure",
+    "2713": "successful",
+    "2714": "failure",
+    "2715": "failure",
+    "2716": "successful",
+    "2717": "already_complete",
+    "2718": "failure",
+    "2719": "failure",
+    "2720": "failure",
+    "2721": "successful",
+    "2722": "failure",
+    "2723": "failure",
+    "2724": "failure",
+    "2725": "failure",
+    "2726": "failure",
+    "2727": "successful",
+    "2728": "successful",
+    "2729": "failure",
+    "2730": "successful",
+    "2731": "successful",
+    "2732": "failure",
+    "2733": "failure",
+    "2734": "successful",
+    "2735": "successful",
+    "2736": "successful",
+    "2737": "failure",
+    "2738": "failure",
+    "2739": "successful",
+    "2740": "failure",
+    "2741": "failure",
+    "2742": "failure",
+    "2743": "failure",
+    "2744": "failure",
+    "2745": "successful",
+    "2746": "successful",
+    "2747": "failure",
+    "2748": "failure",
+    "2749": "successful",
+    "2750": "failure",
+    "2751": "failure",
+    "2752": "successful",
+    "2753": "successful",
+    "2754": "failure",
+    "2755": "successful",
+    "2756": "failure",
+    "2757": "successful",
+    "2758": "successful",
+    "2759": "failure",
+    "2760": "successful",
+    "2761": "failure",
+    "2762": "successful",
+    "2763": "failure",
+    "2764": "failure",
+    "2765": "failure",
+    "2766": "failure",
+    "2767": "failure",
+    "2768": "successful",
+    "2769": "successful",
+    "2770": "failure",
+    "2771": "already_complete",
+    "2772": "successful",
+    "2773": "failure",
+    "2774": "failure",
+    "2775": "failure",
+    "2776": "successful",
+    "2777": "failure",
+    "2778": "failure",
+    "2779": "failure",
+    "2780": "failure",
+    "2781": "successful",
+    "2782": "successful",
+    "2783": "successful",
+    "2784": "failure",
+    "2785": "successful",
+    "2786": "failure",
+    "2787": "infeasible",
+    "2788": "failure",
+    "2789": "failure",
+    "2790": "failure",
+    "2791": "successful",
+    "2792": "failure",
+    "2793": "successful",
+    "2794": "successful",
+    "2795": "infeasible",
+    "2796": "successful",
+    "2797": "successful",
+    "2798": "successful",
+    "2799": "failure",
+    "2800": "failure",
+    "2801": "failure",
+    "2802": "failure",
+    "2803": "failure",
+    "2804": "failure",
+    "2805": "failure",
+    "2806": "failure",
+    "2807": "successful",
+    "2808": "successful",
+    "2809": "failure",
+    "2810": "failure",
+    "2811": "successful",
+    "2812": "already_complete",
+    "2813": "successful",
+    "2814": "successful",
+    "2815": "failure",
+    "2816": "failure",
+    "2817": "successful",
+    "2818": "failure",
+    "2819": "successful",
+    "2820": "failure",
+    "2821": "failure",
+    "2822": "successful",
+    "2823": "failure",
+    "2824": "successful",
+    "2825": "successful",
+    "2826": "failure",
+    "2827": "successful",
+    "2828": "successful",
+    "2829": "failure",
+    "2830": "already_complete",
+    "2831": "successful",
+    "2832": "failure",
+    "2833": "successful",
+    "2834": "successful",
+    "2835": "failure",
+    "2836": "successful",
+    "2837": "successful",
+    "2838": "successful",
+    "2839": "successful",
+    "2840": "successful",
+    "2841": "failure",
+    "2842": "successful",
+    "2843": "failure",
+    "2844": "successful",
+    "2845": "successful",
+    "2846": "successful",
+    "2847": "successful",
+    "2848": "failure",
+    "2849": "failure",
+    "2850": "failure",
+    "2851": "successful",
+    "2852": "failure",
+    "2853": "successful",
+    "2854": "failure",
+    "2855": "failure",
+    "2856": "failure",
+    "2857": "successful",
+    "2858": "failure",
+    "2859": "failure",
+    "2860": "failure",
+    "2861": "successful",
+    "2862": "successful",
+    "2863": "failure",
+    "2864": "failure",
+    "2865": "successful",
+    "2866": "successful",
+    "2867": "failure",
+    "2868": "successful",
+    "2869": "successful",
+    "2870": "failure",
+    "2871": "failure",
+    "2872": "failure",
+    "2873": "already_complete",
+    "2874": "failure",
+    "2875": "failure",
+    "2876": "successful",
+    "2877": "failure",
+    "2878": "successful",
+    "2879": "failure",
+    "2880": "failure",
+    "2881": "successful",
+    "2882": "successful",
+    "2883": "successful",
+    "2884": "successful",
+    "2885": "successful",
+    "2886": "successful",
+    "2887": "failure",
+    "2888": "failure",
+    "2889": "failure",
+    "2890": "successful",
+    "2891": "failure",
+    "2892": "infeasible",
+    "2893": "failure",
+    "2894": "successful",
+    "2895": "failure",
+    "2896": "successful",
+    "2897": "successful",
+    "2898": "failure",
+    "2899": "failure",
+    "2900": "already_complete",
+    "2901": "successful",
+    "2902": "failure",
+    "2903": "successful",
+    "2904": "successful",
+    "2905": "failure",
+    "2906": "successful",
+    "2907": "failure",
+    "2908": "successful",
+    "2909": "failure",
+    "2910": "successful",
+    "2911": "infeasible",
+    "2912": "failure",
+    "2913": "successful",
+    "2914": "failure",
+    "2915": "failure",
+    "2916": "failure",
+    "2917": "successful",
+    "2918": "failure",
+    "2919": "successful",
+    "2920": "successful",
+    "2921": "failure",
+    "2922": "successful",
+    "2923": "failure",
+    "2924": "failure",
+    "2925": "successful",
+    "2926": "successful",
+    "2927": "successful",
+    "2928": "failure",
+    "2929": "already_complete",
+    "2930": "successful",
+    "2931": "successful",
+    "2932": "failure",
+    "2933": "successful",
+    "2934": "infeasible",
+    "2935": "failure",
+    "2936": "successful",
+    "2937": "successful",
+    "2938": "successful",
+    "2939": "failure",
+    "2940": "successful",
+    "2941": "successful",
+    "2942": "failure",
+    "2943": "successful",
+    "2944": "failure",
+    "2945": "failure",
+    "2946": "successful",
+    "2947": "failure",
+    "2948": "failure",
+    "2949": "successful",
+    "2950": "failure",
+    "2951": "successful",
+    "2952": "successful",
+    "2953": "failure",
+    "2954": "successful",
+    "2955": "successful",
+    "2956": "successful",
+    "2957": "successful",
+    "2958": "failure",
+    "2959": "failure",
+    "2960": "failure",
+    "2961": "successful",
+    "2962": "successful",
+    "2963": "failure",
+    "2964": "failure",
+    "2965": "successful",
+    "2966": "successful",
+    "2967": "failure",
+    "2968": "failure",
+    "2969": "failure",
+    "2970": "failure",
+    "2971": "infeasible",
+    "2972": "successful",
+    "2973": "failure",
+    "2974": "failure",
+    "2975": "successful",
+    "2976": "successful",
+    "2977": "failure",
+    "2978": "failure",
+    "2979": "failure",
+    "2980": "failure",
+    "2981": "failure",
+    "2982": "successful",
+    "2983": "successful",
+    "2984": "successful",
+    "2985": "successful",
+    "2986": "successful",
+    "2987": "failure",
+    "2988": "successful",
+    "2989": "successful",
+    "2990": "successful",
+    "2991": "successful",
+    "2992": "successful",
+    "2993": "successful",
+    "2994": "failure",
+    "2995": "successful",
+    "2996": "successful",
+    "2997": "failure",
+    "2998": "successful",
+    "2999": "infeasible",
+    "3000": "failure",
+    "3001": "failure",
+    "3002": "failure",
+    "3003": "successful",
+    "3004": "successful",
+    "3005": "failure",
+    "3006": "failure",
+    "3007": "successful",
+    "3008": "successful",
+    "3009": "successful",
+    "3010": "failure",
+    "3011": "successful",
+    "3012": "successful",
+    "3013": "failure",
+    "3014": "failure",
+    "3015": "successful",
+    "3016": "already_complete",
+    "3017": "already_complete",
+    "3018": "successful",
+    "3019": "successful",
+    "3020": "failure",
+    "3021": "failure",
+    "3022": "successful",
+    "3023": "failure",
+    "3024": "successful",
+    "3025": "failure",
+    "3026": "infeasible",
+    "3027": "infeasible",
+    "3028": "failure",
+    "3029": "successful",
+    "3030": "failure",
+    "3031": "failure",
+    "3032": "successful",
+    "3033": "successful",
+    "3034": "failure",
+    "3035": "failure",
+    "3036": "successful",
+    "3037": "successful",
+    "3038": "failure",
+    "3039": "successful",
+    "3040": "successful",
+    "3041": "failure",
+    "3042": "failure",
+    "3043": "successful",
+    "3044": "successful",
+    "3045": "failure",
+    "3046": "failure",
+    "3047": "failure",
+    "3048": "failure",
+    "3049": "failure",
+    "3050": "failure",
+    "3051": "failure",
+    "3052": "successful",
+    "3053": "failure",
+    "3054": "successful",
+    "3055": "failure",
+    "3056": "failure",
+    "3057": "successful",
+    "3058": "successful",
+    "3059": "successful",
+    "3060": "successful",
+    "3061": "successful",
+    "3062": "failure",
+    "3063": "failure",
+    "3064": "failure",
+    "3065": "successful",
+    "3066": "failure",
+    "3067": "successful",
+    "3068": "successful",
+    "3069": "successful",
+    "3070": "successful",
+    "3071": "successful",
+    "3072": "failure",
+    "3073": "successful",
+    "3074": "successful",
+    "3075": "successful",
+    "3076": "failure",
+    "3077": "failure",
+    "3078": "successful",
+    "3079": "failure",
+    "3080": "successful",
+    "3081": "failure",
+    "3082": "failure",
+    "3083": "failure",
+    "3084": "failure",
+    "3085": "already_complete",
+    "3086": "successful",
+    "3087": "successful",
+    "3088": "successful",
+    "3089": "failure",
+    "3090": "successful",
+    "3091": "infeasible",
+    "3092": "failure",
+    "3093": "failure",
+    "3094": "failure",
+    "3095": "failure",
+    "3096": "successful",
+    "3097": "failure",
+    "3098": "failure",
+    "3099": "failure",
+    "3100": "successful",
+    "3101": "successful",
+    "3102": "failure",
+    "3103": "successful",
+    "3104": "failure",
+    "3105": "failure",
+    "3106": "successful",
+    "3107": "already_complete",
+    "3108": "failure",
+    "3109": "failure",
+    "3110": "successful",
+    "3111": "failure",
+    "3112": "failure",
+    "3113": "failure",
+    "3114": "successful",
+    "3115": "successful",
+    "3116": "failure",
+    "3117": "failure",
+    "3118": "failure",
+    "3119": "successful",
+    "3120": "successful",
+    "3121": "successful",
+    "3122": "successful",
+    "3123": "failure",
+    "3124": "failure",
+    "3125": "successful",
+    "3126": "successful",
+    "3127": "successful",
+    "3128": "successful",
+    "3129": "failure",
+    "3130": "failure",
+    "3131": "failure",
+    "3132": "failure",
+    "3133": "failure",
+    "3134": "failure",
+    "3135": "failure",
+    "3136": "successful",
+    "3137": "successful",
+    "3138": "failure",
+    "3139": "successful",
+    "3140": "failure",
+    "3141": "failure",
+    "3142": "failure",
+    "3143": "failure",
+    "3144": "failure",
+    "3145": "failure",
+    "3146": "failure",
+    "3147": "successful",
+    "3148": "successful",
+    "3149": "successful",
+    "3150": "successful",
+    "3151": "failure",
+    "3152": "failure",
+    "3153": "failure",
+    "3154": "failure",
+    "3155": "failure",
+    "3156": "successful",
+    "3157": "failure",
+    "3158": "successful",
+    "3159": "failure",
+    "3160": "failure",
+    "3161": "failure",
+    "3162": "failure",
+    "3163": "successful",
+    "3164": "successful",
+    "3165": "successful",
+    "3166": "failure",
+    "3167": "successful",
+    "3168": "failure",
+    "3169": "successful",
+    "3170": "successful",
+    "3171": "failure",
+    "3172": "successful",
+    "3173": "failure",
+    "3174": "successful",
+    "3175": "successful",
+    "3176": "successful",
+    "3177": "already_complete",
+    "3178": "failure",
+    "3179": "failure",
+    "3180": "successful",
+    "3181": "successful",
+    "3182": "failure",
+    "3183": "failure",
+    "3184": "successful",
+    "3185": "successful",
+    "3186": "successful",
+    "3187": "successful",
+    "3188": "failure",
+    "3189": "failure",
+    "3190": "failure",
+    "3191": "successful",
+    "3192": "successful",
+    "3193": "failure",
+    "3194": "failure",
+    "3195": "failure",
+    "3196": "successful",
+    "3197": "failure",
+    "3198": "failure",
+    "3199": "successful",
+    "3200": "failure",
+    "3201": "successful",
+    "3202": "successful",
+    "3203": "successful",
+    "3204": "successful",
+    "3205": "successful",
+    "3206": "successful",
+    "3207": "successful",
+    "3208": "successful",
+    "3209": "successful",
+    "3210": "successful",
+    "3211": "successful",
+    "3212": "successful",
+    "3213": "failure",
+    "3214": "successful",
+    "3215": "successful",
+    "3216": "successful",
+    "3217": "failure",
+    "3218": "failure",
+    "3219": "failure",
+    "3220": "failure",
+    "3221": "failure",
+    "3222": "failure",
+    "3223": "successful",
+    "3224": "successful",
+    "3225": "successful",
+    "3226": "infeasible",
+    "3227": "failure",
+    "3228": "successful",
+    "3229": "successful",
+    "3230": "failure",
+    "3231": "failure",
+    "3232": "failure",
+    "3233": "successful",
+    "3234": "failure",
+    "3235": "successful",
+    "3236": "failure",
+    "3237": "failure",
+    "3238": "failure",
+    "3239": "successful",
+    "3240": "successful",
+    "3241": "failure",
+    "3242": "successful",
+    "3243": "failure",
+    "3244": "successful",
+    "3245": "failure",
+    "3246": "infeasible",
+    "3247": "successful",
+    "3248": "failure",
+    "3249": "successful",
+    "3250": "failure",
+    "3251": "successful",
+    "3252": "infeasible",
+    "3253": "successful",
+    "3254": "failure",
+    "3255": "failure",
+    "3256": "successful",
+    "3257": "successful",
+    "3258": "failure",
+    "3259": "successful",
+    "3260": "successful",
+    "3261": "successful",
+    "3262": "successful",
+    "3263": "failure",
+    "3264": "successful",
+    "3265": "successful",
+    "3266": "failure",
+    "3267": "successful",
+    "3268": "failure",
+    "3269": "failure",
+    "3270": "successful",
+    "3271": "successful",
+    "3272": "failure",
+    "3273": "successful",
+    "3274": "failure",
+    "3275": "successful",
+    "3276": "successful",
+    "3277": "successful",
+    "3278": "successful",
+    "3279": "successful",
+    "3280": "successful",
+    "3281": "successful",
+    "3282": "successful",
+    "3283": "successful",
+    "3284": "failure",
+    "3285": "failure",
+    "3286": "failure",
+    "3287": "failure",
+    "3288": "successful",
+    "3289": "failure",
+    "3290": "already_complete",
+    "3291": "already_complete",
+    "3292": "successful",
+    "3293": "failure",
+    "3294": "successful",
+    "3295": "failure",
+    "3296": "failure",
+    "3297": "failure",
+    "3298": "failure",
+    "3299": "successful",
+    "3300": "failure",
+    "3301": "successful",
+    "3302": "infeasible",
+    "3303": "infeasible",
+    "3304": "successful",
+    "3305": "failure",
+    "3306": "failure",
+    "3307": "failure",
+    "3308": "successful",
+    "3309": "successful",
+    "3310": "failure",
+    "3311": "successful",
+    "3312": "already_complete",
+    "3313": "failure",
+    "3314": "failure",
+    "3315": "successful",
+    "3316": "successful",
+    "3317": "successful",
+    "3318": "failure",
+    "3319": "successful",
+    "3320": "failure",
+    "3321": "successful",
+    "3322": "successful",
+    "3323": "successful",
+    "3324": "successful",
+    "3325": "failure",
+    "3326": "failure",
+    "3327": "failure",
+    "3328": "failure",
+    "3329": "successful",
+    "3330": "failure",
+    "3331": "successful",
+    "3332": "failure",
+    "3333": "successful",
+    "3334": "failure",
+    "3335": "failure",
+    "3336": "failure",
+    "3337": "successful",
+    "3338": "successful",
+    "3339": "successful",
+    "3340": "failure",
+    "3341": "failure",
+    "3342": "successful",
+    "3343": "successful",
+    "3344": "failure",
+    "3345": "failure",
+    "3346": "failure",
+    "3347": "successful",
+    "3348": "successful",
+    "3349": "failure",
+    "3350": "successful",
+    "3351": "failure",
+    "3352": "successful",
+    "3353": "failure",
+    "3354": "successful",
+    "3355": "successful",
+    "3356": "failure",
+    "3357": "successful",
+    "3358": "failure",
+    "3359": "failure",
+    "3360": "successful",
+    "3361": "infeasible",
+    "3362": "failure",
+    "3363": "successful",
+    "3364": "failure",
+    "3365": "already_complete",
+    "3366": "failure",
+    "3367": "already_complete",
+    "3368": "failure",
+    "3369": "successful",
+    "3370": "successful",
+    "3371": "successful",
+    "3372": "failure",
+    "3373": "failure",
+    "3374": "failure",
+    "3375": "successful",
+    "3376": "successful",
+    "3377": "failure",
+    "3378": "failure",
+    "3379": "failure",
+    "3380": "successful",
+    "3381": "successful",
+    "3382": "successful",
+    "3383": "failure",
+    "3384": "successful",
+    "3385": "failure",
+    "3386": "failure",
+    "3387": "successful",
+    "3388": "failure",
+    "3389": "successful",
+    "3390": "successful",
+    "3391": "failure",
+    "3392": "failure",
+    "3393": "failure",
+    "3394": "successful",
+    "3395": "successful",
+    "3396": "failure",
+    "3397": "failure",
+    "3398": "failure",
+    "3399": "successful",
+    "3400": "successful",
+    "3401": "successful",
+    "3402": "failure",
+    "3403": "failure",
+    "3404": "successful",
+    "3405": "failure",
+    "3406": "failure",
+    "3407": "failure",
+    "3408": "successful",
+    "3409": "successful",
+    "3410": "failure",
+    "3411": "failure",
+    "3412": "successful",
+    "3413": "successful",
+    "3414": "failure",
+    "3415": "failure",
+    "3416": "successful",
+    "3417": "successful",
+    "3418": "failure",
+    "3419": "failure",
+    "3420": "successful",
+    "3421": "failure",
+    "3422": "failure",
+    "3423": "failure",
+    "3424": "successful",
+    "3425": "successful",
+    "3426": "successful",
+    "3427": "failure",
+    "3428": "failure",
+    "3429": "successful",
+    "3430": "failure",
+    "3431": "failure",
+    "3432": "failure",
+    "3433": "successful",
+    "3434": "successful",
+    "3435": "failure",
+    "3436": "successful",
+    "3437": "failure",
+    "3438": "failure",
+    "3439": "successful",
+    "3440": "successful",
+    "3441": "successful",
+    "3442": "successful",
+    "3443": "failure",
+    "3444": "successful",
+    "3445": "successful",
+    "3446": "failure",
+    "3447": "successful",
+    "3448": "failure",
+    "3449": "successful",
+    "3450": "already_complete",
+    "3451": "successful",
+    "3452": "failure",
+    "3453": "successful",
+    "3454": "failure",
+    "3455": "failure",
+    "3456": "failure",
+    "3457": "failure",
+    "3458": "successful",
+    "3459": "successful",
+    "3460": "successful",
+    "3461": "successful",
+    "3462": "successful",
+    "3463": "successful",
+    "3464": "successful",
+    "3465": "failure",
+    "3466": "failure",
+    "3467": "failure",
+    "3468": "successful",
+    "3469": "failure",
+    "3470": "failure",
+    "3471": "failure",
+    "3472": "failure",
+    "3473": "successful",
+    "3474": "successful",
+    "3475": "successful",
+    "3476": "failure",
+    "3477": "successful",
+    "3478": "successful",
+    "3479": "failure",
+    "3480": "successful",
+    "3481": "successful",
+    "3482": "failure",
+    "3483": "failure",
+    "3484": "failure",
+    "3485": "failure",
+    "3486": "successful",
+    "3487": "successful",
+    "3488": "failure",
+    "3489": "successful",
+    "3490": "successful",
+    "3491": "successful",
+    "3492": "failure",
+    "3493": "successful",
+    "3494": "failure",
+    "3495": "failure",
+    "3496": "failure",
+    "3497": "failure",
+    "3498": "successful",
+    "3499": "failure",
+    "3500": "already_complete",
+    "3501": "successful",
+    "3502": "successful",
+    "3503": "failure",
+    "3504": "failure",
+    "3505": "failure",
+    "3506": "failure",
+    "3507": "failure",
+    "3508": "successful",
+    "3509": "successful",
+    "3510": "failure",
+    "3511": "failure",
+    "3512": "failure",
+    "3513": "failure",
+    "3514": "successful",
+    "3515": "failure",
+    "3516": "successful",
+    "3517": "successful",
+    "3518": "failure",
+    "3519": "failure",
+    "3520": "failure",
+    "3521": "failure",
+    "3522": "successful",
+    "3523": "failure",
+    "3524": "successful",
+    "3525": "successful",
+    "3526": "successful",
+    "3527": "successful",
+    "3528": "successful",
+    "3529": "failure",
+    "3530": "failure",
+    "3531": "failure",
+    "3532": "failure",
+    "3533": "failure",
+    "3534": "failure",
+    "3535": "successful",
+    "3536": "failure",
+    "3537": "failure",
+    "3538": "failure",
+    "3539": "failure",
+    "3540": "successful",
+    "3541": "failure",
+    "3542": "successful",
+    "3543": "failure",
+    "3544": "failure",
+    "3545": "failure",
+    "3546": "successful",
+    "3547": "failure",
+    "3548": "failure",
+    "3549": "failure",
+    "3550": "failure",
+    "3551": "successful",
+    "3552": "failure",
+    "3553": "successful",
+    "3554": "failure",
+    "3555": "successful",
+    "3556": "successful",
+    "3557": "failure",
+    "3558": "failure",
+    "3559": "successful",
+    "3560": "successful",
+    "3561": "successful",
+    "3562": "failure",
+    "3563": "failure",
+    "3564": "successful",
+    "3565": "failure",
+    "3566": "infeasible",
+    "3567": "already_complete",
+    "3568": "failure",
+    "3569": "successful",
+    "3570": "successful",
+    "3571": "successful",
+    "3572": "successful",
+    "3573": "successful",
+    "3574": "failure",
+    "3575": "failure",
+    "3576": "failure",
+    "3577": "failure",
+    "3578": "successful",
+    "3579": "successful",
+    "3580": "failure",
+    "3581": "successful",
+    "3582": "successful",
+    "3583": "failure",
+    "3584": "failure",
+    "3585": "failure",
+    "3586": "failure",
+    "3587": "successful",
+    "3588": "successful",
+    "3589": "failure",
+    "3590": "failure",
+    "3591": "failure",
+    "3592": "failure",
+    "3593": "successful",
+    "3594": "successful",
+    "3595": "successful",
+    "3596": "successful",
+    "3597": "successful",
+    "3598": "failure",
+    "3599": "already_complete",
+    "3600": "failure",
+    "3601": "failure",
+    "3602": "failure",
+    "3603": "successful",
+    "3604": "successful",
+    "3605": "successful",
+    "3606": "failure",
+    "3607": "successful",
+    "3608": "failure",
+    "3609": "successful",
+    "3610": "failure",
+    "3611": "failure",
+    "3612": "successful",
+    "3613": "failure",
+    "3614": "already_complete",
+    "3615": "failure",
+    "3616": "infeasible",
+    "3617": "successful",
+    "3618": "failure",
+    "3619": "successful",
+    "3620": "successful",
+    "3621": "successful",
+    "3622": "failure",
+    "3623": "infeasible",
+    "3624": "successful",
+    "3625": "failure",
+    "3626": "already_complete",
+    "3627": "failure",
+    "3628": "failure",
+    "3629": "successful",
+    "3630": "failure",
+    "3631": "failure",
+    "3632": "failure",
+    "3633": "failure",
+    "3634": "successful",
+    "3635": "failure",
+    "3636": "failure",
+    "3637": "failure",
+    "3638": "already_complete",
+    "3639": "successful",
+    "3640": "failure",
+    "3641": "failure",
+    "3642": "successful",
+    "3643": "failure",
+    "3644": "failure",
+    "3645": "successful",
+    "3646": "failure",
+    "3647": "successful",
+    "3648": "failure",
+    "3649": "successful",
+    "3650": "successful",
+    "3651": "failure",
+    "3652": "failure",
+    "3653": "successful",
+    "3654": "successful",
+    "3655": "failure",
+    "3656": "successful",
+    "3657": "successful",
+    "3658": "failure",
+    "3659": "failure",
+    "3660": "failure",
+    "3661": "failure",
+    "3662": "successful",
+    "3663": "successful",
+    "3664": "successful",
+    "3665": "successful",
+    "3666": "successful",
+    "3667": "already_complete",
+    "3668": "failure",
+    "3669": "failure",
+    "3670": "failure",
+    "3671": "failure",
+    "3672": "successful",
+    "3673": "successful",
+    "3674": "successful",
+    "3675": "failure",
+    "3676": "successful",
+    "3677": "successful",
+    "3678": "failure",
+    "3679": "successful",
+    "3680": "successful",
+    "3681": "failure",
+    "3682": "failure",
+    "3683": "failure",
+    "3684": "failure",
+    "3685": "failure",
+    "3686": "successful",
+    "3687": "failure",
+    "3688": "failure",
+    "3689": "successful",
+    "3690": "failure",
+    "3691": "failure",
+    "3692": "successful",
+    "3693": "failure",
+    "3694": "successful",
+    "3695": "successful",
+    "3696": "successful",
+    "3697": "failure",
+    "3698": "failure",
+    "3699": "infeasible",
+    "3700": "successful",
+    "3701": "successful",
+    "3702": "failure",
+    "3703": "failure",
+    "3704": "already_complete",
+    "3705": "successful",
+    "3706": "successful",
+    "3707": "successful",
+    "3708": "successful",
+    "3709": "successful",
+    "3710": "failure",
+    "3711": "failure",
+    "3712": "failure",
+    "3713": "successful",
+    "3714": "failure",
+    "3715": "failure",
+    "3716": "successful",
+    "3717": "failure",
+    "3718": "failure",
+    "3719": "failure",
+    "3720": "infeasible",
+    "3721": "successful",
+    "3722": "failure",
+    "3723": "successful",
+    "3724": "failure",
+    "3725": "failure",
+    "3726": "failure",
+    "3727": "successful",
+    "3728": "successful",
+    "3729": "successful",
+    "3730": "failure",
+    "3731": "successful",
+    "3732": "successful",
+    "3733": "successful",
+    "3734": "successful",
+    "3735": "successful",
+    "3736": "failure",
+    "3737": "successful",
+    "3738": "failure",
+    "3739": "failure",
+    "3740": "failure",
+    "3741": "successful",
+    "3742": "successful",
+    "3743": "already_complete",
+    "3744": "failure",
+    "3745": "failure",
+    "3746": "already_complete",
+    "3747": "successful",
+    "3748": "successful",
+    "3749": "successful",
+    "3750": "failure",
+    "3751": "failure",
+    "3752": "infeasible",
+    "3753": "successful",
+    "3754": "failure",
+    "3755": "failure",
+    "3756": "successful",
+    "3757": "successful",
+    "3758": "successful",
+    "3759": "failure",
+    "3760": "successful",
+    "3761": "successful",
+    "3762": "successful",
+    "3763": "failure",
+    "3764": "successful",
+    "3765": "successful",
+    "3766": "failure",
+    "3767": "successful",
+    "3768": "already_complete",
+    "3769": "successful",
+    "3770": "successful",
+    "3771": "failure",
+    "3772": "successful",
+    "3773": "successful",
+    "3774": "failure",
+    "3775": "failure",
+    "3776": "successful",
+    "3777": "failure",
+    "3778": "successful",
+    "3779": "failure",
+    "3780": "successful",
+    "3781": "successful",
+    "3782": "failure",
+    "3783": "failure",
+    "3784": "successful",
+    "3785": "failure",
+    "3786": "failure",
+    "3787": "successful",
+    "3788": "successful",
+    "3789": "successful",
+    "3790": "successful",
+    "3791": "already_complete",
+    "3792": "successful",
+    "3793": "failure",
+    "3794": "failure",
+    "3795": "successful",
+    "3796": "successful",
+    "3797": "failure",
+    "3798": "successful",
+    "3799": "successful",
+    "3800": "failure",
+    "3801": "successful",
+    "3802": "successful",
+    "3803": "failure",
+    "3804": "failure",
+    "3805": "failure",
+    "3806": "infeasible",
+    "3807": "successful",
+    "3808": "failure",
+    "3809": "failure",
+    "3810": "failure",
+    "3811": "successful",
+    "3812": "failure",
+    "3813": "failure",
+    "3814": "failure",
+    "3815": "successful",
+    "3816": "successful",
+    "3817": "failure",
+    "3818": "failure",
+    "3819": "failure",
+    "3820": "successful",
+    "3821": "successful",
+    "3822": "successful",
+    "3823": "failure",
+    "3824": "successful",
+    "3825": "successful",
+    "3826": "failure",
+    "3827": "failure",
+    "3828": "failure",
+    "3829": "failure",
+    "3830": "successful",
+    "3831": "failure",
+    "3832": "successful",
+    "3833": "failure",
+    "3834": "successful",
+    "3835": "failure",
+    "3836": "failure",
+    "3837": "successful",
+    "3838": "failure",
+    "3839": "successful",
+    "3840": "failure",
+    "3841": "successful",
+    "3842": "failure",
+    "3843": "failure",
+    "3844": "failure",
+    "3845": "failure",
+    "3846": "failure",
+    "3847": "failure",
+    "3848": "failure",
+    "3849": "failure",
+    "3850": "successful",
+    "3851": "failure",
+    "3852": "failure",
+    "3853": "successful",
+    "3854": "failure",
+    "3855": "successful",
+    "3856": "successful",
+    "3857": "successful",
+    "3858": "successful",
+    "3859": "successful",
+    "3860": "failure",
+    "3861": "successful",
+    "3862": "successful",
+    "3863": "successful",
+    "3864": "successful",
+    "3865": "failure",
+    "3866": "failure",
+    "3867": "failure",
+    "3868": "failure",
+    "3869": "failure",
+    "3870": "successful",
+    "3871": "failure",
+    "3872": "failure",
+    "3873": "already_complete",
+    "3874": "successful",
+    "3875": "successful",
+    "3876": "successful",
+    "3877": "successful",
+    "3878": "failure",
+    "3879": "successful",
+    "3880": "failure",
+    "3881": "successful",
+    "3882": "failure",
+    "3883": "successful",
+    "3884": "failure",
+    "3885": "successful",
+    "3886": "failure",
+    "3887": "successful",
+    "3888": "failure",
+    "3889": "successful",
+    "3890": "successful",
+    "3891": "failure",
+    "3892": "successful",
+    "3893": "failure",
+    "3894": "failure",
+    "3895": "failure",
+    "3896": "successful",
+    "3897": "successful",
+    "3898": "successful",
+    "3899": "successful",
+    "3900": "failure",
+    "3901": "successful",
+    "3902": "successful",
+    "3903": "successful",
+    "3904": "successful",
+    "3905": "successful",
+    "3906": "successful",
+    "3907": "failure",
+    "3908": "failure",
+    "3909": "failure",
+    "3910": "infeasible",
+    "3911": "failure",
+    "3912": "infeasible",
+    "3913": "successful",
+    "3914": "failure",
+    "3915": "successful",
+    "3916": "failure",
+    "3917": "successful",
+    "3918": "successful",
+    "3919": "successful",
+    "3920": "failure",
+    "3921": "failure",
+    "3922": "failure",
+    "3923": "successful",
+    "3924": "failure",
+    "3925": "successful",
+    "3926": "successful",
+    "3927": "failure",
+    "3928": "successful",
+    "3929": "successful",
+    "3930": "failure",
+    "3931": "failure",
+    "3932": "successful",
+    "3933": "failure",
+    "3934": "failure",
+    "3935": "failure",
+    "3936": "successful",
+    "3937": "failure",
+    "3938": "failure",
+    "3939": "successful",
+    "3940": "successful",
+    "3941": "failure",
+    "3942": "successful",
+    "3943": "failure",
+    "3944": "failure",
+    "3945": "successful",
+    "3946": "successful",
+    "3947": "failure",
+    "3948": "infeasible",
+    "3949": "successful",
+    "3950": "successful",
+    "3951": "successful",
+    "3952": "failure",
+    "3953": "failure",
+    "3954": "successful",
+    "3955": "successful",
+    "3956": "successful",
+    "3957": "successful",
+    "3958": "successful",
+    "3959": "failure",
+    "3960": "failure",
+    "3961": "successful",
+    "3962": "failure",
+    "3963": "failure",
+    "3964": "successful",
+    "3965": "successful",
+    "3966": "successful",
+    "3967": "failure",
+    "3968": "failure",
+    "3969": "failure",
+    "3970": "successful",
+    "3971": "failure",
+    "3972": "failure",
+    "3973": "successful",
+    "3974": "infeasible",
+    "3975": "failure",
+    "3976": "failure",
+    "3977": "successful",
+    "3978": "successful",
+    "3979": "failure",
+    "3980": "successful",
+    "3981": "already_complete",
+    "3982": "failure",
+    "3983": "successful",
+    "3984": "successful",
+    "3985": "failure",
+    "3986": "successful",
+    "3987": "infeasible",
+    "3988": "failure",
+    "3989": "failure",
+    "3990": "failure",
+    "3991": "successful",
+    "3992": "failure",
+    "3993": "failure",
+    "3994": "failure",
+    "3995": "failure",
+    "3996": "failure",
+    "3997": "failure",
+    "3998": "failure",
+    "3999": "failure",
+    "4000": "failure",
+    "4001": "failure",
+    "4002": "failure",
+    "4003": "failure",
+    "4004": "successful",
+    "4005": "failure",
+    "4006": "successful",
+    "4007": "failure",
+    "4008": "infeasible",
+    "4009": "failure",
+    "4010": "successful",
+    "4011": "successful",
+    "4012": "failure",
+    "4013": "failure",
+    "4014": "successful",
+    "4015": "failure",
+    "4016": "failure",
+    "4017": "failure",
+    "4018": "successful",
+    "4019": "successful",
+    "4020": "successful",
+    "4021": "infeasible",
+    "4022": "failure",
+    "4023": "successful",
+    "4024": "failure",
+    "4025": "successful",
+    "4026": "successful",
+    "4027": "successful",
+    "4028": "successful",
+    "4029": "successful",
+    "4030": "successful",
+    "4031": "failure",
+    "4032": "failure",
+    "4033": "failure",
+    "4034": "failure",
+    "4035": "failure",
+    "4036": "failure",
+    "4037": "failure",
+    "4038": "failure",
+    "4039": "successful",
+    "4040": "failure",
+    "4041": "failure",
+    "4042": "failure",
+    "4043": "successful",
+    "4044": "successful",
+    "4045": "failure",
+    "4046": "failure",
+    "4047": "failure",
+    "4048": "failure",
+    "4049": "failure",
+    "4050": "failure",
+    "4051": "successful",
+    "4052": "failure",
+    "4053": "successful",
+    "4054": "infeasible",
+    "4055": "successful",
+    "4056": "successful",
+    "4057": "failure",
+    "4058": "successful",
+    "4059": "failure",
+    "4060": "successful",
+    "4061": "failure",
+    "4062": "failure",
+    "4063": "successful",
+    "4064": "successful",
+    "4065": "failure",
+    "4066": "successful",
+    "4067": "failure",
+    "4068": "successful",
+    "4069": "successful",
+    "4070": "successful",
+    "4071": "successful",
+    "4072": "failure",
+    "4073": "failure",
+    "4074": "failure",
+    "4075": "failure",
+    "4076": "failure",
+    "4077": "successful",
+    "4078": "successful",
+    "4079": "successful",
+    "4080": "failure",
+    "4081": "failure",
+    "4082": "already_complete",
+    "4083": "successful",
+    "4084": "failure",
+    "4085": "failure",
+    "4086": "successful",
+    "4087": "failure",
+    "4088": "failure",
+    "4089": "successful",
+    "4090": "already_complete",
+    "4091": "successful",
+    "4092": "failure",
+    "4093": "failure",
+    "4094": "failure",
+    "4095": "successful",
+    "4096": "failure",
+    "4097": "already_complete",
+    "4098": "failure",
+    "4099": "successful",
+    "4100": "successful",
+    "4101": "successful",
+    "4102": "failure",
+    "4103": "successful",
+    "4104": "failure",
+    "4105": "successful",
+    "4106": "failure",
+    "4107": "successful",
+    "4108": "failure",
+    "4109": "infeasible",
+    "4110": "successful",
+    "4111": "successful",
+    "4112": "successful",
+    "4113": "failure",
+    "4114": "successful",
+    "4115": "failure",
+    "4116": "failure",
+    "4117": "successful",
+    "4118": "successful",
+    "4119": "failure",
+    "4120": "failure",
+    "4121": "successful",
+    "4122": "infeasible",
+    "4123": "successful",
+    "4124": "failure",
+    "4125": "failure",
+    "4126": "failure",
+    "4127": "failure",
+    "4128": "failure",
+    "4129": "failure",
+    "4130": "failure",
+    "4131": "failure",
+    "4132": "failure",
+    "4133": "failure",
+    "4134": "failure",
+    "4135": "failure",
+    "4136": "infeasible",
+    "4137": "failure",
+    "4138": "successful",
+    "4139": "failure",
+    "4140": "failure",
+    "4141": "failure",
+    "4142": "failure",
+    "4143": "successful",
+    "4144": "failure",
+    "4145": "failure",
+    "4146": "successful",
+    "4147": "successful",
+    "4148": "failure",
+    "4149": "failure",
+    "4150": "successful",
+    "4151": "failure",
+    "4152": "successful",
+    "4153": "failure",
+    "4154": "successful",
+    "4155": "failure",
+    "4156": "failure",
+    "4157": "failure",
+    "4158": "successful",
+    "4159": "failure",
+    "4160": "failure",
+    "4161": "failure",
+    "4162": "successful",
+    "4163": "failure",
+    "4164": "failure",
+    "4165": "failure",
+    "4166": "successful",
+    "4167": "successful",
+    "4168": "successful",
+    "4169": "failure",
+    "4170": "successful",
+    "4171": "successful",
+    "4172": "successful",
+    "4173": "failure",
+    "4174": "failure",
+    "4175": "failure",
+    "4176": "already_complete",
+    "4177": "successful",
+    "4178": "infeasible",
+    "4179": "failure",
+    "4180": "failure",
+    "4181": "successful",
+    "4182": "failure",
+    "4183": "successful",
+    "4184": "failure",
+    "4185": "failure",
+    "4186": "failure",
+    "4187": "failure",
+    "4188": "successful",
+    "4189": "failure",
+    "4190": "successful",
+    "4191": "infeasible",
+    "4192": "failure",
+    "4193": "failure",
+    "4194": "failure",
+    "4195": "successful",
+    "4196": "failure",
+    "4197": "successful",
+    "4198": "successful",
+    "4199": "failure",
+    "4200": "failure",
+    "4201": "failure",
+    "4202": "successful",
+    "4203": "failure",
+    "4204": "successful",
+    "4205": "failure",
+    "4206": "failure",
+    "4207": "successful",
+    "4208": "successful",
+    "4209": "successful",
+    "4210": "successful",
+    "4211": "successful",
+    "4212": "successful",
+    "4213": "failure",
+    "4214": "successful",
+    "4215": "failure",
+    "4216": "failure",
+    "4217": "failure",
+    "4218": "infeasible",
+    "4219": "failure",
+    "4220": "failure",
+    "4221": "successful",
+    "4222": "failure",
+    "4223": "failure",
+    "4224": "successful",
+    "4225": "failure",
+    "4226": "already_complete",
+    "4227": "failure",
+    "4228": "successful",
+    "4229": "failure",
+    "4230": "successful",
+    "4231": "successful",
+    "4232": "already_complete",
+    "4233": "successful",
+    "4234": "successful",
+    "4235": "successful",
+    "4236": "successful",
+    "4237": "successful",
+    "4238": "failure",
+    "4239": "successful",
+    "4240": "successful",
+    "4241": "failure",
+    "4242": "successful",
+    "4243": "successful",
+    "4244": "successful",
+    "4245": "successful",
+    "4246": "already_complete",
+    "4247": "failure",
+    "4248": "successful",
+    "4249": "successful",
+    "4250": "failure",
+    "4251": "successful",
+    "4252": "failure",
+    "4253": "successful",
+    "4254": "successful",
+    "4255": "failure",
+    "4256": "failure",
+    "4257": "successful",
+    "4258": "failure",
+    "4259": "failure",
+    "4260": "successful",
+    "4261": "failure",
+    "4262": "successful",
+    "4263": "failure",
+    "4264": "successful",
+    "4265": "successful",
+    "4266": "failure",
+    "4267": "failure",
+    "4268": "successful",
+    "4269": "failure",
+    "4270": "failure",
+    "4271": "successful",
+    "4272": "failure",
+    "4273": "successful",
+    "4274": "successful",
+    "4275": "already_complete",
+    "4276": "successful",
+    "4277": "failure",
+    "4278": "successful",
+    "4279": "successful",
+    "4280": "successful",
+    "4281": "failure",
+    "4282": "failure",
+    "4283": "failure",
+    "4284": "failure",
+    "4285": "failure",
+    "4286": "failure",
+    "4287": "successful",
+    "4288": "failure",
+    "4289": "successful",
+    "4290": "successful",
+    "4291": "successful",
+    "4292": "successful",
+    "4293": "successful",
+    "4294": "failure",
+    "4295": "successful",
+    "4296": "successful",
+    "4297": "already_complete",
+    "4298": "successful",
+    "4299": "successful",
+    "4300": "failure",
+    "4301": "successful",
+    "4302": "successful",
+    "4303": "successful",
+    "4304": "successful",
+    "4305": "failure",
+    "4306": "successful",
+    "4307": "successful",
+    "4308": "failure",
+    "4309": "failure",
+    "4310": "infeasible",
+    "4311": "successful",
+    "4312": "successful",
+    "4313": "failure",
+    "4314": "failure",
+    "4315": "failure",
+    "4316": "failure",
+    "4317": "infeasible",
+    "4318": "successful",
+    "4319": "failure",
+    "4320": "successful",
+    "4321": "failure",
+    "4322": "failure",
+    "4323": "failure",
+    "4324": "successful",
+    "4325": "failure",
+    "4326": "failure",
+    "4327": "successful",
+    "4328": "failure",
+    "4329": "successful",
+    "4330": "successful",
+    "4331": "failure",
+    "4332": "successful",
+    "4333": "failure",
+    "4334": "successful",
+    "4335": "failure",
+    "4336": "failure",
+    "4337": "failure",
+    "4338": "failure",
+    "4339": "successful",
+    "4340": "successful",
+    "4341": "failure",
+    "4342": "successful",
+    "4343": "successful",
+    "4344": "infeasible",
+    "4345": "successful",
+    "4346": "successful",
+    "4347": "successful",
+    "4348": "successful",
+    "4349": "successful",
+    "4350": "successful",
+    "4351": "failure",
+    "4352": "failure",
+    "4353": "failure",
+    "4354": "failure",
+    "4355": "failure",
+    "4356": "failure",
+    "4357": "failure",
+    "4358": "successful",
+    "4359": "failure",
+    "4360": "successful",
+    "4361": "failure",
+    "4362": "failure",
+    "4363": "successful",
+    "4364": "successful",
+    "4365": "infeasible",
+    "4366": "failure",
+    "4367": "successful",
+    "4368": "successful",
+    "4369": "failure",
+    "4370": "successful",
+    "4371": "successful",
+    "4372": "successful",
+    "4373": "successful",
+    "4374": "failure",
+    "4375": "infeasible",
+    "4376": "successful",
+    "4377": "successful",
+    "4378": "successful",
+    "4379": "successful",
+    "4380": "failure",
+    "4381": "successful",
+    "4382": "successful",
+    "4383": "successful",
+    "4384": "already_complete",
+    "4385": "successful",
+    "4386": "failure",
+    "4387": "failure",
+    "4388": "failure",
+    "4389": "failure",
+    "4390": "successful",
+    "4391": "successful",
+    "4392": "failure",
+    "4393": "already_complete",
+    "4394": "failure",
+    "4395": "successful",
+    "4396": "successful",
+    "4397": "successful",
+    "4398": "failure",
+    "4399": "failure",
+    "4400": "failure",
+    "4401": "successful",
+    "4402": "failure",
+    "4403": "failure",
+    "4404": "successful",
+    "4405": "failure",
+    "4406": "successful",
+    "4407": "successful",
+    "4408": "failure",
+    "4409": "failure",
+    "4410": "successful",
+    "4411": "successful",
+    "4412": "successful",
+    "4413": "successful",
+    "4414": "failure",
+    "4415": "successful",
+    "4416": "failure",
+    "4417": "failure",
+    "4418": "successful",
+    "4419": "successful",
+    "4420": "failure",
+    "4421": "successful",
+    "4422": "successful",
+    "4423": "already_complete",
+    "4424": "failure",
+    "4425": "failure",
+    "4426": "failure",
+    "4427": "failure",
+    "4428": "successful",
+    "4429": "failure",
+    "4430": "failure",
+    "4431": "failure",
+    "4432": "successful",
+    "4433": "failure",
+    "4434": "successful",
+    "4435": "failure",
+    "4436": "failure",
+    "4437": "failure",
+    "4438": "failure",
+    "4439": "failure",
+    "4440": "successful",
+    "4441": "failure",
+    "4442": "successful",
+    "4443": "successful",
+    "4444": "failure",
+    "4445": "failure",
+    "4446": "failure",
+    "4447": "successful",
+    "4448": "failure",
+    "4449": "successful",
+    "4450": "failure",
+    "4451": "successful",
+    "4452": "failure",
+    "4453": "successful",
+    "4454": "successful",
+    "4455": "successful",
+    "4456": "failure",
+    "4457": "failure",
+    "4458": "failure",
+    "4459": "successful",
+    "4460": "successful",
+    "4461": "failure",
+    "4462": "failure",
+    "4463": "successful",
+    "4464": "successful",
+    "4465": "failure",
+    "4466": "successful",
+    "4467": "successful",
+    "4468": "failure",
+    "4469": "successful",
+    "4470": "failure",
+    "4471": "failure",
+    "4472": "failure",
+    "4473": "failure",
+    "4474": "successful",
+    "4475": "failure",
+    "4476": "failure",
+    "4477": "successful",
+    "4478": "failure",
+    "4479": "successful",
+    "4480": "infeasible",
+    "4481": "successful",
+    "4482": "failure",
+    "4483": "failure",
+    "4484": "failure",
+    "4485": "successful",
+    "4486": "failure",
+    "4487": "successful",
+    "4488": "successful",
+    "4489": "successful",
+    "4490": "successful",
+    "4491": "failure",
+    "4492": "successful",
+    "4493": "successful",
+    "4494": "failure",
+    "4495": "successful",
+    "4496": "successful",
+    "4497": "successful",
+    "4498": "failure",
+    "4499": "successful",
+    "4500": "failure",
+    "4501": "successful",
+    "4502": "failure",
+    "4503": "successful",
+    "4504": "failure",
+    "4505": "successful",
+    "4506": "successful",
+    "4507": "failure",
+    "4508": "successful",
+    "4509": "failure",
+    "4510": "failure",
+    "4511": "successful",
+    "4512": "successful",
+    "4513": "failure",
+    "4514": "failure",
+    "4515": "successful",
+    "4516": "successful",
+    "4517": "failure",
+    "4518": "infeasible",
+    "4519": "successful",
+    "4520": "failure",
+    "4521": "failure",
+    "4522": "failure",
+    "4523": "failure",
+    "4524": "failure",
+    "4525": "successful",
+    "4526": "failure",
+    "4527": "failure",
+    "4528": "failure",
+    "4529": "failure",
+    "4530": "failure",
+    "4531": "failure",
+    "4532": "successful",
+    "4533": "successful",
+    "4534": "successful",
+    "4535": "successful",
+    "4536": "infeasible",
+    "4537": "failure",
+    "4538": "successful",
+    "4539": "successful",
+    "4540": "successful",
+    "4541": "failure",
+    "4542": "successful",
+    "4543": "failure",
+    "4544": "successful",
+    "4545": "failure",
+    "4546": "failure",
+    "4547": "successful",
+    "4548": "failure",
+    "4549": "successful",
+    "4550": "successful",
+    "4551": "successful",
+    "4552": "failure",
+    "4553": "successful",
+    "4554": "successful",
+    "4555": "already_complete",
+    "4556": "successful",
+    "4557": "failure",
+    "4558": "successful",
+    "4559": "failure",
+    "4560": "successful",
+    "4561": "failure",
+    "4562": "failure",
+    "4563": "failure",
+    "4564": "failure",
+    "4565": "successful",
+    "4566": "failure",
+    "4567": "failure",
+    "4568": "failure",
+    "4569": "failure",
+    "4570": "successful",
+    "4571": "successful",
+    "4572": "successful",
+    "4573": "successful",
+    "4574": "successful",
+    "4575": "infeasible",
+    "4576": "failure",
+    "4577": "failure",
+    "4578": "failure",
+    "4579": "failure",
+    "4580": "already_complete",
+    "4581": "failure",
+    "4582": "failure",
+    "4583": "infeasible",
+    "4584": "successful",
+    "4585": "failure",
+    "4586": "infeasible",
+    "4587": "failure",
+    "4588": "failure",
+    "4589": "successful",
+    "4590": "successful",
+    "4591": "successful",
+    "4592": "successful",
+    "4593": "failure",
+    "4594": "successful",
+    "4595": "successful",
+    "4596": "successful",
+    "4597": "failure",
+    "4598": "failure",
+    "4599": "successful",
+    "4600": "failure",
+    "4601": "failure",
+    "4602": "failure",
+    "4603": "failure",
+    "4604": "successful",
+    "4605": "failure",
+    "4606": "failure",
+    "4607": "failure",
+    "4608": "successful",
+    "4609": "successful",
+    "4610": "successful",
+    "4611": "successful",
+    "4612": "failure",
+    "4613": "failure",
+    "4614": "failure",
+    "4615": "successful",
+    "4616": "failure",
+    "4617": "failure",
+    "4618": "successful",
+    "4619": "failure",
+    "4620": "failure",
+    "4621": "failure",
+    "4622": "successful",
+    "4623": "successful",
+    "4624": "successful",
+    "4625": "successful",
+    "4626": "failure",
+    "4627": "failure",
+    "4628": "successful",
+    "4629": "failure",
+    "4630": "failure",
+    "4631": "successful",
+    "4632": "failure",
+    "4633": "successful",
+    "4634": "failure",
+    "4635": "failure",
+    "4636": "successful",
+    "4637": "failure",
+    "4638": "failure",
+    "4639": "successful",
+    "4640": "successful",
+    "4641": "failure",
+    "4642": "failure",
+    "4643": "failure",
+    "4644": "failure",
+    "4645": "failure",
+    "4646": "successful",
+    "4647": "failure",
+    "4648": "already_complete",
+    "4649": "successful",
+    "4650": "successful",
+    "4651": "failure",
+    "4652": "failure",
+    "4653": "failure",
+    "4654": "successful",
+    "4655": "failure",
+    "4656": "failure",
+    "4657": "failure",
+    "4658": "successful",
+    "4659": "successful",
+    "4660": "failure",
+    "4661": "failure",
+    "4662": "failure",
+    "4663": "successful",
+    "4664": "successful",
+    "4665": "successful",
+    "4666": "failure",
+    "4667": "failure",
+    "4668": "failure",
+    "4669": "infeasible",
+    "4670": "successful",
+    "4671": "successful",
+    "4672": "successful",
+    "4673": "successful",
+    "4674": "successful",
+    "4675": "successful",
+    "4676": "successful",
+    "4677": "failure",
+    "4678": "successful",
+    "4679": "failure",
+    "4680": "failure",
+    "4681": "successful",
+    "4682": "failure",
+    "4683": "failure",
+    "4684": "failure",
+    "4685": "failure",
+    "4686": "failure",
+    "4687": "successful",
+    "4688": "failure",
+    "4689": "failure",
+    "4690": "successful",
+    "4691": "failure",
+    "4692": "successful",
+    "4693": "failure",
+    "4694": "successful",
+    "4695": "successful",
+    "4696": "successful",
+    "4697": "successful",
+    "4698": "successful",
+    "4699": "successful",
+    "4700": "successful",
+    "4701": "successful",
+    "4702": "successful",
+    "4703": "failure",
+    "4704": "successful",
+    "4705": "successful",
+    "4706": "failure",
+    "4707": "infeasible",
+    "4708": "failure",
+    "4709": "successful",
+    "4710": "failure",
+    "4711": "successful",
+    "4712": "failure",
+    "4713": "successful",
+    "4714": "failure",
+    "4715": "successful",
+    "4716": "successful",
+    "4717": "failure",
+    "4718": "failure",
+    "4719": "successful",
+    "4720": "successful",
+    "4721": "failure",
+    "4722": "failure",
+    "4723": "successful",
+    "4724": "successful",
+    "4725": "successful",
+    "4726": "successful",
+    "4727": "successful",
+    "4728": "failure",
+    "4729": "failure",
+    "4730": "successful",
+    "4731": "successful",
+    "4732": "successful",
+    "4733": "successful",
+    "4734": "failure",
+    "4735": "already_complete",
+    "4736": "failure",
+    "4737": "failure",
+    "4738": "failure",
+    "4739": "successful",
+    "4740": "failure",
+    "4741": "failure",
+    "4742": "failure",
+    "4743": "failure",
+    "4744": "failure",
+    "4745": "successful",
+    "4746": "failure",
+    "4747": "failure",
+    "4748": "successful",
+    "4749": "failure",
+    "4750": "failure",
+    "4751": "failure",
+    "4752": "successful",
+    "4753": "successful",
+    "4754": "failure",
+    "4755": "successful",
+    "4756": "failure",
+    "4757": "successful",
+    "4758": "failure",
+    "4759": "failure",
+    "4760": "failure",
+    "4761": "successful",
+    "4762": "failure",
+    "4763": "successful",
+    "4764": "infeasible",
+    "4765": "successful",
+    "4766": "failure",
+    "4767": "successful",
+    "4768": "successful",
+    "4769": "failure",
+    "4770": "failure",
+    "4771": "failure",
+    "4772": "successful",
+    "4773": "failure",
+    "4774": "failure",
+    "4775": "successful",
+    "4776": "successful",
+    "4777": "successful",
+    "4778": "failure",
+    "4779": "failure",
+    "4780": "failure",
+    "4781": "successful",
+    "4782": "failure",
+    "4783": "already_complete",
+    "4784": "successful",
+    "4785": "failure",
+    "4786": "failure",
+    "4787": "successful",
+    "4788": "successful",
+    "4789": "failure",
+    "4790": "infeasible",
+    "4791": "successful",
+    "4792": "successful",
+    "4793": "successful",
+    "4794": "failure",
+    "4795": "failure",
+    "4796": "failure",
+    "4797": "failure",
+    "4798": "successful",
+    "4799": "failure",
+    "4800": "successful",
+    "4801": "failure",
+    "4802": "failure",
+    "4803": "failure",
+    "4804": "successful",
+    "4805": "successful",
+    "4806": "failure",
+    "4807": "successful",
+    "4808": "successful",
+    "4809": "failure",
+    "4810": "successful",
+    "4811": "successful",
+    "4812": "failure",
+    "4813": "failure",
+    "4814": "successful",
+    "4815": "failure",
+    "4816": "successful",
+    "4817": "successful",
+    "4818": "failure",
+    "4819": "failure",
+    "4820": "successful",
+    "4821": "successful",
+    "4822": "failure",
+    "4823": "successful",
+    "4824": "failure",
+    "4825": "failure",
+    "4826": "successful",
+    "4827": "failure",
+    "4828": "infeasible",
+    "4829": "already_complete",
+    "4830": "failure",
+    "4831": "successful",
+    "4832": "successful",
+    "4833": "successful",
+    "4834": "failure",
+    "4835": "successful",
+    "4836": "failure",
+    "4837": "failure",
+    "4838": "successful",
+    "4839": "successful",
+    "4840": "successful",
+    "4841": "successful",
+    "4842": "failure",
+    "4843": "failure",
+    "4844": "failure",
+    "4845": "failure",
+    "4846": "failure",
+    "4847": "failure",
+    "4848": "failure",
+    "4849": "successful",
+    "4850": "successful",
+    "4851": "failure",
+    "4852": "successful",
+    "4853": "successful",
+    "4854": "failure",
+    "4855": "successful",
+    "4856": "failure",
+    "4857": "failure",
+    "4858": "failure",
+    "4859": "successful",
+    "4860": "failure",
+    "4861": "failure",
+    "4862": "failure",
+    "4863": "failure",
+    "4864": "successful",
+    "4865": "failure",
+    "4866": "failure",
+    "4867": "successful",
+    "4868": "failure",
+    "4869": "failure",
+    "4870": "failure",
+    "4871": "successful",
+    "4872": "failure",
+    "4873": "failure",
+    "4874": "failure",
+    "4875": "successful",
+    "4876": "failure",
+    "4877": "successful",
+    "4878": "successful",
+    "4879": "successful",
+    "4880": "successful",
+    "4881": "failure",
+    "4882": "successful",
+    "4883": "failure",
+    "4884": "successful",
+    "4885": "successful",
+    "4886": "successful",
+    "4887": "failure",
+    "4888": "failure",
+    "4889": "successful",
+    "4890": "failure",
+    "4891": "failure",
+    "4892": "successful",
+    "4893": "successful",
+    "4894": "failure",
+    "4895": "failure",
+    "4896": "successful",
+    "4897": "failure",
+    "4898": "successful",
+    "4899": "failure",
+    "4900": "failure",
+    "4901": "successful",
+    "4902": "successful",
+    "4903": "successful",
+    "4904": "successful",
+    "4905": "successful",
+    "4906": "failure",
+    "4907": "failure",
+    "4908": "infeasible",
+    "4909": "infeasible",
+    "4910": "successful",
+    "4911": "failure",
+    "4912": "successful",
+    "4913": "successful",
+    "4914": "successful",
+    "4915": "failure",
+    "4916": "failure",
+    "4917": "successful",
+    "4918": "failure",
+    "4919": "successful",
+    "4920": "failure",
+    "4921": "successful",
+    "4922": "successful",
+    "4923": "failure",
+    "4924": "failure",
+    "4925": "failure",
+    "4926": "failure",
+    "4927": "successful",
+    "4928": "failure",
+    "4929": "failure",
+    "4930": "failure",
+    "4931": "successful",
+    "4932": "successful",
+    "4933": "failure",
+    "4934": "failure",
+    "4935": "failure",
+    "4936": "successful",
+    "4937": "successful",
+    "4938": "infeasible",
+    "4939": "successful",
+    "4940": "successful",
+    "4941": "failure",
+    "4942": "successful",
+    "4943": "failure",
+    "4944": "successful",
+    "4945": "failure",
+    "4946": "failure",
+    "4947": "successful",
+    "4948": "failure",
+    "4949": "successful",
+    "4950": "failure",
+    "4951": "infeasible",
+    "4952": "failure",
+    "4953": "successful",
+    "4954": "successful",
+    "4955": "failure",
+    "4956": "infeasible",
+    "4957": "successful",
+    "4958": "failure",
+    "4959": "successful",
+    "4960": "failure",
+    "4961": "failure",
+    "4962": "failure",
+    "4963": "successful",
+    "4964": "failure",
+    "4965": "failure",
+    "4966": "failure",
+    "4967": "successful",
+    "4968": "successful",
+    "4969": "successful",
+    "4970": "successful",
+    "4971": "failure",
+    "4972": "successful",
+    "4973": "successful",
+    "4974": "successful",
+    "4975": "successful",
+    "4976": "failure",
+    "4977": "successful",
+    "4978": "failure",
+    "4979": "failure",
+    "4980": "successful",
+    "4981": "successful",
+    "4982": "successful",
+    "4983": "successful",
+    "4984": "failure",
+    "4985": "failure",
+    "4986": "successful",
+    "4987": "successful",
+    "4988": "successful",
+    "4989": "failure",
+    "4990": "failure",
+    "4991": "successful",
+    "4992": "failure",
+    "4993": "successful",
+    "4994": "successful",
+    "4995": "successful",
+    "4996": "successful",
+    "4997": "successful",
+    "4998": "failure",
+    "4999": "successful",
+    "5000": "failure",
+    "5001": "failure",
+    "5002": "failure",
+    "5003": "successful",
+    "5004": "failure",
+    "5005": "successful",
+    "5006": "successful",
+    "5007": "failure",
+    "5008": "failure",
+    "5009": "already_complete",
+    "5010": "failure",
+    "5011": "failure",
+    "5012": "successful",
+    "5013": "failure",
+    "5014": "successful",
+    "5015": "successful",
+    "5016": "successful",
+    "5017": "successful",
+    "5018": "failure",
+    "5019": "successful",
+    "5020": "successful",
+    "5021": "failure",
+    "5022": "failure",
+    "5023": "failure",
+    "5024": "already_complete",
+    "5025": "successful",
+    "5026": "failure",
+    "5027": "infeasible",
+    "5028": "successful",
+    "5029": "failure",
+    "5030": "successful",
+    "5031": "successful",
+    "5032": "failure",
+    "5033": "successful",
+    "5034": "failure",
+    "5035": "failure",
+    "5036": "failure",
+    "5037": "successful",
+    "5038": "failure",
+    "5039": "successful",
+    "5040": "successful",
+    "5041": "successful",
+    "5042": "successful",
+    "5043": "successful",
+    "5044": "failure",
+    "5045": "failure",
+    "5046": "failure",
+    "5047": "successful",
+    "5048": "failure",
+    "5049": "failure",
+    "5050": "failure",
+    "5051": "successful",
+    "5052": "successful",
+    "5053": "failure",
+    "5054": "failure",
+    "5055": "infeasible",
+    "5056": "failure",
+    "5057": "failure",
+    "5058": "failure",
+    "5059": "failure",
+    "5060": "successful",
+    "5061": "failure",
+    "5062": "successful",
+    "5063": "failure",
+    "5064": "successful",
+    "5065": "failure",
+    "5066": "failure",
+    "5067": "successful",
+    "5068": "successful",
+    "5069": "failure",
+    "5070": "successful",
+    "5071": "successful",
+    "5072": "failure",
+    "5073": "successful",
+    "5074": "failure",
+    "5075": "already_complete",
+    "5076": "failure",
+    "5077": "successful",
+    "5078": "failure",
+    "5079": "failure",
+    "5080": "failure",
+    "5081": "failure",
+    "5082": "failure",
+    "5083": "successful",
+    "5084": "failure",
+    "5085": "failure",
+    "5086": "failure",
+    "5087": "successful",
+    "5088": "successful",
+    "5089": "successful",
+    "5090": "failure",
+    "5091": "failure",
+    "5092": "successful",
+    "5093": "failure",
+    "5094": "successful",
+    "5095": "successful",
+    "5096": "successful",
+    "5097": "failure",
+    "5098": "successful",
+    "5099": "failure",
+    "5100": "failure",
+    "5101": "successful",
+    "5102": "successful",
+    "5103": "failure",
+    "5104": "successful",
+    "5105": "infeasible",
+    "5106": "successful",
+    "5107": "successful",
+    "5108": "failure",
+    "5109": "failure",
+    "5110": "successful",
+    "5111": "failure",
+    "5112": "successful",
+    "5113": "successful",
+    "5114": "failure",
+    "5115": "successful",
+    "5116": "successful",
+    "5117": "successful",
+    "5118": "failure",
+    "5119": "failure",
+    "5120": "failure",
+    "5121": "successful",
+    "5122": "failure",
+    "5123": "failure",
+    "5124": "successful",
+    "5125": "failure",
+    "5126": "successful",
+    "5127": "failure",
+    "5128": "failure",
+    "5129": "successful",
+    "5130": "failure",
+    "5131": "failure",
+    "5132": "successful",
+    "5133": "successful",
+    "5134": "successful",
+    "5135": "failure",
+    "5136": "successful",
+    "5137": "failure",
+    "5138": "failure",
+    "5139": "infeasible",
+    "5140": "successful",
+    "5141": "failure",
+    "5142": "successful",
+    "5143": "successful",
+    "5144": "successful",
+    "5145": "failure",
+    "5146": "successful",
+    "5147": "failure",
+    "5148": "failure",
+    "5149": "already_complete",
+    "5150": "failure",
+    "5151": "successful",
+    "5152": "successful",
+    "5153": "failure",
+    "5154": "failure",
+    "5155": "successful",
+    "5156": "failure",
+    "5157": "already_complete",
+    "5158": "successful",
+    "5159": "failure",
+    "5160": "successful",
+    "5161": "successful",
+    "5162": "successful",
+    "5163": "failure",
+    "5164": "failure",
+    "5165": "failure",
+    "5166": "failure",
+    "5167": "successful",
+    "5168": "failure",
+    "5169": "failure",
+    "5170": "failure",
+    "5171": "failure",
+    "5172": "failure",
+    "5173": "successful",
+    "5174": "infeasible",
+    "5175": "failure",
+    "5176": "failure",
+    "5177": "successful",
+    "5178": "failure",
+    "5179": "failure",
+    "5180": "failure",
+    "5181": "successful",
+    "5182": "failure",
+    "5183": "successful",
+    "5184": "failure",
+    "5185": "failure",
+    "5186": "successful",
+    "5187": "successful",
+    "5188": "already_complete",
+    "5189": "failure",
+    "5190": "successful",
+    "5191": "failure",
+    "5192": "successful",
+    "5193": "failure",
+    "5194": "failure",
+    "5195": "failure",
+    "5196": "successful",
+    "5197": "successful",
+    "5198": "successful",
+    "5199": "failure",
+    "5200": "successful",
+    "5201": "successful",
+    "5202": "successful",
+    "5203": "failure",
+    "5204": "failure",
+    "5205": "failure",
+    "5206": "successful",
+    "5207": "successful",
+    "5208": "successful",
+    "5209": "successful",
+    "5210": "failure",
+    "5211": "successful",
+    "5212": "failure",
+    "5213": "failure",
+    "5214": "failure",
+    "5215": "successful",
+    "5216": "failure",
+    "5217": "failure",
+    "5218": "failure",
+    "5219": "failure",
+    "5220": "failure",
+    "5221": "successful",
+    "5222": "failure",
+    "5223": "successful",
+    "5224": "successful",
+    "5225": "successful",
+    "5226": "already_complete",
+    "5227": "successful",
+    "5228": "failure",
+    "5229": "successful",
+    "5230": "successful",
+    "5231": "failure",
+    "5232": "failure",
+    "5233": "failure",
+    "5234": "failure",
+    "5235": "failure",
+    "5236": "failure",
+    "5237": "failure",
+    "5238": "successful",
+    "5239": "failure",
+    "5240": "successful",
+    "5241": "successful",
+    "5242": "successful",
+    "5243": "successful",
+    "5244": "failure",
+    "5245": "failure",
+    "5246": "successful",
+    "5247": "successful",
+    "5248": "failure",
+    "5249": "failure",
+    "5250": "failure",
+    "5251": "successful",
+    "5252": "already_complete",
+    "5253": "failure",
+    "5254": "successful",
+    "5255": "successful",
+    "5256": "failure",
+    "5257": "successful",
+    "5258": "failure",
+    "5259": "failure",
+    "5260": "failure",
+    "5261": "successful",
+    "5262": "successful",
+    "5263": "failure",
+    "5264": "failure",
+    "5265": "failure",
+    "5266": "failure",
+    "5267": "failure",
+    "5268": "successful",
+    "5269": "failure",
+    "5270": "successful",
+    "5271": "failure",
+    "5272": "successful",
+    "5273": "failure",
+    "5274": "failure",
+    "5275": "successful",
+    "5276": "failure",
+    "5277": "infeasible",
+    "5278": "successful",
+    "5279": "successful",
+    "5280": "failure",
+    "5281": "failure",
+    "5282": "successful",
+    "5283": "failure",
+    "5284": "failure",
+    "5285": "already_complete",
+    "5286": "successful",
+    "5287": "successful",
+    "5288": "successful",
+    "5289": "failure",
+    "5290": "infeasible",
+    "5291": "failure",
+    "5292": "failure",
+    "5293": "successful",
+    "5294": "failure",
+    "5295": "failure",
+    "5296": "failure",
+    "5297": "successful",
+    "5298": "successful",
+    "5299": "already_complete",
+    "5300": "failure",
+    "5301": "already_complete",
+    "5302": "successful",
+    "5303": "failure",
+    "5304": "successful",
+    "5305": "already_complete",
+    "5306": "failure",
+    "5307": "failure",
+    "5308": "successful",
+    "5309": "failure",
+    "5310": "failure",
+    "5311": "successful",
+    "5312": "failure",
+    "5313": "successful",
+    "5314": "successful",
+    "5315": "failure",
+    "5316": "successful",
+    "5317": "successful",
+    "5318": "successful",
+    "5319": "failure",
+    "5320": "successful",
+    "5321": "failure",
+    "5322": "failure",
+    "5323": "failure",
+    "5324": "successful",
+    "5325": "successful",
+    "5326": "failure",
+    "5327": "successful",
+    "5328": "successful",
+    "5329": "successful",
+    "5330": "failure",
+    "5331": "failure",
+    "5332": "successful",
+    "5333": "successful",
+    "5334": "failure",
+    "5335": "successful",
+    "5336": "failure",
+    "5337": "failure",
+    "5338": "successful",
+    "5339": "successful",
+    "5340": "failure",
+    "5341": "failure",
+    "5342": "successful",
+    "5343": "failure",
+    "5344": "failure",
+    "5345": "failure",
+    "5346": "successful",
+    "5347": "successful",
+    "5348": "failure",
+    "5349": "failure",
+    "5350": "failure",
+    "5351": "successful",
+    "5352": "successful",
+    "5353": "failure",
+    "5354": "failure",
+    "5355": "failure",
+    "5356": "successful",
+    "5357": "failure",
+    "5358": "failure",
+    "5359": "successful",
+    "5360": "successful",
+    "5361": "failure",
+    "5362": "failure",
+    "5363": "failure",
+    "5364": "failure",
+    "5365": "successful",
+    "5366": "successful",
+    "5367": "failure",
+    "5368": "successful",
+    "5369": "failure",
+    "5370": "successful",
+    "5371": "successful",
+    "5372": "failure",
+    "5373": "successful",
+    "5374": "failure",
+    "5375": "failure",
+    "5376": "failure",
+    "5377": "successful",
+    "5378": "successful",
+    "5379": "successful",
+    "5380": "successful",
+    "5381": "failure",
+    "5382": "failure",
+    "5383": "failure",
+    "5384": "failure",
+    "5385": "successful",
+    "5386": "successful",
+    "5387": "failure",
+    "5388": "failure",
+    "5389": "successful",
+    "5390": "successful",
+    "5391": "failure",
+    "5392": "failure",
+    "5393": "successful",
+    "5394": "failure",
+    "5395": "failure",
+    "5396": "failure",
+    "5397": "successful",
+    "5398": "failure",
+    "5399": "successful",
+    "5400": "failure",
+    "5401": "failure",
+    "5402": "failure",
+    "5403": "successful",
+    "5404": "successful",
+    "5405": "failure",
+    "5406": "successful",
+    "5407": "failure",
+    "5408": "successful",
+    "5409": "successful",
+    "5410": "failure",
+    "5411": "failure",
+    "5412": "successful",
+    "5413": "successful",
+    "5414": "infeasible",
+    "5415": "failure",
+    "5416": "failure",
+    "5417": "successful",
+    "5418": "successful",
+    "5419": "failure",
+    "5420": "successful",
+    "5421": "successful",
+    "5422": "successful",
+    "5423": "successful",
+    "5424": "failure",
+    "5425": "failure",
+    "5426": "successful",
+    "5427": "successful",
+    "5428": "failure",
+    "5429": "successful",
+    "5430": "failure",
+    "5431": "failure",
+    "5432": "failure",
+    "5433": "failure",
+    "5434": "failure",
+    "5435": "failure",
+    "5436": "failure",
+    "5437": "successful",
+    "5438": "failure",
+    "5439": "successful",
+    "5440": "failure",
+    "5441": "successful",
+    "5442": "failure",
+    "5443": "successful",
+    "5444": "successful",
+    "5445": "successful",
+    "5446": "failure",
+    "5447": "failure",
+    "5448": "successful",
+    "5449": "failure",
+    "5450": "failure",
+    "5451": "failure",
+    "5452": "successful",
+    "5453": "failure",
+    "5454": "successful",
+    "5455": "failure",
+    "5456": "successful",
+    "5457": "failure",
+    "5458": "infeasible",
+    "5459": "failure",
+    "5460": "failure",
+    "5461": "successful",
+    "5462": "successful",
+    "5463": "failure",
+    "5464": "successful",
+    "5465": "failure",
+    "5466": "successful",
+    "5467": "failure",
+    "5468": "successful",
+    "5469": "successful",
+    "5470": "infeasible",
+    "5471": "failure",
+    "5472": "successful",
+    "5473": "failure",
+    "5474": "successful",
+    "5475": "failure",
+    "5476": "failure",
+    "5477": "failure",
+    "5478": "failure",
+    "5479": "successful",
+    "5480": "successful",
+    "5481": "successful",
+    "5482": "failure",
+    "5483": "failure",
+    "5484": "successful",
+    "5485": "failure",
+    "5486": "successful",
+    "5487": "failure",
+    "5488": "successful",
+    "5489": "successful",
+    "5490": "failure",
+    "5491": "failure",
+    "5492": "failure",
+    "5493": "failure",
+    "5494": "failure",
+    "5495": "failure",
+    "5496": "failure",
+    "5497": "successful",
+    "5498": "infeasible",
+    "5499": "successful",
+    "5500": "successful",
+    "5501": "failure",
+    "5502": "failure",
+    "5503": "successful",
+    "5504": "failure",
+    "5505": "failure",
+    "5506": "successful",
+    "5507": "failure",
+    "5508": "successful",
+    "5509": "failure",
+    "5510": "failure",
+    "5511": "failure",
+    "5512": "failure",
+    "5513": "successful",
+    "5514": "failure",
+    "5515": "successful",
+    "5516": "failure",
+    "5517": "failure",
+    "5518": "successful",
+    "5519": "successful",
+    "5520": "failure",
+    "5521": "successful",
+    "5522": "successful",
+    "5523": "failure",
+    "5524": "failure",
+    "5525": "successful",
+    "5526": "failure",
+    "5527": "successful",
+    "5528": "successful",
+    "5529": "failure",
+    "5530": "successful",
+    "5531": "successful",
+    "5532": "failure",
+    "5533": "failure",
+    "5534": "failure",
+    "5535": "failure",
+    "5536": "failure",
+    "5537": "successful",
+    "5538": "failure",
+    "5539": "failure",
+    "5540": "failure",
+    "5541": "successful",
+    "5542": "failure",
+    "5543": "failure",
+    "5544": "successful",
+    "5545": "failure",
+    "5546": "successful",
+    "5547": "successful",
+    "5548": "failure",
+    "5549": "failure",
+    "5550": "infeasible",
+    "5551": "successful",
+    "5552": "failure",
+    "5553": "failure",
+    "5554": "successful",
+    "5555": "failure",
+    "5556": "successful",
+    "5557": "failure",
+    "5558": "successful",
+    "5559": "successful",
+    "5560": "successful",
+    "5561": "successful",
+    "5562": "failure",
+    "5563": "successful",
+    "5564": "failure",
+    "5565": "successful",
+    "5566": "failure",
+    "5567": "failure",
+    "5568": "failure",
+    "5569": "failure",
+    "5570": "successful",
+    "5571": "successful",
+    "5572": "successful",
+    "5573": "failure",
+    "5574": "failure",
+    "5575": "failure",
+    "5576": "successful",
+    "5577": "infeasible",
+    "5578": "failure",
+    "5579": "successful",
+    "5580": "successful",
+    "5581": "failure",
+    "5582": "successful",
+    "5583": "failure",
+    "5584": "failure",
+    "5585": "successful",
+    "5586": "failure",
+    "5587": "failure",
+    "5588": "successful",
+    "5589": "failure",
+    "5590": "failure",
+    "5591": "successful",
+    "5592": "failure",
+    "5593": "failure",
+    "5594": "failure",
+    "5595": "failure",
+    "5596": "failure",
+    "5597": "failure",
+    "5598": "successful",
+    "5599": "successful",
+    "5600": "successful",
+    "5601": "already_complete",
+    "5602": "successful",
+    "5603": "successful",
+    "5604": "successful",
+    "5605": "failure",
+    "5606": "failure",
+    "5607": "already_complete",
+    "5608": "successful",
+    "5609": "failure",
+    "5610": "failure",
+    "5611": "successful",
+    "5612": "successful",
+    "5613": "successful",
+    "5614": "successful",
+    "5615": "failure",
+    "5616": "failure",
+    "5617": "already_complete",
+    "5618": "failure",
+    "5619": "failure",
+    "5620": "failure",
+    "5621": "failure",
+    "5622": "successful",
+    "5623": "successful",
+    "5624": "failure",
+    "5625": "failure",
+    "5626": "successful",
+    "5627": "successful",
+    "5628": "failure",
+    "5629": "successful",
+    "5630": "failure",
+    "5631": "successful",
+    "5632": "already_complete",
+    "5633": "failure",
+    "5634": "failure",
+    "5635": "failure",
+    "5636": "failure",
+    "5637": "failure",
+    "5638": "failure",
+    "5639": "successful",
+    "5640": "successful",
+    "5641": "failure",
+    "5642": "successful",
+    "5643": "successful",
+    "5644": "successful",
+    "5645": "successful",
+    "5646": "failure",
+    "5647": "failure",
+    "5648": "successful",
+    "5649": "successful",
+    "5650": "failure",
+    "5651": "successful",
+    "5652": "successful",
+    "5653": "failure",
+    "5654": "failure",
+    "5655": "failure",
+    "5656": "failure",
+    "5657": "successful",
+    "5658": "failure",
+    "5659": "failure",
+    "5660": "successful",
+    "5661": "successful",
+    "5662": "successful",
+    "5663": "failure",
+    "5664": "infeasible",
+    "5665": "failure",
+    "5666": "failure",
+    "5667": "failure",
+    "5668": "failure",
+    "5669": "successful",
+    "5670": "already_complete",
+    "5671": "successful",
+    "5672": "successful",
+    "5673": "successful",
+    "5674": "failure",
+    "5675": "failure",
+    "5676": "failure",
+    "5677": "successful",
+    "5678": "successful",
+    "5679": "failure",
+    "5680": "failure",
+    "5681": "failure",
+    "5682": "failure",
+    "5683": "successful",
+    "5684": "successful",
+    "5685": "successful",
+    "5686": "successful",
+    "5687": "successful",
+    "5688": "successful",
+    "5689": "successful",
+    "5690": "successful",
+    "5691": "failure",
+    "5692": "failure",
+    "5693": "failure",
+    "5694": "failure",
+    "5695": "successful",
+    "5696": "successful",
+    "5697": "successful",
+    "5698": "failure",
+    "5699": "failure",
+    "5700": "failure",
+    "5701": "successful",
+    "5702": "successful",
+    "5703": "successful",
+    "5704": "failure",
+    "5705": "successful",
+    "5706": "failure",
+    "5707": "successful",
+    "5708": "failure",
+    "5709": "successful",
+    "5710": "failure",
+    "5711": "failure",
+    "5712": "successful",
+    "5713": "failure",
+    "5714": "successful",
+    "5715": "successful",
+    "5716": "failure",
+    "5717": "successful",
+    "5718": "successful",
+    "5719": "successful",
+    "5720": "failure",
+    "5721": "successful",
+    "5722": "successful",
+    "5723": "infeasible",
+    "5724": "failure",
+    "5725": "failure",
+    "5726": "successful",
+    "5727": "already_complete",
+    "5728": "failure",
+    "5729": "failure",
+    "5730": "successful",
+    "5731": "failure",
+    "5732": "successful",
+    "5733": "successful",
+    "5734": "successful",
+    "5735": "failure",
+    "5736": "infeasible",
+    "5737": "failure",
+    "5738": "successful",
+    "5739": "successful",
+    "5740": "successful",
+    "5741": "successful",
+    "5742": "failure",
+    "5743": "successful",
+    "5744": "already_complete",
+    "5745": "failure",
+    "5746": "successful",
+    "5747": "successful",
+    "5748": "successful",
+    "5749": "failure",
+    "5750": "successful",
+    "5751": "successful",
+    "5752": "failure",
+    "5753": "successful",
+    "5754": "successful",
+    "5755": "failure",
+    "5756": "successful",
+    "5757": "failure",
+    "5758": "successful",
+    "5759": "successful",
+    "5760": "successful",
+    "5761": "failure",
+    "5762": "failure",
+    "5763": "already_complete",
+    "5764": "failure",
+    "5765": "failure",
+    "5766": "successful",
+    "5767": "successful",
+    "5768": "already_complete",
+    "5769": "successful",
+    "5770": "successful",
+    "5771": "failure",
+    "5772": "failure",
+    "5773": "failure",
+    "5774": "successful",
+    "5775": "failure",
+    "5776": "successful",
+    "5777": "failure",
+    "5778": "successful",
+    "5779": "failure",
+    "5780": "failure",
+    "5781": "failure",
+    "5782": "successful",
+    "5783": "successful",
+    "5784": "failure",
+    "5785": "failure",
+    "5786": "successful",
+    "5787": "failure",
+    "5788": "failure",
+    "5789": "failure",
+    "5790": "successful",
+    "5791": "failure",
+    "5792": "successful",
+    "5793": "failure",
+    "5794": "successful",
+    "5795": "failure",
+    "5796": "failure",
+    "5797": "successful",
+    "5798": "successful",
+    "5799": "successful",
+    "5800": "failure",
+    "5801": "successful",
+    "5802": "failure",
+    "5803": "successful",
+    "5804": "failure",
+    "5805": "successful",
+    "5806": "failure",
+    "5807": "failure",
+    "5808": "failure",
+    "5809": "failure",
+    "5810": "failure",
+    "5811": "already_complete",
+    "5812": "failure",
+    "5813": "successful",
+    "5814": "failure",
+    "5815": "successful",
+    "5816": "successful",
+    "5817": "failure",
+    "5818": "failure",
+    "5819": "failure",
+    "5820": "successful",
+    "5821": "failure",
+    "5822": "failure",
+    "5823": "successful",
+    "5824": "successful",
+    "5825": "successful",
+    "5826": "successful",
+    "5827": "successful",
+    "5828": "failure",
+    "5829": "failure",
+    "5830": "successful",
+    "5831": "successful",
+    "5832": "failure",
+    "5833": "successful",
+    "5834": "failure",
+    "5835": "failure",
+    "5836": "failure",
+    "5837": "failure",
+    "5838": "failure",
+    "5839": "failure",
+    "5840": "successful",
+    "5841": "failure",
+    "5842": "successful",
+    "5843": "successful",
+    "5844": "failure",
+    "5845": "successful",
+    "5846": "failure",
+    "5847": "successful",
+    "5848": "already_complete",
+    "5849": "successful",
+    "5850": "failure",
+    "5851": "successful",
+    "5852": "successful",
+    "5853": "failure",
+    "5854": "failure",
+    "5855": "failure",
+    "5856": "failure",
+    "5857": "successful",
+    "5858": "failure",
+    "5859": "successful",
+    "5860": "failure",
+    "5861": "successful",
+    "5862": "failure",
+    "5863": "failure",
+    "5864": "failure",
+    "5865": "successful",
+    "5866": "failure",
+    "5867": "failure",
+    "5868": "successful",
+    "5869": "successful",
+    "5870": "successful",
+    "5871": "failure",
+    "5872": "failure",
+    "5873": "successful",
+    "5874": "failure",
+    "5875": "failure",
+    "5876": "failure",
+    "5877": "successful",
+    "5878": "failure",
+    "5879": "failure",
+    "5880": "successful",
+    "5881": "successful",
+    "5882": "successful",
+    "5883": "failure",
+    "5884": "successful",
+    "5885": "failure",
+    "5886": "failure",
+    "5887": "successful",
+    "5888": "failure",
+    "5889": "failure",
+    "5890": "failure",
+    "5891": "failure",
+    "5892": "successful",
+    "5893": "failure",
+    "5894": "successful",
+    "5895": "successful",
+    "5896": "already_complete",
+    "5897": "successful",
+    "5898": "successful",
+    "5899": "successful",
+    "5900": "failure",
+    "5901": "failure",
+    "5902": "successful",
+    "5903": "failure",
+    "5904": "failure",
+    "5905": "successful",
+    "5906": "failure",
+    "5907": "infeasible",
+    "5908": "failure",
+    "5909": "successful",
+    "5910": "failure",
+    "5911": "successful",
+    "5912": "failure",
+    "5913": "failure",
+    "5914": "failure",
+    "5915": "failure",
+    "5916": "successful",
+    "5917": "successful",
+    "5918": "failure",
+    "5919": "failure",
+    "5920": "failure",
+    "5921": "successful",
+    "5922": "infeasible",
+    "5923": "failure",
+    "5924": "failure",
+    "5925": "failure",
+    "5926": "successful",
+    "5927": "failure",
+    "5928": "successful",
+    "5929": "failure",
+    "5930": "successful",
+    "5931": "failure",
+    "5932": "successful",
+    "5933": "failure",
+    "5934": "successful",
+    "5935": "successful",
+    "5936": "failure",
+    "5937": "failure",
+    "5938": "failure",
+    "5939": "failure",
+    "5940": "failure",
+    "5941": "successful",
+    "5942": "failure",
+    "5943": "failure",
+    "5944": "failure",
+    "5945": "failure",
+    "5946": "successful",
+    "5947": "failure",
+    "5948": "successful",
+    "5949": "failure",
+    "5950": "successful",
+    "5951": "successful",
+    "5952": "successful",
+    "5953": "failure",
+    "5954": "failure",
+    "5955": "successful",
+    "5956": "successful",
+    "5957": "failure",
+    "5958": "successful",
+    "5959": "successful",
+    "5960": "successful",
+    "5961": "successful",
+    "5962": "failure",
+    "5963": "successful",
+    "5964": "successful",
+    "5965": "failure",
+    "5966": "failure",
+    "5967": "failure",
+    "5968": "failure",
+    "5969": "failure",
+    "5970": "successful",
+    "5971": "failure",
+    "5972": "successful",
+    "5973": "successful",
+    "5974": "infeasible",
+    "5975": "successful",
+    "5976": "successful",
+    "5977": "successful",
+    "5978": "failure",
+    "5979": "successful",
+    "5980": "failure",
+    "5981": "successful",
+    "5982": "failure",
+    "5983": "successful",
+    "5984": "successful",
+    "5985": "failure",
+    "5986": "failure",
+    "5987": "successful",
+    "5988": "successful",
+    "5989": "successful",
+    "5990": "successful",
+    "5991": "failure",
+    "5992": "successful",
+    "5993": "failure",
+    "5994": "failure",
+    "5995": "successful",
+    "5996": "failure",
+    "5997": "failure",
+    "5998": "failure",
+    "5999": "successful",
+    "6000": "failure",
+    "6001": "successful",
+    "6002": "successful",
+    "6003": "successful",
+    "6004": "successful",
+    "6005": "failure",
+    "6006": "successful",
+    "6007": "successful",
+    "6008": "successful",
+    "6009": "failure",
+    "6010": "failure",
+    "6011": "failure",
+    "6012": "failure",
+    "6013": "failure",
+    "6014": "failure",
+    "6015": "failure",
+    "6016": "failure",
+    "6017": "successful",
+    "6018": "successful",
+    "6019": "failure",
+    "6020": "successful",
+    "6021": "failure",
+    "6022": "successful",
+    "6023": "successful",
+    "6024": "successful",
+    "6025": "failure",
+    "6026": "successful",
+    "6027": "successful",
+    "6028": "successful",
+    "6029": "successful",
+    "6030": "failure",
+    "6031": "successful",
+    "6032": "failure",
+    "6033": "failure",
+    "6034": "failure",
+    "6035": "failure",
+    "6036": "successful",
+    "6037": "failure",
+    "6038": "successful",
+    "6039": "failure",
+    "6040": "failure",
+    "6041": "failure",
+    "6042": "failure",
+    "6043": "successful",
+    "6044": "successful",
+    "6045": "failure",
+    "6046": "failure",
+    "6047": "successful",
+    "6048": "failure",
+    "6049": "successful",
+    "6050": "failure",
+    "6051": "successful",
+    "6052": "failure",
+    "6053": "failure",
+    "6054": "successful",
+    "6055": "already_complete",
+    "6056": "successful",
+    "6057": "successful",
+    "6058": "successful",
+    "6059": "failure",
+    "6060": "failure",
+    "6061": "successful",
+    "6062": "successful",
+    "6063": "successful",
+    "6064": "failure",
+    "6065": "successful",
+    "6066": "successful",
+    "6067": "infeasible",
+    "6068": "failure",
+    "6069": "failure",
+    "6070": "failure",
+    "6071": "failure",
+    "6072": "failure",
+    "6073": "successful",
+    "6074": "successful",
+    "6075": "failure",
+    "6076": "failure",
+    "6077": "successful",
+    "6078": "failure",
+    "6079": "failure",
+    "6080": "successful",
+    "6081": "failure",
+    "6082": "successful",
+    "6083": "failure",
+    "6084": "successful",
+    "6085": "successful",
+    "6086": "failure",
+    "6087": "successful",
+    "6088": "failure",
+    "6089": "failure",
+    "6090": "successful",
+    "6091": "successful",
+    "6092": "failure",
+    "6093": "successful",
+    "6094": "failure",
+    "6095": "failure",
+    "6096": "failure",
+    "6097": "successful",
+    "6098": "failure",
+    "6099": "failure",
+    "6100": "failure",
+    "6101": "successful",
+    "6102": "failure",
+    "6103": "successful",
+    "6104": "successful",
+    "6105": "failure",
+    "6106": "successful",
+    "6107": "failure",
+    "6108": "successful",
+    "6109": "successful",
+    "6110": "successful",
+    "6111": "successful",
+    "6112": "failure",
+    "6113": "successful",
+    "6114": "successful",
+    "6115": "successful",
+    "6116": "failure",
+    "6117": "failure",
+    "6118": "failure",
+    "6119": "failure",
+    "6120": "successful",
+    "6121": "failure",
+    "6122": "successful",
+    "6123": "failure",
+    "6124": "failure",
+    "6125": "successful",
+    "6126": "failure",
+    "6127": "successful",
+    "6128": "failure",
+    "6129": "successful",
+    "6130": "failure",
+    "6131": "infeasible",
+    "6132": "successful",
+    "6133": "successful",
+    "6134": "successful",
+    "6135": "infeasible",
+    "6136": "failure",
+    "6137": "successful",
+    "6138": "successful",
+    "6139": "failure",
+    "6140": "failure",
+    "6141": "failure",
+    "6142": "failure",
+    "6143": "successful",
+    "6144": "failure",
+    "6145": "failure",
+    "6146": "failure",
+    "6147": "failure",
+    "6148": "failure",
+    "6149": "successful",
+    "6150": "successful",
+    "6151": "failure",
+    "6152": "failure",
+    "6153": "infeasible",
+    "6154": "failure",
+    "6155": "failure",
+    "6156": "failure",
+    "6157": "infeasible",
+    "6158": "successful",
+    "6159": "failure",
+    "6160": "successful",
+    "6161": "already_complete",
+    "6162": "failure",
+    "6163": "failure",
+    "6164": "successful",
+    "6165": "already_complete",
+    "6166": "successful",
+    "6167": "failure",
+    "6168": "failure",
+    "6169": "infeasible",
+    "6170": "successful",
+    "6171": "successful",
+    "6172": "failure",
+    "6173": "successful",
+    "6174": "successful",
+    "6175": "successful",
+    "6176": "successful",
+    "6177": "failure",
+    "6178": "failure",
+    "6179": "successful",
+    "6180": "successful",
+    "6181": "successful",
+    "6182": "already_complete",
+    "6183": "infeasible",
+    "6184": "failure",
+    "6185": "failure",
+    "6186": "successful",
+    "6187": "successful",
+    "6188": "successful",
+    "6189": "successful",
+    "6190": "failure",
+    "6191": "failure",
+    "6192": "successful",
+    "6193": "failure",
+    "6194": "successful",
+    "6195": "successful",
+    "6196": "failure",
+    "6197": "failure",
+    "6198": "successful",
+    "6199": "failure",
+    "6200": "successful",
+    "6201": "successful",
+    "6202": "failure",
+    "6203": "failure",
+    "6204": "failure",
+    "6205": "successful",
+    "6206": "failure",
+    "6207": "successful",
+    "6208": "successful",
+    "6209": "successful",
+    "6210": "successful",
+    "6211": "successful",
+    "6212": "successful",
+    "6213": "successful",
+    "6214": "successful",
+    "6215": "failure",
+    "6216": "failure",
+    "6217": "successful",
+    "6218": "failure",
+    "6219": "successful",
+    "6220": "infeasible",
+    "6221": "successful",
+    "6222": "infeasible",
+    "6223": "failure",
+    "6224": "failure",
+    "6225": "successful",
+    "6226": "failure",
+    "6227": "failure",
+    "6228": "successful",
+    "6229": "already_complete",
+    "6230": "failure",
+    "6231": "successful",
+    "6232": "failure",
+    "6233": "failure",
+    "6234": "failure",
+    "6235": "failure",
+    "6236": "failure",
+    "6237": "successful",
+    "6238": "successful",
+    "6239": "failure",
+    "6240": "failure",
+    "6241": "failure",
+    "6242": "failure",
+    "6243": "successful",
+    "6244": "failure",
+    "6245": "failure",
+    "6246": "successful",
+    "6247": "successful",
+    "6248": "successful",
+    "6249": "successful",
+    "6250": "failure",
+    "6251": "successful",
+    "6252": "failure",
+    "6253": "failure",
+    "6254": "successful",
+    "6255": "failure",
+    "6256": "failure",
+    "6257": "failure",
+    "6258": "failure",
+    "6259": "failure",
+    "6260": "failure",
+    "6261": "failure",
+    "6262": "successful",
+    "6263": "successful",
+    "6264": "successful",
+    "6265": "failure",
+    "6266": "successful",
+    "6267": "failure",
+    "6268": "successful",
+    "6269": "failure",
+    "6270": "failure",
+    "6271": "successful",
+    "6272": "failure",
+    "6273": "successful",
+    "6274": "already_complete",
+    "6275": "successful",
+    "6276": "successful",
+    "6277": "successful",
+    "6278": "successful",
+    "6279": "failure",
+    "6280": "failure",
+    "6281": "failure",
+    "6282": "failure",
+    "6283": "successful",
+    "6284": "successful",
+    "6285": "failure",
+    "6286": "successful",
+    "6287": "failure",
+    "6288": "successful",
+    "6289": "failure",
+    "6290": "successful",
+    "6291": "failure",
+    "6292": "successful",
+    "6293": "failure",
+    "6294": "infeasible",
+    "6295": "failure",
+    "6296": "successful",
+    "6297": "successful",
+    "6298": "successful",
+    "6299": "infeasible",
+    "6300": "failure",
+    "6301": "successful",
+    "6302": "failure",
+    "6303": "successful",
+    "6304": "failure",
+    "6305": "successful",
+    "6306": "successful",
+    "6307": "failure",
+    "6308": "successful",
+    "6309": "failure",
+    "6310": "failure",
+    "6311": "failure",
+    "6312": "successful",
+    "6313": "successful",
+    "6314": "failure",
+    "6315": "successful",
+    "6316": "failure",
+    "6317": "failure",
+    "6318": "failure",
+    "6319": "failure",
+    "6320": "failure",
+    "6321": "failure",
+    "6322": "successful",
+    "6323": "failure",
+    "6324": "failure",
+    "6325": "successful",
+    "6326": "failure",
+    "6327": "successful",
+    "6328": "successful",
+    "6329": "successful",
+    "6330": "successful",
+    "6331": "failure",
+    "6332": "failure",
+    "6333": "failure",
+    "6334": "successful",
+    "6335": "failure",
+    "6336": "successful",
+    "6337": "failure",
+    "6338": "successful",
+    "6339": "failure",
+    "6340": "successful",
+    "6341": "failure",
+    "6342": "failure",
+    "6343": "failure",
+    "6344": "failure",
+    "6345": "successful",
+    "6346": "failure",
+    "6347": "failure",
+    "6348": "failure",
+    "6349": "failure",
+    "6350": "failure",
+    "6351": "failure",
+    "6352": "failure",
+    "6353": "failure",
+    "6354": "successful",
+    "6355": "failure",
+    "6356": "successful",
+    "6357": "failure",
+    "6358": "failure",
+    "6359": "failure",
+    "6360": "failure",
+    "6361": "successful",
+    "6362": "successful",
+    "6363": "failure",
+    "6364": "failure",
+    "6365": "successful",
+    "6366": "successful",
+    "6367": "successful",
+    "6368": "successful",
+    "6369": "infeasible",
+    "6370": "successful",
+    "6371": "failure",
+    "6372": "successful",
+    "6373": "successful",
+    "6374": "successful",
+    "6375": "failure",
+    "6376": "successful",
+    "6377": "failure",
+    "6378": "successful",
+    "6379": "successful",
+    "6380": "failure",
+    "6381": "failure",
+    "6382": "successful",
+    "6383": "failure",
+    "6384": "failure",
+    "6385": "failure",
+    "6386": "failure",
+    "6387": "failure",
+    "6388": "successful",
+    "6389": "failure",
+    "6390": "failure",
+    "6391": "failure",
+    "6392": "successful",
+    "6393": "failure",
+    "6394": "successful",
+    "6395": "successful",
+    "6396": "successful",
+    "6397": "already_complete",
+    "6398": "successful",
+    "6399": "successful",
+    "6400": "failure",
+    "6401": "successful",
+    "6402": "successful",
+    "6403": "successful",
+    "6404": "failure",
+    "6405": "successful",
+    "6406": "successful",
+    "6407": "failure",
+    "6408": "failure",
+    "6409": "successful",
+    "6410": "successful",
+    "6411": "failure",
+    "6412": "successful",
+    "6413": "successful",
+    "6414": "failure",
+    "6415": "failure",
+    "6416": "failure",
+    "6417": "failure",
+    "6418": "failure",
+    "6419": "infeasible",
+    "6420": "successful",
+    "6421": "failure",
+    "6422": "failure",
+    "6423": "successful",
+    "6424": "successful",
+    "6425": "successful",
+    "6426": "successful",
+    "6427": "failure",
+    "6428": "infeasible",
+    "6429": "failure",
+    "6430": "failure",
+    "6431": "successful",
+    "6432": "successful",
+    "6433": "successful",
+    "6434": "failure",
+    "6435": "failure",
+    "6436": "successful",
+    "6437": "successful",
+    "6438": "failure",
+    "6439": "successful",
+    "6440": "successful",
+    "6441": "successful",
+    "6442": "successful",
+    "6443": "successful",
+    "6444": "successful",
+    "6445": "successful",
+    "6446": "failure",
+    "6447": "failure",
+    "6448": "failure",
+    "6449": "successful",
+    "6450": "failure",
+    "6451": "successful",
+    "6452": "successful",
+    "6453": "successful",
+    "6454": "failure",
+    "6455": "failure",
+    "6456": "already_complete",
+    "6457": "failure",
+    "6458": "successful",
+    "6459": "successful",
+    "6460": "failure",
+    "6461": "successful",
+    "6462": "failure",
+    "6463": "successful",
+    "6464": "successful",
+    "6465": "successful",
+    "6466": "failure",
+    "6467": "successful",
+    "6468": "failure",
+    "6469": "successful",
+    "6470": "successful",
+    "6471": "failure",
+    "6472": "successful",
+    "6473": "failure",
+    "6474": "successful",
+    "6475": "successful",
+    "6476": "failure",
+    "6477": "failure",
+    "6478": "successful",
+    "6479": "successful",
+    "6480": "successful",
+    "6481": "failure",
+    "6482": "successful",
+    "6483": "successful",
+    "6484": "successful",
+    "6485": "infeasible",
+    "6486": "successful",
+    "6487": "successful",
+    "6488": "failure",
+    "6489": "failure",
+    "6490": "successful",
+    "6491": "failure",
+    "6492": "successful",
+    "6493": "failure",
+    "6494": "successful",
+    "6495": "successful",
+    "6496": "failure",
+    "6497": "successful",
+    "6498": "successful",
+    "6499": "successful",
+    "6500": "infeasible",
+    "6501": "successful",
+    "6502": "successful",
+    "6503": "successful",
+    "6504": "successful",
+    "6505": "failure",
+    "6506": "successful",
+    "6507": "failure",
+    "6508": "successful",
+    "6509": "infeasible",
+    "6510": "failure",
+    "6511": "successful",
+    "6512": "failure",
+    "6513": "failure",
+    "6514": "successful",
+    "6515": "failure",
+    "6516": "successful",
+    "6517": "successful",
+    "6518": "successful",
+    "6519": "successful",
+    "6520": "failure",
+    "6521": "successful",
+    "6522": "successful",
+    "6523": "successful",
+    "6524": "failure",
+    "6525": "failure",
+    "6526": "already_complete",
+    "6527": "failure",
+    "6528": "failure",
+    "6529": "failure",
+    "6530": "successful",
+    "6531": "successful",
+    "6532": "failure",
+    "6533": "failure",
+    "6534": "failure",
+    "6535": "failure",
+    "6536": "successful",
+    "6537": "successful",
+    "6538": "successful",
+    "6539": "failure",
+    "6540": "infeasible",
+    "6541": "already_complete",
+    "6542": "successful",
+    "6543": "failure",
+    "6544": "successful",
+    "6545": "failure",
+    "6546": "failure",
+    "6547": "successful",
+    "6548": "failure",
+    "6549": "failure",
+    "6550": "failure",
+    "6551": "failure",
+    "6552": "successful",
+    "6553": "successful",
+    "6554": "failure",
+    "6555": "successful",
+    "6556": "infeasible",
+    "6557": "successful",
+    "6558": "successful",
+    "6559": "failure",
+    "6560": "successful",
+    "6561": "successful",
+    "6562": "successful",
+    "6563": "failure",
+    "6564": "successful",
+    "6565": "successful",
+    "6566": "failure",
+    "6567": "successful",
+    "6568": "failure",
+    "6569": "failure",
+    "6570": "already_complete",
+    "6571": "failure",
+    "6572": "successful",
+    "6573": "successful",
+    "6574": "failure",
+    "6575": "failure",
+    "6576": "successful",
+    "6577": "already_complete",
+    "6578": "failure",
+    "6579": "successful",
+    "6580": "successful",
+    "6581": "successful",
+    "6582": "successful",
+    "6583": "already_complete",
+    "6584": "successful",
+    "6585": "failure",
+    "6586": "successful",
+    "6587": "failure",
+    "6588": "already_complete",
+    "6589": "failure",
+    "6590": "failure",
+    "6591": "failure",
+    "6592": "infeasible",
+    "6593": "successful",
+    "6594": "failure",
+    "6595": "successful",
+    "6596": "failure",
+    "6597": "successful",
+    "6598": "successful",
+    "6599": "successful",
+    "6600": "failure",
+    "6601": "successful",
+    "6602": "successful",
+    "6603": "failure",
+    "6604": "failure",
+    "6605": "successful",
+    "6606": "successful",
+    "6607": "already_complete",
+    "6608": "already_complete",
+    "6609": "failure",
+    "6610": "failure",
+    "6611": "failure",
+    "6612": "successful",
+    "6613": "failure",
+    "6614": "failure",
+    "6615": "successful",
+    "6616": "failure",
+    "6617": "failure",
+    "6618": "successful",
+    "6619": "failure",
+    "6620": "successful",
+    "6621": "failure",
+    "6622": "failure",
+    "6623": "successful",
+    "6624": "failure",
+    "6625": "successful",
+    "6626": "failure",
+    "6627": "failure",
+    "6628": "failure",
+    "6629": "successful",
+    "6630": "successful",
+    "6631": "infeasible",
+    "6632": "successful",
+    "6633": "successful",
+    "6634": "failure",
+    "6635": "successful",
+    "6636": "failure",
+    "6637": "failure",
+    "6638": "failure",
+    "6639": "failure",
+    "6640": "successful",
+    "6641": "successful",
+    "6642": "failure",
+    "6643": "successful",
+    "6644": "infeasible",
+    "6645": "failure",
+    "6646": "successful",
+    "6647": "failure",
+    "6648": "failure",
+    "6649": "failure",
+    "6650": "failure",
+    "6651": "infeasible",
+    "6652": "successful",
+    "6653": "failure",
+    "6654": "failure",
+    "6655": "successful",
+    "6656": "successful",
+    "6657": "failure",
+    "6658": "failure",
+    "6659": "failure",
+    "6660": "failure",
+    "6661": "successful",
+    "6662": "infeasible",
+    "6663": "failure",
+    "6664": "successful",
+    "6665": "failure",
+    "6666": "failure",
+    "6667": "failure",
+    "6668": "failure",
+    "6669": "failure",
+    "6670": "failure",
+    "6671": "successful",
+    "6672": "failure",
+    "6673": "failure",
+    "6674": "successful",
+    "6675": "failure",
+    "6676": "failure",
+    "6677": "already_complete",
+    "6678": "already_complete",
+    "6679": "successful",
+    "6680": "failure",
+    "6681": "successful",
+    "6682": "successful",
+    "6683": "failure",
+    "6684": "failure",
+    "6685": "failure",
+    "6686": "successful",
+    "6687": "successful",
+    "6688": "successful",
+    "6689": "successful",
+    "6690": "successful",
+    "6691": "failure",
+    "6692": "failure",
+    "6693": "failure",
+    "6694": "successful",
+    "6695": "failure",
+    "6696": "failure",
+    "6697": "already_complete",
+    "6698": "failure",
+    "6699": "failure",
+    "6700": "failure",
+    "6701": "successful",
+    "6702": "failure",
+    "6703": "failure",
+    "6704": "failure",
+    "6705": "successful",
+    "6706": "failure",
+    "6707": "failure",
+    "6708": "successful",
+    "6709": "successful",
+    "6710": "successful",
+    "6711": "failure",
+    "6712": "failure",
+    "6713": "successful",
+    "6714": "successful",
+    "6715": "failure",
+    "6716": "failure",
+    "6717": "successful",
+    "6718": "successful",
+    "6719": "failure",
+    "6720": "failure",
+    "6721": "infeasible",
+    "6722": "successful",
+    "6723": "successful",
+    "6724": "failure",
+    "6725": "failure",
+    "6726": "failure",
+    "6727": "successful",
+    "6728": "successful",
+    "6729": "successful",
+    "6730": "successful",
+    "6731": "already_complete",
+    "6732": "failure",
+    "6733": "failure",
+    "6734": "failure",
+    "6735": "infeasible",
+    "6736": "successful",
+    "6737": "failure",
+    "6738": "failure",
+    "6739": "failure",
+    "6740": "failure",
+    "6741": "failure",
+    "6742": "failure",
+    "6743": "already_complete",
+    "6744": "failure",
+    "6745": "failure",
+    "6746": "failure",
+    "6747": "failure",
+    "6748": "successful",
+    "6749": "successful",
+    "6750": "infeasible",
+    "6751": "successful",
+    "6752": "successful",
+    "6753": "successful",
+    "6754": "failure",
+    "6755": "successful",
+    "6756": "successful",
+    "6757": "failure",
+    "6758": "successful",
+    "6759": "failure",
+    "6760": "successful",
+    "6761": "successful",
+    "6762": "successful",
+    "6763": "failure",
+    "6764": "successful",
+    "6765": "successful",
+    "6766": "failure",
+    "6767": "successful",
+    "6768": "successful",
+    "6769": "successful",
+    "6770": "failure",
+    "6771": "failure",
+    "6772": "successful",
+    "6773": "successful",
+    "6774": "successful",
+    "6775": "successful",
+    "6776": "failure",
+    "6777": "successful",
+    "6778": "successful",
+    "6779": "failure",
+    "6780": "successful",
+    "6781": "failure",
+    "6782": "failure",
+    "6783": "failure",
+    "6784": "failure",
+    "6785": "failure",
+    "6786": "successful",
+    "6787": "successful",
+    "6788": "failure",
+    "6789": "successful",
+    "6790": "successful",
+    "6791": "failure",
+    "6792": "failure",
+    "6793": "failure",
+    "6794": "infeasible",
+    "6795": "infeasible",
+    "6796": "failure",
+    "6797": "failure",
+    "6798": "successful",
+    "6799": "failure",
+    "6800": "successful",
+    "6801": "failure",
+    "6802": "successful",
+    "6803": "successful",
+    "6804": "failure",
+    "6805": "failure",
+    "6806": "failure",
+    "6807": "successful",
+    "6808": "successful",
+    "6809": "infeasible",
+    "6810": "failure",
+    "6811": "successful",
+    "6812": "failure",
+    "6813": "failure",
+    "6814": "successful",
+    "6815": "failure",
+    "6816": "successful",
+    "6817": "failure",
+    "6818": "successful",
+    "6819": "failure",
+    "6820": "failure",
+    "6821": "successful",
+    "6822": "failure",
+    "6823": "failure",
+    "6824": "infeasible",
+    "6825": "failure",
+    "6826": "successful",
+    "6827": "failure",
+    "6828": "successful",
+    "6829": "failure",
+    "6830": "successful",
+    "6831": "failure",
+    "6832": "failure",
+    "6833": "failure",
+    "6834": "failure",
+    "6835": "already_complete",
+    "6836": "failure",
+    "6837": "failure",
+    "6838": "successful",
+    "6839": "successful",
+    "6840": "successful",
+    "6841": "failure",
+    "6842": "already_complete",
+    "6843": "failure",
+    "6844": "failure",
+    "6845": "successful",
+    "6846": "infeasible",
+    "6847": "failure",
+    "6848": "successful",
+    "6849": "failure",
+    "6850": "failure",
+    "6851": "successful",
+    "6852": "failure",
+    "6853": "failure",
+    "6854": "already_complete",
+    "6855": "successful",
+    "6856": "failure",
+    "6857": "successful",
+    "6858": "failure",
+    "6859": "failure",
+    "6860": "failure",
+    "6861": "failure",
+    "6862": "failure",
+    "6863": "failure",
+    "6864": "successful",
+    "6865": "successful",
+    "6866": "failure",
+    "6867": "successful",
+    "6868": "failure",
+    "6869": "successful",
+    "6870": "failure",
+    "6871": "successful",
+    "6872": "successful",
+    "6873": "failure",
+    "6874": "failure",
+    "6875": "failure",
+    "6876": "successful",
+    "6877": "successful",
+    "6878": "infeasible",
+    "6879": "successful",
+    "6880": "failure",
+    "6881": "successful",
+    "6882": "failure",
+    "6883": "successful",
+    "6884": "failure",
+    "6885": "failure",
+    "6886": "successful",
+    "6887": "successful",
+    "6888": "successful",
+    "6889": "failure",
+    "6890": "failure",
+    "6891": "failure",
+    "6892": "failure",
+    "6893": "failure",
+    "6894": "failure",
+    "6895": "failure",
+    "6896": "successful",
+    "6897": "failure",
+    "6898": "failure",
+    "6899": "failure",
+    "6900": "failure",
+    "6901": "successful",
+    "6902": "failure",
+    "6903": "failure",
+    "6904": "failure",
+    "6905": "successful",
+    "6906": "failure",
+    "6907": "successful",
+    "6908": "successful",
+    "6909": "failure",
+    "6910": "failure",
+    "6911": "failure",
+    "6912": "failure",
+    "6913": "failure",
+    "6914": "already_complete",
+    "6915": "successful",
+    "6916": "successful",
+    "6917": "failure",
+    "6918": "successful",
+    "6919": "failure",
+    "6920": "failure",
+    "6921": "successful",
+    "6922": "failure",
+    "6923": "successful",
+    "6924": "failure",
+    "6925": "infeasible",
+    "6926": "successful",
+    "6927": "failure",
+    "6928": "failure",
+    "6929": "successful",
+    "6930": "failure",
+    "6931": "successful",
+    "6932": "successful",
+    "6933": "successful",
+    "6934": "failure",
+    "6935": "successful",
+    "6936": "failure",
+    "6937": "failure",
+    "6938": "successful",
+    "6939": "successful",
+    "6940": "failure",
+    "6941": "failure",
+    "6942": "successful",
+    "6943": "successful",
+    "6944": "successful",
+    "6945": "failure",
+    "6946": "successful",
+    "6947": "already_complete",
+    "6948": "failure",
+    "6949": "successful",
+    "6950": "failure",
+    "6951": "successful",
+    "6952": "successful",
+    "6953": "successful",
+    "6954": "failure",
+    "6955": "failure",
+    "6956": "successful",
+    "6957": "failure",
+    "6958": "failure",
+    "6959": "failure",
+    "6960": "failure",
+    "6961": "successful",
+    "6962": "failure",
+    "6963": "failure",
+    "6964": "failure",
+    "6965": "successful",
+    "6966": "failure",
+    "6967": "failure",
+    "6968": "successful",
+    "6969": "failure",
+    "6970": "failure",
+    "6971": "successful",
+    "6972": "successful",
+    "6973": "successful",
+    "6974": "already_complete",
+    "6975": "failure",
+    "6976": "successful",
+    "6977": "failure",
+    "6978": "failure",
+    "6979": "failure",
+    "6980": "successful",
+    "6981": "failure",
+    "6982": "successful",
+    "6983": "failure",
+    "6984": "failure",
+    "6985": "failure",
+    "6986": "failure",
+    "6987": "failure",
+    "6988": "successful",
+    "6989": "failure",
+    "6990": "failure",
+    "6991": "failure",
+    "6992": "successful",
+    "6993": "failure",
+    "6994": "successful",
+    "6995": "failure",
+    "6996": "successful",
+    "6997": "successful",
+    "6998": "successful",
+    "6999": "infeasible",
+    "7000": "failure",
+    "7001": "successful",
+    "7002": "successful",
+    "7003": "failure",
+    "7004": "failure",
+    "7005": "successful",
+    "7006": "successful",
+    "7007": "failure",
+    "7008": "failure",
+    "7009": "failure",
+    "7010": "successful",
+    "7011": "failure",
+    "7012": "successful",
+    "7013": "failure",
+    "7014": "successful",
+    "7015": "successful",
+    "7016": "failure",
+    "7017": "failure",
+    "7018": "failure",
+    "7019": "successful",
+    "7020": "successful",
+    "7021": "failure",
+    "7022": "successful",
+    "7023": "failure",
+    "7024": "failure",
+    "7025": "failure",
+    "7026": "failure",
+    "7027": "successful",
+    "7028": "infeasible",
+    "7029": "failure",
+    "7030": "failure",
+    "7031": "failure",
+    "7032": "successful",
+    "7033": "failure",
+    "7034": "failure",
+    "7035": "failure",
+    "7036": "infeasible",
+    "7037": "successful",
+    "7038": "failure",
+    "7039": "failure",
+    "7040": "successful",
+    "7041": "failure",
+    "7042": "successful",
+    "7043": "failure",
+    "7044": "infeasible",
+    "7045": "failure",
+    "7046": "successful",
+    "7047": "successful",
+    "7048": "successful",
+    "7049": "failure",
+    "7050": "failure",
+    "7051": "failure",
+    "7052": "successful",
+    "7053": "already_complete",
+    "7054": "failure",
+    "7055": "failure",
+    "7056": "failure",
+    "7057": "failure",
+    "7058": "already_complete",
+    "7059": "already_complete",
+    "7060": "successful",
+    "7061": "failure",
+    "7062": "successful",
+    "7063": "failure",
+    "7064": "failure",
+    "7065": "failure",
+    "7066": "successful",
+    "7067": "successful",
+    "7068": "failure",
+    "7069": "successful",
+    "7070": "successful",
+    "7071": "successful",
+    "7072": "successful",
+    "7073": "successful",
+    "7074": "failure",
+    "7075": "failure",
+    "7076": "failure",
+    "7077": "failure",
+    "7078": "successful",
+    "7079": "failure",
+    "7080": "successful",
+    "7081": "failure",
+    "7082": "infeasible",
+    "7083": "successful",
+    "7084": "successful",
+    "7085": "successful",
+    "7086": "successful",
+    "7087": "failure",
+    "7088": "successful",
+    "7089": "successful",
+    "7090": "failure",
+    "7091": "failure",
+    "7092": "failure",
+    "7093": "failure",
+    "7094": "failure",
+    "7095": "successful",
+    "7096": "failure",
+    "7097": "successful",
+    "7098": "failure",
+    "7099": "successful",
+    "7100": "failure",
+    "7101": "successful",
+    "7102": "successful",
+    "7103": "successful",
+    "7104": "failure",
+    "7105": "failure",
+    "7106": "failure",
+    "7107": "successful",
+    "7108": "failure",
+    "7109": "failure",
+    "7110": "failure",
+    "7111": "failure",
+    "7112": "successful",
+    "7113": "failure",
+    "7114": "failure",
+    "7115": "failure",
+    "7116": "failure",
+    "7117": "failure",
+    "7118": "failure",
+    "7119": "failure",
+    "7120": "failure",
+    "7121": "successful",
+    "7122": "successful",
+    "7123": "failure",
+    "7124": "infeasible",
+    "7125": "failure",
+    "7126": "successful",
+    "7127": "failure",
+    "7128": "already_complete",
+    "7129": "successful",
+    "7130": "failure",
+    "7131": "successful",
+    "7132": "already_complete",
+    "7133": "successful",
+    "7134": "successful",
+    "7135": "failure",
+    "7136": "successful",
+    "7137": "successful",
+    "7138": "infeasible",
+    "7139": "successful",
+    "7140": "failure",
+    "7141": "successful",
+    "7142": "failure",
+    "7143": "successful",
+    "7144": "successful",
+    "7145": "successful",
+    "7146": "failure",
+    "7147": "successful",
+    "7148": "successful",
+    "7149": "failure",
+    "7150": "successful",
+    "7151": "successful",
+    "7152": "successful",
+    "7153": "failure",
+    "7154": "failure",
+    "7155": "failure",
+    "7156": "failure",
+    "7157": "successful",
+    "7158": "successful",
+    "7159": "failure",
+    "7160": "failure",
+    "7161": "successful",
+    "7162": "failure",
+    "7163": "failure",
+    "7164": "successful",
+    "7165": "successful",
+    "7166": "failure",
+    "7167": "failure",
+    "7168": "failure",
+    "7169": "failure",
+    "7170": "infeasible",
+    "7171": "failure",
+    "7172": "infeasible",
+    "7173": "successful",
+    "7174": "failure",
+    "7175": "successful",
+    "7176": "infeasible",
+    "7177": "failure",
+    "7178": "successful",
+    "7179": "successful",
+    "7180": "failure",
+    "7181": "failure",
+    "7182": "successful",
+    "7183": "successful",
+    "7184": "failure",
+    "7185": "already_complete",
+    "7186": "failure",
+    "7187": "successful",
+    "7188": "successful",
+    "7189": "successful",
+    "7190": "failure",
+    "7191": "successful",
+    "7192": "failure",
+    "7193": "failure",
+    "7194": "successful",
+    "7195": "successful",
+    "7196": "failure",
+    "7197": "failure",
+    "7198": "failure",
+    "7199": "successful",
+    "7200": "successful",
+    "7201": "failure",
+    "7202": "successful",
+    "7203": "failure",
+    "7204": "failure",
+    "7205": "successful",
+    "7206": "failure",
+    "7207": "already_complete",
+    "7208": "successful",
+    "7209": "successful",
+    "7210": "successful",
+    "7211": "successful",
+    "7212": "failure",
+    "7213": "successful",
+    "7214": "failure",
+    "7215": "successful",
+    "7216": "failure",
+    "7217": "successful",
+    "7218": "failure",
+    "7219": "successful",
+    "7220": "successful",
+    "7221": "failure",
+    "7222": "successful",
+    "7223": "failure",
+    "7224": "failure",
+    "7225": "failure",
+    "7226": "failure",
+    "7227": "successful",
+    "7228": "failure",
+    "7229": "successful",
+    "7230": "successful",
+    "7231": "successful",
+    "7232": "successful",
+    "7233": "failure",
+    "7234": "successful",
+    "7235": "failure",
+    "7236": "successful",
+    "7237": "failure",
+    "7238": "failure",
+    "7239": "successful",
+    "7240": "successful",
+    "7241": "failure",
+    "7242": "failure",
+    "7243": "failure",
+    "7244": "failure",
+    "7245": "successful",
+    "7246": "successful",
+    "7247": "failure",
+    "7248": "successful",
+    "7249": "failure",
+    "7250": "failure",
+    "7251": "successful",
+    "7252": "failure",
+    "7253": "failure",
+    "7254": "failure",
+    "7255": "successful",
+    "7256": "successful",
+    "7257": "successful",
+    "7258": "failure",
+    "7259": "failure",
+    "7260": "failure",
+    "7261": "failure",
+    "7262": "failure",
+    "7263": "successful",
+    "7264": "successful",
+    "7265": "successful",
+    "7266": "failure",
+    "7267": "failure",
+    "7268": "failure",
+    "7269": "successful",
+    "7270": "failure",
+    "7271": "successful",
+    "7272": "successful",
+    "7273": "failure",
+    "7274": "failure",
+    "7275": "successful",
+    "7276": "failure",
+    "7277": "failure",
+    "7278": "failure",
+    "7279": "successful",
+    "7280": "failure",
+    "7281": "failure",
+    "7282": "failure",
+    "7283": "failure",
+    "7284": "successful",
+    "7285": "successful",
+    "7286": "successful",
+    "7287": "failure",
+    "7288": "successful",
+    "7289": "failure",
+    "7290": "failure",
+    "7291": "successful",
+    "7292": "failure",
+    "7293": "successful",
+    "7294": "successful",
+    "7295": "successful",
+    "7296": "successful",
+    "7297": "failure",
+    "7298": "failure",
+    "7299": "successful",
+    "7300": "failure",
+    "7301": "failure",
+    "7302": "failure",
+    "7303": "successful",
+    "7304": "failure",
+    "7305": "successful",
+    "7306": "successful",
+    "7307": "successful",
+    "7308": "failure",
+    "7309": "failure",
+    "7310": "failure",
+    "7311": "successful",
+    "7312": "successful",
+    "7313": "failure",
+    "7314": "failure",
+    "7315": "successful",
+    "7316": "successful",
+    "7317": "failure",
+    "7318": "failure",
+    "7319": "failure",
+    "7320": "failure",
+    "7321": "successful",
+    "7322": "infeasible",
+    "7323": "already_complete",
+    "7324": "failure",
+    "7325": "successful",
+    "7326": "successful",
+    "7327": "successful",
+    "7328": "failure",
+    "7329": "failure",
+    "7330": "failure",
+    "7331": "successful",
+    "7332": "failure",
+    "7333": "failure",
+    "7334": "already_complete",
+    "7335": "failure",
+    "7336": "failure",
+    "7337": "failure",
+    "7338": "failure",
+    "7339": "failure",
+    "7340": "failure",
+    "7341": "failure",
+    "7342": "infeasible",
+    "7343": "already_complete",
+    "7344": "failure",
+    "7345": "failure",
+    "7346": "successful",
+    "7347": "successful",
+    "7348": "successful",
+    "7349": "successful",
+    "7350": "successful",
+    "7351": "failure",
+    "7352": "successful",
+    "7353": "failure",
+    "7354": "failure",
+    "7355": "failure",
+    "7356": "successful",
+    "7357": "failure",
+    "7358": "failure",
+    "7359": "failure",
+    "7360": "successful",
+    "7361": "successful",
+    "7362": "failure",
+    "7363": "failure",
+    "7364": "failure",
+    "7365": "failure",
+    "7366": "successful",
+    "7367": "failure",
+    "7368": "successful",
+    "7369": "successful",
+    "7370": "failure",
+    "7371": "failure",
+    "7372": "successful",
+    "7373": "successful",
+    "7374": "successful",
+    "7375": "failure",
+    "7376": "failure",
+    "7377": "failure",
+    "7378": "successful",
+    "7379": "failure",
+    "7380": "successful",
+    "7381": "successful",
+    "7382": "successful",
+    "7383": "failure",
+    "7384": "failure",
+    "7385": "failure",
+    "7386": "successful",
+    "7387": "failure",
+    "7388": "failure",
+    "7389": "infeasible",
+    "7390": "failure",
+    "7391": "successful",
+    "7392": "successful",
+    "7393": "failure",
+    "7394": "successful",
+    "7395": "failure",
+    "7396": "failure",
+    "7397": "successful",
+    "7398": "failure",
+    "7399": "infeasible",
+    "7400": "failure",
+    "7401": "successful",
+    "7402": "failure",
+    "7403": "failure",
+    "7404": "successful",
+    "7405": "failure",
+    "7406": "successful",
+    "7407": "failure",
+    "7408": "successful",
+    "7409": "failure",
+    "7410": "successful",
+    "7411": "infeasible",
+    "7412": "failure",
+    "7413": "failure",
+    "7414": "successful",
+    "7415": "failure",
+    "7416": "failure",
+    "7417": "successful",
+    "7418": "failure",
+    "7419": "successful",
+    "7420": "successful",
+    "7421": "failure",
+    "7422": "failure",
+    "7423": "successful",
+    "7424": "failure",
+    "7425": "failure",
+    "7426": "infeasible",
+    "7427": "successful",
+    "7428": "failure",
+    "7429": "failure",
+    "7430": "failure",
+    "7431": "successful",
+    "7432": "successful",
+    "7433": "successful",
+    "7434": "failure",
+    "7435": "failure",
+    "7436": "failure",
+    "7437": "successful",
+    "7438": "successful",
+    "7439": "successful",
+    "7440": "successful",
+    "7441": "failure",
+    "7442": "failure",
+    "7443": "failure",
+    "7444": "failure",
+    "7445": "successful",
+    "7446": "failure",
+    "7447": "failure",
+    "7448": "successful",
+    "7449": "failure",
+    "7450": "successful",
+    "7451": "successful",
+    "7452": "failure",
+    "7453": "failure",
+    "7454": "infeasible",
+    "7455": "already_complete",
+    "7456": "infeasible",
+    "7457": "successful",
+    "7458": "failure",
+    "7459": "successful",
+    "7460": "infeasible",
+    "7461": "failure",
+    "7462": "successful",
+    "7463": "successful",
+    "7464": "failure",
+    "7465": "failure",
+    "7466": "failure",
+    "7467": "successful",
+    "7468": "failure",
+    "7469": "failure",
+    "7470": "failure",
+    "7471": "successful",
+    "7472": "failure",
+    "7473": "successful",
+    "7474": "successful",
+    "7475": "successful",
+    "7476": "successful",
+    "7477": "failure",
+    "7478": "failure",
+    "7479": "failure",
+    "7480": "infeasible",
+    "7481": "failure",
+    "7482": "successful",
+    "7483": "failure",
+    "7484": "successful",
+    "7485": "failure",
+    "7486": "failure",
+    "7487": "successful",
+    "7488": "failure",
+    "7489": "failure",
+    "7490": "failure",
+    "7491": "already_complete",
+    "7492": "successful",
+    "7493": "successful",
+    "7494": "successful",
+    "7495": "already_complete",
+    "7496": "successful",
+    "7497": "failure",
+    "7498": "failure",
+    "7499": "successful",
+    "7500": "failure",
+    "7501": "failure",
+    "7502": "successful",
+    "7503": "failure",
+    "7504": "successful",
+    "7505": "failure",
+    "7506": "successful",
+    "7507": "infeasible",
+    "7508": "failure",
+    "7509": "successful",
+    "7510": "failure",
+    "7511": "successful",
+    "7512": "successful",
+    "7513": "successful",
+    "7514": "failure",
+    "7515": "successful",
+    "7516": "failure",
+    "7517": "failure",
+    "7518": "successful",
+    "7519": "failure",
+    "7520": "successful",
+    "7521": "failure",
+    "7522": "infeasible",
+    "7523": "failure",
+    "7524": "failure",
+    "7525": "failure",
+    "7526": "failure",
+    "7527": "successful",
+    "7528": "failure",
+    "7529": "successful",
+    "7530": "infeasible",
+    "7531": "failure",
+    "7532": "successful",
+    "7533": "successful",
+    "7534": "failure",
+    "7535": "failure",
+    "7536": "failure",
+    "7537": "failure",
+    "7538": "failure",
+    "7539": "failure",
+    "7540": "failure",
+    "7541": "successful",
+    "7542": "successful",
+    "7543": "successful",
+    "7544": "failure",
+    "7545": "successful",
+    "7546": "failure",
+    "7547": "failure",
+    "7548": "failure",
+    "7549": "failure",
+    "7550": "failure",
+    "7551": "failure",
+    "7552": "already_complete",
+    "7553": "failure",
+    "7554": "failure",
+    "7555": "successful",
+    "7556": "failure",
+    "7557": "successful",
+    "7558": "failure",
+    "7559": "failure",
+    "7560": "failure",
+    "7561": "failure",
+    "7562": "successful",
+    "7563": "failure",
+    "7564": "successful",
+    "7565": "successful",
+    "7566": "failure",
+    "7567": "successful",
+    "7568": "successful",
+    "7569": "failure",
+    "7570": "failure",
+    "7571": "failure",
+    "7572": "failure",
+    "7573": "failure",
+    "7574": "failure",
+    "7575": "successful",
+    "7576": "failure",
+    "7577": "failure",
+    "7578": "failure",
+    "7579": "already_complete",
+    "7580": "successful",
+    "7581": "failure",
+    "7582": "successful",
+    "7583": "failure",
+    "7584": "successful",
+    "7585": "already_complete",
+    "7586": "successful",
+    "7587": "successful",
+    "7588": "successful",
+    "7589": "successful",
+    "7590": "failure",
+    "7591": "failure",
+    "7592": "successful",
+    "7593": "successful",
+    "7594": "failure",
+    "7595": "failure",
+    "7596": "failure",
+    "7597": "successful",
+    "7598": "successful",
+    "7599": "successful",
+    "7600": "successful",
+    "7601": "successful",
+    "7602": "successful",
+    "7603": "successful",
+    "7604": "failure",
+    "7605": "failure",
+    "7606": "failure",
+    "7607": "successful",
+    "7608": "infeasible",
+    "7609": "successful",
+    "7610": "successful",
+    "7611": "failure",
+    "7612": "successful",
+    "7613": "successful",
+    "7614": "successful",
+    "7615": "successful",
+    "7616": "failure",
+    "7617": "successful",
+    "7618": "failure",
+    "7619": "successful",
+    "7620": "successful",
+    "7621": "failure",
+    "7622": "successful",
+    "7623": "successful",
+    "7624": "successful",
+    "7625": "infeasible",
+    "7626": "failure",
+    "7627": "successful",
+    "7628": "failure",
+    "7629": "failure",
+    "7630": "successful",
+    "7631": "failure",
+    "7632": "failure",
+    "7633": "successful",
+    "7634": "infeasible",
+    "7635": "failure",
+    "7636": "successful",
+    "7637": "successful",
+    "7638": "failure",
+    "7639": "failure",
+    "7640": "failure",
+    "7641": "successful",
+    "7642": "failure",
+    "7643": "failure",
+    "7644": "failure",
+    "7645": "successful",
+    "7646": "failure",
+    "7647": "failure",
+    "7648": "failure",
+    "7649": "failure",
+    "7650": "successful",
+    "7651": "successful",
+    "7652": "successful",
+    "7653": "successful",
+    "7654": "successful",
+    "7655": "successful",
+    "7656": "failure",
+    "7657": "failure",
+    "7658": "failure",
+    "7659": "failure",
+    "7660": "successful",
+    "7661": "failure",
+    "7662": "successful",
+    "7663": "successful",
+    "7664": "successful",
+    "7665": "failure",
+    "7666": "failure",
+    "7667": "successful",
+    "7668": "failure",
+    "7669": "failure",
+    "7670": "failure",
+    "7671": "successful",
+    "7672": "successful",
+    "7673": "failure",
+    "7674": "successful",
+    "7675": "failure",
+    "7676": "failure",
+    "7677": "successful",
+    "7678": "failure",
+    "7679": "failure",
+    "7680": "failure",
+    "7681": "failure",
+    "7682": "failure",
+    "7683": "failure",
+    "7684": "successful",
+    "7685": "failure",
+    "7686": "failure",
+    "7687": "failure",
+    "7688": "failure",
+    "7689": "successful",
+    "7690": "failure",
+    "7691": "successful",
+    "7692": "successful",
+    "7693": "failure",
+    "7694": "failure",
+    "7695": "failure",
+    "7696": "failure",
+    "7697": "successful",
+    "7698": "successful",
+    "7699": "successful",
+    "7700": "failure",
+    "7701": "successful",
+    "7702": "failure",
+    "7703": "failure",
+    "7704": "failure",
+    "7705": "failure",
+    "7706": "successful",
+    "7707": "successful",
+    "7708": "successful",
+    "7709": "failure",
+    "7710": "successful",
+    "7711": "failure",
+    "7712": "successful",
+    "7713": "failure",
+    "7714": "failure",
+    "7715": "failure",
+    "7716": "successful",
+    "7717": "successful",
+    "7718": "successful",
+    "7719": "successful",
+    "7720": "failure",
+    "7721": "successful",
+    "7722": "failure",
+    "7723": "failure",
+    "7724": "successful",
+    "7725": "successful",
+    "7726": "successful",
+    "7727": "failure",
+    "7728": "successful",
+    "7729": "successful",
+    "7730": "successful",
+    "7731": "failure",
+    "7732": "successful",
+    "7733": "successful",
+    "7734": "failure",
+    "7735": "failure",
+    "7736": "failure",
+    "7737": "failure",
+    "7738": "successful",
+    "7739": "failure",
+    "7740": "successful",
+    "7741": "successful",
+    "7742": "failure",
+    "7743": "failure",
+    "7744": "failure",
+    "7745": "failure",
+    "7746": "successful",
+    "7747": "failure",
+    "7748": "failure",
+    "7749": "failure",
+    "7750": "successful",
+    "7751": "failure",
+    "7752": "successful",
+    "7753": "failure",
+    "7754": "failure",
+    "7755": "successful",
+    "7756": "successful",
+    "7757": "failure",
+    "7758": "failure",
+    "7759": "successful",
+    "7760": "failure",
+    "7761": "failure",
+    "7762": "failure",
+    "7763": "failure",
+    "7764": "successful",
+    "7765": "infeasible",
+    "7766": "successful",
+    "7767": "failure",
+    "7768": "failure",
+    "7769": "failure",
+    "7770": "successful",
+    "7771": "successful",
+    "7772": "failure",
+    "7773": "failure",
+    "7774": "already_complete",
+    "7775": "successful",
+    "7776": "successful",
+    "7777": "failure",
+    "7778": "failure",
+    "7779": "successful",
+    "7780": "successful",
+    "7781": "successful",
+    "7782": "failure",
+    "7783": "failure",
+    "7784": "failure",
+    "7785": "failure",
+    "7786": "failure",
+    "7787": "successful",
+    "7788": "successful",
+    "7789": "successful",
+    "7790": "successful",
+    "7791": "failure",
+    "7792": "successful",
+    "7793": "successful",
+    "7794": "failure",
+    "7795": "failure",
+    "7796": "successful",
+    "7797": "successful",
+    "7798": "failure",
+    "7799": "failure",
+    "7800": "successful",
+    "7801": "failure",
+    "7802": "successful",
+    "7803": "failure",
+    "7804": "successful",
+    "7805": "failure",
+    "7806": "successful",
+    "7807": "successful",
+    "7808": "infeasible",
+    "7809": "failure",
+    "7810": "failure",
+    "7811": "failure",
+    "7812": "successful",
+    "7813": "successful",
+    "7814": "failure",
+    "7815": "failure",
+    "7816": "failure",
+    "7817": "successful",
+    "7818": "successful",
+    "7819": "successful",
+    "7820": "failure",
+    "7821": "failure",
+    "7822": "failure",
+    "7823": "successful",
+    "7824": "failure",
+    "7825": "successful",
+    "7826": "successful",
+    "7827": "successful",
+    "7828": "failure",
+    "7829": "successful",
+    "7830": "failure",
+    "7831": "failure",
+    "7832": "failure",
+    "7833": "already_complete",
+    "7834": "failure",
+    "7835": "failure",
+    "7836": "successful",
+    "7837": "successful",
+    "7838": "successful",
+    "7839": "failure",
+    "7840": "failure",
+    "7841": "successful",
+    "7842": "failure",
+    "7843": "failure",
+    "7844": "failure",
+    "7845": "failure",
+    "7846": "failure",
+    "7847": "failure",
+    "7848": "failure",
+    "7849": "successful",
+    "7850": "failure",
+    "7851": "successful",
+    "7852": "failure",
+    "7853": "failure",
+    "7854": "failure",
+    "7855": "already_complete",
+    "7856": "failure",
+    "7857": "failure",
+    "7858": "failure",
+    "7859": "failure",
+    "7860": "failure",
+    "7861": "failure",
+    "7862": "failure",
+    "7863": "failure",
+    "7864": "failure",
+    "7865": "failure",
+    "7866": "successful",
+    "7867": "successful",
+    "7868": "successful",
+    "7869": "failure",
+    "7870": "successful",
+    "7871": "failure",
+    "7872": "successful",
+    "7873": "successful",
+    "7874": "failure",
+    "7875": "failure",
+    "7876": "failure",
+    "7877": "successful",
+    "7878": "failure",
+    "7879": "failure",
+    "7880": "successful",
+    "7881": "failure",
+    "7882": "successful",
+    "7883": "successful",
+    "7884": "successful",
+    "7885": "successful",
+    "7886": "infeasible",
+    "7887": "failure",
+    "7888": "successful",
+    "7889": "failure",
+    "7890": "successful",
+    "7891": "failure",
+    "7892": "successful",
+    "7893": "failure",
+    "7894": "failure",
+    "7895": "failure",
+    "7896": "successful",
+    "7897": "successful",
+    "7898": "failure",
+    "7899": "failure",
+    "7900": "successful",
+    "7901": "already_complete",
+    "7902": "failure",
+    "7903": "infeasible",
+    "7904": "successful",
+    "7905": "failure",
+    "7906": "failure",
+    "7907": "failure",
+    "7908": "successful",
+    "7909": "successful",
+    "7910": "successful",
+    "7911": "successful",
+    "7912": "failure",
+    "7913": "successful",
+    "7914": "failure",
+    "7915": "successful",
+    "7916": "failure",
+    "7917": "failure",
+    "7918": "failure",
+    "7919": "failure",
+    "7920": "successful",
+    "7921": "failure",
+    "7922": "successful",
+    "7923": "successful",
+    "7924": "failure",
+    "7925": "successful",
+    "7926": "failure",
+    "7927": "successful",
+    "7928": "successful",
+    "7929": "failure",
+    "7930": "successful",
+    "7931": "successful",
+    "7932": "failure",
+    "7933": "failure",
+    "7934": "successful",
+    "7935": "failure",
+    "7936": "successful",
+    "7937": "failure",
+    "7938": "successful",
+    "7939": "failure",
+    "7940": "failure",
+    "7941": "failure",
+    "7942": "failure",
+    "7943": "successful",
+    "7944": "successful",
+    "7945": "successful",
+    "7946": "failure",
+    "7947": "failure",
+    "7948": "successful",
+    "7949": "successful",
+    "7950": "failure",
+    "7951": "failure",
+    "7952": "failure",
+    "7953": "failure",
+    "7954": "successful",
+    "7955": "successful",
+    "7956": "failure",
+    "7957": "successful",
+    "7958": "failure",
+    "7959": "successful",
+    "7960": "successful",
+    "7961": "successful",
+    "7962": "successful",
+    "7963": "successful",
+    "7964": "successful",
+    "7965": "failure",
+    "7966": "failure",
+    "7967": "successful",
+    "7968": "successful",
+    "7969": "successful",
+    "7970": "already_complete",
+    "7971": "failure",
+    "7972": "failure",
+    "7973": "failure",
+    "7974": "successful",
+    "7975": "failure",
+    "7976": "successful",
+    "7977": "failure",
+    "7978": "failure",
+    "7979": "successful",
+    "7980": "successful",
+    "7981": "failure",
+    "7982": "successful",
+    "7983": "successful",
+    "7984": "failure",
+    "7985": "successful",
+    "7986": "failure",
+    "7987": "failure",
+    "7988": "failure",
+    "7989": "successful",
+    "7990": "successful",
+    "7991": "failure",
+    "7992": "successful",
+    "7993": "successful",
+    "7994": "failure",
+    "7995": "successful",
+    "7996": "failure",
+    "7997": "failure",
+    "7998": "successful",
+    "7999": "successful",
+    "8000": "successful",
+    "8001": "failure",
+    "8002": "successful",
+    "8003": "successful",
+    "8004": "successful",
+    "8005": "failure",
+    "8006": "successful",
+    "8007": "failure",
+    "8008": "successful",
+    "8009": "failure",
+    "8010": "successful",
+    "8011": "failure",
+    "8012": "successful",
+    "8013": "successful",
+    "8014": "successful",
+    "8015": "successful",
+    "8016": "failure",
+    "8017": "failure",
+    "8018": "successful",
+    "8019": "successful",
+    "8020": "failure",
+    "8021": "successful",
+    "8022": "failure",
+    "8023": "failure",
+    "8024": "failure",
+    "8025": "failure",
+    "8026": "failure",
+    "8027": "successful",
+    "8028": "successful",
+    "8029": "failure",
+    "8030": "failure",
+    "8031": "failure",
+    "8032": "failure",
+    "8033": "successful",
+    "8034": "failure",
+    "8035": "successful",
+    "8036": "infeasible",
+    "8037": "successful",
+    "8038": "failure",
+    "8039": "failure",
+    "8040": "failure",
+    "8041": "failure",
+    "8042": "failure",
+    "8043": "successful",
+    "8044": "failure",
+    "8045": "successful",
+    "8046": "failure",
+    "8047": "successful",
+    "8048": "successful",
+    "8049": "failure",
+    "8050": "successful",
+    "8051": "failure",
+    "8052": "failure",
+    "8053": "successful",
+    "8054": "successful",
+    "8055": "failure",
+    "8056": "successful",
+    "8057": "successful",
+    "8058": "failure",
+    "8059": "successful",
+    "8060": "failure",
+    "8061": "successful",
+    "8062": "successful",
+    "8063": "successful",
+    "8064": "infeasible",
+    "8065": "successful",
+    "8066": "successful",
+    "8067": "successful",
+    "8068": "successful",
+    "8069": "failure",
+    "8070": "failure",
+    "8071": "failure",
+    "8072": "successful",
+    "8073": "failure",
+    "8074": "successful",
+    "8075": "successful",
+    "8076": "failure",
+    "8077": "successful",
+    "8078": "failure",
+    "8079": "failure",
+    "8080": "successful",
+    "8081": "successful",
+    "8082": "successful",
+    "8083": "failure",
+    "8084": "failure",
+    "8085": "successful",
+    "8086": "failure",
+    "8087": "successful",
+    "8088": "successful",
+    "8089": "successful",
+    "8090": "failure",
+    "8091": "successful",
+    "8092": "successful",
+    "8093": "failure",
+    "8094": "successful",
+    "8095": "failure",
+    "8096": "successful",
+    "8097": "successful",
+    "8098": "failure",
+    "8099": "failure",
+    "8100": "failure",
+    "8101": "successful",
+    "8102": "successful",
+    "8103": "successful",
+    "8104": "successful",
+    "8105": "failure",
+    "8106": "successful",
+    "8107": "successful",
+    "8108": "successful",
+    "8109": "failure",
+    "8110": "failure",
+    "8111": "successful",
+    "8112": "failure",
+    "8113": "failure",
+    "8114": "successful",
+    "8115": "failure",
+    "8116": "failure",
+    "8117": "failure",
+    "8118": "successful",
+    "8119": "failure",
+    "8120": "successful",
+    "8121": "successful",
+    "8122": "successful",
+    "8123": "failure",
+    "8124": "successful",
+    "8125": "successful",
+    "8126": "successful",
+    "8127": "failure",
+    "8128": "failure",
+    "8129": "successful",
+    "8130": "failure",
+    "8131": "failure",
+    "8132": "failure",
+    "8133": "successful",
+    "8134": "failure",
+    "8135": "already_complete",
+    "8136": "already_complete",
+    "8137": "failure",
+    "8138": "failure",
+    "8139": "successful",
+    "8140": "failure",
+    "8141": "successful",
+    "8142": "successful",
+    "8143": "failure",
+    "8144": "failure",
+    "8145": "failure",
+    "8146": "successful",
+    "8147": "successful",
+    "8148": "successful",
+    "8149": "successful",
+    "8150": "successful",
+    "8151": "successful",
+    "8152": "successful",
+    "8153": "successful",
+    "8154": "successful",
+    "8155": "failure",
+    "8156": "failure",
+    "8157": "failure",
+    "8158": "successful",
+    "8159": "failure",
+    "8160": "successful",
+    "8161": "failure",
+    "8162": "failure",
+    "8163": "failure",
+    "8164": "successful",
+    "8165": "failure",
+    "8166": "failure",
+    "8167": "successful",
+    "8168": "successful",
+    "8169": "successful",
+    "8170": "failure",
+    "8171": "failure",
+    "8172": "failure",
+    "8173": "failure",
+    "8174": "failure",
+    "8175": "successful",
+    "8176": "successful",
+    "8177": "infeasible",
+    "8178": "successful",
+    "8179": "successful",
+    "8180": "failure",
+    "8181": "successful",
+    "8182": "failure",
+    "8183": "failure",
+    "8184": "failure",
+    "8185": "failure",
+    "8186": "failure",
+    "8187": "successful",
+    "8188": "failure",
+    "8189": "successful",
+    "8190": "failure",
+    "8191": "infeasible",
+    "8192": "failure",
+    "8193": "successful",
+    "8194": "failure",
+    "8195": "infeasible",
+    "8196": "successful",
+    "8197": "failure",
+    "8198": "failure",
+    "8199": "failure",
+    "8200": "failure",
+    "8201": "failure",
+    "8202": "failure",
+    "8203": "successful",
+    "8204": "failure",
+    "8205": "successful",
+    "8206": "failure",
+    "8207": "successful",
+    "8208": "failure",
+    "8209": "failure",
+    "8210": "failure",
+    "8211": "failure",
+    "8212": "successful",
+    "8213": "failure",
+    "8214": "successful",
+    "8215": "successful",
+    "8216": "successful",
+    "8217": "failure",
+    "8218": "already_complete",
+    "8219": "failure",
+    "8220": "failure",
+    "8221": "failure",
+    "8222": "successful",
+    "8223": "failure",
+    "8224": "failure",
+    "8225": "successful",
+    "8226": "failure",
+    "8227": "failure",
+    "8228": "failure",
+    "8229": "failure",
+    "8230": "successful",
+    "8231": "successful",
+    "8232": "successful",
+    "8233": "failure",
+    "8234": "failure",
+    "8235": "successful",
+    "8236": "failure",
+    "8237": "failure",
+    "8238": "failure",
+    "8239": "failure",
+    "8240": "successful",
+    "8241": "failure",
+    "8242": "failure",
+    "8243": "successful",
+    "8244": "failure",
+    "8245": "successful",
+    "8246": "successful",
+    "8247": "successful",
+    "8248": "failure",
+    "8249": "successful",
+    "8250": "failure",
+    "8251": "failure",
+    "8252": "successful",
+    "8253": "successful",
+    "8254": "failure",
+    "8255": "failure",
+    "8256": "successful",
+    "8257": "successful",
+    "8258": "successful",
+    "8259": "successful",
+    "8260": "successful",
+    "8261": "infeasible",
+    "8262": "successful",
+    "8263": "failure",
+    "8264": "failure",
+    "8265": "failure",
+    "8266": "failure",
+    "8267": "failure",
+    "8268": "failure",
+    "8269": "successful",
+    "8270": "failure",
+    "8271": "failure",
+    "8272": "failure",
+    "8273": "failure",
+    "8274": "failure",
+    "8275": "successful",
+    "8276": "successful",
+    "8277": "successful",
+    "8278": "failure",
+    "8279": "failure",
+    "8280": "successful",
+    "8281": "failure",
+    "8282": "successful",
+    "8283": "successful",
+    "8284": "successful",
+    "8285": "failure",
+    "8286": "failure",
+    "8287": "successful",
+    "8288": "successful",
+    "8289": "successful",
+    "8290": "failure",
+    "8291": "failure",
+    "8292": "failure",
+    "8293": "failure",
+    "8294": "failure",
+    "8295": "successful",
+    "8296": "successful",
+    "8297": "successful",
+    "8298": "failure",
+    "8299": "successful",
+    "8300": "successful",
+    "8301": "successful",
+    "8302": "already_complete",
+    "8303": "failure",
+    "8304": "failure",
+    "8305": "failure",
+    "8306": "successful",
+    "8307": "failure",
+    "8308": "failure",
+    "8309": "infeasible",
+    "8310": "successful",
+    "8311": "failure",
+    "8312": "infeasible",
+    "8313": "failure",
+    "8314": "failure",
+    "8315": "successful",
+    "8316": "successful",
+    "8317": "failure",
+    "8318": "successful",
+    "8319": "failure",
+    "8320": "successful",
+    "8321": "successful",
+    "8322": "failure",
+    "8323": "successful",
+    "8324": "successful",
+    "8325": "failure",
+    "8326": "failure",
+    "8327": "failure",
+    "8328": "successful",
+    "8329": "successful",
+    "8330": "successful",
+    "8331": "successful",
+    "8332": "failure",
+    "8333": "successful",
+    "8334": "failure",
+    "8335": "failure",
+    "8336": "failure",
+    "8337": "failure",
+    "8338": "failure",
+    "8339": "successful",
+    "8340": "successful",
+    "8341": "successful",
+    "8342": "successful",
+    "8343": "successful",
+    "8344": "failure",
+    "8345": "failure",
+    "8346": "successful",
+    "8347": "failure",
+    "8348": "failure",
+    "8349": "failure",
+    "8350": "failure",
+    "8351": "successful",
+    "8352": "successful",
+    "8353": "successful",
+    "8354": "failure",
+    "8355": "failure",
+    "8356": "failure",
+    "8357": "failure",
+    "8358": "infeasible",
+    "8359": "successful",
+    "8360": "successful",
+    "8361": "failure",
+    "8362": "successful",
+    "8363": "successful",
+    "8364": "failure",
+    "8365": "successful",
+    "8366": "failure",
+    "8367": "failure",
+    "8368": "failure",
+    "8369": "failure",
+    "8370": "failure",
+    "8371": "failure",
+    "8372": "successful",
+    "8373": "successful",
+    "8374": "failure",
+    "8375": "successful",
+    "8376": "successful",
+    "8377": "successful",
+    "8378": "failure",
+    "8379": "failure",
+    "8380": "failure",
+    "8381": "successful",
+    "8382": "successful",
+    "8383": "successful",
+    "8384": "failure",
+    "8385": "successful",
+    "8386": "failure",
+    "8387": "successful",
+    "8388": "failure",
+    "8389": "successful",
+    "8390": "successful",
+    "8391": "failure",
+    "8392": "failure",
+    "8393": "successful",
+    "8394": "successful",
+    "8395": "failure",
+    "8396": "successful",
+    "8397": "failure",
+    "8398": "successful",
+    "8399": "infeasible",
+    "8400": "successful",
+    "8401": "failure",
+    "8402": "failure",
+    "8403": "failure",
+    "8404": "successful",
+    "8405": "successful",
+    "8406": "successful",
+    "8407": "successful",
+    "8408": "successful",
+    "8409": "successful",
+    "8410": "successful",
+    "8411": "failure",
+    "8412": "successful",
+    "8413": "failure",
+    "8414": "failure",
+    "8415": "failure",
+    "8416": "successful",
+    "8417": "failure",
+    "8418": "failure",
+    "8419": "successful",
+    "8420": "successful",
+    "8421": "successful",
+    "8422": "failure",
+    "8423": "failure",
+    "8424": "successful",
+    "8425": "failure",
+    "8426": "successful",
+    "8427": "successful",
+    "8428": "successful",
+    "8429": "successful",
+    "8430": "successful",
+    "8431": "failure",
+    "8432": "failure",
+    "8433": "successful",
+    "8434": "failure",
+    "8435": "successful",
+    "8436": "failure",
+    "8437": "successful",
+    "8438": "failure",
+    "8439": "failure",
+    "8440": "failure",
+    "8441": "successful",
+    "8442": "successful",
+    "8443": "failure",
+    "8444": "failure",
+    "8445": "failure",
+    "8446": "failure",
+    "8447": "failure",
+    "8448": "failure",
+    "8449": "successful",
+    "8450": "failure",
+    "8451": "successful",
+    "8452": "failure",
+    "8453": "successful",
+    "8454": "successful",
+    "8455": "failure",
+    "8456": "successful",
+    "8457": "failure",
+    "8458": "failure",
+    "8459": "failure",
+    "8460": "failure",
+    "8461": "failure",
+    "8462": "successful",
+    "8463": "infeasible",
+    "8464": "successful",
+    "8465": "failure",
+    "8466": "failure",
+    "8467": "failure",
+    "8468": "failure",
+    "8469": "successful",
+    "8470": "failure",
+    "8471": "successful",
+    "8472": "successful",
+    "8473": "successful",
+    "8474": "infeasible",
+    "8475": "successful",
+    "8476": "successful",
+    "8477": "failure",
+    "8478": "failure",
+    "8479": "failure",
+    "8480": "failure",
+    "8481": "failure",
+    "8482": "successful",
+    "8483": "successful",
+    "8484": "successful",
+    "8485": "successful",
+    "8486": "failure",
+    "8487": "failure",
+    "8488": "infeasible",
+    "8489": "successful",
+    "8490": "failure",
+    "8491": "successful",
+    "8492": "failure",
+    "8493": "successful",
+    "8494": "already_complete",
+    "8495": "failure",
+    "8496": "failure",
+    "8497": "successful",
+    "8498": "failure",
+    "8499": "failure",
+    "8500": "successful",
+    "8501": "infeasible",
+    "8502": "already_complete",
+    "8503": "failure",
+    "8504": "failure",
+    "8505": "failure",
+    "8506": "successful",
+    "8507": "failure",
+    "8508": "failure",
+    "8509": "failure",
+    "8510": "failure",
+    "8511": "failure",
+    "8512": "successful",
+    "8513": "successful",
+    "8514": "failure",
+    "8515": "failure",
+    "8516": "successful",
+    "8517": "successful",
+    "8518": "failure",
+    "8519": "failure",
+    "8520": "successful",
+    "8521": "failure",
+    "8522": "successful",
+    "8523": "successful",
+    "8524": "failure",
+    "8525": "failure",
+    "8526": "failure",
+    "8527": "failure",
+    "8528": "successful",
+    "8529": "failure",
+    "8530": "successful",
+    "8531": "successful",
+    "8532": "failure",
+    "8533": "failure",
+    "8534": "successful",
+    "8535": "failure",
+    "8536": "failure",
+    "8537": "successful",
+    "8538": "failure",
+    "8539": "successful",
+    "8540": "failure",
+    "8541": "successful",
+    "8542": "successful",
+    "8543": "failure",
+    "8544": "successful",
+    "8545": "successful",
+    "8546": "successful",
+    "8547": "successful",
+    "8548": "successful",
+    "8549": "failure",
+    "8550": "successful",
+    "8551": "successful",
+    "8552": "failure",
+    "8553": "failure",
+    "8554": "successful",
+    "8555": "successful",
+    "8556": "failure",
+    "8557": "successful",
+    "8558": "successful",
+    "8559": "failure",
+    "8560": "successful",
+    "8561": "infeasible",
+    "8562": "failure",
+    "8563": "failure",
+    "8564": "successful",
+    "8565": "successful",
+    "8566": "failure",
+    "8567": "failure",
+    "8568": "successful",
+    "8569": "successful",
+    "8570": "successful",
+    "8571": "successful",
+    "8572": "failure",
+    "8573": "failure",
+    "8574": "successful",
+    "8575": "successful",
+    "8576": "failure",
+    "8577": "failure",
+    "8578": "successful",
+    "8579": "failure",
+    "8580": "failure",
+    "8581": "successful",
+    "8582": "successful",
+    "8583": "successful",
+    "8584": "successful",
+    "8585": "successful",
+    "8586": "successful",
+    "8587": "successful",
+    "8588": "successful",
+    "8589": "successful",
+    "8590": "failure",
+    "8591": "failure",
+    "8592": "successful",
+    "8593": "successful",
+    "8594": "successful",
+    "8595": "successful",
+    "8596": "failure",
+    "8597": "successful",
+    "8598": "failure",
+    "8599": "failure",
+    "8600": "successful",
+    "8601": "successful",
+    "8602": "successful",
+    "8603": "failure",
+    "8604": "failure",
+    "8605": "successful",
+    "8606": "successful",
+    "8607": "successful",
+    "8608": "successful",
+    "8609": "successful",
+    "8610": "successful",
+    "8611": "successful",
+    "8612": "successful",
+    "8613": "successful",
+    "8614": "failure",
+    "8615": "already_complete",
+    "8616": "successful",
+    "8617": "failure",
+    "8618": "failure",
+    "8619": "failure",
+    "8620": "failure",
+    "8621": "successful",
+    "8622": "failure",
+    "8623": "failure",
+    "8624": "successful",
+    "8625": "failure",
+    "8626": "failure",
+    "8627": "successful",
+    "8628": "successful",
+    "8629": "failure",
+    "8630": "failure",
+    "8631": "successful",
+    "8632": "failure",
+    "8633": "failure",
+    "8634": "successful",
+    "8635": "infeasible",
+    "8636": "successful",
+    "8637": "successful",
+    "8638": "failure",
+    "8639": "successful",
+    "8640": "failure",
+    "8641": "successful",
+    "8642": "failure",
+    "8643": "already_complete",
+    "8644": "successful",
+    "8645": "successful",
+    "8646": "failure",
+    "8647": "successful",
+    "8648": "failure",
+    "8649": "successful",
+    "8650": "infeasible",
+    "8651": "failure",
+    "8652": "successful",
+    "8653": "successful",
+    "8654": "successful",
+    "8655": "failure",
+    "8656": "successful",
+    "8657": "successful",
+    "8658": "successful",
+    "8659": "failure",
+    "8660": "failure",
+    "8661": "failure",
+    "8662": "already_complete",
+    "8663": "failure",
+    "8664": "failure",
+    "8665": "failure",
+    "8666": "failure",
+    "8667": "failure",
+    "8668": "successful",
+    "8669": "successful",
+    "8670": "failure",
+    "8671": "failure",
+    "8672": "successful",
+    "8673": "successful",
+    "8674": "successful",
+    "8675": "successful",
+    "8676": "successful",
+    "8677": "successful",
+    "8678": "already_complete",
+    "8679": "failure",
+    "8680": "successful",
+    "8681": "successful",
+    "8682": "successful",
+    "8683": "failure",
+    "8684": "successful",
+    "8685": "failure",
+    "8686": "successful",
+    "8687": "successful",
+    "8688": "successful",
+    "8689": "failure",
+    "8690": "failure",
+    "8691": "failure",
+    "8692": "failure",
+    "8693": "failure",
+    "8694": "failure",
+    "8695": "successful",
+    "8696": "successful",
+    "8697": "successful",
+    "8698": "successful",
+    "8699": "failure",
+    "8700": "failure",
+    "8701": "failure",
+    "8702": "successful",
+    "8703": "failure",
+    "8704": "failure",
+    "8705": "already_complete",
+    "8706": "successful",
+    "8707": "successful",
+    "8708": "failure",
+    "8709": "failure",
+    "8710": "successful",
+    "8711": "failure",
+    "8712": "successful",
+    "8713": "failure",
+    "8714": "successful",
+    "8715": "failure",
+    "8716": "successful",
+    "8717": "infeasible",
+    "8718": "successful",
+    "8719": "successful",
+    "8720": "failure",
+    "8721": "successful",
+    "8722": "failure",
+    "8723": "failure",
+    "8724": "successful",
+    "8725": "failure",
+    "8726": "failure",
+    "8727": "already_complete",
+    "8728": "failure",
+    "8729": "successful",
+    "8730": "already_complete",
+    "8731": "failure",
+    "8732": "failure",
+    "8733": "failure",
+    "8734": "successful",
+    "8735": "failure",
+    "8736": "successful",
+    "8737": "successful",
+    "8738": "failure",
+    "8739": "successful",
+    "8740": "failure",
+    "8741": "failure",
+    "8742": "failure",
+    "8743": "failure",
+    "8744": "successful",
+    "8745": "successful",
+    "8746": "failure",
+    "8747": "failure",
+    "8748": "successful",
+    "8749": "failure",
+    "8750": "successful",
+    "8751": "failure",
+    "8752": "successful",
+    "8753": "failure",
+    "8754": "failure",
+    "8755": "successful",
+    "8756": "successful",
+    "8757": "infeasible",
+    "8758": "failure",
+    "8759": "successful",
+    "8760": "failure",
+    "8761": "successful",
+    "8762": "failure",
+    "8763": "successful",
+    "8764": "failure",
+    "8765": "successful",
+    "8766": "failure",
+    "8767": "successful",
+    "8768": "failure",
+    "8769": "failure",
+    "8770": "failure",
+    "8771": "infeasible",
+    "8772": "successful",
+    "8773": "successful",
+    "8774": "failure",
+    "8775": "successful",
+    "8776": "failure",
+    "8777": "successful",
+    "8778": "failure",
+    "8779": "failure",
+    "8780": "failure",
+    "8781": "successful",
+    "8782": "failure",
+    "8783": "successful",
+    "8784": "successful",
+    "8785": "failure",
+    "8786": "failure",
+    "8787": "successful",
+    "8788": "failure",
+    "8789": "failure",
+    "8790": "failure",
+    "8791": "successful",
+    "8792": "successful",
+    "8793": "failure",
+    "8794": "failure",
+    "8795": "successful",
+    "8796": "failure",
+    "8797": "successful",
+    "8798": "failure",
+    "8799": "failure",
+    "8800": "successful",
+    "8801": "failure",
+    "8802": "already_complete",
+    "8803": "failure",
+    "8804": "successful",
+    "8805": "failure",
+    "8806": "failure",
+    "8807": "successful",
+    "8808": "failure",
+    "8809": "failure",
+    "8810": "failure",
+    "8811": "successful",
+    "8812": "successful",
+    "8813": "failure",
+    "8814": "successful",
+    "8815": "already_complete",
+    "8816": "failure",
+    "8817": "successful",
+    "8818": "failure",
+    "8819": "successful",
+    "8820": "successful",
+    "8821": "failure",
+    "8822": "failure",
+    "8823": "failure",
+    "8824": "successful",
+    "8825": "already_complete",
+    "8826": "successful",
+    "8827": "failure",
+    "8828": "failure",
+    "8829": "successful",
+    "8830": "failure",
+    "8831": "failure",
+    "8832": "failure",
+    "8833": "successful",
+    "8834": "failure",
+    "8835": "successful",
+    "8836": "successful",
+    "8837": "failure",
+    "8838": "successful",
+    "8839": "failure",
+    "8840": "failure",
+    "8841": "failure",
+    "8842": "successful",
+    "8843": "failure",
+    "8844": "successful",
+    "8845": "successful",
+    "8846": "failure",
+    "8847": "successful",
+    "8848": "already_complete",
+    "8849": "already_complete",
+    "8850": "failure",
+    "8851": "failure",
+    "8852": "successful",
+    "8853": "successful",
+    "8854": "successful",
+    "8855": "failure",
+    "8856": "failure",
+    "8857": "failure",
+    "8858": "failure",
+    "8859": "failure",
+    "8860": "failure",
+    "8861": "successful",
+    "8862": "failure",
+    "8863": "failure",
+    "8864": "failure",
+    "8865": "successful",
+    "8866": "failure",
+    "8867": "failure",
+    "8868": "failure",
+    "8869": "successful",
+    "8870": "successful",
+    "8871": "successful",
+    "8872": "failure",
+    "8873": "failure",
+    "8874": "already_complete",
+    "8875": "failure",
+    "8876": "successful",
+    "8877": "successful",
+    "8878": "failure",
+    "8879": "failure",
+    "8880": "failure",
+    "8881": "successful",
+    "8882": "already_complete",
+    "8883": "failure",
+    "8884": "failure",
+    "8885": "failure",
+    "8886": "failure",
+    "8887": "failure",
+    "8888": "successful",
+    "8889": "successful",
+    "8890": "successful",
+    "8891": "successful",
+    "8892": "failure",
+    "8893": "failure",
+    "8894": "successful",
+    "8895": "failure",
+    "8896": "failure",
+    "8897": "successful",
+    "8898": "successful",
+    "8899": "successful",
+    "8900": "successful",
+    "8901": "successful",
+    "8902": "failure",
+    "8903": "failure",
+    "8904": "failure",
+    "8905": "failure",
+    "8906": "failure",
+    "8907": "successful",
+    "8908": "failure",
+    "8909": "successful",
+    "8910": "failure",
+    "8911": "successful",
+    "8912": "successful",
+    "8913": "successful",
+    "8914": "successful",
+    "8915": "failure",
+    "8916": "failure",
+    "8917": "failure",
+    "8918": "successful",
+    "8919": "already_complete",
+    "8920": "already_complete",
+    "8921": "failure",
+    "8922": "successful",
+    "8923": "successful",
+    "8924": "failure",
+    "8925": "failure",
+    "8926": "successful",
+    "8927": "failure",
+    "8928": "successful",
+    "8929": "failure",
+    "8930": "successful",
+    "8931": "failure",
+    "8932": "failure",
+    "8933": "failure",
+    "8934": "failure",
+    "8935": "successful",
+    "8936": "successful",
+    "8937": "successful",
+    "8938": "infeasible",
+    "8939": "failure",
+    "8940": "failure",
+    "8941": "failure",
+    "8942": "failure",
+    "8943": "successful",
+    "8944": "failure",
+    "8945": "failure",
+    "8946": "failure",
+    "8947": "successful",
+    "8948": "successful",
+    "8949": "failure",
+    "8950": "successful",
+    "8951": "successful",
+    "8952": "successful",
+    "8953": "failure",
+    "8954": "successful",
+    "8955": "failure",
+    "8956": "successful",
+    "8957": "successful",
+    "8958": "failure",
+    "8959": "successful",
+    "8960": "successful",
+    "8961": "failure",
+    "8962": "successful",
+    "8963": "failure",
+    "8964": "successful",
+    "8965": "failure",
+    "8966": "successful",
+    "8967": "failure",
+    "8968": "failure",
+    "8969": "successful",
+    "8970": "failure",
+    "8971": "successful",
+    "8972": "failure",
+    "8973": "successful",
+    "8974": "failure",
+    "8975": "successful",
+    "8976": "failure",
+    "8977": "successful",
+    "8978": "successful",
+    "8979": "successful",
+    "8980": "successful",
+    "8981": "successful",
+    "8982": "failure",
+    "8983": "already_complete",
+    "8984": "failure",
+    "8985": "failure",
+    "8986": "failure",
+    "8987": "successful",
+    "8988": "failure",
+    "8989": "successful",
+    "8990": "successful",
+    "8991": "successful",
+    "8992": "failure",
+    "8993": "failure",
+    "8994": "successful",
+    "8995": "successful",
+    "8996": "successful",
+    "8997": "failure",
+    "8998": "successful",
+    "8999": "successful",
+    "9000": "failure",
+    "9001": "successful",
+    "9002": "successful",
+    "9003": "successful",
+    "9004": "successful",
+    "9005": "failure",
+    "9006": "successful",
+    "9007": "failure",
+    "9008": "failure",
+    "9009": "failure",
+    "9010": "failure",
+    "9011": "failure",
+    "9012": "successful",
+    "9013": "infeasible",
+    "9014": "successful",
+    "9015": "successful",
+    "9016": "successful",
+    "9017": "successful",
+    "9018": "failure",
+    "9019": "successful",
+    "9020": "failure",
+    "9021": "failure",
+    "9022": "successful",
+    "9023": "failure",
+    "9024": "failure",
+    "9025": "successful",
+    "9026": "successful",
+    "9027": "failure",
+    "9028": "successful",
+    "9029": "successful",
+    "9030": "failure",
+    "9031": "failure",
+    "9032": "successful",
+    "9033": "already_complete",
+    "9034": "failure",
+    "9035": "failure",
+    "9036": "successful",
+    "9037": "failure",
+    "9038": "failure",
+    "9039": "successful",
+    "9040": "failure",
+    "9041": "failure",
+    "9042": "failure",
+    "9043": "failure",
+    "9044": "successful",
+    "9045": "failure",
+    "9046": "failure",
+    "9047": "failure",
+    "9048": "failure",
+    "9049": "failure",
+    "9050": "infeasible",
+    "9051": "failure",
+    "9052": "failure",
+    "9053": "successful",
+    "9054": "failure",
+    "9055": "successful",
+    "9056": "successful",
+    "9057": "failure",
+    "9058": "failure",
+    "9059": "infeasible",
+    "9060": "successful",
+    "9061": "failure",
+    "9062": "successful",
+    "9063": "already_complete",
+    "9064": "successful",
+    "9065": "failure",
+    "9066": "failure",
+    "9067": "successful",
+    "9068": "failure",
+    "9069": "failure",
+    "9070": "failure",
+    "9071": "failure",
+    "9072": "already_complete",
+    "9073": "infeasible",
+    "9074": "failure",
+    "9075": "failure",
+    "9076": "failure",
+    "9077": "failure",
+    "9078": "failure",
+    "9079": "failure",
+    "9080": "failure",
+    "9081": "failure",
+    "9082": "infeasible",
+    "9083": "successful",
+    "9084": "successful",
+    "9085": "successful",
+    "9086": "failure",
+    "9087": "already_complete",
+    "9088": "successful",
+    "9089": "failure",
+    "9090": "failure",
+    "9091": "failure",
+    "9092": "failure",
+    "9093": "failure",
+    "9094": "successful",
+    "9095": "failure",
+    "9096": "failure",
+    "9097": "failure",
+    "9098": "failure",
+    "9099": "successful",
+    "9100": "successful",
+    "9101": "failure",
+    "9102": "failure",
+    "9103": "failure",
+    "9104": "successful",
+    "9105": "failure",
+    "9106": "failure",
+    "9107": "successful",
+    "9108": "successful",
+    "9109": "successful",
+    "9110": "failure",
+    "9111": "failure",
+    "9112": "successful",
+    "9113": "successful",
+    "9114": "successful",
+    "9115": "successful",
+    "9116": "successful",
+    "9117": "failure",
+    "9118": "failure",
+    "9119": "failure",
+    "9120": "infeasible",
+    "9121": "successful",
+    "9122": "infeasible",
+    "9123": "failure",
+    "9124": "failure",
+    "9125": "successful",
+    "9126": "failure",
+    "9127": "failure",
+    "9128": "failure",
+    "9129": "successful",
+    "9130": "failure",
+    "9131": "successful",
+    "9132": "successful",
+    "9133": "successful",
+    "9134": "successful",
+    "9135": "successful",
+    "9136": "successful",
+    "9137": "successful",
+    "9138": "successful",
+    "9139": "successful",
+    "9140": "failure",
+    "9141": "successful",
+    "9142": "infeasible",
+    "9143": "successful",
+    "9144": "successful",
+    "9145": "successful",
+    "9146": "successful",
+    "9147": "successful",
+    "9148": "successful",
+    "9149": "failure",
+    "9150": "failure",
+    "9151": "failure",
+    "9152": "successful",
+    "9153": "failure",
+    "9154": "infeasible",
+    "9155": "successful",
+    "9156": "successful",
+    "9157": "failure",
+    "9158": "failure",
+    "9159": "failure",
+    "9160": "failure",
+    "9161": "failure",
+    "9162": "failure",
+    "9163": "successful",
+    "9164": "failure",
+    "9165": "failure",
+    "9166": "successful",
+    "9167": "failure",
+    "9168": "failure",
+    "9169": "failure",
+    "9170": "successful",
+    "9171": "failure",
+    "9172": "successful",
+    "9173": "successful",
+    "9174": "successful",
+    "9175": "failure",
+    "9176": "successful",
+    "9177": "successful",
+    "9178": "successful",
+    "9179": "infeasible",
+    "9180": "failure",
+    "9181": "successful",
+    "9182": "failure",
+    "9183": "failure",
+    "9184": "successful",
+    "9185": "successful",
+    "9186": "successful",
+    "9187": "failure",
+    "9188": "failure",
+    "9189": "successful",
+    "9190": "failure",
+    "9191": "failure",
+    "9192": "failure",
+    "9193": "successful",
+    "9194": "failure",
+    "9195": "successful",
+    "9196": "failure",
+    "9197": "failure",
+    "9198": "failure",
+    "9199": "successful",
+    "9200": "failure",
+    "9201": "successful",
+    "9202": "failure",
+    "9203": "successful",
+    "9204": "failure",
+    "9205": "successful",
+    "9206": "failure",
+    "9207": "successful",
+    "9208": "successful",
+    "9209": "failure",
+    "9210": "successful",
+    "9211": "failure",
+    "9212": "failure",
+    "9213": "failure",
+    "9214": "failure",
+    "9215": "successful",
+    "9216": "failure",
+    "9217": "failure",
+    "9218": "successful",
+    "9219": "failure",
+    "9220": "failure",
+    "9221": "failure",
+    "9222": "failure",
+    "9223": "successful",
+    "9224": "failure",
+    "9225": "failure",
+    "9226": "failure",
+    "9227": "failure",
+    "9228": "successful",
+    "9229": "failure",
+    "9230": "failure",
+    "9231": "failure",
+    "9232": "successful",
+    "9233": "successful",
+    "9234": "successful",
+    "9235": "failure",
+    "9236": "failure",
+    "9237": "failure",
+    "9238": "failure",
+    "9239": "failure",
+    "9240": "successful",
+    "9241": "failure",
+    "9242": "successful",
+    "9243": "failure",
+    "9244": "successful",
+    "9245": "successful",
+    "9246": "successful",
+    "9247": "failure",
+    "9248": "failure",
+    "9249": "successful",
+    "9250": "failure",
+    "9251": "successful",
+    "9252": "successful",
+    "9253": "successful",
+    "9254": "successful",
+    "9255": "successful",
+    "9256": "failure",
+    "9257": "failure",
+    "9258": "failure",
+    "9259": "failure",
+    "9260": "failure",
+    "9261": "failure",
+    "9262": "failure",
+    "9263": "successful",
+    "9264": "successful",
+    "9265": "successful",
+    "9266": "successful",
+    "9267": "successful",
+    "9268": "successful",
+    "9269": "successful",
+    "9270": "successful",
+    "9271": "successful",
+    "9272": "failure",
+    "9273": "failure",
+    "9274": "failure",
+    "9275": "successful",
+    "9276": "successful",
+    "9277": "failure",
+    "9278": "successful",
+    "9279": "failure",
+    "9280": "successful",
+    "9281": "successful",
+    "9282": "failure",
+    "9283": "failure",
+    "9284": "failure",
+    "9285": "failure",
+    "9286": "failure",
+    "9287": "successful",
+    "9288": "successful",
+    "9289": "failure",
+    "9290": "failure",
+    "9291": "successful",
+    "9292": "successful",
+    "9293": "successful",
+    "9294": "successful",
+    "9295": "failure",
+    "9296": "failure",
+    "9297": "failure",
+    "9298": "successful",
+    "9299": "successful",
+    "9300": "successful",
+    "9301": "successful",
+    "9302": "successful",
+    "9303": "successful",
+    "9304": "successful",
+    "9305": "successful",
+    "9306": "failure",
+    "9307": "failure",
+    "9308": "failure",
+    "9309": "failure",
+    "9310": "successful",
+    "9311": "successful",
+    "9312": "successful",
+    "9313": "failure",
+    "9314": "successful",
+    "9315": "failure",
+    "9316": "successful",
+    "9317": "failure",
+    "9318": "successful",
+    "9319": "successful",
+    "9320": "failure",
+    "9321": "failure",
+    "9322": "successful",
+    "9323": "failure",
+    "9324": "failure",
+    "9325": "failure",
+    "9326": "failure",
+    "9327": "failure",
+    "9328": "failure",
+    "9329": "failure",
+    "9330": "failure",
+    "9331": "failure",
+    "9332": "failure",
+    "9333": "failure",
+    "9334": "failure",
+    "9335": "failure",
+    "9336": "successful",
+    "9337": "successful",
+    "9338": "successful",
+    "9339": "failure",
+    "9340": "failure",
+    "9341": "failure",
+    "9342": "successful",
+    "9343": "failure",
+    "9344": "successful",
+    "9345": "failure",
+    "9346": "failure",
+    "9347": "successful",
+    "9348": "already_complete",
+    "9349": "failure",
+    "9350": "successful",
+    "9351": "failure",
+    "9352": "successful",
+    "9353": "successful",
+    "9354": "failure",
+    "9355": "already_complete",
+    "9356": "failure",
+    "9357": "successful",
+    "9358": "failure",
+    "9359": "successful",
+    "9360": "successful",
+    "9361": "successful",
+    "9362": "infeasible",
+    "9363": "infeasible",
+    "9364": "failure",
+    "9365": "successful",
+    "9366": "failure",
+    "9367": "failure",
+    "9368": "successful",
+    "9369": "failure",
+    "9370": "successful",
+    "9371": "failure",
+    "9372": "already_complete",
+    "9373": "successful",
+    "9374": "successful",
+    "9375": "failure",
+    "9376": "successful",
+    "9377": "successful",
+    "9378": "successful",
+    "9379": "successful",
+    "9380": "successful",
+    "9381": "successful",
+    "9382": "successful",
+    "9383": "failure",
+    "9384": "successful",
+    "9385": "successful",
+    "9386": "failure",
+    "9387": "successful",
+    "9388": "successful",
+    "9389": "successful",
+    "9390": "failure",
+    "9391": "successful",
+    "9392": "failure",
+    "9393": "successful",
+    "9394": "failure",
+    "9395": "successful",
+    "9396": "successful",
+    "9397": "failure",
+    "9398": "successful",
+    "9399": "already_complete",
+    "9400": "successful",
+    "9401": "failure",
+    "9402": "successful",
+    "9403": "failure",
+    "9404": "already_complete",
+    "9405": "successful",
+    "9406": "successful",
+    "9407": "failure",
+    "9408": "successful",
+    "9409": "failure",
+    "9410": "successful",
+    "9411": "failure",
+    "9412": "failure",
+    "9413": "infeasible",
+    "9414": "successful",
+    "9415": "successful",
+    "9416": "failure",
+    "9417": "successful",
+    "9418": "already_complete",
+    "9419": "failure",
+    "9420": "infeasible",
+    "9421": "failure",
+    "9422": "successful",
+    "9423": "failure",
+    "9424": "failure",
+    "9425": "successful",
+    "9426": "failure",
+    "9427": "successful",
+    "9428": "successful",
+    "9429": "successful",
+    "9430": "failure",
+    "9431": "failure",
+    "9432": "successful",
+    "9433": "failure",
+    "9434": "successful",
+    "9435": "infeasible",
+    "9436": "successful",
+    "9437": "successful",
+    "9438": "failure",
+    "9439": "successful",
+    "9440": "successful",
+    "9441": "failure",
+    "9442": "successful",
+    "9443": "successful",
+    "9444": "successful",
+    "9445": "successful",
+    "9446": "failure",
+    "9447": "failure",
+    "9448": "failure",
+    "9449": "successful",
+    "9450": "successful",
+    "9451": "successful",
+    "9452": "failure",
+    "9453": "already_complete",
+    "9454": "successful",
+    "9455": "successful",
+    "9456": "failure",
+    "9457": "failure",
+    "9458": "failure",
+    "9459": "failure",
+    "9460": "successful",
+    "9461": "failure",
+    "9462": "failure",
+    "9463": "successful",
+    "9464": "successful",
+    "9465": "failure",
+    "9466": "failure",
+    "9467": "successful",
+    "9468": "failure",
+    "9469": "failure",
+    "9470": "failure",
+    "9471": "successful",
+    "9472": "successful",
+    "9473": "successful",
+    "9474": "successful",
+    "9475": "failure",
+    "9476": "failure",
+    "9477": "infeasible",
+    "9478": "failure",
+    "9479": "successful",
+    "9480": "successful",
+    "9481": "failure",
+    "9482": "successful",
+    "9483": "failure",
+    "9484": "successful",
+    "9485": "failure",
+    "9486": "failure",
+    "9487": "failure",
+    "9488": "successful",
+    "9489": "failure",
+    "9490": "successful",
+    "9491": "failure",
+    "9492": "successful",
+    "9493": "successful",
+    "9494": "failure",
+    "9495": "successful",
+    "9496": "successful",
+    "9497": "failure",
+    "9498": "infeasible",
+    "9499": "successful",
+    "9500": "successful",
+    "9501": "successful",
+    "9502": "successful",
+    "9503": "failure",
+    "9504": "successful",
+    "9505": "already_complete",
+    "9506": "failure",
+    "9507": "successful",
+    "9508": "successful",
+    "9509": "failure",
+    "9510": "failure",
+    "9511": "successful",
+    "9512": "failure",
+    "9513": "successful",
+    "9514": "successful",
+    "9515": "failure",
+    "9516": "failure",
+    "9517": "failure",
+    "9518": "failure",
+    "9519": "successful",
+    "9520": "failure",
+    "9521": "successful",
+    "9522": "successful",
+    "9523": "successful",
+    "9524": "successful",
+    "9525": "successful",
+    "9526": "successful",
+    "9527": "successful",
+    "9528": "failure",
+    "9529": "failure",
+    "9530": "failure",
+    "9531": "failure",
+    "9532": "successful",
+    "9533": "successful",
+    "9534": "successful",
+    "9535": "successful",
+    "9536": "successful",
+    "9537": "successful",
+    "9538": "successful",
+    "9539": "failure",
+    "9540": "successful",
+    "9541": "successful",
+    "9542": "successful",
+    "9543": "successful",
+    "9544": "failure",
+    "9545": "successful",
+    "9546": "failure",
+    "9547": "successful",
+    "9548": "successful",
+    "9549": "successful",
+    "9550": "successful",
+    "9551": "failure",
+    "9552": "failure",
+    "9553": "already_complete",
+    "9554": "failure",
+    "9555": "successful",
+    "9556": "failure",
+    "9557": "successful",
+    "9558": "successful",
+    "9559": "infeasible",
+    "9560": "successful",
+    "9561": "failure",
+    "9562": "failure",
+    "9563": "successful",
+    "9564": "failure",
+    "9565": "already_complete",
+    "9566": "failure",
+    "9567": "failure",
+    "9568": "successful",
+    "9569": "failure",
+    "9570": "successful",
+    "9571": "successful",
+    "9572": "already_complete",
+    "9573": "failure",
+    "9574": "successful",
+    "9575": "successful",
+    "9576": "failure",
+    "9577": "successful",
+    "9578": "failure",
+    "9579": "successful",
+    "9580": "successful",
+    "9581": "failure",
+    "9582": "successful",
+    "9583": "successful",
+    "9584": "failure",
+    "9585": "successful",
+    "9586": "successful",
+    "9587": "already_complete",
+    "9588": "failure",
+    "9589": "successful",
+    "9590": "failure",
+    "9591": "successful",
+    "9592": "successful",
+    "9593": "successful",
+    "9594": "failure",
+    "9595": "failure",
+    "9596": "failure",
+    "9597": "failure",
+    "9598": "failure",
+    "9599": "failure",
+    "9600": "successful",
+    "9601": "successful",
+    "9602": "failure",
+    "9603": "failure",
+    "9604": "failure",
+    "9605": "successful",
+    "9606": "successful",
+    "9607": "failure",
+    "9608": "successful",
+    "9609": "failure",
+    "9610": "failure",
+    "9611": "failure",
+    "9612": "failure",
+    "9613": "successful",
+    "9614": "successful",
+    "9615": "successful",
+    "9616": "successful",
+    "9617": "already_complete",
+    "9618": "failure",
+    "9619": "successful",
+    "9620": "successful",
+    "9621": "successful",
+    "9622": "failure",
+    "9623": "failure",
+    "9624": "successful",
+    "9625": "failure",
+    "9626": "failure",
+    "9627": "failure",
+    "9628": "successful",
+    "9629": "successful",
+    "9630": "failure",
+    "9631": "successful",
+    "9632": "successful",
+    "9633": "successful",
+    "9634": "failure",
+    "9635": "failure",
+    "9636": "successful",
+    "9637": "successful",
+    "9638": "successful",
+    "9639": "successful",
+    "9640": "failure",
+    "9641": "successful",
+    "9642": "failure",
+    "9643": "failure",
+    "9644": "successful",
+    "9645": "successful",
+    "9646": "already_complete",
+    "9647": "successful",
+    "9648": "successful",
+    "9649": "successful",
+    "9650": "failure",
+    "9651": "failure",
+    "9652": "failure",
+    "9653": "failure",
+    "9654": "failure",
+    "9655": "successful",
+    "9656": "successful",
+    "9657": "failure",
+    "9658": "successful",
+    "9659": "failure",
+    "9660": "failure",
+    "9661": "failure",
+    "9662": "successful",
+    "9663": "successful",
+    "9664": "failure",
+    "9665": "successful",
+    "9666": "failure",
+    "9667": "successful",
+    "9668": "failure",
+    "9669": "successful",
+    "9670": "failure",
+    "9671": "failure",
+    "9672": "successful",
+    "9673": "successful",
+    "9674": "successful",
+    "9675": "infeasible",
+    "9676": "successful",
+    "9677": "successful",
+    "9678": "successful",
+    "9679": "failure",
+    "9680": "failure",
+    "9681": "successful",
+    "9682": "failure",
+    "9683": "successful",
+    "9684": "failure",
+    "9685": "successful",
+    "9686": "failure",
+    "9687": "successful",
+    "9688": "failure",
+    "9689": "successful",
+    "9690": "successful",
+    "9691": "failure",
+    "9692": "successful",
+    "9693": "successful",
+    "9694": "failure",
+    "9695": "failure",
+    "9696": "failure",
+    "9697": "failure",
+    "9698": "failure",
+    "9699": "failure",
+    "9700": "failure",
+    "9701": "failure",
+    "9702": "failure",
+    "9703": "successful",
+    "9704": "successful",
+    "9705": "successful",
+    "9706": "successful",
+    "9707": "failure",
+    "9708": "successful",
+    "9709": "successful",
+    "9710": "successful",
+    "9711": "successful",
+    "9712": "successful",
+    "9713": "successful",
+    "9714": "successful",
+    "9715": "failure",
+    "9716": "successful",
+    "9717": "failure",
+    "9718": "successful",
+    "9719": "successful",
+    "9720": "failure",
+    "9721": "failure",
+    "9722": "failure",
+    "9723": "failure",
+    "9724": "failure",
+    "9725": "failure",
+    "9726": "failure",
+    "9727": "successful",
+    "9728": "failure",
+    "9729": "failure",
+    "9730": "successful",
+    "9731": "failure",
+    "9732": "failure",
+    "9733": "successful",
+    "9734": "failure",
+    "9735": "failure",
+    "9736": "failure",
+    "9737": "failure",
+    "9738": "successful",
+    "9739": "successful",
+    "9740": "failure",
+    "9741": "failure",
+    "9742": "successful",
+    "9743": "successful",
+    "9744": "failure",
+    "9745": "successful",
+    "9746": "successful",
+    "9747": "successful",
+    "9748": "failure",
+    "9749": "failure",
+    "9750": "failure",
+    "9751": "failure",
+    "9752": "already_complete",
+    "9753": "successful",
+    "9754": "failure",
+    "9755": "failure",
+    "9756": "successful",
+    "9757": "failure",
+    "9758": "failure",
+    "9759": "successful",
+    "9760": "failure",
+    "9761": "infeasible",
+    "9762": "successful",
+    "9763": "failure",
+    "9764": "failure",
+    "9765": "successful",
+    "9766": "successful",
+    "9767": "failure",
+    "9768": "successful",
+    "9769": "failure",
+    "9770": "failure",
+    "9771": "failure",
+    "9772": "failure",
+    "9773": "failure",
+    "9774": "failure",
+    "9775": "failure",
+    "9776": "failure",
+    "9777": "successful",
+    "9778": "failure",
+    "9779": "failure",
+    "9780": "successful",
+    "9781": "successful",
+    "9782": "failure",
+    "9783": "successful",
+    "9784": "successful",
+    "9785": "successful",
+    "9786": "failure",
+    "9787": "successful",
+    "9788": "successful",
+    "9789": "successful",
+    "9790": "successful",
+    "9791": "successful",
+    "9792": "successful",
+    "9793": "failure",
+    "9794": "already_complete",
+    "9795": "failure",
+    "9796": "successful",
+    "9797": "successful",
+    "9798": "failure",
+    "9799": "successful",
+    "9800": "failure",
+    "9801": "failure",
+    "9802": "successful",
+    "9803": "successful",
+    "9804": "successful",
+    "9805": "successful",
+    "9806": "successful",
+    "9807": "failure",
+    "9808": "successful",
+    "9809": "successful",
+    "9810": "already_complete",
+    "9811": "failure",
+    "9812": "failure",
+    "9813": "successful",
+    "9814": "failure",
+    "9815": "successful",
+    "9816": "successful",
+    "9817": "failure",
+    "9818": "failure",
+    "9819": "failure",
+    "9820": "already_complete",
+    "9821": "failure",
+    "9822": "successful",
+    "9823": "failure",
+    "9824": "failure",
+    "9825": "successful",
+    "9826": "successful",
+    "9827": "successful",
+    "9828": "successful",
+    "9829": "successful",
+    "9830": "failure",
+    "9831": "successful",
+    "9832": "infeasible",
+    "9833": "failure",
+    "9834": "failure",
+    "9835": "successful",
+    "9836": "failure",
+    "9837": "failure",
+    "9838": "successful",
+    "9839": "successful",
+    "9840": "failure",
+    "9841": "failure",
+    "9842": "failure",
+    "9843": "failure",
+    "9844": "successful",
+    "9845": "failure",
+    "9846": "failure",
+    "9847": "successful",
+    "9848": "successful",
+    "9849": "failure",
+    "9850": "successful",
+    "9851": "failure",
+    "9852": "successful",
+    "9853": "successful",
+    "9854": "failure",
+    "9855": "failure",
+    "9856": "successful",
+    "9857": "failure",
+    "9858": "failure",
+    "9859": "failure",
+    "9860": "successful",
+    "9861": "successful",
+    "9862": "failure",
+    "9863": "failure",
+    "9864": "successful",
+    "9865": "successful",
+    "9866": "successful",
+    "9867": "successful",
+    "9868": "successful",
+    "9869": "failure",
+    "9870": "failure",
+    "9871": "successful",
+    "9872": "successful",
+    "9873": "failure",
+    "9874": "failure",
+    "9875": "failure",
+    "9876": "successful",
+    "9877": "failure",
+    "9878": "already_complete",
+    "9879": "failure",
+    "9880": "successful",
+    "9881": "failure",
+    "9882": "successful",
+    "9883": "successful",
+    "9884": "successful",
+    "9885": "failure",
+    "9886": "failure",
+    "9887": "successful",
+    "9888": "successful",
+    "9889": "failure",
+    "9890": "successful",
+    "9891": "failure",
+    "9892": "failure",
+    "9893": "successful",
+    "9894": "successful",
+    "9895": "failure",
+    "9896": "failure",
+    "9897": "failure",
+    "9898": "successful",
+    "9899": "successful",
+    "9900": "failure",
+    "9901": "successful",
+    "9902": "successful",
+    "9903": "failure",
+    "9904": "failure",
+    "9905": "failure",
+    "9906": "successful",
+    "9907": "successful",
+    "9908": "successful",
+    "9909": "successful",
+    "9910": "failure",
+    "9911": "failure",
+    "9912": "successful",
+    "9913": "failure",
+    "9914": "successful",
+    "9915": "failure",
+    "9916": "successful",
+    "9917": "successful",
+    "9918": "successful",
+    "9919": "failure",
+    "9920": "successful",
+    "9921": "failure",
+    "9922": "successful",
+    "9923": "failure",
+    "9924": "successful",
+    "9925": "successful",
+    "9926": "successful",
+    "9927": "successful",
+    "9928": "successful",
+    "9929": "successful",
+    "9930": "failure",
+    "9931": "successful",
+    "9932": "successful",
+    "9933": "failure",
+    "9934": "failure",
+    "9935": "successful",
+    "9936": "successful",
+    "9937": "failure",
+    "9938": "failure",
+    "9939": "failure",
+    "9940": "already_complete",
+    "9941": "successful",
+    "9942": "failure",
+    "9943": "successful",
+    "9944": "successful",
+    "9945": "failure",
+    "9946": "failure",
+    "9947": "failure",
+    "9948": "successful",
+    "9949": "failure",
+    "9950": "successful",
+    "9951": "failure",
+    "9952": "successful",
+    "9953": "successful",
+    "9954": "failure",
+    "9955": "failure",
+    "9956": "successful",
+    "9957": "successful",
+    "9958": "failure",
+    "9959": "successful",
+    "9960": "failure",
+    "9961": "failure",
+    "9962": "already_complete",
+    "9963": "successful",
+    "9964": "failure",
+    "9965": "successful",
+    "9966": "successful",
+    "9967": "successful",
+    "9968": "failure",
+    "9969": "successful",
+    "9970": "successful",
+    "9971": "successful",
+    "9972": "successful",
+    "9973": "successful",
+    "9974": "successful",
+    "9975": "successful",
+    "9976": "successful",
+    "9977": "successful",
+    "9978": "failure",
+    "9979": "successful",
+    "9980": "failure",
+    "9981": "failure",
+    "9982": "failure",
+    "9983": "successful",
+    "9984": "failure",
+    "9985": "failure",
+    "9986": "failure",
+    "9987": "successful",
+    "9988": "failure",
+    "9989": "failure",
+    "9990": "successful",
+    "9991": "successful",
+    "9992": "failure",
+    "9993": "successful",
+    "9994": "failure",
+    "9995": "failure",
+    "9996": "failure",
+    "9997": "successful",
+    "9998": "successful",
+    "9999": "successful",
+    "10000": "infeasible",
+    "10001": "failure",
+    "10002": "infeasible",
+    "10003": "failure",
+    "10004": "successful",
+    "10005": "failure",
+    "10006": "successful",
+    "10007": "infeasible",
+    "10008": "successful",
+    "10009": "successful",
+    "10010": "already_complete",
+    "10011": "failure",
+    "10012": "successful",
+    "10013": "successful",
+    "10014": "failure",
+    "10015": "failure",
+    "10016": "successful",
+    "10017": "successful",
+    "10018": "failure",
+    "10019": "failure",
+    "10020": "infeasible",
+    "10021": "failure",
+    "10022": "failure",
+    "10023": "failure",
+    "10024": "failure",
+    "10025": "failure",
+    "10026": "failure",
+    "10027": "failure",
+    "10028": "failure",
+    "10029": "failure",
+    "10030": "failure",
+    "10031": "failure",
+    "10032": "failure",
+    "10033": "failure",
+    "10034": "failure",
+    "10035": "successful",
+    "10036": "infeasible",
+    "10037": "successful",
+    "10038": "successful",
+    "10039": "failure",
+    "10040": "failure",
+    "10041": "successful",
+    "10042": "successful",
+    "10043": "failure",
+    "10044": "failure",
+    "10045": "failure",
+    "10046": "failure",
+    "10047": "infeasible",
+    "10048": "failure",
+    "10049": "failure",
+    "10050": "failure",
+    "10051": "failure",
+    "10052": "infeasible",
+    "10053": "failure",
+    "10054": "failure",
+    "10055": "infeasible",
+    "10056": "failure",
+    "10057": "failure",
+    "10058": "failure",
+    "10059": "failure",
+    "10060": "failure",
+    "10061": "infeasible",
+    "10062": "successful",
+    "10063": "failure",
+    "10064": "failure",
+    "10065": "failure",
+    "10066": "failure",
+    "10067": "failure",
+    "10068": "failure",
+    "10069": "failure",
+    "10070": "failure",
+    "10071": "failure",
+    "10072": "failure",
+    "10073": "failure",
+    "10074": "failure",
+    "10075": "failure",
+    "10076": "successful",
+    "10077": "failure",
+    "10078": "failure",
+    "10079": "failure",
+    "10080": "failure",
+    "10081": "infeasible",
+    "10082": "failure",
+    "10083": "failure",
+    "10084": "infeasible",
+    "10085": "failure",
+    "10086": "failure",
+    "10087": "infeasible",
+    "10088": "successful",
+    "10089": "failure",
+    "10090": "successful",
+    "10091": "failure",
+    "10092": "failure",
+    "10093": "failure",
+    "10094": "infeasible",
+    "10095": "failure",
+    "10096": "failure",
+    "10097": "failure",
+    "10098": "failure",
+    "10099": "failure",
+    "10100": "infeasible",
+    "10101": "failure",
+    "10102": "failure",
+    "10103": "failure",
+    "10104": "already_complete",
+    "10105": "failure",
+    "10106": "failure",
+    "10107": "failure",
+    "10108": "failure",
+    "10109": "failure",
+    "10110": "failure",
+    "10111": "failure",
+    "10112": "failure",
+    "10113": "failure",
+    "10114": "failure",
+    "10115": "failure",
+    "10116": "failure",
+    "10117": "successful",
+    "10118": "failure",
+    "10119": "failure",
+    "10120": "failure",
+    "10121": "failure",
+    "10122": "failure",
+    "10123": "failure",
+    "10124": "failure",
+    "10125": "failure",
+    "10126": "infeasible",
+    "10127": "failure",
+    "10128": "failure",
+    "10129": "successful",
+    "10130": "failure",
+    "10131": "failure",
+    "10132": "successful",
+    "10133": "failure",
+    "10134": "failure",
+    "10135": "successful",
+    "10136": "failure",
+    "10137": "failure",
+    "10138": "infeasible",
+    "10139": "failure",
+    "10140": "failure",
+    "10141": "failure",
+    "10142": "infeasible",
+    "10143": "failure",
+    "10144": "infeasible",
+    "10145": "successful",
+    "10146": "infeasible",
+    "10147": "failure",
+    "10148": "successful",
+    "10149": "failure",
+    "10150": "failure",
+    "10151": "failure",
+    "10152": "failure",
+    "10153": "failure",
+    "10154": "failure",
+    "10155": "failure",
+    "10156": "failure",
+    "10157": "failure",
+    "10158": "infeasible",
+    "10159": "failure",
+    "10160": "failure",
+    "10161": "failure",
+    "10162": "failure",
+    "10163": "failure",
+    "10164": "failure",
+    "10165": "failure",
+    "10166": "failure",
+    "10167": "infeasible",
+    "10168": "failure",
+    "10169": "infeasible",
+    "10170": "failure",
+    "10171": "failure",
+    "10172": "failure",
+    "10173": "failure",
+    "10174": "failure",
+    "10175": "failure",
+    "10176": "failure",
+    "10177": "failure",
+    "10178": "failure",
+    "10179": "failure",
+    "10180": "failure",
+    "10181": "failure",
+    "10182": "failure",
+    "10183": "failure",
+    "10184": "successful",
+    "10185": "failure",
+    "10186": "successful",
+    "10187": "failure",
+    "10188": "failure",
+    "10189": "successful",
+    "10190": "failure",
+    "10191": "failure",
+    "10192": "failure",
+    "10193": "failure",
+    "10194": "failure",
+    "10195": "successful",
+    "10196": "successful",
+    "10197": "failure",
+    "10198": "successful",
+    "10199": "failure",
+    "10200": "failure",
+    "10201": "failure",
+    "10202": "failure",
+    "10203": "failure",
+    "10204": "failure",
+    "10205": "failure",
+    "10206": "failure",
+    "10207": "failure",
+    "10208": "failure",
+    "10209": "failure",
+    "10210": "failure",
+    "10211": "failure",
+    "10212": "failure",
+    "10213": "failure",
+    "10214": "failure",
+    "10215": "failure",
+    "10216": "infeasible",
+    "10217": "failure",
+    "10218": "failure",
+    "10219": "failure",
+    "10220": "failure",
+    "10221": "failure",
+    "10222": "failure",
+    "10223": "failure",
+    "10224": "failure",
+    "10225": "failure",
+    "10226": "failure",
+    "10227": "failure",
+    "10228": "failure",
+    "10229": "failure",
+    "10230": "failure",
+    "10231": "failure",
+    "10232": "failure",
+    "10233": "failure",
+    "10234": "failure",
+    "10235": "failure",
+    "10236": "failure",
+    "10237": "successful",
+    "10238": "failure",
+    "10239": "failure",
+    "10240": "failure",
+    "10241": "failure",
+    "10242": "failure",
+    "10243": "failure",
+    "10244": "failure",
+    "10245": "failure",
+    "10246": "failure",
+    "10247": "infeasible",
+    "10248": "failure",
+    "10249": "failure",
+    "10250": "failure",
+    "10251": "failure",
+    "10252": "failure",
+    "10253": "successful",
+    "10254": "failure",
+    "10255": "failure",
+    "10256": "failure",
+    "10257": "failure",
+    "10258": "failure",
+    "10259": "failure",
+    "10260": "infeasible",
+    "10261": "failure",
+    "10262": "failure",
+    "10263": "failure",
+    "10264": "infeasible",
+    "10265": "failure",
+    "10266": "failure",
+    "10267": "failure",
+    "10268": "failure",
+    "10269": "failure",
+    "10270": "infeasible",
+    "10271": "failure",
+    "10272": "successful",
+    "10273": "failure",
+    "10274": "failure",
+    "10275": "infeasible",
+    "10276": "failure",
+    "10277": "infeasible",
+    "10278": "failure",
+    "10279": "failure",
+    "10280": "failure",
+    "10281": "failure",
+    "10282": "failure",
+    "10283": "failure",
+    "10284": "failure",
+    "10285": "successful",
+    "10286": "already_complete",
+    "10287": "failure",
+    "10288": "successful",
+    "10289": "failure",
+    "10290": "failure",
+    "10291": "failure",
+    "10292": "infeasible",
+    "10293": "failure",
+    "10294": "failure",
+    "10295": "failure",
+    "10296": "failure",
+    "10297": "infeasible",
+    "10298": "successful",
+    "10299": "failure",
+    "10300": "failure",
+    "10301": "failure",
+    "10302": "failure",
+    "10303": "failure",
+    "10304": "successful",
+    "10305": "successful",
+    "10306": "failure",
+    "10307": "failure",
+    "10308": "failure",
+    "10309": "failure",
+    "10310": "failure",
+    "10311": "successful",
+    "10312": "failure",
+    "10313": "failure",
+    "10314": "failure",
+    "10315": "failure",
+    "10316": "failure",
+    "10317": "failure",
+    "10318": "failure",
+    "10319": "failure",
+    "10320": "failure",
+    "10321": "successful",
+    "10322": "successful",
+    "10323": "infeasible",
+    "10324": "infeasible",
+    "10325": "successful",
+    "10326": "failure",
+    "10327": "failure",
+    "10328": "failure",
+    "10329": "failure",
+    "10330": "failure",
+    "10331": "failure",
+    "10332": "infeasible",
+    "10333": "infeasible",
+    "10334": "failure",
+    "10335": "failure",
+    "10336": "failure",
+    "10337": "failure",
+    "10338": "failure",
+    "10339": "failure",
+    "10340": "infeasible",
+    "10341": "infeasible",
+    "10342": "failure",
+    "10343": "infeasible",
+    "10344": "failure",
+    "10345": "failure",
+    "10346": "failure",
+    "10347": "failure",
+    "10348": "failure",
+    "10349": "failure",
+    "10350": "failure",
+    "10351": "failure",
+    "10352": "failure",
+    "10353": "failure",
+    "10354": "failure",
+    "10355": "failure",
+    "10356": "failure",
+    "10357": "successful",
+    "10358": "failure",
+    "10359": "failure",
+    "10360": "failure",
+    "10361": "failure",
+    "10362": "successful",
+    "10363": "failure",
+    "10364": "failure",
+    "10365": "failure",
+    "10366": "failure",
+    "10367": "failure",
+    "10368": "failure",
+    "10369": "failure",
+    "10370": "infeasible",
+    "10371": "infeasible",
+    "10372": "failure",
+    "10373": "failure",
+    "10374": "failure",
+    "10375": "infeasible",
+    "10376": "failure",
+    "10377": "failure",
+    "10378": "failure",
+    "10379": "successful",
+    "10380": "failure",
+    "10381": "failure",
+    "10382": "failure",
+    "10383": "failure",
+    "10384": "failure",
+    "10385": "failure",
+    "10386": "failure",
+    "10387": "successful",
+    "10388": "failure",
+    "10389": "failure",
+    "10390": "successful",
+    "10391": "failure",
+    "10392": "failure",
+    "10393": "successful",
+    "10394": "failure",
+    "10395": "failure",
+    "10396": "failure",
+    "10397": "failure",
+    "10398": "failure",
+    "10399": "successful",
+    "10400": "failure",
+    "10401": "failure",
+    "10402": "failure",
+    "10403": "failure",
+    "10404": "failure",
+    "10405": "successful",
+    "10406": "failure",
+    "10407": "failure",
+    "10408": "failure",
+    "10409": "failure",
+    "10410": "infeasible",
+    "10411": "failure",
+    "10412": "failure",
+    "10413": "failure",
+    "10414": "failure",
+    "10415": "failure",
+    "10416": "failure",
+    "10417": "failure",
+    "10418": "infeasible",
+    "10419": "failure",
+    "10420": "failure",
+    "10421": "failure",
+    "10422": "failure",
+    "10423": "failure",
+    "10424": "failure",
+    "10425": "failure",
+    "10426": "failure",
+    "10427": "failure",
+    "10428": "successful",
+    "10429": "failure",
+    "10430": "failure",
+    "10431": "failure",
+    "10432": "successful",
+    "10433": "successful",
+    "10434": "failure",
+    "10435": "failure",
+    "10436": "failure",
+    "10437": "failure",
+    "10438": "failure",
+    "10439": "failure",
+    "10440": "successful",
+    "10441": "failure",
+    "10442": "failure",
+    "10443": "failure",
+    "10444": "failure",
+    "10445": "failure",
+    "10446": "failure",
+    "10447": "failure",
+    "10448": "failure",
+    "10449": "failure",
+    "10450": "failure",
+    "10451": "failure",
+    "10452": "failure",
+    "10453": "infeasible",
+    "10454": "failure",
+    "10455": "failure",
+    "10456": "infeasible",
+    "10457": "failure",
+    "10458": "failure",
+    "10459": "failure",
+    "10460": "failure",
+    "10461": "failure",
+    "10462": "failure",
+    "10463": "infeasible",
+    "10464": "infeasible",
+    "10465": "infeasible",
+    "10466": "failure",
+    "10467": "failure",
+    "10468": "failure",
+    "10469": "failure",
+    "10470": "failure",
+    "10471": "failure",
+    "10472": "failure",
+    "10473": "failure",
+    "10474": "failure",
+    "10475": "failure",
+    "10476": "successful",
+    "10477": "successful",
+    "10478": "successful",
+    "10479": "failure",
+    "10480": "failure",
+    "10481": "failure",
+    "10482": "infeasible",
+    "10483": "failure",
+    "10484": "failure",
+    "10485": "failure",
+    "10486": "failure",
+    "10487": "failure",
+    "10488": "failure",
+    "10489": "failure",
+    "10490": "failure",
+    "10491": "successful",
+    "10492": "failure",
+    "10493": "failure",
+    "10494": "infeasible",
+    "10495": "failure",
+    "10496": "failure",
+    "10497": "failure",
+    "10498": "failure",
+    "10499": "failure",
+    "10500": "successful",
+    "10501": "failure",
+    "10502": "failure",
+    "10503": "infeasible",
+    "10504": "failure",
+    "10505": "failure",
+    "10506": "successful",
+    "10507": "failure",
+    "10508": "failure",
+    "10509": "failure",
+    "10510": "failure",
+    "10511": "failure",
+    "10512": "failure",
+    "10513": "failure",
+    "10514": "infeasible",
+    "10515": "failure",
+    "10516": "failure",
+    "10517": "failure",
+    "10518": "infeasible",
+    "10519": "failure",
+    "10520": "failure",
+    "10521": "failure",
+    "10522": "successful",
+    "10523": "failure",
+    "10524": "failure",
+    "10525": "failure",
+    "10526": "failure",
+    "10527": "failure",
+    "10528": "failure",
+    "10529": "failure",
+    "10530": "failure",
+    "10531": "failure",
+    "10532": "failure",
+    "10533": "failure",
+    "10534": "failure",
+    "10535": "failure",
+    "10536": "successful",
+    "10537": "failure",
+    "10538": "failure",
+    "10539": "failure",
+    "10540": "failure",
+    "10541": "successful",
+    "10542": "failure",
+    "10543": "failure",
+    "10544": "failure",
+    "10545": "failure",
+    "10546": "failure",
+    "10547": "failure",
+    "10548": "failure",
+    "10549": "infeasible",
+    "10550": "failure",
+    "10551": "successful",
+    "10552": "infeasible",
+    "10553": "failure",
+    "10554": "failure",
+    "10555": "successful",
+    "10556": "failure",
+    "10557": "failure",
+    "10558": "failure",
+    "10559": "failure",
+    "10560": "successful",
+    "10561": "failure",
+    "10562": "failure",
+    "10563": "failure",
+    "10564": "successful",
+    "10565": "failure",
+    "10566": "failure",
+    "10567": "failure",
+    "10568": "failure",
+    "10569": "failure",
+    "10570": "failure",
+    "10571": "failure",
+    "10572": "failure",
+    "10573": "failure",
+    "10574": "failure",
+    "10575": "failure",
+    "10576": "failure",
+    "10577": "failure",
+    "10578": "failure",
+    "10579": "successful",
+    "10580": "successful",
+    "10581": "failure",
+    "10582": "infeasible",
+    "10583": "failure",
+    "10584": "failure",
+    "10585": "failure",
+    "10586": "failure",
+    "10587": "failure",
+    "10588": "successful",
+    "10589": "failure",
+    "10590": "failure",
+    "10591": "failure",
+    "10592": "failure",
+    "10593": "failure",
+    "10594": "infeasible",
+    "10595": "failure",
+    "10596": "successful",
+    "10597": "successful",
+    "10598": "failure",
+    "10599": "failure",
+    "10600": "failure",
+    "10601": "failure",
+    "10602": "infeasible",
+    "10603": "successful",
+    "10604": "failure",
+    "10605": "failure",
+    "10606": "failure",
+    "10607": "infeasible",
+    "10608": "failure",
+    "10609": "successful",
+    "10610": "failure",
+    "10611": "failure",
+    "10612": "failure",
+    "10613": "failure",
+    "10614": "successful",
+    "10615": "successful",
+    "10616": "infeasible",
+    "10617": "failure",
+    "10618": "failure",
+    "10619": "failure",
+    "10620": "failure",
+    "10621": "infeasible",
+    "10622": "failure",
+    "10623": "infeasible",
+    "10624": "failure",
+    "10625": "failure",
+    "10626": "failure",
+    "10627": "failure",
+    "10628": "failure",
+    "10629": "failure",
+    "10630": "failure",
+    "10631": "failure",
+    "10632": "failure",
+    "10633": "failure",
+    "10634": "failure",
+    "10635": "failure",
+    "10636": "failure",
+    "10637": "infeasible",
+    "10638": "failure",
+    "10639": "infeasible",
+    "10640": "infeasible",
+    "10641": "failure",
+    "10642": "infeasible",
+    "10643": "failure",
+    "10644": "infeasible",
+    "10645": "failure",
+    "10646": "failure",
+    "10647": "infeasible",
+    "10648": "failure",
+    "10649": "failure",
+    "10650": "successful",
+    "10651": "failure",
+    "10652": "failure",
+    "10653": "failure",
+    "10654": "successful",
+    "10655": "failure",
+    "10656": "failure",
+    "10657": "failure",
+    "10658": "failure",
+    "10659": "failure",
+    "10660": "failure",
+    "10661": "failure",
+    "10662": "infeasible",
+    "10663": "successful",
+    "10664": "failure",
+    "10665": "failure",
+    "10666": "failure",
+    "10667": "failure",
+    "10668": "infeasible",
+    "10669": "successful",
+    "10670": "failure",
+    "10671": "failure",
+    "10672": "failure",
+    "10673": "already_complete",
+    "10674": "failure",
+    "10675": "failure",
+    "10676": "failure",
+    "10677": "failure",
+    "10678": "failure",
+    "10679": "failure",
+    "10680": "failure",
+    "10681": "failure",
+    "10682": "failure",
+    "10683": "failure",
+    "10684": "failure",
+    "10685": "failure",
+    "10686": "failure",
+    "10687": "failure",
+    "10688": "failure",
+    "10689": "successful",
+    "10690": "failure",
+    "10691": "infeasible",
+    "10692": "successful",
+    "10693": "failure",
+    "10694": "failure",
+    "10695": "failure",
+    "10696": "failure",
+    "10697": "failure",
+    "10698": "successful",
+    "10699": "infeasible",
+    "10700": "successful",
+    "10701": "already_complete",
+    "10702": "failure",
+    "10703": "failure",
+    "10704": "successful",
+    "10705": "failure",
+    "10706": "failure",
+    "10707": "infeasible",
+    "10708": "failure",
+    "10709": "failure",
+    "10710": "failure",
+    "10711": "infeasible",
+    "10712": "failure",
+    "10713": "failure",
+    "10714": "failure",
+    "10715": "failure",
+    "10716": "failure",
+    "10717": "failure",
+    "10718": "failure",
+    "10719": "successful",
+    "10720": "failure",
+    "10721": "infeasible",
+    "10722": "failure",
+    "10723": "failure",
+    "10724": "failure",
+    "10725": "failure",
+    "10726": "failure",
+    "10727": "failure",
+    "10728": "failure",
+    "10729": "failure",
+    "10730": "infeasible",
+    "10731": "failure",
+    "10732": "failure",
+    "10733": "failure",
+    "10734": "failure",
+    "10735": "failure",
+    "10736": "infeasible",
+    "10737": "failure",
+    "10738": "failure",
+    "10739": "failure",
+    "10740": "failure",
+    "10741": "infeasible",
+    "10742": "failure",
+    "10743": "failure",
+    "10744": "failure",
+    "10745": "failure",
+    "10746": "failure",
+    "10747": "failure",
+    "10748": "infeasible",
+    "10749": "failure",
+    "10750": "successful",
+    "10751": "failure",
+    "10752": "failure",
+    "10753": "infeasible",
+    "10754": "failure",
+    "10755": "failure",
+    "10756": "infeasible",
+    "10757": "failure",
+    "10758": "failure",
+    "10759": "failure",
+    "10760": "failure",
+    "10761": "failure",
+    "10762": "failure",
+    "10763": "failure",
+    "10764": "failure",
+    "10765": "failure",
+    "10766": "failure",
+    "10767": "failure",
+    "10768": "failure",
+    "10769": "failure",
+    "10770": "failure",
+    "10771": "failure",
+    "10772": "failure",
+    "10773": "failure",
+    "10774": "failure",
+    "10775": "failure",
+    "10776": "failure",
+    "10777": "failure",
+    "10778": "failure",
+    "10779": "failure",
+    "10780": "successful",
+    "10781": "failure",
+    "10782": "failure",
+    "10783": "failure",
+    "10784": "failure",
+    "10785": "failure",
+    "10786": "failure",
+    "10787": "failure",
+    "10788": "failure",
+    "10789": "failure",
+    "10790": "failure",
+    "10791": "failure",
+    "10792": "failure",
+    "10793": "failure",
+    "10794": "failure",
+    "10795": "failure",
+    "10796": "failure",
+    "10797": "successful",
+    "10798": "failure",
+    "10799": "successful",
+    "10800": "failure",
+    "10801": "failure",
+    "10802": "failure",
+    "10803": "failure",
+    "10804": "failure",
+    "10805": "failure",
+    "10806": "failure",
+    "10807": "infeasible",
+    "10808": "failure",
+    "10809": "failure",
+    "10810": "failure",
+    "10811": "successful",
+    "10812": "failure",
+    "10813": "failure",
+    "10814": "failure",
+    "10815": "successful",
+    "10816": "failure",
+    "10817": "failure",
+    "10818": "failure",
+    "10819": "failure",
+    "10820": "failure",
+    "10821": "failure",
+    "10822": "failure",
+    "10823": "failure",
+    "10824": "failure",
+    "10825": "failure",
+    "10826": "failure",
+    "10827": "failure",
+    "10828": "failure",
+    "10829": "failure",
+    "10830": "successful",
+    "10831": "successful",
+    "10832": "successful",
+    "10833": "failure",
+    "10834": "infeasible",
+    "10835": "failure",
+    "10836": "failure",
+    "10837": "failure",
+    "10838": "failure",
+    "10839": "successful",
+    "10840": "failure",
+    "10841": "failure",
+    "10842": "failure",
+    "10843": "failure",
+    "10844": "failure",
+    "10845": "failure",
+    "10846": "failure",
+    "10847": "failure",
+    "10848": "failure",
+    "10849": "infeasible",
+    "10850": "failure",
+    "10851": "failure",
+    "10852": "failure",
+    "10853": "failure",
+    "10854": "failure",
+    "10855": "failure",
+    "10856": "failure",
+    "10857": "failure",
+    "10858": "failure",
+    "10859": "infeasible",
+    "10860": "failure",
+    "10861": "failure",
+    "10862": "successful",
+    "10863": "failure",
+    "10864": "failure",
+    "10865": "successful",
+    "10866": "failure",
+    "10867": "failure",
+    "10868": "successful",
+    "10869": "failure",
+    "10870": "failure",
+    "10871": "failure",
+    "10872": "failure",
+    "10873": "failure",
+    "10874": "failure",
+    "10875": "successful",
+    "10876": "infeasible",
+    "10877": "failure",
+    "10878": "failure",
+    "10879": "failure",
+    "10880": "failure",
+    "10881": "failure",
+    "10882": "failure",
+    "10883": "failure",
+    "10884": "failure",
+    "10885": "failure",
+    "10886": "failure",
+    "10887": "failure",
+    "10888": "successful",
+    "10889": "failure",
+    "10890": "failure",
+    "10891": "failure",
+    "10892": "successful",
+    "10893": "failure",
+    "10894": "failure",
+    "10895": "failure",
+    "10896": "failure",
+    "10897": "failure",
+    "10898": "infeasible",
+    "10899": "failure",
+    "10900": "failure",
+    "10901": "successful",
+    "10902": "failure",
+    "10903": "failure",
+    "10904": "failure",
+    "10905": "failure",
+    "10906": "failure",
+    "10907": "failure",
+    "10908": "failure",
+    "10909": "failure",
+    "10910": "failure",
+    "10911": "failure",
+    "10912": "failure",
+    "10913": "failure",
+    "10914": "failure",
+    "10915": "failure",
+    "10916": "failure",
+    "10917": "failure",
+    "10918": "failure",
+    "10919": "failure",
+    "10920": "successful",
+    "10921": "failure",
+    "10922": "failure",
+    "10923": "failure",
+    "10924": "failure",
+    "10925": "failure",
+    "10926": "failure",
+    "10927": "failure",
+    "10928": "failure",
+    "10929": "failure",
+    "10930": "failure",
+    "10931": "failure",
+    "10932": "infeasible",
+    "10933": "infeasible",
+    "10934": "failure",
+    "10935": "failure",
+    "10936": "failure",
+    "10937": "successful",
+    "10938": "failure",
+    "10939": "failure",
+    "10940": "failure",
+    "10941": "failure",
+    "10942": "failure",
+    "10943": "failure",
+    "10944": "infeasible",
+    "10945": "failure",
+    "10946": "failure",
+    "10947": "failure",
+    "10948": "failure",
+    "10949": "failure",
+    "10950": "failure",
+    "10951": "failure",
+    "10952": "failure",
+    "10953": "successful",
+    "10954": "failure",
+    "10955": "failure",
+    "10956": "failure",
+    "10957": "failure",
+    "10958": "failure",
+    "10959": "failure",
+    "10960": "failure",
+    "10961": "failure",
+    "10962": "failure",
+    "10963": "failure",
+    "10964": "failure",
+    "10965": "failure",
+    "10966": "failure",
+    "10967": "failure",
+    "10968": "failure",
+    "10969": "failure",
+    "10970": "infeasible",
+    "10971": "infeasible",
+    "10972": "failure",
+    "10973": "failure",
+    "10974": "failure",
+    "10975": "failure",
+    "10976": "failure",
+    "10977": "failure",
+    "10978": "failure",
+    "10979": "failure",
+    "10980": "failure",
+    "10981": "failure",
+    "10982": "failure",
+    "10983": "failure",
+    "10984": "failure",
+    "10985": "failure",
+    "10986": "failure",
+    "10987": "failure",
+    "10988": "failure",
+    "10989": "failure",
+    "10990": "failure",
+    "10991": "failure",
+    "10992": "failure",
+    "10993": "failure",
+    "10994": "failure",
+    "10995": "failure",
+    "10996": "infeasible",
+    "10997": "failure",
+    "10998": "failure",
+    "10999": "failure",
+    "11000": "failure",
+    "11001": "failure",
+    "11002": "failure",
+    "11003": "failure",
+    "11004": "failure",
+    "11005": "failure",
+    "11006": "infeasible",
+    "11007": "failure",
+    "11008": "failure",
+    "11009": "failure",
+    "11010": "failure",
+    "11011": "failure",
+    "11012": "failure",
+    "11013": "failure",
+    "11014": "infeasible",
+    "11015": "failure",
+    "11016": "failure",
+    "11017": "successful",
+    "11018": "failure",
+    "11019": "failure",
+    "11020": "successful",
+    "11021": "infeasible",
+    "11022": "failure",
+    "11023": "infeasible",
+    "11024": "failure",
+    "11025": "failure",
+    "11026": "failure",
+    "11027": "infeasible",
+    "11028": "failure",
+    "11029": "failure",
+    "11030": "successful",
+    "11031": "successful",
+    "11032": "failure",
+    "11033": "failure",
+    "11034": "failure",
+    "11035": "failure",
+    "11036": "failure",
+    "11037": "failure",
+    "11038": "failure",
+    "11039": "failure",
+    "11040": "failure",
+    "11041": "successful",
+    "11042": "successful",
+    "11043": "infeasible",
+    "11044": "failure",
+    "11045": "failure",
+    "11046": "successful",
+    "11047": "failure",
+    "11048": "failure",
+    "11049": "failure",
+    "11050": "failure",
+    "11051": "failure",
+    "11052": "failure",
+    "11053": "failure",
+    "11054": "failure",
+    "11055": "failure",
+    "11056": "failure",
+    "11057": "successful",
+    "11058": "failure",
+    "11059": "failure",
+    "11060": "failure",
+    "11061": "failure",
+    "11062": "failure",
+    "11063": "successful",
+    "11064": "failure",
+    "11065": "failure",
+    "11066": "failure",
+    "11067": "failure",
+    "11068": "failure",
+    "11069": "failure",
+    "11070": "successful",
+    "11071": "failure",
+    "11072": "failure",
+    "11073": "failure",
+    "11074": "failure",
+    "11075": "failure",
+    "11076": "failure",
+    "11077": "failure",
+    "11078": "failure",
+    "11079": "failure",
+    "11080": "failure",
+    "11081": "failure",
+    "11082": "infeasible",
+    "11083": "failure",
+    "11084": "failure",
+    "11085": "failure",
+    "11086": "failure",
+    "11087": "failure",
+    "11088": "failure",
+    "11089": "failure",
+    "11090": "failure",
+    "11091": "successful",
+    "11092": "failure",
+    "11093": "failure",
+    "11094": "failure",
+    "11095": "failure",
+    "11096": "failure",
+    "11097": "failure",
+    "11098": "failure",
+    "11099": "failure",
+    "11100": "infeasible",
+    "11101": "failure",
+    "11102": "failure",
+    "11103": "failure",
+    "11104": "failure",
+    "11105": "failure",
+    "11106": "failure",
+    "11107": "failure",
+    "11108": "failure",
+    "11109": "failure",
+    "11110": "failure",
+    "11111": "failure",
+    "11112": "failure",
+    "11113": "failure",
+    "11114": "already_complete",
+    "11115": "failure",
+    "11116": "failure",
+    "11117": "failure",
+    "11118": "infeasible",
+    "11119": "failure",
+    "11120": "failure",
+    "11121": "infeasible",
+    "11122": "failure",
+    "11123": "infeasible",
+    "11124": "failure",
+    "11125": "infeasible",
+    "11126": "failure",
+    "11127": "failure",
+    "11128": "failure",
+    "11129": "failure",
+    "11130": "failure",
+    "11131": "failure",
+    "11132": "failure",
+    "11133": "failure",
+    "11134": "failure",
+    "11135": "failure",
+    "11136": "failure",
+    "11137": "infeasible",
+    "11138": "failure",
+    "11139": "failure",
+    "11140": "failure",
+    "11141": "infeasible",
+    "11142": "infeasible",
+    "11143": "failure",
+    "11144": "successful",
+    "11145": "failure",
+    "11146": "failure",
+    "11147": "failure",
+    "11148": "failure",
+    "11149": "failure",
+    "11150": "infeasible",
+    "11151": "failure",
+    "11152": "successful",
+    "11153": "failure",
+    "11154": "infeasible",
+    "11155": "failure",
+    "11156": "failure",
+    "11157": "infeasible",
+    "11158": "successful",
+    "11159": "failure",
+    "11160": "failure",
+    "11161": "failure",
+    "11162": "failure",
+    "11163": "successful",
+    "11164": "infeasible",
+    "11165": "failure",
+    "11166": "failure",
+    "11167": "failure",
+    "11168": "successful",
+    "11169": "failure",
+    "11170": "failure",
+    "11171": "failure",
+    "11172": "failure",
+    "11173": "failure",
+    "11174": "failure",
+    "11175": "failure",
+    "11176": "failure",
+    "11177": "failure",
+    "11178": "failure",
+    "11179": "failure",
+    "11180": "failure",
+    "11181": "failure",
+    "11182": "failure",
+    "11183": "failure",
+    "11184": "failure",
+    "11185": "failure",
+    "11186": "failure",
+    "11187": "failure",
+    "11188": "failure",
+    "11189": "failure",
+    "11190": "successful",
+    "11191": "failure",
+    "11192": "infeasible",
+    "11193": "failure",
+    "11194": "failure",
+    "11195": "failure",
+    "11196": "failure",
+    "11197": "failure",
+    "11198": "failure",
+    "11199": "failure",
+    "11200": "failure",
+    "11201": "failure",
+    "11202": "failure",
+    "11203": "failure",
+    "11204": "infeasible",
+    "11205": "failure",
+    "11206": "failure",
+    "11207": "failure",
+    "11208": "failure",
+    "11209": "failure",
+    "11210": "failure",
+    "11211": "failure",
+    "11212": "failure",
+    "11213": "failure",
+    "11214": "failure",
+    "11215": "failure",
+    "11216": "failure",
+    "11217": "failure",
+    "11218": "failure",
+    "11219": "failure",
+    "11220": "failure",
+    "11221": "failure",
+    "11222": "failure",
+    "11223": "infeasible",
+    "11224": "failure",
+    "11225": "successful",
+    "11226": "failure",
+    "11227": "failure",
+    "11228": "failure",
+    "11229": "failure",
+    "11230": "successful",
+    "11231": "failure",
+    "11232": "failure",
+    "11233": "failure",
+    "11234": "failure",
+    "11235": "failure",
+    "11236": "failure",
+    "11237": "failure",
+    "11238": "failure",
+    "11239": "failure",
+    "11240": "successful",
+    "11241": "failure",
+    "11242": "successful",
+    "11243": "failure",
+    "11244": "infeasible",
+    "11245": "failure",
+    "11246": "failure",
+    "11247": "failure",
+    "11248": "failure",
+    "11249": "failure",
+    "11250": "failure",
+    "11251": "failure",
+    "11252": "infeasible",
+    "11253": "failure",
+    "11254": "infeasible",
+    "11255": "failure",
+    "11256": "failure",
+    "11257": "infeasible",
+    "11258": "failure",
+    "11259": "failure",
+    "11260": "failure",
+    "11261": "failure",
+    "11262": "infeasible",
+    "11263": "failure",
+    "11264": "failure",
+    "11265": "failure",
+    "11266": "failure",
+    "11267": "failure",
+    "11268": "failure",
+    "11269": "failure",
+    "11270": "failure",
+    "11271": "failure",
+    "11272": "failure",
+    "11273": "failure",
+    "11274": "failure",
+    "11275": "failure",
+    "11276": "failure",
+    "11277": "failure",
+    "11278": "failure",
+    "11279": "failure",
+    "11280": "failure",
+    "11281": "failure",
+    "11282": "failure",
+    "11283": "failure",
+    "11284": "failure",
+    "11285": "failure",
+    "11286": "failure",
+    "11287": "successful",
+    "11288": "failure",
+    "11289": "failure",
+    "11290": "failure",
+    "11291": "infeasible",
+    "11292": "failure",
+    "11293": "failure",
+    "11294": "failure",
+    "11295": "successful",
+    "11296": "failure",
+    "11297": "successful",
+    "11298": "failure",
+    "11299": "failure",
+    "11300": "failure",
+    "11301": "failure",
+    "11302": "failure",
+    "11303": "successful",
+    "11304": "successful",
+    "11305": "failure",
+    "11306": "failure",
+    "11307": "failure",
+    "11308": "failure",
+    "11309": "infeasible",
+    "11310": "successful",
+    "11311": "failure",
+    "11312": "failure",
+    "11313": "failure",
+    "11314": "failure",
+    "11315": "failure",
+    "11316": "failure",
+    "11317": "failure",
+    "11318": "successful",
+    "11319": "failure",
+    "11320": "infeasible",
+    "11321": "failure",
+    "11322": "failure",
+    "11323": "failure",
+    "11324": "failure",
+    "11325": "failure",
+    "11326": "failure",
+    "11327": "failure",
+    "11328": "failure",
+    "11329": "failure",
+    "11330": "infeasible",
+    "11331": "failure",
+    "11332": "failure",
+    "11333": "failure",
+    "11334": "failure",
+    "11335": "failure",
+    "11336": "failure",
+    "11337": "failure",
+    "11338": "failure",
+    "11339": "failure",
+    "11340": "failure",
+    "11341": "failure",
+    "11342": "failure",
+    "11343": "failure",
+    "11344": "failure",
+    "11345": "failure",
+    "11346": "failure",
+    "11347": "infeasible",
+    "11348": "failure",
+    "11349": "failure",
+    "11350": "failure",
+    "11351": "failure",
+    "11352": "failure",
+    "11353": "failure",
+    "11354": "successful",
+    "11355": "infeasible",
+    "11356": "failure",
+    "11357": "failure",
+    "11358": "successful",
+    "11359": "failure",
+    "11360": "failure",
+    "11361": "failure",
+    "11362": "failure",
+    "11363": "failure",
+    "11364": "successful",
+    "11365": "failure",
+    "11366": "failure",
+    "11367": "infeasible",
+    "11368": "failure",
+    "11369": "failure",
+    "11370": "failure",
+    "11371": "infeasible",
+    "11372": "failure",
+    "11373": "failure",
+    "11374": "failure",
+    "11375": "failure",
+    "11376": "failure",
+    "11377": "failure",
+    "11378": "failure",
+    "11379": "failure",
+    "11380": "failure",
+    "11381": "failure",
+    "11382": "failure",
+    "11383": "failure",
+    "11384": "failure",
+    "11385": "failure",
+    "11386": "failure",
+    "11387": "failure",
+    "11388": "failure",
+    "11389": "failure",
+    "11390": "failure",
+    "11391": "failure",
+    "11392": "failure",
+    "11393": "failure",
+    "11394": "failure",
+    "11395": "failure",
+    "11396": "failure",
+    "11397": "infeasible",
+    "11398": "failure",
+    "11399": "failure",
+    "11400": "successful",
+    "11401": "successful",
+    "11402": "failure",
+    "11403": "infeasible",
+    "11404": "failure",
+    "11405": "failure",
+    "11406": "failure",
+    "11407": "failure",
+    "11408": "failure",
+    "11409": "failure",
+    "11410": "failure",
+    "11411": "failure",
+    "11412": "failure",
+    "11413": "failure",
+    "11414": "failure",
+    "11415": "failure",
+    "11416": "failure",
+    "11417": "failure",
+    "11418": "failure",
+    "11419": "failure",
+    "11420": "infeasible",
+    "11421": "failure",
+    "11422": "failure",
+    "11423": "successful",
+    "11424": "failure",
+    "11425": "failure",
+    "11426": "failure",
+    "11427": "failure",
+    "11428": "infeasible",
+    "11429": "failure",
+    "11430": "failure",
+    "11431": "failure",
+    "11432": "failure",
+    "11433": "failure",
+    "11434": "failure",
+    "11435": "failure",
+    "11436": "infeasible",
+    "11437": "failure",
+    "11438": "successful",
+    "11439": "failure",
+    "11440": "successful",
+    "11441": "infeasible",
+    "11442": "failure",
+    "11443": "failure",
+    "11444": "failure",
+    "11445": "failure",
+    "11446": "failure",
+    "11447": "failure",
+    "11448": "failure",
+    "11449": "infeasible",
+    "11450": "failure",
+    "11451": "infeasible",
+    "11452": "failure",
+    "11453": "failure",
+    "11454": "infeasible",
+    "11455": "failure",
+    "11456": "failure",
+    "11457": "failure",
+    "11458": "failure",
+    "11459": "failure",
+    "11460": "failure",
+    "11461": "successful",
+    "11462": "failure",
+    "11463": "failure",
+    "11464": "failure",
+    "11465": "failure",
+    "11466": "failure",
+    "11467": "failure",
+    "11468": "failure",
+    "11469": "failure",
+    "11470": "failure",
+    "11471": "failure",
+    "11472": "infeasible",
+    "11473": "failure",
+    "11474": "failure",
+    "11475": "failure",
+    "11476": "failure",
+    "11477": "failure",
+    "11478": "failure",
+    "11479": "infeasible",
+    "11480": "failure",
+    "11481": "failure",
+    "11482": "failure",
+    "11483": "failure",
+    "11484": "infeasible",
+    "11485": "failure",
+    "11486": "failure",
+    "11487": "failure",
+    "11488": "failure",
+    "11489": "failure",
+    "11490": "failure",
+    "11491": "failure",
+    "11492": "successful",
+    "11493": "failure",
+    "11494": "failure",
+    "11495": "successful",
+    "11496": "failure",
+    "11497": "failure",
+    "11498": "failure",
+    "11499": "already_complete",
+    "11500": "successful",
+    "11501": "infeasible",
+    "11502": "failure",
+    "11503": "infeasible",
+    "11504": "failure",
+    "11505": "failure",
+    "11506": "failure",
+    "11507": "failure",
+    "11508": "failure",
+    "11509": "failure",
+    "11510": "failure",
+    "11511": "failure",
+    "11512": "successful",
+    "11513": "successful",
+    "11514": "failure",
+    "11515": "failure",
+    "11516": "infeasible",
+    "11517": "infeasible",
+    "11518": "failure",
+    "11519": "failure",
+    "11520": "failure",
+    "11521": "failure",
+    "11522": "infeasible",
+    "11523": "failure",
+    "11524": "failure",
+    "11525": "failure",
+    "11526": "failure",
+    "11527": "successful",
+    "11528": "failure",
+    "11529": "failure",
+    "11530": "failure",
+    "11531": "failure",
+    "11532": "failure",
+    "11533": "failure",
+    "11534": "infeasible",
+    "11535": "failure",
+    "11536": "failure",
+    "11537": "failure",
+    "11538": "infeasible",
+    "11539": "failure",
+    "11540": "failure",
+    "11541": "successful",
+    "11542": "failure",
+    "11543": "failure",
+    "11544": "failure",
+    "11545": "failure",
+    "11546": "failure",
+    "11547": "failure",
+    "11548": "failure",
+    "11549": "failure",
+    "11550": "failure",
+    "11551": "successful",
+    "11552": "failure",
+    "11553": "failure",
+    "11554": "failure",
+    "11555": "failure",
+    "11556": "successful",
+    "11557": "failure",
+    "11558": "failure",
+    "11559": "failure",
+    "11560": "failure",
+    "11561": "infeasible",
+    "11562": "failure",
+    "11563": "failure",
+    "11564": "failure",
+    "11565": "failure",
+    "11566": "failure",
+    "11567": "infeasible",
+    "11568": "failure",
+    "11569": "failure",
+    "11570": "infeasible",
+    "11571": "failure",
+    "11572": "failure",
+    "11573": "failure",
+    "11574": "infeasible",
+    "11575": "failure",
+    "11576": "failure",
+    "11577": "infeasible",
+    "11578": "infeasible",
+    "11579": "failure",
+    "11580": "failure",
+    "11581": "failure",
+    "11582": "failure",
+    "11583": "failure",
+    "11584": "successful",
+    "11585": "failure",
+    "11586": "failure",
+    "11587": "failure",
+    "11588": "failure",
+    "11589": "infeasible",
+    "11590": "failure",
+    "11591": "failure",
+    "11592": "failure",
+    "11593": "failure",
+    "11594": "successful",
+    "11595": "failure",
+    "11596": "failure",
+    "11597": "failure",
+    "11598": "infeasible",
+    "11599": "failure",
+    "11600": "failure",
+    "11601": "failure",
+    "11602": "failure",
+    "11603": "already_complete",
+    "11604": "infeasible",
+    "11605": "failure",
+    "11606": "infeasible",
+    "11607": "infeasible",
+    "11608": "failure",
+    "11609": "successful",
+    "11610": "successful",
+    "11611": "failure",
+    "11612": "failure",
+    "11613": "failure",
+    "11614": "successful",
+    "11615": "failure",
+    "11616": "successful",
+    "11617": "infeasible",
+    "11618": "failure",
+    "11619": "failure",
+    "11620": "failure",
+    "11621": "failure",
+    "11622": "failure",
+    "11623": "failure",
+    "11624": "failure",
+    "11625": "failure",
+    "11626": "successful",
+    "11627": "infeasible",
+    "11628": "failure",
+    "11629": "failure",
+    "11630": "failure",
+    "11631": "failure",
+    "11632": "failure",
+    "11633": "failure",
+    "11634": "failure",
+    "11635": "failure",
+    "11636": "failure",
+    "11637": "failure",
+    "11638": "failure",
+    "11639": "failure",
+    "11640": "successful",
+    "11641": "failure",
+    "11642": "failure",
+    "11643": "successful",
+    "11644": "failure",
+    "11645": "successful",
+    "11646": "infeasible",
+    "11647": "failure",
+    "11648": "infeasible",
+    "11649": "failure",
+    "11650": "failure",
+    "11651": "failure",
+    "11652": "failure",
+    "11653": "successful",
+    "11654": "failure",
+    "11655": "failure",
+    "11656": "successful",
+    "11657": "failure",
+    "11658": "successful",
+    "11659": "failure",
+    "11660": "failure",
+    "11661": "failure",
+    "11662": "infeasible",
+    "11663": "failure",
+    "11664": "failure",
+    "11665": "successful",
+    "11666": "failure",
+    "11667": "infeasible",
+    "11668": "successful",
+    "11669": "successful",
+    "11670": "failure",
+    "11671": "failure",
+    "11672": "failure",
+    "11673": "failure",
+    "11674": "failure",
+    "11675": "failure",
+    "11676": "failure",
+    "11677": "already_complete",
+    "11678": "failure",
+    "11679": "failure",
+    "11680": "failure",
+    "11681": "failure",
+    "11682": "failure",
+    "11683": "failure",
+    "11684": "failure",
+    "11685": "failure",
+    "11686": "failure",
+    "11687": "successful",
+    "11688": "failure",
+    "11689": "failure",
+    "11690": "failure",
+    "11691": "failure",
+    "11692": "failure",
+    "11693": "failure",
+    "11694": "failure",
+    "11695": "failure",
+    "11696": "failure",
+    "11697": "failure",
+    "11698": "failure",
+    "11699": "failure",
+    "11700": "failure",
+    "11701": "failure",
+    "11702": "failure",
+    "11703": "failure",
+    "11704": "failure",
+    "11705": "failure",
+    "11706": "failure",
+    "11707": "failure",
+    "11708": "failure",
+    "11709": "failure",
+    "11710": "failure",
+    "11711": "failure",
+    "11712": "failure",
+    "11713": "infeasible",
+    "11714": "failure",
+    "11715": "failure",
+    "11716": "failure",
+    "11717": "failure",
+    "11718": "successful",
+    "11719": "failure",
+    "11720": "failure",
+    "11721": "failure",
+    "11722": "failure",
+    "11723": "infeasible",
+    "11724": "failure",
+    "11725": "successful",
+    "11726": "failure",
+    "11727": "failure",
+    "11728": "failure",
+    "11729": "failure",
+    "11730": "successful",
+    "11731": "failure",
+    "11732": "failure",
+    "11733": "failure",
+    "11734": "failure",
+    "11735": "failure",
+    "11736": "successful",
+    "11737": "failure",
+    "11738": "failure",
+    "11739": "infeasible",
+    "11740": "failure",
+    "11741": "failure",
+    "11742": "failure",
+    "11743": "failure",
+    "11744": "successful",
+    "11745": "successful",
+    "11746": "failure",
+    "11747": "failure",
+    "11748": "failure",
+    "11749": "failure",
+    "11750": "failure",
+    "11751": "failure",
+    "11752": "failure",
+    "11753": "failure",
+    "11754": "failure",
+    "11755": "failure",
+    "11756": "failure",
+    "11757": "infeasible",
+    "11758": "failure",
+    "11759": "failure",
+    "11760": "infeasible",
+    "11761": "infeasible",
+    "11762": "failure",
+    "11763": "failure",
+    "11764": "failure",
+    "11765": "failure",
+    "11766": "failure",
+    "11767": "failure",
+    "11768": "infeasible",
+    "11769": "infeasible",
+    "11770": "failure",
+    "11771": "failure",
+    "11772": "failure",
+    "11773": "failure",
+    "11774": "failure",
+    "11775": "successful",
+    "11776": "failure",
+    "11777": "failure",
+    "11778": "failure",
+    "11779": "failure",
+    "11780": "successful",
+    "11781": "failure",
+    "11782": "failure",
+    "11783": "failure",
+    "11784": "failure",
+    "11785": "failure",
+    "11786": "failure",
+    "11787": "infeasible",
+    "11788": "failure",
+    "11789": "failure",
+    "11790": "failure",
+    "11791": "failure",
+    "11792": "infeasible",
+    "11793": "failure",
+    "11794": "failure",
+    "11795": "infeasible",
+    "11796": "failure",
+    "11797": "failure",
+    "11798": "failure",
+    "11799": "failure",
+    "11800": "failure",
+    "11801": "failure",
+    "11802": "failure",
+    "11803": "infeasible",
+    "11804": "failure",
+    "11805": "infeasible",
+    "11806": "failure",
+    "11807": "failure",
+    "11808": "failure",
+    "11809": "failure",
+    "11810": "failure",
+    "11811": "failure",
+    "11812": "failure",
+    "11813": "failure",
+    "11814": "failure",
+    "11815": "failure",
+    "11816": "failure",
+    "11817": "failure",
+    "11818": "failure",
+    "11819": "failure",
+    "11820": "failure",
+    "11821": "failure",
+    "11822": "failure",
+    "11823": "failure",
+    "11824": "failure",
+    "11825": "infeasible",
+    "11826": "failure",
+    "11827": "failure",
+    "11828": "failure",
+    "11829": "failure",
+    "11830": "successful",
+    "11831": "failure",
+    "11832": "failure",
+    "11833": "failure",
+    "11834": "failure",
+    "11835": "failure",
+    "11836": "failure",
+    "11837": "failure",
+    "11838": "failure",
+    "11839": "failure",
+    "11840": "failure",
+    "11841": "failure",
+    "11842": "failure",
+    "11843": "successful",
+    "11844": "failure",
+    "11845": "failure",
+    "11846": "failure",
+    "11847": "failure",
+    "11848": "failure",
+    "11849": "failure",
+    "11850": "successful",
+    "11851": "failure",
+    "11852": "failure",
+    "11853": "failure",
+    "11854": "failure",
+    "11855": "failure",
+    "11856": "infeasible",
+    "11857": "failure",
+    "11858": "failure",
+    "11859": "infeasible",
+    "11860": "infeasible",
+    "11861": "failure",
+    "11862": "infeasible",
+    "11863": "failure",
+    "11864": "failure",
+    "11865": "infeasible",
+    "11866": "failure",
+    "11867": "already_complete",
+    "11868": "failure",
+    "11869": "failure",
+    "11870": "infeasible",
+    "11871": "infeasible",
+    "11872": "failure",
+    "11873": "failure",
+    "11874": "failure",
+    "11875": "failure",
+    "11876": "failure",
+    "11877": "failure",
+    "11878": "successful",
+    "11879": "failure",
+    "11880": "failure",
+    "11881": "failure",
+    "11882": "failure",
+    "11883": "failure",
+    "11884": "failure",
+    "11885": "failure",
+    "11886": "failure",
+    "11887": "failure",
+    "11888": "successful",
+    "11889": "failure",
+    "11890": "failure",
+    "11891": "failure",
+    "11892": "failure",
+    "11893": "failure",
+    "11894": "successful",
+    "11895": "failure",
+    "11896": "infeasible",
+    "11897": "failure",
+    "11898": "failure",
+    "11899": "failure",
+    "11900": "infeasible",
+    "11901": "infeasible",
+    "11902": "failure",
+    "11903": "failure",
+    "11904": "failure",
+    "11905": "already_complete",
+    "11906": "failure",
+    "11907": "infeasible",
+    "11908": "failure",
+    "11909": "successful",
+    "11910": "infeasible",
+    "11911": "failure",
+    "11912": "failure",
+    "11913": "failure",
+    "11914": "failure",
+    "11915": "failure",
+    "11916": "failure",
+    "11917": "failure",
+    "11918": "failure",
+    "11919": "failure",
+    "11920": "failure",
+    "11921": "successful",
+    "11922": "failure",
+    "11923": "failure",
+    "11924": "failure",
+    "11925": "failure",
+    "11926": "failure",
+    "11927": "failure",
+    "11928": "failure",
+    "11929": "failure",
+    "11930": "failure",
+    "11931": "failure",
+    "11932": "failure",
+    "11933": "failure",
+    "11934": "failure",
+    "11935": "failure",
+    "11936": "failure",
+    "11937": "failure",
+    "11938": "failure",
+    "11939": "successful",
+    "11940": "successful",
+    "11941": "failure",
+    "11942": "failure",
+    "11943": "failure",
+    "11944": "failure",
+    "11945": "successful",
+    "11946": "failure",
+    "11947": "failure",
+    "11948": "failure",
+    "11949": "failure",
+    "11950": "failure",
+    "11951": "failure",
+    "11952": "failure",
+    "11953": "failure",
+    "11954": "failure",
+    "11955": "failure",
+    "11956": "failure",
+    "11957": "failure",
+    "11958": "infeasible",
+    "11959": "failure",
+    "11960": "successful",
+    "11961": "failure",
+    "11962": "infeasible",
+    "11963": "failure",
+    "11964": "failure",
+    "11965": "failure",
+    "11966": "failure",
+    "11967": "successful",
+    "11968": "failure",
+    "11969": "infeasible",
+    "11970": "failure",
+    "11971": "failure",
+    "11972": "failure",
+    "11973": "failure",
+    "11974": "failure",
+    "11975": "infeasible",
+    "11976": "failure",
+    "11977": "failure",
+    "11978": "failure",
+    "11979": "failure",
+    "11980": "failure",
+    "11981": "failure",
+    "11982": "failure",
+    "11983": "failure",
+    "11984": "failure",
+    "11985": "failure",
+    "11986": "failure",
+    "11987": "failure",
+    "11988": "failure",
+    "11989": "failure",
+    "11990": "failure",
+    "11991": "failure",
+    "11992": "failure",
+    "11993": "failure",
+    "11994": "failure",
+    "11995": "failure",
+    "11996": "failure",
+    "11997": "failure",
+    "11998": "failure",
+    "11999": "successful",
+    "12000": "failure",
+    "12001": "failure",
+    "12002": "failure",
+    "12003": "failure",
+    "12004": "infeasible",
+    "12005": "successful",
+    "12006": "failure",
+    "12007": "already_complete",
+    "12008": "failure",
+    "12009": "failure",
+    "12010": "failure",
+    "12011": "infeasible",
+    "12012": "failure",
+    "12013": "failure",
+    "12014": "failure",
+    "12015": "failure",
+    "12016": "failure",
+    "12017": "infeasible",
+    "12018": "failure",
+    "12019": "failure",
+    "12020": "failure",
+    "12021": "failure",
+    "12022": "failure",
+    "12023": "failure",
+    "12024": "failure",
+    "12025": "infeasible",
+    "12026": "failure",
+    "12027": "failure",
+    "12028": "failure",
+    "12029": "failure",
+    "12030": "failure",
+    "12031": "failure",
+    "12032": "failure",
+    "12033": "failure",
+    "12034": "failure",
+    "12035": "failure",
+    "12036": "failure",
+    "12037": "failure",
+    "12038": "failure",
+    "12039": "failure",
+    "12040": "failure",
+    "12041": "failure",
+    "12042": "failure",
+    "12043": "failure",
+    "12044": "failure",
+    "12045": "infeasible",
+    "12046": "failure",
+    "12047": "successful",
+    "12048": "failure",
+    "12049": "already_complete",
+    "12050": "failure",
+    "12051": "failure",
+    "12052": "failure",
+    "12053": "failure",
+    "12054": "failure",
+    "12055": "failure",
+    "12056": "failure",
+    "12057": "successful",
+    "12058": "failure",
+    "12059": "failure",
+    "12060": "failure",
+    "12061": "failure",
+    "12062": "failure",
+    "12063": "failure",
+    "12064": "failure",
+    "12065": "infeasible",
+    "12066": "failure",
+    "12067": "failure",
+    "12068": "failure",
+    "12069": "failure",
+    "12070": "failure",
+    "12071": "failure",
+    "12072": "failure",
+    "12073": "successful",
+    "12074": "failure",
+    "12075": "failure",
+    "12076": "failure",
+    "12077": "failure",
+    "12078": "failure",
+    "12079": "failure",
+    "12080": "failure",
+    "12081": "failure",
+    "12082": "failure",
+    "12083": "successful",
+    "12084": "failure",
+    "12085": "failure",
+    "12086": "failure",
+    "12087": "failure",
+    "12088": "failure",
+    "12089": "failure",
+    "12090": "failure",
+    "12091": "failure",
+    "12092": "failure",
+    "12093": "failure",
+    "12094": "failure",
+    "12095": "failure",
+    "12096": "failure",
+    "12097": "failure",
+    "12098": "failure",
+    "12099": "infeasible",
+    "12100": "successful",
+    "12101": "failure",
+    "12102": "failure",
+    "12103": "failure",
+    "12104": "failure",
+    "12105": "failure",
+    "12106": "failure",
+    "12107": "failure",
+    "12108": "failure",
+    "12109": "failure",
+    "12110": "successful",
+    "12111": "failure",
+    "12112": "failure",
+    "12113": "infeasible",
+    "12114": "infeasible",
+    "12115": "failure",
+    "12116": "failure",
+    "12117": "failure",
+    "12118": "infeasible",
+    "12119": "failure",
+    "12120": "failure",
+    "12121": "failure",
+    "12122": "failure",
+    "12123": "failure",
+    "12124": "failure",
+    "12125": "failure",
+    "12126": "failure",
+    "12127": "failure",
+    "12128": "infeasible",
+    "12129": "failure",
+    "12130": "failure",
+    "12131": "failure",
+    "12132": "failure",
+    "12133": "failure",
+    "12134": "failure",
+    "12135": "failure",
+    "12136": "failure",
+    "12137": "failure",
+    "12138": "failure",
+    "12139": "infeasible",
+    "12140": "infeasible",
+    "12141": "failure",
+    "12142": "failure",
+    "12143": "failure",
+    "12144": "failure",
+    "12145": "failure",
+    "12146": "failure",
+    "12147": "failure",
+    "12148": "failure",
+    "12149": "failure",
+    "12150": "failure",
+    "12151": "successful",
+    "12152": "failure",
+    "12153": "failure",
+    "12154": "failure",
+    "12155": "failure",
+    "12156": "failure",
+    "12157": "failure",
+    "12158": "failure",
+    "12159": "infeasible",
+    "12160": "failure",
+    "12161": "failure",
+    "12162": "failure",
+    "12163": "failure",
+    "12164": "failure",
+    "12165": "failure",
+    "12166": "failure",
+    "12167": "failure",
+    "12168": "failure",
+    "12169": "failure",
+    "12170": "successful",
+    "12171": "successful",
+    "12172": "failure",
+    "12173": "failure",
+    "12174": "failure",
+    "12175": "failure",
+    "12176": "failure",
+    "12177": "successful",
+    "12178": "infeasible",
+    "12179": "failure",
+    "12180": "failure",
+    "12181": "failure",
+    "12182": "failure",
+    "12183": "successful",
+    "12184": "failure",
+    "12185": "failure",
+    "12186": "failure",
+    "12187": "failure",
+    "12188": "failure",
+    "12189": "successful",
+    "12190": "infeasible",
+    "12191": "failure",
+    "12192": "failure",
+    "12193": "failure",
+    "12194": "failure",
+    "12195": "successful",
+    "12196": "failure",
+    "12197": "failure",
+    "12198": "failure",
+    "12199": "failure",
+    "12200": "failure",
+    "12201": "successful",
+    "12202": "failure",
+    "12203": "failure",
+    "12204": "infeasible",
+    "12205": "failure",
+    "12206": "failure",
+    "12207": "failure",
+    "12208": "failure",
+    "12209": "failure",
+    "12210": "failure",
+    "12211": "failure",
+    "12212": "failure",
+    "12213": "failure",
+    "12214": "failure",
+    "12215": "failure",
+    "12216": "failure",
+    "12217": "successful",
+    "12218": "failure",
+    "12219": "failure",
+    "12220": "failure",
+    "12221": "failure",
+    "12222": "successful",
+    "12223": "successful",
+    "12224": "infeasible",
+    "12225": "successful",
+    "12226": "failure",
+    "12227": "infeasible",
+    "12228": "failure",
+    "12229": "failure",
+    "12230": "failure",
+    "12231": "failure",
+    "12232": "failure",
+    "12233": "failure",
+    "12234": "failure",
+    "12235": "failure",
+    "12236": "failure",
+    "12237": "failure",
+    "12238": "infeasible",
+    "12239": "failure",
+    "12240": "successful",
+    "12241": "failure",
+    "12242": "failure",
+    "12243": "already_complete",
+    "12244": "failure",
+    "12245": "failure",
+    "12246": "successful",
+    "12247": "successful",
+    "12248": "failure",
+    "12249": "failure",
+    "12250": "failure",
+    "12251": "successful",
+    "12252": "failure",
+    "12253": "successful",
+    "12254": "successful",
+    "12255": "infeasible",
+    "12256": "failure",
+    "12257": "failure",
+    "12258": "failure",
+    "12259": "failure",
+    "12260": "failure",
+    "12261": "failure",
+    "12262": "failure",
+    "12263": "failure",
+    "12264": "failure",
+    "12265": "failure",
+    "12266": "infeasible",
+    "12267": "failure",
+    "12268": "failure",
+    "12269": "failure",
+    "12270": "failure",
+    "12271": "failure",
+    "12272": "failure",
+    "12273": "failure",
+    "12274": "failure",
+    "12275": "infeasible",
+    "12276": "failure",
+    "12277": "failure",
+    "12278": "infeasible",
+    "12279": "successful",
+    "12280": "failure",
+    "12281": "failure",
+    "12282": "failure",
+    "12283": "infeasible",
+    "12284": "failure",
+    "12285": "failure",
+    "12286": "failure",
+    "12287": "failure",
+    "12288": "failure",
+    "12289": "failure",
+    "12290": "successful",
+    "12291": "infeasible",
+    "12292": "successful",
+    "12293": "failure",
+    "12294": "successful",
+    "12295": "failure",
+    "12296": "successful",
+    "12297": "failure",
+    "12298": "infeasible",
+    "12299": "failure",
+    "12300": "failure",
+    "12301": "failure",
+    "12302": "failure",
+    "12303": "failure",
+    "12304": "failure",
+    "12305": "failure",
+    "12306": "failure",
+    "12307": "infeasible",
+    "12308": "failure",
+    "12309": "successful",
+    "12310": "failure",
+    "12311": "failure",
+    "12312": "failure",
+    "12313": "infeasible",
+    "12314": "infeasible",
+    "12315": "failure",
+    "12316": "successful",
+    "12317": "failure",
+    "12318": "failure",
+    "12319": "failure",
+    "12320": "failure",
+    "12321": "failure",
+    "12322": "failure",
+    "12323": "successful",
+    "12324": "failure",
+    "12325": "failure",
+    "12326": "failure",
+    "12327": "failure",
+    "12328": "failure",
+    "12329": "failure",
+    "12330": "failure",
+    "12331": "successful",
+    "12332": "failure",
+    "12333": "failure",
+    "12334": "successful",
+    "12335": "failure",
+    "12336": "failure",
+    "12337": "infeasible",
+    "12338": "failure",
+    "12339": "failure",
+    "12340": "failure",
+    "12341": "infeasible",
+    "12342": "failure",
+    "12343": "failure",
+    "12344": "failure",
+    "12345": "successful",
+    "12346": "failure",
+    "12347": "infeasible",
+    "12348": "failure",
+    "12349": "failure",
+    "12350": "successful",
+    "12351": "successful",
+    "12352": "failure",
+    "12353": "failure",
+    "12354": "failure",
+    "12355": "failure",
+    "12356": "failure",
+    "12357": "failure",
+    "12358": "failure",
+    "12359": "failure",
+    "12360": "successful",
+    "12361": "failure",
+    "12362": "failure",
+    "12363": "failure",
+    "12364": "infeasible",
+    "12365": "failure",
+    "12366": "failure",
+    "12367": "infeasible",
+    "12368": "failure",
+    "12369": "failure",
+    "12370": "failure",
+    "12371": "failure",
+    "12372": "failure",
+    "12373": "failure",
+    "12374": "failure",
+    "12375": "successful",
+    "12376": "failure",
+    "12377": "failure",
+    "12378": "failure",
+    "12379": "failure",
+    "12380": "failure",
+    "12381": "failure",
+    "12382": "infeasible",
+    "12383": "failure",
+    "12384": "failure",
+    "12385": "failure",
+    "12386": "failure",
+    "12387": "failure",
+    "12388": "failure",
+    "12389": "infeasible",
+    "12390": "failure",
+    "12391": "successful",
+    "12392": "failure",
+    "12393": "failure",
+    "12394": "failure",
+    "12395": "failure",
+    "12396": "failure",
+    "12397": "failure",
+    "12398": "failure",
+    "12399": "failure",
+    "12400": "failure",
+    "12401": "failure",
+    "12402": "failure",
+    "12403": "successful",
+    "12404": "failure",
+    "12405": "failure",
+    "12406": "failure",
+    "12407": "failure",
+    "12408": "failure",
+    "12409": "failure",
+    "12410": "failure",
+    "12411": "failure",
+    "12412": "failure",
+    "12413": "successful",
+    "12414": "successful",
+    "12415": "failure",
+    "12416": "failure",
+    "12417": "failure",
+    "12418": "infeasible",
+    "12419": "successful",
+    "12420": "successful",
+    "12421": "infeasible",
+    "12422": "successful",
+    "12423": "failure",
+    "12424": "failure",
+    "12425": "failure",
+    "12426": "failure",
+    "12427": "failure",
+    "12428": "failure",
+    "12429": "failure",
+    "12430": "infeasible",
+    "12431": "failure",
+    "12432": "infeasible",
+    "12433": "failure",
+    "12434": "failure",
+    "12435": "failure",
+    "12436": "failure",
+    "12437": "failure",
+    "12438": "failure",
+    "12439": "failure",
+    "12440": "failure",
+    "12441": "failure",
+    "12442": "failure",
+    "12443": "failure",
+    "12444": "failure",
+    "12445": "failure",
+    "12446": "failure",
+    "12447": "failure",
+    "12448": "infeasible",
+    "12449": "failure",
+    "12450": "failure",
+    "12451": "successful",
+    "12452": "infeasible",
+    "12453": "successful",
+    "12454": "failure",
+    "12455": "failure",
+    "12456": "failure",
+    "12457": "failure",
+    "12458": "failure",
+    "12459": "failure",
+    "12460": "failure",
+    "12461": "successful",
+    "12462": "failure",
+    "12463": "failure",
+    "12464": "successful",
+    "12465": "failure",
+    "12466": "failure",
+    "12467": "failure",
+    "12468": "failure",
+    "12469": "failure",
+    "12470": "successful",
+    "12471": "failure",
+    "12472": "failure",
+    "12473": "infeasible",
+    "12474": "failure",
+    "12475": "failure",
+    "12476": "failure",
+    "12477": "infeasible",
+    "12478": "failure",
+    "12479": "failure",
+    "12480": "failure",
+    "12481": "successful",
+    "12482": "failure",
+    "12483": "failure",
+    "12484": "failure",
+    "12485": "failure",
+    "12486": "infeasible",
+    "12487": "failure",
+    "12488": "infeasible",
+    "12489": "failure",
+    "12490": "failure",
+    "12491": "failure",
+    "12492": "failure",
+    "12493": "successful",
+    "12494": "failure",
+    "12495": "failure",
+    "12496": "failure",
+    "12497": "failure",
+    "12498": "failure",
+    "12499": "successful",
+    "12500": "successful",
+    "12501": "successful",
+    "12502": "successful",
+    "12503": "failure",
+    "12504": "failure",
+    "12505": "infeasible",
+    "12506": "failure",
+    "12507": "failure",
+    "12508": "failure",
+    "12509": "failure",
+    "12510": "failure",
+    "12511": "failure",
+    "12512": "failure",
+    "12513": "failure",
+    "12514": "failure",
+    "12515": "failure",
+    "12516": "infeasible",
+    "12517": "failure",
+    "12518": "infeasible",
+    "12519": "failure",
+    "12520": "failure",
+    "12521": "failure",
+    "12522": "failure",
+    "12523": "infeasible",
+    "12524": "failure",
+    "12525": "failure",
+    "12526": "failure",
+    "12527": "failure",
+    "12528": "infeasible",
+    "12529": "failure",
+    "12530": "successful",
+    "12531": "successful",
+    "12532": "infeasible",
+    "12533": "failure",
+    "12534": "failure",
+    "12535": "failure",
+    "12536": "failure",
+    "12537": "failure",
+    "12538": "failure",
+    "12539": "infeasible",
+    "12540": "failure",
+    "12541": "infeasible",
+    "12542": "successful",
+    "12543": "failure",
+    "12544": "failure",
+    "12545": "failure",
+    "12546": "failure",
+    "12547": "failure",
+    "12548": "failure",
+    "12549": "failure",
+    "12550": "failure",
+    "12551": "failure",
+    "12552": "failure",
+    "12553": "successful",
+    "12554": "failure",
+    "12555": "failure",
+    "12556": "failure",
+    "12557": "failure",
+    "12558": "successful",
+    "12559": "infeasible",
+    "12560": "failure",
+    "12561": "failure",
+    "12562": "failure",
+    "12563": "successful",
+    "12564": "failure",
+    "12565": "failure",
+    "12566": "failure",
+    "12567": "failure",
+    "12568": "failure",
+    "12569": "failure",
+    "12570": "failure",
+    "12571": "failure",
+    "12572": "successful",
+    "12573": "successful",
+    "12574": "failure",
+    "12575": "failure",
+    "12576": "failure",
+    "12577": "failure",
+    "12578": "failure",
+    "12579": "failure",
+    "12580": "failure",
+    "12581": "failure",
+    "12582": "failure",
+    "12583": "failure",
+    "12584": "failure",
+    "12585": "infeasible",
+    "12586": "failure",
+    "12587": "successful",
+    "12588": "failure",
+    "12589": "failure",
+    "12590": "failure",
+    "12591": "failure",
+    "12592": "successful",
+    "12593": "already_complete",
+    "12594": "failure",
+    "12595": "failure",
+    "12596": "failure",
+    "12597": "successful",
+    "12598": "failure",
+    "12599": "failure",
+    "12600": "successful",
+    "12601": "failure",
+    "12602": "failure",
+    "12603": "failure",
+    "12604": "failure",
+    "12605": "failure",
+    "12606": "failure",
+    "12607": "successful",
+    "12608": "failure",
+    "12609": "failure",
+    "12610": "successful",
+    "12611": "failure",
+    "12612": "infeasible",
+    "12613": "failure",
+    "12614": "infeasible",
+    "12615": "failure",
+    "12616": "failure",
+    "12617": "failure",
+    "12618": "successful",
+    "12619": "failure",
+    "12620": "failure",
+    "12621": "failure",
+    "12622": "failure",
+    "12623": "failure",
+    "12624": "failure",
+    "12625": "failure",
+    "12626": "failure",
+    "12627": "failure",
+    "12628": "failure",
+    "12629": "failure",
+    "12630": "failure",
+    "12631": "failure",
+    "12632": "failure",
+    "12633": "failure",
+    "12634": "failure",
+    "12635": "infeasible",
+    "12636": "failure",
+    "12637": "failure",
+    "12638": "failure",
+    "12639": "successful",
+    "12640": "successful",
+    "12641": "failure",
+    "12642": "failure",
+    "12643": "successful",
+    "12644": "failure",
+    "12645": "failure",
+    "12646": "failure",
+    "12647": "failure",
+    "12648": "failure",
+    "12649": "infeasible",
+    "12650": "failure",
+    "12651": "failure",
+    "12652": "failure",
+    "12653": "failure",
+    "12654": "infeasible",
+    "12655": "failure",
+    "12656": "failure",
+    "12657": "infeasible",
+    "12658": "failure",
+    "12659": "failure",
+    "12660": "failure",
+    "12661": "successful",
+    "12662": "failure",
+    "12663": "already_complete",
+    "12664": "failure",
+    "12665": "failure",
+    "12666": "failure",
+    "12667": "failure",
+    "12668": "failure",
+    "12669": "failure",
+    "12670": "failure",
+    "12671": "successful",
+    "12672": "failure",
+    "12673": "failure",
+    "12674": "successful",
+    "12675": "failure",
+    "12676": "failure",
+    "12677": "infeasible",
+    "12678": "infeasible",
+    "12679": "failure",
+    "12680": "failure",
+    "12681": "failure",
+    "12682": "failure",
+    "12683": "failure",
+    "12684": "infeasible",
+    "12685": "failure",
+    "12686": "failure",
+    "12687": "successful",
+    "12688": "failure",
+    "12689": "failure",
+    "12690": "failure",
+    "12691": "successful",
+    "12692": "failure",
+    "12693": "infeasible",
+    "12694": "successful",
+    "12695": "successful",
+    "12696": "failure",
+    "12697": "successful",
+    "12698": "failure",
+    "12699": "failure",
+    "12700": "infeasible",
+    "12701": "failure",
+    "12702": "failure",
+    "12703": "failure",
+    "12704": "failure",
+    "12705": "failure",
+    "12706": "failure",
+    "12707": "successful",
+    "12708": "successful",
+    "12709": "successful",
+    "12710": "failure",
+    "12711": "failure",
+    "12712": "failure",
+    "12713": "failure",
+    "12714": "failure",
+    "12715": "failure",
+    "12716": "failure",
+    "12717": "failure",
+    "12718": "failure",
+    "12719": "failure",
+    "12720": "failure",
+    "12721": "failure",
+    "12722": "failure",
+    "12723": "failure",
+    "12724": "failure",
+    "12725": "failure",
+    "12726": "infeasible",
+    "12727": "failure",
+    "12728": "successful",
+    "12729": "failure",
+    "12730": "failure",
+    "12731": "failure",
+    "12732": "failure",
+    "12733": "failure",
+    "12734": "failure",
+    "12735": "failure",
+    "12736": "failure",
+    "12737": "failure",
+    "12738": "successful",
+    "12739": "failure",
+    "12740": "failure",
+    "12741": "failure",
+    "12742": "failure",
+    "12743": "failure",
+    "12744": "failure",
+    "12745": "failure",
+    "12746": "failure",
+    "12747": "failure",
+    "12748": "failure",
+    "12749": "failure",
+    "12750": "failure",
+    "12751": "infeasible",
+    "12752": "failure",
+    "12753": "infeasible",
+    "12754": "failure",
+    "12755": "failure",
+    "12756": "failure",
+    "12757": "failure",
+    "12758": "infeasible",
+    "12759": "failure",
+    "12760": "failure",
+    "12761": "failure",
+    "12762": "failure",
+    "12763": "failure",
+    "12764": "failure",
+    "12765": "infeasible",
+    "12766": "failure",
+    "12767": "failure",
+    "12768": "infeasible",
+    "12769": "failure",
+    "12770": "failure",
+    "12771": "failure",
+    "12772": "failure",
+    "12773": "failure",
+    "12774": "failure",
+    "12775": "failure",
+    "12776": "failure",
+    "12777": "failure",
+    "12778": "failure",
+    "12779": "failure",
+    "12780": "successful",
+    "12781": "successful",
+    "12782": "failure",
+    "12783": "failure",
+    "12784": "failure",
+    "12785": "failure",
+    "12786": "failure",
+    "12787": "failure",
+    "12788": "failure",
+    "12789": "failure",
+    "12790": "failure",
+    "12791": "failure",
+    "12792": "failure",
+    "12793": "failure",
+    "12794": "failure",
+    "12795": "failure",
+    "12796": "failure",
+    "12797": "infeasible",
+    "12798": "failure",
+    "12799": "failure",
+    "12800": "successful",
+    "12801": "failure",
+    "12802": "failure",
+    "12803": "successful",
+    "12804": "failure",
+    "12805": "failure",
+    "12806": "failure",
+    "12807": "failure",
+    "12808": "failure",
+    "12809": "infeasible",
+    "12810": "failure",
+    "12811": "failure",
+    "12812": "failure",
+    "12813": "failure",
+    "12814": "failure",
+    "12815": "failure",
+    "12816": "failure",
+    "12817": "failure",
+    "12818": "failure",
+    "12819": "failure",
+    "12820": "infeasible",
+    "12821": "failure",
+    "12822": "infeasible",
+    "12823": "successful",
+    "12824": "infeasible",
+    "12825": "failure",
+    "12826": "failure",
+    "12827": "failure",
+    "12828": "failure",
+    "12829": "infeasible",
+    "12830": "successful",
+    "12831": "failure",
+    "12832": "failure",
+    "12833": "failure",
+    "12834": "failure",
+    "12835": "failure",
+    "12836": "successful",
+    "12837": "infeasible",
+    "12838": "successful",
+    "12839": "failure",
+    "12840": "successful",
+    "12841": "failure",
+    "12842": "failure",
+    "12843": "failure",
+    "12844": "infeasible",
+    "12845": "failure",
+    "12846": "failure",
+    "12847": "successful",
+    "12848": "failure",
+    "12849": "failure",
+    "12850": "failure",
+    "12851": "failure",
+    "12852": "failure",
+    "12853": "failure",
+    "12854": "failure",
+    "12855": "successful",
+    "12856": "failure",
+    "12857": "failure",
+    "12858": "failure",
+    "12859": "failure",
+    "12860": "failure",
+    "12861": "failure",
+    "12862": "failure",
+    "12863": "failure",
+    "12864": "failure",
+    "12865": "failure",
+    "12866": "failure",
+    "12867": "failure",
+    "12868": "successful",
+    "12869": "failure",
+    "12870": "failure",
+    "12871": "failure",
+    "12872": "failure",
+    "12873": "failure",
+    "12874": "failure",
+    "12875": "failure",
+    "12876": "infeasible",
+    "12877": "infeasible",
+    "12878": "failure",
+    "12879": "failure",
+    "12880": "failure",
+    "12881": "successful",
+    "12882": "failure",
+    "12883": "successful",
+    "12884": "failure",
+    "12885": "failure",
+    "12886": "infeasible",
+    "12887": "infeasible",
+    "12888": "failure",
+    "12889": "failure",
+    "12890": "successful",
+    "12891": "successful",
+    "12892": "failure",
+    "12893": "failure",
+    "12894": "failure",
+    "12895": "infeasible",
+    "12896": "failure",
+    "12897": "infeasible",
+    "12898": "failure",
+    "12899": "failure",
+    "12900": "failure",
+    "12901": "failure",
+    "12902": "already_complete",
+    "12903": "failure",
+    "12904": "successful",
+    "12905": "failure",
+    "12906": "already_complete",
+    "12907": "failure",
+    "12908": "successful",
+    "12909": "failure",
+    "12910": "failure",
+    "12911": "failure",
+    "12912": "failure",
+    "12913": "failure",
+    "12914": "failure",
+    "12915": "infeasible",
+    "12916": "infeasible",
+    "12917": "failure",
+    "12918": "failure",
+    "12919": "infeasible",
+    "12920": "failure",
+    "12921": "failure",
+    "12922": "infeasible",
+    "12923": "failure",
+    "12924": "failure",
+    "12925": "failure",
+    "12926": "infeasible",
+    "12927": "failure",
+    "12928": "failure",
+    "12929": "failure",
+    "12930": "failure",
+    "12931": "failure",
+    "12932": "failure",
+    "12933": "failure",
+    "12934": "failure",
+    "12935": "successful",
+    "12936": "failure",
+    "12937": "failure",
+    "12938": "failure",
+    "12939": "failure",
+    "12940": "failure",
+    "12941": "failure",
+    "12942": "failure",
+    "12943": "successful",
+    "12944": "failure",
+    "12945": "failure",
+    "12946": "failure",
+    "12947": "successful",
+    "12948": "infeasible",
+    "12949": "failure",
+    "12950": "failure",
+    "12951": "failure",
+    "12952": "infeasible",
+    "12953": "failure",
+    "12954": "failure",
+    "12955": "infeasible",
+    "12956": "failure",
+    "12957": "successful",
+    "12958": "failure",
+    "12959": "successful",
+    "12960": "failure",
+    "12961": "failure",
+    "12962": "failure",
+    "12963": "failure",
+    "12964": "failure",
+    "12965": "failure",
+    "12966": "failure",
+    "12967": "successful",
+    "12968": "failure",
+    "12969": "failure",
+    "12970": "failure",
+    "12971": "failure",
+    "12972": "already_complete",
+    "12973": "failure",
+    "12974": "failure",
+    "12975": "failure",
+    "12976": "failure",
+    "12977": "infeasible",
+    "12978": "failure",
+    "12979": "failure",
+    "12980": "failure",
+    "12981": "infeasible",
+    "12982": "successful",
+    "12983": "failure",
+    "12984": "infeasible",
+    "12985": "failure",
+    "12986": "failure",
+    "12987": "failure",
+    "12988": "failure",
+    "12989": "failure",
+    "12990": "failure",
+    "12991": "infeasible",
+    "12992": "failure",
+    "12993": "failure",
+    "12994": "failure",
+    "12995": "failure",
+    "12996": "failure",
+    "12997": "failure",
+    "12998": "successful",
+    "12999": "failure",
+    "13000": "failure",
+    "13001": "failure",
+    "13002": "failure",
+    "13003": "successful",
+    "13004": "failure",
+    "13005": "infeasible",
+    "13006": "failure",
+    "13007": "failure",
+    "13008": "successful",
+    "13009": "failure",
+    "13010": "failure",
+    "13011": "infeasible",
+    "13012": "failure",
+    "13013": "failure",
+    "13014": "failure",
+    "13015": "failure",
+    "13016": "failure",
+    "13017": "failure",
+    "13018": "failure",
+    "13019": "successful",
+    "13020": "failure",
+    "13021": "failure",
+    "13022": "failure",
+    "13023": "failure",
+    "13024": "failure",
+    "13025": "failure",
+    "13026": "failure",
+    "13027": "failure",
+    "13028": "failure",
+    "13029": "failure",
+    "13030": "failure",
+    "13031": "failure",
+    "13032": "failure",
+    "13033": "failure",
+    "13034": "successful",
+    "13035": "failure",
+    "13036": "failure",
+    "13037": "failure",
+    "13038": "failure",
+    "13039": "infeasible",
+    "13040": "failure",
+    "13041": "failure",
+    "13042": "failure",
+    "13043": "failure",
+    "13044": "successful",
+    "13045": "failure",
+    "13046": "failure",
+    "13047": "failure",
+    "13048": "successful",
+    "13049": "failure",
+    "13050": "failure",
+    "13051": "failure",
+    "13052": "failure",
+    "13053": "failure",
+    "13054": "failure",
+    "13055": "failure",
+    "13056": "failure",
+    "13057": "failure",
+    "13058": "failure",
+    "13059": "failure",
+    "13060": "infeasible",
+    "13061": "successful",
+    "13062": "failure",
+    "13063": "failure",
+    "13064": "successful",
+    "13065": "failure",
+    "13066": "failure",
+    "13067": "failure",
+    "13068": "failure",
+    "13069": "failure",
+    "13070": "failure",
+    "13071": "failure",
+    "13072": "failure",
+    "13073": "failure",
+    "13074": "failure",
+    "13075": "successful",
+    "13076": "failure",
+    "13077": "failure",
+    "13078": "infeasible",
+    "13079": "failure",
+    "13080": "failure",
+    "13081": "successful",
+    "13082": "infeasible",
+    "13083": "failure",
+    "13084": "infeasible",
+    "13085": "failure",
+    "13086": "successful",
+    "13087": "failure",
+    "13088": "failure",
+    "13089": "failure",
+    "13090": "failure",
+    "13091": "failure",
+    "13092": "failure",
+    "13093": "failure",
+    "13094": "successful",
+    "13095": "failure",
+    "13096": "infeasible",
+    "13097": "failure",
+    "13098": "failure",
+    "13099": "successful",
+    "13100": "failure",
+    "13101": "failure",
+    "13102": "failure",
+    "13103": "failure",
+    "13104": "failure",
+    "13105": "failure",
+    "13106": "failure",
+    "13107": "failure",
+    "13108": "successful",
+    "13109": "failure",
+    "13110": "successful",
+    "13111": "failure",
+    "13112": "failure",
+    "13113": "failure",
+    "13114": "failure",
+    "13115": "infeasible",
+    "13116": "infeasible",
+    "13117": "failure",
+    "13118": "failure",
+    "13119": "failure",
+    "13120": "failure",
+    "13121": "failure",
+    "13122": "failure",
+    "13123": "successful",
+    "13124": "failure",
+    "13125": "failure",
+    "13126": "failure",
+    "13127": "successful",
+    "13128": "failure",
+    "13129": "failure",
+    "13130": "failure",
+    "13131": "failure",
+    "13132": "infeasible",
+    "13133": "successful",
+    "13134": "failure",
+    "13135": "failure",
+    "13136": "failure",
+    "13137": "failure",
+    "13138": "failure",
+    "13139": "failure",
+    "13140": "failure",
+    "13141": "failure",
+    "13142": "failure",
+    "13143": "failure",
+    "13144": "failure",
+    "13145": "failure",
+    "13146": "infeasible",
+    "13147": "successful",
+    "13148": "failure",
+    "13149": "failure",
+    "13150": "failure",
+    "13151": "failure",
+    "13152": "failure",
+    "13153": "failure",
+    "13154": "successful",
+    "13155": "successful",
+    "13156": "failure",
+    "13157": "failure",
+    "13158": "infeasible",
+    "13159": "failure",
+    "13160": "failure",
+    "13161": "failure",
+    "13162": "failure",
+    "13163": "failure",
+    "13164": "successful",
+    "13165": "failure",
+    "13166": "failure",
+    "13167": "failure",
+    "13168": "failure",
+    "13169": "infeasible",
+    "13170": "infeasible",
+    "13171": "failure",
+    "13172": "successful",
+    "13173": "failure",
+    "13174": "failure",
+    "13175": "successful",
+    "13176": "successful",
+    "13177": "failure",
+    "13178": "failure",
+    "13179": "failure",
+    "13180": "failure",
+    "13181": "failure",
+    "13182": "failure",
+    "13183": "failure",
+    "13184": "failure",
+    "13185": "failure",
+    "13186": "failure",
+    "13187": "failure",
+    "13188": "failure",
+    "13189": "infeasible",
+    "13190": "failure",
+    "13191": "failure",
+    "13192": "successful",
+    "13193": "infeasible",
+    "13194": "successful",
+    "13195": "failure",
+    "13196": "failure",
+    "13197": "successful",
+    "13198": "failure",
+    "13199": "failure",
+    "13200": "failure",
+    "13201": "failure",
+    "13202": "successful",
+    "13203": "failure",
+    "13204": "successful",
+    "13205": "failure",
+    "13206": "failure",
+    "13207": "infeasible",
+    "13208": "failure",
+    "13209": "failure",
+    "13210": "failure",
+    "13211": "successful",
+    "13212": "failure",
+    "13213": "successful",
+    "13214": "failure",
+    "13215": "failure",
+    "13216": "failure",
+    "13217": "failure",
+    "13218": "failure",
+    "13219": "failure",
+    "13220": "failure",
+    "13221": "successful",
+    "13222": "failure",
+    "13223": "infeasible",
+    "13224": "failure",
+    "13225": "failure",
+    "13226": "infeasible",
+    "13227": "successful",
+    "13228": "failure",
+    "13229": "failure",
+    "13230": "failure",
+    "13231": "failure",
+    "13232": "failure",
+    "13233": "successful",
+    "13234": "failure",
+    "13235": "successful",
+    "13236": "failure",
+    "13237": "failure",
+    "13238": "failure",
+    "13239": "successful",
+    "13240": "failure",
+    "13241": "failure",
+    "13242": "failure",
+    "13243": "failure",
+    "13244": "infeasible",
+    "13245": "failure",
+    "13246": "failure",
+    "13247": "successful",
+    "13248": "already_complete",
+    "13249": "failure",
+    "13250": "failure",
+    "13251": "failure",
+    "13252": "failure",
+    "13253": "failure",
+    "13254": "infeasible",
+    "13255": "failure",
+    "13256": "failure",
+    "13257": "failure",
+    "13258": "infeasible",
+    "13259": "failure",
+    "13260": "failure",
+    "13261": "failure",
+    "13262": "failure",
+    "13263": "failure",
+    "13264": "failure",
+    "13265": "failure",
+    "13266": "successful",
+    "13267": "failure",
+    "13268": "failure",
+    "13269": "failure",
+    "13270": "failure",
+    "13271": "successful",
+    "13272": "failure",
+    "13273": "successful",
+    "13274": "failure",
+    "13275": "failure",
+    "13276": "failure",
+    "13277": "failure",
+    "13278": "failure",
+    "13279": "failure",
+    "13280": "failure",
+    "13281": "failure",
+    "13282": "failure",
+    "13283": "failure",
+    "13284": "failure",
+    "13285": "failure",
+    "13286": "failure",
+    "13287": "failure",
+    "13288": "failure",
+    "13289": "successful",
+    "13290": "failure",
+    "13291": "failure",
+    "13292": "failure",
+    "13293": "failure",
+    "13294": "failure",
+    "13295": "failure",
+    "13296": "failure",
+    "13297": "failure",
+    "13298": "infeasible",
+    "13299": "failure",
+    "13300": "failure",
+    "13301": "failure",
+    "13302": "infeasible",
+    "13303": "failure",
+    "13304": "failure",
+    "13305": "failure",
+    "13306": "failure",
+    "13307": "failure",
+    "13308": "failure",
+    "13309": "failure",
+    "13310": "failure",
+    "13311": "failure",
+    "13312": "failure",
+    "13313": "infeasible",
+    "13314": "failure",
+    "13315": "failure",
+    "13316": "failure",
+    "13317": "infeasible",
+    "13318": "failure",
+    "13319": "infeasible",
+    "13320": "infeasible",
+    "13321": "failure",
+    "13322": "failure",
+    "13323": "failure",
+    "13324": "failure",
+    "13325": "failure",
+    "13326": "failure",
+    "13327": "failure",
+    "13328": "failure",
+    "13329": "failure",
+    "13330": "failure",
+    "13331": "failure",
+    "13332": "failure",
+    "13333": "failure",
+    "13334": "failure",
+    "13335": "successful",
+    "13336": "failure",
+    "13337": "failure",
+    "13338": "successful",
+    "13339": "failure",
+    "13340": "failure",
+    "13341": "failure",
+    "13342": "failure",
+    "13343": "successful",
+    "13344": "failure",
+    "13345": "failure",
+    "13346": "successful",
+    "13347": "failure",
+    "13348": "failure",
+    "13349": "failure",
+    "13350": "failure",
+    "13351": "failure",
+    "13352": "infeasible",
+    "13353": "failure",
+    "13354": "failure",
+    "13355": "failure",
+    "13356": "failure",
+    "13357": "failure",
+    "13358": "infeasible",
+    "13359": "failure",
+    "13360": "failure",
+    "13361": "failure",
+    "13362": "failure",
+    "13363": "failure",
+    "13364": "failure",
+    "13365": "failure",
+    "13366": "failure",
+    "13367": "failure",
+    "13368": "failure",
+    "13369": "failure",
+    "13370": "failure",
+    "13371": "failure",
+    "13372": "successful",
+    "13373": "failure",
+    "13374": "infeasible",
+    "13375": "failure",
+    "13376": "failure",
+    "13377": "failure",
+    "13378": "failure",
+    "13379": "failure",
+    "13380": "failure",
+    "13381": "failure",
+    "13382": "failure",
+    "13383": "failure",
+    "13384": "failure",
+    "13385": "failure",
+    "13386": "failure",
+    "13387": "failure",
+    "13388": "failure",
+    "13389": "failure",
+    "13390": "failure",
+    "13391": "failure",
+    "13392": "failure",
+    "13393": "failure",
+    "13394": "failure",
+    "13395": "failure",
+    "13396": "failure",
+    "13397": "failure",
+    "13398": "failure",
+    "13399": "failure",
+    "13400": "failure",
+    "13401": "failure",
+    "13402": "failure",
+    "13403": "failure",
+    "13404": "failure",
+    "13405": "failure",
+    "13406": "failure",
+    "13407": "failure",
+    "13408": "failure",
+    "13409": "failure",
+    "13410": "failure",
+    "13411": "failure",
+    "13412": "infeasible",
+    "13413": "failure",
+    "13414": "failure",
+    "13415": "failure",
+    "13416": "successful",
+    "13417": "failure",
+    "13418": "failure",
+    "13419": "failure",
+    "13420": "failure",
+    "13421": "failure",
+    "13422": "failure",
+    "13423": "failure",
+    "13424": "failure",
+    "13425": "failure",
+    "13426": "failure",
+    "13427": "failure",
+    "13428": "failure",
+    "13429": "failure",
+    "13430": "failure",
+    "13431": "failure",
+    "13432": "failure",
+    "13433": "failure",
+    "13434": "failure",
+    "13435": "failure",
+    "13436": "failure",
+    "13437": "successful",
+    "13438": "successful",
+    "13439": "failure",
+    "13440": "infeasible",
+    "13441": "failure",
+    "13442": "failure",
+    "13443": "failure",
+    "13444": "failure",
+    "13445": "failure",
+    "13446": "failure",
+    "13447": "failure",
+    "13448": "failure",
+    "13449": "failure",
+    "13450": "failure",
+    "13451": "failure",
+    "13452": "failure",
+    "13453": "failure",
+    "13454": "failure",
+    "13455": "infeasible",
+    "13456": "successful",
+    "13457": "successful",
+    "13458": "infeasible",
+    "13459": "failure",
+    "13460": "failure",
+    "13461": "failure",
+    "13462": "failure",
+    "13463": "failure",
+    "13464": "failure",
+    "13465": "failure",
+    "13466": "failure",
+    "13467": "successful",
+    "13468": "successful",
+    "13469": "successful",
+    "13470": "failure",
+    "13471": "failure",
+    "13472": "failure",
+    "13473": "failure",
+    "13474": "failure",
+    "13475": "failure",
+    "13476": "failure",
+    "13477": "failure",
+    "13478": "failure",
+    "13479": "failure",
+    "13480": "failure",
+    "13481": "failure",
+    "13482": "failure",
+    "13483": "infeasible",
+    "13484": "infeasible",
+    "13485": "failure",
+    "13486": "failure",
+    "13487": "failure",
+    "13488": "successful",
+    "13489": "failure",
+    "13490": "failure",
+    "13491": "failure",
+    "13492": "failure",
+    "13493": "failure",
+    "13494": "failure",
+    "13495": "failure",
+    "13496": "failure",
+    "13497": "infeasible",
+    "13498": "failure",
+    "13499": "failure",
+    "13500": "failure",
+    "13501": "successful",
+    "13502": "failure",
+    "13503": "failure",
+    "13504": "failure",
+    "13505": "failure",
+    "13506": "successful",
+    "13507": "failure",
+    "13508": "failure",
+    "13509": "successful",
+    "13510": "infeasible",
+    "13511": "failure",
+    "13512": "failure",
+    "13513": "failure",
+    "13514": "failure",
+    "13515": "failure",
+    "13516": "failure",
+    "13517": "failure",
+    "13518": "failure",
+    "13519": "infeasible",
+    "13520": "failure",
+    "13521": "failure",
+    "13522": "failure",
+    "13523": "failure",
+    "13524": "infeasible",
+    "13525": "failure",
+    "13526": "infeasible",
+    "13527": "infeasible",
+    "13528": "failure",
+    "13529": "failure",
+    "13530": "failure",
+    "13531": "failure",
+    "13532": "failure",
+    "13533": "failure",
+    "13534": "failure",
+    "13535": "failure",
+    "13536": "failure",
+    "13537": "failure",
+    "13538": "failure",
+    "13539": "failure",
+    "13540": "failure",
+    "13541": "successful",
+    "13542": "infeasible",
+    "13543": "failure",
+    "13544": "failure",
+    "13545": "failure",
+    "13546": "failure",
+    "13547": "failure",
+    "13548": "failure",
+    "13549": "failure",
+    "13550": "failure",
+    "13551": "failure",
+    "13552": "infeasible",
+    "13553": "failure",
+    "13554": "failure",
+    "13555": "already_complete",
+    "13556": "failure",
+    "13557": "failure",
+    "13558": "successful",
+    "13559": "failure",
+    "13560": "infeasible",
+    "13561": "failure",
+    "13562": "successful",
+    "13563": "failure",
+    "13564": "failure",
+    "13565": "failure",
+    "13566": "failure",
+    "13567": "failure",
+    "13568": "failure",
+    "13569": "failure",
+    "13570": "failure",
+    "13571": "failure",
+    "13572": "failure",
+    "13573": "infeasible",
+    "13574": "infeasible",
+    "13575": "failure",
+    "13576": "failure",
+    "13577": "failure",
+    "13578": "failure",
+    "13579": "failure",
+    "13580": "failure",
+    "13581": "successful",
+    "13582": "failure",
+    "13583": "failure",
+    "13584": "failure",
+    "13585": "infeasible",
+    "13586": "failure",
+    "13587": "successful",
+    "13588": "failure",
+    "13589": "failure",
+    "13590": "failure",
+    "13591": "failure",
+    "13592": "infeasible",
+    "13593": "failure",
+    "13594": "failure",
+    "13595": "infeasible",
+    "13596": "failure",
+    "13597": "failure",
+    "13598": "failure",
+    "13599": "failure",
+    "13600": "failure",
+    "13601": "infeasible",
+    "13602": "failure",
+    "13603": "failure",
+    "13604": "failure",
+    "13605": "failure",
+    "13606": "failure",
+    "13607": "infeasible",
+    "13608": "failure",
+    "13609": "failure",
+    "13610": "failure",
+    "13611": "failure",
+    "13612": "failure",
+    "13613": "failure",
+    "13614": "failure",
+    "13615": "failure",
+    "13616": "failure",
+    "13617": "failure",
+    "13618": "failure",
+    "13619": "failure",
+    "13620": "successful",
+    "13621": "infeasible",
+    "13622": "already_complete",
+    "13623": "failure",
+    "13624": "infeasible",
+    "13625": "failure",
+    "13626": "failure",
+    "13627": "failure",
+    "13628": "failure",
+    "13629": "failure",
+    "13630": "failure",
+    "13631": "failure",
+    "13632": "failure",
+    "13633": "infeasible",
+    "13634": "failure",
+    "13635": "failure",
+    "13636": "failure",
+    "13637": "successful",
+    "13638": "failure",
+    "13639": "failure",
+    "13640": "successful",
+    "13641": "failure",
+    "13642": "failure",
+    "13643": "successful",
+    "13644": "failure",
+    "13645": "failure",
+    "13646": "failure",
+    "13647": "failure",
+    "13648": "failure",
+    "13649": "failure",
+    "13650": "infeasible",
+    "13651": "failure",
+    "13652": "failure",
+    "13653": "failure",
+    "13654": "failure",
+    "13655": "successful",
+    "13656": "failure",
+    "13657": "failure",
+    "13658": "infeasible",
+    "13659": "failure",
+    "13660": "infeasible",
+    "13661": "failure",
+    "13662": "failure",
+    "13663": "successful",
+    "13664": "failure",
+    "13665": "failure",
+    "13666": "failure",
+    "13667": "failure",
+    "13668": "failure",
+    "13669": "failure",
+    "13670": "failure",
+    "13671": "failure",
+    "13672": "failure",
+    "13673": "failure",
+    "13674": "failure",
+    "13675": "failure",
+    "13676": "failure",
+    "13677": "failure",
+    "13678": "failure",
+    "13679": "failure",
+    "13680": "failure",
+    "13681": "failure",
+    "13682": "successful",
+    "13683": "failure",
+    "13684": "failure",
+    "13685": "failure",
+    "13686": "infeasible",
+    "13687": "failure",
+    "13688": "successful",
+    "13689": "failure",
+    "13690": "failure",
+    "13691": "failure",
+    "13692": "failure",
+    "13693": "failure",
+    "13694": "failure",
+    "13695": "failure",
+    "13696": "failure",
+    "13697": "successful",
+    "13698": "failure",
+    "13699": "failure",
+    "13700": "failure",
+    "13701": "failure",
+    "13702": "failure",
+    "13703": "failure",
+    "13704": "failure",
+    "13705": "failure",
+    "13706": "failure",
+    "13707": "failure",
+    "13708": "failure",
+    "13709": "failure",
+    "13710": "failure",
+    "13711": "failure",
+    "13712": "failure",
+    "13713": "failure",
+    "13714": "failure",
+    "13715": "failure",
+    "13716": "infeasible",
+    "13717": "failure",
+    "13718": "failure",
+    "13719": "failure",
+    "13720": "successful",
+    "13721": "successful",
+    "13722": "failure",
+    "13723": "successful",
+    "13724": "failure",
+    "13725": "failure",
+    "13726": "failure",
+    "13727": "failure",
+    "13728": "failure",
+    "13729": "failure",
+    "13730": "infeasible",
+    "13731": "failure",
+    "13732": "failure",
+    "13733": "failure",
+    "13734": "failure",
+    "13735": "failure",
+    "13736": "failure",
+    "13737": "failure",
+    "13738": "failure",
+    "13739": "failure",
+    "13740": "infeasible",
+    "13741": "failure",
+    "13742": "failure",
+    "13743": "failure",
+    "13744": "failure",
+    "13745": "failure",
+    "13746": "failure",
+    "13747": "failure",
+    "13748": "successful",
+    "13749": "failure",
+    "13750": "failure",
+    "13751": "failure",
+    "13752": "failure",
+    "13753": "failure",
+    "13754": "failure",
+    "13755": "failure",
+    "13756": "failure",
+    "13757": "failure",
+    "13758": "infeasible",
+    "13759": "failure",
+    "13760": "failure",
+    "13761": "failure",
+    "13762": "failure",
+    "13763": "failure",
+    "13764": "failure",
+    "13765": "failure",
+    "13766": "failure",
+    "13767": "failure",
+    "13768": "failure",
+    "13769": "failure",
+    "13770": "failure",
+    "13771": "failure",
+    "13772": "successful",
+    "13773": "failure",
+    "13774": "failure",
+    "13775": "failure",
+    "13776": "infeasible",
+    "13777": "successful",
+    "13778": "failure",
+    "13779": "failure",
+    "13780": "failure",
+    "13781": "failure",
+    "13782": "failure",
+    "13783": "failure",
+    "13784": "failure",
+    "13785": "failure",
+    "13786": "infeasible",
+    "13787": "failure",
+    "13788": "successful",
+    "13789": "failure",
+    "13790": "failure",
+    "13791": "failure",
+    "13792": "failure",
+    "13793": "failure",
+    "13794": "failure",
+    "13795": "successful",
+    "13796": "failure",
+    "13797": "successful",
+    "13798": "failure",
+    "13799": "failure",
+    "13800": "failure",
+    "13801": "successful",
+    "13802": "failure",
+    "13803": "failure",
+    "13804": "failure",
+    "13805": "failure",
+    "13806": "failure",
+    "13807": "successful",
+    "13808": "failure",
+    "13809": "failure",
+    "13810": "successful",
+    "13811": "failure",
+    "13812": "infeasible",
+    "13813": "infeasible",
+    "13814": "failure",
+    "13815": "failure",
+    "13816": "failure",
+    "13817": "failure",
+    "13818": "failure",
+    "13819": "successful",
+    "13820": "failure",
+    "13821": "failure",
+    "13822": "failure",
+    "13823": "failure",
+    "13824": "successful",
+    "13825": "failure",
+    "13826": "failure",
+    "13827": "infeasible",
+    "13828": "failure",
+    "13829": "infeasible",
+    "13830": "failure",
+    "13831": "failure",
+    "13832": "failure",
+    "13833": "failure",
+    "13834": "failure",
+    "13835": "failure",
+    "13836": "infeasible",
+    "13837": "failure",
+    "13838": "failure",
+    "13839": "failure",
+    "13840": "failure",
+    "13841": "failure",
+    "13842": "failure",
+    "13843": "successful",
+    "13844": "failure",
+    "13845": "failure",
+    "13846": "failure",
+    "13847": "failure",
+    "13848": "failure",
+    "13849": "failure",
+    "13850": "failure",
+    "13851": "infeasible",
+    "13852": "failure",
+    "13853": "failure",
+    "13854": "infeasible",
+    "13855": "failure",
+    "13856": "successful",
+    "13857": "failure",
+    "13858": "failure",
+    "13859": "failure",
+    "13860": "failure",
+    "13861": "failure",
+    "13862": "successful",
+    "13863": "failure",
+    "13864": "failure",
+    "13865": "failure",
+    "13866": "failure",
+    "13867": "failure",
+    "13868": "failure",
+    "13869": "failure",
+    "13870": "failure",
+    "13871": "failure",
+    "13872": "successful",
+    "13873": "failure",
+    "13874": "failure",
+    "13875": "failure",
+    "13876": "successful",
+    "13877": "failure",
+    "13878": "failure",
+    "13879": "failure",
+    "13880": "failure",
+    "13881": "successful",
+    "13882": "successful",
+    "13883": "failure",
+    "13884": "failure",
+    "13885": "failure",
+    "13886": "successful",
+    "13887": "failure",
+    "13888": "failure",
+    "13889": "failure",
+    "13890": "failure",
+    "13891": "failure",
+    "13892": "infeasible",
+    "13893": "failure",
+    "13894": "failure",
+    "13895": "failure",
+    "13896": "failure",
+    "13897": "failure",
+    "13898": "failure",
+    "13899": "failure",
+    "13900": "failure",
+    "13901": "failure",
+    "13902": "successful",
+    "13903": "infeasible",
+    "13904": "infeasible",
+    "13905": "infeasible",
+    "13906": "failure",
+    "13907": "successful",
+    "13908": "failure",
+    "13909": "failure",
+    "13910": "failure",
+    "13911": "failure",
+    "13912": "successful",
+    "13913": "infeasible",
+    "13914": "failure",
+    "13915": "successful",
+    "13916": "infeasible",
+    "13917": "failure",
+    "13918": "infeasible",
+    "13919": "failure",
+    "13920": "failure",
+    "13921": "failure",
+    "13922": "failure",
+    "13923": "failure",
+    "13924": "failure",
+    "13925": "failure",
+    "13926": "failure",
+    "13927": "failure",
+    "13928": "failure",
+    "13929": "failure",
+    "13930": "failure",
+    "13931": "failure",
+    "13932": "failure",
+    "13933": "failure",
+    "13934": "infeasible",
+    "13935": "failure",
+    "13936": "failure",
+    "13937": "failure",
+    "13938": "failure",
+    "13939": "failure",
+    "13940": "failure",
+    "13941": "failure",
+    "13942": "successful",
+    "13943": "failure",
+    "13944": "failure",
+    "13945": "failure",
+    "13946": "failure",
+    "13947": "failure",
+    "13948": "failure",
+    "13949": "failure",
+    "13950": "failure",
+    "13951": "failure",
+    "13952": "infeasible",
+    "13953": "failure",
+    "13954": "failure",
+    "13955": "successful",
+    "13956": "failure",
+    "13957": "successful",
+    "13958": "failure",
+    "13959": "failure",
+    "13960": "failure",
+    "13961": "failure",
+    "13962": "failure",
+    "13963": "failure",
+    "13964": "failure",
+    "13965": "failure",
+    "13966": "failure",
+    "13967": "failure",
+    "13968": "successful",
+    "13969": "failure",
+    "13970": "failure",
+    "13971": "failure",
+    "13972": "failure",
+    "13973": "failure",
+    "13974": "infeasible",
+    "13975": "failure",
+    "13976": "failure",
+    "13977": "failure",
+    "13978": "failure",
+    "13979": "failure",
+    "13980": "failure",
+    "13981": "failure",
+    "13982": "failure",
+    "13983": "failure",
+    "13984": "failure",
+    "13985": "failure",
+    "13986": "failure",
+    "13987": "successful",
+    "13988": "failure",
+    "13989": "failure",
+    "13990": "failure",
+    "13991": "failure",
+    "13992": "failure",
+    "13993": "successful",
+    "13994": "failure",
+    "13995": "successful",
+    "13996": "successful",
+    "13997": "failure",
+    "13998": "failure",
+    "13999": "failure",
+    "14000": "failure",
+    "14001": "failure",
+    "14002": "successful",
+    "14003": "failure",
+    "14004": "failure",
+    "14005": "failure",
+    "14006": "failure",
+    "14007": "successful",
+    "14008": "already_complete",
+    "14009": "failure",
+    "14010": "failure",
+    "14011": "infeasible",
+    "14012": "infeasible",
+    "14013": "successful",
+    "14014": "failure",
+    "14015": "failure",
+    "14016": "infeasible",
+    "14017": "infeasible",
+    "14018": "failure",
+    "14019": "failure",
+    "14020": "failure",
+    "14021": "failure",
+    "14022": "failure",
+    "14023": "infeasible",
+    "14024": "failure",
+    "14025": "failure",
+    "14026": "failure",
+    "14027": "failure",
+    "14028": "failure",
+    "14029": "failure",
+    "14030": "failure",
+    "14031": "failure",
+    "14032": "failure",
+    "14033": "infeasible",
+    "14034": "failure",
+    "14035": "failure",
+    "14036": "failure",
+    "14037": "failure",
+    "14038": "failure",
+    "14039": "failure",
+    "14040": "failure",
+    "14041": "failure",
+    "14042": "failure",
+    "14043": "failure",
+    "14044": "failure",
+    "14045": "infeasible",
+    "14046": "failure",
+    "14047": "failure",
+    "14048": "successful",
+    "14049": "failure",
+    "14050": "infeasible",
+    "14051": "failure",
+    "14052": "infeasible",
+    "14053": "failure",
+    "14054": "failure",
+    "14055": "failure",
+    "14056": "failure",
+    "14057": "failure",
+    "14058": "infeasible",
+    "14059": "failure",
+    "14060": "failure",
+    "14061": "failure",
+    "14062": "failure",
+    "14063": "infeasible",
+    "14064": "failure",
+    "14065": "infeasible",
+    "14066": "successful",
+    "14067": "failure",
+    "14068": "failure",
+    "14069": "successful",
+    "14070": "failure",
+    "14071": "successful",
+    "14072": "successful",
+    "14073": "failure",
+    "14074": "failure",
+    "14075": "failure",
+    "14076": "failure",
+    "14077": "failure",
+    "14078": "successful",
+    "14079": "failure",
+    "14080": "failure",
+    "14081": "successful",
+    "14082": "successful",
+    "14083": "failure",
+    "14084": "failure",
+    "14085": "failure",
+    "14086": "failure",
+    "14087": "failure",
+    "14088": "failure",
+    "14089": "failure",
+    "14090": "failure",
+    "14091": "infeasible",
+    "14092": "failure",
+    "14093": "infeasible",
+    "14094": "failure",
+    "14095": "failure",
+    "14096": "infeasible",
+    "14097": "failure",
+    "14098": "failure",
+    "14099": "failure",
+    "14100": "failure",
+    "14101": "failure",
+    "14102": "failure",
+    "14103": "failure",
+    "14104": "failure",
+    "14105": "failure",
+    "14106": "failure",
+    "14107": "failure",
+    "14108": "failure",
+    "14109": "failure",
+    "14110": "failure",
+    "14111": "failure",
+    "14112": "failure",
+    "14113": "failure",
+    "14114": "failure",
+    "14115": "failure",
+    "14116": "successful",
+    "14117": "infeasible",
+    "14118": "failure",
+    "14119": "failure",
+    "14120": "failure",
+    "14121": "successful",
+    "14122": "failure",
+    "14123": "failure",
+    "14124": "failure",
+    "14125": "failure",
+    "14126": "infeasible",
+    "14127": "infeasible",
+    "14128": "failure",
+    "14129": "failure",
+    "14130": "failure",
+    "14131": "failure",
+    "14132": "failure",
+    "14133": "failure",
+    "14134": "infeasible",
+    "14135": "failure",
+    "14136": "failure",
+    "14137": "failure",
+    "14138": "failure",
+    "14139": "failure",
+    "14140": "infeasible",
+    "14141": "failure",
+    "14142": "successful",
+    "14143": "infeasible",
+    "14144": "failure",
+    "14145": "infeasible",
+    "14146": "failure",
+    "14147": "failure",
+    "14148": "failure",
+    "14149": "failure",
+    "14150": "failure",
+    "14151": "failure",
+    "14152": "failure",
+    "14153": "failure",
+    "14154": "failure",
+    "14155": "failure",
+    "14156": "successful",
+    "14157": "failure",
+    "14158": "failure",
+    "14159": "failure",
+    "14160": "failure",
+    "14161": "failure",
+    "14162": "failure",
+    "14163": "failure",
+    "14164": "failure",
+    "14165": "failure",
+    "14166": "infeasible",
+    "14167": "failure",
+    "14168": "failure",
+    "14169": "failure",
+    "14170": "failure",
+    "14171": "failure",
+    "14172": "failure",
+    "14173": "failure",
+    "14174": "failure",
+    "14175": "failure",
+    "14176": "failure",
+    "14177": "failure",
+    "14178": "failure",
+    "14179": "failure",
+    "14180": "failure",
+    "14181": "infeasible",
+    "14182": "infeasible",
+    "14183": "successful",
+    "14184": "failure",
+    "14185": "failure",
+    "14186": "failure",
+    "14187": "failure",
+    "14188": "failure",
+    "14189": "infeasible",
+    "14190": "failure",
+    "14191": "failure",
+    "14192": "failure",
+    "14193": "failure",
+    "14194": "failure",
+    "14195": "failure",
+    "14196": "failure",
+    "14197": "failure",
+    "14198": "failure",
+    "14199": "failure",
+    "14200": "failure",
+    "14201": "failure",
+    "14202": "failure",
+    "14203": "successful",
+    "14204": "successful",
+    "14205": "failure",
+    "14206": "infeasible",
+    "14207": "failure",
+    "14208": "failure",
+    "14209": "failure",
+    "14210": "failure",
+    "14211": "successful",
+    "14212": "failure",
+    "14213": "failure",
+    "14214": "failure",
+    "14215": "failure",
+    "14216": "successful",
+    "14217": "failure",
+    "14218": "failure",
+    "14219": "failure",
+    "14220": "infeasible",
+    "14221": "failure",
+    "14222": "successful",
+    "14223": "failure",
+    "14224": "infeasible",
+    "14225": "infeasible",
+    "14226": "failure",
+    "14227": "failure",
+    "14228": "failure",
+    "14229": "failure",
+    "14230": "failure",
+    "14231": "infeasible",
+    "14232": "failure",
+    "14233": "failure",
+    "14234": "failure",
+    "14235": "failure",
+    "14236": "failure",
+    "14237": "failure",
+    "14238": "failure",
+    "14239": "infeasible",
+    "14240": "failure",
+    "14241": "failure",
+    "14242": "failure",
+    "14243": "failure",
+    "14244": "failure",
+    "14245": "successful",
+    "14246": "failure",
+    "14247": "failure",
+    "14248": "infeasible",
+    "14249": "failure",
+    "14250": "failure",
+    "14251": "failure",
+    "14252": "failure",
+    "14253": "failure",
+    "14254": "failure",
+    "14255": "failure",
+    "14256": "failure",
+    "14257": "failure",
+    "14258": "failure",
+    "14259": "failure",
+    "14260": "failure",
+    "14261": "failure",
+    "14262": "failure",
+    "14263": "failure",
+    "14264": "failure",
+    "14265": "failure",
+    "14266": "failure",
+    "14267": "successful",
+    "14268": "failure",
+    "14269": "failure",
+    "14270": "failure",
+    "14271": "failure",
+    "14272": "successful",
+    "14273": "failure",
+    "14274": "failure",
+    "14275": "failure",
+    "14276": "failure",
+    "14277": "infeasible",
+    "14278": "failure",
+    "14279": "failure",
+    "14280": "successful",
+    "14281": "successful",
+    "14282": "failure",
+    "14283": "failure",
+    "14284": "successful",
+    "14285": "failure",
+    "14286": "failure",
+    "14287": "successful",
+    "14288": "failure",
+    "14289": "failure",
+    "14290": "failure",
+    "14291": "failure",
+    "14292": "failure",
+    "14293": "failure",
+    "14294": "successful",
+    "14295": "failure",
+    "14296": "failure",
+    "14297": "infeasible",
+    "14298": "failure",
+    "14299": "infeasible",
+    "14300": "failure",
+    "14301": "failure",
+    "14302": "infeasible",
+    "14303": "failure",
+    "14304": "failure",
+    "14305": "failure",
+    "14306": "infeasible",
+    "14307": "failure",
+    "14308": "infeasible",
+    "14309": "failure",
+    "14310": "failure",
+    "14311": "failure",
+    "14312": "infeasible",
+    "14313": "failure",
+    "14314": "failure",
+    "14315": "successful",
+    "14316": "failure",
+    "14317": "failure",
+    "14318": "failure",
+    "14319": "failure",
+    "14320": "failure",
+    "14321": "failure",
+    "14322": "failure",
+    "14323": "failure",
+    "14324": "infeasible",
+    "14325": "failure",
+    "14326": "failure",
+    "14327": "failure",
+    "14328": "failure",
+    "14329": "failure",
+    "14330": "successful",
+    "14331": "failure",
+    "14332": "failure",
+    "14333": "infeasible",
+    "14334": "failure",
+    "14335": "infeasible",
+    "14336": "failure",
+    "14337": "failure",
+    "14338": "failure",
+    "14339": "infeasible",
+    "14340": "failure",
+    "14341": "failure",
+    "14342": "failure",
+    "14343": "failure",
+    "14344": "successful",
+    "14345": "infeasible",
+    "14346": "failure",
+    "14347": "failure",
+    "14348": "successful",
+    "14349": "infeasible",
+    "14350": "failure",
+    "14351": "failure",
+    "14352": "failure",
+    "14353": "failure",
+    "14354": "failure",
+    "14355": "infeasible",
+    "14356": "failure",
+    "14357": "failure",
+    "14358": "failure",
+    "14359": "failure",
+    "14360": "failure",
+    "14361": "successful",
+    "14362": "failure",
+    "14363": "failure",
+    "14364": "failure",
+    "14365": "failure",
+    "14366": "failure",
+    "14367": "successful",
+    "14368": "successful",
+    "14369": "failure",
+    "14370": "successful",
+    "14371": "failure",
+    "14372": "failure",
+    "14373": "failure",
+    "14374": "failure",
+    "14375": "failure",
+    "14376": "successful",
+    "14377": "failure",
+    "14378": "failure",
+    "14379": "failure",
+    "14380": "failure",
+    "14381": "failure",
+    "14382": "failure",
+    "14383": "failure",
+    "14384": "successful",
+    "14385": "failure",
+    "14386": "failure",
+    "14387": "failure",
+    "14388": "failure",
+    "14389": "failure",
+    "14390": "failure",
+    "14391": "failure",
+    "14392": "failure",
+    "14393": "failure",
+    "14394": "failure",
+    "14395": "failure",
+    "14396": "failure",
+    "14397": "failure",
+    "14398": "failure",
+    "14399": "failure",
+    "14400": "failure",
+    "14401": "failure",
+    "14402": "failure",
+    "14403": "failure",
+    "14404": "infeasible",
+    "14405": "failure",
+    "14406": "failure",
+    "14407": "failure",
+    "14408": "failure",
+    "14409": "failure",
+    "14410": "successful",
+    "14411": "failure",
+    "14412": "failure",
+    "14413": "failure",
+    "14414": "failure",
+    "14415": "failure",
+    "14416": "successful",
+    "14417": "failure",
+    "14418": "infeasible",
+    "14419": "failure",
+    "14420": "failure",
+    "14421": "failure",
+    "14422": "failure",
+    "14423": "failure",
+    "14424": "infeasible",
+    "14425": "failure",
+    "14426": "infeasible",
+    "14427": "failure",
+    "14428": "failure",
+    "14429": "infeasible",
+    "14430": "failure",
+    "14431": "failure",
+    "14432": "failure",
+    "14433": "failure",
+    "14434": "failure",
+    "14435": "failure",
+    "14436": "infeasible",
+    "14437": "failure",
+    "14438": "failure",
+    "14439": "failure",
+    "14440": "infeasible",
+    "14441": "successful",
+    "14442": "successful",
+    "14443": "failure",
+    "14444": "failure",
+    "14445": "failure",
+    "14446": "infeasible",
+    "14447": "failure",
+    "14448": "infeasible",
+    "14449": "successful",
+    "14450": "failure",
+    "14451": "failure",
+    "14452": "failure",
+    "14453": "successful",
+    "14454": "failure",
+    "14455": "failure",
+    "14456": "infeasible",
+    "14457": "failure",
+    "14458": "infeasible",
+    "14459": "infeasible",
+    "14460": "failure",
+    "14461": "successful",
+    "14462": "failure",
+    "14463": "failure",
+    "14464": "infeasible",
+    "14465": "failure",
+    "14466": "failure",
+    "14467": "infeasible",
+    "14468": "failure",
+    "14469": "failure",
+    "14470": "successful",
+    "14471": "successful",
+    "14472": "failure",
+    "14473": "successful",
+    "14474": "failure",
+    "14475": "successful",
+    "14476": "failure",
+    "14477": "failure",
+    "14478": "failure",
+    "14479": "failure",
+    "14480": "failure",
+    "14481": "failure",
+    "14482": "successful",
+    "14483": "failure",
+    "14484": "failure",
+    "14485": "infeasible",
+    "14486": "failure",
+    "14487": "failure",
+    "14488": "failure",
+    "14489": "failure",
+    "14490": "failure",
+    "14491": "failure",
+    "14492": "failure",
+    "14493": "failure",
+    "14494": "failure",
+    "14495": "failure",
+    "14496": "failure",
+    "14497": "failure",
+    "14498": "failure",
+    "14499": "failure",
+    "14500": "failure",
+    "14501": "failure",
+    "14502": "successful",
+    "14503": "infeasible",
+    "14504": "failure",
+    "14505": "failure",
+    "14506": "failure",
+    "14507": "infeasible",
+    "14508": "failure",
+    "14509": "failure",
+    "14510": "failure",
+    "14511": "failure",
+    "14512": "failure",
+    "14513": "successful",
+    "14514": "failure",
+    "14515": "successful",
+    "14516": "failure",
+    "14517": "infeasible",
+    "14518": "failure",
+    "14519": "failure",
+    "14520": "failure",
+    "14521": "failure",
+    "14522": "successful",
+    "14523": "failure",
+    "14524": "failure",
+    "14525": "failure",
+    "14526": "infeasible",
+    "14527": "failure",
+    "14528": "failure",
+    "14529": "failure",
+    "14530": "successful",
+    "14531": "infeasible",
+    "14532": "failure",
+    "14533": "successful",
+    "14534": "failure",
+    "14535": "failure",
+    "14536": "failure",
+    "14537": "failure",
+    "14538": "failure",
+    "14539": "failure",
+    "14540": "failure",
+    "14541": "failure",
+    "14542": "already_complete",
+    "14543": "failure",
+    "14544": "failure",
+    "14545": "failure",
+    "14546": "failure",
+    "14547": "infeasible",
+    "14548": "failure",
+    "14549": "infeasible",
+    "14550": "failure",
+    "14551": "failure",
+    "14552": "failure",
+    "14553": "failure",
+    "14554": "failure",
+    "14555": "failure",
+    "14556": "failure",
+    "14557": "infeasible",
+    "14558": "failure",
+    "14559": "failure",
+    "14560": "infeasible",
+    "14561": "successful",
+    "14562": "failure",
+    "14563": "failure",
+    "14564": "failure",
+    "14565": "failure",
+    "14566": "successful",
+    "14567": "failure",
+    "14568": "failure",
+    "14569": "failure",
+    "14570": "failure",
+    "14571": "failure",
+    "14572": "failure",
+    "14573": "failure",
+    "14574": "infeasible",
+    "14575": "failure",
+    "14576": "failure",
+    "14577": "failure",
+    "14578": "failure",
+    "14579": "failure",
+    "14580": "successful",
+    "14581": "failure",
+    "14582": "failure",
+    "14583": "failure",
+    "14584": "failure",
+    "14585": "already_complete",
+    "14586": "failure",
+    "14587": "failure",
+    "14588": "failure",
+    "14589": "failure",
+    "14590": "failure",
+    "14591": "failure",
+    "14592": "successful",
+    "14593": "failure",
+    "14594": "failure",
+    "14595": "failure",
+    "14596": "failure",
+    "14597": "failure",
+    "14598": "failure",
+    "14599": "successful",
+    "14600": "infeasible",
+    "14601": "infeasible",
+    "14602": "failure",
+    "14603": "failure",
+    "14604": "infeasible",
+    "14605": "failure",
+    "14606": "infeasible",
+    "14607": "failure",
+    "14608": "failure",
+    "14609": "failure",
+    "14610": "failure",
+    "14611": "failure",
+    "14612": "failure",
+    "14613": "successful",
+    "14614": "successful",
+    "14615": "successful",
+    "14616": "failure",
+    "14617": "failure",
+    "14618": "failure",
+    "14619": "failure",
+    "14620": "successful",
+    "14621": "failure",
+    "14622": "failure",
+    "14623": "infeasible",
+    "14624": "failure",
+    "14625": "failure",
+    "14626": "failure",
+    "14627": "failure",
+    "14628": "failure",
+    "14629": "failure",
+    "14630": "failure",
+    "14631": "failure",
+    "14632": "failure",
+    "14633": "failure",
+    "14634": "failure",
+    "14635": "failure",
+    "14636": "failure",
+    "14637": "successful",
+    "14638": "failure",
+    "14639": "successful",
+    "14640": "successful",
+    "14641": "failure",
+    "14642": "failure",
+    "14643": "failure",
+    "14644": "failure",
+    "14645": "infeasible",
+    "14646": "failure",
+    "14647": "failure",
+    "14648": "failure",
+    "14649": "failure",
+    "14650": "failure",
+    "14651": "failure",
+    "14652": "failure",
+    "14653": "failure",
+    "14654": "failure",
+    "14655": "failure",
+    "14656": "failure",
+    "14657": "failure",
+    "14658": "failure",
+    "14659": "failure",
+    "14660": "successful",
+    "14661": "failure",
+    "14662": "failure",
+    "14663": "failure",
+    "14664": "failure",
+    "14665": "successful",
+    "14666": "infeasible",
+    "14667": "failure",
+    "14668": "successful",
+    "14669": "failure",
+    "14670": "failure",
+    "14671": "successful",
+    "14672": "failure",
+    "14673": "infeasible",
+    "14674": "successful",
+    "14675": "failure",
+    "14676": "failure",
+    "14677": "successful",
+    "14678": "failure",
+    "14679": "failure",
+    "14680": "failure",
+    "14681": "failure",
+    "14682": "successful",
+    "14683": "successful",
+    "14684": "failure",
+    "14685": "already_complete",
+    "14686": "failure",
+    "14687": "failure",
+    "14688": "failure",
+    "14689": "failure",
+    "14690": "successful",
+    "14691": "failure",
+    "14692": "failure",
+    "14693": "failure",
+    "14694": "failure",
+    "14695": "failure",
+    "14696": "failure",
+    "14697": "failure",
+    "14698": "failure",
+    "14699": "failure",
+    "14700": "failure",
+    "14701": "failure",
+    "14702": "failure",
+    "14703": "failure",
+    "14704": "failure",
+    "14705": "infeasible",
+    "14706": "failure",
+    "14707": "failure",
+    "14708": "failure",
+    "14709": "failure",
+    "14710": "failure",
+    "14711": "failure",
+    "14712": "successful",
+    "14713": "failure",
+    "14714": "failure",
+    "14715": "infeasible",
+    "14716": "failure",
+    "14717": "failure",
+    "14718": "failure",
+    "14719": "failure",
+    "14720": "infeasible",
+    "14721": "failure",
+    "14722": "successful",
+    "14723": "failure",
+    "14724": "failure",
+    "14725": "failure",
+    "14726": "failure",
+    "14727": "failure",
+    "14728": "failure",
+    "14729": "failure",
+    "14730": "failure",
+    "14731": "failure",
+    "14732": "failure",
+    "14733": "failure",
+    "14734": "failure",
+    "14735": "failure",
+    "14736": "failure",
+    "14737": "failure",
+    "14738": "failure",
+    "14739": "failure",
+    "14740": "failure",
+    "14741": "failure",
+    "14742": "failure",
+    "14743": "failure",
+    "14744": "failure",
+    "14745": "failure",
+    "14746": "failure",
+    "14747": "failure",
+    "14748": "failure",
+    "14749": "infeasible",
+    "14750": "failure",
+    "14751": "failure",
+    "14752": "failure",
+    "14753": "failure",
+    "14754": "failure",
+    "14755": "failure",
+    "14756": "infeasible",
+    "14757": "successful",
+    "14758": "failure",
+    "14759": "successful",
+    "14760": "failure",
+    "14761": "failure",
+    "14762": "failure",
+    "14763": "failure",
+    "14764": "failure",
+    "14765": "infeasible",
+    "14766": "infeasible",
+    "14767": "successful",
+    "14768": "failure",
+    "14769": "failure",
+    "14770": "failure",
+    "14771": "failure",
+    "14772": "infeasible",
+    "14773": "failure",
+    "14774": "failure",
+    "14775": "failure",
+    "14776": "failure",
+    "14777": "failure",
+    "14778": "failure",
+    "14779": "failure",
+    "14780": "failure",
+    "14781": "failure",
+    "14782": "failure",
+    "14783": "failure",
+    "14784": "failure",
+    "14785": "successful",
+    "14786": "infeasible",
+    "14787": "failure",
+    "14788": "failure",
+    "14789": "failure",
+    "14790": "failure",
+    "14791": "failure",
+    "14792": "failure",
+    "14793": "failure",
+    "14794": "failure",
+    "14795": "failure",
+    "14796": "failure",
+    "14797": "failure",
+    "14798": "successful",
+    "14799": "failure",
+    "14800": "infeasible",
+    "14801": "failure",
+    "14802": "failure",
+    "14803": "failure",
+    "14804": "failure",
+    "14805": "failure",
+    "14806": "infeasible",
+    "14807": "failure",
+    "14808": "failure",
+    "14809": "failure",
+    "14810": "infeasible",
+    "14811": "infeasible",
+    "14812": "failure",
+    "14813": "failure",
+    "14814": "failure",
+    "14815": "failure",
+    "14816": "failure",
+    "14817": "successful",
+    "14818": "failure",
+    "14819": "failure",
+    "14820": "failure",
+    "14821": "failure",
+    "14822": "failure",
+    "14823": "failure",
+    "14824": "failure",
+    "14825": "failure",
+    "14826": "failure",
+    "14827": "failure",
+    "14828": "successful",
+    "14829": "failure",
+    "14830": "failure",
+    "14831": "infeasible",
+    "14832": "failure",
+    "14833": "failure",
+    "14834": "failure",
+    "14835": "failure",
+    "14836": "failure",
+    "14837": "failure",
+    "14838": "failure",
+    "14839": "failure",
+    "14840": "failure",
+    "14841": "failure",
+    "14842": "failure",
+    "14843": "failure",
+    "14844": "failure",
+    "14845": "failure",
+    "14846": "failure",
+    "14847": "failure",
+    "14848": "failure",
+    "14849": "infeasible",
+    "14850": "failure",
+    "14851": "failure",
+    "14852": "successful",
+    "14853": "failure",
+    "14854": "failure",
+    "14855": "failure",
+    "14856": "failure",
+    "14857": "failure",
+    "14858": "successful",
+    "14859": "failure",
+    "14860": "failure",
+    "14861": "successful",
+    "14862": "failure",
+    "14863": "failure",
+    "14864": "infeasible",
+    "14865": "successful",
+    "14866": "failure",
+    "14867": "failure",
+    "14868": "infeasible",
+    "14869": "successful",
+    "14870": "failure",
+    "14871": "failure",
+    "14872": "failure",
+    "14873": "failure",
+    "14874": "failure",
+    "14875": "infeasible",
+    "14876": "successful",
+    "14877": "failure",
+    "14878": "failure",
+    "14879": "failure",
+    "14880": "failure",
+    "14881": "failure",
+    "14882": "failure",
+    "14883": "failure",
+    "14884": "failure",
+    "14885": "failure",
+    "14886": "failure",
+    "14887": "failure",
+    "14888": "failure",
+    "14889": "failure",
+    "14890": "failure",
+    "14891": "failure",
+    "14892": "failure",
+    "14893": "failure",
+    "14894": "failure",
+    "14895": "failure",
+    "14896": "failure",
+    "14897": "failure",
+    "14898": "infeasible",
+    "14899": "failure",
+    "14900": "failure",
+    "14901": "failure",
+    "14902": "failure",
+    "14903": "failure",
+    "14904": "failure",
+    "14905": "successful",
+    "14906": "failure",
+    "14907": "failure",
+    "14908": "successful",
+    "14909": "failure",
+    "14910": "failure",
+    "14911": "failure",
+    "14912": "failure",
+    "14913": "failure",
+    "14914": "successful",
+    "14915": "failure",
+    "14916": "infeasible",
+    "14917": "failure",
+    "14918": "successful",
+    "14919": "failure",
+    "14920": "failure",
+    "14921": "failure",
+    "14922": "infeasible",
+    "14923": "failure",
+    "14924": "failure",
+    "14925": "failure",
+    "14926": "failure",
+    "14927": "failure",
+    "14928": "infeasible",
+    "14929": "failure",
+    "14930": "failure",
+    "14931": "infeasible",
+    "14932": "failure",
+    "14933": "failure",
+    "14934": "failure",
+    "14935": "failure",
+    "14936": "failure",
+    "14937": "successful",
+    "14938": "infeasible",
+    "14939": "failure",
+    "14940": "failure",
+    "14941": "failure",
+    "14942": "successful",
+    "14943": "successful",
+    "14944": "infeasible",
+    "14945": "failure",
+    "14946": "failure",
+    "14947": "successful",
+    "14948": "failure",
+    "14949": "failure",
+    "14950": "successful",
+    "14951": "failure",
+    "14952": "failure",
+    "14953": "successful",
+    "14954": "failure",
+    "14955": "failure",
+    "14956": "failure",
+    "14957": "successful",
+    "14958": "failure",
+    "14959": "failure",
+    "14960": "failure",
+    "14961": "successful",
+    "14962": "failure",
+    "14963": "failure",
+    "14964": "failure",
+    "14965": "failure",
+    "14966": "failure",
+    "14967": "failure",
+    "14968": "successful",
+    "14969": "failure",
+    "14970": "failure",
+    "14971": "failure",
+    "14972": "successful",
+    "14973": "failure",
+    "14974": "infeasible",
+    "14975": "failure",
+    "14976": "infeasible",
+    "14977": "failure",
+    "14978": "failure",
+    "14979": "failure",
+    "14980": "failure",
+    "14981": "failure",
+    "14982": "failure",
+    "14983": "failure",
+    "14984": "failure",
+    "14985": "successful",
+    "14986": "successful",
+    "14987": "failure",
+    "14988": "infeasible",
+    "14989": "failure",
+    "14990": "infeasible",
+    "14991": "failure",
+    "14992": "failure",
+    "14993": "successful",
+    "14994": "failure",
+    "14995": "failure",
+    "14996": "successful",
+    "14997": "failure",
+    "14998": "failure",
+    "14999": "failure",
+    "15000": "failure",
+    "15001": "failure",
+    "15002": "failure",
+    "15003": "failure",
+    "15004": "failure",
+    "15005": "failure",
+    "15006": "successful",
+    "15007": "failure",
+    "15008": "failure",
+    "15009": "infeasible",
+    "15010": "failure",
+    "15011": "successful",
+    "15012": "successful",
+    "15013": "failure",
+    "15014": "successful",
+    "15015": "infeasible",
+    "15016": "failure",
+    "15017": "failure",
+    "15018": "failure",
+    "15019": "failure",
+    "15020": "failure",
+    "15021": "failure",
+    "15022": "failure",
+    "15023": "failure",
+    "15024": "failure",
+    "15025": "failure",
+    "15026": "failure",
+    "15027": "failure",
+    "15028": "failure",
+    "15029": "failure",
+    "15030": "failure",
+    "15031": "failure",
+    "15032": "successful",
+    "15033": "failure",
+    "15034": "failure",
+    "15035": "failure",
+    "15036": "failure",
+    "15037": "failure",
+    "15038": "failure",
+    "15039": "successful",
+    "15040": "infeasible",
+    "15041": "failure",
+    "15042": "failure",
+    "15043": "failure",
+    "15044": "successful",
+    "15045": "failure",
+    "15046": "successful",
+    "15047": "infeasible",
+    "15048": "successful",
+    "15049": "failure",
+    "15050": "failure",
+    "15051": "infeasible",
+    "15052": "infeasible",
+    "15053": "failure",
+    "15054": "failure",
+    "15055": "failure",
+    "15056": "failure",
+    "15057": "successful",
+    "15058": "failure",
+    "15059": "failure",
+    "15060": "failure",
+    "15061": "infeasible",
+    "15062": "failure",
+    "15063": "successful",
+    "15064": "failure",
+    "15065": "failure",
+    "15066": "failure",
+    "15067": "successful",
+    "15068": "failure",
+    "15069": "failure",
+    "15070": "failure",
+    "15071": "failure",
+    "15072": "successful",
+    "15073": "failure",
+    "15074": "failure",
+    "15075": "failure",
+    "15076": "failure",
+    "15077": "failure",
+    "15078": "failure",
+    "15079": "failure",
+    "15080": "failure",
+    "15081": "infeasible",
+    "15082": "infeasible",
+    "15083": "failure",
+    "15084": "failure",
+    "15085": "infeasible",
+    "15086": "failure",
+    "15087": "failure",
+    "15088": "successful",
+    "15089": "successful",
+    "15090": "failure",
+    "15091": "failure",
+    "15092": "failure",
+    "15093": "failure",
+    "15094": "successful",
+    "15095": "failure",
+    "15096": "failure",
+    "15097": "failure",
+    "15098": "failure",
+    "15099": "failure",
+    "15100": "failure",
+    "15101": "failure",
+    "15102": "infeasible",
+    "15103": "failure",
+    "15104": "failure",
+    "15105": "successful",
+    "15106": "failure",
+    "15107": "failure",
+    "15108": "failure",
+    "15109": "failure",
+    "15110": "failure",
+    "15111": "failure",
+    "15112": "failure",
+    "15113": "infeasible",
+    "15114": "failure",
+    "15115": "successful",
+    "15116": "failure",
+    "15117": "failure",
+    "15118": "failure",
+    "15119": "failure",
+    "15120": "failure",
+    "15121": "failure",
+    "15122": "failure",
+    "15123": "infeasible",
+    "15124": "failure",
+    "15125": "failure",
+    "15126": "failure",
+    "15127": "failure",
+    "15128": "failure",
+    "15129": "failure",
+    "15130": "failure",
+    "15131": "failure",
+    "15132": "failure",
+    "15133": "failure",
+    "15134": "failure",
+    "15135": "successful",
+    "15136": "failure",
+    "15137": "failure",
+    "15138": "failure",
+    "15139": "failure",
+    "15140": "failure",
+    "15141": "failure",
+    "15142": "successful",
+    "15143": "successful",
+    "15144": "failure",
+    "15145": "failure",
+    "15146": "failure",
+    "15147": "failure",
+    "15148": "failure",
+    "15149": "failure",
+    "15150": "failure",
+    "15151": "infeasible",
+    "15152": "failure",
+    "15153": "failure",
+    "15154": "failure",
+    "15155": "failure",
+    "15156": "failure",
+    "15157": "failure",
+    "15158": "failure",
+    "15159": "failure",
+    "15160": "failure",
+    "15161": "failure",
+    "15162": "failure",
+    "15163": "failure",
+    "15164": "failure",
+    "15165": "failure",
+    "15166": "failure",
+    "15167": "failure",
+    "15168": "failure",
+    "15169": "infeasible",
+    "15170": "failure",
+    "15171": "successful",
+    "15172": "failure",
+    "15173": "failure",
+    "15174": "failure",
+    "15175": "failure",
+    "15176": "failure",
+    "15177": "failure",
+    "15178": "failure",
+    "15179": "failure",
+    "15180": "failure",
+    "15181": "failure",
+    "15182": "failure",
+    "15183": "infeasible",
+    "15184": "failure",
+    "15185": "failure",
+    "15186": "failure",
+    "15187": "failure",
+    "15188": "successful",
+    "15189": "failure",
+    "15190": "infeasible",
+    "15191": "failure",
+    "15192": "failure",
+    "15193": "successful",
+    "15194": "failure",
+    "15195": "failure",
+    "15196": "failure",
+    "15197": "failure",
+    "15198": "successful",
+    "15199": "infeasible",
+    "15200": "failure",
+    "15201": "failure",
+    "15202": "failure",
+    "15203": "failure",
+    "15204": "failure",
+    "15205": "failure",
+    "15206": "failure",
+    "15207": "failure",
+    "15208": "failure",
+    "15209": "infeasible",
+    "15210": "failure",
+    "15211": "failure",
+    "15212": "failure",
+    "15213": "infeasible",
+    "15214": "failure",
+    "15215": "infeasible",
+    "15216": "failure",
+    "15217": "failure",
+    "15218": "failure",
+    "15219": "successful",
+    "15220": "successful",
+    "15221": "successful",
+    "15222": "successful",
+    "15223": "failure",
+    "15224": "infeasible",
+    "15225": "successful",
+    "15226": "failure",
+    "15227": "failure",
+    "15228": "failure",
+    "15229": "failure",
+    "15230": "failure",
+    "15231": "infeasible",
+    "15232": "failure",
+    "15233": "failure",
+    "15234": "failure",
+    "15235": "failure",
+    "15236": "infeasible",
+    "15237": "failure",
+    "15238": "failure",
+    "15239": "failure",
+    "15240": "failure",
+    "15241": "failure",
+    "15242": "failure",
+    "15243": "failure",
+    "15244": "infeasible",
+    "15245": "failure",
+    "15246": "failure",
+    "15247": "failure",
+    "15248": "failure",
+    "15249": "infeasible",
+    "15250": "failure",
+    "15251": "failure",
+    "15252": "failure",
+    "15253": "failure",
+    "15254": "failure",
+    "15255": "successful",
+    "15256": "failure",
+    "15257": "failure",
+    "15258": "infeasible",
+    "15259": "failure",
+    "15260": "failure",
+    "15261": "infeasible",
+    "15262": "failure",
+    "15263": "infeasible",
+    "15264": "failure",
+    "15265": "failure",
+    "15266": "successful",
+    "15267": "failure",
+    "15268": "failure",
+    "15269": "failure",
+    "15270": "failure",
+    "15271": "failure",
+    "15272": "failure",
+    "15273": "failure",
+    "15274": "successful",
+    "15275": "failure",
+    "15276": "failure",
+    "15277": "infeasible",
+    "15278": "failure",
+    "15279": "successful",
+    "15280": "failure",
+    "15281": "failure",
+    "15282": "failure",
+    "15283": "failure",
+    "15284": "failure",
+    "15285": "failure",
+    "15286": "failure",
+    "15287": "failure",
+    "15288": "failure",
+    "15289": "failure",
+    "15290": "successful",
+    "15291": "failure",
+    "15292": "infeasible",
+    "15293": "failure",
+    "15294": "failure",
+    "15295": "failure",
+    "15296": "failure",
+    "15297": "failure",
+    "15298": "failure",
+    "15299": "successful",
+    "15300": "failure",
+    "15301": "failure",
+    "15302": "failure",
+    "15303": "successful",
+    "15304": "failure",
+    "15305": "failure",
+    "15306": "failure",
+    "15307": "failure",
+    "15308": "failure",
+    "15309": "failure",
+    "15310": "infeasible",
+    "15311": "failure",
+    "15312": "failure",
+    "15313": "failure",
+    "15314": "failure",
+    "15315": "failure",
+    "15316": "failure",
+    "15317": "failure",
+    "15318": "successful",
+    "15319": "failure",
+    "15320": "failure",
+    "15321": "successful",
+    "15322": "failure",
+    "15323": "failure",
+    "15324": "failure",
+    "15325": "failure",
+    "15326": "successful",
+    "15327": "failure",
+    "15328": "failure",
+    "15329": "failure",
+    "15330": "successful",
+    "15331": "failure",
+    "15332": "infeasible",
+    "15333": "failure",
+    "15334": "successful",
+    "15335": "successful",
+    "15336": "successful",
+    "15337": "failure",
+    "15338": "failure",
+    "15339": "failure",
+    "15340": "failure",
+    "15341": "failure",
+    "15342": "failure",
+    "15343": "failure",
+    "15344": "infeasible",
+    "15345": "failure",
+    "15346": "failure",
+    "15347": "failure",
+    "15348": "infeasible",
+    "15349": "failure",
+    "15350": "infeasible",
+    "15351": "failure",
+    "15352": "failure",
+    "15353": "failure",
+    "15354": "failure",
+    "15355": "failure",
+    "15356": "failure",
+    "15357": "failure",
+    "15358": "failure",
+    "15359": "failure",
+    "15360": "failure",
+    "15361": "failure",
+    "15362": "failure",
+    "15363": "failure",
+    "15364": "failure",
+    "15365": "failure",
+    "15366": "failure",
+    "15367": "failure",
+    "15368": "successful",
+    "15369": "failure",
+    "15370": "failure",
+    "15371": "failure",
+    "15372": "failure",
+    "15373": "failure",
+    "15374": "failure",
+    "15375": "failure",
+    "15376": "failure",
+    "15377": "failure",
+    "15378": "infeasible",
+    "15379": "failure",
+    "15380": "successful",
+    "15381": "failure",
+    "15382": "infeasible",
+    "15383": "failure",
+    "15384": "infeasible",
+    "15385": "failure",
+    "15386": "failure",
+    "15387": "failure",
+    "15388": "failure",
+    "15389": "failure",
+    "15390": "failure",
+    "15391": "failure",
+    "15392": "failure",
+    "15393": "failure",
+    "15394": "infeasible",
+    "15395": "failure",
+    "15396": "failure",
+    "15397": "failure",
+    "15398": "infeasible",
+    "15399": "failure",
+    "15400": "failure",
+    "15401": "failure",
+    "15402": "failure",
+    "15403": "successful",
+    "15404": "failure",
+    "15405": "failure",
+    "15406": "successful",
+    "15407": "failure",
+    "15408": "successful",
+    "15409": "successful",
+    "15410": "successful",
+    "15411": "failure",
+    "15412": "failure",
+    "15413": "failure",
+    "15414": "successful",
+    "15415": "failure",
+    "15416": "failure",
+    "15417": "failure",
+    "15418": "successful",
+    "15419": "failure",
+    "15420": "failure",
+    "15421": "failure",
+    "15422": "successful",
+    "15423": "successful",
+    "15424": "failure",
+    "15425": "successful",
+    "15426": "failure",
+    "15427": "failure",
+    "15428": "failure",
+    "15429": "failure",
+    "15430": "failure",
+    "15431": "failure",
+    "15432": "failure",
+    "15433": "failure",
+    "15434": "infeasible",
+    "15435": "successful",
+    "15436": "failure",
+    "15437": "successful",
+    "15438": "failure",
+    "15439": "infeasible",
+    "15440": "failure",
+    "15441": "failure",
+    "15442": "successful",
+    "15443": "failure",
+    "15444": "failure",
+    "15445": "failure",
+    "15446": "failure",
+    "15447": "failure",
+    "15448": "failure",
+    "15449": "failure",
+    "15450": "failure",
+    "15451": "infeasible",
+    "15452": "failure",
+    "15453": "failure",
+    "15454": "infeasible",
+    "15455": "failure",
+    "15456": "infeasible",
+    "15457": "successful",
+    "15458": "infeasible",
+    "15459": "failure",
+    "15460": "infeasible",
+    "15461": "infeasible",
+    "15462": "infeasible",
+    "15463": "failure",
+    "15464": "failure",
+    "15465": "failure",
+    "15466": "failure",
+    "15467": "failure",
+    "15468": "failure",
+    "15469": "failure",
+    "15470": "failure",
+    "15471": "infeasible",
+    "15472": "failure",
+    "15473": "failure",
+    "15474": "infeasible",
+    "15475": "infeasible",
+    "15476": "infeasible",
+    "15477": "failure",
+    "15478": "successful",
+    "15479": "failure",
+    "15480": "failure",
+    "15481": "failure",
+    "15482": "infeasible",
+    "15483": "failure",
+    "15484": "failure",
+    "15485": "failure",
+    "15486": "failure",
+    "15487": "failure",
+    "15488": "failure",
+    "15489": "failure",
+    "15490": "failure",
+    "15491": "failure",
+    "15492": "infeasible",
+    "15493": "failure",
+    "15494": "failure",
+    "15495": "failure",
+    "15496": "failure",
+    "15497": "failure",
+    "15498": "failure",
+    "15499": "failure",
+    "15500": "failure",
+    "15501": "infeasible",
+    "15502": "failure",
+    "15503": "failure",
+    "15504": "successful",
+    "15505": "infeasible",
+    "15506": "failure",
+    "15507": "failure",
+    "15508": "infeasible",
+    "15509": "failure",
+    "15510": "failure",
+    "15511": "successful",
+    "15512": "failure",
+    "15513": "failure",
+    "15514": "failure",
+    "15515": "failure",
+    "15516": "failure",
+    "15517": "failure",
+    "15518": "failure",
+    "15519": "failure",
+    "15520": "failure",
+    "15521": "failure",
+    "15522": "failure",
+    "15523": "failure",
+    "15524": "failure",
+    "15525": "failure",
+    "15526": "failure",
+    "15527": "failure",
+    "15528": "successful",
+    "15529": "failure",
+    "15530": "successful",
+    "15531": "failure",
+    "15532": "failure",
+    "15533": "successful",
+    "15534": "failure",
+    "15535": "failure",
+    "15536": "failure",
+    "15537": "successful",
+    "15538": "failure",
+    "15539": "failure",
+    "15540": "failure",
+    "15541": "failure",
+    "15542": "failure",
+    "15543": "failure",
+    "15544": "failure",
+    "15545": "infeasible",
+    "15546": "failure",
+    "15547": "failure",
+    "15548": "successful",
+    "15549": "already_complete",
+    "15550": "failure",
+    "15551": "failure",
+    "15552": "failure",
+    "15553": "failure",
+    "15554": "infeasible",
+    "15555": "failure",
+    "15556": "infeasible",
+    "15557": "failure",
+    "15558": "failure",
+    "15559": "infeasible",
+    "15560": "failure",
+    "15561": "successful",
+    "15562": "failure",
+    "15563": "failure",
+    "15564": "failure",
+    "15565": "failure",
+    "15566": "failure",
+    "15567": "failure",
+    "15568": "failure",
+    "15569": "failure",
+    "15570": "failure",
+    "15571": "failure",
+    "15572": "failure",
+    "15573": "failure",
+    "15574": "failure",
+    "15575": "failure",
+    "15576": "failure",
+    "15577": "failure",
+    "15578": "failure",
+    "15579": "failure",
+    "15580": "failure",
+    "15581": "failure",
+    "15582": "failure",
+    "15583": "failure",
+    "15584": "successful",
+    "15585": "failure",
+    "15586": "failure",
+    "15587": "failure",
+    "15588": "failure",
+    "15589": "failure",
+    "15590": "failure",
+    "15591": "failure",
+    "15592": "infeasible",
+    "15593": "failure",
+    "15594": "failure",
+    "15595": "infeasible",
+    "15596": "failure",
+    "15597": "failure",
+    "15598": "failure",
+    "15599": "failure",
+    "15600": "failure",
+    "15601": "failure",
+    "15602": "failure",
+    "15603": "failure",
+    "15604": "infeasible",
+    "15605": "failure",
+    "15606": "failure",
+    "15607": "failure",
+    "15608": "failure",
+    "15609": "failure",
+    "15610": "infeasible",
+    "15611": "failure",
+    "15612": "failure",
+    "15613": "failure",
+    "15614": "failure",
+    "15615": "failure",
+    "15616": "failure",
+    "15617": "failure",
+    "15618": "failure",
+    "15619": "failure",
+    "15620": "failure",
+    "15621": "failure",
+    "15622": "failure",
+    "15623": "failure",
+    "15624": "failure",
+    "15625": "failure",
+    "15626": "failure",
+    "15627": "failure",
+    "15628": "failure",
+    "15629": "failure",
+    "15630": "failure",
+    "15631": "successful",
+    "15632": "failure",
+    "15633": "failure",
+    "15634": "failure",
+    "15635": "failure",
+    "15636": "infeasible",
+    "15637": "failure",
+    "15638": "failure",
+    "15639": "failure",
+    "15640": "failure",
+    "15641": "failure",
+    "15642": "failure",
+    "15643": "failure",
+    "15644": "infeasible",
+    "15645": "successful",
+    "15646": "successful",
+    "15647": "failure",
+    "15648": "failure",
+    "15649": "failure",
+    "15650": "failure",
+    "15651": "failure",
+    "15652": "failure",
+    "15653": "failure",
+    "15654": "failure",
+    "15655": "failure",
+    "15656": "failure",
+    "15657": "successful",
+    "15658": "failure",
+    "15659": "failure",
+    "15660": "failure",
+    "15661": "failure",
+    "15662": "infeasible",
+    "15663": "failure",
+    "15664": "failure",
+    "15665": "failure",
+    "15666": "successful",
+    "15667": "failure",
+    "15668": "failure",
+    "15669": "failure",
+    "15670": "failure",
+    "15671": "failure",
+    "15672": "successful",
+    "15673": "failure",
+    "15674": "failure",
+    "15675": "failure",
+    "15676": "failure",
+    "15677": "failure",
+    "15678": "failure",
+    "15679": "failure",
+    "15680": "infeasible",
+    "15681": "failure",
+    "15682": "successful",
+    "15683": "failure",
+    "15684": "failure",
+    "15685": "failure",
+    "15686": "failure",
+    "15687": "infeasible",
+    "15688": "failure",
+    "15689": "failure",
+    "15690": "failure",
+    "15691": "failure",
+    "15692": "failure",
+    "15693": "failure",
+    "15694": "failure",
+    "15695": "failure",
+    "15696": "infeasible",
+    "15697": "failure",
+    "15698": "failure",
+    "15699": "successful",
+    "15700": "failure",
+    "15701": "failure",
+    "15702": "failure",
+    "15703": "failure",
+    "15704": "successful",
+    "15705": "failure",
+    "15706": "successful",
+    "15707": "successful",
+    "15708": "failure",
+    "15709": "failure",
+    "15710": "failure",
+    "15711": "failure",
+    "15712": "failure",
+    "15713": "failure",
+    "15714": "failure",
+    "15715": "failure",
+    "15716": "failure",
+    "15717": "failure",
+    "15718": "failure",
+    "15719": "failure",
+    "15720": "failure",
+    "15721": "failure",
+    "15722": "failure",
+    "15723": "failure",
+    "15724": "failure",
+    "15725": "failure",
+    "15726": "successful",
+    "15727": "failure",
+    "15728": "infeasible",
+    "15729": "failure",
+    "15730": "successful",
+    "15731": "successful",
+    "15732": "failure",
+    "15733": "failure",
+    "15734": "successful",
+    "15735": "failure",
+    "15736": "failure",
+    "15737": "infeasible",
+    "15738": "failure",
+    "15739": "failure",
+    "15740": "failure",
+    "15741": "successful",
+    "15742": "successful",
+    "15743": "failure",
+    "15744": "failure",
+    "15745": "failure",
+    "15746": "failure",
+    "15747": "failure",
+    "15748": "failure",
+    "15749": "failure",
+    "15750": "failure",
+    "15751": "failure",
+    "15752": "failure",
+    "15753": "failure",
+    "15754": "failure",
+    "15755": "failure",
+    "15756": "failure",
+    "15757": "failure",
+    "15758": "failure",
+    "15759": "successful",
+    "15760": "successful",
+    "15761": "failure",
+    "15762": "failure",
+    "15763": "failure",
+    "15764": "failure",
+    "15765": "failure",
+    "15766": "successful",
+    "15767": "failure",
+    "15768": "failure",
+    "15769": "failure",
+    "15770": "failure",
+    "15771": "failure",
+    "15772": "infeasible",
+    "15773": "infeasible",
+    "15774": "failure",
+    "15775": "successful",
+    "15776": "failure",
+    "15777": "failure",
+    "15778": "failure",
+    "15779": "failure",
+    "15780": "failure",
+    "15781": "failure",
+    "15782": "successful",
+    "15783": "infeasible",
+    "15784": "failure",
+    "15785": "failure",
+    "15786": "failure",
+    "15787": "failure",
+    "15788": "failure",
+    "15789": "successful",
+    "15790": "failure",
+    "15791": "successful",
+    "15792": "failure",
+    "15793": "failure",
+    "15794": "failure",
+    "15795": "successful",
+    "15796": "failure",
+    "15797": "failure",
+    "15798": "successful",
+    "15799": "infeasible",
+    "15800": "failure",
+    "15801": "failure",
+    "15802": "failure",
+    "15803": "infeasible",
+    "15804": "failure",
+    "15805": "infeasible",
+    "15806": "failure",
+    "15807": "successful",
+    "15808": "failure",
+    "15809": "failure",
+    "15810": "successful",
+    "15811": "successful",
+    "15812": "failure",
+    "15813": "failure",
+    "15814": "failure",
+    "15815": "failure",
+    "15816": "failure",
+    "15817": "failure",
+    "15818": "failure",
+    "15819": "successful",
+    "15820": "failure",
+    "15821": "failure",
+    "15822": "failure",
+    "15823": "failure",
+    "15824": "failure",
+    "15825": "failure",
+    "15826": "failure",
+    "15827": "failure",
+    "15828": "failure",
+    "15829": "successful",
+    "15830": "failure",
+    "15831": "failure",
+    "15832": "failure",
+    "15833": "failure",
+    "15834": "failure",
+    "15835": "successful",
+    "15836": "failure",
+    "15837": "failure",
+    "15838": "failure",
+    "15839": "failure",
+    "15840": "failure",
+    "15841": "failure",
+    "15842": "failure",
+    "15843": "failure",
+    "15844": "failure",
+    "15845": "failure",
+    "15846": "failure",
+    "15847": "failure",
+    "15848": "failure",
+    "15849": "failure",
+    "15850": "failure",
+    "15851": "failure",
+    "15852": "infeasible",
+    "15853": "infeasible",
+    "15854": "failure",
+    "15855": "infeasible",
+    "15856": "failure",
+    "15857": "failure",
+    "15858": "failure",
+    "15859": "failure",
+    "15860": "failure",
+    "15861": "successful",
+    "15862": "already_complete",
+    "15863": "failure",
+    "15864": "failure",
+    "15865": "failure",
+    "15866": "failure",
+    "15867": "failure",
+    "15868": "failure",
+    "15869": "failure",
+    "15870": "failure",
+    "15871": "successful",
+    "15872": "failure",
+    "15873": "successful",
+    "15874": "failure",
+    "15875": "failure",
+    "15876": "failure",
+    "15877": "failure",
+    "15878": "failure",
+    "15879": "failure",
+    "15880": "failure",
+    "15881": "failure",
+    "15882": "failure",
+    "15883": "failure",
+    "15884": "successful",
+    "15885": "failure",
+    "15886": "already_complete",
+    "15887": "failure",
+    "15888": "failure",
+    "15889": "successful",
+    "15890": "failure",
+    "15891": "failure",
+    "15892": "failure",
+    "15893": "failure",
+    "15894": "failure",
+    "15895": "failure",
+    "15896": "failure",
+    "15897": "failure",
+    "15898": "infeasible",
+    "15899": "failure",
+    "15900": "failure",
+    "15901": "failure",
+    "15902": "failure",
+    "15903": "failure",
+    "15904": "failure",
+    "15905": "failure",
+    "15906": "failure",
+    "15907": "failure",
+    "15908": "failure",
+    "15909": "successful",
+    "15910": "failure",
+    "15911": "failure",
+    "15912": "failure",
+    "15913": "failure",
+    "15914": "failure",
+    "15915": "failure",
+    "15916": "failure",
+    "15917": "infeasible",
+    "15918": "failure",
+    "15919": "failure",
+    "15920": "failure",
+    "15921": "failure",
+    "15922": "failure",
+    "15923": "successful",
+    "15924": "failure",
+    "15925": "successful",
+    "15926": "failure",
+    "15927": "failure",
+    "15928": "failure",
+    "15929": "successful",
+    "15930": "failure",
+    "15931": "successful",
+    "15932": "failure",
+    "15933": "failure",
+    "15934": "failure",
+    "15935": "infeasible",
+    "15936": "failure",
+    "15937": "infeasible",
+    "15938": "failure",
+    "15939": "already_complete",
+    "15940": "failure",
+    "15941": "failure",
+    "15942": "failure",
+    "15943": "failure",
+    "15944": "failure",
+    "15945": "successful",
+    "15946": "failure",
+    "15947": "failure",
+    "15948": "failure",
+    "15949": "failure",
+    "15950": "successful",
+    "15951": "infeasible",
+    "15952": "failure",
+    "15953": "failure",
+    "15954": "successful",
+    "15955": "failure",
+    "15956": "infeasible",
+    "15957": "successful",
+    "15958": "failure",
+    "15959": "failure",
+    "15960": "failure",
+    "15961": "failure",
+    "15962": "successful",
+    "15963": "failure",
+    "15964": "failure",
+    "15965": "successful",
+    "15966": "failure",
+    "15967": "failure",
+    "15968": "failure",
+    "15969": "failure",
+    "15970": "failure",
+    "15971": "failure",
+    "15972": "failure",
+    "15973": "failure",
+    "15974": "failure",
+    "15975": "failure",
+    "15976": "failure",
+    "15977": "successful",
+    "15978": "infeasible",
+    "15979": "failure",
+    "15980": "infeasible",
+    "15981": "failure",
+    "15982": "infeasible",
+    "15983": "failure",
+    "15984": "successful",
+    "15985": "successful",
+    "15986": "successful",
+    "15987": "failure",
+    "15988": "failure",
+    "15989": "failure",
+    "15990": "failure",
+    "15991": "failure",
+    "15992": "failure",
+    "15993": "infeasible",
+    "15994": "successful",
+    "15995": "failure",
+    "15996": "failure",
+    "15997": "failure",
+    "15998": "failure",
+    "15999": "failure",
+    "16000": "infeasible",
+    "16001": "failure",
+    "16002": "failure",
+    "16003": "failure",
+    "16004": "failure",
+    "16005": "failure",
+    "16006": "failure",
+    "16007": "failure",
+    "16008": "failure",
+    "16009": "failure",
+    "16010": "failure",
+    "16011": "failure",
+    "16012": "infeasible",
+    "16013": "successful",
+    "16014": "failure",
+    "16015": "failure",
+    "16016": "infeasible",
+    "16017": "failure",
+    "16018": "failure",
+    "16019": "successful",
+    "16020": "failure",
+    "16021": "failure",
+    "16022": "failure",
+    "16023": "successful",
+    "16024": "infeasible",
+    "16025": "failure",
+    "16026": "failure",
+    "16027": "failure",
+    "16028": "failure",
+    "16029": "failure",
+    "16030": "infeasible",
+    "16031": "failure",
+    "16032": "failure",
+    "16033": "failure",
+    "16034": "failure",
+    "16035": "failure",
+    "16036": "failure",
+    "16037": "failure",
+    "16038": "failure",
+    "16039": "successful",
+    "16040": "failure",
+    "16041": "failure",
+    "16042": "failure",
+    "16043": "failure",
+    "16044": "failure",
+    "16045": "failure",
+    "16046": "failure",
+    "16047": "failure",
+    "16048": "failure",
+    "16049": "failure",
+    "16050": "failure",
+    "16051": "failure",
+    "16052": "failure",
+    "16053": "failure",
+    "16054": "failure",
+    "16055": "failure",
+    "16056": "failure",
+    "16057": "failure",
+    "16058": "successful",
+    "16059": "failure",
+    "16060": "failure",
+    "16061": "successful",
+    "16062": "failure",
+    "16063": "failure",
+    "16064": "failure",
+    "16065": "failure",
+    "16066": "failure",
+    "16067": "failure",
+    "16068": "failure",
+    "16069": "infeasible",
+    "16070": "failure",
+    "16071": "failure",
+    "16072": "infeasible",
+    "16073": "infeasible",
+    "16074": "failure",
+    "16075": "infeasible",
+    "16076": "failure",
+    "16077": "infeasible",
+    "16078": "failure",
+    "16079": "failure",
+    "16080": "failure",
+    "16081": "failure",
+    "16082": "failure",
+    "16083": "failure",
+    "16084": "failure",
+    "16085": "failure",
+    "16086": "successful",
+    "16087": "failure",
+    "16088": "failure",
+    "16089": "failure",
+    "16090": "failure",
+    "16091": "failure",
+    "16092": "infeasible",
+    "16093": "failure",
+    "16094": "failure",
+    "16095": "failure",
+    "16096": "failure",
+    "16097": "failure",
+    "16098": "failure",
+    "16099": "failure",
+    "16100": "infeasible",
+    "16101": "failure",
+    "16102": "failure",
+    "16103": "failure",
+    "16104": "failure",
+    "16105": "failure",
+    "16106": "failure",
+    "16107": "failure",
+    "16108": "successful",
+    "16109": "failure",
+    "16110": "failure",
+    "16111": "failure",
+    "16112": "failure",
+    "16113": "failure",
+    "16114": "failure",
+    "16115": "failure",
+    "16116": "failure",
+    "16117": "infeasible",
+    "16118": "failure",
+    "16119": "failure",
+    "16120": "failure",
+    "16121": "failure",
+    "16122": "failure",
+    "16123": "failure",
+    "16124": "failure",
+    "16125": "failure",
+    "16126": "failure",
+    "16127": "failure",
+    "16128": "failure",
+    "16129": "failure",
+    "16130": "failure",
+    "16131": "failure",
+    "16132": "failure",
+    "16133": "failure",
+    "16134": "successful",
+    "16135": "failure",
+    "16136": "successful",
+    "16137": "failure",
+    "16138": "failure",
+    "16139": "failure",
+    "16140": "failure",
+    "16141": "successful",
+    "16142": "failure",
+    "16143": "infeasible",
+    "16144": "failure",
+    "16145": "failure",
+    "16146": "failure",
+    "16147": "failure",
+    "16148": "failure",
+    "16149": "failure",
+    "16150": "successful",
+    "16151": "successful",
+    "16152": "failure",
+    "16153": "failure",
+    "16154": "failure",
+    "16155": "infeasible",
+    "16156": "failure",
+    "16157": "failure",
+    "16158": "failure",
+    "16159": "successful",
+    "16160": "successful",
+    "16161": "successful",
+    "16162": "successful",
+    "16163": "failure",
+    "16164": "failure",
+    "16165": "failure",
+    "16166": "failure",
+    "16167": "failure",
+    "16168": "failure",
+    "16169": "failure",
+    "16170": "failure",
+    "16171": "failure",
+    "16172": "failure",
+    "16173": "successful",
+    "16174": "failure",
+    "16175": "failure",
+    "16176": "infeasible",
+    "16177": "failure",
+    "16178": "failure",
+    "16179": "failure",
+    "16180": "failure",
+    "16181": "failure",
+    "16182": "failure",
+    "16183": "failure",
+    "16184": "failure",
+    "16185": "failure",
+    "16186": "failure",
+    "16187": "failure",
+    "16188": "failure",
+    "16189": "failure",
+    "16190": "infeasible",
+    "16191": "failure",
+    "16192": "failure",
+    "16193": "failure",
+    "16194": "failure",
+    "16195": "infeasible",
+    "16196": "failure",
+    "16197": "failure",
+    "16198": "failure",
+    "16199": "failure",
+    "16200": "failure",
+    "16201": "failure",
+    "16202": "successful",
+    "16203": "successful",
+    "16204": "failure",
+    "16205": "failure",
+    "16206": "successful",
+    "16207": "failure",
+    "16208": "failure",
+    "16209": "failure",
+    "16210": "failure",
+    "16211": "failure",
+    "16212": "failure",
+    "16213": "failure",
+    "16214": "failure",
+    "16215": "failure",
+    "16216": "failure",
+    "16217": "successful",
+    "16218": "failure",
+    "16219": "failure",
+    "16220": "failure",
+    "16221": "failure",
+    "16222": "failure",
+    "16223": "failure",
+    "16224": "failure",
+    "16225": "failure",
+    "16226": "failure",
+    "16227": "failure",
+    "16228": "successful",
+    "16229": "failure",
+    "16230": "infeasible",
+    "16231": "failure",
+    "16232": "failure",
+    "16233": "failure",
+    "16234": "failure",
+    "16235": "failure",
+    "16236": "failure",
+    "16237": "successful",
+    "16238": "failure",
+    "16239": "failure",
+    "16240": "failure",
+    "16241": "failure",
+    "16242": "failure",
+    "16243": "failure",
+    "16244": "failure",
+    "16245": "failure",
+    "16246": "failure",
+    "16247": "failure",
+    "16248": "failure",
+    "16249": "successful",
+    "16250": "infeasible",
+    "16251": "failure",
+    "16252": "infeasible",
+    "16253": "failure",
+    "16254": "failure",
+    "16255": "failure",
+    "16256": "infeasible",
+    "16257": "failure",
+    "16258": "failure",
+    "16259": "failure",
+    "16260": "failure",
+    "16261": "failure",
+    "16262": "successful",
+    "16263": "failure",
+    "16264": "failure",
+    "16265": "failure",
+    "16266": "failure",
+    "16267": "infeasible",
+    "16268": "failure",
+    "16269": "failure",
+    "16270": "failure",
+    "16271": "infeasible",
+    "16272": "failure",
+    "16273": "failure",
+    "16274": "infeasible",
+    "16275": "failure",
+    "16276": "failure",
+    "16277": "failure",
+    "16278": "failure",
+    "16279": "successful",
+    "16280": "failure",
+    "16281": "failure",
+    "16282": "failure",
+    "16283": "failure",
+    "16284": "failure",
+    "16285": "already_complete",
+    "16286": "infeasible",
+    "16287": "failure",
+    "16288": "infeasible",
+    "16289": "failure",
+    "16290": "failure",
+    "16291": "failure",
+    "16292": "failure",
+    "16293": "failure",
+    "16294": "failure",
+    "16295": "failure",
+    "16296": "failure",
+    "16297": "failure",
+    "16298": "failure",
+    "16299": "successful",
+    "16300": "failure",
+    "16301": "failure",
+    "16302": "failure",
+    "16303": "failure",
+    "16304": "successful",
+    "16305": "infeasible",
+    "16306": "failure",
+    "16307": "successful",
+    "16308": "failure",
+    "16309": "failure",
+    "16310": "failure",
+    "16311": "failure",
+    "16312": "failure",
+    "16313": "failure",
+    "16314": "failure",
+    "16315": "successful",
+    "16316": "failure",
+    "16317": "failure",
+    "16318": "failure",
+    "16319": "failure",
+    "16320": "failure",
+    "16321": "infeasible",
+    "16322": "failure",
+    "16323": "failure",
+    "16324": "failure",
+    "16325": "failure",
+    "16326": "infeasible",
+    "16327": "failure",
+    "16328": "failure",
+    "16329": "infeasible",
+    "16330": "successful",
+    "16331": "infeasible",
+    "16332": "failure",
+    "16333": "successful",
+    "16334": "failure",
+    "16335": "failure",
+    "16336": "failure",
+    "16337": "failure",
+    "16338": "successful",
+    "16339": "infeasible",
+    "16340": "successful",
+    "16341": "successful",
+    "16342": "successful",
+    "16343": "failure",
+    "16344": "failure",
+    "16345": "failure",
+    "16346": "failure",
+    "16347": "failure",
+    "16348": "infeasible",
+    "16349": "failure",
+    "16350": "failure",
+    "16351": "failure",
+    "16352": "failure",
+    "16353": "failure",
+    "16354": "failure",
+    "16355": "successful",
+    "16356": "infeasible",
+    "16357": "failure",
+    "16358": "failure",
+    "16359": "failure",
+    "16360": "failure",
+    "16361": "successful",
+    "16362": "infeasible",
+    "16363": "successful",
+    "16364": "successful",
+    "16365": "failure",
+    "16366": "failure",
+    "16367": "failure",
+    "16368": "failure",
+    "16369": "infeasible",
+    "16370": "failure",
+    "16371": "successful",
+    "16372": "infeasible",
+    "16373": "infeasible",
+    "16374": "successful",
+    "16375": "failure",
+    "16376": "failure",
+    "16377": "failure",
+    "16378": "infeasible",
+    "16379": "failure",
+    "16380": "failure",
+    "16381": "failure",
+    "16382": "failure",
+    "16383": "failure",
+    "16384": "failure",
+    "16385": "failure",
+    "16386": "failure",
+    "16387": "failure",
+    "16388": "successful",
+    "16389": "failure",
+    "16390": "failure",
+    "16391": "failure",
+    "16392": "infeasible",
+    "16393": "failure",
+    "16394": "failure",
+    "16395": "failure",
+    "16396": "failure",
+    "16397": "failure",
+    "16398": "failure",
+    "16399": "infeasible",
+    "16400": "infeasible",
+    "16401": "failure",
+    "16402": "successful",
+    "16403": "failure",
+    "16404": "successful",
+    "16405": "failure",
+    "16406": "failure",
+    "16407": "infeasible",
+    "16408": "failure",
+    "16409": "failure",
+    "16410": "successful",
+    "16411": "failure",
+    "16412": "failure",
+    "16413": "failure",
+    "16414": "failure",
+    "16415": "failure",
+    "16416": "failure",
+    "16417": "failure",
+    "16418": "failure",
+    "16419": "infeasible",
+    "16420": "failure",
+    "16421": "failure",
+    "16422": "failure",
+    "16423": "failure",
+    "16424": "infeasible",
+    "16425": "failure",
+    "16426": "failure",
+    "16427": "failure",
+    "16428": "failure",
+    "16429": "failure",
+    "16430": "failure",
+    "16431": "failure",
+    "16432": "failure",
+    "16433": "failure",
+    "16434": "failure",
+    "16435": "failure",
+    "16436": "infeasible",
+    "16437": "failure",
+    "16438": "failure",
+    "16439": "failure",
+    "16440": "failure",
+    "16441": "failure",
+    "16442": "failure",
+    "16443": "failure",
+    "16444": "failure",
+    "16445": "failure",
+    "16446": "successful",
+    "16447": "failure",
+    "16448": "failure",
+    "16449": "failure",
+    "16450": "failure",
+    "16451": "failure",
+    "16452": "successful",
+    "16453": "successful",
+    "16454": "failure",
+    "16455": "failure",
+    "16456": "failure",
+    "16457": "infeasible",
+    "16458": "infeasible",
+    "16459": "successful",
+    "16460": "successful",
+    "16461": "failure",
+    "16462": "successful",
+    "16463": "failure",
+    "16464": "failure",
+    "16465": "failure",
+    "16466": "failure",
+    "16467": "failure",
+    "16468": "failure",
+    "16469": "successful",
+    "16470": "failure",
+    "16471": "failure",
+    "16472": "failure",
+    "16473": "successful",
+    "16474": "failure",
+    "16475": "failure",
+    "16476": "failure",
+    "16477": "failure",
+    "16478": "successful",
+    "16479": "infeasible",
+    "16480": "failure",
+    "16481": "failure",
+    "16482": "failure",
+    "16483": "failure",
+    "16484": "failure",
+    "16485": "failure",
+    "16486": "successful",
+    "16487": "failure",
+    "16488": "failure",
+    "16489": "failure",
+    "16490": "failure",
+    "16491": "infeasible",
+    "16492": "failure",
+    "16493": "infeasible",
+    "16494": "failure",
+    "16495": "failure",
+    "16496": "failure",
+    "16497": "failure",
+    "16498": "failure",
+    "16499": "failure",
+    "16500": "successful",
+    "16501": "failure",
+    "16502": "failure",
+    "16503": "failure",
+    "16504": "failure",
+    "16505": "failure",
+    "16506": "failure",
+    "16507": "failure",
+    "16508": "failure",
+    "16509": "failure",
+    "16510": "failure",
+    "16511": "failure",
+    "16512": "failure",
+    "16513": "failure",
+    "16514": "failure",
+    "16515": "failure",
+    "16516": "failure",
+    "16517": "failure",
+    "16518": "successful",
+    "16519": "failure",
+    "16520": "failure",
+    "16521": "failure",
+    "16522": "failure",
+    "16523": "failure",
+    "16524": "failure",
+    "16525": "failure",
+    "16526": "failure",
+    "16527": "failure",
+    "16528": "infeasible",
+    "16529": "failure",
+    "16530": "failure",
+    "16531": "failure",
+    "16532": "failure",
+    "16533": "failure",
+    "16534": "failure",
+    "16535": "failure",
+    "16536": "infeasible",
+    "16537": "successful",
+    "16538": "successful",
+    "16539": "failure",
+    "16540": "failure",
+    "16541": "failure",
+    "16542": "failure",
+    "16543": "failure",
+    "16544": "failure",
+    "16545": "failure",
+    "16546": "failure",
+    "16547": "infeasible",
+    "16548": "failure",
+    "16549": "failure",
+    "16550": "failure",
+    "16551": "failure",
+    "16552": "failure",
+    "16553": "failure",
+    "16554": "infeasible",
+    "16555": "failure",
+    "16556": "infeasible",
+    "16557": "failure",
+    "16558": "failure",
+    "16559": "failure",
+    "16560": "successful",
+    "16561": "failure",
+    "16562": "successful",
+    "16563": "failure",
+    "16564": "failure",
+    "16565": "infeasible",
+    "16566": "failure",
+    "16567": "successful",
+    "16568": "failure",
+    "16569": "successful",
+    "16570": "failure",
+    "16571": "failure",
+    "16572": "successful",
+    "16573": "failure",
+    "16574": "failure",
+    "16575": "failure",
+    "16576": "failure",
+    "16577": "already_complete",
+    "16578": "infeasible",
+    "16579": "failure",
+    "16580": "failure",
+    "16581": "failure",
+    "16582": "successful",
+    "16583": "failure",
+    "16584": "failure",
+    "16585": "infeasible",
+    "16586": "successful",
+    "16587": "successful",
+    "16588": "failure",
+    "16589": "failure",
+    "16590": "failure",
+    "16591": "infeasible",
+    "16592": "failure",
+    "16593": "failure",
+    "16594": "failure",
+    "16595": "failure",
+    "16596": "successful",
+    "16597": "failure",
+    "16598": "successful",
+    "16599": "failure",
+    "16600": "successful",
+    "16601": "infeasible",
+    "16602": "failure",
+    "16603": "failure",
+    "16604": "failure",
+    "16605": "failure",
+    "16606": "failure",
+    "16607": "successful",
+    "16608": "failure",
+    "16609": "infeasible",
+    "16610": "infeasible",
+    "16611": "failure",
+    "16612": "failure",
+    "16613": "failure",
+    "16614": "successful",
+    "16615": "failure",
+    "16616": "failure",
+    "16617": "failure",
+    "16618": "failure",
+    "16619": "failure",
+    "16620": "failure",
+    "16621": "failure",
+    "16622": "failure",
+    "16623": "successful",
+    "16624": "failure",
+    "16625": "failure",
+    "16626": "successful",
+    "16627": "failure",
+    "16628": "infeasible",
+    "16629": "successful",
+    "16630": "failure",
+    "16631": "failure",
+    "16632": "failure",
+    "16633": "failure",
+    "16634": "failure",
+    "16635": "failure",
+    "16636": "failure",
+    "16637": "failure",
+    "16638": "failure",
+    "16639": "failure",
+    "16640": "successful",
+    "16641": "failure",
+    "16642": "failure",
+    "16643": "failure",
+    "16644": "failure",
+    "16645": "failure",
+    "16646": "failure",
+    "16647": "failure",
+    "16648": "failure",
+    "16649": "failure",
+    "16650": "failure",
+    "16651": "failure",
+    "16652": "failure",
+    "16653": "failure",
+    "16654": "failure",
+    "16655": "failure",
+    "16656": "failure",
+    "16657": "successful",
+    "16658": "failure",
+    "16659": "failure",
+    "16660": "failure",
+    "16661": "failure",
+    "16662": "failure",
+    "16663": "failure",
+    "16664": "failure",
+    "16665": "failure",
+    "16666": "failure",
+    "16667": "failure",
+    "16668": "failure",
+    "16669": "failure",
+    "16670": "failure",
+    "16671": "failure",
+    "16672": "failure",
+    "16673": "infeasible",
+    "16674": "failure",
+    "16675": "failure",
+    "16676": "failure",
+    "16677": "failure",
+    "16678": "failure",
+    "16679": "failure",
+    "16680": "failure",
+    "16681": "failure",
+    "16682": "failure",
+    "16683": "failure",
+    "16684": "failure",
+    "16685": "infeasible",
+    "16686": "failure",
+    "16687": "failure",
+    "16688": "failure",
+    "16689": "failure",
+    "16690": "infeasible",
+    "16691": "successful",
+    "16692": "successful",
+    "16693": "failure",
+    "16694": "failure",
+    "16695": "failure",
+    "16696": "failure",
+    "16697": "failure",
+    "16698": "successful",
+    "16699": "successful",
+    "16700": "failure",
+    "16701": "failure",
+    "16702": "failure",
+    "16703": "successful",
+    "16704": "failure",
+    "16705": "failure",
+    "16706": "successful",
+    "16707": "failure",
+    "16708": "failure",
+    "16709": "failure",
+    "16710": "failure",
+    "16711": "failure",
+    "16712": "failure",
+    "16713": "failure",
+    "16714": "successful",
+    "16715": "failure",
+    "16716": "failure",
+    "16717": "failure",
+    "16718": "failure",
+    "16719": "successful",
+    "16720": "failure",
+    "16721": "failure",
+    "16722": "failure",
+    "16723": "failure",
+    "16724": "failure",
+    "16725": "failure",
+    "16726": "infeasible",
+    "16727": "successful",
+    "16728": "failure",
+    "16729": "failure",
+    "16730": "failure",
+    "16731": "failure",
+    "16732": "infeasible",
+    "16733": "failure",
+    "16734": "failure",
+    "16735": "failure",
+    "16736": "infeasible",
+    "16737": "failure",
+    "16738": "infeasible",
+    "16739": "failure",
+    "16740": "failure",
+    "16741": "failure",
+    "16742": "failure",
+    "16743": "failure",
+    "16744": "infeasible",
+    "16745": "failure",
+    "16746": "failure",
+    "16747": "failure",
+    "16748": "failure",
+    "16749": "failure",
+    "16750": "failure",
+    "16751": "failure",
+    "16752": "failure",
+    "16753": "failure",
+    "16754": "failure",
+    "16755": "failure",
+    "16756": "failure",
+    "16757": "failure",
+    "16758": "successful",
+    "16759": "failure",
+    "16760": "failure",
+    "16761": "failure",
+    "16762": "failure",
+    "16763": "failure",
+    "16764": "failure",
+    "16765": "successful",
+    "16766": "successful",
+    "16767": "successful",
+    "16768": "infeasible",
+    "16769": "failure",
+    "16770": "failure",
+    "16771": "failure",
+    "16772": "failure",
+    "16773": "failure",
+    "16774": "failure",
+    "16775": "failure",
+    "16776": "infeasible",
+    "16777": "failure",
+    "16778": "failure",
+    "16779": "failure",
+    "16780": "failure",
+    "16781": "failure",
+    "16782": "failure",
+    "16783": "failure",
+    "16784": "successful",
+    "16785": "failure",
+    "16786": "failure",
+    "16787": "successful",
+    "16788": "failure",
+    "16789": "failure",
+    "16790": "successful",
+    "16791": "failure",
+    "16792": "failure",
+    "16793": "failure",
+    "16794": "failure",
+    "16795": "failure",
+    "16796": "failure",
+    "16797": "failure",
+    "16798": "successful",
+    "16799": "failure",
+    "16800": "failure",
+    "16801": "failure",
+    "16802": "failure",
+    "16803": "failure",
+    "16804": "successful",
+    "16805": "failure",
+    "16806": "failure",
+    "16807": "failure",
+    "16808": "failure",
+    "16809": "failure",
+    "16810": "failure",
+    "16811": "failure",
+    "16812": "failure",
+    "16813": "infeasible",
+    "16814": "failure",
+    "16815": "failure",
+    "16816": "successful",
+    "16817": "failure",
+    "16818": "failure",
+    "16819": "failure",
+    "16820": "failure",
+    "16821": "failure",
+    "16822": "successful",
+    "16823": "failure",
+    "16824": "failure",
+    "16825": "successful",
+    "16826": "successful",
+    "16827": "failure",
+    "16828": "failure",
+    "16829": "failure",
+    "16830": "failure",
+    "16831": "failure",
+    "16832": "failure",
+    "16833": "failure",
+    "16834": "failure",
+    "16835": "infeasible",
+    "16836": "failure",
+    "16837": "failure",
+    "16838": "failure",
+    "16839": "failure",
+    "16840": "failure",
+    "16841": "failure",
+    "16842": "failure",
+    "16843": "failure",
+    "16844": "failure",
+    "16845": "failure",
+    "16846": "failure",
+    "16847": "infeasible",
+    "16848": "successful",
+    "16849": "failure",
+    "16850": "failure",
+    "16851": "failure",
+    "16852": "failure",
+    "16853": "successful",
+    "16854": "successful",
+    "16855": "failure",
+    "16856": "failure",
+    "16857": "failure",
+    "16858": "failure",
+    "16859": "failure",
+    "16860": "failure",
+    "16861": "failure",
+    "16862": "failure",
+    "16863": "failure",
+    "16864": "failure",
+    "16865": "successful",
+    "16866": "failure",
+    "16867": "failure",
+    "16868": "failure",
+    "16869": "failure",
+    "16870": "failure",
+    "16871": "failure",
+    "16872": "failure",
+    "16873": "failure",
+    "16874": "failure",
+    "16875": "failure",
+    "16876": "failure",
+    "16877": "failure",
+    "16878": "infeasible",
+    "16879": "failure",
+    "16880": "failure",
+    "16881": "successful",
+    "16882": "failure",
+    "16883": "failure",
+    "16884": "infeasible",
+    "16885": "failure",
+    "16886": "failure",
+    "16887": "failure",
+    "16888": "infeasible",
+    "16889": "failure",
+    "16890": "successful",
+    "16891": "failure",
+    "16892": "successful",
+    "16893": "failure",
+    "16894": "failure",
+    "16895": "failure",
+    "16896": "failure",
+    "16897": "failure",
+    "16898": "failure",
+    "16899": "failure",
+    "16900": "failure",
+    "16901": "failure",
+    "16902": "failure",
+    "16903": "successful",
+    "16904": "failure",
+    "16905": "failure",
+    "16906": "failure",
+    "16907": "failure",
+    "16908": "failure",
+    "16909": "failure",
+    "16910": "failure",
+    "16911": "failure",
+    "16912": "failure",
+    "16913": "infeasible",
+    "16914": "successful",
+    "16915": "failure",
+    "16916": "failure",
+    "16917": "infeasible",
+    "16918": "failure",
+    "16919": "failure",
+    "16920": "failure",
+    "16921": "infeasible",
+    "16922": "failure",
+    "16923": "failure",
+    "16924": "failure",
+    "16925": "failure",
+    "16926": "failure",
+    "16927": "failure",
+    "16928": "failure",
+    "16929": "infeasible",
+    "16930": "failure",
+    "16931": "failure",
+    "16932": "failure",
+    "16933": "failure",
+    "16934": "infeasible",
+    "16935": "failure",
+    "16936": "failure",
+    "16937": "failure",
+    "16938": "failure",
+    "16939": "infeasible",
+    "16940": "failure",
+    "16941": "failure",
+    "16942": "failure",
+    "16943": "failure",
+    "16944": "failure",
+    "16945": "failure",
+    "16946": "failure",
+    "16947": "failure",
+    "16948": "failure",
+    "16949": "failure",
+    "16950": "failure",
+    "16951": "failure",
+    "16952": "failure",
+    "16953": "failure",
+    "16954": "failure",
+    "16955": "infeasible",
+    "16956": "failure",
+    "16957": "infeasible",
+    "16958": "successful",
+    "16959": "infeasible",
+    "16960": "failure",
+    "16961": "failure",
+    "16962": "failure",
+    "16963": "failure",
+    "16964": "failure",
+    "16965": "failure",
+    "16966": "failure",
+    "16967": "infeasible",
+    "16968": "failure",
+    "16969": "infeasible",
+    "16970": "successful",
+    "16971": "failure",
+    "16972": "failure",
+    "16973": "failure",
+    "16974": "failure",
+    "16975": "failure",
+    "16976": "failure",
+    "16977": "failure",
+    "16978": "failure",
+    "16979": "failure",
+    "16980": "failure",
+    "16981": "failure",
+    "16982": "failure",
+    "16983": "failure",
+    "16984": "successful",
+    "16985": "failure",
+    "16986": "failure",
+    "16987": "failure",
+    "16988": "infeasible",
+    "16989": "failure",
+    "16990": "successful",
+    "16991": "failure",
+    "16992": "failure",
+    "16993": "infeasible",
+    "16994": "failure",
+    "16995": "failure",
+    "16996": "failure",
+    "16997": "failure",
+    "16998": "failure",
+    "16999": "failure",
+    "17000": "failure",
+    "17001": "failure",
+    "17002": "failure",
+    "17003": "successful",
+    "17004": "failure",
+    "17005": "failure",
+    "17006": "failure",
+    "17007": "failure",
+    "17008": "failure",
+    "17009": "failure",
+    "17010": "failure",
+    "17011": "failure",
+    "17012": "failure",
+    "17013": "failure",
+    "17014": "failure",
+    "17015": "failure",
+    "17016": "failure",
+    "17017": "failure",
+    "17018": "failure",
+    "17019": "failure",
+    "17020": "failure",
+    "17021": "failure",
+    "17022": "failure",
+    "17023": "successful",
+    "17024": "failure",
+    "17025": "failure",
+    "17026": "failure",
+    "17027": "failure",
+    "17028": "failure",
+    "17029": "failure",
+    "17030": "failure",
+    "17031": "failure",
+    "17032": "failure",
+    "17033": "failure",
+    "17034": "failure",
+    "17035": "successful",
+    "17036": "failure",
+    "17037": "failure",
+    "17038": "failure",
+    "17039": "failure",
+    "17040": "failure",
+    "17041": "infeasible",
+    "17042": "failure",
+    "17043": "failure",
+    "17044": "failure",
+    "17045": "failure",
+    "17046": "failure",
+    "17047": "failure",
+    "17048": "failure",
+    "17049": "failure",
+    "17050": "failure",
+    "17051": "failure",
+    "17052": "failure",
+    "17053": "infeasible",
+    "17054": "failure",
+    "17055": "successful",
+    "17056": "failure",
+    "17057": "infeasible",
+    "17058": "failure",
+    "17059": "failure",
+    "17060": "failure",
+    "17061": "failure",
+    "17062": "failure",
+    "17063": "failure",
+    "17064": "failure",
+    "17065": "successful",
+    "17066": "failure",
+    "17067": "failure",
+    "17068": "successful",
+    "17069": "failure",
+    "17070": "successful",
+    "17071": "failure",
+    "17072": "failure",
+    "17073": "failure",
+    "17074": "successful",
+    "17075": "failure",
+    "17076": "failure",
+    "17077": "failure",
+    "17078": "infeasible",
+    "17079": "successful",
+    "17080": "failure",
+    "17081": "failure",
+    "17082": "failure",
+    "17083": "failure",
+    "17084": "failure",
+    "17085": "failure",
+    "17086": "failure",
+    "17087": "failure",
+    "17088": "failure",
+    "17089": "failure",
+    "17090": "infeasible",
+    "17091": "successful",
+    "17092": "failure",
+    "17093": "failure",
+    "17094": "failure",
+    "17095": "failure",
+    "17096": "failure",
+    "17097": "infeasible",
+    "17098": "failure",
+    "17099": "failure",
+    "17100": "failure",
+    "17101": "failure",
+    "17102": "successful",
+    "17103": "failure",
+    "17104": "failure",
+    "17105": "infeasible",
+    "17106": "failure",
+    "17107": "failure",
+    "17108": "infeasible",
+    "17109": "infeasible",
+    "17110": "failure",
+    "17111": "failure",
+    "17112": "failure",
+    "17113": "failure",
+    "17114": "failure",
+    "17115": "failure",
+    "17116": "successful",
+    "17117": "failure",
+    "17118": "failure",
+    "17119": "failure",
+    "17120": "failure",
+    "17121": "infeasible",
+    "17122": "failure",
+    "17123": "failure",
+    "17124": "failure",
+    "17125": "failure",
+    "17126": "failure",
+    "17127": "failure",
+    "17128": "infeasible",
+    "17129": "failure",
+    "17130": "failure",
+    "17131": "failure",
+    "17132": "failure",
+    "17133": "failure",
+    "17134": "successful",
+    "17135": "failure",
+    "17136": "failure",
+    "17137": "failure",
+    "17138": "successful",
+    "17139": "failure",
+    "17140": "failure",
+    "17141": "successful",
+    "17142": "failure",
+    "17143": "failure",
+    "17144": "failure",
+    "17145": "successful",
+    "17146": "failure",
+    "17147": "failure",
+    "17148": "failure",
+    "17149": "failure",
+    "17150": "failure",
+    "17151": "failure",
+    "17152": "failure",
+    "17153": "failure",
+    "17154": "successful",
+    "17155": "failure",
+    "17156": "failure",
+    "17157": "successful",
+    "17158": "failure",
+    "17159": "failure",
+    "17160": "failure",
+    "17161": "successful",
+    "17162": "failure",
+    "17163": "failure",
+    "17164": "failure",
+    "17165": "failure",
+    "17166": "failure",
+    "17167": "successful",
+    "17168": "failure",
+    "17169": "failure",
+    "17170": "failure",
+    "17171": "successful",
+    "17172": "failure",
+    "17173": "failure",
+    "17174": "successful",
+    "17175": "failure",
+    "17176": "failure",
+    "17177": "failure",
+    "17178": "failure",
+    "17179": "failure",
+    "17180": "failure",
+    "17181": "failure",
+    "17182": "failure",
+    "17183": "failure",
+    "17184": "successful",
+    "17185": "infeasible",
+    "17186": "failure",
+    "17187": "infeasible",
+    "17188": "failure",
+    "17189": "failure",
+    "17190": "infeasible",
+    "17191": "failure",
+    "17192": "failure",
+    "17193": "infeasible",
+    "17194": "infeasible",
+    "17195": "failure",
+    "17196": "failure",
+    "17197": "failure",
+    "17198": "failure",
+    "17199": "infeasible",
+    "17200": "failure",
+    "17201": "failure",
+    "17202": "infeasible",
+    "17203": "already_complete",
+    "17204": "infeasible",
+    "17205": "failure",
+    "17206": "failure",
+    "17207": "failure",
+    "17208": "failure",
+    "17209": "failure",
+    "17210": "failure",
+    "17211": "failure",
+    "17212": "failure",
+    "17213": "failure",
+    "17214": "failure",
+    "17215": "failure",
+    "17216": "failure",
+    "17217": "failure",
+    "17218": "failure",
+    "17219": "failure",
+    "17220": "successful",
+    "17221": "failure",
+    "17222": "failure",
+    "17223": "failure",
+    "17224": "failure",
+    "17225": "infeasible",
+    "17226": "failure",
+    "17227": "successful",
+    "17228": "successful",
+    "17229": "failure",
+    "17230": "failure",
+    "17231": "failure",
+    "17232": "infeasible",
+    "17233": "failure",
+    "17234": "infeasible",
+    "17235": "failure",
+    "17236": "failure",
+    "17237": "successful",
+    "17238": "failure",
+    "17239": "failure",
+    "17240": "failure",
+    "17241": "failure",
+    "17242": "failure",
+    "17243": "failure",
+    "17244": "failure",
+    "17245": "failure",
+    "17246": "failure",
+    "17247": "failure",
+    "17248": "failure",
+    "17249": "failure",
+    "17250": "failure",
+    "17251": "successful",
+    "17252": "failure",
+    "17253": "failure",
+    "17254": "failure",
+    "17255": "failure",
+    "17256": "successful",
+    "17257": "failure",
+    "17258": "failure",
+    "17259": "failure",
+    "17260": "failure",
+    "17261": "failure",
+    "17262": "failure",
+    "17263": "failure",
+    "17264": "failure",
+    "17265": "failure",
+    "17266": "successful",
+    "17267": "successful",
+    "17268": "failure",
+    "17269": "failure",
+    "17270": "failure",
+    "17271": "failure",
+    "17272": "failure",
+    "17273": "failure",
+    "17274": "failure",
+    "17275": "failure",
+    "17276": "failure",
+    "17277": "failure",
+    "17278": "failure",
+    "17279": "failure",
+    "17280": "failure",
+    "17281": "failure",
+    "17282": "failure",
+    "17283": "failure",
+    "17284": "failure",
+    "17285": "failure",
+    "17286": "successful",
+    "17287": "failure",
+    "17288": "failure",
+    "17289": "successful",
+    "17290": "failure",
+    "17291": "failure",
+    "17292": "failure",
+    "17293": "failure",
+    "17294": "failure",
+    "17295": "failure",
+    "17296": "successful",
+    "17297": "failure",
+    "17298": "failure",
+    "17299": "failure",
+    "17300": "failure",
+    "17301": "failure",
+    "17302": "failure",
+    "17303": "failure",
+    "17304": "failure",
+    "17305": "failure",
+    "17306": "failure",
+    "17307": "failure",
+    "17308": "failure",
+    "17309": "failure",
+    "17310": "failure",
+    "17311": "failure",
+    "17312": "infeasible",
+    "17313": "failure",
+    "17314": "failure",
+    "17315": "failure",
+    "17316": "successful",
+    "17317": "failure",
+    "17318": "failure",
+    "17319": "failure",
+    "17320": "failure",
+    "17321": "failure",
+    "17322": "failure",
+    "17323": "failure",
+    "17324": "infeasible",
+    "17325": "failure",
+    "17326": "infeasible",
+    "17327": "failure",
+    "17328": "failure",
+    "17329": "failure",
+    "17330": "failure",
+    "17331": "failure",
+    "17332": "successful",
+    "17333": "failure",
+    "17334": "failure",
+    "17335": "failure",
+    "17336": "failure",
+    "17337": "successful",
+    "17338": "infeasible",
+    "17339": "failure",
+    "17340": "failure",
+    "17341": "infeasible",
+    "17342": "failure",
+    "17343": "failure",
+    "17344": "failure",
+    "17345": "failure",
+    "17346": "failure",
+    "17347": "failure",
+    "17348": "failure",
+    "17349": "failure",
+    "17350": "failure",
+    "17351": "infeasible",
+    "17352": "failure",
+    "17353": "failure",
+    "17354": "failure",
+    "17355": "failure",
+    "17356": "failure",
+    "17357": "failure",
+    "17358": "failure",
+    "17359": "failure",
+    "17360": "failure",
+    "17361": "successful",
+    "17362": "failure",
+    "17363": "failure",
+    "17364": "failure",
+    "17365": "infeasible",
+    "17366": "failure",
+    "17367": "infeasible",
+    "17368": "failure",
+    "17369": "failure",
+    "17370": "failure",
+    "17371": "failure",
+    "17372": "failure",
+    "17373": "failure",
+    "17374": "failure",
+    "17375": "failure",
+    "17376": "failure",
+    "17377": "infeasible",
+    "17378": "failure",
+    "17379": "failure",
+    "17380": "failure",
+    "17381": "failure",
+    "17382": "successful",
+    "17383": "successful",
+    "17384": "failure",
+    "17385": "failure",
+    "17386": "failure",
+    "17387": "failure",
+    "17388": "failure",
+    "17389": "failure",
+    "17390": "failure",
+    "17391": "failure",
+    "17392": "failure",
+    "17393": "failure",
+    "17394": "already_complete",
+    "17395": "successful",
+    "17396": "failure",
+    "17397": "failure",
+    "17398": "infeasible",
+    "17399": "failure",
+    "17400": "failure",
+    "17401": "successful",
+    "17402": "failure",
+    "17403": "failure",
+    "17404": "failure",
+    "17405": "infeasible",
+    "17406": "failure",
+    "17407": "failure",
+    "17408": "failure",
+    "17409": "failure",
+    "17410": "failure",
+    "17411": "failure",
+    "17412": "failure",
+    "17413": "infeasible",
+    "17414": "successful",
+    "17415": "infeasible",
+    "17416": "failure",
+    "17417": "failure",
+    "17418": "failure",
+    "17419": "failure",
+    "17420": "failure",
+    "17421": "failure",
+    "17422": "infeasible",
+    "17423": "failure",
+    "17424": "failure",
+    "17425": "failure",
+    "17426": "failure",
+    "17427": "infeasible",
+    "17428": "failure",
+    "17429": "failure",
+    "17430": "failure",
+    "17431": "failure",
+    "17432": "failure",
+    "17433": "failure",
+    "17434": "failure",
+    "17435": "failure",
+    "17436": "failure",
+    "17437": "failure",
+    "17438": "failure",
+    "17439": "successful",
+    "17440": "failure",
+    "17441": "failure",
+    "17442": "failure",
+    "17443": "failure",
+    "17444": "failure",
+    "17445": "failure",
+    "17446": "failure",
+    "17447": "failure",
+    "17448": "failure",
+    "17449": "failure",
+    "17450": "failure",
+    "17451": "failure",
+    "17452": "failure",
+    "17453": "failure",
+    "17454": "successful",
+    "17455": "failure",
+    "17456": "failure",
+    "17457": "failure",
+    "17458": "successful",
+    "17459": "failure",
+    "17460": "successful",
+    "17461": "failure",
+    "17462": "failure",
+    "17463": "successful",
+    "17464": "failure",
+    "17465": "successful",
+    "17466": "failure",
+    "17467": "failure",
+    "17468": "failure",
+    "17469": "failure",
+    "17470": "failure",
+    "17471": "failure",
+    "17472": "failure",
+    "17473": "failure",
+    "17474": "failure",
+    "17475": "successful",
+    "17476": "failure",
+    "17477": "failure",
+    "17478": "failure",
+    "17479": "failure",
+    "17480": "already_complete",
+    "17481": "failure",
+    "17482": "failure",
+    "17483": "failure",
+    "17484": "failure",
+    "17485": "failure",
+    "17486": "failure",
+    "17487": "failure",
+    "17488": "failure",
+    "17489": "failure",
+    "17490": "infeasible",
+    "17491": "failure",
+    "17492": "successful",
+    "17493": "already_complete",
+    "17494": "failure",
+    "17495": "successful",
+    "17496": "infeasible",
+    "17497": "failure",
+    "17498": "failure",
+    "17499": "failure",
+    "17500": "failure",
+    "17501": "infeasible",
+    "17502": "failure",
+    "17503": "failure",
+    "17504": "successful",
+    "17505": "failure",
+    "17506": "failure",
+    "17507": "failure",
+    "17508": "failure",
+    "17509": "failure",
+    "17510": "failure",
+    "17511": "failure",
+    "17512": "failure",
+    "17513": "failure",
+    "17514": "failure",
+    "17515": "failure",
+    "17516": "successful",
+    "17517": "failure",
+    "17518": "failure",
+    "17519": "failure",
+    "17520": "failure",
+    "17521": "failure",
+    "17522": "successful",
+    "17523": "failure",
+    "17524": "infeasible",
+    "17525": "successful",
+    "17526": "infeasible",
+    "17527": "failure",
+    "17528": "successful",
+    "17529": "infeasible",
+    "17530": "failure",
+    "17531": "successful",
+    "17532": "failure",
+    "17533": "failure",
+    "17534": "failure",
+    "17535": "failure",
+    "17536": "failure",
+    "17537": "failure",
+    "17538": "failure",
+    "17539": "failure",
+    "17540": "failure",
+    "17541": "failure",
+    "17542": "infeasible",
+    "17543": "infeasible",
+    "17544": "failure",
+    "17545": "failure",
+    "17546": "failure",
+    "17547": "failure",
+    "17548": "failure",
+    "17549": "failure",
+    "17550": "failure",
+    "17551": "failure",
+    "17552": "successful",
+    "17553": "successful",
+    "17554": "successful",
+    "17555": "failure",
+    "17556": "failure",
+    "17557": "successful",
+    "17558": "failure",
+    "17559": "failure",
+    "17560": "failure",
+    "17561": "failure",
+    "17562": "failure",
+    "17563": "failure",
+    "17564": "failure",
+    "17565": "successful",
+    "17566": "failure",
+    "17567": "failure",
+    "17568": "successful",
+    "17569": "successful",
+    "17570": "failure",
+    "17571": "failure",
+    "17572": "infeasible",
+    "17573": "failure",
+    "17574": "failure",
+    "17575": "failure",
+    "17576": "failure",
+    "17577": "failure",
+    "17578": "failure",
+    "17579": "successful",
+    "17580": "failure",
+    "17581": "failure",
+    "17582": "failure",
+    "17583": "failure",
+    "17584": "failure",
+    "17585": "successful",
+    "17586": "failure",
+    "17587": "failure",
+    "17588": "successful",
+    "17589": "successful",
+    "17590": "failure",
+    "17591": "failure",
+    "17592": "failure",
+    "17593": "failure",
+    "17594": "failure",
+    "17595": "failure",
+    "17596": "infeasible",
+    "17597": "failure",
+    "17598": "failure",
+    "17599": "successful",
+    "17600": "infeasible",
+    "17601": "failure",
+    "17602": "failure",
+    "17603": "failure",
+    "17604": "failure",
+    "17605": "failure",
+    "17606": "failure",
+    "17607": "infeasible",
+    "17608": "failure",
+    "17609": "failure",
+    "17610": "failure",
+    "17611": "failure",
+    "17612": "failure",
+    "17613": "failure",
+    "17614": "failure",
+    "17615": "failure",
+    "17616": "failure",
+    "17617": "failure",
+    "17618": "successful",
+    "17619": "successful",
+    "17620": "failure",
+    "17621": "successful",
+    "17622": "failure",
+    "17623": "failure",
+    "17624": "failure",
+    "17625": "failure",
+    "17626": "infeasible",
+    "17627": "failure",
+    "17628": "failure",
+    "17629": "failure",
+    "17630": "successful",
+    "17631": "infeasible",
+    "17632": "failure",
+    "17633": "failure",
+    "17634": "failure",
+    "17635": "failure",
+    "17636": "infeasible",
+    "17637": "failure",
+    "17638": "failure",
+    "17639": "failure",
+    "17640": "failure",
+    "17641": "infeasible",
+    "17642": "failure",
+    "17643": "successful",
+    "17644": "failure",
+    "17645": "failure",
+    "17646": "failure",
+    "17647": "failure",
+    "17648": "failure",
+    "17649": "failure",
+    "17650": "failure",
+    "17651": "failure",
+    "17652": "failure",
+    "17653": "failure",
+    "17654": "failure",
+    "17655": "failure",
+    "17656": "failure",
+    "17657": "failure",
+    "17658": "failure",
+    "17659": "failure",
+    "17660": "failure",
+    "17661": "failure",
+    "17662": "failure",
+    "17663": "infeasible",
+    "17664": "failure",
+    "17665": "failure",
+    "17666": "infeasible",
+    "17667": "already_complete",
+    "17668": "failure",
+    "17669": "failure",
+    "17670": "successful",
+    "17671": "failure",
+    "17672": "failure",
+    "17673": "failure",
+    "17674": "already_complete",
+    "17675": "failure",
+    "17676": "successful",
+    "17677": "failure",
+    "17678": "failure",
+    "17679": "failure",
+    "17680": "failure",
+    "17681": "failure",
+    "17682": "infeasible",
+    "17683": "failure",
+    "17684": "failure",
+    "17685": "successful",
+    "17686": "failure",
+    "17687": "failure",
+    "17688": "failure",
+    "17689": "successful",
+    "17690": "failure",
+    "17691": "failure",
+    "17692": "failure",
+    "17693": "failure",
+    "17694": "failure",
+    "17695": "failure",
+    "17696": "failure",
+    "17697": "failure",
+    "17698": "infeasible",
+    "17699": "failure",
+    "17700": "infeasible",
+    "17701": "failure",
+    "17702": "failure",
+    "17703": "failure",
+    "17704": "successful",
+    "17705": "infeasible",
+    "17706": "failure",
+    "17707": "failure",
+    "17708": "failure",
+    "17709": "failure",
+    "17710": "successful",
+    "17711": "successful",
+    "17712": "failure",
+    "17713": "failure",
+    "17714": "failure",
+    "17715": "failure",
+    "17716": "infeasible",
+    "17717": "failure",
+    "17718": "failure",
+    "17719": "failure",
+    "17720": "failure",
+    "17721": "successful",
+    "17722": "infeasible",
+    "17723": "failure",
+    "17724": "infeasible",
+    "17725": "failure",
+    "17726": "failure",
+    "17727": "failure",
+    "17728": "failure",
+    "17729": "failure",
+    "17730": "failure",
+    "17731": "failure",
+    "17732": "successful",
+    "17733": "failure",
+    "17734": "failure",
+    "17735": "infeasible",
+    "17736": "failure",
+    "17737": "failure",
+    "17738": "failure",
+    "17739": "infeasible",
+    "17740": "failure",
+    "17741": "failure",
+    "17742": "failure",
+    "17743": "failure",
+    "17744": "failure",
+    "17745": "failure",
+    "17746": "failure",
+    "17747": "failure",
+    "17748": "failure",
+    "17749": "successful",
+    "17750": "failure",
+    "17751": "infeasible",
+    "17752": "failure",
+    "17753": "failure",
+    "17754": "failure",
+    "17755": "failure",
+    "17756": "failure",
+    "17757": "failure",
+    "17758": "failure",
+    "17759": "failure",
+    "17760": "failure",
+    "17761": "infeasible",
+    "17762": "failure",
+    "17763": "successful",
+    "17764": "failure",
+    "17765": "successful",
+    "17766": "failure",
+    "17767": "failure",
+    "17768": "failure",
+    "17769": "failure",
+    "17770": "successful",
+    "17771": "failure",
+    "17772": "failure",
+    "17773": "failure",
+    "17774": "failure",
+    "17775": "successful",
+    "17776": "successful",
+    "17777": "failure",
+    "17778": "failure",
+    "17779": "successful",
+    "17780": "successful",
+    "17781": "failure",
+    "17782": "infeasible",
+    "17783": "infeasible",
+    "17784": "failure",
+    "17785": "already_complete",
+    "17786": "failure",
+    "17787": "failure",
+    "17788": "failure",
+    "17789": "failure",
+    "17790": "successful",
+    "17791": "failure",
+    "17792": "failure",
+    "17793": "failure",
+    "17794": "failure",
+    "17795": "failure",
+    "17796": "failure",
+    "17797": "failure",
+    "17798": "failure",
+    "17799": "failure",
+    "17800": "failure",
+    "17801": "failure",
+    "17802": "failure",
+    "17803": "failure",
+    "17804": "failure",
+    "17805": "failure",
+    "17806": "failure",
+    "17807": "failure",
+    "17808": "failure",
+    "17809": "failure",
+    "17810": "failure",
+    "17811": "infeasible",
+    "17812": "failure",
+    "17813": "failure",
+    "17814": "failure",
+    "17815": "infeasible",
+    "17816": "failure",
+    "17817": "failure",
+    "17818": "failure",
+    "17819": "failure",
+    "17820": "successful",
+    "17821": "failure",
+    "17822": "failure",
+    "17823": "successful",
+    "17824": "failure",
+    "17825": "failure",
+    "17826": "failure",
+    "17827": "failure",
+    "17828": "failure",
+    "17829": "failure",
+    "17830": "infeasible",
+    "17831": "successful",
+    "17832": "failure",
+    "17833": "failure",
+    "17834": "failure",
+    "17835": "failure",
+    "17836": "successful",
+    "17837": "failure",
+    "17838": "failure",
+    "17839": "failure",
+    "17840": "failure",
+    "17841": "failure",
+    "17842": "failure",
+    "17843": "failure",
+    "17844": "failure",
+    "17845": "failure",
+    "17846": "failure",
+    "17847": "failure",
+    "17848": "failure",
+    "17849": "failure",
+    "17850": "failure",
+    "17851": "failure",
+    "17852": "successful",
+    "17853": "failure",
+    "17854": "failure",
+    "17855": "failure",
+    "17856": "failure",
+    "17857": "failure",
+    "17858": "already_complete",
+    "17859": "failure",
+    "17860": "failure",
+    "17861": "failure",
+    "17862": "failure",
+    "17863": "failure",
+    "17864": "infeasible",
+    "17865": "failure",
+    "17866": "failure",
+    "17867": "failure",
+    "17868": "successful",
+    "17869": "failure",
+    "17870": "failure",
+    "17871": "failure",
+    "17872": "failure",
+    "17873": "failure",
+    "17874": "failure",
+    "17875": "failure",
+    "17876": "failure",
+    "17877": "successful",
+    "17878": "infeasible",
+    "17879": "failure",
+    "17880": "failure",
+    "17881": "already_complete",
+    "17882": "failure",
+    "17883": "failure",
+    "17884": "failure",
+    "17885": "successful",
+    "17886": "failure",
+    "17887": "successful",
+    "17888": "failure",
+    "17889": "failure",
+    "17890": "failure",
+    "17891": "failure",
+    "17892": "successful",
+    "17893": "failure",
+    "17894": "successful",
+    "17895": "failure",
+    "17896": "failure",
+    "17897": "failure",
+    "17898": "failure",
+    "17899": "successful",
+    "17900": "failure",
+    "17901": "failure",
+    "17902": "failure",
+    "17903": "failure",
+    "17904": "failure",
+    "17905": "failure",
+    "17906": "failure",
+    "17907": "failure",
+    "17908": "failure",
+    "17909": "infeasible",
+    "17910": "infeasible",
+    "17911": "failure",
+    "17912": "failure",
+    "17913": "infeasible",
+    "17914": "failure",
+    "17915": "infeasible",
+    "17916": "failure",
+    "17917": "failure",
+    "17918": "infeasible",
+    "17919": "failure",
+    "17920": "failure",
+    "17921": "failure",
+    "17922": "failure",
+    "17923": "infeasible",
+    "17924": "infeasible",
+    "17925": "failure",
+    "17926": "failure",
+    "17927": "failure",
+    "17928": "failure",
+    "17929": "failure",
+    "17930": "successful",
+    "17931": "failure",
+    "17932": "failure",
+    "17933": "failure",
+    "17934": "failure",
+    "17935": "failure",
+    "17936": "failure",
+    "17937": "failure",
+    "17938": "failure",
+    "17939": "failure",
+    "17940": "failure",
+    "17941": "successful",
+    "17942": "failure",
+    "17943": "infeasible",
+    "17944": "failure",
+    "17945": "failure",
+    "17946": "failure",
+    "17947": "failure",
+    "17948": "failure",
+    "17949": "failure",
+    "17950": "failure",
+    "17951": "failure",
+    "17952": "successful",
+    "17953": "failure",
+    "17954": "failure",
+    "17955": "failure",
+    "17956": "failure",
+    "17957": "failure",
+    "17958": "failure",
+    "17959": "failure",
+    "17960": "failure",
+    "17961": "failure",
+    "17962": "failure",
+    "17963": "failure",
+    "17964": "failure",
+    "17965": "successful",
+    "17966": "failure",
+    "17967": "infeasible",
+    "17968": "infeasible",
+    "17969": "failure",
+    "17970": "failure",
+    "17971": "failure",
+    "17972": "failure",
+    "17973": "failure",
+    "17974": "failure",
+    "17975": "failure",
+    "17976": "failure",
+    "17977": "successful",
+    "17978": "failure",
+    "17979": "failure",
+    "17980": "failure",
+    "17981": "failure",
+    "17982": "failure",
+    "17983": "failure",
+    "17984": "failure",
+    "17985": "failure",
+    "17986": "failure",
+    "17987": "failure",
+    "17988": "failure",
+    "17989": "infeasible",
+    "17990": "failure",
+    "17991": "failure",
+    "17992": "failure",
+    "17993": "failure",
+    "17994": "failure",
+    "17995": "failure",
+    "17996": "successful",
+    "17997": "failure",
+    "17998": "failure",
+    "17999": "failure",
+    "18000": "failure",
+    "18001": "failure",
+    "18002": "failure",
+    "18003": "failure",
+    "18004": "successful",
+    "18005": "failure",
+    "18006": "successful",
+    "18007": "failure",
+    "18008": "failure",
+    "18009": "failure",
+    "18010": "failure",
+    "18011": "failure",
+    "18012": "failure",
+    "18013": "failure",
+    "18014": "failure",
+    "18015": "failure",
+    "18016": "failure",
+    "18017": "successful",
+    "18018": "failure",
+    "18019": "successful",
+    "18020": "failure",
+    "18021": "failure",
+    "18022": "failure",
+    "18023": "failure",
+    "18024": "failure",
+    "18025": "successful",
+    "18026": "failure",
+    "18027": "failure",
+    "18028": "failure",
+    "18029": "failure",
+    "18030": "failure",
+    "18031": "failure",
+    "18032": "failure",
+    "18033": "failure",
+    "18034": "failure",
+    "18035": "failure",
+    "18036": "failure",
+    "18037": "failure",
+    "18038": "failure",
+    "18039": "failure",
+    "18040": "failure",
+    "18041": "failure",
+    "18042": "failure",
+    "18043": "failure",
+    "18044": "failure",
+    "18045": "failure",
+    "18046": "failure",
+    "18047": "failure",
+    "18048": "failure",
+    "18049": "failure",
+    "18050": "failure",
+    "18051": "failure",
+    "18052": "successful",
+    "18053": "failure",
+    "18054": "failure",
+    "18055": "failure",
+    "18056": "infeasible",
+    "18057": "failure",
+    "18058": "failure",
+    "18059": "failure",
+    "18060": "failure",
+    "18061": "failure",
+    "18062": "infeasible",
+    "18063": "failure",
+    "18064": "infeasible",
+    "18065": "failure",
+    "18066": "failure",
+    "18067": "failure",
+    "18068": "failure",
+    "18069": "successful",
+    "18070": "failure",
+    "18071": "failure",
+    "18072": "failure",
+    "18073": "failure",
+    "18074": "infeasible",
+    "18075": "failure",
+    "18076": "already_complete",
+    "18077": "failure",
+    "18078": "failure",
+    "18079": "failure",
+    "18080": "failure",
+    "18081": "failure",
+    "18082": "failure",
+    "18083": "failure",
+    "18084": "failure",
+    "18085": "failure",
+    "18086": "failure",
+    "18087": "failure",
+    "18088": "successful",
+    "18089": "failure",
+    "18090": "failure",
+    "18091": "failure",
+    "18092": "failure",
+    "18093": "failure",
+    "18094": "failure",
+    "18095": "successful",
+    "18096": "failure",
+    "18097": "failure",
+    "18098": "failure",
+    "18099": "infeasible",
+    "18100": "failure",
+    "18101": "successful",
+    "18102": "failure",
+    "18103": "failure",
+    "18104": "failure",
+    "18105": "failure",
+    "18106": "infeasible",
+    "18107": "failure",
+    "18108": "failure",
+    "18109": "successful",
+    "18110": "failure",
+    "18111": "already_complete",
+    "18112": "infeasible",
+    "18113": "successful",
+    "18114": "failure",
+    "18115": "infeasible",
+    "18116": "infeasible",
+    "18117": "failure",
+    "18118": "failure",
+    "18119": "failure",
+    "18120": "failure",
+    "18121": "failure",
+    "18122": "failure",
+    "18123": "failure",
+    "18124": "failure",
+    "18125": "failure",
+    "18126": "failure",
+    "18127": "failure",
+    "18128": "failure",
+    "18129": "failure",
+    "18130": "failure",
+    "18131": "failure",
+    "18132": "failure",
+    "18133": "infeasible",
+    "18134": "failure",
+    "18135": "failure",
+    "18136": "failure",
+    "18137": "successful",
+    "18138": "already_complete",
+    "18139": "successful",
+    "18140": "successful",
+    "18141": "failure",
+    "18142": "failure",
+    "18143": "failure",
+    "18144": "failure",
+    "18145": "failure",
+    "18146": "failure",
+    "18147": "failure",
+    "18148": "failure",
+    "18149": "failure",
+    "18150": "infeasible",
+    "18151": "infeasible",
+    "18152": "failure",
+    "18153": "failure",
+    "18154": "failure",
+    "18155": "failure",
+    "18156": "failure",
+    "18157": "already_complete",
+    "18158": "failure",
+    "18159": "failure",
+    "18160": "failure",
+    "18161": "failure",
+    "18162": "successful",
+    "18163": "failure",
+    "18164": "failure",
+    "18165": "failure",
+    "18166": "failure",
+    "18167": "successful",
+    "18168": "failure",
+    "18169": "failure",
+    "18170": "failure",
+    "18171": "failure",
+    "18172": "failure",
+    "18173": "successful",
+    "18174": "failure",
+    "18175": "already_complete",
+    "18176": "successful",
+    "18177": "successful",
+    "18178": "successful",
+    "18179": "failure",
+    "18180": "failure",
+    "18181": "failure",
+    "18182": "successful",
+    "18183": "infeasible",
+    "18184": "failure",
+    "18185": "failure",
+    "18186": "failure",
+    "18187": "successful",
+    "18188": "failure",
+    "18189": "failure",
+    "18190": "failure",
+    "18191": "failure",
+    "18192": "failure",
+    "18193": "failure",
+    "18194": "failure",
+    "18195": "failure",
+    "18196": "failure",
+    "18197": "failure",
+    "18198": "successful",
+    "18199": "failure",
+    "18200": "failure",
+    "18201": "failure",
+    "18202": "failure",
+    "18203": "successful",
+    "18204": "failure",
+    "18205": "infeasible",
+    "18206": "failure",
+    "18207": "failure",
+    "18208": "failure",
+    "18209": "failure",
+    "18210": "failure",
+    "18211": "failure",
+    "18212": "failure",
+    "18213": "failure",
+    "18214": "failure",
+    "18215": "failure",
+    "18216": "failure",
+    "18217": "successful",
+    "18218": "failure",
+    "18219": "successful",
+    "18220": "successful",
+    "18221": "already_complete",
+    "18222": "successful",
+    "18223": "failure",
+    "18224": "failure",
+    "18225": "infeasible",
+    "18226": "failure",
+    "18227": "successful",
+    "18228": "failure",
+    "18229": "already_complete",
+    "18230": "failure",
+    "18231": "failure",
+    "18232": "failure",
+    "18233": "failure",
+    "18234": "failure",
+    "18235": "successful",
+    "18236": "failure",
+    "18237": "failure",
+    "18238": "successful",
+    "18239": "failure",
+    "18240": "infeasible",
+    "18241": "failure",
+    "18242": "failure",
+    "18243": "failure",
+    "18244": "failure",
+    "18245": "failure",
+    "18246": "failure",
+    "18247": "failure",
+    "18248": "failure",
+    "18249": "failure",
+    "18250": "failure",
+    "18251": "successful",
+    "18252": "failure",
+    "18253": "failure",
+    "18254": "failure",
+    "18255": "failure",
+    "18256": "failure",
+    "18257": "failure",
+    "18258": "failure",
+    "18259": "failure",
+    "18260": "failure",
+    "18261": "failure",
+    "18262": "infeasible",
+    "18263": "failure",
+    "18264": "failure",
+    "18265": "failure",
+    "18266": "successful",
+    "18267": "failure",
+    "18268": "failure",
+    "18269": "failure",
+    "18270": "failure",
+    "18271": "failure",
+    "18272": "failure",
+    "18273": "failure",
+    "18274": "failure",
+    "18275": "failure",
+    "18276": "failure",
+    "18277": "failure",
+    "18278": "infeasible",
+    "18279": "failure",
+    "18280": "failure",
+    "18281": "infeasible",
+    "18282": "failure",
+    "18283": "failure",
+    "18284": "failure",
+    "18285": "failure",
+    "18286": "successful",
+    "18287": "failure",
+    "18288": "failure",
+    "18289": "failure",
+    "18290": "failure",
+    "18291": "failure",
+    "18292": "failure",
+    "18293": "failure",
+    "18294": "failure",
+    "18295": "failure",
+    "18296": "failure",
+    "18297": "failure",
+    "18298": "failure",
+    "18299": "infeasible",
+    "18300": "failure",
+    "18301": "infeasible",
+    "18302": "failure",
+    "18303": "infeasible",
+    "18304": "failure",
+    "18305": "failure",
+    "18306": "successful",
+    "18307": "failure",
+    "18308": "successful",
+    "18309": "failure",
+    "18310": "failure",
+    "18311": "failure",
+    "18312": "failure",
+    "18313": "failure",
+    "18314": "infeasible",
+    "18315": "failure",
+    "18316": "failure",
+    "18317": "successful",
+    "18318": "failure",
+    "18319": "failure",
+    "18320": "infeasible",
+    "18321": "successful",
+    "18322": "failure",
+    "18323": "failure",
+    "18324": "failure",
+    "18325": "failure",
+    "18326": "failure",
+    "18327": "failure",
+    "18328": "failure",
+    "18329": "failure",
+    "18330": "failure",
+    "18331": "successful",
+    "18332": "infeasible",
+    "18333": "successful",
+    "18334": "failure",
+    "18335": "failure",
+    "18336": "failure",
+    "18337": "failure",
+    "18338": "successful",
+    "18339": "infeasible",
+    "18340": "failure",
+    "18341": "failure",
+    "18342": "failure",
+    "18343": "successful",
+    "18344": "failure",
+    "18345": "failure",
+    "18346": "failure",
+    "18347": "failure",
+    "18348": "failure",
+    "18349": "failure",
+    "18350": "failure",
+    "18351": "failure",
+    "18352": "failure",
+    "18353": "failure",
+    "18354": "failure",
+    "18355": "infeasible",
+    "18356": "failure",
+    "18357": "failure",
+    "18358": "infeasible",
+    "18359": "failure",
+    "18360": "failure",
+    "18361": "failure",
+    "18362": "failure",
+    "18363": "failure",
+    "18364": "failure",
+    "18365": "successful",
+    "18366": "failure",
+    "18367": "failure",
+    "18368": "failure",
+    "18369": "infeasible",
+    "18370": "failure",
+    "18371": "failure",
+    "18372": "failure",
+    "18373": "failure",
+    "18374": "failure",
+    "18375": "failure",
+    "18376": "failure",
+    "18377": "failure",
+    "18378": "failure",
+    "18379": "failure",
+    "18380": "successful",
+    "18381": "successful",
+    "18382": "failure",
+    "18383": "failure",
+    "18384": "failure",
+    "18385": "failure",
+    "18386": "failure",
+    "18387": "failure",
+    "18388": "infeasible",
+    "18389": "failure",
+    "18390": "infeasible",
+    "18391": "failure",
+    "18392": "successful",
+    "18393": "failure",
+    "18394": "failure",
+    "18395": "infeasible",
+    "18396": "failure",
+    "18397": "failure",
+    "18398": "failure",
+    "18399": "failure",
+    "18400": "failure",
+    "18401": "failure",
+    "18402": "failure",
+    "18403": "failure",
+    "18404": "failure",
+    "18405": "failure",
+    "18406": "infeasible",
+    "18407": "failure",
+    "18408": "failure",
+    "18409": "failure",
+    "18410": "failure",
+    "18411": "infeasible",
+    "18412": "successful",
+    "18413": "failure",
+    "18414": "failure",
+    "18415": "failure",
+    "18416": "failure",
+    "18417": "failure",
+    "18418": "failure",
+    "18419": "failure",
+    "18420": "failure",
+    "18421": "failure",
+    "18422": "failure",
+    "18423": "failure",
+    "18424": "failure",
+    "18425": "failure",
+    "18426": "infeasible",
+    "18427": "failure",
+    "18428": "infeasible",
+    "18429": "infeasible",
+    "18430": "successful",
+    "18431": "failure",
+    "18432": "failure",
+    "18433": "failure",
+    "18434": "infeasible",
+    "18435": "failure",
+    "18436": "failure",
+    "18437": "failure",
+    "18438": "failure",
+    "18439": "successful",
+    "18440": "failure",
+    "18441": "infeasible",
+    "18442": "failure",
+    "18443": "failure",
+    "18444": "failure",
+    "18445": "failure",
+    "18446": "failure",
+    "18447": "failure",
+    "18448": "failure",
+    "18449": "failure",
+    "18450": "failure",
+    "18451": "failure",
+    "18452": "failure",
+    "18453": "failure",
+    "18454": "failure",
+    "18455": "failure",
+    "18456": "failure",
+    "18457": "infeasible",
+    "18458": "failure",
+    "18459": "failure",
+    "18460": "failure",
+    "18461": "failure",
+    "18462": "failure",
+    "18463": "failure",
+    "18464": "infeasible",
+    "18465": "infeasible",
+    "18466": "failure",
+    "18467": "failure",
+    "18468": "infeasible",
+    "18469": "failure",
+    "18470": "already_complete",
+    "18471": "failure",
+    "18472": "failure",
+    "18473": "failure",
+    "18474": "failure",
+    "18475": "failure",
+    "18476": "infeasible",
+    "18477": "failure",
+    "18478": "failure",
+    "18479": "failure",
+    "18480": "failure",
+    "18481": "failure",
+    "18482": "failure",
+    "18483": "infeasible",
+    "18484": "successful",
+    "18485": "failure",
+    "18486": "successful",
+    "18487": "failure",
+    "18488": "failure",
+    "18489": "infeasible",
+    "18490": "failure",
+    "18491": "failure",
+    "18492": "failure",
+    "18493": "failure",
+    "18494": "failure",
+    "18495": "successful",
+    "18496": "infeasible",
+    "18497": "successful",
+    "18498": "failure",
+    "18499": "failure",
+    "18500": "failure",
+    "18501": "failure",
+    "18502": "failure",
+    "18503": "failure",
+    "18504": "failure",
+    "18505": "failure",
+    "18506": "failure",
+    "18507": "failure",
+    "18508": "failure",
+    "18509": "successful",
+    "18510": "failure",
+    "18511": "failure",
+    "18512": "failure",
+    "18513": "successful",
+    "18514": "failure",
+    "18515": "failure",
+    "18516": "failure",
+    "18517": "failure",
+    "18518": "failure",
+    "18519": "failure",
+    "18520": "failure",
+    "18521": "infeasible",
+    "18522": "failure",
+    "18523": "failure",
+    "18524": "failure",
+    "18525": "successful",
+    "18526": "failure",
+    "18527": "failure",
+    "18528": "failure",
+    "18529": "failure",
+    "18530": "failure",
+    "18531": "failure",
+    "18532": "failure",
+    "18533": "failure",
+    "18534": "failure",
+    "18535": "failure",
+    "18536": "infeasible",
+    "18537": "failure",
+    "18538": "failure",
+    "18539": "failure",
+    "18540": "failure",
+    "18541": "failure",
+    "18542": "infeasible",
+    "18543": "infeasible",
+    "18544": "failure",
+    "18545": "successful",
+    "18546": "failure",
+    "18547": "failure",
+    "18548": "failure",
+    "18549": "failure",
+    "18550": "successful",
+    "18551": "failure",
+    "18552": "failure",
+    "18553": "failure",
+    "18554": "failure",
+    "18555": "failure",
+    "18556": "failure",
+    "18557": "failure",
+    "18558": "failure",
+    "18559": "failure",
+    "18560": "failure",
+    "18561": "failure",
+    "18562": "failure",
+    "18563": "failure",
+    "18564": "infeasible",
+    "18565": "failure",
+    "18566": "failure",
+    "18567": "failure",
+    "18568": "infeasible",
+    "18569": "failure",
+    "18570": "failure",
+    "18571": "failure",
+    "18572": "successful",
+    "18573": "failure",
+    "18574": "successful",
+    "18575": "failure",
+    "18576": "failure",
+    "18577": "failure",
+    "18578": "successful",
+    "18579": "failure",
+    "18580": "failure",
+    "18581": "infeasible",
+    "18582": "failure",
+    "18583": "failure",
+    "18584": "failure",
+    "18585": "failure",
+    "18586": "failure",
+    "18587": "successful",
+    "18588": "infeasible",
+    "18589": "failure",
+    "18590": "failure",
+    "18591": "failure",
+    "18592": "failure",
+    "18593": "failure",
+    "18594": "successful",
+    "18595": "infeasible",
+    "18596": "failure",
+    "18597": "failure",
+    "18598": "successful",
+    "18599": "failure",
+    "18600": "failure",
+    "18601": "failure",
+    "18602": "failure",
+    "18603": "failure",
+    "18604": "failure",
+    "18605": "failure",
+    "18606": "failure",
+    "18607": "failure",
+    "18608": "failure",
+    "18609": "infeasible",
+    "18610": "failure",
+    "18611": "failure",
+    "18612": "failure",
+    "18613": "infeasible",
+    "18614": "failure",
+    "18615": "failure",
+    "18616": "successful",
+    "18617": "failure",
+    "18618": "failure",
+    "18619": "failure",
+    "18620": "successful",
+    "18621": "successful",
+    "18622": "failure",
+    "18623": "successful",
+    "18624": "infeasible",
+    "18625": "successful",
+    "18626": "failure",
+    "18627": "failure",
+    "18628": "failure",
+    "18629": "failure",
+    "18630": "failure",
+    "18631": "failure",
+    "18632": "failure",
+    "18633": "failure",
+    "18634": "failure",
+    "18635": "failure",
+    "18636": "failure",
+    "18637": "failure",
+    "18638": "failure",
+    "18639": "failure",
+    "18640": "failure",
+    "18641": "successful",
+    "18642": "failure",
+    "18643": "successful",
+    "18644": "failure",
+    "18645": "failure",
+    "18646": "failure",
+    "18647": "successful",
+    "18648": "failure",
+    "18649": "failure",
+    "18650": "failure",
+    "18651": "successful",
+    "18652": "failure",
+    "18653": "successful",
+    "18654": "failure",
+    "18655": "failure",
+    "18656": "infeasible",
+    "18657": "infeasible",
+    "18658": "failure",
+    "18659": "failure",
+    "18660": "failure",
+    "18661": "failure",
+    "18662": "failure",
+    "18663": "failure",
+    "18664": "successful",
+    "18665": "failure",
+    "18666": "failure",
+    "18667": "infeasible",
+    "18668": "failure",
+    "18669": "failure",
+    "18670": "failure",
+    "18671": "failure",
+    "18672": "failure",
+    "18673": "failure",
+    "18674": "failure",
+    "18675": "successful",
+    "18676": "failure",
+    "18677": "failure",
+    "18678": "successful",
+    "18679": "failure",
+    "18680": "failure",
+    "18681": "failure",
+    "18682": "failure",
+    "18683": "failure",
+    "18684": "failure",
+    "18685": "failure",
+    "18686": "failure",
+    "18687": "infeasible",
+    "18688": "failure",
+    "18689": "failure",
+    "18690": "failure",
+    "18691": "infeasible",
+    "18692": "failure",
+    "18693": "failure",
+    "18694": "infeasible",
+    "18695": "successful",
+    "18696": "failure",
+    "18697": "successful",
+    "18698": "failure",
+    "18699": "infeasible",
+    "18700": "failure",
+    "18701": "successful",
+    "18702": "failure",
+    "18703": "failure",
+    "18704": "successful",
+    "18705": "failure",
+    "18706": "infeasible",
+    "18707": "failure",
+    "18708": "failure",
+    "18709": "failure",
+    "18710": "failure",
+    "18711": "successful",
+    "18712": "failure",
+    "18713": "failure",
+    "18714": "failure",
+    "18715": "failure",
+    "18716": "failure",
+    "18717": "successful",
+    "18718": "successful",
+    "18719": "successful",
+    "18720": "failure",
+    "18721": "successful",
+    "18722": "failure",
+    "18723": "failure",
+    "18724": "failure",
+    "18725": "failure",
+    "18726": "failure",
+    "18727": "failure",
+    "18728": "successful",
+    "18729": "failure",
+    "18730": "failure",
+    "18731": "failure",
+    "18732": "failure",
+    "18733": "failure",
+    "18734": "successful",
+    "18735": "failure",
+    "18736": "failure",
+    "18737": "failure",
+    "18738": "failure",
+    "18739": "failure",
+    "18740": "infeasible",
+    "18741": "successful",
+    "18742": "infeasible",
+    "18743": "failure",
+    "18744": "failure",
+    "18745": "failure",
+    "18746": "failure",
+    "18747": "failure",
+    "18748": "failure",
+    "18749": "failure",
+    "18750": "successful",
+    "18751": "failure",
+    "18752": "failure",
+    "18753": "successful",
+    "18754": "infeasible",
+    "18755": "failure",
+    "18756": "failure",
+    "18757": "failure",
+    "18758": "failure",
+    "18759": "failure",
+    "18760": "failure",
+    "18761": "failure",
+    "18762": "infeasible",
+    "18763": "successful",
+    "18764": "failure",
+    "18765": "failure",
+    "18766": "failure",
+    "18767": "failure",
+    "18768": "failure",
+    "18769": "failure",
+    "18770": "failure",
+    "18771": "failure",
+    "18772": "successful",
+    "18773": "failure",
+    "18774": "failure",
+    "18775": "already_complete",
+    "18776": "failure",
+    "18777": "failure",
+    "18778": "failure",
+    "18779": "infeasible",
+    "18780": "failure",
+    "18781": "failure",
+    "18782": "failure",
+    "18783": "failure",
+    "18784": "failure",
+    "18785": "failure",
+    "18786": "infeasible",
+    "18787": "failure",
+    "18788": "failure",
+    "18789": "failure",
+    "18790": "failure",
+    "18791": "failure",
+    "18792": "failure",
+    "18793": "failure",
+    "18794": "infeasible",
+    "18795": "failure",
+    "18796": "successful",
+    "18797": "failure",
+    "18798": "failure",
+    "18799": "failure",
+    "18800": "successful",
+    "18801": "failure",
+    "18802": "failure",
+    "18803": "infeasible",
+    "18804": "successful",
+    "18805": "failure",
+    "18806": "failure",
+    "18807": "failure",
+    "18808": "successful",
+    "18809": "failure",
+    "18810": "failure",
+    "18811": "failure",
+    "18812": "failure",
+    "18813": "failure",
+    "18814": "failure",
+    "18815": "failure",
+    "18816": "failure",
+    "18817": "failure",
+    "18818": "successful",
+    "18819": "failure",
+    "18820": "failure",
+    "18821": "failure",
+    "18822": "infeasible",
+    "18823": "infeasible",
+    "18824": "failure",
+    "18825": "failure",
+    "18826": "infeasible",
+    "18827": "failure",
+    "18828": "successful",
+    "18829": "failure",
+    "18830": "failure",
+    "18831": "failure",
+    "18832": "failure",
+    "18833": "failure",
+    "18834": "failure",
+    "18835": "failure",
+    "18836": "already_complete",
+    "18837": "failure",
+    "18838": "failure",
+    "18839": "infeasible",
+    "18840": "failure",
+    "18841": "infeasible",
+    "18842": "failure",
+    "18843": "failure",
+    "18844": "failure",
+    "18845": "failure",
+    "18846": "failure",
+    "18847": "failure",
+    "18848": "successful",
+    "18849": "failure",
+    "18850": "failure",
+    "18851": "failure",
+    "18852": "failure",
+    "18853": "successful",
+    "18854": "failure",
+    "18855": "failure",
+    "18856": "failure",
+    "18857": "failure",
+    "18858": "infeasible",
+    "18859": "failure",
+    "18860": "failure",
+    "18861": "failure",
+    "18862": "successful",
+    "18863": "failure",
+    "18864": "failure",
+    "18865": "failure",
+    "18866": "failure",
+    "18867": "failure",
+    "18868": "failure",
+    "18869": "infeasible",
+    "18870": "failure",
+    "18871": "failure",
+    "18872": "failure",
+    "18873": "failure",
+    "18874": "failure",
+    "18875": "failure",
+    "18876": "failure",
+    "18877": "successful",
+    "18878": "failure",
+    "18879": "successful",
+    "18880": "failure",
+    "18881": "failure",
+    "18882": "successful",
+    "18883": "infeasible",
+    "18884": "failure",
+    "18885": "failure",
+    "18886": "failure",
+    "18887": "failure",
+    "18888": "failure",
+    "18889": "failure",
+    "18890": "failure",
+    "18891": "infeasible",
+    "18892": "failure",
+    "18893": "failure",
+    "18894": "successful",
+    "18895": "failure",
+    "18896": "failure",
+    "18897": "failure",
+    "18898": "failure",
+    "18899": "failure",
+    "18900": "failure",
+    "18901": "failure",
+    "18902": "failure",
+    "18903": "failure",
+    "18904": "failure",
+    "18905": "failure",
+    "18906": "failure",
+    "18907": "failure",
+    "18908": "failure",
+    "18909": "failure",
+    "18910": "failure",
+    "18911": "failure",
+    "18912": "failure",
+    "18913": "infeasible",
+    "18914": "failure",
+    "18915": "failure",
+    "18916": "failure",
+    "18917": "infeasible",
+    "18918": "failure",
+    "18919": "infeasible",
+    "18920": "failure",
+    "18921": "failure",
+    "18922": "failure",
+    "18923": "infeasible",
+    "18924": "infeasible",
+    "18925": "failure",
+    "18926": "failure",
+    "18927": "failure",
+    "18928": "failure",
+    "18929": "failure",
+    "18930": "failure",
+    "18931": "failure",
+    "18932": "failure",
+    "18933": "successful",
+    "18934": "failure",
+    "18935": "failure",
+    "18936": "failure",
+    "18937": "failure",
+    "18938": "failure",
+    "18939": "failure",
+    "18940": "failure",
+    "18941": "failure",
+    "18942": "failure",
+    "18943": "successful",
+    "18944": "failure",
+    "18945": "failure",
+    "18946": "infeasible",
+    "18947": "failure",
+    "18948": "infeasible",
+    "18949": "failure",
+    "18950": "failure",
+    "18951": "failure",
+    "18952": "infeasible",
+    "18953": "failure",
+    "18954": "failure",
+    "18955": "failure",
+    "18956": "failure",
+    "18957": "failure",
+    "18958": "failure",
+    "18959": "failure",
+    "18960": "successful",
+    "18961": "failure",
+    "18962": "failure",
+    "18963": "failure",
+    "18964": "infeasible",
+    "18965": "infeasible",
+    "18966": "failure",
+    "18967": "failure",
+    "18968": "failure",
+    "18969": "failure",
+    "18970": "failure",
+    "18971": "failure",
+    "18972": "failure",
+    "18973": "failure",
+    "18974": "failure",
+    "18975": "failure",
+    "18976": "failure",
+    "18977": "failure",
+    "18978": "failure",
+    "18979": "infeasible",
+    "18980": "infeasible",
+    "18981": "failure",
+    "18982": "failure",
+    "18983": "failure",
+    "18984": "infeasible",
+    "18985": "failure",
+    "18986": "failure",
+    "18987": "failure",
+    "18988": "failure",
+    "18989": "failure",
+    "18990": "failure",
+    "18991": "failure",
+    "18992": "failure",
+    "18993": "failure",
+    "18994": "failure",
+    "18995": "failure",
+    "18996": "already_complete",
+    "18997": "failure",
+    "18998": "infeasible",
+    "18999": "successful",
+    "19000": "failure",
+    "19001": "failure",
+    "19002": "failure",
+    "19003": "failure",
+    "19004": "failure",
+    "19005": "failure",
+    "19006": "already_complete",
+    "19007": "failure",
+    "19008": "failure",
+    "19009": "successful",
+    "19010": "successful",
+    "19011": "failure",
+    "19012": "failure",
+    "19013": "infeasible",
+    "19014": "failure",
+    "19015": "failure",
+    "19016": "failure",
+    "19017": "failure",
+    "19018": "failure",
+    "19019": "failure",
+    "19020": "failure",
+    "19021": "failure",
+    "19022": "failure",
+    "19023": "successful",
+    "19024": "failure",
+    "19025": "successful",
+    "19026": "successful",
+    "19027": "failure",
+    "19028": "failure",
+    "19029": "failure",
+    "19030": "failure",
+    "19031": "failure",
+    "19032": "failure",
+    "19033": "failure",
+    "19034": "failure",
+    "19035": "successful",
+    "19036": "failure",
+    "19037": "successful",
+    "19038": "failure",
+    "19039": "successful",
+    "19040": "failure",
+    "19041": "failure",
+    "19042": "infeasible",
+    "19043": "failure",
+    "19044": "failure",
+    "19045": "successful",
+    "19046": "failure",
+    "19047": "failure",
+    "19048": "failure",
+    "19049": "failure",
+    "19050": "failure",
+    "19051": "failure",
+    "19052": "failure",
+    "19053": "failure",
+    "19054": "failure",
+    "19055": "failure",
+    "19056": "failure",
+    "19057": "failure",
+    "19058": "failure",
+    "19059": "failure",
+    "19060": "successful",
+    "19061": "failure",
+    "19062": "failure",
+    "19063": "failure",
+    "19064": "failure",
+    "19065": "failure",
+    "19066": "failure",
+    "19067": "failure",
+    "19068": "failure",
+    "19069": "failure",
+    "19070": "failure",
+    "19071": "failure",
+    "19072": "successful",
+    "19073": "failure",
+    "19074": "failure",
+    "19075": "failure",
+    "19076": "failure",
+    "19077": "failure",
+    "19078": "failure",
+    "19079": "failure",
+    "19080": "failure",
+    "19081": "failure",
+    "19082": "failure",
+    "19083": "failure",
+    "19084": "failure",
+    "19085": "failure",
+    "19086": "successful",
+    "19087": "successful",
+    "19088": "failure",
+    "19089": "failure",
+    "19090": "failure",
+    "19091": "failure",
+    "19092": "failure",
+    "19093": "failure",
+    "19094": "failure",
+    "19095": "failure",
+    "19096": "failure",
+    "19097": "failure",
+    "19098": "infeasible",
+    "19099": "failure",
+    "19100": "failure",
+    "19101": "infeasible",
+    "19102": "failure",
+    "19103": "failure",
+    "19104": "failure",
+    "19105": "failure",
+    "19106": "successful",
+    "19107": "successful",
+    "19108": "failure",
+    "19109": "failure",
+    "19110": "failure",
+    "19111": "failure",
+    "19112": "infeasible",
+    "19113": "successful",
+    "19114": "failure",
+    "19115": "infeasible",
+    "19116": "failure",
+    "19117": "failure",
+    "19118": "failure",
+    "19119": "failure",
+    "19120": "infeasible",
+    "19121": "failure",
+    "19122": "failure",
+    "19123": "failure",
+    "19124": "failure",
+    "19125": "failure",
+    "19126": "failure",
+    "19127": "failure",
+    "19128": "failure",
+    "19129": "failure",
+    "19130": "failure",
+    "19131": "failure",
+    "19132": "failure",
+    "19133": "failure",
+    "19134": "failure",
+    "19135": "successful",
+    "19136": "successful",
+    "19137": "failure",
+    "19138": "failure",
+    "19139": "failure",
+    "19140": "failure",
+    "19141": "failure",
+    "19142": "failure",
+    "19143": "successful",
+    "19144": "successful",
+    "19145": "failure",
+    "19146": "infeasible",
+    "19147": "failure",
+    "19148": "failure",
+    "19149": "failure",
+    "19150": "failure",
+    "19151": "failure",
+    "19152": "failure",
+    "19153": "successful",
+    "19154": "failure",
+    "19155": "failure",
+    "19156": "failure",
+    "19157": "failure",
+    "19158": "successful",
+    "19159": "infeasible",
+    "19160": "failure",
+    "19161": "infeasible",
+    "19162": "failure",
+    "19163": "successful",
+    "19164": "failure",
+    "19165": "failure",
+    "19166": "failure",
+    "19167": "infeasible",
+    "19168": "failure",
+    "19169": "failure",
+    "19170": "failure",
+    "19171": "failure",
+    "19172": "failure",
+    "19173": "infeasible",
+    "19174": "failure",
+    "19175": "failure",
+    "19176": "failure",
+    "19177": "failure",
+    "19178": "failure",
+    "19179": "failure",
+    "19180": "failure",
+    "19181": "failure",
+    "19182": "failure",
+    "19183": "failure",
+    "19184": "failure",
+    "19185": "failure",
+    "19186": "failure",
+    "19187": "successful",
+    "19188": "failure",
+    "19189": "failure",
+    "19190": "successful",
+    "19191": "failure",
+    "19192": "failure",
+    "19193": "failure",
+    "19194": "failure",
+    "19195": "failure",
+    "19196": "failure",
+    "19197": "failure",
+    "19198": "failure",
+    "19199": "failure",
+    "19200": "infeasible",
+    "19201": "failure",
+    "19202": "failure",
+    "19203": "failure",
+    "19204": "infeasible",
+    "19205": "failure",
+    "19206": "failure",
+    "19207": "successful",
+    "19208": "failure",
+    "19209": "failure",
+    "19210": "failure",
+    "19211": "failure",
+    "19212": "failure",
+    "19213": "infeasible",
+    "19214": "failure",
+    "19215": "failure",
+    "19216": "failure",
+    "19217": "infeasible",
+    "19218": "successful",
+    "19219": "failure",
+    "19220": "successful",
+    "19221": "failure",
+    "19222": "failure",
+    "19223": "successful",
+    "19224": "infeasible",
+    "19225": "failure",
+    "19226": "failure",
+    "19227": "failure",
+    "19228": "successful",
+    "19229": "failure",
+    "19230": "failure",
+    "19231": "failure",
+    "19232": "successful",
+    "19233": "failure",
+    "19234": "failure",
+    "19235": "failure",
+    "19236": "failure",
+    "19237": "failure",
+    "19238": "infeasible",
+    "19239": "failure",
+    "19240": "failure",
+    "19241": "failure",
+    "19242": "successful",
+    "19243": "failure",
+    "19244": "failure",
+    "19245": "failure",
+    "19246": "infeasible",
+    "19247": "failure",
+    "19248": "failure",
+    "19249": "failure",
+    "19250": "failure",
+    "19251": "failure",
+    "19252": "failure",
+    "19253": "failure",
+    "19254": "failure",
+    "19255": "infeasible",
+    "19256": "failure",
+    "19257": "failure",
+    "19258": "failure",
+    "19259": "failure",
+    "19260": "failure",
+    "19261": "failure",
+    "19262": "failure",
+    "19263": "failure",
+    "19264": "successful",
+    "19265": "failure",
+    "19266": "failure",
+    "19267": "failure",
+    "19268": "failure",
+    "19269": "failure",
+    "19270": "failure",
+    "19271": "failure",
+    "19272": "successful",
+    "19273": "failure",
+    "19274": "failure",
+    "19275": "failure",
+    "19276": "failure",
+    "19277": "failure",
+    "19278": "failure",
+    "19279": "failure",
+    "19280": "failure",
+    "19281": "failure",
+    "19282": "failure",
+    "19283": "successful",
+    "19284": "failure",
+    "19285": "successful",
+    "19286": "failure",
+    "19287": "failure",
+    "19288": "failure",
+    "19289": "failure",
+    "19290": "successful",
+    "19291": "failure",
+    "19292": "failure",
+    "19293": "failure",
+    "19294": "failure",
+    "19295": "failure",
+    "19296": "failure",
+    "19297": "failure",
+    "19298": "failure",
+    "19299": "failure",
+    "19300": "failure",
+    "19301": "failure",
+    "19302": "failure",
+    "19303": "failure",
+    "19304": "failure",
+    "19305": "infeasible",
+    "19306": "failure",
+    "19307": "failure",
+    "19308": "failure",
+    "19309": "infeasible",
+    "19310": "infeasible",
+    "19311": "failure",
+    "19312": "failure",
+    "19313": "failure",
+    "19314": "failure",
+    "19315": "failure",
+    "19316": "failure",
+    "19317": "failure",
+    "19318": "failure",
+    "19319": "failure",
+    "19320": "failure",
+    "19321": "failure",
+    "19322": "failure",
+    "19323": "failure",
+    "19324": "failure",
+    "19325": "failure",
+    "19326": "failure",
+    "19327": "failure",
+    "19328": "infeasible",
+    "19329": "infeasible",
+    "19330": "failure",
+    "19331": "successful",
+    "19332": "infeasible",
+    "19333": "failure",
+    "19334": "failure",
+    "19335": "failure",
+    "19336": "failure",
+    "19337": "failure",
+    "19338": "failure",
+    "19339": "failure",
+    "19340": "failure",
+    "19341": "failure",
+    "19342": "failure",
+    "19343": "infeasible",
+    "19344": "failure",
+    "19345": "failure",
+    "19346": "infeasible",
+    "19347": "failure",
+    "19348": "failure",
+    "19349": "failure",
+    "19350": "failure",
+    "19351": "successful",
+    "19352": "failure",
+    "19353": "infeasible",
+    "19354": "failure",
+    "19355": "infeasible",
+    "19356": "failure",
+    "19357": "failure",
+    "19358": "failure",
+    "19359": "failure",
+    "19360": "failure",
+    "19361": "infeasible",
+    "19362": "successful",
+    "19363": "successful",
+    "19364": "failure",
+    "19365": "infeasible",
+    "19366": "infeasible",
+    "19367": "successful",
+    "19368": "successful",
+    "19369": "failure",
+    "19370": "failure",
+    "19371": "failure",
+    "19372": "failure",
+    "19373": "failure",
+    "19374": "failure",
+    "19375": "successful",
+    "19376": "failure",
+    "19377": "failure",
+    "19378": "failure",
+    "19379": "failure",
+    "19380": "failure",
+    "19381": "failure",
+    "19382": "failure",
+    "19383": "failure",
+    "19384": "successful",
+    "19385": "failure",
+    "19386": "failure",
+    "19387": "failure",
+    "19388": "failure",
+    "19389": "failure",
+    "19390": "failure",
+    "19391": "failure",
+    "19392": "failure",
+    "19393": "failure",
+    "19394": "failure",
+    "19395": "successful",
+    "19396": "failure",
+    "19397": "failure",
+    "19398": "failure",
+    "19399": "failure",
+    "19400": "failure",
+    "19401": "failure",
+    "19402": "failure",
+    "19403": "successful",
+    "19404": "failure",
+    "19405": "failure",
+    "19406": "failure",
+    "19407": "failure",
+    "19408": "failure",
+    "19409": "failure",
+    "19410": "failure",
+    "19411": "infeasible",
+    "19412": "failure",
+    "19413": "failure",
+    "19414": "failure",
+    "19415": "failure",
+    "19416": "infeasible",
+    "19417": "failure",
+    "19418": "failure",
+    "19419": "infeasible",
+    "19420": "failure",
+    "19421": "failure",
+    "19422": "failure",
+    "19423": "failure",
+    "19424": "failure",
+    "19425": "failure",
+    "19426": "successful",
+    "19427": "failure",
+    "19428": "successful",
+    "19429": "successful",
+    "19430": "failure",
+    "19431": "failure",
+    "19432": "failure",
+    "19433": "failure",
+    "19434": "failure",
+    "19435": "failure",
+    "19436": "failure",
+    "19437": "failure",
+    "19438": "successful",
+    "19439": "failure",
+    "19440": "failure",
+    "19441": "infeasible",
+    "19442": "failure",
+    "19443": "failure",
+    "19444": "infeasible",
+    "19445": "failure",
+    "19446": "failure",
+    "19447": "failure",
+    "19448": "infeasible",
+    "19449": "infeasible",
+    "19450": "failure",
+    "19451": "failure",
+    "19452": "failure",
+    "19453": "failure",
+    "19454": "successful",
+    "19455": "failure",
+    "19456": "failure",
+    "19457": "failure",
+    "19458": "failure",
+    "19459": "failure",
+    "19460": "successful",
+    "19461": "failure",
+    "19462": "failure",
+    "19463": "failure",
+    "19464": "failure",
+    "19465": "failure",
+    "19466": "failure",
+    "19467": "failure",
+    "19468": "failure",
+    "19469": "failure",
+    "19470": "failure",
+    "19471": "failure",
+    "19472": "failure",
+    "19473": "failure",
+    "19474": "failure",
+    "19475": "failure",
+    "19476": "failure",
+    "19477": "failure",
+    "19478": "failure",
+    "19479": "failure",
+    "19480": "failure",
+    "19481": "failure",
+    "19482": "failure",
+    "19483": "successful",
+    "19484": "failure",
+    "19485": "failure",
+    "19486": "failure",
+    "19487": "failure",
+    "19488": "failure",
+    "19489": "failure",
+    "19490": "successful",
+    "19491": "infeasible",
+    "19492": "failure",
+    "19493": "failure",
+    "19494": "failure",
+    "19495": "failure",
+    "19496": "failure",
+    "19497": "failure",
+    "19498": "failure",
+    "19499": "failure",
+    "19500": "successful",
+    "19501": "failure",
+    "19502": "failure",
+    "19503": "successful",
+    "19504": "failure",
+    "19505": "failure",
+    "19506": "failure",
+    "19507": "failure",
+    "19508": "failure",
+    "19509": "failure",
+    "19510": "infeasible",
+    "19511": "failure",
+    "19512": "failure",
+    "19513": "failure",
+    "19514": "failure",
+    "19515": "failure",
+    "19516": "failure",
+    "19517": "failure",
+    "19518": "failure",
+    "19519": "failure",
+    "19520": "infeasible",
+    "19521": "failure",
+    "19522": "failure",
+    "19523": "failure",
+    "19524": "infeasible",
+    "19525": "failure",
+    "19526": "infeasible",
+    "19527": "failure",
+    "19528": "failure",
+    "19529": "failure",
+    "19530": "failure",
+    "19531": "failure",
+    "19532": "failure",
+    "19533": "failure",
+    "19534": "failure",
+    "19535": "failure",
+    "19536": "successful",
+    "19537": "failure",
+    "19538": "failure",
+    "19539": "failure",
+    "19540": "infeasible",
+    "19541": "failure",
+    "19542": "failure",
+    "19543": "failure",
+    "19544": "failure",
+    "19545": "failure",
+    "19546": "failure",
+    "19547": "infeasible",
+    "19548": "failure",
+    "19549": "failure",
+    "19550": "failure",
+    "19551": "failure",
+    "19552": "failure",
+    "19553": "failure",
+    "19554": "failure",
+    "19555": "failure",
+    "19556": "failure",
+    "19557": "failure",
+    "19558": "failure",
+    "19559": "failure",
+    "19560": "failure",
+    "19561": "failure",
+    "19562": "failure",
+    "19563": "failure",
+    "19564": "infeasible",
+    "19565": "failure",
+    "19566": "failure",
+    "19567": "failure",
+    "19568": "successful",
+    "19569": "failure",
+    "19570": "infeasible",
+    "19571": "successful",
+    "19572": "failure",
+    "19573": "failure",
+    "19574": "failure",
+    "19575": "failure",
+    "19576": "failure",
+    "19577": "failure",
+    "19578": "failure",
+    "19579": "failure",
+    "19580": "failure",
+    "19581": "failure",
+    "19582": "failure",
+    "19583": "failure",
+    "19584": "failure",
+    "19585": "failure",
+    "19586": "successful",
+    "19587": "failure",
+    "19588": "failure",
+    "19589": "failure",
+    "19590": "failure",
+    "19591": "successful",
+    "19592": "failure",
+    "19593": "successful",
+    "19594": "failure",
+    "19595": "successful",
+    "19596": "failure",
+    "19597": "failure",
+    "19598": "failure",
+    "19599": "failure",
+    "19600": "failure",
+    "19601": "failure",
+    "19602": "failure",
+    "19603": "failure",
+    "19604": "failure",
+    "19605": "failure",
+    "19606": "successful",
+    "19607": "failure",
+    "19608": "failure",
+    "19609": "failure",
+    "19610": "failure",
+    "19611": "failure",
+    "19612": "failure",
+    "19613": "failure",
+    "19614": "failure",
+    "19615": "failure",
+    "19616": "failure",
+    "19617": "failure",
+    "19618": "failure",
+    "19619": "failure",
+    "19620": "failure",
+    "19621": "successful",
+    "19622": "failure",
+    "19623": "infeasible",
+    "19624": "successful",
+    "19625": "infeasible",
+    "19626": "successful",
+    "19627": "infeasible",
+    "19628": "successful",
+    "19629": "failure",
+    "19630": "failure",
+    "19631": "failure",
+    "19632": "infeasible",
+    "19633": "infeasible",
+    "19634": "failure",
+    "19635": "failure",
+    "19636": "failure",
+    "19637": "infeasible",
+    "19638": "failure",
+    "19639": "failure",
+    "19640": "failure",
+    "19641": "failure",
+    "19642": "failure",
+    "19643": "failure",
+    "19644": "failure",
+    "19645": "failure",
+    "19646": "failure",
+    "19647": "infeasible",
+    "19648": "infeasible",
+    "19649": "failure",
+    "19650": "failure",
+    "19651": "failure",
+    "19652": "failure",
+    "19653": "failure",
+    "19654": "failure",
+    "19655": "failure",
+    "19656": "failure",
+    "19657": "infeasible",
+    "19658": "failure",
+    "19659": "failure",
+    "19660": "infeasible",
+    "19661": "failure",
+    "19662": "successful",
+    "19663": "failure",
+    "19664": "failure",
+    "19665": "failure",
+    "19666": "failure",
+    "19667": "failure",
+    "19668": "failure",
+    "19669": "successful",
+    "19670": "failure",
+    "19671": "failure",
+    "19672": "failure",
+    "19673": "failure",
+    "19674": "failure",
+    "19675": "failure",
+    "19676": "failure",
+    "19677": "successful",
+    "19678": "failure",
+    "19679": "failure",
+    "19680": "failure",
+    "19681": "failure",
+    "19682": "failure",
+    "19683": "failure",
+    "19684": "infeasible",
+    "19685": "failure",
+    "19686": "failure",
+    "19687": "failure",
+    "19688": "failure",
+    "19689": "failure",
+    "19690": "failure",
+    "19691": "failure",
+    "19692": "failure",
+    "19693": "failure",
+    "19694": "successful",
+    "19695": "failure",
+    "19696": "failure",
+    "19697": "failure",
+    "19698": "failure",
+    "19699": "failure",
+    "19700": "failure",
+    "19701": "failure",
+    "19702": "failure",
+    "19703": "failure",
+    "19704": "failure",
+    "19705": "failure",
+    "19706": "failure",
+    "19707": "failure",
+    "19708": "failure",
+    "19709": "failure",
+    "19710": "failure",
+    "19711": "failure",
+    "19712": "infeasible",
+    "19713": "failure",
+    "19714": "failure",
+    "19715": "successful",
+    "19716": "successful",
+    "19717": "failure",
+    "19718": "failure",
+    "19719": "infeasible",
+    "19720": "failure",
+    "19721": "failure",
+    "19722": "infeasible",
+    "19723": "failure",
+    "19724": "failure",
+    "19725": "failure",
+    "19726": "failure",
+    "19727": "successful",
+    "19728": "failure",
+    "19729": "successful",
+    "19730": "failure",
+    "19731": "failure",
+    "19732": "failure",
+    "19733": "failure",
+    "19734": "infeasible",
+    "19735": "failure",
+    "19736": "failure",
+    "19737": "failure",
+    "19738": "failure",
+    "19739": "infeasible",
+    "19740": "failure",
+    "19741": "failure",
+    "19742": "failure",
+    "19743": "failure",
+    "19744": "failure",
+    "19745": "failure",
+    "19746": "failure",
+    "19747": "failure",
+    "19748": "successful",
+    "19749": "failure",
+    "19750": "failure",
+    "19751": "failure",
+    "19752": "failure",
+    "19753": "failure",
+    "19754": "successful",
+    "19755": "failure",
+    "19756": "failure",
+    "19757": "failure",
+    "19758": "failure",
+    "19759": "infeasible",
+    "19760": "failure",
+    "19761": "failure",
+    "19762": "failure",
+    "19763": "failure",
+    "19764": "failure",
+    "19765": "failure",
+    "19766": "infeasible",
+    "19767": "failure",
+    "19768": "infeasible",
+    "19769": "failure",
+    "19770": "failure",
+    "19771": "infeasible",
+    "19772": "failure",
+    "19773": "failure",
+    "19774": "successful",
+    "19775": "failure",
+    "19776": "failure",
+    "19777": "infeasible",
+    "19778": "failure",
+    "19779": "failure",
+    "19780": "failure",
+    "19781": "infeasible",
+    "19782": "failure",
+    "19783": "successful",
+    "19784": "failure",
+    "19785": "failure",
+    "19786": "infeasible",
+    "19787": "failure",
+    "19788": "failure",
+    "19789": "failure",
+    "19790": "failure",
+    "19791": "failure",
+    "19792": "failure",
+    "19793": "failure",
+    "19794": "failure",
+    "19795": "failure",
+    "19796": "failure",
+    "19797": "failure",
+    "19798": "failure",
+    "19799": "failure",
+    "19800": "infeasible",
+    "19801": "failure",
+    "19802": "failure",
+    "19803": "failure",
+    "19804": "infeasible",
+    "19805": "successful",
+    "19806": "failure",
+    "19807": "failure",
+    "19808": "failure",
+    "19809": "failure",
+    "19810": "failure",
+    "19811": "infeasible",
+    "19812": "failure",
+    "19813": "failure",
+    "19814": "failure",
+    "19815": "failure",
+    "19816": "infeasible",
+    "19817": "failure",
+    "19818": "failure",
+    "19819": "infeasible",
+    "19820": "failure",
+    "19821": "failure",
+    "19822": "failure",
+    "19823": "infeasible",
+    "19824": "failure",
+    "19825": "failure",
+    "19826": "failure",
+    "19827": "failure",
+    "19828": "failure",
+    "19829": "failure",
+    "19830": "failure",
+    "19831": "failure",
+    "19832": "failure",
+    "19833": "failure",
+    "19834": "failure",
+    "19835": "infeasible",
+    "19836": "failure",
+    "19837": "failure",
+    "19838": "failure",
+    "19839": "failure",
+    "19840": "infeasible",
+    "19841": "infeasible",
+    "19842": "failure",
+    "19843": "failure",
+    "19844": "failure",
+    "19845": "failure",
+    "19846": "failure",
+    "19847": "failure",
+    "19848": "failure",
+    "19849": "infeasible",
+    "19850": "failure",
+    "19851": "failure",
+    "19852": "failure",
+    "19853": "failure",
+    "19854": "failure",
+    "19855": "successful",
+    "19856": "failure",
+    "19857": "failure",
+    "19858": "successful",
+    "19859": "failure",
+    "19860": "failure",
+    "19861": "failure",
+    "19862": "failure",
+    "19863": "failure",
+    "19864": "failure",
+    "19865": "failure",
+    "19866": "failure",
+    "19867": "failure",
+    "19868": "failure",
+    "19869": "failure",
+    "19870": "failure",
+    "19871": "failure",
+    "19872": "infeasible",
+    "19873": "successful",
+    "19874": "successful",
+    "19875": "successful",
+    "19876": "failure",
+    "19877": "successful",
+    "19878": "failure",
+    "19879": "failure",
+    "19880": "failure",
+    "19881": "failure",
+    "19882": "failure",
+    "19883": "failure",
+    "19884": "failure",
+    "19885": "failure",
+    "19886": "failure",
+    "19887": "infeasible",
+    "19888": "failure",
+    "19889": "failure",
+    "19890": "failure",
+    "19891": "failure",
+    "19892": "failure",
+    "19893": "failure",
+    "19894": "failure",
+    "19895": "failure",
+    "19896": "failure",
+    "19897": "failure",
+    "19898": "failure",
+    "19899": "failure",
+    "19900": "successful",
+    "19901": "failure",
+    "19902": "failure",
+    "19903": "infeasible",
+    "19904": "failure",
+    "19905": "failure",
+    "19906": "failure",
+    "19907": "failure",
+    "19908": "successful",
+    "19909": "failure",
+    "19910": "infeasible",
+    "19911": "failure",
+    "19912": "successful",
+    "19913": "failure",
+    "19914": "failure",
+    "19915": "failure",
+    "19916": "failure",
+    "19917": "infeasible",
+    "19918": "failure",
+    "19919": "successful",
+    "19920": "failure",
+    "19921": "failure",
+    "19922": "successful",
+    "19923": "failure",
+    "19924": "failure",
+    "19925": "failure",
+    "19926": "failure",
+    "19927": "failure",
+    "19928": "infeasible",
+    "19929": "failure",
+    "19930": "failure",
+    "19931": "failure",
+    "19932": "failure",
+    "19933": "successful",
+    "19934": "failure",
+    "19935": "infeasible",
+    "19936": "failure",
+    "19937": "failure",
+    "19938": "successful",
+    "19939": "failure",
+    "19940": "failure",
+    "19941": "failure",
+    "19942": "failure",
+    "19943": "failure",
+    "19944": "failure",
+    "19945": "failure",
+    "19946": "successful",
+    "19947": "failure",
+    "19948": "failure",
+    "19949": "failure",
+    "19950": "successful",
+    "19951": "failure",
+    "19952": "successful",
+    "19953": "failure",
+    "19954": "failure",
+    "19955": "infeasible",
+    "19956": "failure",
+    "19957": "failure",
+    "19958": "failure",
+    "19959": "failure",
+    "19960": "failure",
+    "19961": "failure",
+    "19962": "failure",
+    "19963": "failure",
+    "19964": "failure",
+    "19965": "failure",
+    "19966": "failure",
+    "19967": "failure",
+    "19968": "failure",
+    "19969": "failure",
+    "19970": "successful",
+    "19971": "failure",
+    "19972": "successful",
+    "19973": "failure",
+    "19974": "failure",
+    "19975": "failure",
+    "19976": "failure",
+    "19977": "failure",
+    "19978": "failure",
+    "19979": "failure",
+    "19980": "failure",
+    "19981": "failure",
+    "19982": "failure",
+    "19983": "failure",
+    "19984": "failure",
+    "19985": "failure",
+    "19986": "failure",
+    "19987": "successful",
+    "19988": "failure",
+    "19989": "failure",
+    "19990": "infeasible",
+    "19991": "failure",
+    "19992": "failure",
+    "19993": "infeasible",
+    "19994": "failure",
+    "19995": "failure",
+    "19996": "failure",
+    "19997": "successful",
+    "19998": "failure",
+    "19999": "failure",
+    "20000": "failure",
+    "20001": "failure",
+    "20002": "failure",
+    "20003": "failure",
+    "20004": "failure",
+    "20005": "failure",
+    "20006": "failure",
+    "20007": "failure",
+    "20008": "failure",
+    "20009": "failure",
+    "20010": "failure",
+    "20011": "infeasible",
+    "20012": "failure",
+    "20013": "failure",
+    "20014": "failure",
+    "20015": "failure",
+    "20016": "already_complete",
+    "20017": "failure",
+    "20018": "failure",
+    "20019": "infeasible",
+    "20020": "failure",
+    "20021": "failure",
+    "20022": "failure",
+    "20023": "failure",
+    "20024": "failure",
+    "20025": "successful",
+    "20026": "failure",
+    "20027": "failure",
+    "20028": "failure",
+    "20029": "failure",
+    "20030": "failure",
+    "20031": "failure",
+    "20032": "successful",
+    "20033": "failure",
+    "20034": "failure",
+    "20035": "failure",
+    "20036": "failure",
+    "20037": "failure",
+    "20038": "failure",
+    "20039": "successful",
+    "20040": "successful",
+    "20041": "failure",
+    "20042": "failure",
+    "20043": "failure",
+    "20044": "failure",
+    "20045": "successful",
+    "20046": "failure",
+    "20047": "failure",
+    "20048": "failure",
+    "20049": "failure",
+    "20050": "failure",
+    "20051": "infeasible",
+    "20052": "failure",
+    "20053": "failure",
+    "20054": "failure",
+    "20055": "infeasible",
+    "20056": "failure",
+    "20057": "failure",
+    "20058": "failure",
+    "20059": "failure",
+    "20060": "failure",
+    "20061": "infeasible",
+    "20062": "failure",
+    "20063": "failure",
+    "20064": "failure",
+    "20065": "failure",
+    "20066": "failure",
+    "20067": "failure",
+    "20068": "failure",
+    "20069": "failure",
+    "20070": "failure",
+    "20071": "failure",
+    "20072": "failure",
+    "20073": "failure",
+    "20074": "failure",
+    "20075": "failure",
+    "20076": "failure",
+    "20077": "failure",
+    "20078": "failure",
+    "20079": "failure",
+    "20080": "failure",
+    "20081": "failure",
+    "20082": "failure",
+    "20083": "infeasible",
+    "20084": "infeasible",
+    "20085": "failure",
+    "20086": "failure",
+    "20087": "failure",
+    "20088": "successful",
+    "20089": "failure",
+    "20090": "failure",
+    "20091": "failure",
+    "20092": "failure",
+    "20093": "failure",
+    "20094": "failure",
+    "20095": "failure",
+    "20096": "failure",
+    "20097": "failure",
+    "20098": "failure",
+    "20099": "failure",
+    "20100": "failure",
+    "20101": "successful",
+    "20102": "failure",
+    "20103": "successful",
+    "20104": "failure",
+    "20105": "failure",
+    "20106": "failure",
+    "20107": "failure",
+    "20108": "failure",
+    "20109": "successful",
+    "20110": "failure",
+    "20111": "successful",
+    "20112": "failure",
+    "20113": "failure",
+    "20114": "failure",
+    "20115": "infeasible",
+    "20116": "failure",
+    "20117": "failure",
+    "20118": "failure",
+    "20119": "failure",
+    "20120": "failure",
+    "20121": "failure",
+    "20122": "infeasible",
+    "20123": "failure",
+    "20124": "infeasible",
+    "20125": "failure",
+    "20126": "failure",
+    "20127": "failure",
+    "20128": "successful",
+    "20129": "successful",
+    "20130": "failure",
+    "20131": "failure",
+    "20132": "failure",
+    "20133": "failure",
+    "20134": "failure",
+    "20135": "failure",
+    "20136": "successful",
+    "20137": "failure",
+    "20138": "failure",
+    "20139": "failure",
+    "20140": "failure",
+    "20141": "failure",
+    "20142": "infeasible",
+    "20143": "infeasible",
+    "20144": "failure",
+    "20145": "failure",
+    "20146": "failure",
+    "20147": "failure",
+    "20148": "failure",
+    "20149": "failure",
+    "20150": "failure",
+    "20151": "failure",
+    "20152": "infeasible",
+    "20153": "failure",
+    "20154": "failure",
+    "20155": "failure",
+    "20156": "failure",
+    "20157": "failure",
+    "20158": "failure",
+    "20159": "failure",
+    "20160": "failure",
+    "20161": "failure",
+    "20162": "successful",
+    "20163": "failure",
+    "20164": "infeasible",
+    "20165": "successful",
+    "20166": "failure",
+    "20167": "failure",
+    "20168": "failure",
+    "20169": "failure",
+    "20170": "infeasible",
+    "20171": "failure",
+    "20172": "failure",
+    "20173": "successful",
+    "20174": "failure",
+    "20175": "infeasible",
+    "20176": "infeasible",
+    "20177": "infeasible",
+    "20178": "infeasible",
+    "20179": "infeasible",
+    "20180": "failure",
+    "20181": "failure",
+    "20182": "failure",
+    "20183": "infeasible",
+    "20184": "failure",
+    "20185": "failure",
+    "20186": "failure",
+    "20187": "failure",
+    "20188": "failure",
+    "20189": "infeasible",
+    "20190": "failure",
+    "20191": "failure",
+    "20192": "failure",
+    "20193": "failure",
+    "20194": "infeasible",
+    "20195": "failure",
+    "20196": "infeasible",
+    "20197": "failure",
+    "20198": "failure",
+    "20199": "failure",
+    "20200": "successful",
+    "20201": "failure",
+    "20202": "failure",
+    "20203": "failure",
+    "20204": "failure",
+    "20205": "failure",
+    "20206": "failure",
+    "20207": "failure",
+    "20208": "infeasible",
+    "20209": "successful",
+    "20210": "infeasible",
+    "20211": "failure",
+    "20212": "successful",
+    "20213": "successful",
+    "20214": "successful",
+    "20215": "failure",
+    "20216": "failure",
+    "20217": "failure",
+    "20218": "failure",
+    "20219": "successful",
+    "20220": "failure",
+    "20221": "failure",
+    "20222": "failure",
+    "20223": "failure",
+    "20224": "failure",
+    "20225": "failure",
+    "20226": "failure",
+    "20227": "failure",
+    "20228": "failure",
+    "20229": "infeasible",
+    "20230": "successful",
+    "20231": "successful",
+    "20232": "successful",
+    "20233": "failure",
+    "20234": "successful",
+    "20235": "failure",
+    "20236": "failure",
+    "20237": "failure",
+    "20238": "failure",
+    "20239": "failure",
+    "20240": "successful",
+    "20241": "failure",
+    "20242": "failure",
+    "20243": "successful",
+    "20244": "failure",
+    "20245": "infeasible",
+    "20246": "failure",
+    "20247": "failure",
+    "20248": "failure",
+    "20249": "failure",
+    "20250": "failure",
+    "20251": "failure",
+    "20252": "failure",
+    "20253": "successful",
+    "20254": "infeasible",
+    "20255": "failure",
+    "20256": "failure",
+    "20257": "failure",
+    "20258": "successful",
+    "20259": "failure",
+    "20260": "infeasible",
+    "20261": "failure",
+    "20262": "failure",
+    "20263": "failure",
+    "20264": "failure",
+    "20265": "failure",
+    "20266": "failure",
+    "20267": "failure",
+    "20268": "failure",
+    "20269": "failure",
+    "20270": "successful",
+    "20271": "failure",
+    "20272": "failure",
+    "20273": "failure",
+    "20274": "failure",
+    "20275": "failure",
+    "20276": "failure",
+    "20277": "failure",
+    "20278": "infeasible",
+    "20279": "failure",
+    "20280": "infeasible",
+    "20281": "successful",
+    "20282": "failure",
+    "20283": "failure",
+    "20284": "failure",
+    "20285": "failure",
+    "20286": "successful",
+    "20287": "failure",
+    "20288": "failure",
+    "20289": "failure",
+    "20290": "failure",
+    "20291": "failure",
+    "20292": "failure",
+    "20293": "failure",
+    "20294": "failure",
+    "20295": "failure",
+    "20296": "failure",
+    "20297": "failure",
+    "20298": "failure",
+    "20299": "successful",
+    "20300": "failure",
+    "20301": "failure",
+    "20302": "failure",
+    "20303": "failure",
+    "20304": "failure",
+    "20305": "already_complete",
+    "20306": "infeasible",
+    "20307": "failure",
+    "20308": "failure",
+    "20309": "failure",
+    "20310": "successful",
+    "20311": "failure",
+    "20312": "successful",
+    "20313": "failure",
+    "20314": "failure",
+    "20315": "failure",
+    "20316": "failure",
+    "20317": "successful",
+    "20318": "failure",
+    "20319": "failure",
+    "20320": "infeasible",
+    "20321": "failure",
+    "20322": "infeasible",
+    "20323": "failure",
+    "20324": "failure",
+    "20325": "infeasible",
+    "20326": "failure",
+    "20327": "failure",
+    "20328": "failure",
+    "20329": "failure",
+    "20330": "successful",
+    "20331": "failure",
+    "20332": "failure",
+    "20333": "failure",
+    "20334": "failure",
+    "20335": "failure",
+    "20336": "failure",
+    "20337": "successful",
+    "20338": "infeasible",
+    "20339": "failure",
+    "20340": "failure",
+    "20341": "failure",
+    "20342": "failure",
+    "20343": "failure",
+    "20344": "successful",
+    "20345": "already_complete",
+    "20346": "infeasible",
+    "20347": "failure",
+    "20348": "failure",
+    "20349": "failure",
+    "20350": "failure",
+    "20351": "failure",
+    "20352": "failure",
+    "20353": "failure",
+    "20354": "failure",
+    "20355": "failure",
+    "20356": "failure",
+    "20357": "failure",
+    "20358": "infeasible",
+    "20359": "failure",
+    "20360": "infeasible",
+    "20361": "failure",
+    "20362": "failure",
+    "20363": "failure",
+    "20364": "failure",
+    "20365": "failure",
+    "20366": "successful",
+    "20367": "successful",
+    "20368": "infeasible",
+    "20369": "failure",
+    "20370": "failure",
+    "20371": "failure",
+    "20372": "failure",
+    "20373": "failure",
+    "20374": "failure",
+    "20375": "failure",
+    "20376": "failure",
+    "20377": "successful",
+    "20378": "infeasible",
+    "20379": "failure",
+    "20380": "failure",
+    "20381": "successful",
+    "20382": "failure",
+    "20383": "failure",
+    "20384": "failure",
+    "20385": "successful",
+    "20386": "failure",
+    "20387": "failure",
+    "20388": "successful",
+    "20389": "failure",
+    "20390": "failure",
+    "20391": "failure",
+    "20392": "failure",
+    "20393": "failure",
+    "20394": "successful",
+    "20395": "failure",
+    "20396": "failure",
+    "20397": "infeasible",
+    "20398": "failure",
+    "20399": "failure",
+    "20400": "failure",
+    "20401": "failure",
+    "20402": "failure",
+    "20403": "failure",
+    "20404": "failure",
+    "20405": "failure",
+    "20406": "failure",
+    "20407": "infeasible",
+    "20408": "failure",
+    "20409": "failure",
+    "20410": "failure",
+    "20411": "failure",
+    "20412": "failure",
+    "20413": "failure",
+    "20414": "failure",
+    "20415": "failure",
+    "20416": "failure",
+    "20417": "failure",
+    "20418": "successful",
+    "20419": "failure",
+    "20420": "failure",
+    "20421": "failure",
+    "20422": "failure",
+    "20423": "failure",
+    "20424": "failure",
+    "20425": "failure",
+    "20426": "failure",
+    "20427": "infeasible",
+    "20428": "failure",
+    "20429": "failure",
+    "20430": "failure",
+    "20431": "failure",
+    "20432": "failure",
+    "20433": "failure",
+    "20434": "successful",
+    "20435": "failure",
+    "20436": "failure",
+    "20437": "failure",
+    "20438": "failure",
+    "20439": "failure",
+    "20440": "failure",
+    "20441": "infeasible",
+    "20442": "failure",
+    "20443": "failure",
+    "20444": "failure",
+    "20445": "failure",
+    "20446": "infeasible",
+    "20447": "failure",
+    "20448": "successful",
+    "20449": "failure",
+    "20450": "failure",
+    "20451": "failure",
+    "20452": "successful",
+    "20453": "failure",
+    "20454": "infeasible",
+    "20455": "failure",
+    "20456": "failure",
+    "20457": "successful",
+    "20458": "failure",
+    "20459": "failure",
+    "20460": "failure",
+    "20461": "failure",
+    "20462": "failure",
+    "20463": "infeasible",
+    "20464": "failure",
+    "20465": "failure",
+    "20466": "failure",
+    "20467": "successful",
+    "20468": "failure",
+    "20469": "failure",
+    "20470": "failure",
+    "20471": "failure",
+    "20472": "failure",
+    "20473": "failure",
+    "20474": "successful",
+    "20475": "failure",
+    "20476": "failure",
+    "20477": "failure",
+    "20478": "successful",
+    "20479": "failure",
+    "20480": "failure",
+    "20481": "successful",
+    "20482": "failure",
+    "20483": "failure",
+    "20484": "failure",
+    "20485": "failure",
+    "20486": "failure",
+    "20487": "failure",
+    "20488": "failure",
+    "20489": "failure",
+    "20490": "failure",
+    "20491": "failure",
+    "20492": "failure",
+    "20493": "failure",
+    "20494": "failure",
+    "20495": "failure",
+    "20496": "failure",
+    "20497": "failure",
+    "20498": "failure",
+    "20499": "failure",
+    "20500": "failure",
+    "20501": "failure",
+    "20502": "successful",
+    "20503": "failure",
+    "20504": "failure",
+    "20505": "successful",
+    "20506": "infeasible",
+    "20507": "failure",
+    "20508": "failure",
+    "20509": "failure",
+    "20510": "failure",
+    "20511": "failure",
+    "20512": "successful",
+    "20513": "failure",
+    "20514": "failure",
+    "20515": "failure",
+    "20516": "failure",
+    "20517": "failure",
+    "20518": "failure",
+    "20519": "failure",
+    "20520": "failure",
+    "20521": "failure",
+    "20522": "successful",
+    "20523": "failure",
+    "20524": "successful",
+    "20525": "failure",
+    "20526": "failure",
+    "20527": "failure",
+    "20528": "failure",
+    "20529": "failure",
+    "20530": "successful",
+    "20531": "failure",
+    "20532": "failure",
+    "20533": "failure",
+    "20534": "failure",
+    "20535": "failure",
+    "20536": "failure",
+    "20537": "already_complete",
+    "20538": "failure",
+    "20539": "failure",
+    "20540": "failure",
+    "20541": "failure",
+    "20542": "failure",
+    "20543": "failure",
+    "20544": "failure",
+    "20545": "failure",
+    "20546": "failure",
+    "20547": "failure",
+    "20548": "failure",
+    "20549": "failure",
+    "20550": "failure",
+    "20551": "failure",
+    "20552": "failure",
+    "20553": "failure",
+    "20554": "failure",
+    "20555": "failure",
+    "20556": "failure",
+    "20557": "failure",
+    "20558": "failure",
+    "20559": "infeasible",
+    "20560": "failure",
+    "20561": "already_complete",
+    "20562": "failure",
+    "20563": "failure",
+    "20564": "successful",
+    "20565": "failure",
+    "20566": "failure",
+    "20567": "failure",
+    "20568": "failure",
+    "20569": "failure",
+    "20570": "failure",
+    "20571": "already_complete",
+    "20572": "failure",
+    "20573": "failure",
+    "20574": "failure",
+    "20575": "failure",
+    "20576": "failure",
+    "20577": "failure",
+    "20578": "failure",
+    "20579": "failure",
+    "20580": "successful",
+    "20581": "failure",
+    "20582": "failure",
+    "20583": "infeasible",
+    "20584": "failure",
+    "20585": "failure",
+    "20586": "failure",
+    "20587": "successful",
+    "20588": "failure",
+    "20589": "failure",
+    "20590": "failure",
+    "20591": "failure",
+    "20592": "failure",
+    "20593": "failure",
+    "20594": "failure",
+    "20595": "successful",
+    "20596": "failure",
+    "20597": "failure",
+    "20598": "failure",
+    "20599": "failure",
+    "20600": "failure",
+    "20601": "failure",
+    "20602": "infeasible",
+    "20603": "failure",
+    "20604": "failure",
+    "20605": "successful",
+    "20606": "infeasible",
+    "20607": "infeasible",
+    "20608": "failure",
+    "20609": "failure",
+    "20610": "infeasible",
+    "20611": "failure",
+    "20612": "successful",
+    "20613": "failure",
+    "20614": "failure",
+    "20615": "failure",
+    "20616": "failure",
+    "20617": "failure",
+    "20618": "failure",
+    "20619": "failure",
+    "20620": "failure",
+    "20621": "failure",
+    "20622": "successful",
+    "20623": "failure",
+    "20624": "failure",
+    "20625": "failure",
+    "20626": "failure",
+    "20627": "failure",
+    "20628": "failure",
+    "20629": "failure",
+    "20630": "failure",
+    "20631": "failure",
+    "20632": "failure",
+    "20633": "failure",
+    "20634": "failure",
+    "20635": "failure",
+    "20636": "infeasible",
+    "20637": "failure",
+    "20638": "failure",
+    "20639": "failure",
+    "20640": "failure",
+    "20641": "successful",
+    "20642": "failure",
+    "20643": "successful",
+    "20644": "failure",
+    "20645": "successful",
+    "20646": "successful",
+    "20647": "failure",
+    "20648": "failure",
+    "20649": "failure",
+    "20650": "successful",
+    "20651": "failure",
+    "20652": "failure",
+    "20653": "successful",
+    "20654": "failure",
+    "20655": "failure",
+    "20656": "failure",
+    "20657": "failure",
+    "20658": "failure",
+    "20659": "successful",
+    "20660": "successful",
+    "20661": "failure",
+    "20662": "infeasible",
+    "20663": "infeasible",
+    "20664": "failure",
+    "20665": "failure",
+    "20666": "failure",
+    "20667": "infeasible",
+    "20668": "infeasible",
+    "20669": "failure",
+    "20670": "failure",
+    "20671": "failure",
+    "20672": "failure",
+    "20673": "failure",
+    "20674": "successful",
+    "20675": "infeasible",
+    "20676": "failure",
+    "20677": "failure",
+    "20678": "failure",
+    "20679": "failure",
+    "20680": "successful",
+    "20681": "failure",
+    "20682": "failure",
+    "20683": "failure",
+    "20684": "failure",
+    "20685": "failure",
+    "20686": "infeasible",
+    "20687": "failure",
+    "20688": "failure",
+    "20689": "failure",
+    "20690": "failure",
+    "20691": "successful",
+    "20692": "failure",
+    "20693": "failure",
+    "20694": "successful",
+    "20695": "failure",
+    "20696": "failure",
+    "20697": "failure",
+    "20698": "failure",
+    "20699": "failure",
+    "20700": "failure",
+    "20701": "failure",
+    "20702": "infeasible",
+    "20703": "failure",
+    "20704": "infeasible",
+    "20705": "failure",
+    "20706": "failure",
+    "20707": "failure",
+    "20708": "infeasible",
+    "20709": "failure",
+    "20710": "successful",
+    "20711": "infeasible",
+    "20712": "failure",
+    "20713": "failure",
+    "20714": "failure",
+    "20715": "already_complete",
+    "20716": "infeasible",
+    "20717": "failure",
+    "20718": "failure",
+    "20719": "failure",
+    "20720": "failure",
+    "20721": "failure",
+    "20722": "failure",
+    "20723": "failure",
+    "20724": "failure",
+    "20725": "infeasible",
+    "20726": "failure",
+    "20727": "infeasible",
+    "20728": "failure",
+    "20729": "failure",
+    "20730": "failure",
+    "20731": "infeasible",
+    "20732": "failure",
+    "20733": "failure",
+    "20734": "failure",
+    "20735": "failure",
+    "20736": "successful",
+    "20737": "failure",
+    "20738": "failure",
+    "20739": "failure",
+    "20740": "failure",
+    "20741": "failure",
+    "20742": "failure",
+    "20743": "infeasible",
+    "20744": "failure",
+    "20745": "infeasible",
+    "20746": "failure",
+    "20747": "failure",
+    "20748": "failure",
+    "20749": "failure",
+    "20750": "successful",
+    "20751": "successful",
+    "20752": "failure",
+    "20753": "failure",
+    "20754": "failure",
+    "20755": "infeasible",
+    "20756": "failure",
+    "20757": "infeasible",
+    "20758": "failure",
+    "20759": "successful",
+    "20760": "failure",
+    "20761": "failure",
+    "20762": "failure",
+    "20763": "failure",
+    "20764": "infeasible",
+    "20765": "successful",
+    "20766": "failure",
+    "20767": "failure",
+    "20768": "successful",
+    "20769": "infeasible",
+    "20770": "failure",
+    "20771": "failure",
+    "20772": "failure",
+    "20773": "failure",
+    "20774": "infeasible",
+    "20775": "failure",
+    "20776": "failure",
+    "20777": "infeasible",
+    "20778": "failure",
+    "20779": "failure",
+    "20780": "successful",
+    "20781": "failure",
+    "20782": "infeasible",
+    "20783": "failure",
+    "20784": "failure",
+    "20785": "failure",
+    "20786": "failure",
+    "20787": "failure",
+    "20788": "failure",
+    "20789": "failure",
+    "20790": "failure",
+    "20791": "failure",
+    "20792": "infeasible",
+    "20793": "failure",
+    "20794": "failure",
+    "20795": "failure",
+    "20796": "failure",
+    "20797": "failure",
+    "20798": "infeasible",
+    "20799": "failure",
+    "20800": "infeasible",
+    "20801": "failure",
+    "20802": "failure",
+    "20803": "failure",
+    "20804": "failure",
+    "20805": "failure",
+    "20806": "failure",
+    "20807": "failure",
+    "20808": "failure",
+    "20809": "failure",
+    "20810": "failure",
+    "20811": "failure",
+    "20812": "failure",
+    "20813": "infeasible",
+    "20814": "failure",
+    "20815": "failure",
+    "20816": "failure",
+    "20817": "infeasible",
+    "20818": "failure",
+    "20819": "failure",
+    "20820": "infeasible",
+    "20821": "failure",
+    "20822": "failure",
+    "20823": "failure",
+    "20824": "failure",
+    "20825": "failure",
+    "20826": "failure",
+    "20827": "failure",
+    "20828": "infeasible",
+    "20829": "failure",
+    "20830": "failure",
+    "20831": "failure",
+    "20832": "failure",
+    "20833": "failure",
+    "20834": "failure",
+    "20835": "failure",
+    "20836": "failure",
+    "20837": "failure",
+    "20838": "failure",
+    "20839": "failure",
+    "20840": "failure",
+    "20841": "failure",
+    "20842": "successful",
+    "20843": "failure",
+    "20844": "failure",
+    "20845": "failure",
+    "20846": "failure",
+    "20847": "infeasible",
+    "20848": "already_complete",
+    "20849": "failure",
+    "20850": "failure",
+    "20851": "failure",
+    "20852": "failure",
+    "20853": "failure",
+    "20854": "infeasible",
+    "20855": "failure",
+    "20856": "failure",
+    "20857": "failure",
+    "20858": "failure",
+    "20859": "failure",
+    "20860": "successful",
+    "20861": "failure",
+    "20862": "failure",
+    "20863": "failure",
+    "20864": "successful",
+    "20865": "failure",
+    "20866": "failure",
+    "20867": "infeasible",
+    "20868": "failure",
+    "20869": "failure",
+    "20870": "failure",
+    "20871": "failure",
+    "20872": "infeasible",
+    "20873": "failure",
+    "20874": "successful",
+    "20875": "failure",
+    "20876": "failure",
+    "20877": "failure",
+    "20878": "failure",
+    "20879": "infeasible",
+    "20880": "successful",
+    "20881": "failure",
+    "20882": "infeasible",
+    "20883": "infeasible",
+    "20884": "failure",
+    "20885": "failure",
+    "20886": "failure",
+    "20887": "failure",
+    "20888": "failure",
+    "20889": "failure",
+    "20890": "failure",
+    "20891": "failure",
+    "20892": "failure",
+    "20893": "failure",
+    "20894": "infeasible",
+    "20895": "failure",
+    "20896": "failure",
+    "20897": "failure",
+    "20898": "failure",
+    "20899": "successful",
+    "20900": "failure",
+    "20901": "failure",
+    "20902": "failure",
+    "20903": "failure",
+    "20904": "failure",
+    "20905": "successful",
+    "20906": "infeasible",
+    "20907": "successful",
+    "20908": "failure",
+    "20909": "failure",
+    "20910": "failure",
+    "20911": "failure",
+    "20912": "failure",
+    "20913": "failure",
+    "20914": "failure",
+    "20915": "infeasible",
+    "20916": "failure",
+    "20917": "failure",
+    "20918": "infeasible",
+    "20919": "failure",
+    "20920": "failure",
+    "20921": "infeasible",
+    "20922": "failure",
+    "20923": "failure",
+    "20924": "failure",
+    "20925": "failure",
+    "20926": "successful",
+    "20927": "failure",
+    "20928": "failure",
+    "20929": "failure",
+    "20930": "failure",
+    "20931": "failure",
+    "20932": "failure",
+    "20933": "failure",
+    "20934": "infeasible",
+    "20935": "failure",
+    "20936": "failure",
+    "20937": "failure",
+    "20938": "infeasible",
+    "20939": "failure",
+    "20940": "infeasible",
+    "20941": "failure",
+    "20942": "failure",
+    "20943": "failure",
+    "20944": "failure",
+    "20945": "failure",
+    "20946": "failure",
+    "20947": "failure",
+    "20948": "failure",
+    "20949": "failure",
+    "20950": "failure",
+    "20951": "failure",
+    "20952": "infeasible",
+    "20953": "failure",
+    "20954": "failure",
+    "20955": "failure",
+    "20956": "failure",
+    "20957": "failure",
+    "20958": "failure",
+    "20959": "failure",
+    "20960": "failure",
+    "20961": "failure",
+    "20962": "failure",
+    "20963": "failure",
+    "20964": "failure",
+    "20965": "failure",
+    "20966": "successful",
+    "20967": "failure",
+    "20968": "failure",
+    "20969": "failure",
+    "20970": "failure",
+    "20971": "failure",
+    "20972": "failure",
+    "20973": "failure",
+    "20974": "failure",
+    "20975": "failure",
+    "20976": "failure",
+    "20977": "failure",
+    "20978": "failure",
+    "20979": "failure",
+    "20980": "failure",
+    "20981": "failure",
+    "20982": "failure",
+    "20983": "failure",
+    "20984": "failure",
+    "20985": "successful",
+    "20986": "failure",
+    "20987": "failure",
+    "20988": "failure",
+    "20989": "failure",
+    "20990": "failure",
+    "20991": "failure",
+    "20992": "failure",
+    "20993": "failure",
+    "20994": "failure",
+    "20995": "failure",
+    "20996": "successful",
+    "20997": "failure",
+    "20998": "failure",
+    "20999": "failure",
+    "21000": "failure",
+    "21001": "failure",
+    "21002": "failure",
+    "21003": "failure",
+    "21004": "failure",
+    "21005": "failure",
+    "21006": "failure",
+    "21007": "failure",
+    "21008": "failure",
+    "21009": "failure",
+    "21010": "failure",
+    "21011": "successful",
+    "21012": "failure",
+    "21013": "infeasible",
+    "21014": "failure",
+    "21015": "failure",
+    "21016": "failure",
+    "21017": "failure",
+    "21018": "failure",
+    "21019": "failure",
+    "21020": "failure",
+    "21021": "failure",
+    "21022": "failure",
+    "21023": "failure",
+    "21024": "failure",
+    "21025": "failure",
+    "21026": "failure",
+    "21027": "failure",
+    "21028": "successful",
+    "21029": "failure",
+    "21030": "infeasible",
+    "21031": "failure",
+    "21032": "failure",
+    "21033": "successful",
+    "21034": "successful",
+    "21035": "failure",
+    "21036": "failure",
+    "21037": "failure",
+    "21038": "failure",
+    "21039": "failure",
+    "21040": "failure",
+    "21041": "failure",
+    "21042": "failure",
+    "21043": "failure",
+    "21044": "failure",
+    "21045": "failure",
+    "21046": "failure",
+    "21047": "failure",
+    "21048": "failure",
+    "21049": "failure",
+    "21050": "failure",
+    "21051": "infeasible",
+    "21052": "failure",
+    "21053": "failure",
+    "21054": "failure",
+    "21055": "failure",
+    "21056": "infeasible",
+    "21057": "failure",
+    "21058": "failure",
+    "21059": "failure",
+    "21060": "failure",
+    "21061": "failure",
+    "21062": "failure",
+    "21063": "failure",
+    "21064": "failure",
+    "21065": "failure",
+    "21066": "failure",
+    "21067": "failure",
+    "21068": "already_complete",
+    "21069": "failure",
+    "21070": "failure",
+    "21071": "failure",
+    "21072": "successful",
+    "21073": "failure",
+    "21074": "successful",
+    "21075": "failure",
+    "21076": "failure",
+    "21077": "failure",
+    "21078": "failure",
+    "21079": "failure",
+    "21080": "failure",
+    "21081": "failure",
+    "21082": "failure",
+    "21083": "failure",
+    "21084": "infeasible",
+    "21085": "failure",
+    "21086": "failure",
+    "21087": "failure",
+    "21088": "failure",
+    "21089": "failure",
+    "21090": "failure",
+    "21091": "failure",
+    "21092": "failure",
+    "21093": "failure",
+    "21094": "successful",
+    "21095": "failure",
+    "21096": "failure",
+    "21097": "infeasible",
+    "21098": "failure",
+    "21099": "infeasible",
+    "21100": "failure",
+    "21101": "failure",
+    "21102": "failure",
+    "21103": "infeasible",
+    "21104": "failure",
+    "21105": "failure",
+    "21106": "failure",
+    "21107": "failure",
+    "21108": "failure",
+    "21109": "infeasible",
+    "21110": "failure",
+    "21111": "failure",
+    "21112": "failure",
+    "21113": "failure",
+    "21114": "failure",
+    "21115": "failure",
+    "21116": "failure",
+    "21117": "failure",
+    "21118": "failure",
+    "21119": "failure",
+    "21120": "failure",
+    "21121": "failure",
+    "21122": "successful",
+    "21123": "failure",
+    "21124": "failure",
+    "21125": "failure",
+    "21126": "failure",
+    "21127": "failure",
+    "21128": "failure",
+    "21129": "failure",
+    "21130": "failure",
+    "21131": "failure",
+    "21132": "failure",
+    "21133": "failure",
+    "21134": "successful",
+    "21135": "failure",
+    "21136": "failure",
+    "21137": "failure",
+    "21138": "failure",
+    "21139": "successful",
+    "21140": "infeasible",
+    "21141": "failure",
+    "21142": "failure",
+    "21143": "failure",
+    "21144": "successful",
+    "21145": "failure",
+    "21146": "failure",
+    "21147": "failure",
+    "21148": "failure",
+    "21149": "failure",
+    "21150": "already_complete",
+    "21151": "successful",
+    "21152": "failure",
+    "21153": "failure",
+    "21154": "failure",
+    "21155": "successful",
+    "21156": "failure",
+    "21157": "failure",
+    "21158": "failure",
+    "21159": "successful",
+    "21160": "failure",
+    "21161": "failure",
+    "21162": "failure",
+    "21163": "failure",
+    "21164": "successful",
+    "21165": "failure",
+    "21166": "failure",
+    "21167": "failure",
+    "21168": "already_complete",
+    "21169": "failure",
+    "21170": "failure",
+    "21171": "infeasible",
+    "21172": "failure",
+    "21173": "failure",
+    "21174": "failure",
+    "21175": "failure",
+    "21176": "infeasible",
+    "21177": "failure",
+    "21178": "failure",
+    "21179": "failure",
+    "21180": "infeasible",
+    "21181": "failure",
+    "21182": "successful",
+    "21183": "failure",
+    "21184": "failure",
+    "21185": "failure",
+    "21186": "failure",
+    "21187": "failure",
+    "21188": "failure",
+    "21189": "failure",
+    "21190": "infeasible",
+    "21191": "infeasible",
+    "21192": "failure",
+    "21193": "successful",
+    "21194": "failure",
+    "21195": "failure",
+    "21196": "failure",
+    "21197": "failure",
+    "21198": "failure",
+    "21199": "failure",
+    "21200": "failure",
+    "21201": "failure",
+    "21202": "failure",
+    "21203": "failure",
+    "21204": "failure",
+    "21205": "failure",
+    "21206": "failure",
+    "21207": "successful",
+    "21208": "failure",
+    "21209": "failure",
+    "21210": "failure",
+    "21211": "successful",
+    "21212": "infeasible",
+    "21213": "failure",
+    "21214": "infeasible",
+    "21215": "failure",
+    "21216": "failure",
+    "21217": "failure",
+    "21218": "already_complete",
+    "21219": "failure",
+    "21220": "failure",
+    "21221": "failure",
+    "21222": "failure",
+    "21223": "failure",
+    "21224": "successful",
+    "21225": "failure",
+    "21226": "failure",
+    "21227": "infeasible",
+    "21228": "successful",
+    "21229": "failure",
+    "21230": "failure",
+    "21231": "failure",
+    "21232": "infeasible",
+    "21233": "failure",
+    "21234": "failure",
+    "21235": "failure",
+    "21236": "failure",
+    "21237": "failure",
+    "21238": "failure",
+    "21239": "failure",
+    "21240": "failure",
+    "21241": "failure",
+    "21242": "failure",
+    "21243": "failure",
+    "21244": "failure",
+    "21245": "failure",
+    "21246": "failure",
+    "21247": "failure",
+    "21248": "failure",
+    "21249": "failure",
+    "21250": "successful",
+    "21251": "failure",
+    "21252": "successful",
+    "21253": "failure",
+    "21254": "failure",
+    "21255": "failure",
+    "21256": "failure",
+    "21257": "failure",
+    "21258": "failure",
+    "21259": "failure",
+    "21260": "failure",
+    "21261": "failure",
+    "21262": "infeasible",
+    "21263": "failure",
+    "21264": "failure",
+    "21265": "successful",
+    "21266": "successful",
+    "21267": "successful",
+    "21268": "infeasible",
+    "21269": "failure",
+    "21270": "infeasible",
+    "21271": "failure",
+    "21272": "failure",
+    "21273": "failure",
+    "21274": "failure",
+    "21275": "failure",
+    "21276": "failure",
+    "21277": "failure",
+    "21278": "failure",
+    "21279": "infeasible",
+    "21280": "failure",
+    "21281": "failure",
+    "21282": "failure",
+    "21283": "failure",
+    "21284": "successful",
+    "21285": "failure",
+    "21286": "failure",
+    "21287": "already_complete",
+    "21288": "failure",
+    "21289": "successful",
+    "21290": "failure",
+    "21291": "failure",
+    "21292": "failure",
+    "21293": "failure",
+    "21294": "failure",
+    "21295": "failure",
+    "21296": "failure",
+    "21297": "failure",
+    "21298": "failure",
+    "21299": "failure",
+    "21300": "failure",
+    "21301": "failure",
+    "21302": "failure",
+    "21303": "failure",
+    "21304": "failure",
+    "21305": "failure",
+    "21306": "failure",
+    "21307": "failure",
+    "21308": "infeasible",
+    "21309": "failure",
+    "21310": "failure",
+    "21311": "failure",
+    "21312": "failure",
+    "21313": "successful",
+    "21314": "failure",
+    "21315": "failure",
+    "21316": "failure",
+    "21317": "failure",
+    "21318": "failure",
+    "21319": "failure",
+    "21320": "successful",
+    "21321": "successful",
+    "21322": "failure",
+    "21323": "failure",
+    "21324": "failure",
+    "21325": "failure",
+    "21326": "failure",
+    "21327": "failure",
+    "21328": "successful",
+    "21329": "failure",
+    "21330": "infeasible",
+    "21331": "successful",
+    "21332": "failure",
+    "21333": "successful",
+    "21334": "successful",
+    "21335": "failure",
+    "21336": "failure",
+    "21337": "successful",
+    "21338": "infeasible",
+    "21339": "failure",
+    "21340": "failure",
+    "21341": "failure",
+    "21342": "failure",
+    "21343": "failure",
+    "21344": "failure",
+    "21345": "failure",
+    "21346": "failure",
+    "21347": "infeasible",
+    "21348": "failure",
+    "21349": "failure",
+    "21350": "failure",
+    "21351": "failure",
+    "21352": "failure",
+    "21353": "failure",
+    "21354": "failure",
+    "21355": "successful",
+    "21356": "infeasible",
+    "21357": "successful",
+    "21358": "failure",
+    "21359": "failure",
+    "21360": "infeasible",
+    "21361": "successful",
+    "21362": "failure",
+    "21363": "failure",
+    "21364": "already_complete",
+    "21365": "failure",
+    "21366": "infeasible",
+    "21367": "successful",
+    "21368": "infeasible",
+    "21369": "failure",
+    "21370": "failure",
+    "21371": "successful",
+    "21372": "failure",
+    "21373": "failure",
+    "21374": "failure",
+    "21375": "failure",
+    "21376": "failure",
+    "21377": "infeasible",
+    "21378": "failure",
+    "21379": "infeasible",
+    "21380": "failure",
+    "21381": "failure",
+    "21382": "failure",
+    "21383": "infeasible",
+    "21384": "failure",
+    "21385": "failure",
+    "21386": "infeasible",
+    "21387": "failure",
+    "21388": "failure",
+    "21389": "successful",
+    "21390": "failure",
+    "21391": "infeasible",
+    "21392": "failure",
+    "21393": "failure",
+    "21394": "failure",
+    "21395": "failure",
+    "21396": "failure",
+    "21397": "successful",
+    "21398": "failure",
+    "21399": "failure",
+    "21400": "successful",
+    "21401": "failure",
+    "21402": "failure",
+    "21403": "failure",
+    "21404": "infeasible",
+    "21405": "failure",
+    "21406": "failure",
+    "21407": "failure",
+    "21408": "failure",
+    "21409": "failure",
+    "21410": "failure",
+    "21411": "successful",
+    "21412": "infeasible",
+    "21413": "failure",
+    "21414": "failure",
+    "21415": "failure",
+    "21416": "failure",
+    "21417": "failure",
+    "21418": "failure",
+    "21419": "failure",
+    "21420": "failure",
+    "21421": "failure",
+    "21422": "infeasible",
+    "21423": "failure",
+    "21424": "failure",
+    "21425": "successful",
+    "21426": "failure",
+    "21427": "failure",
+    "21428": "failure",
+    "21429": "failure",
+    "21430": "failure",
+    "21431": "infeasible",
+    "21432": "failure",
+    "21433": "successful",
+    "21434": "failure",
+    "21435": "failure",
+    "21436": "failure",
+    "21437": "failure",
+    "21438": "failure",
+    "21439": "failure",
+    "21440": "failure",
+    "21441": "failure",
+    "21442": "failure",
+    "21443": "failure",
+    "21444": "failure",
+    "21445": "failure",
+    "21446": "failure",
+    "21447": "failure",
+    "21448": "failure",
+    "21449": "successful",
+    "21450": "infeasible",
+    "21451": "infeasible",
+    "21452": "infeasible",
+    "21453": "failure",
+    "21454": "successful",
+    "21455": "infeasible",
+    "21456": "failure",
+    "21457": "failure",
+    "21458": "failure",
+    "21459": "failure",
+    "21460": "failure",
+    "21461": "failure",
+    "21462": "failure",
+    "21463": "failure",
+    "21464": "failure",
+    "21465": "failure",
+    "21466": "failure",
+    "21467": "failure",
+    "21468": "failure",
+    "21469": "failure",
+    "21470": "failure",
+    "21471": "failure",
+    "21472": "failure",
+    "21473": "infeasible",
+    "21474": "failure",
+    "21475": "failure",
+    "21476": "failure",
+    "21477": "failure",
+    "21478": "successful",
+    "21479": "failure",
+    "21480": "failure",
+    "21481": "failure",
+    "21482": "failure",
+    "21483": "infeasible",
+    "21484": "failure",
+    "21485": "failure",
+    "21486": "failure",
+    "21487": "failure",
+    "21488": "failure",
+    "21489": "successful",
+    "21490": "failure",
+    "21491": "successful",
+    "21492": "failure",
+    "21493": "successful",
+    "21494": "failure",
+    "21495": "failure",
+    "21496": "failure",
+    "21497": "failure",
+    "21498": "failure",
+    "21499": "successful",
+    "21500": "failure",
+    "21501": "infeasible",
+    "21502": "failure",
+    "21503": "failure",
+    "21504": "failure",
+    "21505": "failure",
+    "21506": "failure",
+    "21507": "failure",
+    "21508": "failure",
+    "21509": "failure",
+    "21510": "failure",
+    "21511": "failure",
+    "21512": "infeasible",
+    "21513": "failure",
+    "21514": "failure",
+    "21515": "failure",
+    "21516": "failure",
+    "21517": "failure",
+    "21518": "successful",
+    "21519": "failure",
+    "21520": "failure",
+    "21521": "failure",
+    "21522": "failure",
+    "21523": "failure",
+    "21524": "failure",
+    "21525": "successful",
+    "21526": "failure",
+    "21527": "failure",
+    "21528": "failure",
+    "21529": "failure",
+    "21530": "failure",
+    "21531": "failure",
+    "21532": "failure",
+    "21533": "failure",
+    "21534": "failure",
+    "21535": "failure",
+    "21536": "failure",
+    "21537": "failure",
+    "21538": "failure",
+    "21539": "failure",
+    "21540": "failure",
+    "21541": "failure",
+    "21542": "failure",
+    "21543": "failure",
+    "21544": "failure",
+    "21545": "successful",
+    "21546": "infeasible",
+    "21547": "failure",
+    "21548": "infeasible",
+    "21549": "failure",
+    "21550": "successful",
+    "21551": "failure",
+    "21552": "failure",
+    "21553": "failure",
+    "21554": "failure",
+    "21555": "failure",
+    "21556": "failure",
+    "21557": "infeasible",
+    "21558": "successful",
+    "21559": "failure",
+    "21560": "failure",
+    "21561": "failure",
+    "21562": "failure",
+    "21563": "failure",
+    "21564": "infeasible",
+    "21565": "failure",
+    "21566": "failure",
+    "21567": "successful",
+    "21568": "failure",
+    "21569": "failure",
+    "21570": "failure",
+    "21571": "failure",
+    "21572": "failure",
+    "21573": "failure",
+    "21574": "failure",
+    "21575": "failure",
+    "21576": "failure",
+    "21577": "successful",
+    "21578": "failure",
+    "21579": "failure",
+    "21580": "failure",
+    "21581": "failure",
+    "21582": "failure",
+    "21583": "failure",
+    "21584": "failure",
+    "21585": "failure",
+    "21586": "infeasible",
+    "21587": "failure",
+    "21588": "failure",
+    "21589": "failure",
+    "21590": "failure",
+    "21591": "failure",
+    "21592": "failure",
+    "21593": "failure",
+    "21594": "failure",
+    "21595": "failure",
+    "21596": "failure",
+    "21597": "failure",
+    "21598": "successful",
+    "21599": "failure",
+    "21600": "failure",
+    "21601": "failure",
+    "21602": "failure",
+    "21603": "failure",
+    "21604": "failure",
+    "21605": "successful",
+    "21606": "successful",
+    "21607": "failure",
+    "21608": "failure",
+    "21609": "failure",
+    "21610": "failure",
+    "21611": "failure",
+    "21612": "failure",
+    "21613": "failure",
+    "21614": "failure",
+    "21615": "failure",
+    "21616": "successful",
+    "21617": "failure",
+    "21618": "failure",
+    "21619": "failure",
+    "21620": "failure",
+    "21621": "failure",
+    "21622": "successful",
+    "21623": "failure",
+    "21624": "failure",
+    "21625": "infeasible",
+    "21626": "successful",
+    "21627": "failure",
+    "21628": "failure",
+    "21629": "failure",
+    "21630": "failure",
+    "21631": "failure",
+    "21632": "failure",
+    "21633": "failure",
+    "21634": "successful",
+    "21635": "successful",
+    "21636": "failure",
+    "21637": "infeasible",
+    "21638": "failure",
+    "21639": "failure",
+    "21640": "failure",
+    "21641": "failure",
+    "21642": "failure",
+    "21643": "infeasible",
+    "21644": "successful",
+    "21645": "failure",
+    "21646": "failure",
+    "21647": "failure",
+    "21648": "failure",
+    "21649": "failure",
+    "21650": "failure",
+    "21651": "failure",
+    "21652": "failure",
+    "21653": "failure",
+    "21654": "failure",
+    "21655": "failure",
+    "21656": "infeasible",
+    "21657": "failure",
+    "21658": "failure",
+    "21659": "failure",
+    "21660": "failure",
+    "21661": "failure",
+    "21662": "failure",
+    "21663": "infeasible",
+    "21664": "failure",
+    "21665": "failure",
+    "21666": "failure",
+    "21667": "failure",
+    "21668": "failure",
+    "21669": "failure",
+    "21670": "successful",
+    "21671": "failure",
+    "21672": "failure",
+    "21673": "failure",
+    "21674": "failure",
+    "21675": "failure",
+    "21676": "failure",
+    "21677": "failure",
+    "21678": "successful",
+    "21679": "failure",
+    "21680": "failure",
+    "21681": "failure",
+    "21682": "successful",
+    "21683": "infeasible",
+    "21684": "failure",
+    "21685": "infeasible",
+    "21686": "failure",
+    "21687": "failure",
+    "21688": "failure",
+    "21689": "successful",
+    "21690": "infeasible",
+    "21691": "successful",
+    "21692": "failure",
+    "21693": "failure",
+    "21694": "failure",
+    "21695": "failure",
+    "21696": "failure",
+    "21697": "failure",
+    "21698": "failure",
+    "21699": "failure",
+    "21700": "failure",
+    "21701": "failure",
+    "21702": "failure",
+    "21703": "failure",
+    "21704": "failure",
+    "21705": "failure",
+    "21706": "failure",
+    "21707": "failure",
+    "21708": "failure",
+    "21709": "infeasible",
+    "21710": "failure",
+    "21711": "failure",
+    "21712": "failure",
+    "21713": "failure",
+    "21714": "failure",
+    "21715": "infeasible",
+    "21716": "failure",
+    "21717": "failure",
+    "21718": "failure",
+    "21719": "failure",
+    "21720": "failure",
+    "21721": "infeasible",
+    "21722": "failure",
+    "21723": "failure",
+    "21724": "successful",
+    "21725": "failure",
+    "21726": "successful",
+    "21727": "failure",
+    "21728": "failure",
+    "21729": "failure",
+    "21730": "failure",
+    "21731": "failure",
+    "21732": "failure",
+    "21733": "failure",
+    "21734": "failure",
+    "21735": "failure",
+    "21736": "failure",
+    "21737": "infeasible",
+    "21738": "failure",
+    "21739": "failure",
+    "21740": "failure",
+    "21741": "failure",
+    "21742": "failure",
+    "21743": "successful",
+    "21744": "failure",
+    "21745": "failure",
+    "21746": "failure",
+    "21747": "failure",
+    "21748": "failure",
+    "21749": "successful",
+    "21750": "failure",
+    "21751": "failure",
+    "21752": "failure",
+    "21753": "failure",
+    "21754": "failure",
+    "21755": "failure",
+    "21756": "failure",
+    "21757": "failure",
+    "21758": "infeasible",
+    "21759": "failure",
+    "21760": "failure",
+    "21761": "successful",
+    "21762": "failure",
+    "21763": "infeasible",
+    "21764": "failure",
+    "21765": "successful",
+    "21766": "failure",
+    "21767": "failure",
+    "21768": "failure",
+    "21769": "infeasible",
+    "21770": "failure",
+    "21771": "failure",
+    "21772": "failure",
+    "21773": "failure",
+    "21774": "failure",
+    "21775": "failure",
+    "21776": "failure",
+    "21777": "infeasible",
+    "21778": "successful",
+    "21779": "failure",
+    "21780": "failure",
+    "21781": "failure",
+    "21782": "failure",
+    "21783": "failure",
+    "21784": "failure",
+    "21785": "failure",
+    "21786": "failure",
+    "21787": "infeasible",
+    "21788": "failure",
+    "21789": "failure",
+    "21790": "failure",
+    "21791": "failure",
+    "21792": "failure",
+    "21793": "failure",
+    "21794": "failure",
+    "21795": "infeasible",
+    "21796": "infeasible",
+    "21797": "failure",
+    "21798": "failure",
+    "21799": "infeasible",
+    "21800": "failure",
+    "21801": "failure",
+    "21802": "failure",
+    "21803": "failure",
+    "21804": "failure",
+    "21805": "failure",
+    "21806": "failure",
+    "21807": "failure",
+    "21808": "failure",
+    "21809": "infeasible",
+    "21810": "failure",
+    "21811": "failure",
+    "21812": "failure",
+    "21813": "failure",
+    "21814": "infeasible",
+    "21815": "failure",
+    "21816": "failure",
+    "21817": "failure",
+    "21818": "failure",
+    "21819": "failure",
+    "21820": "failure",
+    "21821": "infeasible",
+    "21822": "failure",
+    "21823": "failure",
+    "21824": "failure",
+    "21825": "successful",
+    "21826": "failure",
+    "21827": "failure",
+    "21828": "failure",
+    "21829": "failure",
+    "21830": "failure",
+    "21831": "failure",
+    "21832": "infeasible",
+    "21833": "failure",
+    "21834": "failure",
+    "21835": "failure",
+    "21836": "failure",
+    "21837": "failure",
+    "21838": "failure",
+    "21839": "failure",
+    "21840": "failure",
+    "21841": "failure",
+    "21842": "failure",
+    "21843": "failure",
+    "21844": "failure",
+    "21845": "infeasible",
+    "21846": "failure",
+    "21847": "failure",
+    "21848": "infeasible",
+    "21849": "successful",
+    "21850": "failure",
+    "21851": "failure",
+    "21852": "failure",
+    "21853": "failure",
+    "21854": "failure",
+    "21855": "failure",
+    "21856": "failure",
+    "21857": "failure",
+    "21858": "failure",
+    "21859": "failure",
+    "21860": "failure",
+    "21861": "successful",
+    "21862": "failure",
+    "21863": "failure",
+    "21864": "failure",
+    "21865": "successful",
+    "21866": "successful",
+    "21867": "failure",
+    "21868": "failure",
+    "21869": "infeasible",
+    "21870": "failure",
+    "21871": "failure",
+    "21872": "failure",
+    "21873": "failure",
+    "21874": "failure",
+    "21875": "failure",
+    "21876": "failure",
+    "21877": "failure",
+    "21878": "failure",
+    "21879": "failure",
+    "21880": "successful",
+    "21881": "successful",
+    "21882": "failure",
+    "21883": "failure",
+    "21884": "failure",
+    "21885": "failure",
+    "21886": "failure",
+    "21887": "failure",
+    "21888": "failure",
+    "21889": "failure",
+    "21890": "failure",
+    "21891": "failure",
+    "21892": "failure",
+    "21893": "infeasible",
+    "21894": "failure",
+    "21895": "failure",
+    "21896": "successful",
+    "21897": "infeasible",
+    "21898": "failure",
+    "21899": "failure",
+    "21900": "infeasible",
+    "21901": "infeasible",
+    "21902": "failure",
+    "21903": "successful",
+    "21904": "infeasible",
+    "21905": "failure",
+    "21906": "failure",
+    "21907": "failure",
+    "21908": "failure",
+    "21909": "failure",
+    "21910": "successful",
+    "21911": "failure",
+    "21912": "failure",
+    "21913": "successful",
+    "21914": "failure",
+    "21915": "failure",
+    "21916": "failure",
+    "21917": "failure",
+    "21918": "failure",
+    "21919": "failure",
+    "21920": "failure",
+    "21921": "failure",
+    "21922": "failure",
+    "21923": "failure",
+    "21924": "failure",
+    "21925": "failure",
+    "21926": "failure",
+    "21927": "infeasible",
+    "21928": "failure",
+    "21929": "failure",
+    "21930": "failure",
+    "21931": "failure",
+    "21932": "failure",
+    "21933": "failure",
+    "21934": "failure",
+    "21935": "failure",
+    "21936": "successful",
+    "21937": "failure",
+    "21938": "failure",
+    "21939": "infeasible",
+    "21940": "failure",
+    "21941": "failure",
+    "21942": "failure",
+    "21943": "failure",
+    "21944": "failure",
+    "21945": "failure",
+    "21946": "failure",
+    "21947": "failure",
+    "21948": "failure",
+    "21949": "failure",
+    "21950": "failure",
+    "21951": "failure",
+    "21952": "failure",
+    "21953": "failure",
+    "21954": "failure",
+    "21955": "successful",
+    "21956": "failure",
+    "21957": "failure",
+    "21958": "failure",
+    "21959": "successful",
+    "21960": "failure",
+    "21961": "failure",
+    "21962": "failure",
+    "21963": "failure",
+    "21964": "failure",
+    "21965": "successful",
+    "21966": "failure",
+    "21967": "failure",
+    "21968": "failure",
+    "21969": "failure",
+    "21970": "failure",
+    "21971": "failure",
+    "21972": "infeasible",
+    "21973": "failure",
+    "21974": "failure",
+    "21975": "failure",
+    "21976": "failure",
+    "21977": "failure",
+    "21978": "failure",
+    "21979": "failure",
+    "21980": "failure",
+    "21981": "infeasible",
+    "21982": "failure",
+    "21983": "failure",
+    "21984": "failure",
+    "21985": "failure",
+    "21986": "failure",
+    "21987": "successful",
+    "21988": "failure",
+    "21989": "failure",
+    "21990": "failure",
+    "21991": "failure",
+    "21992": "failure",
+    "21993": "failure",
+    "21994": "infeasible",
+    "21995": "failure",
+    "21996": "failure",
+    "21997": "failure",
+    "21998": "failure",
+    "21999": "failure",
+    "22000": "failure",
+    "22001": "failure",
+    "22002": "successful",
+    "22003": "failure",
+    "22004": "successful",
+    "22005": "failure",
+    "22006": "failure",
+    "22007": "failure",
+    "22008": "failure",
+    "22009": "failure",
+    "22010": "failure",
+    "22011": "failure",
+    "22012": "failure",
+    "22013": "failure",
+    "22014": "successful",
+    "22015": "successful",
+    "22016": "failure",
+    "22017": "failure",
+    "22018": "successful",
+    "22019": "failure",
+    "22020": "infeasible",
+    "22021": "failure",
+    "22022": "failure",
+    "22023": "failure",
+    "22024": "failure",
+    "22025": "failure",
+    "22026": "failure",
+    "22027": "failure",
+    "22028": "failure",
+    "22029": "failure",
+    "22030": "failure",
+    "22031": "successful",
+    "22032": "failure",
+    "22033": "failure",
+    "22034": "failure",
+    "22035": "failure",
+    "22036": "failure",
+    "22037": "failure",
+    "22038": "failure",
+    "22039": "failure",
+    "22040": "failure",
+    "22041": "failure",
+    "22042": "failure",
+    "22043": "failure",
+    "22044": "failure",
+    "22045": "failure",
+    "22046": "failure",
+    "22047": "successful",
+    "22048": "failure",
+    "22049": "failure",
+    "22050": "failure",
+    "22051": "failure",
+    "22052": "failure",
+    "22053": "failure",
+    "22054": "failure",
+    "22055": "failure",
+    "22056": "failure",
+    "22057": "failure",
+    "22058": "failure",
+    "22059": "failure",
+    "22060": "failure",
+    "22061": "successful",
+    "22062": "failure",
+    "22063": "failure",
+    "22064": "failure",
+    "22065": "infeasible",
+    "22066": "failure",
+    "22067": "successful",
+    "22068": "failure",
+    "22069": "failure",
+    "22070": "infeasible",
+    "22071": "successful",
+    "22072": "failure",
+    "22073": "infeasible",
+    "22074": "failure",
+    "22075": "already_complete",
+    "22076": "failure",
+    "22077": "failure",
+    "22078": "failure",
+    "22079": "failure",
+    "22080": "failure",
+    "22081": "failure",
+    "22082": "failure",
+    "22083": "failure",
+    "22084": "failure",
+    "22085": "failure",
+    "22086": "failure",
+    "22087": "failure",
+    "22088": "successful",
+    "22089": "failure",
+    "22090": "failure",
+    "22091": "failure",
+    "22092": "failure",
+    "22093": "failure",
+    "22094": "failure",
+    "22095": "failure",
+    "22096": "failure",
+    "22097": "failure",
+    "22098": "successful",
+    "22099": "failure",
+    "22100": "failure",
+    "22101": "failure",
+    "22102": "failure",
+    "22103": "successful",
+    "22104": "infeasible",
+    "22105": "failure",
+    "22106": "failure",
+    "22107": "failure",
+    "22108": "failure",
+    "22109": "failure",
+    "22110": "failure",
+    "22111": "failure",
+    "22112": "failure",
+    "22113": "failure",
+    "22114": "failure",
+    "22115": "failure",
+    "22116": "failure",
+    "22117": "infeasible",
+    "22118": "successful",
+    "22119": "failure",
+    "22120": "failure",
+    "22121": "failure",
+    "22122": "failure",
+    "22123": "failure",
+    "22124": "failure",
+    "22125": "failure",
+    "22126": "failure",
+    "22127": "failure",
+    "22128": "successful",
+    "22129": "infeasible",
+    "22130": "failure",
+    "22131": "failure",
+    "22132": "failure",
+    "22133": "failure",
+    "22134": "failure",
+    "22135": "failure",
+    "22136": "failure",
+    "22137": "failure",
+    "22138": "failure",
+    "22139": "failure",
+    "22140": "failure",
+    "22141": "failure",
+    "22142": "infeasible",
+    "22143": "failure",
+    "22144": "failure",
+    "22145": "failure",
+    "22146": "failure",
+    "22147": "failure",
+    "22148": "failure",
+    "22149": "successful",
+    "22150": "failure",
+    "22151": "failure",
+    "22152": "failure",
+    "22153": "failure",
+    "22154": "failure",
+    "22155": "failure",
+    "22156": "failure",
+    "22157": "failure",
+    "22158": "failure",
+    "22159": "failure",
+    "22160": "failure",
+    "22161": "failure",
+    "22162": "failure",
+    "22163": "successful",
+    "22164": "failure",
+    "22165": "infeasible",
+    "22166": "failure",
+    "22167": "failure",
+    "22168": "infeasible",
+    "22169": "failure",
+    "22170": "failure",
+    "22171": "failure",
+    "22172": "failure",
+    "22173": "failure",
+    "22174": "failure",
+    "22175": "failure",
+    "22176": "failure",
+    "22177": "failure",
+    "22178": "failure",
+    "22179": "failure",
+    "22180": "failure",
+    "22181": "failure",
+    "22182": "failure",
+    "22183": "failure",
+    "22184": "failure",
+    "22185": "failure",
+    "22186": "failure",
+    "22187": "failure",
+    "22188": "successful",
+    "22189": "failure",
+    "22190": "already_complete",
+    "22191": "infeasible",
+    "22192": "infeasible",
+    "22193": "failure",
+    "22194": "failure",
+    "22195": "failure",
+    "22196": "failure",
+    "22197": "failure",
+    "22198": "failure",
+    "22199": "successful",
+    "22200": "failure",
+    "22201": "failure",
+    "22202": "failure",
+    "22203": "infeasible",
+    "22204": "failure",
+    "22205": "infeasible",
+    "22206": "failure",
+    "22207": "failure",
+    "22208": "failure",
+    "22209": "failure",
+    "22210": "failure",
+    "22211": "failure",
+    "22212": "failure",
+    "22213": "failure",
+    "22214": "failure",
+    "22215": "failure",
+    "22216": "failure",
+    "22217": "failure",
+    "22218": "infeasible",
+    "22219": "failure",
+    "22220": "failure",
+    "22221": "failure",
+    "22222": "failure",
+    "22223": "failure",
+    "22224": "failure",
+    "22225": "failure",
+    "22226": "failure",
+    "22227": "failure",
+    "22228": "infeasible",
+    "22229": "failure",
+    "22230": "successful",
+    "22231": "failure",
+    "22232": "successful",
+    "22233": "successful",
+    "22234": "successful",
+    "22235": "infeasible",
+    "22236": "already_complete",
+    "22237": "already_complete",
+    "22238": "failure",
+    "22239": "failure",
+    "22240": "failure",
+    "22241": "failure",
+    "22242": "failure",
+    "22243": "failure",
+    "22244": "infeasible",
+    "22245": "failure",
+    "22246": "failure",
+    "22247": "failure",
+    "22248": "failure",
+    "22249": "failure",
+    "22250": "successful",
+    "22251": "failure",
+    "22252": "failure",
+    "22253": "failure",
+    "22254": "successful",
+    "22255": "failure",
+    "22256": "successful",
+    "22257": "successful",
+    "22258": "already_complete",
+    "22259": "failure",
+    "22260": "failure",
+    "22261": "failure",
+    "22262": "failure",
+    "22263": "failure",
+    "22264": "failure",
+    "22265": "failure",
+    "22266": "failure",
+    "22267": "failure",
+    "22268": "failure",
+    "22269": "failure",
+    "22270": "successful",
+    "22271": "failure",
+    "22272": "successful",
+    "22273": "failure",
+    "22274": "failure",
+    "22275": "failure",
+    "22276": "failure",
+    "22277": "failure",
+    "22278": "failure",
+    "22279": "successful",
+    "22280": "successful",
+    "22281": "failure",
+    "22282": "failure",
+    "22283": "failure",
+    "22284": "failure",
+    "22285": "failure",
+    "22286": "failure",
+    "22287": "already_complete",
+    "22288": "successful",
+    "22289": "failure",
+    "22290": "failure",
+    "22291": "failure",
+    "22292": "failure",
+    "22293": "failure",
+    "22294": "already_complete",
+    "22295": "failure",
+    "22296": "failure",
+    "22297": "failure",
+    "22298": "successful",
+    "22299": "failure",
+    "22300": "failure",
+    "22301": "failure",
+    "22302": "failure",
+    "22303": "failure",
+    "22304": "failure",
+    "22305": "failure",
+    "22306": "failure",
+    "22307": "successful",
+    "22308": "infeasible",
+    "22309": "failure",
+    "22310": "failure",
+    "22311": "failure",
+    "22312": "failure",
+    "22313": "failure",
+    "22314": "failure",
+    "22315": "already_complete",
+    "22316": "failure",
+    "22317": "failure",
+    "22318": "failure",
+    "22319": "failure",
+    "22320": "infeasible",
+    "22321": "failure",
+    "22322": "successful",
+    "22323": "failure",
+    "22324": "failure",
+    "22325": "successful",
+    "22326": "failure",
+    "22327": "failure",
+    "22328": "failure",
+    "22329": "failure",
+    "22330": "infeasible",
+    "22331": "failure",
+    "22332": "failure",
+    "22333": "failure",
+    "22334": "failure",
+    "22335": "failure",
+    "22336": "failure",
+    "22337": "successful",
+    "22338": "failure",
+    "22339": "failure",
+    "22340": "failure",
+    "22341": "failure",
+    "22342": "failure",
+    "22343": "failure",
+    "22344": "infeasible",
+    "22345": "failure",
+    "22346": "failure",
+    "22347": "infeasible",
+    "22348": "failure",
+    "22349": "failure",
+    "22350": "failure",
+    "22351": "failure",
+    "22352": "failure",
+    "22353": "infeasible",
+    "22354": "failure",
+    "22355": "successful",
+    "22356": "failure",
+    "22357": "infeasible",
+    "22358": "failure",
+    "22359": "failure",
+    "22360": "failure",
+    "22361": "failure",
+    "22362": "successful",
+    "22363": "failure",
+    "22364": "successful",
+    "22365": "failure",
+    "22366": "failure",
+    "22367": "failure",
+    "22368": "failure",
+    "22369": "failure",
+    "22370": "failure",
+    "22371": "failure",
+    "22372": "infeasible",
+    "22373": "failure",
+    "22374": "failure",
+    "22375": "failure",
+    "22376": "successful",
+    "22377": "failure",
+    "22378": "failure",
+    "22379": "failure",
+    "22380": "failure",
+    "22381": "infeasible",
+    "22382": "infeasible",
+    "22383": "failure",
+    "22384": "infeasible",
+    "22385": "failure",
+    "22386": "failure",
+    "22387": "failure",
+    "22388": "failure",
+    "22389": "infeasible",
+    "22390": "failure",
+    "22391": "failure",
+    "22392": "infeasible",
+    "22393": "successful",
+    "22394": "failure",
+    "22395": "failure",
+    "22396": "infeasible",
+    "22397": "successful",
+    "22398": "failure",
+    "22399": "infeasible",
+    "22400": "failure",
+    "22401": "failure",
+    "22402": "failure",
+    "22403": "failure",
+    "22404": "successful",
+    "22405": "failure",
+    "22406": "failure",
+    "22407": "successful",
+    "22408": "successful",
+    "22409": "failure",
+    "22410": "failure",
+    "22411": "failure",
+    "22412": "failure",
+    "22413": "failure",
+    "22414": "failure",
+    "22415": "failure",
+    "22416": "failure",
+    "22417": "failure",
+    "22418": "successful",
+    "22419": "failure",
+    "22420": "failure",
+    "22421": "infeasible",
+    "22422": "failure",
+    "22423": "failure",
+    "22424": "failure",
+    "22425": "failure",
+    "22426": "failure",
+    "22427": "failure",
+    "22428": "infeasible",
+    "22429": "failure",
+    "22430": "failure",
+    "22431": "failure",
+    "22432": "successful",
+    "22433": "failure",
+    "22434": "failure",
+    "22435": "failure",
+    "22436": "failure",
+    "22437": "infeasible",
+    "22438": "failure",
+    "22439": "failure",
+    "22440": "failure",
+    "22441": "successful",
+    "22442": "failure",
+    "22443": "failure",
+    "22444": "infeasible",
+    "22445": "failure",
+    "22446": "failure",
+    "22447": "failure",
+    "22448": "failure",
+    "22449": "successful",
+    "22450": "failure",
+    "22451": "failure",
+    "22452": "successful",
+    "22453": "failure",
+    "22454": "failure",
+    "22455": "failure",
+    "22456": "failure",
+    "22457": "failure",
+    "22458": "failure",
+    "22459": "failure",
+    "22460": "failure",
+    "22461": "failure",
+    "22462": "failure",
+    "22463": "failure",
+    "22464": "infeasible",
+    "22465": "failure",
+    "22466": "failure",
+    "22467": "failure",
+    "22468": "failure",
+    "22469": "failure",
+    "22470": "failure",
+    "22471": "failure",
+    "22472": "failure",
+    "22473": "failure",
+    "22474": "failure",
+    "22475": "failure",
+    "22476": "infeasible",
+    "22477": "infeasible",
+    "22478": "failure",
+    "22479": "failure",
+    "22480": "failure",
+    "22481": "failure",
+    "22482": "failure",
+    "22483": "infeasible",
+    "22484": "failure",
+    "22485": "failure",
+    "22486": "failure",
+    "22487": "failure",
+    "22488": "failure",
+    "22489": "failure",
+    "22490": "failure",
+    "22491": "failure",
+    "22492": "failure",
+    "22493": "failure",
+    "22494": "failure",
+    "22495": "failure",
+    "22496": "failure",
+    "22497": "failure",
+    "22498": "failure",
+    "22499": "already_complete",
+    "22500": "failure",
+    "22501": "failure",
+    "22502": "failure",
+    "22503": "failure",
+    "22504": "successful",
+    "22505": "failure",
+    "22506": "failure",
+    "22507": "successful",
+    "22508": "infeasible",
+    "22509": "failure",
+    "22510": "failure",
+    "22511": "failure",
+    "22512": "failure",
+    "22513": "failure",
+    "22514": "infeasible",
+    "22515": "failure",
+    "22516": "already_complete",
+    "22517": "failure",
+    "22518": "failure",
+    "22519": "failure",
+    "22520": "infeasible",
+    "22521": "failure",
+    "22522": "failure",
+    "22523": "failure",
+    "22524": "failure",
+    "22525": "failure",
+    "22526": "successful",
+    "22527": "successful",
+    "22528": "failure",
+    "22529": "failure",
+    "22530": "failure",
+    "22531": "infeasible",
+    "22532": "failure",
+    "22533": "infeasible",
+    "22534": "failure",
+    "22535": "failure",
+    "22536": "failure",
+    "22537": "failure",
+    "22538": "successful",
+    "22539": "failure",
+    "22540": "failure",
+    "22541": "failure",
+    "22542": "failure",
+    "22543": "successful",
+    "22544": "failure",
+    "22545": "successful",
+    "22546": "successful",
+    "22547": "successful",
+    "22548": "infeasible",
+    "22549": "successful",
+    "22550": "failure",
+    "22551": "failure",
+    "22552": "failure",
+    "22553": "successful",
+    "22554": "failure",
+    "22555": "infeasible",
+    "22556": "failure",
+    "22557": "infeasible",
+    "22558": "failure",
+    "22559": "failure",
+    "22560": "failure",
+    "22561": "failure",
+    "22562": "failure",
+    "22563": "failure",
+    "22564": "successful",
+    "22565": "infeasible",
+    "22566": "failure",
+    "22567": "failure",
+    "22568": "failure",
+    "22569": "failure",
+    "22570": "failure",
+    "22571": "successful",
+    "22572": "successful",
+    "22573": "failure",
+    "22574": "failure",
+    "22575": "failure",
+    "22576": "failure",
+    "22577": "failure",
+    "22578": "failure",
+    "22579": "failure",
+    "22580": "failure",
+    "22581": "failure",
+    "22582": "failure",
+    "22583": "failure",
+    "22584": "failure",
+    "22585": "failure",
+    "22586": "failure",
+    "22587": "failure",
+    "22588": "failure",
+    "22589": "successful",
+    "22590": "failure",
+    "22591": "failure",
+    "22592": "infeasible",
+    "22593": "failure",
+    "22594": "successful",
+    "22595": "failure",
+    "22596": "successful",
+    "22597": "infeasible",
+    "22598": "failure",
+    "22599": "failure",
+    "22600": "successful",
+    "22601": "failure",
+    "22602": "failure",
+    "22603": "failure",
+    "22604": "failure",
+    "22605": "failure",
+    "22606": "failure",
+    "22607": "infeasible",
+    "22608": "failure",
+    "22609": "failure",
+    "22610": "failure",
+    "22611": "failure",
+    "22612": "failure",
+    "22613": "failure",
+    "22614": "successful",
+    "22615": "failure",
+    "22616": "failure",
+    "22617": "failure",
+    "22618": "failure",
+    "22619": "infeasible",
+    "22620": "failure",
+    "22621": "failure",
+    "22622": "failure",
+    "22623": "failure",
+    "22624": "failure",
+    "22625": "failure",
+    "22626": "failure",
+    "22627": "failure",
+    "22628": "failure",
+    "22629": "failure",
+    "22630": "failure",
+    "22631": "infeasible",
+    "22632": "failure",
+    "22633": "failure",
+    "22634": "failure",
+    "22635": "failure",
+    "22636": "failure",
+    "22637": "failure",
+    "22638": "failure",
+    "22639": "failure",
+    "22640": "successful",
+    "22641": "infeasible",
+    "22642": "failure",
+    "22643": "failure",
+    "22644": "failure",
+    "22645": "failure",
+    "22646": "failure",
+    "22647": "failure",
+    "22648": "failure",
+    "22649": "failure",
+    "22650": "failure",
+    "22651": "failure",
+    "22652": "failure",
+    "22653": "failure",
+    "22654": "failure",
+    "22655": "failure",
+    "22656": "already_complete",
+    "22657": "failure",
+    "22658": "failure",
+    "22659": "failure",
+    "22660": "failure",
+    "22661": "failure",
+    "22662": "failure",
+    "22663": "failure",
+    "22664": "failure",
+    "22665": "failure",
+    "22666": "failure",
+    "22667": "failure",
+    "22668": "failure",
+    "22669": "failure",
+    "22670": "failure",
+    "22671": "failure",
+    "22672": "failure",
+    "22673": "failure",
+    "22674": "failure",
+    "22675": "failure",
+    "22676": "failure",
+    "22677": "failure",
+    "22678": "failure",
+    "22679": "failure",
+    "22680": "failure",
+    "22681": "successful",
+    "22682": "failure",
+    "22683": "successful",
+    "22684": "failure",
+    "22685": "failure",
+    "22686": "failure",
+    "22687": "failure",
+    "22688": "failure",
+    "22689": "failure",
+    "22690": "failure",
+    "22691": "failure",
+    "22692": "failure",
+    "22693": "failure",
+    "22694": "failure",
+    "22695": "failure",
+    "22696": "failure",
+    "22697": "failure",
+    "22698": "failure",
+    "22699": "failure",
+    "22700": "failure",
+    "22701": "successful",
+    "22702": "failure",
+    "22703": "failure",
+    "22704": "failure",
+    "22705": "failure",
+    "22706": "failure",
+    "22707": "failure",
+    "22708": "failure",
+    "22709": "successful",
+    "22710": "failure",
+    "22711": "failure",
+    "22712": "failure",
+    "22713": "failure",
+    "22714": "failure",
+    "22715": "failure",
+    "22716": "successful",
+    "22717": "failure",
+    "22718": "failure",
+    "22719": "failure",
+    "22720": "infeasible",
+    "22721": "failure",
+    "22722": "failure",
+    "22723": "failure",
+    "22724": "successful",
+    "22725": "failure",
+    "22726": "failure",
+    "22727": "infeasible",
+    "22728": "successful",
+    "22729": "failure",
+    "22730": "failure",
+    "22731": "failure",
+    "22732": "failure",
+    "22733": "failure",
+    "22734": "failure",
+    "22735": "successful",
+    "22736": "failure",
+    "22737": "failure",
+    "22738": "failure",
+    "22739": "failure",
+    "22740": "successful",
+    "22741": "failure",
+    "22742": "infeasible",
+    "22743": "failure",
+    "22744": "failure",
+    "22745": "failure",
+    "22746": "failure",
+    "22747": "infeasible",
+    "22748": "infeasible",
+    "22749": "failure",
+    "22750": "failure",
+    "22751": "failure",
+    "22752": "failure",
+    "22753": "failure",
+    "22754": "successful",
+    "22755": "failure",
+    "22756": "failure",
+    "22757": "infeasible",
+    "22758": "successful",
+    "22759": "failure",
+    "22760": "failure",
+    "22761": "failure",
+    "22762": "successful",
+    "22763": "failure",
+    "22764": "already_complete",
+    "22765": "infeasible",
+    "22766": "failure",
+    "22767": "failure",
+    "22768": "failure",
+    "22769": "failure",
+    "22770": "failure",
+    "22771": "infeasible",
+    "22772": "failure",
+    "22773": "failure",
+    "22774": "successful",
+    "22775": "failure",
+    "22776": "failure",
+    "22777": "failure",
+    "22778": "failure",
+    "22779": "failure",
+    "22780": "failure",
+    "22781": "failure",
+    "22782": "failure",
+    "22783": "failure",
+    "22784": "successful",
+    "22785": "infeasible",
+    "22786": "failure",
+    "22787": "infeasible",
+    "22788": "failure",
+    "22789": "successful",
+    "22790": "failure",
+    "22791": "failure",
+    "22792": "failure",
+    "22793": "failure",
+    "22794": "infeasible",
+    "22795": "failure",
+    "22796": "failure",
+    "22797": "failure",
+    "22798": "failure",
+    "22799": "successful",
+    "22800": "failure",
+    "22801": "already_complete",
+    "22802": "failure",
+    "22803": "successful",
+    "22804": "failure",
+    "22805": "failure",
+    "22806": "already_complete",
+    "22807": "failure",
+    "22808": "failure",
+    "22809": "failure",
+    "22810": "failure",
+    "22811": "failure",
+    "22812": "failure",
+    "22813": "failure",
+    "22814": "failure",
+    "22815": "failure",
+    "22816": "failure",
+    "22817": "failure",
+    "22818": "failure",
+    "22819": "failure",
+    "22820": "failure",
+    "22821": "failure",
+    "22822": "failure",
+    "22823": "failure",
+    "22824": "failure",
+    "22825": "failure",
+    "22826": "failure",
+    "22827": "failure",
+    "22828": "failure",
+    "22829": "failure",
+    "22830": "failure",
+    "22831": "infeasible",
+    "22832": "failure",
+    "22833": "failure",
+    "22834": "failure",
+    "22835": "failure",
+    "22836": "infeasible",
+    "22837": "failure",
+    "22838": "failure",
+    "22839": "failure",
+    "22840": "failure",
+    "22841": "failure",
+    "22842": "failure",
+    "22843": "failure",
+    "22844": "infeasible",
+    "22845": "failure",
+    "22846": "failure",
+    "22847": "failure",
+    "22848": "failure",
+    "22849": "failure",
+    "22850": "failure",
+    "22851": "infeasible",
+    "22852": "failure",
+    "22853": "failure",
+    "22854": "failure",
+    "22855": "failure",
+    "22856": "failure",
+    "22857": "failure",
+    "22858": "successful",
+    "22859": "failure",
+    "22860": "failure",
+    "22861": "successful",
+    "22862": "successful",
+    "22863": "failure",
+    "22864": "failure",
+    "22865": "infeasible",
+    "22866": "failure",
+    "22867": "failure",
+    "22868": "infeasible",
+    "22869": "failure",
+    "22870": "failure",
+    "22871": "failure",
+    "22872": "successful",
+    "22873": "infeasible",
+    "22874": "failure",
+    "22875": "failure",
+    "22876": "already_complete",
+    "22877": "failure",
+    "22878": "failure",
+    "22879": "successful",
+    "22880": "failure",
+    "22881": "failure",
+    "22882": "infeasible",
+    "22883": "failure",
+    "22884": "failure",
+    "22885": "failure",
+    "22886": "failure",
+    "22887": "failure",
+    "22888": "failure",
+    "22889": "successful",
+    "22890": "failure",
+    "22891": "failure",
+    "22892": "failure",
+    "22893": "failure",
+    "22894": "infeasible",
+    "22895": "failure",
+    "22896": "failure",
+    "22897": "failure",
+    "22898": "failure",
+    "22899": "failure",
+    "22900": "failure",
+    "22901": "infeasible",
+    "22902": "failure",
+    "22903": "failure",
+    "22904": "failure",
+    "22905": "failure",
+    "22906": "failure",
+    "22907": "failure",
+    "22908": "failure",
+    "22909": "failure",
+    "22910": "successful",
+    "22911": "failure",
+    "22912": "failure",
+    "22913": "failure",
+    "22914": "successful",
+    "22915": "failure",
+    "22916": "failure",
+    "22917": "failure",
+    "22918": "failure",
+    "22919": "successful",
+    "22920": "failure",
+    "22921": "failure",
+    "22922": "failure",
+    "22923": "failure",
+    "22924": "failure",
+    "22925": "failure",
+    "22926": "failure",
+    "22927": "already_complete",
+    "22928": "failure",
+    "22929": "failure",
+    "22930": "failure",
+    "22931": "failure",
+    "22932": "failure",
+    "22933": "failure",
+    "22934": "failure",
+    "22935": "failure",
+    "22936": "failure",
+    "22937": "failure",
+    "22938": "failure",
+    "22939": "failure",
+    "22940": "failure",
+    "22941": "infeasible",
+    "22942": "failure",
+    "22943": "failure",
+    "22944": "failure",
+    "22945": "successful",
+    "22946": "failure",
+    "22947": "failure",
+    "22948": "failure",
+    "22949": "successful",
+    "22950": "failure",
+    "22951": "failure",
+    "22952": "failure",
+    "22953": "failure",
+    "22954": "failure",
+    "22955": "failure",
+    "22956": "failure",
+    "22957": "successful",
+    "22958": "failure",
+    "22959": "failure",
+    "22960": "failure",
+    "22961": "successful",
+    "22962": "failure",
+    "22963": "failure",
+    "22964": "infeasible",
+    "22965": "failure",
+    "22966": "failure",
+    "22967": "failure",
+    "22968": "infeasible",
+    "22969": "successful",
+    "22970": "failure",
+    "22971": "failure",
+    "22972": "infeasible",
+    "22973": "infeasible",
+    "22974": "failure",
+    "22975": "infeasible",
+    "22976": "infeasible",
+    "22977": "failure",
+    "22978": "failure",
+    "22979": "infeasible",
+    "22980": "infeasible",
+    "22981": "failure",
+    "22982": "failure",
+    "22983": "infeasible",
+    "22984": "failure",
+    "22985": "failure",
+    "22986": "failure",
+    "22987": "infeasible",
+    "22988": "failure",
+    "22989": "failure",
+    "22990": "failure",
+    "22991": "failure",
+    "22992": "failure",
+    "22993": "infeasible",
+    "22994": "failure",
+    "22995": "failure",
+    "22996": "successful",
+    "22997": "failure",
+    "22998": "failure",
+    "22999": "failure",
+    "23000": "failure",
+    "23001": "failure",
+    "23002": "infeasible",
+    "23003": "failure",
+    "23004": "failure",
+    "23005": "failure",
+    "23006": "failure",
+    "23007": "infeasible",
+    "23008": "failure",
+    "23009": "failure",
+    "23010": "failure",
+    "23011": "successful",
+    "23012": "failure",
+    "23013": "failure",
+    "23014": "failure",
+    "23015": "infeasible",
+    "23016": "successful",
+    "23017": "infeasible",
+    "23018": "infeasible",
+    "23019": "failure",
+    "23020": "successful",
+    "23021": "failure",
+    "23022": "failure",
+    "23023": "infeasible",
+    "23024": "failure",
+    "23025": "successful",
+    "23026": "failure",
+    "23027": "failure",
+    "23028": "successful",
+    "23029": "infeasible",
+    "23030": "failure",
+    "23031": "failure",
+    "23032": "failure",
+    "23033": "failure",
+    "23034": "failure",
+    "23035": "failure",
+    "23036": "failure",
+    "23037": "failure",
+    "23038": "successful",
+    "23039": "failure",
+    "23040": "infeasible",
+    "23041": "failure",
+    "23042": "successful",
+    "23043": "failure",
+    "23044": "failure",
+    "23045": "failure",
+    "23046": "failure",
+    "23047": "failure",
+    "23048": "failure",
+    "23049": "infeasible",
+    "23050": "successful",
+    "23051": "failure",
+    "23052": "failure",
+    "23053": "failure",
+    "23054": "failure",
+    "23055": "failure",
+    "23056": "failure",
+    "23057": "successful",
+    "23058": "failure",
+    "23059": "failure",
+    "23060": "failure",
+    "23061": "failure",
+    "23062": "failure",
+    "23063": "failure",
+    "23064": "failure",
+    "23065": "failure",
+    "23066": "failure",
+    "23067": "failure",
+    "23068": "infeasible",
+    "23069": "failure",
+    "23070": "failure",
+    "23071": "failure",
+    "23072": "infeasible",
+    "23073": "failure",
+    "23074": "failure",
+    "23075": "infeasible",
+    "23076": "infeasible",
+    "23077": "failure",
+    "23078": "failure",
+    "23079": "failure",
+    "23080": "failure",
+    "23081": "failure",
+    "23082": "failure",
+    "23083": "failure",
+    "23084": "infeasible",
+    "23085": "successful",
+    "23086": "failure",
+    "23087": "failure",
+    "23088": "failure",
+    "23089": "failure",
+    "23090": "failure",
+    "23091": "failure",
+    "23092": "failure",
+    "23093": "infeasible",
+    "23094": "failure",
+    "23095": "failure",
+    "23096": "failure",
+    "23097": "failure",
+    "23098": "failure",
+    "23099": "failure",
+    "23100": "failure",
+    "23101": "infeasible",
+    "23102": "failure",
+    "23103": "failure",
+    "23104": "failure",
+    "23105": "failure",
+    "23106": "failure",
+    "23107": "failure",
+    "23108": "failure",
+    "23109": "failure",
+    "23110": "failure",
+    "23111": "failure",
+    "23112": "failure",
+    "23113": "infeasible",
+    "23114": "infeasible",
+    "23115": "successful",
+    "23116": "failure",
+    "23117": "failure",
+    "23118": "failure",
+    "23119": "failure",
+    "23120": "successful",
+    "23121": "failure",
+    "23122": "failure",
+    "23123": "successful",
+    "23124": "failure",
+    "23125": "failure",
+    "23126": "failure",
+    "23127": "failure",
+    "23128": "failure",
+    "23129": "failure",
+    "23130": "failure",
+    "23131": "failure",
+    "23132": "failure",
+    "23133": "failure",
+    "23134": "failure",
+    "23135": "infeasible",
+    "23136": "successful",
+    "23137": "failure",
+    "23138": "failure",
+    "23139": "failure",
+    "23140": "failure",
+    "23141": "failure",
+    "23142": "failure",
+    "23143": "failure",
+    "23144": "failure",
+    "23145": "failure",
+    "23146": "failure",
+    "23147": "failure",
+    "23148": "failure",
+    "23149": "failure",
+    "23150": "failure",
+    "23151": "failure",
+    "23152": "failure",
+    "23153": "successful",
+    "23154": "failure",
+    "23155": "failure",
+    "23156": "failure",
+    "23157": "failure",
+    "23158": "failure",
+    "23159": "failure",
+    "23160": "failure",
+    "23161": "failure",
+    "23162": "successful",
+    "23163": "failure",
+    "23164": "failure",
+    "23165": "failure",
+    "23166": "infeasible",
+    "23167": "failure",
+    "23168": "already_complete",
+    "23169": "failure",
+    "23170": "successful",
+    "23171": "successful",
+    "23172": "failure",
+    "23173": "failure",
+    "23174": "failure",
+    "23175": "failure",
+    "23176": "failure",
+    "23177": "successful",
+    "23178": "failure",
+    "23179": "failure",
+    "23180": "failure",
+    "23181": "successful",
+    "23182": "failure",
+    "23183": "failure",
+    "23184": "failure",
+    "23185": "failure",
+    "23186": "failure",
+    "23187": "failure",
+    "23188": "failure",
+    "23189": "failure",
+    "23190": "failure",
+    "23191": "failure",
+    "23192": "infeasible",
+    "23193": "failure",
+    "23194": "failure",
+    "23195": "failure",
+    "23196": "failure",
+    "23197": "failure",
+    "23198": "failure",
+    "23199": "failure",
+    "23200": "failure",
+    "23201": "successful",
+    "23202": "infeasible",
+    "23203": "failure",
+    "23204": "failure",
+    "23205": "infeasible",
+    "23206": "failure",
+    "23207": "successful",
+    "23208": "failure",
+    "23209": "already_complete",
+    "23210": "successful",
+    "23211": "infeasible",
+    "23212": "failure",
+    "23213": "failure",
+    "23214": "failure",
+    "23215": "failure",
+    "23216": "failure",
+    "23217": "infeasible",
+    "23218": "failure",
+    "23219": "successful",
+    "23220": "infeasible",
+    "23221": "failure",
+    "23222": "failure",
+    "23223": "failure",
+    "23224": "infeasible",
+    "23225": "failure",
+    "23226": "failure",
+    "23227": "failure",
+    "23228": "failure",
+    "23229": "failure",
+    "23230": "failure",
+    "23231": "successful",
+    "23232": "failure",
+    "23233": "failure",
+    "23234": "failure",
+    "23235": "already_complete",
+    "23236": "failure",
+    "23237": "failure",
+    "23238": "failure",
+    "23239": "failure",
+    "23240": "infeasible",
+    "23241": "failure",
+    "23242": "failure",
+    "23243": "failure",
+    "23244": "failure",
+    "23245": "infeasible",
+    "23246": "failure",
+    "23247": "failure",
+    "23248": "failure",
+    "23249": "failure",
+    "23250": "failure",
+    "23251": "failure",
+    "23252": "failure",
+    "23253": "failure",
+    "23254": "failure",
+    "23255": "failure",
+    "23256": "failure",
+    "23257": "already_complete",
+    "23258": "successful",
+    "23259": "failure",
+    "23260": "failure",
+    "23261": "failure",
+    "23262": "failure",
+    "23263": "failure",
+    "23264": "infeasible",
+    "23265": "failure",
+    "23266": "failure",
+    "23267": "successful",
+    "23268": "failure",
+    "23269": "failure",
+    "23270": "failure",
+    "23271": "failure",
+    "23272": "failure",
+    "23273": "failure",
+    "23274": "failure",
+    "23275": "failure",
+    "23276": "failure",
+    "23277": "failure",
+    "23278": "failure",
+    "23279": "failure",
+    "23280": "failure",
+    "23281": "failure",
+    "23282": "failure",
+    "23283": "failure",
+    "23284": "failure",
+    "23285": "successful",
+    "23286": "failure",
+    "23287": "failure",
+    "23288": "failure",
+    "23289": "failure",
+    "23290": "failure",
+    "23291": "failure",
+    "23292": "successful",
+    "23293": "failure",
+    "23294": "failure",
+    "23295": "failure",
+    "23296": "failure",
+    "23297": "failure",
+    "23298": "failure",
+    "23299": "failure",
+    "23300": "failure",
+    "23301": "failure",
+    "23302": "failure",
+    "23303": "failure",
+    "23304": "failure",
+    "23305": "failure",
+    "23306": "failure",
+    "23307": "failure",
+    "23308": "failure",
+    "23309": "infeasible",
+    "23310": "successful",
+    "23311": "failure",
+    "23312": "failure",
+    "23313": "failure",
+    "23314": "failure",
+    "23315": "successful",
+    "23316": "infeasible",
+    "23317": "successful",
+    "23318": "failure",
+    "23319": "failure",
+    "23320": "failure",
+    "23321": "failure",
+    "23322": "failure",
+    "23323": "failure",
+    "23324": "failure",
+    "23325": "failure",
+    "23326": "failure",
+    "23327": "failure",
+    "23328": "failure",
+    "23329": "successful",
+    "23330": "successful",
+    "23331": "infeasible",
+    "23332": "failure",
+    "23333": "failure",
+    "23334": "failure",
+    "23335": "successful",
+    "23336": "failure",
+    "23337": "failure",
+    "23338": "infeasible",
+    "23339": "failure",
+    "23340": "successful",
+    "23341": "failure",
+    "23342": "failure",
+    "23343": "failure",
+    "23344": "failure",
+    "23345": "failure",
+    "23346": "failure",
+    "23347": "failure",
+    "23348": "failure",
+    "23349": "failure",
+    "23350": "failure",
+    "23351": "successful",
+    "23352": "failure",
+    "23353": "failure",
+    "23354": "failure",
+    "23355": "failure",
+    "23356": "infeasible",
+    "23357": "infeasible",
+    "23358": "failure",
+    "23359": "successful",
+    "23360": "failure",
+    "23361": "failure",
+    "23362": "infeasible",
+    "23363": "failure",
+    "23364": "failure",
+    "23365": "failure",
+    "23366": "successful",
+    "23367": "failure",
+    "23368": "failure",
+    "23369": "successful",
+    "23370": "failure",
+    "23371": "failure",
+    "23372": "failure",
+    "23373": "failure",
+    "23374": "failure",
+    "23375": "infeasible",
+    "23376": "failure",
+    "23377": "failure",
+    "23378": "failure",
+    "23379": "infeasible",
+    "23380": "failure",
+    "23381": "failure",
+    "23382": "infeasible",
+    "23383": "failure",
+    "23384": "successful",
+    "23385": "failure",
+    "23386": "failure",
+    "23387": "failure",
+    "23388": "failure",
+    "23389": "failure",
+    "23390": "failure",
+    "23391": "failure",
+    "23392": "failure",
+    "23393": "failure",
+    "23394": "successful",
+    "23395": "failure",
+    "23396": "failure",
+    "23397": "successful",
+    "23398": "failure",
+    "23399": "failure",
+    "23400": "failure",
+    "23401": "failure",
+    "23402": "already_complete",
+    "23403": "failure",
+    "23404": "failure",
+    "23405": "failure",
+    "23406": "failure",
+    "23407": "failure",
+    "23408": "failure",
+    "23409": "failure",
+    "23410": "failure",
+    "23411": "failure",
+    "23412": "successful",
+    "23413": "failure",
+    "23414": "failure",
+    "23415": "failure",
+    "23416": "failure",
+    "23417": "infeasible",
+    "23418": "failure",
+    "23419": "failure",
+    "23420": "failure",
+    "23421": "failure",
+    "23422": "failure",
+    "23423": "infeasible",
+    "23424": "failure",
+    "23425": "failure",
+    "23426": "failure",
+    "23427": "failure",
+    "23428": "failure",
+    "23429": "failure",
+    "23430": "failure",
+    "23431": "failure",
+    "23432": "failure",
+    "23433": "successful",
+    "23434": "failure",
+    "23435": "failure",
+    "23436": "failure",
+    "23437": "infeasible",
+    "23438": "failure",
+    "23439": "already_complete",
+    "23440": "failure",
+    "23441": "infeasible",
+    "23442": "failure",
+    "23443": "failure",
+    "23444": "failure",
+    "23445": "failure",
+    "23446": "successful",
+    "23447": "failure",
+    "23448": "failure",
+    "23449": "failure",
+    "23450": "failure",
+    "23451": "failure",
+    "23452": "failure",
+    "23453": "failure",
+    "23454": "infeasible",
+    "23455": "failure",
+    "23456": "failure",
+    "23457": "successful",
+    "23458": "failure",
+    "23459": "failure",
+    "23460": "successful",
+    "23461": "failure",
+    "23462": "failure",
+    "23463": "failure",
+    "23464": "failure",
+    "23465": "failure",
+    "23466": "failure",
+    "23467": "failure",
+    "23468": "failure",
+    "23469": "failure",
+    "23470": "infeasible",
+    "23471": "failure",
+    "23472": "failure",
+    "23473": "failure",
+    "23474": "failure",
+    "23475": "infeasible",
+    "23476": "failure",
+    "23477": "failure",
+    "23478": "failure",
+    "23479": "failure",
+    "23480": "infeasible",
+    "23481": "failure",
+    "23482": "infeasible",
+    "23483": "failure",
+    "23484": "failure",
+    "23485": "failure",
+    "23486": "failure",
+    "23487": "failure",
+    "23488": "failure",
+    "23489": "successful",
+    "23490": "failure",
+    "23491": "failure",
+    "23492": "failure",
+    "23493": "failure",
+    "23494": "failure",
+    "23495": "failure",
+    "23496": "failure",
+    "23497": "failure",
+    "23498": "failure",
+    "23499": "failure",
+    "23500": "failure",
+    "23501": "failure",
+    "23502": "failure",
+    "23503": "successful",
+    "23504": "failure",
+    "23505": "successful",
+    "23506": "failure",
+    "23507": "failure",
+    "23508": "successful",
+    "23509": "failure",
+    "23510": "successful",
+    "23511": "infeasible",
+    "23512": "failure",
+    "23513": "failure",
+    "23514": "already_complete",
+    "23515": "failure",
+    "23516": "failure",
+    "23517": "successful",
+    "23518": "successful",
+    "23519": "failure",
+    "23520": "failure",
+    "23521": "failure",
+    "23522": "failure",
+    "23523": "infeasible",
+    "23524": "failure",
+    "23525": "failure",
+    "23526": "failure",
+    "23527": "failure",
+    "23528": "successful",
+    "23529": "failure",
+    "23530": "infeasible",
+    "23531": "failure",
+    "23532": "failure",
+    "23533": "failure",
+    "23534": "failure",
+    "23535": "failure",
+    "23536": "failure",
+    "23537": "failure",
+    "23538": "successful",
+    "23539": "failure",
+    "23540": "infeasible",
+    "23541": "successful",
+    "23542": "infeasible",
+    "23543": "failure",
+    "23544": "failure",
+    "23545": "failure",
+    "23546": "failure",
+    "23547": "failure",
+    "23548": "infeasible",
+    "23549": "infeasible",
+    "23550": "successful",
+    "23551": "successful",
+    "23552": "failure",
+    "23553": "failure",
+    "23554": "failure",
+    "23555": "infeasible",
+    "23556": "failure",
+    "23557": "failure",
+    "23558": "failure",
+    "23559": "failure",
+    "23560": "failure",
+    "23561": "successful",
+    "23562": "successful",
+    "23563": "failure",
+    "23564": "failure",
+    "23565": "successful",
+    "23566": "failure",
+    "23567": "infeasible",
+    "23568": "failure",
+    "23569": "failure",
+    "23570": "successful",
+    "23571": "failure",
+    "23572": "failure",
+    "23573": "failure",
+    "23574": "failure",
+    "23575": "infeasible",
+    "23576": "failure",
+    "23577": "failure",
+    "23578": "failure",
+    "23579": "failure",
+    "23580": "failure",
+    "23581": "infeasible",
+    "23582": "failure",
+    "23583": "failure",
+    "23584": "failure",
+    "23585": "failure",
+    "23586": "successful",
+    "23587": "failure",
+    "23588": "infeasible",
+    "23589": "failure",
+    "23590": "failure",
+    "23591": "infeasible",
+    "23592": "successful",
+    "23593": "infeasible",
+    "23594": "failure",
+    "23595": "failure",
+    "23596": "failure",
+    "23597": "failure",
+    "23598": "successful",
+    "23599": "failure",
+    "23600": "failure",
+    "23601": "failure",
+    "23602": "infeasible",
+    "23603": "failure",
+    "23604": "failure",
+    "23605": "failure",
+    "23606": "failure",
+    "23607": "failure",
+    "23608": "failure",
+    "23609": "failure",
+    "23610": "failure",
+    "23611": "infeasible",
+    "23612": "failure",
+    "23613": "infeasible",
+    "23614": "infeasible",
+    "23615": "infeasible",
+    "23616": "failure",
+    "23617": "failure",
+    "23618": "failure",
+    "23619": "infeasible",
+    "23620": "failure",
+    "23621": "failure",
+    "23622": "failure",
+    "23623": "failure",
+    "23624": "failure",
+    "23625": "successful",
+    "23626": "failure",
+    "23627": "infeasible",
+    "23628": "failure",
+    "23629": "failure",
+    "23630": "failure",
+    "23631": "failure",
+    "23632": "failure",
+    "23633": "failure",
+    "23634": "failure",
+    "23635": "infeasible",
+    "23636": "failure",
+    "23637": "failure",
+    "23638": "failure",
+    "23639": "failure",
+    "23640": "failure",
+    "23641": "failure",
+    "23642": "failure",
+    "23643": "failure",
+    "23644": "failure",
+    "23645": "infeasible",
+    "23646": "failure",
+    "23647": "failure",
+    "23648": "failure",
+    "23649": "failure",
+    "23650": "failure",
+    "23651": "failure",
+    "23652": "failure",
+    "23653": "failure",
+    "23654": "failure",
+    "23655": "infeasible",
+    "23656": "failure",
+    "23657": "failure",
+    "23658": "failure",
+    "23659": "failure",
+    "23660": "failure",
+    "23661": "successful",
+    "23662": "failure",
+    "23663": "failure",
+    "23664": "failure",
+    "23665": "failure",
+    "23666": "failure",
+    "23667": "failure",
+    "23668": "failure",
+    "23669": "infeasible",
+    "23670": "failure",
+    "23671": "failure",
+    "23672": "failure",
+    "23673": "failure",
+    "23674": "failure",
+    "23675": "failure",
+    "23676": "failure",
+    "23677": "failure",
+    "23678": "failure",
+    "23679": "infeasible",
+    "23680": "failure",
+    "23681": "failure",
+    "23682": "failure",
+    "23683": "failure",
+    "23684": "infeasible",
+    "23685": "failure",
+    "23686": "successful",
+    "23687": "failure",
+    "23688": "failure",
+    "23689": "failure",
+    "23690": "successful",
+    "23691": "failure",
+    "23692": "failure",
+    "23693": "failure",
+    "23694": "failure",
+    "23695": "failure",
+    "23696": "failure",
+    "23697": "failure",
+    "23698": "successful",
+    "23699": "failure",
+    "23700": "failure",
+    "23701": "successful",
+    "23702": "failure",
+    "23703": "failure",
+    "23704": "failure",
+    "23705": "failure",
+    "23706": "failure",
+    "23707": "failure",
+    "23708": "failure",
+    "23709": "failure",
+    "23710": "failure",
+    "23711": "failure",
+    "23712": "infeasible",
+    "23713": "infeasible",
+    "23714": "failure",
+    "23715": "failure",
+    "23716": "failure",
+    "23717": "failure",
+    "23718": "failure",
+    "23719": "failure",
+    "23720": "failure",
+    "23721": "infeasible",
+    "23722": "failure",
+    "23723": "infeasible",
+    "23724": "failure",
+    "23725": "failure",
+    "23726": "failure",
+    "23727": "failure",
+    "23728": "failure",
+    "23729": "failure",
+    "23730": "infeasible",
+    "23731": "failure",
+    "23732": "failure",
+    "23733": "infeasible",
+    "23734": "failure",
+    "23735": "failure",
+    "23736": "failure",
+    "23737": "failure",
+    "23738": "failure",
+    "23739": "infeasible",
+    "23740": "failure",
+    "23741": "failure",
+    "23742": "failure",
+    "23743": "failure",
+    "23744": "successful",
+    "23745": "successful",
+    "23746": "failure",
+    "23747": "failure",
+    "23748": "failure",
+    "23749": "successful",
+    "23750": "failure",
+    "23751": "failure",
+    "23752": "failure",
+    "23753": "failure",
+    "23754": "failure",
+    "23755": "failure",
+    "23756": "failure",
+    "23757": "failure",
+    "23758": "failure",
+    "23759": "failure",
+    "23760": "failure",
+    "23761": "failure",
+    "23762": "already_complete",
+    "23763": "failure",
+    "23764": "failure",
+    "23765": "failure",
+    "23766": "failure",
+    "23767": "successful",
+    "23768": "failure",
+    "23769": "failure",
+    "23770": "failure",
+    "23771": "failure",
+    "23772": "infeasible",
+    "23773": "failure",
+    "23774": "failure",
+    "23775": "failure",
+    "23776": "failure",
+    "23777": "infeasible",
+    "23778": "failure",
+    "23779": "failure",
+    "23780": "failure",
+    "23781": "failure",
+    "23782": "infeasible",
+    "23783": "failure",
+    "23784": "failure",
+    "23785": "failure",
+    "23786": "failure",
+    "23787": "failure",
+    "23788": "failure",
+    "23789": "successful",
+    "23790": "failure",
+    "23791": "successful",
+    "23792": "successful",
+    "23793": "failure",
+    "23794": "failure",
+    "23795": "successful",
+    "23796": "failure",
+    "23797": "failure",
+    "23798": "failure",
+    "23799": "failure",
+    "23800": "successful",
+    "23801": "failure",
+    "23802": "failure",
+    "23803": "failure",
+    "23804": "infeasible",
+    "23805": "failure",
+    "23806": "failure",
+    "23807": "failure",
+    "23808": "failure",
+    "23809": "failure",
+    "23810": "failure",
+    "23811": "failure",
+    "23812": "infeasible",
+    "23813": "failure",
+    "23814": "failure",
+    "23815": "failure",
+    "23816": "failure",
+    "23817": "failure",
+    "23818": "failure",
+    "23819": "failure",
+    "23820": "infeasible",
+    "23821": "failure",
+    "23822": "failure",
+    "23823": "successful",
+    "23824": "successful",
+    "23825": "infeasible",
+    "23826": "failure",
+    "23827": "failure",
+    "23828": "failure",
+    "23829": "failure",
+    "23830": "failure",
+    "23831": "failure",
+    "23832": "failure",
+    "23833": "failure",
+    "23834": "failure",
+    "23835": "failure",
+    "23836": "failure",
+    "23837": "failure",
+    "23838": "successful",
+    "23839": "failure",
+    "23840": "failure",
+    "23841": "failure",
+    "23842": "failure",
+    "23843": "failure",
+    "23844": "failure",
+    "23845": "infeasible",
+    "23846": "failure",
+    "23847": "failure",
+    "23848": "infeasible",
+    "23849": "failure",
+    "23850": "failure",
+    "23851": "successful",
+    "23852": "failure",
+    "23853": "failure",
+    "23854": "infeasible",
+    "23855": "failure",
+    "23856": "failure",
+    "23857": "failure",
+    "23858": "infeasible",
+    "23859": "failure",
+    "23860": "infeasible",
+    "23861": "failure",
+    "23862": "failure",
+    "23863": "failure",
+    "23864": "failure",
+    "23865": "failure",
+    "23866": "failure",
+    "23867": "failure",
+    "23868": "failure",
+    "23869": "failure",
+    "23870": "failure",
+    "23871": "failure",
+    "23872": "failure",
+    "23873": "failure",
+    "23874": "successful",
+    "23875": "infeasible",
+    "23876": "failure",
+    "23877": "infeasible",
+    "23878": "already_complete",
+    "23879": "failure",
+    "23880": "failure",
+    "23881": "infeasible",
+    "23882": "failure",
+    "23883": "failure",
+    "23884": "failure",
+    "23885": "failure",
+    "23886": "failure",
+    "23887": "failure",
+    "23888": "failure",
+    "23889": "failure",
+    "23890": "infeasible",
+    "23891": "failure",
+    "23892": "failure",
+    "23893": "successful",
+    "23894": "failure",
+    "23895": "failure",
+    "23896": "failure",
+    "23897": "infeasible",
+    "23898": "failure",
+    "23899": "failure",
+    "23900": "failure",
+    "23901": "infeasible",
+    "23902": "failure",
+    "23903": "successful",
+    "23904": "successful",
+    "23905": "failure",
+    "23906": "failure",
+    "23907": "failure",
+    "23908": "failure",
+    "23909": "failure",
+    "23910": "failure",
+    "23911": "failure",
+    "23912": "failure",
+    "23913": "failure",
+    "23914": "failure",
+    "23915": "infeasible",
+    "23916": "infeasible",
+    "23917": "failure",
+    "23918": "failure",
+    "23919": "successful",
+    "23920": "failure",
+    "23921": "failure",
+    "23922": "failure",
+    "23923": "failure",
+    "23924": "failure",
+    "23925": "failure",
+    "23926": "failure",
+    "23927": "failure",
+    "23928": "successful",
+    "23929": "failure",
+    "23930": "infeasible",
+    "23931": "failure",
+    "23932": "failure",
+    "23933": "failure",
+    "23934": "successful",
+    "23935": "failure",
+    "23936": "failure",
+    "23937": "failure",
+    "23938": "failure",
+    "23939": "infeasible",
+    "23940": "failure",
+    "23941": "failure",
+    "23942": "failure",
+    "23943": "successful",
+    "23944": "failure",
+    "23945": "failure",
+    "23946": "failure",
+    "23947": "failure",
+    "23948": "successful",
+    "23949": "failure",
+    "23950": "failure",
+    "23951": "failure",
+    "23952": "infeasible",
+    "23953": "failure",
+    "23954": "failure",
+    "23955": "infeasible",
+    "23956": "successful",
+    "23957": "failure",
+    "23958": "infeasible",
+    "23959": "failure",
+    "23960": "failure",
+    "23961": "failure",
+    "23962": "failure",
+    "23963": "failure",
+    "23964": "failure",
+    "23965": "failure",
+    "23966": "failure",
+    "23967": "failure",
+    "23968": "failure",
+    "23969": "failure",
+    "23970": "failure",
+    "23971": "failure",
+    "23972": "failure",
+    "23973": "infeasible",
+    "23974": "failure",
+    "23975": "failure",
+    "23976": "failure",
+    "23977": "failure",
+    "23978": "failure",
+    "23979": "failure",
+    "23980": "failure",
+    "23981": "failure",
+    "23982": "failure",
+    "23983": "failure",
+    "23984": "failure",
+    "23985": "failure",
+    "23986": "successful",
+    "23987": "successful",
+    "23988": "failure",
+    "23989": "failure",
+    "23990": "failure",
+    "23991": "failure",
+    "23992": "failure",
+    "23993": "successful",
+    "23994": "failure",
+    "23995": "successful",
+    "23996": "successful",
+    "23997": "infeasible",
+    "23998": "failure",
+    "23999": "infeasible",
+    "24000": "failure",
+    "24001": "successful",
+    "24002": "failure",
+    "24003": "infeasible",
+    "24004": "successful",
+    "24005": "failure",
+    "24006": "failure",
+    "24007": "failure",
+    "24008": "failure",
+    "24009": "failure",
+    "24010": "failure",
+    "24011": "failure",
+    "24012": "failure",
+    "24013": "infeasible",
+    "24014": "failure",
+    "24015": "failure",
+    "24016": "failure",
+    "24017": "failure",
+    "24018": "failure",
+    "24019": "failure",
+    "24020": "successful",
+    "24021": "failure",
+    "24022": "failure",
+    "24023": "failure",
+    "24024": "failure",
+    "24025": "failure",
+    "24026": "failure",
+    "24027": "failure",
+    "24028": "successful",
+    "24029": "failure",
+    "24030": "already_complete",
+    "24031": "failure",
+    "24032": "failure",
+    "24033": "failure",
+    "24034": "failure",
+    "24035": "failure",
+    "24036": "failure",
+    "24037": "infeasible",
+    "24038": "failure",
+    "24039": "successful",
+    "24040": "failure",
+    "24041": "failure",
+    "24042": "infeasible",
+    "24043": "failure",
+    "24044": "failure",
+    "24045": "failure",
+    "24046": "failure",
+    "24047": "successful",
+    "24048": "successful",
+    "24049": "failure",
+    "24050": "failure",
+    "24051": "successful",
+    "24052": "failure",
+    "24053": "successful",
+    "24054": "failure",
+    "24055": "successful",
+    "24056": "infeasible",
+    "24057": "failure",
+    "24058": "failure",
+    "24059": "infeasible",
+    "24060": "successful",
+    "24061": "failure",
+    "24062": "failure",
+    "24063": "failure",
+    "24064": "failure",
+    "24065": "failure",
+    "24066": "failure",
+    "24067": "infeasible",
+    "24068": "infeasible",
+    "24069": "failure",
+    "24070": "already_complete",
+    "24071": "failure",
+    "24072": "failure",
+    "24073": "successful",
+    "24074": "failure",
+    "24075": "failure",
+    "24076": "failure",
+    "24077": "failure",
+    "24078": "successful",
+    "24079": "successful",
+    "24080": "failure",
+    "24081": "failure",
+    "24082": "failure",
+    "24083": "infeasible",
+    "24084": "infeasible",
+    "24085": "infeasible",
+    "24086": "failure",
+    "24087": "successful",
+    "24088": "failure",
+    "24089": "failure",
+    "24090": "failure",
+    "24091": "infeasible",
+    "24092": "failure",
+    "24093": "failure",
+    "24094": "failure",
+    "24095": "failure",
+    "24096": "failure",
+    "24097": "infeasible",
+    "24098": "failure",
+    "24099": "failure",
+    "24100": "failure",
+    "24101": "failure",
+    "24102": "failure",
+    "24103": "failure",
+    "24104": "failure",
+    "24105": "failure",
+    "24106": "failure",
+    "24107": "infeasible",
+    "24108": "successful",
+    "24109": "failure",
+    "24110": "infeasible",
+    "24111": "successful",
+    "24112": "failure",
+    "24113": "failure",
+    "24114": "failure",
+    "24115": "failure",
+    "24116": "failure",
+    "24117": "failure",
+    "24118": "infeasible",
+    "24119": "failure",
+    "24120": "failure",
+    "24121": "failure",
+    "24122": "successful",
+    "24123": "failure",
+    "24124": "infeasible",
+    "24125": "infeasible",
+    "24126": "failure",
+    "24127": "failure",
+    "24128": "failure",
+    "24129": "failure",
+    "24130": "failure",
+    "24131": "infeasible",
+    "24132": "successful",
+    "24133": "failure",
+    "24134": "failure",
+    "24135": "failure",
+    "24136": "failure",
+    "24137": "successful",
+    "24138": "failure",
+    "24139": "failure",
+    "24140": "failure",
+    "24141": "failure",
+    "24142": "failure",
+    "24143": "failure",
+    "24144": "failure",
+    "24145": "failure",
+    "24146": "failure",
+    "24147": "failure",
+    "24148": "successful",
+    "24149": "failure",
+    "24150": "failure",
+    "24151": "failure",
+    "24152": "failure",
+    "24153": "failure",
+    "24154": "failure",
+    "24155": "failure",
+    "24156": "failure",
+    "24157": "failure",
+    "24158": "failure",
+    "24159": "infeasible",
+    "24160": "failure",
+    "24161": "failure",
+    "24162": "infeasible",
+    "24163": "infeasible",
+    "24164": "infeasible",
+    "24165": "failure",
+    "24166": "infeasible",
+    "24167": "failure",
+    "24168": "failure",
+    "24169": "failure",
+    "24170": "failure",
+    "24171": "failure",
+    "24172": "failure",
+    "24173": "failure",
+    "24174": "failure",
+    "24175": "already_complete",
+    "24176": "failure",
+    "24177": "failure",
+    "24178": "failure",
+    "24179": "failure",
+    "24180": "failure",
+    "24181": "failure",
+    "24182": "failure",
+    "24183": "failure",
+    "24184": "failure",
+    "24185": "failure",
+    "24186": "failure",
+    "24187": "failure",
+    "24188": "failure",
+    "24189": "failure",
+    "24190": "failure",
+    "24191": "infeasible",
+    "24192": "failure",
+    "24193": "failure",
+    "24194": "failure",
+    "24195": "failure",
+    "24196": "failure",
+    "24197": "failure",
+    "24198": "failure",
+    "24199": "failure",
+    "24200": "successful",
+    "24201": "failure",
+    "24202": "infeasible",
+    "24203": "successful",
+    "24204": "failure",
+    "24205": "infeasible",
+    "24206": "failure",
+    "24207": "failure",
+    "24208": "failure",
+    "24209": "failure",
+    "24210": "failure",
+    "24211": "failure",
+    "24212": "failure",
+    "24213": "failure",
+    "24214": "failure",
+    "24215": "failure",
+    "24216": "failure",
+    "24217": "failure",
+    "24218": "failure",
+    "24219": "failure",
+    "24220": "failure",
+    "24221": "failure",
+    "24222": "failure",
+    "24223": "failure",
+    "24224": "failure",
+    "24225": "infeasible",
+    "24226": "failure",
+    "24227": "failure",
+    "24228": "failure",
+    "24229": "failure",
+    "24230": "failure",
+    "24231": "infeasible",
+    "24232": "failure",
+    "24233": "failure",
+    "24234": "failure",
+    "24235": "failure",
+    "24236": "failure",
+    "24237": "failure",
+    "24238": "failure",
+    "24239": "failure",
+    "24240": "failure",
+    "24241": "successful",
+    "24242": "successful",
+    "24243": "failure",
+    "24244": "failure",
+    "24245": "failure",
+    "24246": "infeasible",
+    "24247": "failure",
+    "24248": "infeasible",
+    "24249": "failure",
+    "24250": "infeasible",
+    "24251": "failure",
+    "24252": "failure",
+    "24253": "failure",
+    "24254": "failure",
+    "24255": "failure",
+    "24256": "failure",
+    "24257": "failure",
+    "24258": "failure",
+    "24259": "failure",
+    "24260": "failure",
+    "24261": "failure",
+    "24262": "failure",
+    "24263": "failure",
+    "24264": "failure",
+    "24265": "infeasible",
+    "24266": "failure",
+    "24267": "failure",
+    "24268": "failure",
+    "24269": "successful",
+    "24270": "failure",
+    "24271": "failure",
+    "24272": "failure",
+    "24273": "failure",
+    "24274": "failure",
+    "24275": "failure",
+    "24276": "failure",
+    "24277": "failure",
+    "24278": "failure",
+    "24279": "failure",
+    "24280": "failure",
+    "24281": "infeasible",
+    "24282": "failure",
+    "24283": "failure",
+    "24284": "failure",
+    "24285": "failure",
+    "24286": "failure",
+    "24287": "successful",
+    "24288": "infeasible",
+    "24289": "failure",
+    "24290": "failure",
+    "24291": "successful",
+    "24292": "infeasible",
+    "24293": "failure",
+    "24294": "failure",
+    "24295": "successful",
+    "24296": "failure",
+    "24297": "infeasible",
+    "24298": "failure",
+    "24299": "failure",
+    "24300": "failure",
+    "24301": "failure",
+    "24302": "failure",
+    "24303": "failure",
+    "24304": "failure",
+    "24305": "failure",
+    "24306": "failure",
+    "24307": "failure",
+    "24308": "successful",
+    "24309": "failure",
+    "24310": "failure",
+    "24311": "failure",
+    "24312": "failure",
+    "24313": "failure",
+    "24314": "failure",
+    "24315": "failure",
+    "24316": "failure",
+    "24317": "infeasible",
+    "24318": "failure",
+    "24319": "failure",
+    "24320": "failure",
+    "24321": "failure",
+    "24322": "failure",
+    "24323": "failure",
+    "24324": "failure",
+    "24325": "failure",
+    "24326": "failure",
+    "24327": "failure",
+    "24328": "successful",
+    "24329": "failure",
+    "24330": "successful",
+    "24331": "failure",
+    "24332": "failure",
+    "24333": "failure",
+    "24334": "failure",
+    "24335": "failure",
+    "24336": "failure",
+    "24337": "failure",
+    "24338": "failure",
+    "24339": "failure",
+    "24340": "failure",
+    "24341": "infeasible",
+    "24342": "failure",
+    "24343": "failure",
+    "24344": "failure",
+    "24345": "failure",
+    "24346": "failure",
+    "24347": "failure",
+    "24348": "successful",
+    "24349": "failure",
+    "24350": "failure",
+    "24351": "failure",
+    "24352": "infeasible",
+    "24353": "failure",
+    "24354": "successful",
+    "24355": "successful",
+    "24356": "failure",
+    "24357": "infeasible",
+    "24358": "failure",
+    "24359": "failure",
+    "24360": "infeasible",
+    "24361": "successful",
+    "24362": "failure",
+    "24363": "successful",
+    "24364": "failure",
+    "24365": "failure",
+    "24366": "failure",
+    "24367": "failure",
+    "24368": "failure",
+    "24369": "failure",
+    "24370": "successful",
+    "24371": "failure",
+    "24372": "infeasible",
+    "24373": "successful",
+    "24374": "failure",
+    "24375": "failure",
+    "24376": "failure",
+    "24377": "infeasible",
+    "24378": "failure",
+    "24379": "failure",
+    "24380": "successful",
+    "24381": "failure",
+    "24382": "failure",
+    "24383": "failure",
+    "24384": "infeasible",
+    "24385": "failure",
+    "24386": "successful",
+    "24387": "failure",
+    "24388": "failure",
+    "24389": "failure",
+    "24390": "infeasible",
+    "24391": "failure",
+    "24392": "failure",
+    "24393": "failure",
+    "24394": "failure",
+    "24395": "failure",
+    "24396": "failure",
+    "24397": "failure",
+    "24398": "failure",
+    "24399": "failure",
+    "24400": "failure",
+    "24401": "failure",
+    "24402": "failure",
+    "24403": "failure",
+    "24404": "failure",
+    "24405": "infeasible",
+    "24406": "failure",
+    "24407": "failure",
+    "24408": "failure",
+    "24409": "failure",
+    "24410": "failure",
+    "24411": "failure",
+    "24412": "infeasible",
+    "24413": "failure",
+    "24414": "infeasible",
+    "24415": "failure",
+    "24416": "failure",
+    "24417": "failure",
+    "24418": "failure",
+    "24419": "failure",
+    "24420": "failure",
+    "24421": "failure",
+    "24422": "failure",
+    "24423": "failure",
+    "24424": "failure",
+    "24425": "failure",
+    "24426": "failure",
+    "24427": "failure",
+    "24428": "successful",
+    "24429": "failure",
+    "24430": "successful",
+    "24431": "failure",
+    "24432": "failure",
+    "24433": "failure",
+    "24434": "failure",
+    "24435": "failure",
+    "24436": "failure",
+    "24437": "failure",
+    "24438": "failure",
+    "24439": "failure",
+    "24440": "failure",
+    "24441": "failure",
+    "24442": "infeasible",
+    "24443": "failure",
+    "24444": "failure",
+    "24445": "failure",
+    "24446": "failure",
+    "24447": "failure",
+    "24448": "failure",
+    "24449": "failure",
+    "24450": "infeasible",
+    "24451": "infeasible",
+    "24452": "already_complete",
+    "24453": "failure",
+    "24454": "failure",
+    "24455": "failure",
+    "24456": "failure",
+    "24457": "failure",
+    "24458": "failure",
+    "24459": "failure",
+    "24460": "failure",
+    "24461": "failure",
+    "24462": "successful",
+    "24463": "failure",
+    "24464": "failure",
+    "24465": "failure",
+    "24466": "failure",
+    "24467": "failure",
+    "24468": "failure",
+    "24469": "successful",
+    "24470": "already_complete",
+    "24471": "successful",
+    "24472": "successful",
+    "24473": "failure",
+    "24474": "failure",
+    "24475": "failure",
+    "24476": "failure",
+    "24477": "failure",
+    "24478": "successful",
+    "24479": "failure",
+    "24480": "failure",
+    "24481": "failure",
+    "24482": "infeasible",
+    "24483": "failure",
+    "24484": "failure",
+    "24485": "failure",
+    "24486": "failure",
+    "24487": "failure",
+    "24488": "failure",
+    "24489": "failure",
+    "24490": "failure",
+    "24491": "successful",
+    "24492": "failure",
+    "24493": "failure",
+    "24494": "failure",
+    "24495": "failure",
+    "24496": "failure",
+    "24497": "failure",
+    "24498": "failure",
+    "24499": "failure",
+    "24500": "failure",
+    "24501": "failure",
+    "24502": "infeasible",
+    "24503": "successful",
+    "24504": "failure",
+    "24505": "failure",
+    "24506": "failure",
+    "24507": "failure",
+    "24508": "successful",
+    "24509": "failure",
+    "24510": "failure",
+    "24511": "failure",
+    "24512": "failure",
+    "24513": "infeasible",
+    "24514": "failure",
+    "24515": "failure",
+    "24516": "successful",
+    "24517": "failure",
+    "24518": "infeasible",
+    "24519": "failure",
+    "24520": "failure",
+    "24521": "infeasible",
+    "24522": "failure",
+    "24523": "failure",
+    "24524": "failure",
+    "24525": "failure",
+    "24526": "failure",
+    "24527": "failure",
+    "24528": "failure",
+    "24529": "failure",
+    "24530": "failure",
+    "24531": "failure",
+    "24532": "failure",
+    "24533": "successful",
+    "24534": "failure",
+    "24535": "failure",
+    "24536": "failure",
+    "24537": "successful",
+    "24538": "successful",
+    "24539": "successful",
+    "24540": "successful",
+    "24541": "failure",
+    "24542": "failure",
+    "24543": "failure",
+    "24544": "failure",
+    "24545": "infeasible",
+    "24546": "failure",
+    "24547": "failure",
+    "24548": "failure",
+    "24549": "failure",
+    "24550": "already_complete",
+    "24551": "failure",
+    "24552": "failure",
+    "24553": "failure",
+    "24554": "failure",
+    "24555": "failure",
+    "24556": "failure",
+    "24557": "infeasible",
+    "24558": "failure",
+    "24559": "failure",
+    "24560": "failure",
+    "24561": "failure",
+    "24562": "failure",
+    "24563": "failure",
+    "24564": "failure",
+    "24565": "failure",
+    "24566": "failure",
+    "24567": "failure",
+    "24568": "failure",
+    "24569": "failure",
+    "24570": "failure",
+    "24571": "failure",
+    "24572": "failure",
+    "24573": "infeasible",
+    "24574": "successful",
+    "24575": "failure",
+    "24576": "failure",
+    "24577": "failure",
+    "24578": "successful",
+    "24579": "failure",
+    "24580": "failure",
+    "24581": "successful",
+    "24582": "failure",
+    "24583": "failure",
+    "24584": "failure",
+    "24585": "failure",
+    "24586": "failure",
+    "24587": "failure",
+    "24588": "failure",
+    "24589": "failure",
+    "24590": "failure",
+    "24591": "failure",
+    "24592": "failure",
+    "24593": "failure",
+    "24594": "failure",
+    "24595": "failure",
+    "24596": "failure",
+    "24597": "failure",
+    "24598": "failure",
+    "24599": "successful",
+    "24600": "successful",
+    "24601": "failure",
+    "24602": "failure",
+    "24603": "successful",
+    "24604": "infeasible",
+    "24605": "failure",
+    "24606": "failure",
+    "24607": "failure",
+    "24608": "successful",
+    "24609": "failure",
+    "24610": "failure",
+    "24611": "failure",
+    "24612": "infeasible",
+    "24613": "failure",
+    "24614": "failure",
+    "24615": "failure",
+    "24616": "failure",
+    "24617": "failure",
+    "24618": "failure",
+    "24619": "failure",
+    "24620": "failure",
+    "24621": "failure",
+    "24622": "infeasible",
+    "24623": "failure",
+    "24624": "failure",
+    "24625": "failure",
+    "24626": "failure",
+    "24627": "failure",
+    "24628": "failure",
+    "24629": "failure",
+    "24630": "failure",
+    "24631": "failure",
+    "24632": "failure",
+    "24633": "failure",
+    "24634": "failure",
+    "24635": "failure",
+    "24636": "failure",
+    "24637": "failure",
+    "24638": "failure",
+    "24639": "failure",
+    "24640": "failure",
+    "24641": "failure",
+    "24642": "failure",
+    "24643": "failure",
+    "24644": "failure",
+    "24645": "failure",
+    "24646": "failure",
+    "24647": "failure",
+    "24648": "failure",
+    "24649": "failure",
+    "24650": "successful",
+    "24651": "failure",
+    "24652": "failure",
+    "24653": "successful",
+    "24654": "failure",
+    "24655": "failure",
+    "24656": "successful",
+    "24657": "failure",
+    "24658": "failure",
+    "24659": "failure",
+    "24660": "failure",
+    "24661": "failure",
+    "24662": "infeasible",
+    "24663": "failure",
+    "24664": "failure",
+    "24665": "failure",
+    "24666": "failure",
+    "24667": "failure",
+    "24668": "failure",
+    "24669": "failure",
+    "24670": "failure",
+    "24671": "failure",
+    "24672": "failure",
+    "24673": "failure",
+    "24674": "failure",
+    "24675": "infeasible",
+    "24676": "failure",
+    "24677": "failure",
+    "24678": "failure",
+    "24679": "failure",
+    "24680": "failure",
+    "24681": "failure",
+    "24682": "failure",
+    "24683": "infeasible",
+    "24684": "failure",
+    "24685": "failure",
+    "24686": "failure",
+    "24687": "successful",
+    "24688": "successful",
+    "24689": "failure",
+    "24690": "failure",
+    "24691": "failure",
+    "24692": "failure",
+    "24693": "failure",
+    "24694": "failure",
+    "24695": "failure",
+    "24696": "failure",
+    "24697": "failure",
+    "24698": "failure",
+    "24699": "successful",
+    "24700": "successful",
+    "24701": "successful",
+    "24702": "failure",
+    "24703": "failure",
+    "24704": "failure",
+    "24705": "failure",
+    "24706": "failure",
+    "24707": "failure",
+    "24708": "failure",
+    "24709": "failure",
+    "24710": "failure",
+    "24711": "failure",
+    "24712": "failure",
+    "24713": "failure",
+    "24714": "failure",
+    "24715": "failure",
+    "24716": "infeasible",
+    "24717": "failure",
+    "24718": "failure",
+    "24719": "failure",
+    "24720": "failure",
+    "24721": "failure",
+    "24722": "infeasible",
+    "24723": "failure",
+    "24724": "failure",
+    "24725": "failure",
+    "24726": "failure",
+    "24727": "infeasible",
+    "24728": "failure",
+    "24729": "failure",
+    "24730": "failure",
+    "24731": "failure",
+    "24732": "failure",
+    "24733": "failure",
+    "24734": "failure",
+    "24735": "failure",
+    "24736": "failure",
+    "24737": "failure",
+    "24738": "failure",
+    "24739": "failure",
+    "24740": "failure",
+    "24741": "infeasible",
+    "24742": "failure",
+    "24743": "failure",
+    "24744": "successful",
+    "24745": "infeasible",
+    "24746": "failure",
+    "24747": "failure",
+    "24748": "successful",
+    "24749": "failure",
+    "24750": "failure",
+    "24751": "infeasible",
+    "24752": "failure",
+    "24753": "failure",
+    "24754": "failure",
+    "24755": "failure",
+    "24756": "failure",
+    "24757": "failure",
+    "24758": "failure",
+    "24759": "failure",
+    "24760": "failure",
+    "24761": "failure",
+    "24762": "failure",
+    "24763": "failure",
+    "24764": "failure",
+    "24765": "infeasible",
+    "24766": "failure",
+    "24767": "infeasible",
+    "24768": "failure",
+    "24769": "infeasible",
+    "24770": "infeasible",
+    "24771": "failure",
+    "24772": "failure",
+    "24773": "failure",
+    "24774": "failure",
+    "24775": "failure",
+    "24776": "failure",
+    "24777": "failure",
+    "24778": "failure",
+    "24779": "infeasible",
+    "24780": "failure",
+    "24781": "failure",
+    "24782": "failure",
+    "24783": "failure",
+    "24784": "failure",
+    "24785": "failure",
+    "24786": "failure",
+    "24787": "failure",
+    "24788": "failure",
+    "24789": "failure",
+    "24790": "failure",
+    "24791": "successful",
+    "24792": "failure",
+    "24793": "failure",
+    "24794": "failure",
+    "24795": "failure",
+    "24796": "failure",
+    "24797": "failure",
+    "24798": "failure",
+    "24799": "failure",
+    "24800": "failure",
+    "24801": "failure",
+    "24802": "failure",
+    "24803": "failure",
+    "24804": "failure",
+    "24805": "successful",
+    "24806": "failure",
+    "24807": "failure",
+    "24808": "failure",
+    "24809": "infeasible",
+    "24810": "infeasible",
+    "24811": "successful",
+    "24812": "failure",
+    "24813": "failure",
+    "24814": "failure",
+    "24815": "failure",
+    "24816": "failure",
+    "24817": "failure",
+    "24818": "failure",
+    "24819": "failure",
+    "24820": "failure",
+    "24821": "infeasible",
+    "24822": "failure",
+    "24823": "failure",
+    "24824": "failure",
+    "24825": "failure",
+    "24826": "failure",
+    "24827": "failure",
+    "24828": "failure",
+    "24829": "failure",
+    "24830": "failure",
+    "24831": "failure",
+    "24832": "failure",
+    "24833": "failure",
+    "24834": "failure",
+    "24835": "infeasible",
+    "24836": "failure",
+    "24837": "failure",
+    "24838": "successful",
+    "24839": "failure",
+    "24840": "failure",
+    "24841": "failure",
+    "24842": "failure",
+    "24843": "infeasible",
+    "24844": "successful",
+    "24845": "failure",
+    "24846": "failure",
+    "24847": "failure",
+    "24848": "infeasible",
+    "24849": "infeasible",
+    "24850": "failure",
+    "24851": "failure",
+    "24852": "failure",
+    "24853": "failure",
+    "24854": "infeasible",
+    "24855": "successful",
+    "24856": "failure",
+    "24857": "already_complete",
+    "24858": "failure",
+    "24859": "failure",
+    "24860": "failure",
+    "24861": "failure",
+    "24862": "failure",
+    "24863": "failure",
+    "24864": "failure",
+    "24865": "failure",
+    "24866": "failure",
+    "24867": "failure",
+    "24868": "failure",
+    "24869": "failure",
+    "24870": "failure",
+    "24871": "failure",
+    "24872": "failure",
+    "24873": "failure",
+    "24874": "failure",
+    "24875": "failure",
+    "24876": "failure",
+    "24877": "failure",
+    "24878": "failure",
+    "24879": "failure",
+    "24880": "already_complete",
+    "24881": "already_complete",
+    "24882": "failure",
+    "24883": "failure",
+    "24884": "failure",
+    "24885": "failure",
+    "24886": "failure",
+    "24887": "failure",
+    "24888": "failure",
+    "24889": "failure",
+    "24890": "failure",
+    "24891": "failure",
+    "24892": "failure",
+    "24893": "failure",
+    "24894": "failure",
+    "24895": "failure",
+    "24896": "failure",
+    "24897": "failure",
+    "24898": "failure",
+    "24899": "successful",
+    "24900": "failure",
+    "24901": "failure",
+    "24902": "failure",
+    "24903": "failure",
+    "24904": "infeasible",
+    "24905": "failure",
+    "24906": "failure",
+    "24907": "failure",
+    "24908": "failure",
+    "24909": "failure",
+    "24910": "failure",
+    "24911": "failure",
+    "24912": "infeasible",
+    "24913": "failure",
+    "24914": "failure",
+    "24915": "successful",
+    "24916": "failure",
+    "24917": "successful",
+    "24918": "failure",
+    "24919": "successful",
+    "24920": "failure",
+    "24921": "successful",
+    "24922": "failure",
+    "24923": "failure",
+    "24924": "failure",
+    "24925": "successful",
+    "24926": "failure",
+    "24927": "failure",
+    "24928": "failure",
+    "24929": "failure",
+    "24930": "failure",
+    "24931": "failure",
+    "24932": "failure",
+    "24933": "failure",
+    "24934": "failure",
+    "24935": "failure",
+    "24936": "failure",
+    "24937": "failure",
+    "24938": "failure",
+    "24939": "already_complete",
+    "24940": "failure",
+    "24941": "failure",
+    "24942": "failure",
+    "24943": "failure",
+    "24944": "infeasible",
+    "24945": "failure",
+    "24946": "failure",
+    "24947": "failure",
+    "24948": "failure",
+    "24949": "failure",
+    "24950": "failure",
+    "24951": "failure",
+    "24952": "failure",
+    "24953": "infeasible",
+    "24954": "already_complete",
+    "24955": "failure",
+    "24956": "failure",
+    "24957": "infeasible",
+    "24958": "failure",
+    "24959": "failure",
+    "24960": "failure",
+    "24961": "infeasible",
+    "24962": "failure",
+    "24963": "failure",
+    "24964": "failure",
+    "24965": "infeasible",
+    "24966": "failure",
+    "24967": "failure",
+    "24968": "infeasible",
+    "24969": "successful",
+    "24970": "failure",
+    "24971": "already_complete",
+    "24972": "failure",
+    "24973": "failure",
+    "24974": "failure",
+    "24975": "failure",
+    "24976": "infeasible",
+    "24977": "failure",
+    "24978": "failure",
+    "24979": "failure",
+    "24980": "successful",
+    "24981": "failure",
+    "24982": "failure",
+    "24983": "failure",
+    "24984": "failure",
+    "24985": "failure",
+    "24986": "infeasible",
+    "24987": "failure",
+    "24988": "failure",
+    "24989": "failure",
+    "24990": "failure",
+    "24991": "failure",
+    "24992": "failure",
+    "24993": "failure",
+    "24994": "failure",
+    "24995": "failure",
+    "24996": "failure",
+    "24997": "failure",
+    "24998": "failure",
+    "24999": "failure",
+    "25000": "infeasible",
+    "25001": "failure",
+    "25002": "failure",
+    "25003": "failure",
+    "25004": "failure",
+    "25005": "failure",
+    "25006": "failure",
+    "25007": "failure",
+    "25008": "infeasible",
+    "25009": "failure",
+    "25010": "infeasible",
+    "25011": "failure",
+    "25012": "failure",
+    "25013": "failure",
+    "25014": "infeasible",
+    "25015": "failure",
+    "25016": "failure",
+    "25017": "failure",
+    "25018": "infeasible",
+    "25019": "failure",
+    "25020": "failure",
+    "25021": "failure",
+    "25022": "failure",
+    "25023": "failure",
+    "25024": "infeasible",
+    "25025": "failure",
+    "25026": "failure",
+    "25027": "failure",
+    "25028": "failure",
+    "25029": "failure",
+    "25030": "failure",
+    "25031": "failure",
+    "25032": "successful",
+    "25033": "failure",
+    "25034": "infeasible",
+    "25035": "failure",
+    "25036": "successful",
+    "25037": "failure",
+    "25038": "failure",
+    "25039": "failure",
+    "25040": "infeasible",
+    "25041": "failure",
+    "25042": "failure",
+    "25043": "failure",
+    "25044": "failure",
+    "25045": "failure",
+    "25046": "successful",
+    "25047": "infeasible",
+    "25048": "failure",
+    "25049": "failure",
+    "25050": "failure",
+    "25051": "failure",
+    "25052": "already_complete",
+    "25053": "failure",
+    "25054": "failure",
+    "25055": "failure",
+    "25056": "failure",
+    "25057": "infeasible",
+    "25058": "failure",
+    "25059": "failure",
+    "25060": "failure",
+    "25061": "successful",
+    "25062": "failure",
+    "25063": "infeasible",
+    "25064": "failure",
+    "25065": "failure",
+    "25066": "infeasible",
+    "25067": "failure",
+    "25068": "failure",
+    "25069": "failure",
+    "25070": "failure",
+    "25071": "failure",
+    "25072": "failure",
+    "25073": "failure",
+    "25074": "failure",
+    "25075": "failure",
+    "25076": "failure",
+    "25077": "already_complete",
+    "25078": "failure",
+    "25079": "failure",
+    "25080": "failure",
+    "25081": "already_complete",
+    "25082": "failure",
+    "25083": "failure",
+    "25084": "infeasible",
+    "25085": "failure",
+    "25086": "failure",
+    "25087": "failure",
+    "25088": "successful",
+    "25089": "failure",
+    "25090": "failure",
+    "25091": "failure",
+    "25092": "failure",
+    "25093": "successful",
+    "25094": "failure",
+    "25095": "failure",
+    "25096": "failure",
+    "25097": "failure",
+    "25098": "failure",
+    "25099": "failure",
+    "25100": "failure",
+    "25101": "failure",
+    "25102": "failure",
+    "25103": "successful",
+    "25104": "failure",
+    "25105": "failure",
+    "25106": "failure",
+    "25107": "failure",
+    "25108": "failure",
+    "25109": "failure",
+    "25110": "failure",
+    "25111": "failure",
+    "25112": "already_complete",
+    "25113": "successful",
+    "25114": "infeasible",
+    "25115": "failure",
+    "25116": "failure",
+    "25117": "failure",
+    "25118": "failure",
+    "25119": "failure",
+    "25120": "failure",
+    "25121": "failure",
+    "25122": "failure",
+    "25123": "failure",
+    "25124": "failure",
+    "25125": "failure",
+    "25126": "failure",
+    "25127": "failure",
+    "25128": "failure",
+    "25129": "failure",
+    "25130": "failure",
+    "25131": "failure",
+    "25132": "successful",
+    "25133": "failure",
+    "25134": "failure",
+    "25135": "failure",
+    "25136": "successful",
+    "25137": "failure",
+    "25138": "successful",
+    "25139": "failure",
+    "25140": "failure",
+    "25141": "infeasible",
+    "25142": "failure",
+    "25143": "successful",
+    "25144": "failure",
+    "25145": "successful",
+    "25146": "failure",
+    "25147": "successful",
+    "25148": "failure",
+    "25149": "failure",
+    "25150": "failure",
+    "25151": "failure",
+    "25152": "failure",
+    "25153": "failure",
+    "25154": "failure",
+    "25155": "successful",
+    "25156": "failure",
+    "25157": "infeasible",
+    "25158": "infeasible",
+    "25159": "failure",
+    "25160": "failure",
+    "25161": "failure",
+    "25162": "failure",
+    "25163": "failure",
+    "25164": "infeasible",
+    "25165": "failure",
+    "25166": "failure",
+    "25167": "failure",
+    "25168": "failure",
+    "25169": "failure",
+    "25170": "failure",
+    "25171": "failure",
+    "25172": "successful",
+    "25173": "failure",
+    "25174": "failure",
+    "25175": "failure",
+    "25176": "failure",
+    "25177": "failure",
+    "25178": "successful",
+    "25179": "failure",
+    "25180": "infeasible",
+    "25181": "failure",
+    "25182": "failure",
+    "25183": "failure",
+    "25184": "failure",
+    "25185": "failure",
+    "25186": "infeasible",
+    "25187": "failure",
+    "25188": "failure",
+    "25189": "failure",
+    "25190": "failure",
+    "25191": "failure",
+    "25192": "infeasible",
+    "25193": "failure",
+    "25194": "infeasible",
+    "25195": "failure",
+    "25196": "failure",
+    "25197": "failure",
+    "25198": "infeasible",
+    "25199": "failure",
+    "25200": "failure",
+    "25201": "failure",
+    "25202": "infeasible",
+    "25203": "failure",
+    "25204": "infeasible",
+    "25205": "infeasible",
+    "25206": "failure",
+    "25207": "successful",
+    "25208": "infeasible",
+    "25209": "failure",
+    "25210": "failure",
+    "25211": "already_complete",
+    "25212": "successful",
+    "25213": "failure",
+    "25214": "failure",
+    "25215": "failure",
+    "25216": "failure",
+    "25217": "failure",
+    "25218": "failure",
+    "25219": "failure",
+    "25220": "failure",
+    "25221": "failure",
+    "25222": "failure",
+    "25223": "failure",
+    "25224": "infeasible",
+    "25225": "failure",
+    "25226": "failure",
+    "25227": "failure",
+    "25228": "failure",
+    "25229": "successful",
+    "25230": "failure",
+    "25231": "failure",
+    "25232": "failure",
+    "25233": "failure",
+    "25234": "failure",
+    "25235": "failure",
+    "25236": "failure",
+    "25237": "failure",
+    "25238": "failure",
+    "25239": "successful",
+    "25240": "failure",
+    "25241": "failure",
+    "25242": "failure",
+    "25243": "failure",
+    "25244": "successful",
+    "25245": "failure",
+    "25246": "failure",
+    "25247": "failure",
+    "25248": "failure",
+    "25249": "failure",
+    "25250": "failure",
+    "25251": "failure",
+    "25252": "failure",
+    "25253": "successful",
+    "25254": "failure",
+    "25255": "successful",
+    "25256": "infeasible",
+    "25257": "failure",
+    "25258": "failure",
+    "25259": "failure",
+    "25260": "failure",
+    "25261": "failure",
+    "25262": "failure",
+    "25263": "failure",
+    "25264": "failure",
+    "25265": "failure",
+    "25266": "failure",
+    "25267": "failure",
+    "25268": "failure",
+    "25269": "successful",
+    "25270": "failure",
+    "25271": "failure",
+    "25272": "failure",
+    "25273": "failure",
+    "25274": "failure",
+    "25275": "failure",
+    "25276": "failure",
+    "25277": "failure",
+    "25278": "successful",
+    "25279": "failure",
+    "25280": "failure",
+    "25281": "already_complete",
+    "25282": "failure",
+    "25283": "failure",
+    "25284": "failure",
+    "25285": "failure",
+    "25286": "already_complete",
+    "25287": "failure",
+    "25288": "failure",
+    "25289": "infeasible",
+    "25290": "failure",
+    "25291": "failure",
+    "25292": "infeasible",
+    "25293": "failure",
+    "25294": "failure",
+    "25295": "failure",
+    "25296": "failure",
+    "25297": "successful",
+    "25298": "failure",
+    "25299": "successful",
+    "25300": "failure",
+    "25301": "infeasible",
+    "25302": "failure",
+    "25303": "failure",
+    "25304": "infeasible",
+    "25305": "infeasible",
+    "25306": "failure",
+    "25307": "successful",
+    "25308": "failure",
+    "25309": "successful",
+    "25310": "failure",
+    "25311": "failure",
+    "25312": "infeasible",
+    "25313": "failure",
+    "25314": "already_complete",
+    "25315": "successful",
+    "25316": "failure",
+    "25317": "failure",
+    "25318": "successful",
+    "25319": "failure",
+    "25320": "failure",
+    "25321": "failure",
+    "25322": "infeasible",
+    "25323": "failure",
+    "25324": "failure",
+    "25325": "successful",
+    "25326": "failure",
+    "25327": "failure",
+    "25328": "failure",
+    "25329": "failure",
+    "25330": "failure",
+    "25331": "failure",
+    "25332": "failure",
+    "25333": "failure",
+    "25334": "failure",
+    "25335": "failure",
+    "25336": "infeasible",
+    "25337": "failure",
+    "25338": "failure",
+    "25339": "infeasible",
+    "25340": "failure",
+    "25341": "failure",
+    "25342": "failure",
+    "25343": "failure",
+    "25344": "failure",
+    "25345": "successful",
+    "25346": "failure",
+    "25347": "failure",
+    "25348": "failure",
+    "25349": "failure",
+    "25350": "failure",
+    "25351": "failure",
+    "25352": "failure",
+    "25353": "failure",
+    "25354": "failure",
+    "25355": "failure",
+    "25356": "failure",
+    "25357": "failure",
+    "25358": "successful",
+    "25359": "infeasible",
+    "25360": "failure",
+    "25361": "failure",
+    "25362": "failure",
+    "25363": "failure",
+    "25364": "failure",
+    "25365": "failure",
+    "25366": "failure",
+    "25367": "failure",
+    "25368": "failure",
+    "25369": "failure",
+    "25370": "failure",
+    "25371": "failure",
+    "25372": "failure",
+    "25373": "failure",
+    "25374": "failure",
+    "25375": "failure",
+    "25376": "failure",
+    "25377": "failure",
+    "25378": "infeasible",
+    "25379": "successful",
+    "25380": "infeasible",
+    "25381": "successful",
+    "25382": "successful",
+    "25383": "failure",
+    "25384": "failure",
+    "25385": "failure",
+    "25386": "failure",
+    "25387": "infeasible",
+    "25388": "failure",
+    "25389": "failure",
+    "25390": "failure",
+    "25391": "failure",
+    "25392": "failure",
+    "25393": "failure",
+    "25394": "failure",
+    "25395": "failure",
+    "25396": "failure",
+    "25397": "failure",
+    "25398": "failure",
+    "25399": "infeasible",
+    "25400": "failure",
+    "25401": "failure",
+    "25402": "failure",
+    "25403": "failure",
+    "25404": "failure",
+    "25405": "failure",
+    "25406": "failure",
+    "25407": "successful",
+    "25408": "failure",
+    "25409": "failure",
+    "25410": "failure",
+    "25411": "failure",
+    "25412": "failure",
+    "25413": "infeasible",
+    "25414": "successful",
+    "25415": "failure",
+    "25416": "failure",
+    "25417": "failure",
+    "25418": "infeasible",
+    "25419": "failure",
+    "25420": "failure",
+    "25421": "failure",
+    "25422": "failure",
+    "25423": "failure",
+    "25424": "infeasible",
+    "25425": "failure",
+    "25426": "failure",
+    "25427": "failure",
+    "25428": "successful",
+    "25429": "failure",
+    "25430": "failure",
+    "25431": "failure",
+    "25432": "failure",
+    "25433": "failure",
+    "25434": "failure",
+    "25435": "infeasible",
+    "25436": "failure",
+    "25437": "failure",
+    "25438": "failure",
+    "25439": "failure",
+    "25440": "failure",
+    "25441": "failure",
+    "25442": "failure",
+    "25443": "failure",
+    "25444": "failure",
+    "25445": "successful",
+    "25446": "infeasible",
+    "25447": "failure",
+    "25448": "infeasible",
+    "25449": "failure",
+    "25450": "failure",
+    "25451": "successful",
+    "25452": "successful",
+    "25453": "failure",
+    "25454": "failure",
+    "25455": "infeasible",
+    "25456": "failure",
+    "25457": "successful",
+    "25458": "failure",
+    "25459": "failure",
+    "25460": "failure",
+    "25461": "failure",
+    "25462": "failure",
+    "25463": "failure",
+    "25464": "failure",
+    "25465": "failure",
+    "25466": "failure",
+    "25467": "failure",
+    "25468": "successful",
+    "25469": "failure",
+    "25470": "failure",
+    "25471": "failure",
+    "25472": "failure",
+    "25473": "failure",
+    "25474": "failure",
+    "25475": "failure",
+    "25476": "failure",
+    "25477": "failure",
+    "25478": "failure",
+    "25479": "failure",
+    "25480": "failure",
+    "25481": "failure",
+    "25482": "infeasible",
+    "25483": "failure",
+    "25484": "failure",
+    "25485": "failure",
+    "25486": "failure",
+    "25487": "failure",
+    "25488": "failure",
+    "25489": "failure",
+    "25490": "failure",
+    "25491": "failure",
+    "25492": "failure",
+    "25493": "failure",
+    "25494": "failure",
+    "25495": "infeasible",
+    "25496": "failure",
+    "25497": "infeasible",
+    "25498": "failure",
+    "25499": "failure",
+    "25500": "failure",
+    "25501": "failure",
+    "25502": "failure",
+    "25503": "failure",
+    "25504": "failure",
+    "25505": "failure",
+    "25506": "failure",
+    "25507": "failure",
+    "25508": "infeasible",
+    "25509": "failure",
+    "25510": "failure",
+    "25511": "failure",
+    "25512": "failure",
+    "25513": "successful",
+    "25514": "failure",
+    "25515": "failure",
+    "25516": "failure",
+    "25517": "failure",
+    "25518": "failure",
+    "25519": "failure",
+    "25520": "failure",
+    "25521": "failure",
+    "25522": "failure",
+    "25523": "failure",
+    "25524": "failure",
+    "25525": "failure",
+    "25526": "failure",
+    "25527": "failure",
+    "25528": "failure",
+    "25529": "failure",
+    "25530": "failure",
+    "25531": "failure",
+    "25532": "failure",
+    "25533": "successful",
+    "25534": "failure",
+    "25535": "already_complete",
+    "25536": "failure",
+    "25537": "successful",
+    "25538": "failure",
+    "25539": "failure",
+    "25540": "failure",
+    "25541": "failure",
+    "25542": "failure",
+    "25543": "failure",
+    "25544": "failure",
+    "25545": "failure",
+    "25546": "failure",
+    "25547": "failure",
+    "25548": "failure",
+    "25549": "failure",
+    "25550": "failure",
+    "25551": "failure",
+    "25552": "failure",
+    "25553": "failure",
+    "25554": "failure",
+    "25555": "infeasible",
+    "25556": "successful",
+    "25557": "successful",
+    "25558": "successful",
+    "25559": "failure",
+    "25560": "failure",
+    "25561": "failure",
+    "25562": "failure",
+    "25563": "infeasible",
+    "25564": "failure",
+    "25565": "failure",
+    "25566": "failure",
+    "25567": "failure",
+    "25568": "infeasible",
+    "25569": "failure",
+    "25570": "successful",
+    "25571": "failure",
+    "25572": "failure",
+    "25573": "failure",
+    "25574": "failure",
+    "25575": "failure",
+    "25576": "successful",
+    "25577": "failure",
+    "25578": "failure",
+    "25579": "failure",
+    "25580": "failure",
+    "25581": "failure",
+    "25582": "infeasible",
+    "25583": "failure",
+    "25584": "infeasible",
+    "25585": "failure",
+    "25586": "failure",
+    "25587": "failure",
+    "25588": "infeasible",
+    "25589": "failure",
+    "25590": "failure",
+    "25591": "infeasible",
+    "25592": "failure",
+    "25593": "failure",
+    "25594": "failure",
+    "25595": "infeasible",
+    "25596": "failure",
+    "25597": "failure",
+    "25598": "failure",
+    "25599": "infeasible",
+    "25600": "failure",
+    "25601": "successful",
+    "25602": "infeasible",
+    "25603": "failure",
+    "25604": "failure",
+    "25605": "failure",
+    "25606": "successful",
+    "25607": "failure",
+    "25608": "failure",
+    "25609": "failure",
+    "25610": "infeasible",
+    "25611": "failure",
+    "25612": "failure",
+    "25613": "failure",
+    "25614": "failure",
+    "25615": "failure",
+    "25616": "successful",
+    "25617": "infeasible",
+    "25618": "failure",
+    "25619": "failure",
+    "25620": "failure",
+    "25621": "failure",
+    "25622": "failure",
+    "25623": "successful",
+    "25624": "failure",
+    "25625": "failure",
+    "25626": "infeasible",
+    "25627": "successful",
+    "25628": "successful",
+    "25629": "failure",
+    "25630": "failure",
+    "25631": "failure",
+    "25632": "failure",
+    "25633": "failure",
+    "25634": "failure",
+    "25635": "successful",
+    "25636": "failure",
+    "25637": "failure",
+    "25638": "successful",
+    "25639": "failure",
+    "25640": "infeasible",
+    "25641": "failure",
+    "25642": "failure",
+    "25643": "infeasible",
+    "25644": "failure",
+    "25645": "failure",
+    "25646": "failure",
+    "25647": "failure",
+    "25648": "successful",
+    "25649": "successful",
+    "25650": "failure",
+    "25651": "infeasible",
+    "25652": "failure",
+    "25653": "failure",
+    "25654": "failure",
+    "25655": "failure",
+    "25656": "infeasible",
+    "25657": "failure",
+    "25658": "failure",
+    "25659": "failure",
+    "25660": "failure",
+    "25661": "failure",
+    "25662": "failure",
+    "25663": "failure",
+    "25664": "failure",
+    "25665": "failure",
+    "25666": "infeasible",
+    "25667": "failure",
+    "25668": "infeasible",
+    "25669": "infeasible",
+    "25670": "failure",
+    "25671": "failure",
+    "25672": "failure",
+    "25673": "successful",
+    "25674": "failure",
+    "25675": "successful",
+    "25676": "failure",
+    "25677": "failure",
+    "25678": "failure",
+    "25679": "failure",
+    "25680": "failure",
+    "25681": "failure",
+    "25682": "failure",
+    "25683": "failure",
+    "25684": "failure",
+    "25685": "failure",
+    "25686": "failure",
+    "25687": "failure",
+    "25688": "failure",
+    "25689": "failure",
+    "25690": "infeasible",
+    "25691": "failure",
+    "25692": "failure",
+    "25693": "failure",
+    "25694": "failure",
+    "25695": "failure",
+    "25696": "successful",
+    "25697": "failure",
+    "25698": "failure",
+    "25699": "failure",
+    "25700": "infeasible",
+    "25701": "failure",
+    "25702": "failure",
+    "25703": "failure",
+    "25704": "failure",
+    "25705": "failure",
+    "25706": "successful",
+    "25707": "successful",
+    "25708": "failure",
+    "25709": "failure",
+    "25710": "successful",
+    "25711": "failure",
+    "25712": "failure",
+    "25713": "failure",
+    "25714": "failure",
+    "25715": "failure",
+    "25716": "failure",
+    "25717": "infeasible",
+    "25718": "failure",
+    "25719": "successful",
+    "25720": "failure",
+    "25721": "failure",
+    "25722": "failure",
+    "25723": "failure",
+    "25724": "successful",
+    "25725": "failure",
+    "25726": "failure",
+    "25727": "failure",
+    "25728": "failure",
+    "25729": "already_complete",
+    "25730": "failure",
+    "25731": "successful",
+    "25732": "failure",
+    "25733": "failure",
+    "25734": "failure",
+    "25735": "failure",
+    "25736": "failure",
+    "25737": "failure",
+    "25738": "failure",
+    "25739": "successful",
+    "25740": "failure",
+    "25741": "successful",
+    "25742": "failure",
+    "25743": "successful",
+    "25744": "successful",
+    "25745": "failure",
+    "25746": "successful",
+    "25747": "failure",
+    "25748": "failure",
+    "25749": "failure",
+    "25750": "failure",
+    "25751": "failure",
+    "25752": "failure",
+    "25753": "failure",
+    "25754": "infeasible",
+    "25755": "infeasible",
+    "25756": "failure",
+    "25757": "infeasible",
+    "25758": "failure",
+    "25759": "failure",
+    "25760": "infeasible",
+    "25761": "failure",
+    "25762": "failure",
+    "25763": "failure",
+    "25764": "failure",
+    "25765": "failure",
+    "25766": "failure",
+    "25767": "successful",
+    "25768": "failure",
+    "25769": "failure",
+    "25770": "failure",
+    "25771": "failure",
+    "25772": "failure",
+    "25773": "failure",
+    "25774": "failure",
+    "25775": "failure",
+    "25776": "failure",
+    "25777": "successful",
+    "25778": "failure",
+    "25779": "failure",
+    "25780": "successful",
+    "25781": "failure",
+    "25782": "successful",
+    "25783": "failure",
+    "25784": "successful",
+    "25785": "successful",
+    "25786": "failure",
+    "25787": "already_complete",
+    "25788": "failure",
+    "25789": "failure",
+    "25790": "failure",
+    "25791": "failure",
+    "25792": "infeasible",
+    "25793": "successful",
+    "25794": "successful",
+    "25795": "failure",
+    "25796": "failure",
+    "25797": "successful",
+    "25798": "failure",
+    "25799": "successful",
+    "25800": "failure",
+    "25801": "failure",
+    "25802": "failure",
+    "25803": "failure",
+    "25804": "failure",
+    "25805": "failure",
+    "25806": "failure",
+    "25807": "failure",
+    "25808": "successful",
+    "25809": "failure",
+    "25810": "failure",
+    "25811": "failure",
+    "25812": "failure",
+    "25813": "failure",
+    "25814": "failure",
+    "25815": "failure",
+    "25816": "failure",
+    "25817": "failure",
+    "25818": "failure",
+    "25819": "failure",
+    "25820": "failure",
+    "25821": "infeasible",
+    "25822": "failure",
+    "25823": "failure",
+    "25824": "failure",
+    "25825": "failure",
+    "25826": "failure",
+    "25827": "failure",
+    "25828": "successful",
+    "25829": "failure",
+    "25830": "failure",
+    "25831": "failure",
+    "25832": "failure",
+    "25833": "failure",
+    "25834": "failure",
+    "25835": "failure",
+    "25836": "failure",
+    "25837": "failure",
+    "25838": "failure",
+    "25839": "infeasible",
+    "25840": "failure",
+    "25841": "failure",
+    "25842": "infeasible",
+    "25843": "successful",
+    "25844": "successful",
+    "25845": "failure",
+    "25846": "failure",
+    "25847": "failure",
+    "25848": "failure",
+    "25849": "failure",
+    "25850": "failure",
+    "25851": "failure",
+    "25852": "failure",
+    "25853": "failure",
+    "25854": "failure",
+    "25855": "failure",
+    "25856": "failure",
+    "25857": "failure",
+    "25858": "failure",
+    "25859": "failure",
+    "25860": "failure",
+    "25861": "failure",
+    "25862": "failure",
+    "25863": "infeasible",
+    "25864": "failure",
+    "25865": "failure",
+    "25866": "failure",
+    "25867": "failure",
+    "25868": "successful",
+    "25869": "failure",
+    "25870": "infeasible",
+    "25871": "failure",
+    "25872": "failure",
+    "25873": "failure",
+    "25874": "successful",
+    "25875": "failure",
+    "25876": "failure",
+    "25877": "failure",
+    "25878": "successful",
+    "25879": "failure",
+    "25880": "failure",
+    "25881": "failure",
+    "25882": "failure",
+    "25883": "failure",
+    "25884": "failure",
+    "25885": "failure",
+    "25886": "failure",
+    "25887": "already_complete",
+    "25888": "failure",
+    "25889": "infeasible",
+    "25890": "failure",
+    "25891": "failure",
+    "25892": "failure",
+    "25893": "successful",
+    "25894": "infeasible",
+    "25895": "failure",
+    "25896": "failure",
+    "25897": "failure",
+    "25898": "failure",
+    "25899": "failure",
+    "25900": "failure",
+    "25901": "failure",
+    "25902": "failure",
+    "25903": "failure",
+    "25904": "failure",
+    "25905": "infeasible",
+    "25906": "failure",
+    "25907": "successful",
+    "25908": "failure",
+    "25909": "failure",
+    "25910": "failure",
+    "25911": "infeasible",
+    "25912": "successful",
+    "25913": "failure",
+    "25914": "failure",
+    "25915": "infeasible",
+    "25916": "failure",
+    "25917": "failure",
+    "25918": "failure",
+    "25919": "successful",
+    "25920": "failure",
+    "25921": "infeasible",
+    "25922": "failure",
+    "25923": "failure",
+    "25924": "failure",
+    "25925": "failure",
+    "25926": "failure",
+    "25927": "failure",
+    "25928": "failure",
+    "25929": "failure",
+    "25930": "successful",
+    "25931": "failure",
+    "25932": "failure",
+    "25933": "failure",
+    "25934": "successful",
+    "25935": "failure",
+    "25936": "failure",
+    "25937": "failure",
+    "25938": "failure",
+    "25939": "failure",
+    "25940": "infeasible",
+    "25941": "infeasible",
+    "25942": "failure",
+    "25943": "failure",
+    "25944": "failure",
+    "25945": "failure",
+    "25946": "infeasible",
+    "25947": "failure",
+    "25948": "failure",
+    "25949": "failure",
+    "25950": "successful",
+    "25951": "failure",
+    "25952": "successful",
+    "25953": "infeasible",
+    "25954": "infeasible",
+    "25955": "failure",
+    "25956": "failure",
+    "25957": "failure",
+    "25958": "failure",
+    "25959": "failure",
+    "25960": "failure",
+    "25961": "failure",
+    "25962": "failure",
+    "25963": "failure",
+    "25964": "infeasible",
+    "25965": "failure",
+    "25966": "failure",
+    "25967": "failure",
+    "25968": "failure",
+    "25969": "failure",
+    "25970": "failure",
+    "25971": "failure",
+    "25972": "failure",
+    "25973": "failure",
+    "25974": "failure",
+    "25975": "failure",
+    "25976": "failure",
+    "25977": "failure",
+    "25978": "infeasible",
+    "25979": "failure",
+    "25980": "failure",
+    "25981": "infeasible",
+    "25982": "failure",
+    "25983": "failure",
+    "25984": "successful",
+    "25985": "failure",
+    "25986": "infeasible",
+    "25987": "failure",
+    "25988": "failure",
+    "25989": "failure",
+    "25990": "failure",
+    "25991": "failure",
+    "25992": "failure",
+    "25993": "failure",
+    "25994": "infeasible",
+    "25995": "failure",
+    "25996": "failure",
+    "25997": "failure",
+    "25998": "failure",
+    "25999": "failure",
+    "26000": "failure",
+    "26001": "failure",
+    "26002": "failure",
+    "26003": "failure",
+    "26004": "failure",
+    "26005": "failure",
+    "26006": "failure",
+    "26007": "successful",
+    "26008": "infeasible",
+    "26009": "successful",
+    "26010": "failure",
+    "26011": "successful",
+    "26012": "failure",
+    "26013": "successful",
+    "26014": "failure",
+    "26015": "failure",
+    "26016": "failure",
+    "26017": "failure",
+    "26018": "failure",
+    "26019": "failure",
+    "26020": "failure",
+    "26021": "failure",
+    "26022": "failure",
+    "26023": "failure",
+    "26024": "infeasible",
+    "26025": "infeasible",
+    "26026": "successful",
+    "26027": "failure",
+    "26028": "failure",
+    "26029": "failure",
+    "26030": "failure",
+    "26031": "failure",
+    "26032": "failure",
+    "26033": "failure",
+    "26034": "failure",
+    "26035": "failure",
+    "26036": "failure",
+    "26037": "failure",
+    "26038": "failure",
+    "26039": "failure",
+    "26040": "failure",
+    "26041": "failure",
+    "26042": "failure",
+    "26043": "failure",
+    "26044": "successful",
+    "26045": "failure",
+    "26046": "failure",
+    "26047": "failure",
+    "26048": "failure",
+    "26049": "failure",
+    "26050": "failure",
+    "26051": "infeasible",
+    "26052": "infeasible",
+    "26053": "infeasible",
+    "26054": "failure",
+    "26055": "failure",
+    "26056": "failure",
+    "26057": "failure",
+    "26058": "successful",
+    "26059": "failure",
+    "26060": "failure",
+    "26061": "failure",
+    "26062": "infeasible",
+    "26063": "failure",
+    "26064": "failure",
+    "26065": "successful",
+    "26066": "successful",
+    "26067": "failure",
+    "26068": "failure",
+    "26069": "failure",
+    "26070": "infeasible",
+    "26071": "failure",
+    "26072": "infeasible",
+    "26073": "failure",
+    "26074": "failure",
+    "26075": "failure",
+    "26076": "failure",
+    "26077": "already_complete",
+    "26078": "successful",
+    "26079": "failure",
+    "26080": "successful",
+    "26081": "failure",
+    "26082": "infeasible",
+    "26083": "failure",
+    "26084": "failure",
+    "26085": "failure",
+    "26086": "failure",
+    "26087": "failure",
+    "26088": "failure",
+    "26089": "failure",
+    "26090": "failure",
+    "26091": "failure",
+    "26092": "failure",
+    "26093": "failure",
+    "26094": "failure",
+    "26095": "failure",
+    "26096": "failure",
+    "26097": "failure",
+    "26098": "failure",
+    "26099": "failure",
+    "26100": "failure",
+    "26101": "failure",
+    "26102": "successful",
+    "26103": "failure",
+    "26104": "failure",
+    "26105": "failure",
+    "26106": "failure",
+    "26107": "failure",
+    "26108": "failure",
+    "26109": "failure",
+    "26110": "failure",
+    "26111": "failure",
+    "26112": "infeasible",
+    "26113": "infeasible",
+    "26114": "failure",
+    "26115": "successful",
+    "26116": "failure",
+    "26117": "failure",
+    "26118": "failure",
+    "26119": "successful",
+    "26120": "failure",
+    "26121": "failure",
+    "26122": "failure",
+    "26123": "failure",
+    "26124": "failure",
+    "26125": "failure",
+    "26126": "failure",
+    "26127": "failure",
+    "26128": "failure",
+    "26129": "failure",
+    "26130": "successful",
+    "26131": "infeasible",
+    "26132": "failure",
+    "26133": "failure",
+    "26134": "failure",
+    "26135": "infeasible",
+    "26136": "failure",
+    "26137": "failure",
+    "26138": "failure",
+    "26139": "failure",
+    "26140": "failure",
+    "26141": "failure",
+    "26142": "failure",
+    "26143": "successful",
+    "26144": "failure",
+    "26145": "failure",
+    "26146": "failure",
+    "26147": "failure",
+    "26148": "failure",
+    "26149": "failure",
+    "26150": "failure",
+    "26151": "failure",
+    "26152": "successful",
+    "26153": "failure",
+    "26154": "successful",
+    "26155": "failure",
+    "26156": "failure",
+    "26157": "failure",
+    "26158": "infeasible",
+    "26159": "failure",
+    "26160": "failure",
+    "26161": "failure",
+    "26162": "failure",
+    "26163": "failure",
+    "26164": "successful",
+    "26165": "failure",
+    "26166": "infeasible",
+    "26167": "successful",
+    "26168": "failure",
+    "26169": "infeasible",
+    "26170": "failure",
+    "26171": "failure",
+    "26172": "failure",
+    "26173": "failure",
+    "26174": "failure",
+    "26175": "failure",
+    "26176": "failure",
+    "26177": "successful",
+    "26178": "infeasible",
+    "26179": "failure",
+    "26180": "failure",
+    "26181": "successful",
+    "26182": "failure",
+    "26183": "failure",
+    "26184": "failure",
+    "26185": "failure",
+    "26186": "failure",
+    "26187": "failure",
+    "26188": "successful",
+    "26189": "failure",
+    "26190": "failure",
+    "26191": "failure",
+    "26192": "failure",
+    "26193": "failure",
+    "26194": "failure",
+    "26195": "successful",
+    "26196": "failure",
+    "26197": "failure",
+    "26198": "infeasible",
+    "26199": "failure",
+    "26200": "failure",
+    "26201": "failure",
+    "26202": "successful",
+    "26203": "failure",
+    "26204": "failure",
+    "26205": "failure",
+    "26206": "failure",
+    "26207": "failure",
+    "26208": "infeasible",
+    "26209": "successful",
+    "26210": "failure",
+    "26211": "already_complete",
+    "26212": "failure",
+    "26213": "failure",
+    "26214": "failure",
+    "26215": "failure",
+    "26216": "failure",
+    "26217": "infeasible",
+    "26218": "failure",
+    "26219": "failure",
+    "26220": "failure",
+    "26221": "failure",
+    "26222": "failure",
+    "26223": "failure",
+    "26224": "successful",
+    "26225": "failure",
+    "26226": "failure",
+    "26227": "infeasible",
+    "26228": "successful",
+    "26229": "failure",
+    "26230": "failure",
+    "26231": "failure",
+    "26232": "failure",
+    "26233": "failure",
+    "26234": "failure",
+    "26235": "failure",
+    "26236": "infeasible",
+    "26237": "failure",
+    "26238": "failure",
+    "26239": "failure",
+    "26240": "failure",
+    "26241": "failure",
+    "26242": "successful",
+    "26243": "failure",
+    "26244": "successful",
+    "26245": "failure",
+    "26246": "failure",
+    "26247": "failure",
+    "26248": "failure",
+    "26249": "failure",
+    "26250": "failure",
+    "26251": "failure",
+    "26252": "failure",
+    "26253": "failure",
+    "26254": "failure",
+    "26255": "failure",
+    "26256": "failure",
+    "26257": "failure",
+    "26258": "failure",
+    "26259": "failure",
+    "26260": "failure",
+    "26261": "failure",
+    "26262": "failure",
+    "26263": "failure",
+    "26264": "failure",
+    "26265": "failure",
+    "26266": "infeasible",
+    "26267": "failure",
+    "26268": "failure",
+    "26269": "failure",
+    "26270": "failure",
+    "26271": "failure",
+    "26272": "failure",
+    "26273": "infeasible",
+    "26274": "failure",
+    "26275": "failure",
+    "26276": "successful",
+    "26277": "failure",
+    "26278": "failure",
+    "26279": "failure",
+    "26280": "failure",
+    "26281": "failure",
+    "26282": "failure",
+    "26283": "failure",
+    "26284": "successful",
+    "26285": "infeasible",
+    "26286": "infeasible",
+    "26287": "failure",
+    "26288": "failure",
+    "26289": "failure",
+    "26290": "failure",
+    "26291": "failure",
+    "26292": "failure",
+    "26293": "successful",
+    "26294": "failure",
+    "26295": "failure",
+    "26296": "successful",
+    "26297": "failure",
+    "26298": "failure",
+    "26299": "infeasible",
+    "26300": "failure",
+    "26301": "failure",
+    "26302": "infeasible",
+    "26303": "failure",
+    "26304": "failure",
+    "26305": "failure",
+    "26306": "failure",
+    "26307": "failure",
+    "26308": "successful",
+    "26309": "failure",
+    "26310": "infeasible",
+    "26311": "failure",
+    "26312": "failure",
+    "26313": "failure",
+    "26314": "failure",
+    "26315": "failure",
+    "26316": "failure",
+    "26317": "failure",
+    "26318": "failure",
+    "26319": "failure",
+    "26320": "failure",
+    "26321": "successful",
+    "26322": "failure",
+    "26323": "failure",
+    "26324": "failure",
+    "26325": "failure",
+    "26326": "failure",
+    "26327": "successful",
+    "26328": "failure",
+    "26329": "failure",
+    "26330": "failure",
+    "26331": "failure",
+    "26332": "failure",
+    "26333": "failure",
+    "26334": "infeasible",
+    "26335": "failure",
+    "26336": "failure",
+    "26337": "successful",
+    "26338": "failure",
+    "26339": "failure",
+    "26340": "failure",
+    "26341": "failure",
+    "26342": "failure",
+    "26343": "failure",
+    "26344": "infeasible",
+    "26345": "failure",
+    "26346": "successful",
+    "26347": "infeasible",
+    "26348": "failure",
+    "26349": "failure",
+    "26350": "failure",
+    "26351": "failure",
+    "26352": "failure",
+    "26353": "successful",
+    "26354": "failure",
+    "26355": "failure",
+    "26356": "failure",
+    "26357": "failure",
+    "26358": "failure",
+    "26359": "failure",
+    "26360": "infeasible",
+    "26361": "failure",
+    "26362": "failure",
+    "26363": "failure",
+    "26364": "infeasible",
+    "26365": "infeasible",
+    "26366": "failure",
+    "26367": "failure",
+    "26368": "successful",
+    "26369": "failure",
+    "26370": "infeasible",
+    "26371": "failure",
+    "26372": "failure",
+    "26373": "failure",
+    "26374": "failure",
+    "26375": "failure",
+    "26376": "failure",
+    "26377": "infeasible",
+    "26378": "failure",
+    "26379": "failure",
+    "26380": "failure",
+    "26381": "failure",
+    "26382": "failure",
+    "26383": "failure",
+    "26384": "failure",
+    "26385": "failure",
+    "26386": "failure",
+    "26387": "failure",
+    "26388": "failure",
+    "26389": "failure",
+    "26390": "failure",
+    "26391": "failure",
+    "26392": "failure",
+    "26393": "failure",
+    "26394": "failure",
+    "26395": "failure",
+    "26396": "failure",
+    "26397": "infeasible",
+    "26398": "infeasible",
+    "26399": "failure",
+    "26400": "failure",
+    "26401": "infeasible",
+    "26402": "failure",
+    "26403": "failure",
+    "26404": "failure",
+    "26405": "infeasible",
+    "26406": "failure",
+    "26407": "successful",
+    "26408": "failure",
+    "26409": "failure",
+    "26410": "infeasible",
+    "26411": "failure",
+    "26412": "failure",
+    "26413": "failure",
+    "26414": "failure",
+    "26415": "failure",
+    "26416": "failure",
+    "26417": "failure",
+    "26418": "successful",
+    "26419": "failure",
+    "26420": "successful",
+    "26421": "successful",
+    "26422": "failure",
+    "26423": "failure",
+    "26424": "successful",
+    "26425": "failure",
+    "26426": "failure",
+    "26427": "successful",
+    "26428": "failure",
+    "26429": "failure",
+    "26430": "failure",
+    "26431": "failure",
+    "26432": "failure",
+    "26433": "failure",
+    "26434": "infeasible",
+    "26435": "failure",
+    "26436": "successful",
+    "26437": "failure",
+    "26438": "failure",
+    "26439": "failure",
+    "26440": "failure",
+    "26441": "failure",
+    "26442": "failure",
+    "26443": "failure",
+    "26444": "failure",
+    "26445": "successful",
+    "26446": "failure",
+    "26447": "successful",
+    "26448": "failure",
+    "26449": "failure",
+    "26450": "failure",
+    "26451": "failure",
+    "26452": "failure",
+    "26453": "failure",
+    "26454": "failure",
+    "26455": "successful",
+    "26456": "failure",
+    "26457": "failure",
+    "26458": "failure",
+    "26459": "failure",
+    "26460": "infeasible",
+    "26461": "failure",
+    "26462": "infeasible",
+    "26463": "failure",
+    "26464": "failure",
+    "26465": "failure",
+    "26466": "failure",
+    "26467": "failure",
+    "26468": "failure",
+    "26469": "failure",
+    "26470": "infeasible",
+    "26471": "infeasible",
+    "26472": "failure",
+    "26473": "failure",
+    "26474": "failure",
+    "26475": "infeasible",
+    "26476": "infeasible",
+    "26477": "failure",
+    "26478": "failure",
+    "26479": "failure",
+    "26480": "failure",
+    "26481": "failure",
+    "26482": "infeasible",
+    "26483": "failure",
+    "26484": "failure",
+    "26485": "infeasible",
+    "26486": "failure",
+    "26487": "successful",
+    "26488": "failure",
+    "26489": "failure",
+    "26490": "failure",
+    "26491": "failure",
+    "26492": "failure",
+    "26493": "failure",
+    "26494": "failure",
+    "26495": "failure",
+    "26496": "failure",
+    "26497": "failure",
+    "26498": "failure",
+    "26499": "failure",
+    "26500": "failure",
+    "26501": "failure",
+    "26502": "failure",
+    "26503": "failure",
+    "26504": "failure",
+    "26505": "failure",
+    "26506": "failure",
+    "26507": "failure",
+    "26508": "failure",
+    "26509": "failure",
+    "26510": "failure",
+    "26511": "successful",
+    "26512": "infeasible",
+    "26513": "failure",
+    "26514": "successful",
+    "26515": "successful",
+    "26516": "failure",
+    "26517": "failure",
+    "26518": "failure",
+    "26519": "failure",
+    "26520": "failure",
+    "26521": "failure",
+    "26522": "failure",
+    "26523": "failure",
+    "26524": "failure",
+    "26525": "successful",
+    "26526": "failure",
+    "26527": "failure",
+    "26528": "failure",
+    "26529": "failure",
+    "26530": "failure",
+    "26531": "failure",
+    "26532": "failure",
+    "26533": "infeasible",
+    "26534": "failure",
+    "26535": "failure",
+    "26536": "already_complete",
+    "26537": "failure",
+    "26538": "infeasible",
+    "26539": "failure",
+    "26540": "infeasible",
+    "26541": "infeasible",
+    "26542": "infeasible",
+    "26543": "failure",
+    "26544": "successful",
+    "26545": "failure",
+    "26546": "failure",
+    "26547": "failure",
+    "26548": "failure",
+    "26549": "infeasible",
+    "26550": "failure",
+    "26551": "failure",
+    "26552": "infeasible",
+    "26553": "failure",
+    "26554": "failure",
+    "26555": "failure",
+    "26556": "failure",
+    "26557": "failure",
+    "26558": "failure",
+    "26559": "successful",
+    "26560": "failure",
+    "26561": "failure",
+    "26562": "infeasible",
+    "26563": "successful",
+    "26564": "already_complete",
+    "26565": "successful",
+    "26566": "failure",
+    "26567": "successful",
+    "26568": "failure",
+    "26569": "failure",
+    "26570": "failure",
+    "26571": "failure",
+    "26572": "failure",
+    "26573": "failure",
+    "26574": "failure",
+    "26575": "failure",
+    "26576": "failure",
+    "26577": "failure",
+    "26578": "failure",
+    "26579": "failure",
+    "26580": "failure",
+    "26581": "failure",
+    "26582": "failure",
+    "26583": "failure",
+    "26584": "failure",
+    "26585": "failure",
+    "26586": "failure",
+    "26587": "failure",
+    "26588": "successful",
+    "26589": "failure",
+    "26590": "failure",
+    "26591": "failure",
+    "26592": "failure",
+    "26593": "failure",
+    "26594": "failure",
+    "26595": "infeasible",
+    "26596": "failure",
+    "26597": "failure",
+    "26598": "failure",
+    "26599": "successful",
+    "26600": "failure",
+    "26601": "failure",
+    "26602": "failure",
+    "26603": "failure",
+    "26604": "failure",
+    "26605": "failure",
+    "26606": "failure",
+    "26607": "failure",
+    "26608": "failure",
+    "26609": "failure",
+    "26610": "successful",
+    "26611": "failure",
+    "26612": "failure",
+    "26613": "failure",
+    "26614": "failure",
+    "26615": "successful",
+    "26616": "failure",
+    "26617": "failure",
+    "26618": "failure",
+    "26619": "failure",
+    "26620": "successful",
+    "26621": "failure",
+    "26622": "failure",
+    "26623": "failure",
+    "26624": "infeasible",
+    "26625": "infeasible",
+    "26626": "failure",
+    "26627": "failure",
+    "26628": "failure",
+    "26629": "failure",
+    "26630": "failure",
+    "26631": "infeasible",
+    "26632": "failure",
+    "26633": "failure",
+    "26634": "failure",
+    "26635": "failure",
+    "26636": "failure",
+    "26637": "failure",
+    "26638": "infeasible",
+    "26639": "failure",
+    "26640": "failure",
+    "26641": "failure",
+    "26642": "failure",
+    "26643": "failure",
+    "26644": "successful",
+    "26645": "failure",
+    "26646": "failure",
+    "26647": "infeasible",
+    "26648": "successful",
+    "26649": "failure",
+    "26650": "failure",
+    "26651": "successful",
+    "26652": "failure",
+    "26653": "successful",
+    "26654": "successful",
+    "26655": "failure",
+    "26656": "failure",
+    "26657": "failure",
+    "26658": "failure",
+    "26659": "failure",
+    "26660": "failure",
+    "26661": "failure",
+    "26662": "failure",
+    "26663": "failure",
+    "26664": "infeasible",
+    "26665": "failure",
+    "26666": "successful",
+    "26667": "infeasible",
+    "26668": "infeasible",
+    "26669": "successful",
+    "26670": "failure",
+    "26671": "successful",
+    "26672": "failure",
+    "26673": "failure",
+    "26674": "failure",
+    "26675": "failure",
+    "26676": "failure",
+    "26677": "failure",
+    "26678": "failure",
+    "26679": "successful",
+    "26680": "failure",
+    "26681": "failure",
+    "26682": "failure",
+    "26683": "failure",
+    "26684": "failure",
+    "26685": "failure",
+    "26686": "failure",
+    "26687": "successful",
+    "26688": "failure",
+    "26689": "successful",
+    "26690": "failure",
+    "26691": "failure",
+    "26692": "failure",
+    "26693": "infeasible",
+    "26694": "failure",
+    "26695": "failure",
+    "26696": "failure",
+    "26697": "failure",
+    "26698": "failure",
+    "26699": "failure",
+    "26700": "already_complete",
+    "26701": "infeasible",
+    "26702": "failure",
+    "26703": "failure",
+    "26704": "failure",
+    "26705": "failure",
+    "26706": "failure",
+    "26707": "infeasible",
+    "26708": "failure",
+    "26709": "failure",
+    "26710": "infeasible",
+    "26711": "infeasible",
+    "26712": "infeasible",
+    "26713": "failure",
+    "26714": "successful",
+    "26715": "failure",
+    "26716": "failure",
+    "26717": "successful",
+    "26718": "failure",
+    "26719": "failure",
+    "26720": "failure",
+    "26721": "failure",
+    "26722": "failure",
+    "26723": "failure",
+    "26724": "failure",
+    "26725": "successful",
+    "26726": "failure",
+    "26727": "failure",
+    "26728": "failure",
+    "26729": "failure",
+    "26730": "successful",
+    "26731": "failure",
+    "26732": "successful",
+    "26733": "failure",
+    "26734": "failure",
+    "26735": "failure",
+    "26736": "successful",
+    "26737": "failure",
+    "26738": "failure",
+    "26739": "failure",
+    "26740": "failure",
+    "26741": "successful",
+    "26742": "failure",
+    "26743": "failure",
+    "26744": "failure",
+    "26745": "failure",
+    "26746": "successful",
+    "26747": "failure",
+    "26748": "failure",
+    "26749": "failure",
+    "26750": "failure",
+    "26751": "failure",
+    "26752": "failure",
+    "26753": "failure",
+    "26754": "successful",
+    "26755": "failure",
+    "26756": "failure",
+    "26757": "failure",
+    "26758": "infeasible",
+    "26759": "failure",
+    "26760": "failure",
+    "26761": "failure",
+    "26762": "failure",
+    "26763": "failure",
+    "26764": "failure",
+    "26765": "failure",
+    "26766": "failure",
+    "26767": "failure",
+    "26768": "infeasible",
+    "26769": "failure",
+    "26770": "failure",
+    "26771": "successful",
+    "26772": "successful",
+    "26773": "failure",
+    "26774": "failure",
+    "26775": "failure",
+    "26776": "infeasible",
+    "26777": "successful",
+    "26778": "failure",
+    "26779": "failure",
+    "26780": "failure",
+    "26781": "failure",
+    "26782": "failure",
+    "26783": "failure",
+    "26784": "failure",
+    "26785": "successful",
+    "26786": "failure",
+    "26787": "failure",
+    "26788": "failure",
+    "26789": "failure",
+    "26790": "infeasible",
+    "26791": "failure",
+    "26792": "failure",
+    "26793": "failure",
+    "26794": "failure",
+    "26795": "failure",
+    "26796": "failure",
+    "26797": "failure",
+    "26798": "failure",
+    "26799": "failure",
+    "26800": "failure",
+    "26801": "failure",
+    "26802": "failure",
+    "26803": "failure",
+    "26804": "failure",
+    "26805": "failure",
+    "26806": "infeasible",
+    "26807": "failure",
+    "26808": "failure",
+    "26809": "failure",
+    "26810": "failure",
+    "26811": "failure",
+    "26812": "failure",
+    "26813": "failure",
+    "26814": "successful",
+    "26815": "failure",
+    "26816": "failure",
+    "26817": "failure",
+    "26818": "failure",
+    "26819": "failure",
+    "26820": "failure",
+    "26821": "failure",
+    "26822": "failure",
+    "26823": "failure",
+    "26824": "failure",
+    "26825": "infeasible",
+    "26826": "failure",
+    "26827": "failure",
+    "26828": "failure",
+    "26829": "failure",
+    "26830": "infeasible",
+    "26831": "failure",
+    "26832": "failure",
+    "26833": "failure",
+    "26834": "failure",
+    "26835": "failure",
+    "26836": "failure",
+    "26837": "successful",
+    "26838": "infeasible",
+    "26839": "failure",
+    "26840": "failure",
+    "26841": "failure",
+    "26842": "failure",
+    "26843": "failure",
+    "26844": "failure",
+    "26845": "successful",
+    "26846": "successful",
+    "26847": "failure",
+    "26848": "failure",
+    "26849": "failure",
+    "26850": "failure",
+    "26851": "failure",
+    "26852": "failure",
+    "26853": "failure",
+    "26854": "failure",
+    "26855": "failure",
+    "26856": "failure",
+    "26857": "already_complete",
+    "26858": "failure",
+    "26859": "failure",
+    "26860": "infeasible",
+    "26861": "successful",
+    "26862": "failure",
+    "26863": "failure",
+    "26864": "infeasible",
+    "26865": "failure",
+    "26866": "failure",
+    "26867": "failure",
+    "26868": "failure",
+    "26869": "failure",
+    "26870": "failure",
+    "26871": "failure",
+    "26872": "failure",
+    "26873": "failure",
+    "26874": "failure",
+    "26875": "failure",
+    "26876": "failure",
+    "26877": "failure",
+    "26878": "failure",
+    "26879": "failure",
+    "26880": "failure",
+    "26881": "failure",
+    "26882": "successful",
+    "26883": "failure",
+    "26884": "failure",
+    "26885": "failure",
+    "26886": "failure",
+    "26887": "successful",
+    "26888": "failure",
+    "26889": "failure",
+    "26890": "failure",
+    "26891": "failure",
+    "26892": "failure",
+    "26893": "failure",
+    "26894": "failure",
+    "26895": "failure",
+    "26896": "infeasible",
+    "26897": "failure",
+    "26898": "failure",
+    "26899": "failure",
+    "26900": "failure",
+    "26901": "successful",
+    "26902": "failure",
+    "26903": "failure",
+    "26904": "failure",
+    "26905": "successful",
+    "26906": "failure",
+    "26907": "failure",
+    "26908": "failure",
+    "26909": "failure",
+    "26910": "failure",
+    "26911": "failure",
+    "26912": "failure",
+    "26913": "failure",
+    "26914": "failure",
+    "26915": "failure",
+    "26916": "infeasible",
+    "26917": "failure",
+    "26918": "successful",
+    "26919": "failure",
+    "26920": "successful",
+    "26921": "successful",
+    "26922": "failure",
+    "26923": "failure",
+    "26924": "failure",
+    "26925": "failure",
+    "26926": "failure",
+    "26927": "failure",
+    "26928": "failure",
+    "26929": "failure",
+    "26930": "successful",
+    "26931": "failure",
+    "26932": "infeasible",
+    "26933": "failure",
+    "26934": "failure",
+    "26935": "failure",
+    "26936": "already_complete",
+    "26937": "failure",
+    "26938": "failure",
+    "26939": "failure",
+    "26940": "failure",
+    "26941": "infeasible",
+    "26942": "failure",
+    "26943": "failure",
+    "26944": "failure",
+    "26945": "failure",
+    "26946": "failure",
+    "26947": "failure",
+    "26948": "failure",
+    "26949": "failure",
+    "26950": "failure",
+    "26951": "failure",
+    "26952": "failure",
+    "26953": "failure",
+    "26954": "failure",
+    "26955": "failure",
+    "26956": "failure",
+    "26957": "failure",
+    "26958": "failure",
+    "26959": "successful",
+    "26960": "failure",
+    "26961": "failure",
+    "26962": "failure",
+    "26963": "failure",
+    "26964": "failure",
+    "26965": "failure",
+    "26966": "failure",
+    "26967": "failure",
+    "26968": "failure",
+    "26969": "failure",
+    "26970": "failure",
+    "26971": "failure",
+    "26972": "failure",
+    "26973": "failure",
+    "26974": "successful",
+    "26975": "failure",
+    "26976": "infeasible",
+    "26977": "failure",
+    "26978": "failure",
+    "26979": "failure",
+    "26980": "failure",
+    "26981": "failure",
+    "26982": "failure",
+    "26983": "infeasible",
+    "26984": "failure",
+    "26985": "failure",
+    "26986": "failure",
+    "26987": "failure",
+    "26988": "failure",
+    "26989": "failure",
+    "26990": "failure",
+    "26991": "failure",
+    "26992": "failure",
+    "26993": "infeasible",
+    "26994": "failure",
+    "26995": "failure",
+    "26996": "failure",
+    "26997": "failure",
+    "26998": "failure",
+    "26999": "failure",
+    "27000": "failure",
+    "27001": "failure",
+    "27002": "successful",
+    "27003": "failure",
+    "27004": "failure",
+    "27005": "failure",
+    "27006": "failure",
+    "27007": "successful",
+    "27008": "failure",
+    "27009": "failure",
+    "27010": "failure",
+    "27011": "failure",
+    "27012": "failure",
+    "27013": "failure",
+    "27014": "failure",
+    "27015": "successful",
+    "27016": "failure",
+    "27017": "failure",
+    "27018": "infeasible",
+    "27019": "infeasible",
+    "27020": "failure",
+    "27021": "successful",
+    "27022": "failure",
+    "27023": "failure",
+    "27024": "failure",
+    "27025": "failure",
+    "27026": "failure",
+    "27027": "failure",
+    "27028": "failure",
+    "27029": "failure",
+    "27030": "infeasible",
+    "27031": "failure",
+    "27032": "failure",
+    "27033": "failure",
+    "27034": "failure",
+    "27035": "failure",
+    "27036": "failure",
+    "27037": "failure",
+    "27038": "failure",
+    "27039": "failure",
+    "27040": "failure",
+    "27041": "failure",
+    "27042": "failure",
+    "27043": "failure",
+    "27044": "failure",
+    "27045": "infeasible",
+    "27046": "successful",
+    "27047": "failure",
+    "27048": "failure",
+    "27049": "failure",
+    "27050": "failure",
+    "27051": "failure",
+    "27052": "infeasible",
+    "27053": "infeasible",
+    "27054": "infeasible",
+    "27055": "failure",
+    "27056": "failure",
+    "27057": "already_complete",
+    "27058": "failure",
+    "27059": "failure",
+    "27060": "failure",
+    "27061": "failure",
+    "27062": "failure",
+    "27063": "failure",
+    "27064": "failure",
+    "27065": "failure",
+    "27066": "failure",
+    "27067": "failure",
+    "27068": "infeasible",
+    "27069": "failure",
+    "27070": "infeasible",
+    "27071": "failure",
+    "27072": "failure",
+    "27073": "failure",
+    "27074": "failure",
+    "27075": "failure",
+    "27076": "infeasible",
+    "27077": "failure",
+    "27078": "failure",
+    "27079": "failure",
+    "27080": "failure",
+    "27081": "infeasible",
+    "27082": "failure",
+    "27083": "failure",
+    "27084": "failure",
+    "27085": "failure",
+    "27086": "failure",
+    "27087": "successful",
+    "27088": "failure",
+    "27089": "failure",
+    "27090": "failure",
+    "27091": "failure",
+    "27092": "successful",
+    "27093": "failure",
+    "27094": "failure",
+    "27095": "failure",
+    "27096": "failure",
+    "27097": "successful",
+    "27098": "failure",
+    "27099": "failure",
+    "27100": "successful",
+    "27101": "failure",
+    "27102": "failure",
+    "27103": "failure",
+    "27104": "successful",
+    "27105": "failure",
+    "27106": "failure",
+    "27107": "failure",
+    "27108": "successful",
+    "27109": "successful",
+    "27110": "infeasible",
+    "27111": "infeasible",
+    "27112": "failure",
+    "27113": "failure",
+    "27114": "failure",
+    "27115": "failure",
+    "27116": "failure",
+    "27117": "failure",
+    "27118": "successful",
+    "27119": "failure",
+    "27120": "failure",
+    "27121": "failure",
+    "27122": "failure",
+    "27123": "failure",
+    "27124": "failure",
+    "27125": "failure",
+    "27126": "failure",
+    "27127": "failure",
+    "27128": "failure",
+    "27129": "failure",
+    "27130": "failure",
+    "27131": "successful",
+    "27132": "failure",
+    "27133": "failure",
+    "27134": "failure",
+    "27135": "failure",
+    "27136": "failure",
+    "27137": "failure",
+    "27138": "failure",
+    "27139": "failure",
+    "27140": "failure",
+    "27141": "failure",
+    "27142": "failure",
+    "27143": "infeasible",
+    "27144": "failure",
+    "27145": "failure",
+    "27146": "failure",
+    "27147": "failure",
+    "27148": "failure",
+    "27149": "failure",
+    "27150": "failure",
+    "27151": "failure",
+    "27152": "infeasible",
+    "27153": "failure",
+    "27154": "failure",
+    "27155": "successful",
+    "27156": "infeasible",
+    "27157": "failure",
+    "27158": "failure",
+    "27159": "failure",
+    "27160": "failure",
+    "27161": "failure",
+    "27162": "failure",
+    "27163": "failure",
+    "27164": "failure",
+    "27165": "failure",
+    "27166": "failure",
+    "27167": "failure",
+    "27168": "failure",
+    "27169": "failure",
+    "27170": "failure",
+    "27171": "failure",
+    "27172": "failure",
+    "27173": "failure",
+    "27174": "failure",
+    "27175": "successful",
+    "27176": "infeasible",
+    "27177": "failure",
+    "27178": "infeasible",
+    "27179": "failure",
+    "27180": "successful",
+    "27181": "failure",
+    "27182": "failure",
+    "27183": "failure",
+    "27184": "failure",
+    "27185": "failure",
+    "27186": "failure",
+    "27187": "failure",
+    "27188": "infeasible",
+    "27189": "failure",
+    "27190": "failure",
+    "27191": "failure",
+    "27192": "failure",
+    "27193": "already_complete",
+    "27194": "failure",
+    "27195": "failure",
+    "27196": "failure",
+    "27197": "failure",
+    "27198": "failure",
+    "27199": "failure",
+    "27200": "failure",
+    "27201": "failure",
+    "27202": "failure",
+    "27203": "failure",
+    "27204": "successful",
+    "27205": "failure",
+    "27206": "failure",
+    "27207": "successful",
+    "27208": "failure",
+    "27209": "failure",
+    "27210": "successful",
+    "27211": "failure",
+    "27212": "infeasible",
+    "27213": "failure",
+    "27214": "infeasible",
+    "27215": "failure",
+    "27216": "failure",
+    "27217": "failure",
+    "27218": "successful",
+    "27219": "failure",
+    "27220": "failure",
+    "27221": "infeasible",
+    "27222": "failure",
+    "27223": "successful",
+    "27224": "failure",
+    "27225": "successful",
+    "27226": "failure",
+    "27227": "failure",
+    "27228": "failure",
+    "27229": "successful",
+    "27230": "successful",
+    "27231": "failure",
+    "27232": "failure",
+    "27233": "successful",
+    "27234": "failure",
+    "27235": "successful",
+    "27236": "failure",
+    "27237": "infeasible",
+    "27238": "infeasible",
+    "27239": "failure",
+    "27240": "failure",
+    "27241": "successful",
+    "27242": "failure",
+    "27243": "failure",
+    "27244": "failure",
+    "27245": "failure",
+    "27246": "failure",
+    "27247": "failure",
+    "27248": "failure",
+    "27249": "failure",
+    "27250": "failure",
+    "27251": "failure",
+    "27252": "failure",
+    "27253": "failure",
+    "27254": "failure",
+    "27255": "failure",
+    "27256": "failure",
+    "27257": "failure",
+    "27258": "successful",
+    "27259": "failure",
+    "27260": "failure",
+    "27261": "infeasible",
+    "27262": "successful",
+    "27263": "failure",
+    "27264": "failure",
+    "27265": "failure",
+    "27266": "failure",
+    "27267": "failure",
+    "27268": "failure",
+    "27269": "failure",
+    "27270": "failure",
+    "27271": "failure",
+    "27272": "infeasible",
+    "27273": "failure",
+    "27274": "failure",
+    "27275": "failure",
+    "27276": "failure",
+    "27277": "failure",
+    "27278": "failure",
+    "27279": "failure",
+    "27280": "failure",
+    "27281": "successful",
+    "27282": "failure",
+    "27283": "failure",
+    "27284": "failure",
+    "27285": "failure",
+    "27286": "already_complete",
+    "27287": "failure",
+    "27288": "failure",
+    "27289": "failure",
+    "27290": "failure",
+    "27291": "failure",
+    "27292": "failure",
+    "27293": "failure",
+    "27294": "failure",
+    "27295": "failure",
+    "27296": "failure",
+    "27297": "failure",
+    "27298": "failure",
+    "27299": "failure",
+    "27300": "infeasible",
+    "27301": "failure",
+    "27302": "failure",
+    "27303": "failure",
+    "27304": "failure",
+    "27305": "successful",
+    "27306": "failure",
+    "27307": "successful",
+    "27308": "infeasible",
+    "27309": "failure",
+    "27310": "failure",
+    "27311": "failure",
+    "27312": "failure",
+    "27313": "failure",
+    "27314": "failure",
+    "27315": "successful",
+    "27316": "failure",
+    "27317": "failure",
+    "27318": "infeasible",
+    "27319": "infeasible",
+    "27320": "successful",
+    "27321": "failure",
+    "27322": "failure",
+    "27323": "failure",
+    "27324": "successful",
+    "27325": "failure",
+    "27326": "successful",
+    "27327": "failure",
+    "27328": "failure",
+    "27329": "failure",
+    "27330": "failure",
+    "27331": "failure",
+    "27332": "failure",
+    "27333": "infeasible",
+    "27334": "failure",
+    "27335": "failure",
+    "27336": "failure",
+    "27337": "failure",
+    "27338": "failure",
+    "27339": "successful",
+    "27340": "successful",
+    "27341": "failure",
+    "27342": "infeasible",
+    "27343": "failure",
+    "27344": "failure",
+    "27345": "successful",
+    "27346": "failure",
+    "27347": "failure",
+    "27348": "infeasible",
+    "27349": "failure",
+    "27350": "failure",
+    "27351": "successful",
+    "27352": "failure",
+    "27353": "failure",
+    "27354": "failure",
+    "27355": "failure",
+    "27356": "failure",
+    "27357": "failure",
+    "27358": "failure",
+    "27359": "successful",
+    "27360": "successful",
+    "27361": "failure",
+    "27362": "failure",
+    "27363": "failure",
+    "27364": "failure",
+    "27365": "failure",
+    "27366": "failure",
+    "27367": "failure",
+    "27368": "failure",
+    "27369": "infeasible",
+    "27370": "failure",
+    "27371": "failure",
+    "27372": "failure",
+    "27373": "successful",
+    "27374": "successful",
+    "27375": "failure",
+    "27376": "failure",
+    "27377": "failure",
+    "27378": "infeasible",
+    "27379": "failure",
+    "27380": "failure",
+    "27381": "failure",
+    "27382": "failure",
+    "27383": "failure",
+    "27384": "failure",
+    "27385": "failure",
+    "27386": "failure",
+    "27387": "failure",
+    "27388": "failure",
+    "27389": "failure",
+    "27390": "failure",
+    "27391": "failure",
+    "27392": "failure",
+    "27393": "failure",
+    "27394": "failure",
+    "27395": "failure",
+    "27396": "failure",
+    "27397": "failure",
+    "27398": "successful",
+    "27399": "failure",
+    "27400": "failure",
+    "27401": "successful",
+    "27402": "failure",
+    "27403": "failure",
+    "27404": "failure",
+    "27405": "failure",
+    "27406": "failure",
+    "27407": "failure",
+    "27408": "failure",
+    "27409": "failure",
+    "27410": "failure",
+    "27411": "failure",
+    "27412": "failure",
+    "27413": "failure",
+    "27414": "successful",
+    "27415": "failure",
+    "27416": "failure",
+    "27417": "failure",
+    "27418": "failure",
+    "27419": "failure",
+    "27420": "infeasible",
+    "27421": "successful",
+    "27422": "failure",
+    "27423": "successful",
+    "27424": "failure",
+    "27425": "failure",
+    "27426": "failure",
+    "27427": "failure",
+    "27428": "failure",
+    "27429": "infeasible",
+    "27430": "failure",
+    "27431": "failure",
+    "27432": "infeasible",
+    "27433": "failure",
+    "27434": "failure",
+    "27435": "failure",
+    "27436": "failure",
+    "27437": "failure",
+    "27438": "successful",
+    "27439": "failure",
+    "27440": "failure",
+    "27441": "failure",
+    "27442": "failure",
+    "27443": "failure",
+    "27444": "failure",
+    "27445": "failure",
+    "27446": "failure",
+    "27447": "failure",
+    "27448": "successful",
+    "27449": "failure",
+    "27450": "failure",
+    "27451": "infeasible",
+    "27452": "failure",
+    "27453": "failure",
+    "27454": "failure",
+    "27455": "failure",
+    "27456": "failure",
+    "27457": "failure",
+    "27458": "failure",
+    "27459": "infeasible",
+    "27460": "failure",
+    "27461": "successful",
+    "27462": "failure",
+    "27463": "failure",
+    "27464": "successful",
+    "27465": "failure",
+    "27466": "failure",
+    "27467": "failure",
+    "27468": "failure",
+    "27469": "failure",
+    "27470": "failure",
+    "27471": "failure",
+    "27472": "successful",
+    "27473": "failure",
+    "27474": "failure",
+    "27475": "successful",
+    "27476": "infeasible",
+    "27477": "failure",
+    "27478": "failure",
+    "27479": "failure",
+    "27480": "failure",
+    "27481": "failure",
+    "27482": "failure",
+    "27483": "failure",
+    "27484": "failure",
+    "27485": "failure",
+    "27486": "successful",
+    "27487": "failure",
+    "27488": "infeasible",
+    "27489": "failure",
+    "27490": "failure",
+    "27491": "failure",
+    "27492": "failure",
+    "27493": "failure",
+    "27494": "failure",
+    "27495": "failure",
+    "27496": "failure",
+    "27497": "failure",
+    "27498": "failure",
+    "27499": "failure",
+    "27500": "failure",
+    "27501": "failure",
+    "27502": "infeasible",
+    "27503": "failure",
+    "27504": "failure",
+    "27505": "failure",
+    "27506": "failure",
+    "27507": "failure",
+    "27508": "failure",
+    "27509": "infeasible",
+    "27510": "failure",
+    "27511": "failure",
+    "27512": "failure",
+    "27513": "failure",
+    "27514": "failure",
+    "27515": "failure",
+    "27516": "failure",
+    "27517": "failure",
+    "27518": "failure",
+    "27519": "failure",
+    "27520": "successful",
+    "27521": "failure",
+    "27522": "failure",
+    "27523": "failure",
+    "27524": "failure",
+    "27525": "failure",
+    "27526": "failure",
+    "27527": "successful",
+    "27528": "failure",
+    "27529": "failure",
+    "27530": "failure",
+    "27531": "failure",
+    "27532": "successful",
+    "27533": "failure",
+    "27534": "failure",
+    "27535": "successful",
+    "27536": "failure",
+    "27537": "failure",
+    "27538": "failure",
+    "27539": "failure",
+    "27540": "failure",
+    "27541": "infeasible",
+    "27542": "infeasible",
+    "27543": "failure",
+    "27544": "failure",
+    "27545": "failure",
+    "27546": "infeasible",
+    "27547": "failure",
+    "27548": "failure",
+    "27549": "failure",
+    "27550": "infeasible",
+    "27551": "failure",
+    "27552": "failure",
+    "27553": "failure",
+    "27554": "failure",
+    "27555": "failure",
+    "27556": "failure",
+    "27557": "failure",
+    "27558": "failure",
+    "27559": "infeasible",
+    "27560": "failure",
+    "27561": "failure",
+    "27562": "failure",
+    "27563": "failure",
+    "27564": "failure",
+    "27565": "failure",
+    "27566": "failure",
+    "27567": "failure",
+    "27568": "successful",
+    "27569": "failure",
+    "27570": "failure",
+    "27571": "failure",
+    "27572": "failure",
+    "27573": "failure",
+    "27574": "infeasible",
+    "27575": "failure",
+    "27576": "failure",
+    "27577": "failure",
+    "27578": "failure",
+    "27579": "failure",
+    "27580": "failure",
+    "27581": "failure",
+    "27582": "failure",
+    "27583": "failure",
+    "27584": "failure",
+    "27585": "failure",
+    "27586": "failure",
+    "27587": "successful",
+    "27588": "infeasible",
+    "27589": "failure",
+    "27590": "failure",
+    "27591": "failure",
+    "27592": "failure",
+    "27593": "failure",
+    "27594": "failure",
+    "27595": "failure",
+    "27596": "failure",
+    "27597": "failure",
+    "27598": "failure",
+    "27599": "failure",
+    "27600": "failure",
+    "27601": "failure",
+    "27602": "failure",
+    "27603": "failure",
+    "27604": "failure",
+    "27605": "failure",
+    "27606": "failure",
+    "27607": "successful",
+    "27608": "failure",
+    "27609": "failure",
+    "27610": "failure",
+    "27611": "failure",
+    "27612": "successful",
+    "27613": "successful",
+    "27614": "failure",
+    "27615": "infeasible",
+    "27616": "failure",
+    "27617": "infeasible",
+    "27618": "failure",
+    "27619": "failure",
+    "27620": "failure",
+    "27621": "failure",
+    "27622": "failure",
+    "27623": "infeasible",
+    "27624": "failure",
+    "27625": "failure",
+    "27626": "infeasible",
+    "27627": "failure",
+    "27628": "failure",
+    "27629": "failure",
+    "27630": "infeasible",
+    "27631": "infeasible",
+    "27632": "successful",
+    "27633": "infeasible",
+    "27634": "failure",
+    "27635": "failure",
+    "27636": "failure",
+    "27637": "failure",
+    "27638": "infeasible",
+    "27639": "failure",
+    "27640": "failure",
+    "27641": "infeasible",
+    "27642": "successful",
+    "27643": "failure",
+    "27644": "failure",
+    "27645": "failure",
+    "27646": "failure",
+    "27647": "failure",
+    "27648": "failure",
+    "27649": "failure",
+    "27650": "failure",
+    "27651": "failure",
+    "27652": "failure",
+    "27653": "failure",
+    "27654": "failure",
+    "27655": "failure",
+    "27656": "failure",
+    "27657": "failure",
+    "27658": "successful",
+    "27659": "failure",
+    "27660": "failure",
+    "27661": "failure",
+    "27662": "infeasible",
+    "27663": "infeasible",
+    "27664": "successful",
+    "27665": "infeasible",
+    "27666": "failure",
+    "27667": "failure",
+    "27668": "failure",
+    "27669": "failure",
+    "27670": "failure",
+    "27671": "failure",
+    "27672": "failure",
+    "27673": "successful",
+    "27674": "failure",
+    "27675": "failure",
+    "27676": "infeasible",
+    "27677": "failure",
+    "27678": "failure",
+    "27679": "failure",
+    "27680": "failure",
+    "27681": "failure",
+    "27682": "successful",
+    "27683": "failure",
+    "27684": "failure",
+    "27685": "failure",
+    "27686": "failure",
+    "27687": "successful",
+    "27688": "failure",
+    "27689": "failure",
+    "27690": "failure",
+    "27691": "failure",
+    "27692": "failure",
+    "27693": "infeasible",
+    "27694": "failure",
+    "27695": "failure",
+    "27696": "successful",
+    "27697": "failure",
+    "27698": "failure",
+    "27699": "failure",
+    "27700": "failure",
+    "27701": "failure",
+    "27702": "failure",
+    "27703": "failure",
+    "27704": "infeasible",
+    "27705": "failure",
+    "27706": "failure",
+    "27707": "failure",
+    "27708": "infeasible",
+    "27709": "failure",
+    "27710": "failure",
+    "27711": "failure",
+    "27712": "failure",
+    "27713": "failure",
+    "27714": "failure",
+    "27715": "infeasible",
+    "27716": "failure",
+    "27717": "failure",
+    "27718": "failure",
+    "27719": "failure",
+    "27720": "failure",
+    "27721": "failure",
+    "27722": "failure",
+    "27723": "failure",
+    "27724": "failure",
+    "27725": "failure",
+    "27726": "infeasible",
+    "27727": "successful",
+    "27728": "failure",
+    "27729": "infeasible",
+    "27730": "failure",
+    "27731": "infeasible",
+    "27732": "failure",
+    "27733": "failure",
+    "27734": "failure",
+    "27735": "successful",
+    "27736": "successful",
+    "27737": "failure",
+    "27738": "failure",
+    "27739": "failure",
+    "27740": "failure",
+    "27741": "failure",
+    "27742": "failure",
+    "27743": "failure",
+    "27744": "failure",
+    "27745": "failure",
+    "27746": "successful",
+    "27747": "failure",
+    "27748": "failure",
+    "27749": "failure",
+    "27750": "failure",
+    "27751": "failure",
+    "27752": "failure",
+    "27753": "failure",
+    "27754": "infeasible",
+    "27755": "failure",
+    "27756": "failure",
+    "27757": "failure",
+    "27758": "failure",
+    "27759": "successful",
+    "27760": "failure",
+    "27761": "failure",
+    "27762": "successful",
+    "27763": "successful",
+    "27764": "failure",
+    "27765": "failure",
+    "27766": "already_complete",
+    "27767": "failure",
+    "27768": "failure",
+    "27769": "failure",
+    "27770": "failure",
+    "27771": "failure",
+    "27772": "failure",
+    "27773": "failure",
+    "27774": "failure",
+    "27775": "failure",
+    "27776": "failure",
+    "27777": "failure",
+    "27778": "failure",
+    "27779": "failure",
+    "27780": "infeasible",
+    "27781": "failure",
+    "27782": "successful",
+    "27783": "failure",
+    "27784": "failure",
+    "27785": "failure",
+    "27786": "failure",
+    "27787": "failure",
+    "27788": "failure",
+    "27789": "failure",
+    "27790": "failure",
+    "27791": "failure",
+    "27792": "failure",
+    "27793": "failure",
+    "27794": "failure",
+    "27795": "infeasible",
+    "27796": "failure",
+    "27797": "successful",
+    "27798": "failure",
+    "27799": "failure",
+    "27800": "failure",
+    "27801": "failure",
+    "27802": "failure",
+    "27803": "failure",
+    "27804": "failure",
+    "27805": "successful",
+    "27806": "failure",
+    "27807": "failure",
+    "27808": "failure",
+    "27809": "failure",
+    "27810": "failure",
+    "27811": "failure",
+    "27812": "failure",
+    "27813": "successful",
+    "27814": "failure",
+    "27815": "failure",
+    "27816": "failure",
+    "27817": "failure",
+    "27818": "failure",
+    "27819": "successful",
+    "27820": "failure",
+    "27821": "failure",
+    "27822": "failure",
+    "27823": "failure",
+    "27824": "failure",
+    "27825": "failure",
+    "27826": "failure",
+    "27827": "infeasible",
+    "27828": "failure",
+    "27829": "infeasible",
+    "27830": "failure",
+    "27831": "successful",
+    "27832": "failure",
+    "27833": "failure",
+    "27834": "failure",
+    "27835": "infeasible",
+    "27836": "failure",
+    "27837": "infeasible",
+    "27838": "failure",
+    "27839": "failure",
+    "27840": "failure",
+    "27841": "failure",
+    "27842": "failure",
+    "27843": "failure",
+    "27844": "failure",
+    "27845": "failure",
+    "27846": "infeasible",
+    "27847": "failure",
+    "27848": "failure",
+    "27849": "failure",
+    "27850": "failure",
+    "27851": "infeasible",
+    "27852": "failure",
+    "27853": "failure",
+    "27854": "failure",
+    "27855": "failure",
+    "27856": "infeasible",
+    "27857": "failure",
+    "27858": "failure",
+    "27859": "infeasible",
+    "27860": "failure",
+    "27861": "failure",
+    "27862": "failure",
+    "27863": "failure",
+    "27864": "failure",
+    "27865": "failure",
+    "27866": "failure",
+    "27867": "failure",
+    "27868": "infeasible",
+    "27869": "failure",
+    "27870": "failure",
+    "27871": "failure",
+    "27872": "infeasible",
+    "27873": "failure",
+    "27874": "failure",
+    "27875": "failure",
+    "27876": "successful",
+    "27877": "successful",
+    "27878": "infeasible",
+    "27879": "failure",
+    "27880": "failure",
+    "27881": "failure",
+    "27882": "failure",
+    "27883": "already_complete",
+    "27884": "failure",
+    "27885": "failure",
+    "27886": "failure",
+    "27887": "failure",
+    "27888": "failure",
+    "27889": "failure",
+    "27890": "failure",
+    "27891": "failure",
+    "27892": "failure",
+    "27893": "failure",
+    "27894": "failure",
+    "27895": "failure",
+    "27896": "failure",
+    "27897": "failure",
+    "27898": "already_complete",
+    "27899": "failure",
+    "27900": "successful",
+    "27901": "failure",
+    "27902": "failure",
+    "27903": "failure",
+    "27904": "failure",
+    "27905": "failure",
+    "27906": "failure",
+    "27907": "infeasible",
+    "27908": "failure",
+    "27909": "successful",
+    "27910": "failure",
+    "27911": "infeasible",
+    "27912": "failure",
+    "27913": "failure",
+    "27914": "failure",
+    "27915": "failure",
+    "27916": "failure",
+    "27917": "failure",
+    "27918": "failure",
+    "27919": "failure",
+    "27920": "infeasible",
+    "27921": "infeasible",
+    "27922": "failure",
+    "27923": "failure",
+    "27924": "failure",
+    "27925": "failure",
+    "27926": "failure",
+    "27927": "failure",
+    "27928": "failure",
+    "27929": "failure",
+    "27930": "successful",
+    "27931": "failure",
+    "27932": "failure",
+    "27933": "infeasible",
+    "27934": "failure",
+    "27935": "failure",
+    "27936": "failure",
+    "27937": "failure",
+    "27938": "failure",
+    "27939": "failure",
+    "27940": "infeasible",
+    "27941": "failure",
+    "27942": "infeasible",
+    "27943": "failure",
+    "27944": "failure",
+    "27945": "failure",
+    "27946": "failure",
+    "27947": "failure",
+    "27948": "failure",
+    "27949": "failure",
+    "27950": "failure",
+    "27951": "failure",
+    "27952": "failure",
+    "27953": "failure",
+    "27954": "successful",
+    "27955": "infeasible",
+    "27956": "failure",
+    "27957": "infeasible",
+    "27958": "failure",
+    "27959": "failure",
+    "27960": "infeasible",
+    "27961": "failure",
+    "27962": "failure",
+    "27963": "failure",
+    "27964": "failure",
+    "27965": "failure",
+    "27966": "failure",
+    "27967": "failure",
+    "27968": "failure",
+    "27969": "failure",
+    "27970": "successful",
+    "27971": "failure",
+    "27972": "failure",
+    "27973": "infeasible",
+    "27974": "failure",
+    "27975": "infeasible",
+    "27976": "failure",
+    "27977": "failure",
+    "27978": "failure",
+    "27979": "failure",
+    "27980": "failure",
+    "27981": "failure",
+    "27982": "failure",
+    "27983": "failure",
+    "27984": "failure",
+    "27985": "failure",
+    "27986": "failure",
+    "27987": "infeasible",
+    "27988": "failure",
+    "27989": "failure",
+    "27990": "failure",
+    "27991": "failure",
+    "27992": "failure",
+    "27993": "infeasible",
+    "27994": "failure",
+    "27995": "infeasible",
+    "27996": "failure",
+    "27997": "already_complete",
+    "27998": "failure",
+    "27999": "already_complete",
+    "28000": "failure",
+    "28001": "failure",
+    "28002": "failure",
+    "28003": "failure",
+    "28004": "failure",
+    "28005": "failure",
+    "28006": "failure",
+    "28007": "failure",
+    "28008": "failure",
+    "28009": "successful",
+    "28010": "successful",
+    "28011": "successful",
+    "28012": "failure",
+    "28013": "infeasible",
+    "28014": "infeasible",
+    "28015": "infeasible",
+    "28016": "failure",
+    "28017": "infeasible",
+    "28018": "failure",
+    "28019": "failure",
+    "28020": "failure",
+    "28021": "failure",
+    "28022": "infeasible",
+    "28023": "failure",
+    "28024": "infeasible",
+    "28025": "failure",
+    "28026": "failure",
+    "28027": "failure",
+    "28028": "failure",
+    "28029": "failure",
+    "28030": "infeasible",
+    "28031": "infeasible",
+    "28032": "failure",
+    "28033": "infeasible",
+    "28034": "failure",
+    "28035": "failure",
+    "28036": "failure",
+    "28037": "failure",
+    "28038": "successful",
+    "28039": "failure",
+    "28040": "failure",
+    "28041": "failure",
+    "28042": "failure",
+    "28043": "failure",
+    "28044": "failure",
+    "28045": "failure",
+    "28046": "failure",
+    "28047": "successful",
+    "28048": "failure",
+    "28049": "failure",
+    "28050": "failure",
+    "28051": "failure",
+    "28052": "failure",
+    "28053": "failure",
+    "28054": "failure",
+    "28055": "failure",
+    "28056": "failure",
+    "28057": "failure",
+    "28058": "failure",
+    "28059": "failure",
+    "28060": "failure",
+    "28061": "failure",
+    "28062": "failure",
+    "28063": "successful",
+    "28064": "successful",
+    "28065": "failure",
+    "28066": "failure",
+    "28067": "failure",
+    "28068": "failure",
+    "28069": "failure",
+    "28070": "failure",
+    "28071": "failure",
+    "28072": "failure",
+    "28073": "failure",
+    "28074": "failure",
+    "28075": "failure",
+    "28076": "successful",
+    "28077": "infeasible",
+    "28078": "failure",
+    "28079": "failure",
+    "28080": "failure",
+    "28081": "failure",
+    "28082": "failure",
+    "28083": "failure",
+    "28084": "failure",
+    "28085": "failure",
+    "28086": "failure",
+    "28087": "failure",
+    "28088": "successful",
+    "28089": "failure",
+    "28090": "successful",
+    "28091": "failure",
+    "28092": "failure",
+    "28093": "failure",
+    "28094": "failure",
+    "28095": "failure",
+    "28096": "failure",
+    "28097": "failure",
+    "28098": "failure",
+    "28099": "failure",
+    "28100": "failure",
+    "28101": "failure",
+    "28102": "successful",
+    "28103": "successful",
+    "28104": "failure",
+    "28105": "infeasible",
+    "28106": "failure",
+    "28107": "failure",
+    "28108": "failure",
+    "28109": "failure",
+    "28110": "failure",
+    "28111": "failure",
+    "28112": "failure",
+    "28113": "failure",
+    "28114": "failure",
+    "28115": "infeasible",
+    "28116": "infeasible",
+    "28117": "successful",
+    "28118": "failure",
+    "28119": "failure",
+    "28120": "failure",
+    "28121": "infeasible",
+    "28122": "failure",
+    "28123": "failure",
+    "28124": "failure",
+    "28125": "failure",
+    "28126": "failure",
+    "28127": "failure",
+    "28128": "failure",
+    "28129": "failure",
+    "28130": "failure",
+    "28131": "infeasible",
+    "28132": "successful",
+    "28133": "successful",
+    "28134": "failure",
+    "28135": "failure",
+    "28136": "failure",
+    "28137": "failure",
+    "28138": "failure",
+    "28139": "failure",
+    "28140": "failure",
+    "28141": "successful",
+    "28142": "failure",
+    "28143": "infeasible",
+    "28144": "failure",
+    "28145": "infeasible",
+    "28146": "failure",
+    "28147": "failure",
+    "28148": "failure",
+    "28149": "failure",
+    "28150": "failure",
+    "28151": "failure",
+    "28152": "failure",
+    "28153": "successful",
+    "28154": "infeasible",
+    "28155": "failure",
+    "28156": "failure",
+    "28157": "failure",
+    "28158": "failure",
+    "28159": "successful",
+    "28160": "failure",
+    "28161": "failure",
+    "28162": "successful",
+    "28163": "successful",
+    "28164": "failure",
+    "28165": "failure",
+    "28166": "successful",
+    "28167": "failure",
+    "28168": "successful",
+    "28169": "failure",
+    "28170": "infeasible",
+    "28171": "failure",
+    "28172": "failure",
+    "28173": "infeasible",
+    "28174": "successful",
+    "28175": "successful",
+    "28176": "failure",
+    "28177": "failure",
+    "28178": "failure",
+    "28179": "successful",
+    "28180": "failure",
+    "28181": "infeasible",
+    "28182": "infeasible",
+    "28183": "failure",
+    "28184": "failure",
+    "28185": "infeasible",
+    "28186": "failure",
+    "28187": "failure",
+    "28188": "failure",
+    "28189": "failure",
+    "28190": "infeasible",
+    "28191": "failure",
+    "28192": "successful",
+    "28193": "failure",
+    "28194": "infeasible",
+    "28195": "failure",
+    "28196": "failure",
+    "28197": "failure",
+    "28198": "infeasible",
+    "28199": "failure",
+    "28200": "failure",
+    "28201": "failure",
+    "28202": "successful",
+    "28203": "failure",
+    "28204": "failure",
+    "28205": "failure",
+    "28206": "failure",
+    "28207": "failure",
+    "28208": "failure",
+    "28209": "successful",
+    "28210": "failure",
+    "28211": "failure",
+    "28212": "failure",
+    "28213": "failure",
+    "28214": "failure",
+    "28215": "failure",
+    "28216": "failure",
+    "28217": "failure",
+    "28218": "failure",
+    "28219": "failure",
+    "28220": "failure",
+    "28221": "failure",
+    "28222": "successful",
+    "28223": "failure",
+    "28224": "failure",
+    "28225": "failure",
+    "28226": "failure",
+    "28227": "failure",
+    "28228": "successful",
+    "28229": "failure",
+    "28230": "failure",
+    "28231": "failure",
+    "28232": "failure",
+    "28233": "failure",
+    "28234": "failure",
+    "28235": "failure",
+    "28236": "failure",
+    "28237": "successful",
+    "28238": "failure",
+    "28239": "failure",
+    "28240": "failure",
+    "28241": "failure",
+    "28242": "successful",
+    "28243": "failure",
+    "28244": "failure",
+    "28245": "successful",
+    "28246": "infeasible",
+    "28247": "failure",
+    "28248": "failure",
+    "28249": "infeasible",
+    "28250": "failure",
+    "28251": "failure",
+    "28252": "failure",
+    "28253": "failure",
+    "28254": "failure",
+    "28255": "failure",
+    "28256": "failure",
+    "28257": "successful",
+    "28258": "failure",
+    "28259": "infeasible",
+    "28260": "failure",
+    "28261": "failure",
+    "28262": "failure",
+    "28263": "infeasible",
+    "28264": "infeasible",
+    "28265": "failure",
+    "28266": "failure",
+    "28267": "failure",
+    "28268": "failure",
+    "28269": "failure",
+    "28270": "failure",
+    "28271": "failure",
+    "28272": "failure",
+    "28273": "failure",
+    "28274": "failure",
+    "28275": "failure",
+    "28276": "failure",
+    "28277": "successful",
+    "28278": "failure",
+    "28279": "already_complete",
+    "28280": "failure",
+    "28281": "failure",
+    "28282": "failure",
+    "28283": "infeasible",
+    "28284": "failure",
+    "28285": "failure",
+    "28286": "successful",
+    "28287": "failure",
+    "28288": "failure",
+    "28289": "failure",
+    "28290": "failure",
+    "28291": "successful",
+    "28292": "failure",
+    "28293": "failure",
+    "28294": "failure",
+    "28295": "failure",
+    "28296": "failure",
+    "28297": "failure",
+    "28298": "failure",
+    "28299": "failure",
+    "28300": "failure",
+    "28301": "failure",
+    "28302": "failure",
+    "28303": "successful",
+    "28304": "failure",
+    "28305": "failure",
+    "28306": "failure",
+    "28307": "failure",
+    "28308": "infeasible",
+    "28309": "failure",
+    "28310": "failure",
+    "28311": "failure",
+    "28312": "failure",
+    "28313": "successful",
+    "28314": "failure",
+    "28315": "successful",
+    "28316": "failure",
+    "28317": "failure",
+    "28318": "failure",
+    "28319": "failure",
+    "28320": "failure",
+    "28321": "infeasible",
+    "28322": "failure",
+    "28323": "failure",
+    "28324": "failure",
+    "28325": "failure",
+    "28326": "failure",
+    "28327": "failure",
+    "28328": "infeasible",
+    "28329": "failure",
+    "28330": "failure",
+    "28331": "failure",
+    "28332": "failure",
+    "28333": "failure",
+    "28334": "failure",
+    "28335": "failure",
+    "28336": "failure",
+    "28337": "successful",
+    "28338": "failure",
+    "28339": "failure",
+    "28340": "failure",
+    "28341": "failure",
+    "28342": "failure",
+    "28343": "failure",
+    "28344": "failure",
+    "28345": "infeasible",
+    "28346": "successful",
+    "28347": "failure",
+    "28348": "failure",
+    "28349": "failure",
+    "28350": "failure",
+    "28351": "failure",
+    "28352": "failure",
+    "28353": "failure",
+    "28354": "infeasible",
+    "28355": "failure",
+    "28356": "failure",
+    "28357": "failure",
+    "28358": "infeasible",
+    "28359": "failure",
+    "28360": "failure",
+    "28361": "failure",
+    "28362": "failure",
+    "28363": "failure",
+    "28364": "infeasible",
+    "28365": "failure",
+    "28366": "failure",
+    "28367": "successful",
+    "28368": "failure",
+    "28369": "infeasible",
+    "28370": "failure",
+    "28371": "successful",
+    "28372": "failure",
+    "28373": "failure",
+    "28374": "failure",
+    "28375": "failure",
+    "28376": "failure",
+    "28377": "successful",
+    "28378": "failure",
+    "28379": "failure",
+    "28380": "already_complete",
+    "28381": "failure",
+    "28382": "failure",
+    "28383": "failure",
+    "28384": "failure",
+    "28385": "infeasible",
+    "28386": "failure",
+    "28387": "failure",
+    "28388": "failure",
+    "28389": "failure",
+    "28390": "infeasible",
+    "28391": "infeasible",
+    "28392": "failure",
+    "28393": "failure",
+    "28394": "failure",
+    "28395": "failure",
+    "28396": "failure",
+    "28397": "failure",
+    "28398": "failure",
+    "28399": "infeasible",
+    "28400": "successful",
+    "28401": "failure",
+    "28402": "failure",
+    "28403": "infeasible",
+    "28404": "failure",
+    "28405": "failure",
+    "28406": "infeasible",
+    "28407": "failure",
+    "28408": "infeasible",
+    "28409": "failure",
+    "28410": "failure",
+    "28411": "failure",
+    "28412": "failure",
+    "28413": "failure",
+    "28414": "failure",
+    "28415": "already_complete",
+    "28416": "failure",
+    "28417": "failure",
+    "28418": "failure",
+    "28419": "failure",
+    "28420": "failure",
+    "28421": "failure",
+    "28422": "failure",
+    "28423": "failure",
+    "28424": "failure",
+    "28425": "failure",
+    "28426": "failure",
+    "28427": "failure",
+    "28428": "failure",
+    "28429": "failure",
+    "28430": "failure",
+    "28431": "failure",
+    "28432": "infeasible",
+    "28433": "failure",
+    "28434": "failure",
+    "28435": "failure",
+    "28436": "failure",
+    "28437": "successful",
+    "28438": "failure",
+    "28439": "failure",
+    "28440": "failure",
+    "28441": "failure",
+    "28442": "infeasible",
+    "28443": "failure",
+    "28444": "failure",
+    "28445": "successful",
+    "28446": "failure",
+    "28447": "failure",
+    "28448": "failure",
+    "28449": "failure",
+    "28450": "failure",
+    "28451": "failure",
+    "28452": "failure",
+    "28453": "failure",
+    "28454": "failure",
+    "28455": "failure",
+    "28456": "failure",
+    "28457": "failure",
+    "28458": "failure",
+    "28459": "infeasible",
+    "28460": "failure",
+    "28461": "infeasible",
+    "28462": "failure",
+    "28463": "successful",
+    "28464": "failure",
+    "28465": "failure",
+    "28466": "infeasible",
+    "28467": "failure",
+    "28468": "successful",
+    "28469": "failure",
+    "28470": "failure",
+    "28471": "failure",
+    "28472": "failure",
+    "28473": "failure",
+    "28474": "failure",
+    "28475": "infeasible",
+    "28476": "successful",
+    "28477": "failure",
+    "28478": "failure",
+    "28479": "failure",
+    "28480": "successful",
+    "28481": "failure",
+    "28482": "failure",
+    "28483": "failure",
+    "28484": "failure",
+    "28485": "failure",
+    "28486": "failure",
+    "28487": "failure",
+    "28488": "successful",
+    "28489": "failure",
+    "28490": "failure",
+    "28491": "failure",
+    "28492": "failure",
+    "28493": "failure",
+    "28494": "failure",
+    "28495": "failure",
+    "28496": "failure",
+    "28497": "failure",
+    "28498": "successful",
+    "28499": "infeasible",
+    "28500": "failure",
+    "28501": "infeasible",
+    "28502": "failure",
+    "28503": "successful",
+    "28504": "failure",
+    "28505": "failure",
+    "28506": "failure",
+    "28507": "failure",
+    "28508": "failure",
+    "28509": "failure",
+    "28510": "failure",
+    "28511": "failure",
+    "28512": "failure",
+    "28513": "failure",
+    "28514": "failure",
+    "28515": "failure",
+    "28516": "failure",
+    "28517": "failure",
+    "28518": "failure",
+    "28519": "failure",
+    "28520": "infeasible",
+    "28521": "failure",
+    "28522": "failure",
+    "28523": "failure",
+    "28524": "failure",
+    "28525": "successful",
+    "28526": "failure",
+    "28527": "successful",
+    "28528": "failure",
+    "28529": "failure",
+    "28530": "successful",
+    "28531": "failure",
+    "28532": "failure",
+    "28533": "failure",
+    "28534": "failure",
+    "28535": "failure",
+    "28536": "infeasible",
+    "28537": "failure",
+    "28538": "failure",
+    "28539": "failure",
+    "28540": "failure",
+    "28541": "infeasible",
+    "28542": "failure",
+    "28543": "successful",
+    "28544": "failure",
+    "28545": "failure",
+    "28546": "failure",
+    "28547": "failure",
+    "28548": "already_complete",
+    "28549": "failure",
+    "28550": "infeasible",
+    "28551": "failure",
+    "28552": "infeasible",
+    "28553": "successful",
+    "28554": "successful",
+    "28555": "failure",
+    "28556": "failure",
+    "28557": "failure",
+    "28558": "failure",
+    "28559": "failure",
+    "28560": "failure",
+    "28561": "failure",
+    "28562": "successful",
+    "28563": "infeasible",
+    "28564": "failure",
+    "28565": "failure",
+    "28566": "successful",
+    "28567": "failure",
+    "28568": "failure",
+    "28569": "successful",
+    "28570": "failure",
+    "28571": "failure",
+    "28572": "failure",
+    "28573": "failure",
+    "28574": "infeasible",
+    "28575": "failure",
+    "28576": "failure",
+    "28577": "failure",
+    "28578": "infeasible",
+    "28579": "infeasible",
+    "28580": "failure",
+    "28581": "failure",
+    "28582": "failure",
+    "28583": "failure",
+    "28584": "infeasible",
+    "28585": "successful",
+    "28586": "successful",
+    "28587": "failure",
+    "28588": "successful",
+    "28589": "failure",
+    "28590": "failure",
+    "28591": "failure",
+    "28592": "failure",
+    "28593": "failure",
+    "28594": "failure",
+    "28595": "successful",
+    "28596": "failure",
+    "28597": "failure",
+    "28598": "failure",
+    "28599": "infeasible",
+    "28600": "failure",
+    "28601": "successful",
+    "28602": "failure",
+    "28603": "failure",
+    "28604": "already_complete",
+    "28605": "failure",
+    "28606": "infeasible",
+    "28607": "failure",
+    "28608": "infeasible",
+    "28609": "failure",
+    "28610": "failure",
+    "28611": "failure",
+    "28612": "failure",
+    "28613": "successful",
+    "28614": "failure",
+    "28615": "successful",
+    "28616": "failure",
+    "28617": "failure",
+    "28618": "failure",
+    "28619": "successful",
+    "28620": "failure",
+    "28621": "failure",
+    "28622": "failure",
+    "28623": "successful",
+    "28624": "failure",
+    "28625": "failure",
+    "28626": "failure",
+    "28627": "successful",
+    "28628": "failure",
+    "28629": "failure",
+    "28630": "infeasible",
+    "28631": "failure",
+    "28632": "failure",
+    "28633": "failure",
+    "28634": "failure",
+    "28635": "failure",
+    "28636": "failure",
+    "28637": "failure",
+    "28638": "failure",
+    "28639": "failure",
+    "28640": "failure",
+    "28641": "failure",
+    "28642": "failure",
+    "28643": "failure",
+    "28644": "failure",
+    "28645": "failure",
+    "28646": "failure",
+    "28647": "infeasible",
+    "28648": "failure",
+    "28649": "failure",
+    "28650": "failure",
+    "28651": "failure",
+    "28652": "failure",
+    "28653": "infeasible",
+    "28654": "failure",
+    "28655": "failure",
+    "28656": "failure",
+    "28657": "failure",
+    "28658": "failure",
+    "28659": "failure",
+    "28660": "failure",
+    "28661": "failure",
+    "28662": "failure",
+    "28663": "failure",
+    "28664": "failure",
+    "28665": "failure",
+    "28666": "failure",
+    "28667": "failure",
+    "28668": "failure",
+    "28669": "failure",
+    "28670": "failure",
+    "28671": "failure",
+    "28672": "successful",
+    "28673": "failure",
+    "28674": "failure",
+    "28675": "failure",
+    "28676": "failure",
+    "28677": "failure",
+    "28678": "failure",
+    "28679": "failure",
+    "28680": "failure",
+    "28681": "failure",
+    "28682": "failure",
+    "28683": "failure",
+    "28684": "successful",
+    "28685": "infeasible",
+    "28686": "failure",
+    "28687": "failure",
+    "28688": "failure",
+    "28689": "failure",
+    "28690": "failure",
+    "28691": "infeasible",
+    "28692": "failure",
+    "28693": "failure",
+    "28694": "successful",
+    "28695": "infeasible",
+    "28696": "infeasible",
+    "28697": "failure",
+    "28698": "failure",
+    "28699": "failure",
+    "28700": "failure",
+    "28701": "failure",
+    "28702": "failure",
+    "28703": "successful",
+    "28704": "failure",
+    "28705": "failure",
+    "28706": "successful",
+    "28707": "failure",
+    "28708": "failure",
+    "28709": "infeasible",
+    "28710": "failure",
+    "28711": "failure",
+    "28712": "failure",
+    "28713": "failure",
+    "28714": "successful",
+    "28715": "failure",
+    "28716": "infeasible",
+    "28717": "failure",
+    "28718": "failure",
+    "28719": "failure",
+    "28720": "failure",
+    "28721": "failure",
+    "28722": "failure",
+    "28723": "failure",
+    "28724": "failure",
+    "28725": "failure",
+    "28726": "failure",
+    "28727": "failure",
+    "28728": "failure",
+    "28729": "failure",
+    "28730": "failure",
+    "28731": "failure",
+    "28732": "failure",
+    "28733": "infeasible",
+    "28734": "failure",
+    "28735": "successful",
+    "28736": "failure",
+    "28737": "failure",
+    "28738": "failure",
+    "28739": "failure",
+    "28740": "failure",
+    "28741": "failure",
+    "28742": "failure",
+    "28743": "failure",
+    "28744": "failure",
+    "28745": "failure",
+    "28746": "successful",
+    "28747": "failure",
+    "28748": "failure",
+    "28749": "successful",
+    "28750": "failure",
+    "28751": "failure",
+    "28752": "failure",
+    "28753": "successful",
+    "28754": "failure",
+    "28755": "failure",
+    "28756": "failure",
+    "28757": "failure",
+    "28758": "failure",
+    "28759": "failure",
+    "28760": "failure",
+    "28761": "failure",
+    "28762": "infeasible",
+    "28763": "failure",
+    "28764": "failure",
+    "28765": "failure",
+    "28766": "failure",
+    "28767": "failure",
+    "28768": "failure",
+    "28769": "failure",
+    "28770": "failure",
+    "28771": "failure",
+    "28772": "failure",
+    "28773": "failure",
+    "28774": "failure",
+    "28775": "failure",
+    "28776": "failure",
+    "28777": "failure",
+    "28778": "successful",
+    "28779": "failure",
+    "28780": "failure",
+    "28781": "failure",
+    "28782": "failure",
+    "28783": "failure",
+    "28784": "failure",
+    "28785": "failure",
+    "28786": "failure",
+    "28787": "failure",
+    "28788": "failure",
+    "28789": "failure",
+    "28790": "failure",
+    "28791": "failure",
+    "28792": "failure",
+    "28793": "failure",
+    "28794": "successful",
+    "28795": "failure",
+    "28796": "failure",
+    "28797": "failure",
+    "28798": "failure",
+    "28799": "failure",
+    "28800": "infeasible",
+    "28801": "failure",
+    "28802": "failure",
+    "28803": "failure",
+    "28804": "failure",
+    "28805": "failure",
+    "28806": "successful",
+    "28807": "failure",
+    "28808": "infeasible",
+    "28809": "failure",
+    "28810": "failure",
+    "28811": "failure",
+    "28812": "failure",
+    "28813": "failure",
+    "28814": "failure",
+    "28815": "failure",
+    "28816": "failure",
+    "28817": "failure",
+    "28818": "failure",
+    "28819": "failure",
+    "28820": "failure",
+    "28821": "failure",
+    "28822": "failure",
+    "28823": "failure",
+    "28824": "failure",
+    "28825": "failure",
+    "28826": "failure",
+    "28827": "failure",
+    "28828": "failure",
+    "28829": "failure",
+    "28830": "failure",
+    "28831": "failure",
+    "28832": "successful",
+    "28833": "failure",
+    "28834": "failure",
+    "28835": "failure",
+    "28836": "failure",
+    "28837": "failure",
+    "28838": "successful",
+    "28839": "failure",
+    "28840": "failure",
+    "28841": "failure",
+    "28842": "failure",
+    "28843": "failure",
+    "28844": "failure",
+    "28845": "failure",
+    "28846": "failure",
+    "28847": "failure",
+    "28848": "failure",
+    "28849": "failure",
+    "28850": "infeasible",
+    "28851": "failure",
+    "28852": "failure",
+    "28853": "failure",
+    "28854": "failure",
+    "28855": "failure",
+    "28856": "failure",
+    "28857": "failure",
+    "28858": "failure",
+    "28859": "failure",
+    "28860": "successful",
+    "28861": "failure",
+    "28862": "failure",
+    "28863": "failure",
+    "28864": "failure",
+    "28865": "failure",
+    "28866": "failure",
+    "28867": "successful",
+    "28868": "failure",
+    "28869": "successful",
+    "28870": "infeasible",
+    "28871": "failure",
+    "28872": "failure",
+    "28873": "successful",
+    "28874": "failure",
+    "28875": "failure",
+    "28876": "failure",
+    "28877": "infeasible",
+    "28878": "failure",
+    "28879": "failure",
+    "28880": "successful",
+    "28881": "failure",
+    "28882": "failure",
+    "28883": "infeasible",
+    "28884": "failure",
+    "28885": "failure",
+    "28886": "failure",
+    "28887": "failure",
+    "28888": "failure",
+    "28889": "infeasible",
+    "28890": "failure",
+    "28891": "infeasible",
+    "28892": "failure",
+    "28893": "failure",
+    "28894": "failure",
+    "28895": "failure",
+    "28896": "failure",
+    "28897": "failure",
+    "28898": "failure",
+    "28899": "failure",
+    "28900": "failure",
+    "28901": "infeasible",
+    "28902": "failure",
+    "28903": "successful",
+    "28904": "failure",
+    "28905": "failure",
+    "28906": "failure",
+    "28907": "failure",
+    "28908": "failure",
+    "28909": "failure",
+    "28910": "failure",
+    "28911": "failure",
+    "28912": "failure",
+    "28913": "failure",
+    "28914": "failure",
+    "28915": "failure",
+    "28916": "failure",
+    "28917": "failure",
+    "28918": "failure",
+    "28919": "successful",
+    "28920": "failure",
+    "28921": "failure",
+    "28922": "failure",
+    "28923": "failure",
+    "28924": "failure",
+    "28925": "infeasible",
+    "28926": "failure",
+    "28927": "failure",
+    "28928": "failure",
+    "28929": "failure",
+    "28930": "failure",
+    "28931": "failure",
+    "28932": "failure",
+    "28933": "failure",
+    "28934": "failure",
+    "28935": "failure",
+    "28936": "infeasible",
+    "28937": "failure",
+    "28938": "successful",
+    "28939": "failure",
+    "28940": "failure",
+    "28941": "failure",
+    "28942": "successful",
+    "28943": "infeasible",
+    "28944": "failure",
+    "28945": "successful",
+    "28946": "failure",
+    "28947": "infeasible",
+    "28948": "failure",
+    "28949": "failure",
+    "28950": "failure",
+    "28951": "failure",
+    "28952": "infeasible",
+    "28953": "failure",
+    "28954": "failure",
+    "28955": "failure",
+    "28956": "failure",
+    "28957": "failure",
+    "28958": "successful",
+    "28959": "failure",
+    "28960": "failure",
+    "28961": "successful",
+    "28962": "failure",
+    "28963": "failure",
+    "28964": "failure",
+    "28965": "failure",
+    "28966": "failure",
+    "28967": "failure",
+    "28968": "failure",
+    "28969": "failure",
+    "28970": "successful",
+    "28971": "already_complete",
+    "28972": "failure",
+    "28973": "failure",
+    "28974": "failure",
+    "28975": "failure",
+    "28976": "failure",
+    "28977": "failure",
+    "28978": "failure",
+    "28979": "successful",
+    "28980": "failure",
+    "28981": "failure",
+    "28982": "successful",
+    "28983": "failure",
+    "28984": "failure",
+    "28985": "infeasible",
+    "28986": "failure",
+    "28987": "failure",
+    "28988": "failure",
+    "28989": "failure",
+    "28990": "failure",
+    "28991": "failure",
+    "28992": "failure",
+    "28993": "failure",
+    "28994": "failure",
+    "28995": "failure",
+    "28996": "failure",
+    "28997": "failure",
+    "28998": "failure",
+    "28999": "failure",
+    "29000": "failure",
+    "29001": "infeasible",
+    "29002": "failure",
+    "29003": "failure",
+    "29004": "failure",
+    "29005": "failure",
+    "29006": "failure",
+    "29007": "failure",
+    "29008": "failure",
+    "29009": "failure",
+    "29010": "failure",
+    "29011": "failure",
+    "29012": "successful",
+    "29013": "failure",
+    "29014": "failure",
+    "29015": "failure",
+    "29016": "failure",
+    "29017": "failure",
+    "29018": "successful",
+    "29019": "failure",
+    "29020": "failure",
+    "29021": "failure",
+    "29022": "already_complete",
+    "29023": "failure",
+    "29024": "infeasible",
+    "29025": "failure",
+    "29026": "failure",
+    "29027": "failure",
+    "29028": "failure",
+    "29029": "failure",
+    "29030": "successful",
+    "29031": "successful",
+    "29032": "failure",
+    "29033": "failure",
+    "29034": "infeasible",
+    "29035": "failure",
+    "29036": "failure",
+    "29037": "infeasible",
+    "29038": "already_complete",
+    "29039": "failure",
+    "29040": "failure",
+    "29041": "failure",
+    "29042": "infeasible",
+    "29043": "failure",
+    "29044": "failure",
+    "29045": "successful",
+    "29046": "failure",
+    "29047": "failure",
+    "29048": "failure",
+    "29049": "successful",
+    "29050": "failure",
+    "29051": "failure",
+    "29052": "failure",
+    "29053": "infeasible",
+    "29054": "failure",
+    "29055": "failure",
+    "29056": "failure",
+    "29057": "infeasible",
+    "29058": "failure",
+    "29059": "successful",
+    "29060": "failure",
+    "29061": "failure",
+    "29062": "failure",
+    "29063": "failure",
+    "29064": "failure",
+    "29065": "successful",
+    "29066": "failure",
+    "29067": "failure",
+    "29068": "failure",
+    "29069": "failure",
+    "29070": "failure",
+    "29071": "failure",
+    "29072": "failure",
+    "29073": "failure",
+    "29074": "successful",
+    "29075": "failure",
+    "29076": "failure",
+    "29077": "failure",
+    "29078": "failure",
+    "29079": "failure",
+    "29080": "failure",
+    "29081": "failure",
+    "29082": "infeasible",
+    "29083": "successful",
+    "29084": "failure",
+    "29085": "failure",
+    "29086": "failure",
+    "29087": "failure",
+    "29088": "failure",
+    "29089": "successful",
+    "29090": "failure",
+    "29091": "successful",
+    "29092": "failure",
+    "29093": "failure",
+    "29094": "failure",
+    "29095": "infeasible",
+    "29096": "failure",
+    "29097": "failure",
+    "29098": "successful",
+    "29099": "failure",
+    "29100": "failure",
+    "29101": "failure",
+    "29102": "infeasible",
+    "29103": "failure",
+    "29104": "failure",
+    "29105": "failure",
+    "29106": "failure",
+    "29107": "failure",
+    "29108": "failure",
+    "29109": "failure",
+    "29110": "failure",
+    "29111": "failure",
+    "29112": "failure",
+    "29113": "failure",
+    "29114": "failure",
+    "29115": "failure",
+    "29116": "failure",
+    "29117": "failure",
+    "29118": "failure",
+    "29119": "failure",
+    "29120": "failure",
+    "29121": "failure",
+    "29122": "failure",
+    "29123": "failure",
+    "29124": "infeasible",
+    "29125": "failure",
+    "29126": "failure",
+    "29127": "failure",
+    "29128": "failure",
+    "29129": "failure",
+    "29130": "failure",
+    "29131": "successful",
+    "29132": "failure",
+    "29133": "failure",
+    "29134": "failure",
+    "29135": "failure",
+    "29136": "successful",
+    "29137": "failure",
+    "29138": "successful",
+    "29139": "infeasible",
+    "29140": "failure",
+    "29141": "failure",
+    "29142": "infeasible",
+    "29143": "failure",
+    "29144": "failure",
+    "29145": "failure",
+    "29146": "failure",
+    "29147": "failure",
+    "29148": "failure",
+    "29149": "failure",
+    "29150": "failure",
+    "29151": "failure",
+    "29152": "failure",
+    "29153": "failure",
+    "29154": "failure",
+    "29155": "successful",
+    "29156": "failure",
+    "29157": "failure",
+    "29158": "failure",
+    "29159": "failure",
+    "29160": "failure",
+    "29161": "failure",
+    "29162": "infeasible",
+    "29163": "failure",
+    "29164": "failure",
+    "29165": "failure",
+    "29166": "failure",
+    "29167": "infeasible",
+    "29168": "infeasible",
+    "29169": "failure",
+    "29170": "successful",
+    "29171": "failure",
+    "29172": "failure",
+    "29173": "failure",
+    "29174": "successful",
+    "29175": "failure",
+    "29176": "failure",
+    "29177": "failure",
+    "29178": "failure",
+    "29179": "successful",
+    "29180": "failure",
+    "29181": "already_complete",
+    "29182": "successful",
+    "29183": "failure",
+    "29184": "failure",
+    "29185": "failure",
+    "29186": "failure",
+    "29187": "failure",
+    "29188": "failure",
+    "29189": "failure",
+    "29190": "failure",
+    "29191": "failure",
+    "29192": "failure",
+    "29193": "successful",
+    "29194": "failure",
+    "29195": "successful",
+    "29196": "failure",
+    "29197": "failure",
+    "29198": "infeasible",
+    "29199": "failure",
+    "29200": "failure",
+    "29201": "failure",
+    "29202": "infeasible",
+    "29203": "failure",
+    "29204": "failure",
+    "29205": "failure",
+    "29206": "already_complete",
+    "29207": "failure",
+    "29208": "failure",
+    "29209": "failure",
+    "29210": "failure",
+    "29211": "failure",
+    "29212": "failure",
+    "29213": "failure",
+    "29214": "failure",
+    "29215": "failure",
+    "29216": "failure",
+    "29217": "failure",
+    "29218": "failure",
+    "29219": "failure",
+    "29220": "failure",
+    "29221": "failure",
+    "29222": "failure",
+    "29223": "failure",
+    "29224": "failure",
+    "29225": "successful",
+    "29226": "failure",
+    "29227": "failure",
+    "29228": "failure",
+    "29229": "infeasible",
+    "29230": "infeasible",
+    "29231": "failure",
+    "29232": "failure",
+    "29233": "failure",
+    "29234": "failure",
+    "29235": "failure",
+    "29236": "failure",
+    "29237": "failure",
+    "29238": "failure",
+    "29239": "failure",
+    "29240": "infeasible",
+    "29241": "failure",
+    "29242": "failure",
+    "29243": "infeasible",
+    "29244": "failure",
+    "29245": "failure",
+    "29246": "failure",
+    "29247": "failure",
+    "29248": "failure",
+    "29249": "failure",
+    "29250": "failure",
+    "29251": "failure",
+    "29252": "failure",
+    "29253": "failure",
+    "29254": "failure",
+    "29255": "failure",
+    "29256": "successful",
+    "29257": "failure",
+    "29258": "failure",
+    "29259": "failure",
+    "29260": "failure",
+    "29261": "failure",
+    "29262": "failure",
+    "29263": "successful",
+    "29264": "failure",
+    "29265": "failure",
+    "29266": "failure",
+    "29267": "successful",
+    "29268": "failure",
+    "29269": "failure",
+    "29270": "failure",
+    "29271": "failure",
+    "29272": "failure",
+    "29273": "failure",
+    "29274": "failure",
+    "29275": "failure",
+    "29276": "failure",
+    "29277": "failure",
+    "29278": "successful",
+    "29279": "infeasible",
+    "29280": "failure",
+    "29281": "successful",
+    "29282": "successful",
+    "29283": "failure",
+    "29284": "failure",
+    "29285": "failure",
+    "29286": "failure",
+    "29287": "failure",
+    "29288": "failure",
+    "29289": "failure",
+    "29290": "successful",
+    "29291": "failure",
+    "29292": "failure",
+    "29293": "failure",
+    "29294": "failure",
+    "29295": "failure",
+    "29296": "failure",
+    "29297": "infeasible",
+    "29298": "infeasible",
+    "29299": "failure",
+    "29300": "failure",
+    "29301": "failure",
+    "29302": "failure",
+    "29303": "failure",
+    "29304": "failure",
+    "29305": "failure",
+    "29306": "failure",
+    "29307": "failure",
+    "29308": "successful",
+    "29309": "failure",
+    "29310": "failure",
+    "29311": "failure",
+    "29312": "failure",
+    "29313": "failure",
+    "29314": "failure",
+    "29315": "failure",
+    "29316": "failure",
+    "29317": "failure",
+    "29318": "failure",
+    "29319": "infeasible",
+    "29320": "failure",
+    "29321": "failure",
+    "29322": "failure",
+    "29323": "failure",
+    "29324": "failure",
+    "29325": "failure",
+    "29326": "failure",
+    "29327": "failure",
+    "29328": "failure",
+    "29329": "failure",
+    "29330": "failure",
+    "29331": "failure",
+    "29332": "failure",
+    "29333": "failure",
+    "29334": "failure",
+    "29335": "failure",
+    "29336": "failure",
+    "29337": "failure",
+    "29338": "failure",
+    "29339": "failure",
+    "29340": "failure",
+    "29341": "failure",
+    "29342": "failure",
+    "29343": "failure",
+    "29344": "failure",
+    "29345": "failure",
+    "29346": "failure",
+    "29347": "failure",
+    "29348": "failure",
+    "29349": "successful",
+    "29350": "failure",
+    "29351": "successful",
+    "29352": "failure",
+    "29353": "failure",
+    "29354": "successful",
+    "29355": "failure",
+    "29356": "failure",
+    "29357": "failure",
+    "29358": "failure",
+    "29359": "infeasible",
+    "29360": "failure",
+    "29361": "failure",
+    "29362": "failure",
+    "29363": "successful",
+    "29364": "failure",
+    "29365": "failure",
+    "29366": "failure",
+    "29367": "failure",
+    "29368": "failure",
+    "29369": "failure",
+    "29370": "failure",
+    "29371": "failure",
+    "29372": "failure",
+    "29373": "failure",
+    "29374": "failure",
+    "29375": "failure",
+    "29376": "failure",
+    "29377": "failure",
+    "29378": "successful",
+    "29379": "failure",
+    "29380": "failure",
+    "29381": "failure",
+    "29382": "failure",
+    "29383": "failure",
+    "29384": "failure",
+    "29385": "successful",
+    "29386": "failure",
+    "29387": "failure",
+    "29388": "failure",
+    "29389": "successful",
+    "29390": "infeasible",
+    "29391": "failure",
+    "29392": "failure",
+    "29393": "failure",
+    "29394": "failure",
+    "29395": "failure",
+    "29396": "failure",
+    "29397": "successful",
+    "29398": "failure",
+    "29399": "failure",
+    "29400": "failure",
+    "29401": "failure",
+    "29402": "failure",
+    "29403": "failure",
+    "29404": "failure",
+    "29405": "infeasible",
+    "29406": "failure",
+    "29407": "failure",
+    "29408": "failure",
+    "29409": "failure",
+    "29410": "failure",
+    "29411": "failure",
+    "29412": "failure",
+    "29413": "failure",
+    "29414": "failure",
+    "29415": "infeasible",
+    "29416": "failure",
+    "29417": "failure",
+    "29418": "failure",
+    "29419": "successful",
+    "29420": "failure",
+    "29421": "infeasible",
+    "29422": "failure",
+    "29423": "failure",
+    "29424": "failure",
+    "29425": "successful",
+    "29426": "failure",
+    "29427": "infeasible",
+    "29428": "successful",
+    "29429": "successful",
+    "29430": "successful",
+    "29431": "failure",
+    "29432": "successful",
+    "29433": "failure",
+    "29434": "failure",
+    "29435": "failure",
+    "29436": "failure",
+    "29437": "failure",
+    "29438": "failure",
+    "29439": "failure",
+    "29440": "successful",
+    "29441": "failure",
+    "29442": "failure",
+    "29443": "failure",
+    "29444": "infeasible",
+    "29445": "failure",
+    "29446": "failure",
+    "29447": "infeasible",
+    "29448": "failure",
+    "29449": "failure",
+    "29450": "infeasible",
+    "29451": "failure",
+    "29452": "failure",
+    "29453": "failure",
+    "29454": "failure",
+    "29455": "failure",
+    "29456": "infeasible",
+    "29457": "infeasible",
+    "29458": "failure",
+    "29459": "failure",
+    "29460": "failure",
+    "29461": "failure",
+    "29462": "failure",
+    "29463": "infeasible",
+    "29464": "infeasible",
+    "29465": "failure",
+    "29466": "failure",
+    "29467": "failure",
+    "29468": "successful",
+    "29469": "failure",
+    "29470": "failure",
+    "29471": "failure",
+    "29472": "infeasible",
+    "29473": "failure",
+    "29474": "failure",
+    "29475": "failure",
+    "29476": "failure",
+    "29477": "failure",
+    "29478": "infeasible",
+    "29479": "failure",
+    "29480": "failure",
+    "29481": "failure",
+    "29482": "failure",
+    "29483": "infeasible",
+    "29484": "failure",
+    "29485": "failure",
+    "29486": "failure",
+    "29487": "failure",
+    "29488": "failure",
+    "29489": "successful",
+    "29490": "failure",
+    "29491": "failure",
+    "29492": "failure",
+    "29493": "failure",
+    "29494": "failure",
+    "29495": "failure",
+    "29496": "successful",
+    "29497": "failure",
+    "29498": "failure",
+    "29499": "successful",
+    "29500": "failure",
+    "29501": "failure",
+    "29502": "failure",
+    "29503": "failure",
+    "29504": "failure",
+    "29505": "infeasible",
+    "29506": "failure",
+    "29507": "failure",
+    "29508": "failure",
+    "29509": "failure",
+    "29510": "infeasible",
+    "29511": "successful",
+    "29512": "failure",
+    "29513": "failure",
+    "29514": "failure",
+    "29515": "successful",
+    "29516": "failure",
+    "29517": "infeasible",
+    "29518": "failure",
+    "29519": "failure",
+    "29520": "failure",
+    "29521": "failure",
+    "29522": "failure",
+    "29523": "infeasible",
+    "29524": "failure",
+    "29525": "failure",
+    "29526": "successful",
+    "29527": "failure",
+    "29528": "infeasible",
+    "29529": "failure",
+    "29530": "failure",
+    "29531": "failure",
+    "29532": "failure",
+    "29533": "failure",
+    "29534": "successful",
+    "29535": "failure",
+    "29536": "failure",
+    "29537": "failure",
+    "29538": "failure",
+    "29539": "failure",
+    "29540": "failure",
+    "29541": "failure",
+    "29542": "successful",
+    "29543": "failure",
+    "29544": "infeasible",
+    "29545": "failure",
+    "29546": "failure",
+    "29547": "failure",
+    "29548": "failure",
+    "29549": "failure",
+    "29550": "failure",
+    "29551": "failure",
+    "29552": "successful",
+    "29553": "failure",
+    "29554": "failure",
+    "29555": "failure",
+    "29556": "failure",
+    "29557": "infeasible",
+    "29558": "failure",
+    "29559": "failure",
+    "29560": "failure",
+    "29561": "infeasible",
+    "29562": "failure",
+    "29563": "failure",
+    "29564": "failure",
+    "29565": "failure",
+    "29566": "failure",
+    "29567": "failure",
+    "29568": "failure",
+    "29569": "failure",
+    "29570": "failure",
+    "29571": "failure",
+    "29572": "failure",
+    "29573": "failure",
+    "29574": "failure",
+    "29575": "infeasible",
+    "29576": "failure",
+    "29577": "failure",
+    "29578": "failure",
+    "29579": "failure",
+    "29580": "infeasible",
+    "29581": "failure",
+    "29582": "failure",
+    "29583": "infeasible",
+    "29584": "infeasible",
+    "29585": "failure",
+    "29586": "failure",
+    "29587": "failure",
+    "29588": "failure",
+    "29589": "failure",
+    "29590": "failure",
+    "29591": "failure",
+    "29592": "failure",
+    "29593": "failure",
+    "29594": "failure",
+    "29595": "failure",
+    "29596": "failure",
+    "29597": "failure",
+    "29598": "infeasible",
+    "29599": "successful",
+    "29600": "infeasible",
+    "29601": "failure",
+    "29602": "failure",
+    "29603": "failure",
+    "29604": "failure",
+    "29605": "failure",
+    "29606": "infeasible",
+    "29607": "failure",
+    "29608": "failure",
+    "29609": "infeasible",
+    "29610": "failure",
+    "29611": "failure",
+    "29612": "failure",
+    "29613": "failure",
+    "29614": "successful",
+    "29615": "failure",
+    "29616": "successful",
+    "29617": "failure",
+    "29618": "successful",
+    "29619": "infeasible",
+    "29620": "failure",
+    "29621": "successful",
+    "29622": "failure",
+    "29623": "failure",
+    "29624": "failure",
+    "29625": "infeasible",
+    "29626": "failure",
+    "29627": "failure",
+    "29628": "infeasible",
+    "29629": "failure",
+    "29630": "failure",
+    "29631": "failure",
+    "29632": "failure",
+    "29633": "failure",
+    "29634": "failure",
+    "29635": "infeasible",
+    "29636": "failure",
+    "29637": "failure",
+    "29638": "failure",
+    "29639": "failure",
+    "29640": "failure",
+    "29641": "failure",
+    "29642": "failure",
+    "29643": "successful",
+    "29644": "failure",
+    "29645": "failure",
+    "29646": "failure",
+    "29647": "failure",
+    "29648": "failure",
+    "29649": "failure",
+    "29650": "failure",
+    "29651": "failure",
+    "29652": "failure",
+    "29653": "failure",
+    "29654": "failure",
+    "29655": "failure",
+    "29656": "failure",
+    "29657": "successful",
+    "29658": "failure",
+    "29659": "failure",
+    "29660": "failure",
+    "29661": "failure",
+    "29662": "already_complete",
+    "29663": "failure",
+    "29664": "failure",
+    "29665": "failure",
+    "29666": "failure",
+    "29667": "successful",
+    "29668": "failure",
+    "29669": "failure",
+    "29670": "failure",
+    "29671": "failure",
+    "29672": "failure",
+    "29673": "failure",
+    "29674": "infeasible",
+    "29675": "failure",
+    "29676": "successful",
+    "29677": "failure",
+    "29678": "failure",
+    "29679": "failure",
+    "29680": "failure",
+    "29681": "failure",
+    "29682": "failure",
+    "29683": "infeasible",
+    "29684": "failure",
+    "29685": "successful",
+    "29686": "failure",
+    "29687": "failure",
+    "29688": "failure",
+    "29689": "failure",
+    "29690": "failure",
+    "29691": "failure",
+    "29692": "failure",
+    "29693": "failure",
+    "29694": "failure",
+    "29695": "failure",
+    "29696": "failure",
+    "29697": "successful",
+    "29698": "failure",
+    "29699": "failure",
+    "29700": "failure",
+    "29701": "failure",
+    "29702": "failure",
+    "29703": "failure",
+    "29704": "infeasible",
+    "29705": "failure",
+    "29706": "failure",
+    "29707": "failure",
+    "29708": "failure",
+    "29709": "failure",
+    "29710": "infeasible",
+    "29711": "failure",
+    "29712": "successful",
+    "29713": "successful",
+    "29714": "failure",
+    "29715": "failure",
+    "29716": "failure",
+    "29717": "failure",
+    "29718": "failure",
+    "29719": "successful",
+    "29720": "infeasible",
+    "29721": "successful",
+    "29722": "successful",
+    "29723": "failure",
+    "29724": "failure",
+    "29725": "failure",
+    "29726": "successful",
+    "29727": "failure",
+    "29728": "failure",
+    "29729": "successful",
+    "29730": "successful",
+    "29731": "failure",
+    "29732": "failure",
+    "29733": "failure",
+    "29734": "failure",
+    "29735": "failure",
+    "29736": "failure",
+    "29737": "infeasible",
+    "29738": "successful",
+    "29739": "failure",
+    "29740": "failure",
+    "29741": "failure",
+    "29742": "failure",
+    "29743": "failure",
+    "29744": "failure",
+    "29745": "failure",
+    "29746": "failure",
+    "29747": "failure",
+    "29748": "successful",
+    "29749": "infeasible",
+    "29750": "infeasible",
+    "29751": "failure",
+    "29752": "failure",
+    "29753": "failure",
+    "29754": "failure",
+    "29755": "failure",
+    "29756": "failure",
+    "29757": "infeasible",
+    "29758": "failure",
+    "29759": "failure",
+    "29760": "failure",
+    "29761": "failure",
+    "29762": "failure",
+    "29763": "failure",
+    "29764": "failure",
+    "29765": "failure",
+    "29766": "successful",
+    "29767": "failure",
+    "29768": "failure",
+    "29769": "failure",
+    "29770": "failure",
+    "29771": "infeasible",
+    "29772": "infeasible",
+    "29773": "failure",
+    "29774": "failure",
+    "29775": "failure",
+    "29776": "failure",
+    "29777": "failure",
+    "29778": "failure",
+    "29779": "failure",
+    "29780": "failure",
+    "29781": "infeasible",
+    "29782": "failure",
+    "29783": "failure",
+    "29784": "successful",
+    "29785": "infeasible",
+    "29786": "failure",
+    "29787": "failure",
+    "29788": "failure",
+    "29789": "failure",
+    "29790": "failure",
+    "29791": "infeasible",
+    "29792": "infeasible",
+    "29793": "failure",
+    "29794": "infeasible",
+    "29795": "failure",
+    "29796": "failure",
+    "29797": "failure",
+    "29798": "failure",
+    "29799": "failure",
+    "29800": "failure",
+    "29801": "failure",
+    "29802": "failure",
+    "29803": "failure",
+    "29804": "failure",
+    "29805": "failure",
+    "29806": "failure",
+    "29807": "successful",
+    "29808": "failure",
+    "29809": "failure",
+    "29810": "failure",
+    "29811": "failure",
+    "29812": "infeasible",
+    "29813": "failure",
+    "29814": "failure",
+    "29815": "failure",
+    "29816": "failure",
+    "29817": "failure",
+    "29818": "infeasible",
+    "29819": "infeasible",
+    "29820": "successful",
+    "29821": "infeasible",
+    "29822": "failure",
+    "29823": "failure",
+    "29824": "failure",
+    "29825": "failure",
+    "29826": "failure",
+    "29827": "failure",
+    "29828": "failure",
+    "29829": "failure",
+    "29830": "failure",
+    "29831": "failure",
+    "29832": "failure",
+    "29833": "infeasible",
+    "29834": "failure",
+    "29835": "failure",
+    "29836": "failure",
+    "29837": "failure",
+    "29838": "failure",
+    "29839": "failure",
+    "29840": "infeasible",
+    "29841": "failure",
+    "29842": "failure",
+    "29843": "failure",
+    "29844": "failure",
+    "29845": "failure",
+    "29846": "failure",
+    "29847": "failure",
+    "29848": "failure",
+    "29849": "infeasible",
+    "29850": "failure",
+    "29851": "infeasible",
+    "29852": "failure",
+    "29853": "failure",
+    "29854": "failure",
+    "29855": "failure",
+    "29856": "failure",
+    "29857": "failure",
+    "29858": "failure",
+    "29859": "failure",
+    "29860": "failure",
+    "29861": "successful",
+    "29862": "failure",
+    "29863": "failure",
+    "29864": "failure",
+    "29865": "failure",
+    "29866": "failure",
+    "29867": "failure",
+    "29868": "failure",
+    "29869": "failure",
+    "29870": "failure",
+    "29871": "failure",
+    "29872": "failure",
+    "29873": "successful",
+    "29874": "failure",
+    "29875": "infeasible",
+    "29876": "infeasible",
+    "29877": "failure",
+    "29878": "failure",
+    "29879": "failure",
+    "29880": "infeasible",
+    "29881": "failure",
+    "29882": "successful",
+    "29883": "failure",
+    "29884": "infeasible",
+    "29885": "successful",
+    "29886": "failure",
+    "29887": "successful",
+    "29888": "successful",
+    "29889": "failure",
+    "29890": "failure",
+    "29891": "failure",
+    "29892": "failure",
+    "29893": "failure",
+    "29894": "failure",
+    "29895": "failure",
+    "29896": "successful",
+    "29897": "successful",
+    "29898": "failure",
+    "29899": "failure",
+    "29900": "failure",
+    "29901": "failure",
+    "29902": "failure",
+    "29903": "failure",
+    "29904": "failure",
+    "29905": "failure",
+    "29906": "failure",
+    "29907": "failure",
+    "29908": "successful",
+    "29909": "failure",
+    "29910": "failure",
+    "29911": "infeasible",
+    "29912": "successful",
+    "29913": "failure",
+    "29914": "failure",
+    "29915": "failure",
+    "29916": "failure",
+    "29917": "failure",
+    "29918": "failure",
+    "29919": "infeasible",
+    "29920": "failure",
+    "29921": "successful",
+    "29922": "failure",
+    "29923": "failure",
+    "29924": "failure",
+    "29925": "successful",
+    "29926": "failure",
+    "29927": "failure",
+    "29928": "successful",
+    "29929": "failure",
+    "29930": "successful",
+    "29931": "successful",
+    "29932": "already_complete",
+    "29933": "failure",
+    "29934": "failure",
+    "29935": "failure",
+    "29936": "infeasible",
+    "29937": "failure",
+    "29938": "failure",
+    "29939": "successful",
+    "29940": "failure",
+    "29941": "failure",
+    "29942": "failure",
+    "29943": "successful",
+    "29944": "failure",
+    "29945": "failure",
+    "29946": "failure",
+    "29947": "failure",
+    "29948": "failure",
+    "29949": "failure",
+    "29950": "failure",
+    "29951": "failure",
+    "29952": "failure",
+    "29953": "infeasible",
+    "29954": "failure",
+    "29955": "failure",
+    "29956": "failure",
+    "29957": "failure",
+    "29958": "failure",
+    "29959": "failure",
+    "29960": "failure",
+    "29961": "failure",
+    "29962": "failure",
+    "29963": "failure",
+    "29964": "failure",
+    "29965": "failure",
+    "29966": "infeasible",
+    "29967": "successful",
+    "29968": "failure",
+    "29969": "failure",
+    "29970": "failure",
+    "29971": "infeasible",
+    "29972": "failure",
+    "29973": "successful",
+    "29974": "failure",
+    "29975": "failure",
+    "29976": "failure",
+    "29977": "failure",
+    "29978": "failure",
+    "29979": "failure",
+    "29980": "failure",
+    "29981": "failure",
+    "29982": "failure",
+    "29983": "failure",
+    "29984": "failure",
+    "29985": "failure",
+    "29986": "failure",
+    "29987": "successful",
+    "29988": "failure",
+    "29989": "successful",
+    "29990": "infeasible",
+    "29991": "infeasible",
+    "29992": "infeasible",
+    "29993": "successful",
+    "29994": "failure",
+    "29995": "failure",
+    "29996": "failure",
+    "29997": "infeasible",
+    "29998": "failure",
+    "29999": "failure",
+    "30000": "infeasible",
+    "30001": "failure",
+    "30002": "failure",
+    "30003": "failure",
+    "30004": "failure",
+    "30005": "failure",
+    "30006": "failure",
+    "30007": "failure",
+    "30008": "failure",
+    "30009": "failure",
+    "30010": "infeasible",
+    "30011": "failure",
+    "30012": "infeasible",
+    "30013": "failure",
+    "30014": "failure",
+    "30015": "failure",
+    "30016": "successful",
+    "30017": "failure",
+    "30018": "failure",
+    "30019": "failure",
+    "30020": "failure",
+    "30021": "failure",
+    "30022": "failure",
+    "30023": "failure",
+    "30024": "successful",
+    "30025": "failure",
+    "30026": "successful",
+    "30027": "infeasible",
+    "30028": "failure",
+    "30029": "infeasible",
+    "30030": "failure",
+    "30031": "infeasible",
+    "30032": "failure",
+    "30033": "failure",
+    "30034": "failure",
+    "30035": "successful",
+    "30036": "infeasible",
+    "30037": "infeasible",
+    "30038": "failure",
+    "30039": "failure",
+    "30040": "failure",
+    "30041": "failure",
+    "30042": "failure",
+    "30043": "infeasible",
+    "30044": "failure",
+    "30045": "failure",
+    "30046": "failure",
+    "30047": "failure",
+    "30048": "failure",
+    "30049": "failure",
+    "30050": "successful",
+    "30051": "successful",
+    "30052": "failure",
+    "30053": "failure",
+    "30054": "infeasible",
+    "30055": "successful",
+    "30056": "failure",
+    "30057": "infeasible",
+    "30058": "failure",
+    "30059": "failure",
+    "30060": "infeasible",
+    "30061": "infeasible",
+    "30062": "failure",
+    "30063": "failure",
+    "30064": "failure",
+    "30065": "infeasible",
+    "30066": "successful",
+    "30067": "failure",
+    "30068": "infeasible",
+    "30069": "failure",
+    "30070": "infeasible",
+    "30071": "failure",
+    "30072": "failure",
+    "30073": "failure",
+    "30074": "failure",
+    "30075": "infeasible",
+    "30076": "successful",
+    "30077": "successful",
+    "30078": "failure",
+    "30079": "successful",
+    "30080": "failure",
+    "30081": "failure",
+    "30082": "successful",
+    "30083": "infeasible",
+    "30084": "successful",
+    "30085": "failure",
+    "30086": "successful",
+    "30087": "failure",
+    "30088": "failure",
+    "30089": "failure",
+    "30090": "failure",
+    "30091": "successful",
+    "30092": "failure",
+    "30093": "successful",
+    "30094": "failure",
+    "30095": "failure",
+    "30096": "failure",
+    "30097": "failure",
+    "30098": "failure",
+    "30099": "failure",
+    "30100": "infeasible",
+    "30101": "failure",
+    "30102": "failure",
+    "30103": "failure",
+    "30104": "successful",
+    "30105": "failure",
+    "30106": "failure",
+    "30107": "failure",
+    "30108": "successful",
+    "30109": "failure",
+    "30110": "failure",
+    "30111": "failure",
+    "30112": "successful",
+    "30113": "failure",
+    "30114": "failure",
+    "30115": "successful",
+    "30116": "failure",
+    "30117": "failure",
+    "30118": "failure",
+    "30119": "infeasible",
+    "30120": "failure",
+    "30121": "successful",
+    "30122": "failure",
+    "30123": "successful",
+    "30124": "failure",
+    "30125": "failure",
+    "30126": "failure",
+    "30127": "failure",
+    "30128": "infeasible",
+    "30129": "failure",
+    "30130": "failure",
+    "30131": "failure",
+    "30132": "infeasible",
+    "30133": "failure",
+    "30134": "infeasible",
+    "30135": "failure",
+    "30136": "failure",
+    "30137": "failure",
+    "30138": "successful",
+    "30139": "failure",
+    "30140": "infeasible",
+    "30141": "failure",
+    "30142": "infeasible",
+    "30143": "failure",
+    "30144": "failure",
+    "30145": "failure",
+    "30146": "failure",
+    "30147": "failure",
+    "30148": "successful",
+    "30149": "successful",
+    "30150": "infeasible",
+    "30151": "failure",
+    "30152": "failure",
+    "30153": "failure",
+    "30154": "failure",
+    "30155": "failure",
+    "30156": "failure",
+    "30157": "failure",
+    "30158": "failure",
+    "30159": "infeasible",
+    "30160": "failure",
+    "30161": "successful",
+    "30162": "failure",
+    "30163": "failure",
+    "30164": "infeasible",
+    "30165": "failure",
+    "30166": "failure",
+    "30167": "failure",
+    "30168": "failure",
+    "30169": "failure",
+    "30170": "infeasible",
+    "30171": "infeasible",
+    "30172": "failure",
+    "30173": "failure",
+    "30174": "failure",
+    "30175": "failure",
+    "30176": "failure",
+    "30177": "failure",
+    "30178": "failure",
+    "30179": "failure",
+    "30180": "failure",
+    "30181": "failure",
+    "30182": "failure",
+    "30183": "infeasible",
+    "30184": "failure",
+    "30185": "failure",
+    "30186": "failure",
+    "30187": "failure",
+    "30188": "infeasible",
+    "30189": "failure",
+    "30190": "failure",
+    "30191": "failure",
+    "30192": "failure",
+    "30193": "failure",
+    "30194": "failure",
+    "30195": "failure",
+    "30196": "failure",
+    "30197": "failure",
+    "30198": "failure",
+    "30199": "infeasible",
+    "30200": "failure",
+    "30201": "successful",
+    "30202": "successful",
+    "30203": "failure",
+    "30204": "failure",
+    "30205": "failure",
+    "30206": "infeasible",
+    "30207": "successful",
+    "30208": "failure",
+    "30209": "failure",
+    "30210": "failure",
+    "30211": "failure",
+    "30212": "failure",
+    "30213": "failure",
+    "30214": "failure",
+    "30215": "failure",
+    "30216": "failure",
+    "30217": "failure",
+    "30218": "failure",
+    "30219": "failure",
+    "30220": "successful",
+    "30221": "failure",
+    "30222": "infeasible",
+    "30223": "failure",
+    "30224": "failure",
+    "30225": "failure",
+    "30226": "failure",
+    "30227": "failure",
+    "30228": "failure",
+    "30229": "failure",
+    "30230": "failure",
+    "30231": "failure",
+    "30232": "successful",
+    "30233": "failure",
+    "30234": "failure",
+    "30235": "failure",
+    "30236": "failure",
+    "30237": "failure",
+    "30238": "successful",
+    "30239": "successful",
+    "30240": "failure",
+    "30241": "failure",
+    "30242": "failure",
+    "30243": "failure",
+    "30244": "successful",
+    "30245": "failure",
+    "30246": "failure",
+    "30247": "failure",
+    "30248": "failure",
+    "30249": "failure",
+    "30250": "failure",
+    "30251": "failure",
+    "30252": "failure",
+    "30253": "successful",
+    "30254": "failure",
+    "30255": "failure",
+    "30256": "failure",
+    "30257": "failure",
+    "30258": "failure",
+    "30259": "failure",
+    "30260": "failure",
+    "30261": "failure",
+    "30262": "failure",
+    "30263": "failure",
+    "30264": "failure",
+    "30265": "failure",
+    "30266": "failure",
+    "30267": "infeasible",
+    "30268": "failure",
+    "30269": "infeasible",
+    "30270": "failure",
+    "30271": "failure",
+    "30272": "infeasible",
+    "30273": "failure",
+    "30274": "failure",
+    "30275": "failure",
+    "30276": "failure",
+    "30277": "failure",
+    "30278": "failure",
+    "30279": "failure",
+    "30280": "failure",
+    "30281": "infeasible",
+    "30282": "infeasible",
+    "30283": "infeasible",
+    "30284": "failure",
+    "30285": "infeasible",
+    "30286": "successful",
+    "30287": "failure",
+    "30288": "failure",
+    "30289": "successful",
+    "30290": "failure",
+    "30291": "infeasible",
+    "30292": "infeasible",
+    "30293": "failure",
+    "30294": "failure",
+    "30295": "infeasible",
+    "30296": "failure",
+    "30297": "successful",
+    "30298": "failure",
+    "30299": "infeasible",
+    "30300": "successful",
+    "30301": "failure",
+    "30302": "failure",
+    "30303": "failure",
+    "30304": "failure",
+    "30305": "failure",
+    "30306": "failure",
+    "30307": "failure",
+    "30308": "infeasible",
+    "30309": "failure",
+    "30310": "failure",
+    "30311": "failure",
+    "30312": "failure",
+    "30313": "failure",
+    "30314": "failure",
+    "30315": "failure",
+    "30316": "successful",
+    "30317": "failure",
+    "30318": "failure",
+    "30319": "failure",
+    "30320": "failure",
+    "30321": "successful",
+    "30322": "successful",
+    "30323": "failure",
+    "30324": "failure",
+    "30325": "failure",
+    "30326": "failure",
+    "30327": "failure",
+    "30328": "failure",
+    "30329": "failure",
+    "30330": "failure",
+    "30331": "failure",
+    "30332": "failure",
+    "30333": "infeasible",
+    "30334": "successful",
+    "30335": "failure",
+    "30336": "failure",
+    "30337": "failure",
+    "30338": "failure",
+    "30339": "failure",
+    "30340": "infeasible",
+    "30341": "failure",
+    "30342": "infeasible",
+    "30343": "failure",
+    "30344": "failure",
+    "30345": "failure",
+    "30346": "infeasible",
+    "30347": "failure",
+    "30348": "failure",
+    "30349": "failure",
+    "30350": "failure",
+    "30351": "failure",
+    "30352": "failure",
+    "30353": "infeasible",
+    "30354": "failure",
+    "30355": "failure",
+    "30356": "failure",
+    "30357": "failure",
+    "30358": "failure",
+    "30359": "infeasible",
+    "30360": "failure",
+    "30361": "failure",
+    "30362": "infeasible",
+    "30363": "successful",
+    "30364": "failure",
+    "30365": "failure",
+    "30366": "infeasible",
+    "30367": "failure",
+    "30368": "failure",
+    "30369": "failure",
+    "30370": "failure",
+    "30371": "failure",
+    "30372": "infeasible",
+    "30373": "failure",
+    "30374": "failure",
+    "30375": "failure",
+    "30376": "infeasible",
+    "30377": "failure",
+    "30378": "failure",
+    "30379": "failure",
+    "30380": "failure",
+    "30381": "failure",
+    "30382": "infeasible",
+    "30383": "failure",
+    "30384": "failure",
+    "30385": "failure",
+    "30386": "successful",
+    "30387": "failure",
+    "30388": "failure",
+    "30389": "failure",
+    "30390": "failure",
+    "30391": "failure",
+    "30392": "failure",
+    "30393": "failure",
+    "30394": "failure",
+    "30395": "failure",
+    "30396": "successful",
+    "30397": "successful",
+    "30398": "failure",
+    "30399": "failure",
+    "30400": "failure",
+    "30401": "failure",
+    "30402": "failure",
+    "30403": "successful",
+    "30404": "failure",
+    "30405": "failure",
+    "30406": "failure",
+    "30407": "successful",
+    "30408": "failure",
+    "30409": "infeasible",
+    "30410": "failure",
+    "30411": "failure",
+    "30412": "successful",
+    "30413": "successful",
+    "30414": "infeasible",
+    "30415": "failure",
+    "30416": "failure",
+    "30417": "failure",
+    "30418": "failure",
+    "30419": "failure",
+    "30420": "successful",
+    "30421": "failure",
+    "30422": "failure",
+    "30423": "infeasible",
+    "30424": "failure",
+    "30425": "failure",
+    "30426": "failure",
+    "30427": "successful",
+    "30428": "failure",
+    "30429": "failure",
+    "30430": "failure",
+    "30431": "failure",
+    "30432": "failure",
+    "30433": "failure",
+    "30434": "failure",
+    "30435": "failure",
+    "30436": "failure",
+    "30437": "failure",
+    "30438": "failure",
+    "30439": "failure",
+    "30440": "successful",
+    "30441": "failure",
+    "30442": "failure",
+    "30443": "failure",
+    "30444": "failure",
+    "30445": "failure",
+    "30446": "failure",
+    "30447": "failure",
+    "30448": "failure",
+    "30449": "failure",
+    "30450": "failure",
+    "30451": "failure",
+    "30452": "failure",
+    "30453": "failure",
+    "30454": "failure",
+    "30455": "failure",
+    "30456": "successful",
+    "30457": "successful",
+    "30458": "failure",
+    "30459": "failure",
+    "30460": "failure",
+    "30461": "failure",
+    "30462": "successful",
+    "30463": "failure",
+    "30464": "failure",
+    "30465": "failure",
+    "30466": "failure",
+    "30467": "failure",
+    "30468": "failure",
+    "30469": "infeasible",
+    "30470": "failure",
+    "30471": "failure",
+    "30472": "failure",
+    "30473": "failure",
+    "30474": "failure",
+    "30475": "failure",
+    "30476": "failure",
+    "30477": "failure",
+    "30478": "failure",
+    "30479": "successful",
+    "30480": "failure",
+    "30481": "failure",
+    "30482": "failure",
+    "30483": "failure",
+    "30484": "infeasible",
+    "30485": "failure",
+    "30486": "failure",
+    "30487": "failure",
+    "30488": "failure",
+    "30489": "failure",
+    "30490": "failure",
+    "30491": "failure",
+    "30492": "failure",
+    "30493": "failure",
+    "30494": "failure",
+    "30495": "failure",
+    "30496": "failure",
+    "30497": "infeasible",
+    "30498": "failure",
+    "30499": "failure",
+    "30500": "failure",
+    "30501": "failure",
+    "30502": "infeasible",
+    "30503": "infeasible",
+    "30504": "failure",
+    "30505": "failure",
+    "30506": "failure",
+    "30507": "infeasible",
+    "30508": "failure",
+    "30509": "failure",
+    "30510": "failure",
+    "30511": "failure",
+    "30512": "failure",
+    "30513": "failure",
+    "30514": "failure",
+    "30515": "failure",
+    "30516": "failure",
+    "30517": "failure",
+    "30518": "failure",
+    "30519": "failure",
+    "30520": "failure",
+    "30521": "failure",
+    "30522": "failure",
+    "30523": "infeasible",
+    "30524": "successful",
+    "30525": "failure",
+    "30526": "successful",
+    "30527": "failure",
+    "30528": "failure",
+    "30529": "successful",
+    "30530": "failure",
+    "30531": "failure",
+    "30532": "infeasible",
+    "30533": "successful",
+    "30534": "failure",
+    "30535": "failure",
+    "30536": "infeasible",
+    "30537": "successful",
+    "30538": "failure",
+    "30539": "failure",
+    "30540": "failure",
+    "30541": "failure",
+    "30542": "failure",
+    "30543": "failure",
+    "30544": "failure",
+    "30545": "successful",
+    "30546": "failure",
+    "30547": "failure",
+    "30548": "successful",
+    "30549": "infeasible",
+    "30550": "failure",
+    "30551": "failure",
+    "30552": "failure",
+    "30553": "failure",
+    "30554": "failure",
+    "30555": "failure",
+    "30556": "failure",
+    "30557": "failure",
+    "30558": "failure",
+    "30559": "failure",
+    "30560": "failure",
+    "30561": "failure",
+    "30562": "failure",
+    "30563": "failure",
+    "30564": "failure",
+    "30565": "failure",
+    "30566": "failure",
+    "30567": "successful",
+    "30568": "failure",
+    "30569": "failure",
+    "30570": "failure",
+    "30571": "failure",
+    "30572": "failure",
+    "30573": "failure",
+    "30574": "failure",
+    "30575": "successful",
+    "30576": "failure",
+    "30577": "failure",
+    "30578": "failure",
+    "30579": "failure"
+  }
+}
\ No newline at end of file
diff --git a/robometer/dataset_upload/dataset_helpers/soar_vlm_labeling_reqs.txt b/robometer/dataset_upload/dataset_helpers/soar_vlm_labeling_reqs.txt
new file mode 100644
index 0000000000000000000000000000000000000000..b284cc25e32976ad1c6358ef4eb7882e2cf03049
--- /dev/null
+++ b/robometer/dataset_upload/dataset_helpers/soar_vlm_labeling_reqs.txt
@@ -0,0 +1,22 @@
+# Requirements for SOAR VLM labeling script
+# Install with: pip install -r vlm_labeling_requirements.txt
+
+# Core ML libraries
+torch>=2.0.0
+transformers>=4.37.0
+accelerate>=0.26.0
+
+# Qwen VL utilities
+qwen-vl-utils
+
+# Image processing
+Pillow>=10.0.0
+torchvision>=0.15.0
+
+# Dataset loading
+tensorflow-datasets>=4.9.0
+tensorflow>=2.13.0
+
+# Utilities
+numpy>=1.24.0
+tqdm>=4.65.0
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/koch_arm_ut_dallas_loader.cpython-310.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/koch_arm_ut_dallas_loader.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..92666f0c9d1b98de7b1b88335edab75ff3374df3
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/koch_arm_ut_dallas_loader.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/mw_collected_loader.cpython-311.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/mw_collected_loader.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..977cbbf8ad557ce37842dfa9c51ad0cbd6c063e8
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/mw_collected_loader.cpython-311.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/mw_task_annotations.cpython-311.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/mw_task_annotations.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6ae76ba5f6a396db86e3f045c4754aad3e4c2ec1
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/mw_task_annotations.cpython-311.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/robofac_loader.cpython-310.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/robofac_loader.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f4bb767d9394d5e9ffbcd832121496e5214aeff0
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/robofac_loader.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/usc_franka_policy_ranking_loader.cpython-310.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/usc_franka_policy_ranking_loader.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..cf12f95e19e05bcdc5aa1469e6b74096a6a89f61
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/usc_franka_policy_ranking_loader.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/usc_xarm_policy_ranking_loader.cpython-310.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/usc_xarm_policy_ranking_loader.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84bdb791db56719ba191263b5e37795d599271e5
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/usc_xarm_policy_ranking_loader.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/__pycache__/utd_so101_loader.cpython-310.pyc b/robometer/dataset_upload/dataset_loaders/__pycache__/utd_so101_loader.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..28f3325f4f7f9c2a05aa858849d9a91b7d7328b4
Binary files /dev/null and b/robometer/dataset_upload/dataset_loaders/__pycache__/utd_so101_loader.cpython-310.pyc differ
diff --git a/robometer/dataset_upload/dataset_loaders/agibotworld_loader.py b/robometer/dataset_upload/dataset_loaders/agibotworld_loader.py
new file mode 100755
index 0000000000000000000000000000000000000000..c189b0170c5eefeb0e0c220c97bab0b0f45773a1
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/agibotworld_loader.py
@@ -0,0 +1,619 @@
+"""
+AgiBotWorld dataset loader for the generic dataset converter for Robometer model training.
+This module contains AgiBotWorld-specific logic for loading and processing data using
+HuggingFace streaming to efficiently handle large datasets.
+"""
+
+import json
+import os
+from functools import partial
+from multiprocessing import Pool, cpu_count
+from pathlib import Path
+from typing import Any
+
+import h5py
+import numpy as np
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    create_trajectory_video_optimized,
+    load_sentence_transformer_model,
+    generate_unique_id,
+)
+from dataset_upload.video_helpers import load_video_frames
+from tqdm import tqdm
+
+import datasets as hfds
+from datasets import Dataset, load_dataset
+
+# Episode/task helpers built earlier
+from dataset_upload.data_scripts.agibot.agibot_helper import get_episode_record
+
+# ------------------------------
+# Small utilities
+# ------------------------------
+
+CAMERA_KEYS = {
+    "head_color",
+    "head_left_fisheye_color",
+    "head_right_fisheye_color",
+    "head_center_fisheye_color",
+}
+
+
+def _stable_shard_for_episode(episode_id: str, shard_modulus: int = 1000) -> str:
+    """Return a stable top-level shard name based on episode id.
+
+    Keeps at most ~shard_modulus episode directories per shard.
+    """
+
+    try:
+        idx = int(episode_id)
+    except Exception:
+        idx = abs(hash(episode_id))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _parse_episode_and_camera(key: str) -> tuple[str, str | None]:
+    """Parse __key__ like '678985/videos/head_color' -> ('678985', 'head_color')."""
+    parts = key.split("/")
+    if len(parts) < 3:
+        return parts[0], None
+    return parts[0], parts[2]
+
+
+def _build_video_paths(
+    output_dir: str,
+    dataset_name: str,
+    episode_id: str,
+    subtask_idx: int,
+    camera: str,
+) -> tuple[str, str]:
+    """Return (full_path, relative_path) using a two-level shard + per-episode layout.
+
+    Layout:
+      <output>/<dataset>/<shard_X>/<episode_id>/clip_<k>@<camera>.mp4
+    This avoids >1k files per directory while keeping resume-friendly structure.
+    """
+    shard_dir = _stable_shard_for_episode(episode_id)
+    episode_dir = os.path.join(output_dir, dataset_name.lower(), shard_dir, f"episode_{episode_id}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip_{subtask_idx}@{camera}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_name.lower(), shard_dir, f"episode_{episode_id}", filename)
+    return full_path, rel_path
+
+
+def _collect_unique_texts_for_batch(records: list[tuple[str, dict]]) -> list[str]:
+    """Collect unique instruction texts from a list of (episode_id, record) pairs."""
+    texts: list[str] = []
+    seen: set = set()
+    for _episode_id, rec in records:
+        # Full trajectory instruction
+        full_text = rec.get("task_name") or rec.get("task_description") or ""
+        if full_text and full_text not in seen:
+            seen.add(full_text)
+            texts.append(full_text)
+
+        # Subtasks
+        actions = rec.get("label_info", {}).get("action_config", [])
+        for a in actions:
+            t = (a or {}).get("action_text", "").strip()
+            if t and t not in seen:
+                seen.add(t)
+                texts.append(t)
+    return texts
+
+
+def _encode_texts(texts: list[str], model) -> dict[str, Any]:
+    """Encode a list of texts to vectors using a preloaded model, with caching."""
+    if not texts:
+        return {}
+    vectors = model.encode(texts)
+    return dict(zip(texts, vectors, strict=False))
+
+
+def _frames_for_subrange(frames: np.ndarray, start: int, end: int) -> np.ndarray:
+    """Return frames[start:end] with guardrails; [start, end) semantics."""
+    start = max(int(start), 0)
+    end = min(int(end), len(frames))
+    if start >= end:
+        return np.empty((0,), dtype=object)
+    return frames[start:end]
+
+
+def _process_single_stream_sample(
+    sample: dict[str, Any],
+    embeddings: dict[str, Any],
+    output_dir: str,
+    dataset_name: str,
+    max_frames: int,
+    fps: int,
+) -> list[dict]:
+    """Process one streaming sample: returns zero or more HF entries.
+
+    This function does not load any language model; it expects embeddings for
+    the relevant instruction texts to be provided.
+    """
+
+    result_entries: list[dict] = []
+
+    # Extract key and keep only camera samples we care about
+    key = sample.get("__key__") or ""
+    episode_id, camera = _parse_episode_and_camera(key)
+    if not camera or camera not in CAMERA_KEYS:
+        return result_entries
+
+    # Load associated episode record for task/subtasks
+    try:
+        _json_path, rec = get_episode_record(episode_id)
+    except Exception:
+        return result_entries
+
+    # Get video bytes (dataset exposes only 'mp4', '__key__', '__url__')
+    video_bytes = sample.get("mp4")
+    if not video_bytes:
+        return result_entries
+
+    # Decode the video to frames once
+    try:
+        frames = load_video_frames(video_bytes)
+    except Exception:
+        return result_entries
+
+    if frames is None or len(frames) == 0:
+        return result_entries
+
+    # Build entries: full + subtasks
+    # Full trajectory
+    full_text = rec.get("task_name") or rec.get("task_description") or ""
+    if full_text:
+        subtask_idx = 0
+        full_out_path, rel_path = _build_video_paths(output_dir, dataset_name, episode_id, subtask_idx, camera)
+        # Create video if missing
+        if not os.path.exists(full_out_path):
+            _ = create_trajectory_video_optimized(frames, full_out_path, max_frames=max_frames, fps=fps)
+
+        lang_vec = embeddings.get(full_text)
+        if lang_vec is None:
+            # As a fallback, keep empty vector to avoid crashing
+            lang_vec = np.zeros((384,), dtype=np.float32)
+
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,  # Not used by create_hf_trajectory now since we already wrote, but pass for API
+            "task": full_text,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+        entry = create_hf_trajectory(
+            traj_dict=traj_dict,
+            video_path=full_out_path,
+            lang_vector=lang_vec,
+            max_frames=max_frames,
+            dataset_name=dataset_name,
+            use_video=True,
+            fps=fps,
+        )
+        if entry:
+            entry["frames"] = rel_path
+            result_entries.append(entry)
+
+    # Subtasks
+    actions = rec.get("label_info", {}).get("action_config", [])
+    for i, a in enumerate(actions, start=1):
+        if not isinstance(a, dict):
+            continue
+        text = (a.get("action_text") or "").strip()
+        if not text:
+            continue
+        start = a.get("start_frame", 0)
+        end = a.get("end_frame", len(frames))
+        sub_frames = _frames_for_subrange(frames, start, end)
+        if sub_frames.size == 0:
+            continue
+
+        out_path, rel_path = _build_video_paths(output_dir, dataset_name, episode_id, i, camera)
+        if not os.path.exists(out_path):
+            _ = create_trajectory_video_optimized(sub_frames, out_path, max_frames=max_frames, fps=fps)
+
+        lang_vec = embeddings.get(text)
+        if lang_vec is None:
+            lang_vec = np.zeros((384,), dtype=np.float32)
+
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": sub_frames,
+            "task": text,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+        entry = create_hf_trajectory(
+            traj_dict=traj_dict,
+            video_path=out_path,
+            lang_vector=lang_vec,
+            max_frames=max_frames,
+            dataset_name=dataset_name,
+            use_video=True,
+            fps=fps,
+        )
+        if entry:
+            entry["frames"] = rel_path
+            result_entries.append(entry)
+
+    return result_entries
+
+
+def convert_agibotworld_streaming_to_hf(
+    dataset_name: str,
+    output_dir: str,
+    dataset_label: str = "agibotworld",
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Stream AgiBotWorld, extract camera videos, and write HF entries.
+
+    Returns a datasets.Dataset built from the collected entries. All videos are
+    saved to disk under output_dir.
+    """
+
+    # Load streaming dataset
+    ds = load_dataset(dataset_name, streaming=True, split="train")
+    # Some shards expose PNG frames instead of MP4. Widen features so casting
+    # does not fail during iteration; we'll simply skip non-MP4 samples.
+    widened = hfds.Features({
+        "__key__": hfds.Value("string"),
+        "__url__": hfds.Value("string"),
+        "mp4": hfds.Value("binary"),
+        "png": hfds.Value("binary"),
+    })
+    try:
+        ds = ds.cast(widened)
+    except Exception:
+        pass
+
+    # Determine workers
+    if num_workers == -1:
+        num_workers = max(1, min(cpu_count(), 8))
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Language model for batch embedding
+    lang_model = load_sentence_transformer_model()
+
+    entries: list[dict] = []
+    processed = 0  # number of streaming samples actually flushed/processed
+    default_batch_size = 64
+    batch_size = default_batch_size if (max_trajectories is None) else min(default_batch_size, max_trajectories)
+    batch_samples: list[dict[str, Any]] = []
+    batch_records: list[tuple[str, dict]] = []
+
+    # Simple live stats
+    seen_samples = 0
+    skipped_camera = 0
+    skipped_no_record = 0
+    skipped_no_mp4 = 0
+
+    def flush_batch():
+        nonlocal entries, processed, batch_samples, batch_records
+        if not batch_samples:
+            return
+
+        # Collect unique texts and encode once
+        unique_texts = _collect_unique_texts_for_batch(batch_records)
+        emb_map = _encode_texts(unique_texts, lang_model)
+
+        if num_workers == 1:
+            for sample in tqdm(batch_samples, desc="Batch (seq)", leave=False):
+                res = _process_single_stream_sample(
+                    sample=sample,
+                    embeddings=emb_map,
+                    output_dir=output_dir,
+                    dataset_name=dataset_label,
+                    max_frames=max_frames,
+                    fps=fps,
+                )
+                # res is a list; extend and update decode_fail if nothing produced due to decode error
+                entries.extend(res)
+        else:
+            with Pool(processes=num_workers) as pool:
+                worker = partial(
+                    _process_single_stream_sample,
+                    embeddings=emb_map,
+                    output_dir=output_dir,
+                    dataset_name=dataset_label,
+                    max_frames=max_frames,
+                    fps=fps,
+                )
+                for res in tqdm(
+                    pool.imap_unordered(worker, batch_samples),
+                    total=len(batch_samples),
+                    desc=f"Batch (workers={num_workers})",
+                    leave=False,
+                ):
+                    entries.extend(res)
+
+        processed += len(batch_samples)
+        batch_samples = []
+        batch_records = []
+
+    print(f"Streaming {dataset_name}; workers={num_workers}, batch_size={batch_size}")
+    stream_pbar = tqdm(desc="Streaming samples", unit="sample", dynamic_ncols=True)
+
+    for sample in ds:
+        if max_trajectories is not None and processed >= max_trajectories:
+            break
+
+        key = sample.get("__key__", "")
+        episode_id, camera = _parse_episode_and_camera(key)
+        seen_samples += 1
+        stream_pbar.update(1)
+        if not camera or camera not in CAMERA_KEYS:
+            skipped_camera += 1
+            continue
+
+        # Ensure episode record exists; gather for embedding planning
+        try:
+            _json_path, rec = get_episode_record(episode_id)
+        except Exception:
+            skipped_no_record += 1
+            continue
+
+        # Require mp4 content; if absent (e.g., png-only shard), skip early
+        if not sample.get("mp4"):
+            skipped_no_mp4 += 1
+            continue
+
+        batch_samples.append(sample)
+        batch_records.append((episode_id, rec))
+
+        if len(batch_samples) >= batch_size:
+            flush_batch()
+
+        # If user asked for a very small number, don't wait for another full batch
+        if max_trajectories is not None and (processed + len(batch_samples)) >= max_trajectories:
+            flush_batch()
+            break
+
+    # Final flush
+    flush_batch()
+    stream_pbar.close()
+
+    # Build HF dataset from entries
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    # datasets can infer features; rely on default
+    print(
+        f"Done. seen={seen_samples}, entries={len(entries)}, "
+        f"skipped_camera={skipped_camera}, skipped_no_record={skipped_no_record}, "
+        f"skipped_no_mp4={skipped_no_mp4}"
+    )
+    return Dataset.from_list(entries)
+
+
+def load_agibotworld_dataset(dataset_name_or_path: str, max_trajectories: int = 100) -> dict[str, list[dict]]:
+    """Load AgiBotWorld dataset using HuggingFace streaming and extract head_color.mp4 files.
+
+    Args:
+        dataset_name_or_path: HuggingFace dataset name (e.g. "agibot-world/AgiBotWorld-Alpha")
+                             or local path to the dataset
+
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+
+    print(f"Loading AgiBotWorld dataset from: {dataset_name_or_path}")
+    print("=" * 100)
+    print("LOADING AGIBOTWORLD DATASET")
+    print("=" * 100)
+
+    task_data = {}
+
+    # Check if it's a local path or HuggingFace dataset name
+    if os.path.exists(dataset_name_or_path):
+        # Local dataset
+        task_data = _load_local_agibotworld(dataset_name_or_path, max_trajectories)
+    else:
+        # HuggingFace dataset - use streaming
+        task_data = _load_streaming_agibotworld(dataset_name_or_path, max_trajectories)
+
+    print(
+        f"Loaded {sum(len(trajectories) for trajectories in task_data.values())} trajectories from {len(task_data)} tasks"
+    )
+    return task_data
+
+
+# NOTE: As the dataset is too large, we did not test this function extensively and it may be out of date.
+def _load_local_agibotworld(base_path: str, max_trajectories: int = 100, max_frames: int = 32) -> dict[str, list[dict]]:
+    """Load AgiBotWorld dataset from local files, starting with task_info JSON files."""
+    base_path = Path(base_path)
+    task_data = {}
+
+    # Define required directories
+    observations_dir = base_path / "observations"
+    task_info_dir = base_path / "task_info"
+    proprio_stats_dir = base_path / "proprio_stats"
+
+    if not observations_dir.exists():
+        raise FileNotFoundError(f"Observations directory not found: {observations_dir}")
+    if not task_info_dir.exists():
+        raise FileNotFoundError(f"Task info directory not found: {task_info_dir}")
+
+    # Start by iterating over task_info JSON files to get proper task names
+    task_info_files = list(task_info_dir.glob("*.json"))
+
+    if not task_info_files:
+        raise FileNotFoundError(f"No task info JSON files found in: {task_info_dir}")
+
+    print(f"Found {len(task_info_files)} task info files")
+
+    total_trajectories = 0
+
+    for task_info_file in tqdm(task_info_files, desc="Processing tasks"):
+        if total_trajectories >= max_trajectories:
+            print(f"Reached max_trajectories limit ({max_trajectories}), stopping...")
+            break
+
+        # Extract task ID from filename (e.g., "task_392.json" -> "392")
+        task_id = task_info_file.stem.replace("task_", "")
+
+        # Load task information from JSON
+        task_info = _load_task_info(task_info_file)
+
+        if not task_info:
+            print(f"Skipping task {task_id} - no valid task info")
+            continue
+
+        # Extract proper task name from first episode (they should all have the same task)
+        if task_info and len(task_info) > 0:
+            first_episode = task_info[0]
+            task_name = first_episode.get("task_name", f"Task {task_id}")
+            first_episode.get("task_description", f"AgiBotWorld Task {task_id}")
+        else:
+            task_name = f"Task {task_id}"
+
+        print(f"Processing task {task_id}: '{task_name}'")
+
+        # Get the corresponding task directory
+        task_dir = observations_dir / task_id
+        if not task_dir.exists():
+            print(f"Task directory not found: {task_dir}, skipping...")
+            continue
+
+        trajectories = []
+
+        # Process episodes based on the information in task_info JSON
+        for episode_info in task_info:
+            if total_trajectories >= max_trajectories:
+                break
+
+            episode_id = str(episode_info.get("episode_id", ""))
+            if not episode_id:
+                continue
+
+            # Check if episode directory exists
+            episode_dir = task_dir / episode_id
+            if not episode_dir.exists():
+                print(f"Episode directory not found: {episode_dir}, skipping episode {episode_id}")
+                continue
+
+            # Look for head_color.mp4 file
+            videos_dir = episode_dir / "videos"
+            head_color_video = videos_dir / "head_color.mp4"
+
+            if head_color_video.exists():
+                # Load proprioceptive data
+                proprio_file = proprio_stats_dir / task_id / episode_id / "proprio_stats.h5"
+                actions = _load_actions_from_h5(proprio_file)
+
+                # Process video: resize to 256x256 and downsample frames
+                try:
+                    processed_frames = load_video_frames(head_color_video)
+
+                    trajectory = {
+                        "frames": processed_frames,  # Processed video frames
+                        "actions": actions,
+                        "is_robot": True,  # AgiBotWorld is robot data
+                        "task": task_name,  # Use the descriptive task name from JSON
+                        "optimal": "optimal",  # Assume all AgiBotWorld trajectories are optimal
+                    }
+                except Exception as e:
+                    print(f"  ❌ Failed to process video {head_color_video}: {e}")
+                    continue
+
+                trajectories.append(trajectory)
+                total_trajectories += 1
+
+                print(f"  ✅ Loaded episode {episode_id} ({total_trajectories}/{max_trajectories})")
+            else:
+                print(f"  ❌ head_color.mp4 not found for episode {episode_id}")
+
+        if trajectories:
+            # Use proper task name from JSON instead of generic "task_{id}"
+            task_data[task_name] = trajectories
+            print(f"Added {len(trajectories)} trajectories for task '{task_name}'")
+
+    print(f"Loaded {total_trajectories} total trajectories from {len(task_data)} tasks")
+    return task_data
+
+
+def _load_streaming_agibotworld(dataset_name: str, max_trajectories: int = 100) -> dict[str, list[dict]]:
+    """Legacy helper no longer used. Kept for compatibility."""
+    raise NotImplementedError("Use convert_agibotworld_streaming_to_hf() for streaming conversion.")
+
+
+def _load_task_info(task_info_file: Path) -> list[dict]:
+    """Load task information from JSON file."""
+    if not task_info_file.exists():
+        print(f"Task info file not found: {task_info_file}")
+        return []
+
+    try:
+        with open(task_info_file) as f:
+            task_info = json.load(f)
+        return task_info if isinstance(task_info, list) else [task_info]
+    except Exception as e:
+        print(f"Error loading task info from {task_info_file}: {e}")
+        return []
+
+
+def _load_actions_from_h5(proprio_file: Path) -> np.ndarray:
+    """Load actions from proprioceptive H5 file."""
+    if not proprio_file.exists():
+        print(f"Proprioceptive file not found: {proprio_file}")
+        return np.array([])
+
+    try:
+        with h5py.File(proprio_file, "r") as f:
+            # According to AgiBotWorld docs, actions are stored under /action
+            if "action" in f:
+                action_group = f["action"]
+
+                # Try to extract joint actions (most common for manipulation)
+                if "joint" in action_group and "position" in action_group["joint"]:
+                    actions = action_group["joint"]["position"][:]
+                elif "end" in action_group and "position" in action_group["end"]:
+                    # Use end-effector positions if joint positions not available
+                    end_positions = action_group["end"]["position"][:]
+                    end_orientations = (
+                        action_group["end"]["orientation"][:] if "orientation" in action_group["end"] else None
+                    )
+
+                    if end_orientations is not None:
+                        # Concatenate position and orientation for full 6-DOF actions
+                        # Reshape orientations from [N, 2, 4] to [N, 8] (both arms)
+                        end_orientations_flat = end_orientations.reshape(end_orientations.shape[0], -1)
+                        # Reshape positions from [N, 2, 3] to [N, 6]
+                        end_positions_flat = end_positions.reshape(end_positions.shape[0], -1)
+                        actions = np.concatenate([end_positions_flat, end_orientations_flat], axis=1)
+                    else:
+                        actions = end_positions.reshape(end_positions.shape[0], -1)
+                else:
+                    print(f"No recognizable action data found in {proprio_file}")
+                    return np.array([])
+
+                return actions
+            else:
+                print(f"No action group found in {proprio_file}")
+                return np.array([])
+
+    except Exception as e:
+        print(f"Error loading actions from {proprio_file}: {e}")
+        return np.array([])
diff --git a/robometer/dataset_upload/dataset_loaders/autoeval_loader.py b/robometer/dataset_upload/dataset_loaders/autoeval_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..44dddb6499e0d64926fffb82a2fc5b3d04e0333d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/autoeval_loader.py
@@ -0,0 +1,85 @@
+import os
+import pickle
+from collections import defaultdict
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from tqdm import tqdm
+from dataset_upload.helpers import generate_unique_id
+
+
+class AutoEvalFrameLoader:
+    """Pickle-able loader that decodes image frames from a pickled episode file on demand.
+
+    Expects each pickle to contain a sequence of steps with observations under
+    f.obs['image_primary'] (RGB arrays or encodable).
+    """
+
+    def __init__(self, pickle_path: str) -> None:
+        self.pickle_path = pickle_path
+
+    def __call__(self) -> np.ndarray:
+        with open(self.pickle_path, "rb") as f:
+            ep = pickle.load(f)
+        frames = np.asarray(ep.obs["image_primary"], dtype=np.uint8)
+        return frames
+
+
+def _make_traj(pickle_path: Path, task: str, success: bool) -> dict[str, Any]:
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    traj["task"] = task
+    traj["frames"] = AutoEvalFrameLoader(str(pickle_path))
+    traj["is_robot"] = True
+    traj["quality_label"] = "successful" if success else "failure"
+    traj["partial_success"] = 1 if success else 0
+    traj["data_source"] = "autoeval"
+    traj["preference_group_id"] = None
+    traj["preference_rank"] = None
+    return traj
+
+
+def load_autoeval_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    """Load AutoEval pickled episodes and return paired trajectories per task.
+
+    Assumes structure like:
+      <dataset_path>/eval_data/<episode_folder>/*.pkl
+
+    We treat each subfolder under eval_data as a group, and within it we look
+    for pairs of success/failure pickles for the same episode id when possible.
+    We also print counts of successes and failures and keep only paired ones.
+    """
+
+    root = Path(os.path.expanduser(dataset_path))
+    eval_root = root / "eval_data"
+    if not eval_root.exists():
+        raise FileNotFoundError(f"AutoEval eval_data folder not found: {eval_root}")
+
+    pkl_files = list(eval_root.rglob("*.pkl"))
+
+    success_count = 0
+    total_count = 0
+
+    task_data: dict[str, list[dict]] = defaultdict(list)
+    task_success_set = set()
+
+    for pkl in tqdm(pkl_files, desc="Loading AutoEval dataset", total=len(pkl_files)):
+        with open(pkl, "rb") as f:
+            ep = pickle.load(f)
+            success = ep.success
+            task = ep.language_command
+        if success is None or task is None:
+            continue
+
+        success_count += int(success)
+        total_count += 1
+        if bool(success):
+            task_success_set.add(task)
+        task_data[task].append(_make_traj(pkl, task, success=bool(success)))
+
+    # make sure all tasks have at least one success
+    for task in task_data.keys():
+        assert task in task_success_set, f"Task {task} is not in task_success_set"
+    print(f"AutoEval: successes={success_count}, total={total_count}")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/egocot_loader.py b/robometer/dataset_upload/dataset_loaders/egocot_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..846d324c6d28bd82ec628df24f9cc66ba28c0e54
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/egocot_loader.py
@@ -0,0 +1,188 @@
+import os
+import json
+from collections import defaultdict
+from tqdm import tqdm
+
+import numpy as np
+from dataset_upload.helpers import generate_unique_id
+
+trajectory_info_template = {
+    "id": [],
+    "task": [],
+    # "lang_vector": [],
+    "data_source": None,
+    "frames": None,
+    "is_robot": None,
+    "quality_label": None,
+    "partial_success": None,  # in [0, 1]
+}
+
+
+class EgoCOTFrameloader:
+    """Pickle-able loader that reads EgoCoT frames from disk on demand.
+
+    Stores only simple fields so it can be safely passed across processes.
+    """
+
+    def __init__(self, frames_path: str) -> None:
+        self.frames_path = frames_path
+
+    def __call__(self) -> np.ndarray:
+        """Load frames from disk when called.
+
+        Returns:
+            np.ndarray of shape (T, H, W, 3), dtype uint8
+        """
+        # Load the numpy array containing 8 consecutive frames
+        frames = np.load(self.frames_path)
+
+        # Ensure the frames are in the correct format
+        # EgoCoT frames are stored as numpy arrays with 8 consecutive frames
+        assert frames.ndim == 4, f"Expected 4D array, got {frames.ndim}D array"
+
+        # Ensure shape and dtype sanity
+        if not isinstance(frames, np.ndarray) or frames.ndim != 4 or frames.shape[1] != 3:
+            raise ValueError(f"Unexpected frames shape for {self.frames_path}: {getattr(frames, 'shape', None)}")
+
+        # Ensure uint8
+        if frames.dtype != np.uint8:
+            # Convert from float to uint8 if necessary
+            if frames.dtype in [np.float32, np.float64]:
+                if frames.max() <= 1.0:
+                    # Values in [0, 1] range - just scale to [0, 255]
+                    frames = (frames * 255).astype(np.uint8)
+                else:
+                    # Values appear to be ImageNet normalized - denormalize first
+                    # ImageNet normalization: (pixel - mean) / std
+                    # Reverse: pixel * std + mean
+                    imagenet_mean = np.array([0.485, 0.456, 0.406]).reshape(1, 3, 1, 1)
+                    imagenet_std = np.array([0.229, 0.224, 0.225]).reshape(1, 3, 1, 1)
+
+                    # Denormalize: multiply by std and add mean
+                    frames = frames * imagenet_std + imagenet_mean
+
+                    # Clamp to [0, 1] and convert to [0, 255]
+                    frames = np.clip(frames, 0, 1)
+                    frames = (frames * 255).astype(np.uint8)
+            else:
+                frames = frames.astype(np.uint8, copy=False)
+
+        # now convert frames to (T, H, W, 3) from (T, C, H, W)
+        frames = frames.transpose(0, 2, 3, 1)
+
+        return frames
+
+
+def create_new_trajectory(frames_path: str, caption: str) -> dict:
+    """Create a new trajectory from EgoCoT data."""
+    trajectory_info = {}
+    trajectory_info["id"] = generate_unique_id()
+    trajectory_info["task"] = caption  # Use caption as the task description
+    trajectory_info["frames"] = EgoCOTFrameloader(frames_path)
+    trajectory_info["is_robot"] = False  # EgoCoT is human egocentric data
+    trajectory_info["quality_label"] = "successful"
+    trajectory_info["partial_success"] = 1
+    trajectory_info["data_source"] = "egocot"
+    return trajectory_info
+
+
+def load_egocot_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    """Load EgoCoT dataset from results.json and .npy frame files (EgoCOT_clear).
+
+    Expected layout (example):
+        <dataset_path>/
+          results.json
+          EgoCOT_clear/
+            EGO_0000.npy
+            EGO_0001.npy
+
+    Args:
+        dataset_path: Path to a directory containing one or more EgoCoT result folders
+
+    Returns:
+        Dictionary mapping task descriptions to lists of trajectory dictionaries
+    """
+    # Locate results.json files
+    json_files = []
+    for root, dirs, files in os.walk(dataset_path):
+        for file in files:
+            if file.lower() == "results.json":
+                json_files.append(os.path.join(root, file))
+
+    if not json_files:
+        raise FileNotFoundError(f"No results.json files found in {dataset_path}")
+
+    task_data = defaultdict(list)
+    total_trajectories = 0
+
+    for json_file in json_files:
+        print(f"Loading annotations from {json_file}")
+
+        with open(json_file, "r") as f:
+            annotations = json.load(f)
+
+        # Normalize JSON structures
+        if isinstance(annotations, list):
+            data_items = annotations
+        elif isinstance(annotations, dict):
+            if "data" in annotations:
+                data_items = annotations["data"]
+            elif "results" in annotations:
+                data_items = annotations["results"]
+            elif "annotations" in annotations:
+                data_items = annotations["annotations"]
+            elif "samples" in annotations:
+                data_items = annotations["samples"]
+            elif "image" in annotations:
+                data_items = [annotations]
+            else:
+                raise ValueError(f"Unexpected JSON structure in {json_file}: cannot find data list")
+        else:
+            raise ValueError(f"Unexpected JSON structure in {json_file}")
+
+        for item in tqdm(data_items, desc="Processing trajectories"):
+            # Extract required fields (handle common variants and typos)
+            image_filename = item.get("image")
+            # this caption is the llm generated one that's much more detailed from EgoCOT's processing
+            caption = item.get("planing").split("\n")[0][1:]
+            if not caption:
+                caption = item.get("caption")  # use the backup original caption
+            score = item.get("score")
+
+            if not image_filename or not caption:
+                print(f"Skipping item with missing image or caption: {item}")
+                continue
+
+            # Construct full path to the .npy file, prioritizing EgoCOT_clear next to results.json
+            base_dir = os.path.dirname(json_file)
+            candidate_paths = [
+                os.path.join(base_dir, "EgoCOT_clear", image_filename) if image_filename else None,
+                # os.path.join(base_dir, image_filename) if image_filename else None,
+                # os.path.join(dataset_path, "EgoCOT_clear", image_filename) if image_filename else None,
+                # os.path.join(dataset_path, image_filename) if image_filename else None,
+            ]
+
+            frames_path = None
+            for cand in candidate_paths:
+                if cand and os.path.exists(cand):
+                    frames_path = cand
+                    break
+
+            if frames_path is None:
+                print(f"Warning: .npy frame file not found for: {image_filename}")
+                continue
+
+            if not frames_path.lower().endswith(".npy"):
+                print(f"Warning: expected .npy file, got: {frames_path}. Skipping.")
+                continue
+
+            # Create trajectory
+            trajectory = create_new_trajectory(frames_path, caption)
+
+            # Group by task/caption for organization
+            task_key = caption[:50] + "..." if len(caption) > 50 else caption
+            task_data[task_key].append(trajectory)
+            total_trajectories += 1
+
+    print(f"Loaded {total_trajectories} trajectories from {len(task_data)} unique tasks")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/egodex_loader.py b/robometer/dataset_upload/dataset_loaders/egodex_loader.py
new file mode 100755
index 0000000000000000000000000000000000000000..a88c562306acbab7191b4c5f6a410802f86fd3a4
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/egodex_loader.py
@@ -0,0 +1,146 @@
+"""
+EgoDex dataset loader for the generic dataset converter for Robometer model training.
+
+This module provides a simple, readable loader inspired by the LIBERO loader:
+- Discovers (HDF5, MP4) pairs per task directory
+- Lazily loads frames via a small frame-loader callable
+- Extracts a task description and pose actions from HDF5
+- Returns a dictionary mapping task names to lists of trajectory dicts
+"""
+
+import os
+from pathlib import Path
+from re import A
+
+import h5py
+import numpy as np
+from dataset_upload.helpers import generate_unique_id
+from dataset_upload.video_helpers import load_video_frames
+from tqdm import tqdm
+
+
+class EgoDexFrameLoader:
+    """Pickle-able frame loader for EgoDex videos."""
+
+    def __init__(self, mp4_path: str):
+        self.mp4_path = mp4_path
+
+    def __call__(self) -> np.ndarray:
+        """Load frames from the MP4 file when called."""
+        return load_video_frames(Path(self.mp4_path), max_frames=1800)  # 30hz * 60s = 1800 frames
+
+
+def _discover_trajectory_files(dataset_path: Path) -> list[tuple[Path, Path, str]]:
+    """Discover all (HDF5, MP4) pairs grouped by task directory."""
+    trajectory_files: list[tuple[Path, Path, str]] = []
+    for task_dir in dataset_path.iterdir():
+        if not task_dir.is_dir():
+            continue
+        task_name = task_dir.name
+        for hdf5_file in task_dir.glob("*.hdf5"):
+            mp4_file = hdf5_file.with_suffix(".mp4")
+            if mp4_file.exists():
+                trajectory_files.append((hdf5_file, mp4_file, task_name))
+            else:
+                print(f"Warning: Missing MP4 file for {hdf5_file}")
+    return trajectory_files
+
+
+def _load_hdf5_data(hdf5_path: Path) -> tuple[np.ndarray, str]:
+    """Load pose data and task description from EgoDex HDF5 file."""
+    with h5py.File(hdf5_path, "r") as f:
+        task_description = ""
+        try:
+            if "llm_description" in f.attrs:
+                if "which_llm_description" in f.attrs:
+                    which_desc = f.attrs["which_llm_description"]
+                    if int(which_desc) == 2 and "llm_description2" in f.attrs:
+                        desc = f.attrs["llm_description2"]
+                    else:
+                        desc = f.attrs["llm_description"]
+                else:
+                    desc = f.attrs["llm_description"]
+                if isinstance(desc, bytes):
+                    task_description = desc.decode("utf-8")
+                else:
+                    task_description = str(desc)
+        except Exception as e:
+            print(f"Warning: Could not load task description from {hdf5_path}: {e}")
+        pose_data = _extract_pose_actions(f)
+    return pose_data, task_description
+
+
+def _extract_pose_actions(hdf5_file) -> np.ndarray:
+    """Extract pose actions (positions) from EgoDex HDF5."""
+    actions: list[np.ndarray] = []
+    pose_keys = [
+        "transforms/leftHand",
+        "transforms/rightHand",
+        "transforms/leftIndexFingerTip",
+        "transforms/rightIndexFingerTip",
+        "transforms/camera",
+    ]
+    for key in pose_keys:
+        if key in hdf5_file:
+            transform_data = hdf5_file[key][:]  # (N, 4, 4)
+            positions = transform_data[:, :3, 3]
+            actions.append(positions)
+    if not actions:
+        if "transforms/camera" in hdf5_file:
+            camera_transforms = hdf5_file["transforms/camera"][:]
+            camera_positions = camera_transforms[:, :3, 3]
+            actions.append(camera_positions)
+        else:
+            print("Warning: No pose data found, creating dummy actions")
+            actions.append(np.zeros((100, 3)))
+    return np.concatenate(actions, axis=1)
+
+
+def load_egodex_dataset(dataset_path: str, max_trajectories: int = 100) -> dict[str, list[dict]]:
+    """Load EgoDex dataset and organize by task, without a separate iterator class."""
+    print(f"Loading EgoDex dataset from: {dataset_path}")
+    print("=" * 100)
+    print("LOADING EGODEX DATASET")
+    print("=" * 100)
+
+    dataset_path = Path(os.path.expanduser(dataset_path))
+    if not dataset_path.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    traj_files = _discover_trajectory_files(dataset_path)
+    print(f"Found {len(traj_files)} trajectory pairs")
+
+    task_data: dict[str, list[dict]] = {}
+    loaded_count = 0
+
+    for hdf5_path, mp4_path, task_name in tqdm(traj_files, desc="Processing trajectories"):
+        if max_trajectories is not None and loaded_count >= max_trajectories and max_trajectories != -1:
+            break
+        pose_data, task_description = _load_hdf5_data(hdf5_path)
+
+        if "description unavailable" in task_description.lower():
+            print(f"Skipping task {hdf5_path} because description is: {task_description}")
+            continue
+        frame_loader = EgoDexFrameLoader(str(mp4_path))
+
+        assert task_description is not None
+
+        trajectory = {
+            "frames": frame_loader,
+            # "actions": pose_data,
+            "is_robot": False,
+            "task": task_description,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+            "task_name": task_name,
+            "id": generate_unique_id(),
+        }
+
+        task_data.setdefault(task_name, []).append(trajectory)
+        loaded_count += 1
+
+    total_trajectories = sum(len(v) for v in task_data.values())
+    print(f"Loaded {total_trajectories} trajectories from {len(task_data)} tasks")
+
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/epic_loader.py b/robometer/dataset_upload/dataset_loaders/epic_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..2b7873b6737d93216212864efefb5d0c7bed5375
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/epic_loader.py
@@ -0,0 +1,268 @@
+# Epic kitchens 100
+import csv
+import os
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+from datasets import Dataset
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    create_output_directory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+
+@dataclass
+class EpicClip:
+    participant_id: str
+    video_id: str
+    narration_id: str
+    start_frame: int
+    stop_frame: int
+    narration: str
+
+
+def _read_epic_csv(csv_path: Path) -> list[EpicClip]:
+    clips: list[EpicClip] = []
+    with open(csv_path, "r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            try:
+                clips.append(
+                    EpicClip(
+                        participant_id=row["participant_id"].strip(),
+                        video_id=row["video_id"].strip(),
+                        start_frame=int(row["start_frame"]),
+                        stop_frame=int(row["stop_frame"]),
+                        narration=row["narration"].strip(),
+                        narration_id=row["narration_id"].strip(),
+                    )
+                )
+            except Exception:
+                continue
+    return clips
+
+
+def _video_path_for_clip(dataset_path: Path, clip: EpicClip) -> Path:
+    # video_id maps to video basename (without .MP4). Participant folder contains videos/ with .MP4
+    # Example: <dataset_path>/P01/videos/<video_id>.MP4
+    return dataset_path / clip.participant_id / "videos" / f"{clip.video_id}.MP4"
+
+
+def _read_video_segment(video_path: Path, start_frame: int, stop_frame: int) -> np.ndarray:
+    cap = cv2.VideoCapture(str(video_path))
+    if not cap.isOpened():
+        raise FileNotFoundError(f"Cannot open video: {video_path}")
+
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    start = max(0, min(start_frame, total - 1))
+    end = max(start + 1, min(stop_frame, total))
+
+    # Seek to start
+    cap.set(cv2.CAP_PROP_POS_FRAMES, start)
+
+    frames: list[np.ndarray] = []
+    idx = start
+    while idx < end:
+        ok, frame_bgr = cap.read()
+        if not ok:
+            break
+        frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+        frames.append(frame_rgb)
+        idx += 1
+
+    cap.release()
+    return np.asarray(frames, dtype=np.uint8)
+
+
+def _process_single_epic_clip(args: tuple[Any, ...]) -> dict | None:
+    (
+        clip,
+        dataset_name,
+        root,
+        output_dir,
+        max_frames,
+        fps,
+        shortest_edge_size,
+        center_crop,
+        lang_vec,
+    ) = args
+
+    video_path = _video_path_for_clip(root, clip)  # derive from output_dir -> dataset root
+
+    if not video_path.exists():
+        return None
+
+    # skip anything > 1000
+    if clip.stop_frame - clip.start_frame > 1000:
+        print("Skipping clip because it's too long, length is", clip.stop_frame - clip.start_frame)
+        return None
+    frames = _read_video_segment(video_path, clip.start_frame, clip.stop_frame)
+    if frames.size == 0:
+        return None
+
+    traj = {
+        "id": generate_unique_id(),
+        "task": clip.narration,
+        "frames": frames,
+        "is_robot": False,
+        "quality_label": "successful",
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    out_dir = os.path.join(output_dir, dataset_name.lower(), clip.participant_id)
+    os.makedirs(out_dir, exist_ok=True)
+    out_video = os.path.join(out_dir, f"{clip.narration_id}.mp4")
+
+    entry = create_hf_trajectory(
+        traj_dict=traj,
+        video_path=out_video,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_name,
+        use_video=True,
+        fps=fps,
+        shortest_edge_size=shortest_edge_size,
+        center_crop=center_crop,
+    )
+    if entry:
+        entry["frames"] = os.path.relpath(out_video, output_dir)
+    return entry
+
+
+def convert_epic_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+    shortest_edge_size: int = 240,
+    center_crop: bool = False,
+) -> Dataset:
+    """Convert EPIC-KITCHENS to HF format by writing videos directly (H2R/OXE-style)."""
+
+    create_output_directory(output_dir)
+    root = Path(dataset_path)
+    csv_path = root / "EPIC_100_train.csv"
+    if not csv_path.exists():
+        raise FileNotFoundError(f"EPIC_100_train.csv not found at {csv_path}")
+
+    clips = _read_epic_csv(csv_path)
+    if not clips:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Determine workers
+    if num_workers == -1:
+        try:
+            from multiprocessing import cpu_count as _cpu_count
+
+            num_workers = min(_cpu_count(), 8)
+        except Exception:
+            num_workers = 1
+    elif num_workers == 0:
+        num_workers = 1
+
+    batch_size = 6
+    entries: list[dict] = []
+    produced = 0
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+    file_batch: list[EpicClip] = []
+    vec_batch: list[np.ndarray] = []
+
+    from tqdm import tqdm
+
+    for idx, clip in tqdm(enumerate(clips), desc="iterating through EPIC-KITCHENS Clips", total=len(clips)):
+        if produced >= max_limit:
+            break
+
+        # Precompute language vector
+        if clip.narration not in lang_cache:
+            lang_cache[clip.narration] = lang_model.encode(clip.narration)
+        lang_vec = lang_cache[clip.narration]
+
+        file_batch.append(clip)
+        vec_batch.append(lang_vec)
+
+        if len(file_batch) >= batch_size or idx + 1 == len(clips):
+            worker_args = [
+                (
+                    clip,
+                    dataset_name,
+                    root,
+                    output_dir,
+                    max_frames,
+                    fps,
+                    shortest_edge_size,
+                    center_crop,
+                    vec,
+                )
+                for clip, vec in zip(file_batch, vec_batch)
+            ]
+
+            if num_workers == 1:
+                for args in worker_args:
+                    entry = _process_single_epic_clip(args)
+                    if entry:
+                        entries.append(entry)
+                        produced += 1
+                        if produced >= max_limit:
+                            break
+            else:
+                from multiprocessing import Pool
+                from tqdm import tqdm
+
+                with Pool(processes=num_workers) as pool:
+                    results = list(
+                        tqdm(
+                            pool.imap_unordered(_process_single_epic_clip, worker_args),
+                            total=len(worker_args),
+                            desc=f"Processing EPIC clips (workers={num_workers})",
+                        )
+                    )
+                for entry in results:
+                    if entry:
+                        entries.append(entry)
+                        produced += 1
+                        if produced >= max_limit:
+                            break
+
+            file_batch = []
+            vec_batch = []
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/failsafe_loader.py b/robometer/dataset_upload/dataset_loaders/failsafe_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..cd8ab48d4ee57d9ca04a417a636d18e3cfe1599b
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/failsafe_loader.py
@@ -0,0 +1,202 @@
+import glob
+import json
+import os
+import random
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+from dataset_upload.helpers import generate_unique_id
+
+TASK_TO_INSTRUCTION = {
+    "FailPickCube-v1": "Pick up the red cube",
+    "FailPushCube-v1": "Push and move a cube to a goal region in front of it",
+    "FailStackCube-v1": "Pick up a red cube and stack it on top of a green cube and let go of the cube without it falling",
+}
+
+
+class FailSafeFrameListLoader:
+    """Pickle-able loader that reads a list of image paths on demand.
+
+    Returns np.ndarray (T, H, W, 3) uint8.
+    """
+
+    def __init__(self, image_paths: list[str]) -> None:
+        self.image_paths = image_paths
+        assert len(image_paths) > 0
+
+    def __call__(self) -> np.ndarray:
+        frames: list[np.ndarray] = []
+        for p in self.image_paths:
+            img_bgr = cv2.imread(p, cv2.IMREAD_COLOR)
+            if img_bgr is None:
+                continue
+            img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+            frames.append(img_rgb)
+        if not frames:
+            return np.empty((0, 0, 0, 3), dtype=np.uint8)
+        frames_np = np.asarray(frames, dtype=np.uint8)
+        return frames_np
+
+
+def _sorted_pngs(dir_path: Path) -> list[str]:
+    files = [str(p) for p in dir_path.glob("*.png")]
+
+    def _key(s: str) -> tuple:
+        name = os.path.splitext(os.path.basename(s))[0]
+        try:
+            return (int(name),)
+        except Exception:
+            return (name,)
+
+    files.sort(key=_key)
+    return files
+
+
+def _make_traj(
+    image_paths: list[str], task: str, instruction: str, is_success: bool, sub_task: str | None = None
+) -> dict[str, Any]:
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    # Combine main instruction with optional sub_task for clarity
+    if sub_task:
+        traj["task"] = sub_task
+    else:
+        traj["task"] = instruction
+    traj["frames"] = FailSafeFrameListLoader(image_paths)
+    traj["is_robot"] = True
+    traj["quality_label"] = "successful" if is_success else "failure"
+    traj["data_source"] = "failsafe"
+    traj["preference_group_id"] = None
+    traj["preference_rank"] = None
+    return traj
+
+
+def _gather_full_episodes(task_dir: Path, view: str, instruction: str) -> list[dict]:
+    episodes: list[dict] = []
+    # Seeds are numbered directories directly under task_dir
+    for seed_dir in sorted([p for p in task_dir.iterdir() if p.is_dir()]):
+        # Ground truth (success)
+        gt_view_dir = seed_dir / "Ground_Truth" / view
+        if gt_view_dir.exists():
+            imgs = _sorted_pngs(gt_view_dir)
+            assert len(imgs) > 0
+            if imgs:
+                episodes.append(_make_traj(imgs, task_dir.name, instruction, is_success=True))
+
+        # Failures: any subfolder except Ground_Truth
+        for attempt_dir in sorted([p for p in seed_dir.iterdir() if p.is_dir() and p.name != "Ground_Truth"]):
+            view_dir = attempt_dir / view
+            if view_dir.exists():
+                assert len(imgs) > 0
+                imgs = _sorted_pngs(view_dir)
+                if imgs:
+                    episodes.append(_make_traj(imgs, task_dir.name, instruction, is_success=False))
+    return episodes
+
+
+def _gather_sub_episodes_from_json(dataset_root: Path, view: str) -> list[dict]:
+    episodes: list[dict] = []
+    # JSON files like vla_data_FailPickCube-v1.json, vla_data_GT_PickCube-v1.json etc.
+    json_dir = dataset_root / "json_files"
+    if not json_dir.exists():
+        json_dir = dataset_root  # fallback if jsons are at root
+
+    json_files = glob.glob(str(json_dir / "vla_data_*.json"))
+    for jf in sorted(json_files):
+        try:
+            with open(jf, "r") as f:
+                data = json.load(f)
+        except Exception:
+            continue
+        if not isinstance(data, list):
+            continue
+        # sub sample 1/3 for 3 views
+        for entry in random.sample(data, len(data) // 3):
+            task_key = entry.get("task")
+            instruction = entry.get("instruction") or TASK_TO_INSTRUCTION.get(task_key, task_key or "")
+            sub_task = entry.get("sub_task")
+            failure_type = entry.get("failure_type", "None")
+            # Image list is relative to dataset root
+            imgs_rel = entry.get("image", [])
+            if not imgs_rel:
+                continue
+            # Filter by desired view: ensure each path contains "/<view>/"
+            if view:
+                imgs_rel = [p for p in imgs_rel if f"/{view}/" in p or f"\\{view}\\" in p]
+                if len(imgs_rel) == 0:
+                    continue
+            image_paths = [str((dataset_root / p).resolve()) for p in imgs_rel]
+            is_success = (failure_type is None) or (str(failure_type).lower() == "none")
+            episodes.append(
+                _make_traj(image_paths, task_key or "failsafe", instruction, is_success=is_success, sub_task=sub_task)
+            )
+    return episodes
+
+
+def load_failsafe_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    """Load FailSafe dataset from local folders and JSON sub-trajectory annotations.
+
+    Args:
+        dataset_path: Root directory containing FailPickCube-v1/ FailPushCube-v1/ FailStackCube-v1/ and jsons
+
+    Returns:
+        Mapping: instruction string -> list of trajectory dicts
+    """
+    views = ["front", "side", "wrist"]
+    include_sub_trajectories = True
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"FailSafe dataset path not found: {root}")
+
+    task_dirs = [
+        p for p in [root / "FailPickCube-v1", root / "FailPushCube-v1", root / "FailStackCube-v1"] if p.exists()
+    ]
+
+    task_data: dict[str, list[dict]] = {}
+
+    # Sub-trajectory episodes from JSON
+    if include_sub_trajectories:
+        for view in views:
+            # sample one view
+            sub_episodes = _gather_sub_episodes_from_json(root, view=view)
+            print(f"Found {len(sub_episodes)} sub-trajectories for {view} after sampling 1/3 of the data")
+            for traj in sub_episodes:
+                task = traj["task"]
+                task_data.setdefault(task, []).append(traj)
+
+    # Full episodes
+    for tdir in task_dirs:
+        instruction = TASK_TO_INSTRUCTION.get(tdir.name, tdir.name)
+        print(f"Gathering full episodes for {instruction}")
+        for view in views:
+            episodes = _gather_full_episodes(tdir, view=view, instruction=instruction)
+            print(f"Found {len(episodes)} episodes for {instruction} {view}")
+            if episodes:
+                task_data.setdefault(instruction, []).extend(episodes)
+
+    # only keep tasks that have both failed and successful trajectories
+    task_data_paired = {}
+    for task, trajectories in task_data.items():
+        failed_trajectories = [t for t in trajectories if t["quality_label"] == "failure"]
+        successful_trajectories = [t for t in trajectories if t["quality_label"] == "successful"]
+        if len(failed_trajectories) == 0 or len(successful_trajectories) == 0:
+            continue
+        task_data_paired[task] = failed_trajectories + successful_trajectories
+
+    print(
+        f"Found {len(task_data_paired)} tasks with both failed and successful trajectories from originally {len(task_data)} tasks"
+    )
+
+    # print how many failed and successful trajectories there are
+    failed_trajectories = [
+        sum([1 for t in traj if t["quality_label"] == "failure"]) for traj in task_data_paired.values()
+    ]
+    successful_trajectories = [
+        sum([1 for t in traj if t["quality_label"] == "successful"]) for traj in task_data_paired.values()
+    ]
+    print(f"Found {sum(failed_trajectories)} failed trajectories")
+    print(f"Found {sum(successful_trajectories)} successful trajectories")
+    return task_data_paired
diff --git a/robometer/dataset_upload/dataset_loaders/fino_net_loader.py b/robometer/dataset_upload/dataset_loaders/fino_net_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..52d6e53f9eb99168fb865fdad6b2467e9f0427e0
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/fino_net_loader.py
@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+"""
+FinoNet dataset loader for the generic dataset converter for Robometer model training.
+https://huggingface.co/datasets/jesbu1/fino-net-dataset
+This module contains FinoNet-specific logic for loading and processing image sequences.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from PIL import Image
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+from tqdm import tqdm
+from datasets import Dataset
+
+
+# Task mapping from task names to instructions
+TASK_TO_INSTRUCTION = {
+    "put_on": "put the single block on the table onto the stack",
+    "put_in": "put the object on the table into the container",
+    "place": "place the left object on the table onto the stack",
+    "pour": "pour the contents of the cup into the receptacle on the table without spilling",
+    "push": "push the object towards the right without knocking it over",
+}
+
+
+def _load_annotation_files(dataset_path: Path) -> dict[str, dict[int, int]]:
+    """Load annotation files for all tasks.
+
+    Returns:
+        Dictionary mapping task name to {episode_number: label} where label is 0 for success, 1 for failure
+    """
+    annotations = {}
+
+    # The annotation files are in the root directory
+    annotation_files = {
+        "put_on": "put_on_annotation.txt",
+        "put_in": "put_in_annotation.txt",
+        "place": "place_annotation.txt",
+        "pour": "pour_annotation.txt",
+        "push": "push_annotation.txt",
+    }
+
+    for task_name, filename in annotation_files.items():
+        annot_file = dataset_path / filename
+        if not annot_file.exists():
+            print(f"Warning: {filename} not found, skipping {task_name}")
+            continue
+
+        task_annotations = {}
+        with open(annot_file, "r") as f:
+            for i, line in enumerate(f):
+                line = line.strip()
+                if not line:
+                    continue
+                # Parse format: name, label
+                parts = line.split(",")
+                if len(parts) >= 2 and i > 0:  # i ==0 is the header line
+                    episode_num = int(parts[0].strip())
+                    label = int(parts[1].strip())
+                    task_annotations[episode_num] = label
+
+        annotations[task_name] = task_annotations
+        print(f"Loaded {len(task_annotations)} annotations for {task_name}")
+
+    return annotations
+
+
+def _load_episode_images(episode_dir: Path) -> list[Path]:
+    """Load all image files from an episode directory, sorted by frame number.
+
+    Args:
+        episode_dir: Path to episode directory containing PNG files
+
+    Returns:
+        List of image file paths sorted by frame number
+    """
+    if not episode_dir.exists():
+        return []
+
+    # Find all PNG files
+    image_files = []
+    for img_file in episode_dir.glob("*.png"):
+        image_files.append(img_file)
+
+    # Sort by frame number (e.g., frame0000000.png, frame0000024.png)
+    def get_frame_num(path: Path) -> int:
+        name = path.stem  # e.g., "frame0000000"
+        try:
+            return int(name.replace("frame", ""))
+        except:
+            return 0
+
+    image_files.sort(key=get_frame_num)
+    return image_files
+
+
+def _load_image_as_numpy(img_path: Path) -> np.ndarray:
+    """Load a PNG image and return as numpy array in RGB format."""
+    with Image.open(img_path) as img:
+        # Convert to RGB if needed
+        if img.mode != "RGB":
+            img = img.convert("RGB")
+        # Return as numpy array
+        return np.array(img)
+
+
+def _discover_episodes(dataset_path: Path) -> list[tuple[str, int, int]]:
+    """Discover all episodes in the FinoNet dataset structure.
+
+    Expected structure (after unzipping failure.zip):
+    dataset_path/
+        failnet_dataset/
+            rgb_imgs/
+                put_on/
+                    9/
+                        frame0000000.png
+                        frame0000024.png
+                        ...
+                put_in/
+                place/
+                pour/
+                push/
+
+    Returns:
+        List of tuples: (task_name, episode_number, label)
+    """
+    episodes = []
+
+    # Load annotations
+    annotations = _load_annotation_files(dataset_path)
+
+    # Find the unzipped dataset directory
+    rgb_imgs_dir = dataset_path / "failnet_dataset" / "rgb_imgs"
+    if not rgb_imgs_dir.exists():
+        print(f"Warning: rgb_imgs directory not found at {rgb_imgs_dir}")
+        return episodes
+
+    # Iterate over task directories
+    for task_dir in rgb_imgs_dir.iterdir():
+        if not task_dir.is_dir():
+            continue
+
+        task_name = task_dir.name
+        if task_name not in annotations:
+            print(f"Skipping task {task_name} (no annotations)")
+            continue
+
+        task_annotations = annotations[task_name]
+
+        # Iterate over episode subdirectories
+        for episode_dir in task_dir.iterdir():
+            if not episode_dir.is_dir():
+                continue
+
+            try:
+                episode_num = int(episode_dir.name)
+            except ValueError:
+                continue
+
+            # Get label from annotations
+            if episode_num not in task_annotations:
+                print(f"Warning: Episode {episode_num} for task {task_name} not in annotations")
+                continue
+
+            label = task_annotations[episode_num]
+            episodes.append((task_name, episode_num, label))
+
+    print(f"Discovered {len(episodes)} episodes across {len(annotations)} tasks")
+    return episodes
+
+
+def _process_single_episode(args):
+    """Worker to process a single episode into a trajectory entry.
+
+    Returns a single entry dict or empty list if failed.
+    """
+    (
+        task_name,
+        episode_num,
+        label,
+        dataset_name,
+        output_dir,
+        max_frames,
+        fps,
+        task_instruction,
+        lang_vec,
+        rgb_imgs_dir,
+    ) = args
+
+    try:
+        # Load images for this episode
+        episode_dir = rgb_imgs_dir / task_name / str(episode_num)
+        image_files = _load_episode_images(episode_dir)
+
+        if not image_files:
+            print(f"Warning: No images found for episode {episode_num} in task {task_name}")
+            return []
+
+        # Load all frames into memory
+        frames = []
+        for img_path in image_files:
+            frame = _load_image_as_numpy(img_path)
+            frames.append(frame)
+
+        frames = np.array(frames)  # Shape: (T, H, W, 3)
+
+        # skip first 10 frames because they typically don't show the arm
+        frames = frames[10:]
+
+        # Determine quality label (0 = success, 1 = failure)
+        quality_label = "failed" if label == 1 else "successful"
+
+        # Create video path
+        episode_video_dir = os.path.join(output_dir, dataset_name.lower(), task_name, f"episode_{episode_num:06d}")
+        os.makedirs(episode_video_dir, exist_ok=True)
+        video_filename = "clip.mp4"
+        full_video_path = os.path.join(episode_video_dir, video_filename)
+        rel_video_path = os.path.join(dataset_name.lower(), task_name, f"episode_{episode_num:06d}", video_filename)
+        # Create trajectory dict
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,
+            "task": task_instruction,
+            "is_robot": True,
+            "quality_label": quality_label,
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+
+        # Create HF trajectory entry
+        entry = create_hf_trajectory(
+            traj_dict=traj_dict,
+            video_path=full_video_path,
+            lang_vector=lang_vec,
+            max_frames=max_frames,
+            dataset_name=dataset_name,
+            use_video=True,
+            fps=fps,
+        )
+
+        if entry:
+            entry["frames"] = rel_video_path
+            return [entry]
+
+        return []
+
+    except Exception as e:
+        print(f"Error processing episode {episode_num} for task {task_name}: {e}")
+        return []
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    """Deterministically bucket an index into a shard directory name."""
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def convert_fino_net_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert the FinoNet dataset to HF format by writing videos directly.
+
+    This follows the streaming approach: iterate episodes, write videos,
+    assemble entries, and return a Dataset at the end.
+    """
+
+    if dataset_name is None:
+        raise ValueError("dataset_name is required")
+
+    base_path = Path(dataset_path)
+    if not base_path.exists():
+        raise FileNotFoundError(f"FinoNet dataset path not found: {base_path}")
+
+    # Discover all episodes
+    episodes = _discover_episodes(base_path)
+    if len(episodes) == 0:
+        # Return empty dataset
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+            "partial_success": [],
+        })
+
+    # Limit trajectories if specified
+    if max_trajectories is not None and max_trajectories > 0:
+        episodes = episodes[:max_trajectories]
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Determine workers
+    if num_workers == -1:
+        try:
+            from multiprocessing import cpu_count as _cpu_count
+
+            num_workers = min(_cpu_count(), 8)
+        except Exception:
+            num_workers = 1
+    elif num_workers == 0:
+        num_workers = 1
+
+    batch_size = 64
+
+    entries: list[dict[str, Any]] = []
+    produced_count = 0
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories <= 0) else int(max_trajectories)
+
+    print(f"Found {len(episodes)} episodes; processing in batches of {batch_size} with {num_workers} workers...")
+
+    # Path to rgb_imgs directory
+    rgb_imgs_dir = base_path / "failnet_dataset" / "rgb_imgs"
+
+    # Process in batches
+    episode_batch: list[tuple[str, int, int]] = []
+    info_batch: list[tuple[str, Any]] = []  # (task_instruction, lang_vec)
+
+    for idx, (task_name, episode_num, label) in enumerate(tqdm(episodes, desc="Queuing FinoNet episodes")):
+        if produced_count >= max_limit:
+            break
+
+        # Get task instruction
+        if task_name not in TASK_TO_INSTRUCTION:
+            print(f"Skipping unknown task: {task_name}")
+            continue
+
+        task_instruction = TASK_TO_INSTRUCTION[task_name]
+
+        # Get or create language embedding
+        if task_instruction not in lang_cache:
+            lang_cache[task_instruction] = lang_model.encode(task_instruction)
+        lang_vec = lang_cache[task_instruction]
+
+        episode_batch.append((task_name, episode_num, label))
+        info_batch.append((task_instruction, lang_vec))
+
+        if len(episode_batch) >= batch_size or idx + 1 == len(episodes):
+            # Build worker args
+            worker_args = list(
+                zip(
+                    [t for (t, _, _) in episode_batch],
+                    [e for (_, e, _) in episode_batch],
+                    [l for (_, _, l) in episode_batch],
+                    [dataset_name] * len(episode_batch),
+                    [output_dir] * len(episode_batch),
+                    [max_frames] * len(episode_batch),
+                    [fps] * len(episode_batch),
+                    [ti for (ti, _) in info_batch],
+                    [lv for (_, lv) in info_batch],
+                    [rgb_imgs_dir] * len(episode_batch),
+                    strict=False,
+                )
+            )
+
+            if num_workers == 1:
+                # Sequential processing
+                for args in worker_args:
+                    entries.extend(_process_single_episode(args))
+                    produced_count += 1
+                    if produced_count >= max_limit:
+                        break
+            else:
+                from multiprocessing import Pool
+
+                with Pool(processes=num_workers) as pool:
+                    results = list(
+                        tqdm(
+                            pool.imap_unordered(_process_single_episode, worker_args),
+                            total=len(worker_args),
+                            desc=f"Processing batch (workers={num_workers})",
+                        )
+                    )
+                for res in results:
+                    entries.extend(res)
+                    produced_count += 1
+                    if produced_count >= max_limit:
+                        break
+
+            # Clear batch
+            episode_batch = []
+            info_batch = []
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+            "partial_success": [],
+        })
+
+    print(f"Successfully created {len(entries)} entries")
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/galaxea_loader.py b/robometer/dataset_upload/dataset_loaders/galaxea_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..b6e22eb930168a6b1c46ea71cda68a94556fa3a4
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/galaxea_loader.py
@@ -0,0 +1,302 @@
+import os
+import gc
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from datasets import Dataset, concatenate_datasets
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+from tqdm import tqdm
+
+# Disable GPUs for TensorFlow in this loader to avoid CUDA context issues in workers
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+
+import tensorflow_datasets as tfds
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _build_galaxea_video_paths(
+    output_dir: str,
+    dataset_label: str,
+    episode_idx: int,
+    view_key: str,
+) -> tuple[str, str]:
+    shard_dir = _stable_shard_for_index(episode_idx)
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip@{view_key}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _parse_low_level_english(instruction: bytes | str) -> str | None:
+    """Galaxea's language_instruction format: "high@low_cn@low_en". Return low_en."""
+    try:
+        instruction = instruction.decode("utf-8")
+        parts = instruction.split("@")
+        if len(parts) >= 3:
+            return parts[2].strip()
+        # If not delimited, return as-is
+        return instruction.strip()
+    except Exception:
+        return None
+
+
+def _process_single_galaxea_episode(args):
+    episode, ep_idx, task, lang_vec, output_dir, dataset_name, max_frames, fps, valid_img_keys = args
+
+    episode_entries = []
+    first_step = next(episode)
+    assert len(valid_img_keys) == 1, (
+        "Galaxea only has one valid image key for now. No support for multiple because of the way we iterate over the episode."
+    )
+    for img_key in valid_img_keys:
+        # Validate key presence
+        if img_key not in first_step["observation"]:
+            continue
+        # Prune trivial black frames
+        if np.all(first_step["observation"][img_key] == 0):
+            continue
+
+        frames = [first_step["observation"][img_key]] + [
+            s["observation"][img_key] for s in episode if img_key in s["observation"]
+        ]
+        if not frames:
+            continue
+        # skip anything > 800 frames for now because memory usage
+        elif len(frames) > 1000:
+            print(f"Skipping episode {ep_idx} because it's too long, length is {len(frames)}")
+            del frames
+            continue
+
+        full_path, rel_path = _build_galaxea_video_paths(
+            output_dir=output_dir,
+            dataset_label=dataset_name,
+            episode_idx=ep_idx,
+            view_key=img_key,
+        )
+
+        # Pass frames as list avoid doubling memory from np.stack
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,  # Pass as list, let create_hf_trajectory handle it
+            "task": task,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+        try:
+            entry = create_hf_trajectory(
+                traj_dict=traj_dict,
+                video_path=full_path,
+                lang_vector=lang_vec,
+                max_frames=max_frames,
+                dataset_name=dataset_name,
+                use_video=True,
+                fps=fps,
+            )
+        except Exception as e:
+            print(f"Warning: Failed to create HF trajectory for ep {ep_idx}: {e}")
+            continue
+        if entry:
+            entry["frames"] = rel_path
+            episode_entries.append(entry)
+        del frames
+
+    return episode_entries
+
+
+def convert_galaxea_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert Galaxea RLDS datasets to HF format by writing videos directly (OXE-style).
+
+    Args:
+        dataset_path: Root path that contains an 'rlds' directory with builders.
+        dataset_name: Name to tag the resulting dataset (e.g., 'galaxea').
+        output_dir: Where to write video files and dataset.
+        max_trajectories: Limit number of produced trajectories (None/-1 for all).
+        max_frames: Max frames per video.
+        fps: Video fps.
+    """
+
+    # Normalize and checks
+    if dataset_name is None:
+        raise ValueError("dataset_name is required")
+
+    root = Path(os.path.expanduser(dataset_path)) / "rlds"
+    if not root.exists():
+        raise FileNotFoundError(f"'rlds' directory not found under: {dataset_path}")
+
+    # Determine workers
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    rlds_name = dataset_name.replace("galaxea_", "")
+
+    # Find builder directory/version: root/rlds_name/<version>
+    ds_root = root / rlds_name
+    versions = os.listdir(str(ds_root)) if ds_root.exists() else []
+    if len(versions) == 0:
+        raise ValueError(f"No versions found for {rlds_name} in {ds_root}")
+
+    builder = None
+    for version in versions:
+        if "incomplete" in version:
+            continue
+        try:
+            builder = tfds.builder_from_directory(f"{ds_root}/{version}")
+            break
+        except Exception:
+            continue
+    if builder is None:
+        raise ValueError(f"No valid builder found for {rlds_name} in {ds_root}")
+
+    # to keep memory usage low, use 1 worker for decoding and interleave files
+    dataset = builder.as_dataset(split="train", shuffle_files=False)
+
+    # Determine valid image observation keys for Galaxea (head and both wrists)
+    valid_img_keys = [
+        "image_camera_head",
+    ]
+
+    # Batch/process episodes
+    batch_size = 1
+    num_workers = min(num_workers, 1)
+    entries: list[dict[str, Any]] = []
+    produced = 0
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+    episode_batch = []
+    info_batch = []
+
+    # split up
+    for ep_idx, episode in enumerate(tqdm(dataset, desc=f"Processing {rlds_name} episodes")):
+        if produced >= max_limit:
+            break
+
+        # Materialize first step for language instruction
+        try:
+            first_step = next(iter(tfds.as_numpy(episode["steps"])))
+        except StopIteration:
+            continue
+
+        # Galaxea stores 'language_instruction' at step-level; parse low-level English
+        task = None
+        if "language_instruction" in first_step:
+            task = _parse_low_level_english(first_step["language_instruction"])  # type: ignore[index]
+        if not task:
+            continue
+
+        # Precompute embedding
+        if task not in lang_cache:
+            lang_cache[task] = lang_model.encode(task)
+        lang_vec = lang_cache[task]
+
+        # Convert episode to numpy (list of steps)
+        try:
+            # episode_np = list(tfds.as_numpy(episode["steps"]))
+            episode_np = iter(tfds.as_numpy(episode["steps"]))
+        except Exception as e:
+            print(f"Warning: Failed to convert episode {ep_idx} to numpy: {e}")
+            continue
+
+        episode_batch.append(episode_np)
+        info_batch.append((ep_idx, task, lang_vec))
+
+        if len(episode_batch) >= batch_size or ep_idx + 1 == len(dataset):
+            if num_workers == 1:
+                for args in zip(
+                    episode_batch,
+                    [i for (i, _, _) in info_batch],
+                    [t for (_, t, _) in info_batch],
+                    [v for (_, _, v) in info_batch],
+                    [output_dir] * len(episode_batch),
+                    [dataset_name] * len(episode_batch),
+                    [max_frames] * len(episode_batch),
+                    [fps] * len(episode_batch),
+                    [valid_img_keys] * len(episode_batch),
+                    strict=False,
+                ):
+                    episode_entries = _process_single_galaxea_episode(args)
+                    entries.extend(episode_entries)
+                    produced += len(episode_entries)
+                    if produced >= max_limit:
+                        break
+            else:
+                raise ValueError("num_workers > 1 not supported for Galaxea due to the way the frame loader works.")
+                # from multiprocessing import Pool
+
+                # worker_args = list(
+                #    zip(
+                #        episode_batch,
+                #        [i for (i, _, _) in info_batch],
+                #        [t for (_, t, _) in info_batch],
+                #        [v for (_, _, v) in info_batch],
+                #        [output_dir] * len(episode_batch),
+                #        [dataset_name] * len(episode_batch),
+                #        [rlds_name] * len(episode_batch),
+                #        [max_frames] * len(episode_batch),
+                #        [fps] * len(episode_batch),
+                #        [valid_img_keys] * len(episode_batch),
+                #        strict=False,
+                #    )
+                # )
+
+                # with Pool(processes=num_workers) as pool:
+                #    results = list(
+                #        tqdm(
+                #            pool.imap_unordered(_process_single_galaxea_episode, worker_args),
+                #            total=len(worker_args),
+                #            desc=f"Processing batch (workers={num_workers})",
+                #        )
+                #    )
+                # for res in results:
+                #    entries.extend(res)
+                #    produced += len(res)
+                #    if produced >= max_limit:
+                #        break
+
+            episode_batch = []
+            info_batch = []
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/h2r_loader.py b/robometer/dataset_upload/dataset_loaders/h2r_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad7b963ba7d44abaccb6175d46303a4c8327c7b3
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/h2r_loader.py
@@ -0,0 +1,380 @@
+#!/usr/bin/env python3
+"""
+H2R dataset loader for the generic dataset converter for Robometer model training.
+https://huggingface.co/datasets/dannyXSC/HumanAndRobot
+Human2Robot: Learning Robot Actions from Paired Human-Robot Videos
+This module contains H2R-specific logic for loading and processing HDF5 files.
+
+Updated to support OXE-style streaming conversion: write videos and build
+HF entries on the fly, and return a ready `datasets.Dataset` to be pushed
+or saved by the caller.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import h5py
+import numpy as np
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+from tqdm import tqdm
+from datasets import Dataset
+
+
+class H2RFrameLoader:
+    """Pickle-able loader that reads H2R frames from an HDF5 dataset on demand.
+
+    Stores only simple fields so it can be safely passed across processes.
+    """
+
+    def __init__(self, hdf5_path: str, convert_to_rgb: bool = True):
+        self.hdf5_path = hdf5_path
+        self.convert_to_rgb = convert_to_rgb
+
+    def __call__(self) -> tuple[np.ndarray, np.ndarray]:
+        """Load frames from HDF5 when called.
+
+        Returns:
+            np.ndarray of shape (T, H, W, 3), dtype uint8
+        """
+        with h5py.File(self.hdf5_path, "r") as f:
+            human_frames = f["/cam_data/human_camera"][:]
+            robot_frames = f["/cam_data/robot_camera"][:]
+
+        if self.convert_to_rgb:
+            human_frames = human_frames[..., [2, 1, 0]]
+            robot_frames = robot_frames[..., [2, 1, 0]]
+
+        # Ensure shape and dtype sanity
+        if not isinstance(human_frames, np.ndarray) or human_frames.ndim != 4 or human_frames.shape[-1] != 3:
+            raise ValueError(f"Unexpected frames shape for {self.hdf5_path}: {getattr(human_frames, 'shape', None)}")
+
+        if not isinstance(robot_frames, np.ndarray) or robot_frames.ndim != 4 or robot_frames.shape[-1] != 3:
+            raise ValueError(f"Unexpected frames shape for {self.hdf5_path}: {getattr(robot_frames, 'shape', None)}")
+
+        # Ensure uint8
+        if human_frames.dtype != np.uint8:
+            human_frames = human_frames.astype(np.uint8, copy=False)
+        if robot_frames.dtype != np.uint8:
+            robot_frames = robot_frames.astype(np.uint8, copy=False)
+
+        return human_frames, robot_frames
+
+
+# Task mapping from folder names to task descriptions
+FOLDER_TO_TASK_NAME = {
+    "grab_both_cubes_v1": "pick up each cube individually and place them onto the plate",
+    "grab_cube2_v1": "pick up the cube and place it onto the plate",
+    "grab_cup_v1": "move the cup from left to right",
+    "grab_pencil1_v1": "pick up the marker and place it on the plate",
+    "grab_pencil2_v1": "pick up the marker and place it on the plate",
+    "grab_pencil_v1": "pick up the marker and place it on the plate",
+    "grab_two_cubes2_v1": "pick up the green cube and place it onto the plate",
+    "grab_to_plate1_and_back_v1": "put the red cube on the darker plate",
+    "grab_to_plate1_v1": "pick up the red cube and place it onto the darker plate",
+    "grab_to_plate2_v1": "pick up the red cube and place it onto the lighter plate",
+    "grab_to_plate2_and_back_v1": "put the red cube on the yellow plate",
+    "grab_to_plate2_and_pull_v1": "put the cube on the plate, then pull the plate from bottom to top",
+    "pull_plate_grab_cube": "pull the plate from bottom to top, then pick up the cube and place it onto the plate",
+    "pull_plate_v1": "pull the plate from bottom to top",
+    "push_box_common_v1": "push the box from left to right",
+    "push_box_random_v1": "push the box from left to right",
+    "push_box_two_v1": "push the tissues from left to right",
+    "push_plate_v1": "push the plate from top to bottom",
+    # "roll": "pick up the brush and write on the table", # skipped because it's weird
+    # "writing": "write aimlessly on the desk", # skipped because writing aimlessly is not helpful for reward modeling
+}
+
+
+def _get_task_name_from_folder(folder_name: str) -> str:
+    """Convert folder name to task name using the mapping."""
+    # First try to find exact match
+    if folder_name in FOLDER_TO_TASK_NAME:
+        return FOLDER_TO_TASK_NAME[folder_name]
+    else:
+        return None
+
+
+def _discover_h2r_files(dataset_path: Path) -> list[tuple[Path, str]]:
+    """Discover all video files in the H2R dataset structure.
+
+    Expected structure:
+    dataset_path/
+        folder_name_1/
+            hdf5_file_1.hdf5
+            hdf5_file_2.hdf5
+            hdf5_file_3.hdf5
+            ...
+        folder_name_2/
+            hdf5_file_1.hdf5
+            hdf5_file_2.hdf5
+            hdf5_file_3.hdf5
+            ...
+        ...
+
+    Returns:
+        List of tuples: (hdf5_file_path, task_name)
+    """
+    trajectory_files: list[tuple[Path, str]] = []
+    for folder in dataset_path.iterdir():
+        if folder.is_dir():
+            for file in folder.glob("*.hdf5"):
+                trajectory_files.append((file, folder.name))
+
+    return trajectory_files
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    """Deterministically bucket an index into a shard directory name.
+
+    Matches the naming style used in the OXE loader for consistent layout.
+    """
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _build_h2r_video_paths(
+    output_dir: str,
+    dataset_label: str,
+    episode_idx: int,
+    role: str,
+) -> tuple[str, str]:
+    shard_dir = _stable_shard_for_index(episode_idx)
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip@{role}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_single_h2r_file(args):
+    """Worker to process a single H2R HDF5 file into up to two entries.
+
+    Returns a list of entries (human and/or robot), each with relative frame paths.
+    """
+    (
+        file_path,
+        folder_name,
+        ep_idx,
+        dataset_name,
+        output_dir,
+        max_frames,
+        fps,
+        task,
+        lang_vec,
+    ) = args
+
+    entries: list[dict[str, Any]] = []
+
+    # Load frames for this file (human and robot)
+    human_frames, robot_frames = H2RFrameLoader(str(file_path))()
+
+    # HUMAN entry
+    full_h_path, rel_h_path = _build_h2r_video_paths(
+        output_dir=output_dir,
+        dataset_label=dataset_name,
+        episode_idx=ep_idx,
+        role="human",
+    )
+    human_traj = {
+        "id": generate_unique_id(),
+        "frames": human_frames,
+        "task": task,
+        "is_robot": False,
+        "quality_label": "successful",
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+    human_entry = create_hf_trajectory(
+        traj_dict=human_traj,
+        video_path=full_h_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_name,
+        use_video=True,
+        fps=fps,
+    )
+    if human_entry:
+        human_entry["frames"] = rel_h_path
+        entries.append(human_entry)
+
+    # ROBOT entry
+    full_r_path, rel_r_path = _build_h2r_video_paths(
+        output_dir=output_dir,
+        dataset_label=dataset_name,
+        episode_idx=ep_idx,
+        role="robot",
+    )
+    robot_traj = {
+        "id": generate_unique_id(),
+        "frames": robot_frames,
+        "task": task,
+        "is_robot": True,
+        "quality_label": "successful",
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+    robot_entry = create_hf_trajectory(
+        traj_dict=robot_traj,
+        video_path=full_r_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_name,
+        use_video=True,
+        fps=fps,
+    )
+    if robot_entry:
+        robot_entry["frames"] = rel_r_path
+        entries.append(robot_entry)
+
+    return entries
+
+
+def convert_h2r_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert the H2R dataset to HF format by writing videos directly.
+
+    This mirrors the OXE loader's streaming approach: iterate files, write videos,
+    assemble entries, and return a Dataset at the end.
+    """
+
+    if dataset_name is None:
+        raise ValueError("dataset_name is required")
+
+    base_path = Path(dataset_path)
+    if not base_path.exists():
+        raise FileNotFoundError(f"H2R dataset path not found: {base_path}")
+
+    discovered = _discover_h2r_files(base_path)
+    if len(discovered) == 0:
+        # Return an empty dataset with expected columns
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            # keep schema compatible with helpers/create_hf_trajectory usage
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    # Language model and cache (avoid recomputing for identical tasks)
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Determine workers and batching (match OXE approach to control memory)
+    if num_workers == -1:
+        try:
+            from multiprocessing import cpu_count as _cpu_count
+
+            num_workers = min(_cpu_count(), 8)
+        except Exception:
+            num_workers = 1
+    elif num_workers == 0:
+        num_workers = 1
+
+    batch_size = 64
+
+    entries: list[dict[str, Any]] = []
+    produced_pairs = 0  # count by file; each file can produce up to 2 entries
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+    print(f"Found {len(discovered)} HDF5 files; processing in batches of {batch_size} with {num_workers} workers...")
+
+    # Process files in batches
+    file_batch: list[tuple[Path, str]] = []
+    info_batch: list[tuple[int, str, Any]] = []  # (ep_idx, task, lang_vec)
+
+    for ep_idx, (file_path, folder_name) in enumerate(tqdm(discovered, desc="Queuing H2R files")):
+        if produced_pairs >= max_limit:
+            break
+
+        task = _get_task_name_from_folder(folder_name)
+        if task is None:
+            print("Skipping file: ", file_path)
+            continue
+        if task not in lang_cache:
+            lang_cache[task] = lang_model.encode(task)
+        lang_vec = lang_cache[task]
+
+        file_batch.append((file_path, folder_name))
+        info_batch.append((ep_idx, task, lang_vec))
+
+        if len(file_batch) >= batch_size or ep_idx + 1 == len(discovered):
+            # Build worker args
+            worker_args = list(
+                zip(
+                    [f for (f, _) in file_batch],
+                    [fn for (_, fn) in file_batch],
+                    [i for (i, _, _) in info_batch],
+                    [dataset_name] * len(file_batch),
+                    [output_dir] * len(file_batch),
+                    [max_frames] * len(file_batch),
+                    [fps] * len(file_batch),
+                    [t for (_, t, _) in info_batch],
+                    [lv for (_, _, lv) in info_batch],
+                    strict=False,
+                )
+            )
+
+            if num_workers == 1:
+                # Sequential processing
+                for args in worker_args:
+                    entries.extend(_process_single_h2r_file(args))
+                    produced_pairs += 1
+                    if produced_pairs >= max_limit:
+                        break
+            else:
+                from multiprocessing import Pool
+
+                with Pool(processes=num_workers) as pool:
+                    results = list(
+                        tqdm(
+                            pool.imap_unordered(_process_single_h2r_file, worker_args),
+                            total=len(worker_args),
+                            desc=f"Processing batch (workers={num_workers})",
+                        )
+                    )
+                for res in results:
+                    entries.extend(res)
+                    produced_pairs += 1
+                    if produced_pairs >= max_limit:
+                        break
+
+            # Clear batch
+            file_batch = []
+            info_batch = []
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/hand_paired_loader.py b/robometer/dataset_upload/dataset_loaders/hand_paired_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..4d3276c9e1beadeb1b014bb2b959d2364c1c880d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/hand_paired_loader.py
@@ -0,0 +1,168 @@
+"""
+Loader for HAND_paired_data dataset containing paired robot and human demonstrations.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+from dataset_upload.helpers import generate_unique_id
+
+
+CAMERA_VIEWS = ["external_imgs", "over_shoulder_imgs"]
+
+
+class HandPairedFrameLoader:
+    """Pickle-able loader that reads a list of JPG image paths on demand (RGB, uint8)."""
+
+    def __init__(self, image_paths: list[str]) -> None:
+        if not image_paths:
+            raise ValueError("image_paths must be non-empty")
+        self.image_paths = image_paths
+
+    def __call__(self) -> np.ndarray:
+        frames: list[np.ndarray] = []
+        for p in self.image_paths:
+            img_bgr = cv2.imread(p, cv2.IMREAD_COLOR)
+            if img_bgr is None:
+                continue
+            img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+            frames.append(img_rgb)
+        if not frames:
+            return np.empty((0, 0, 0, 3), dtype=np.uint8)
+        frames_np = np.asarray(frames, dtype=np.uint8)
+        return frames_np
+
+
+def _sorted_jpgs(dir_path: Path) -> list[str]:
+    """Return sorted list of JPG file paths from a directory."""
+    paths = [p for p in dir_path.glob("*.jpg")]
+
+    def _key(p: Path):
+        # Extract number from filenames like "im_0.jpg", "im_1.jpg", etc.
+        name = p.stem
+        try:
+            # Handle "im_X" format
+            if "_" in name:
+                return int(name.split("_")[-1])
+            return int(name)
+        except Exception:
+            return 0
+
+    paths.sort(key=_key)
+    return [str(p) for p in paths]
+
+
+def _parse_task_name(folder_name: str) -> str:
+    """Convert folder name to human-readable task instruction.
+
+    Examples:
+        blend_carrot -> blend carrot
+        close_microwave_hand -> close microwave
+    """
+    # Remove '_hand' suffix if present
+    task = folder_name.replace("_hand", "")
+    # Replace underscores with spaces
+    task = task.replace("_", " ")
+    return task
+
+
+def _is_human_task(folder_name: str) -> bool:
+    """Check if this is a human demonstration task."""
+    return folder_name.endswith("_hand")
+
+
+def _make_traj(image_paths: list[str], task_text: str, is_robot: bool) -> dict[str, Any]:
+    """Create a trajectory dictionary."""
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    traj["task"] = task_text
+    traj["frames"] = HandPairedFrameLoader(image_paths)
+    traj["is_robot"] = is_robot
+    traj["quality_label"] = "successful"  # Assuming all demonstrations are successful
+    traj["data_source"] = "hand_paired"
+    traj["preference_group_id"] = None
+    traj["preference_rank"] = None
+    return traj
+
+
+def load_hand_paired_dataset(dataset_path: str, dataset_name: str) -> dict[str, list[dict]]:
+    """Load HAND_paired_data dataset from local folders.
+
+    Args:
+        dataset_path: Root directory containing task folders (e.g., blend_carrot, blend_carrot_hand, etc.)
+
+    Structure:
+        dataset_path/
+            blend_carrot/
+                traj0/
+                    external_imgs/
+                        im_0.jpg, im_1.jpg, ...
+                    over_shoulder_imgs/
+                        im_0.jpg, im_1.jpg, ...
+                traj1/
+                    ...
+            blend_carrot_hand/
+                traj0/
+                    ...
+            close_microwave/
+                ...
+            close_microwave_hand/
+                ...
+
+    Returns:
+        Mapping: task instruction -> list of trajectory dicts
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"HAND_paired dataset path not found: {root}")
+
+    # Get all task directories
+    task_dirs = [p for p in root.iterdir() if p.is_dir()]
+
+    task_data: dict[str, list[dict]] = {}
+
+    dataset_name = dataset_name.replace("hand_paired_", "")
+
+    for task_dir in task_dirs:
+        print(f"Processing task: {task_dir.name}")
+        task_name = _parse_task_name(task_dir.name)
+        is_robot = not _is_human_task(task_dir.name)
+        if dataset_name == "robot":
+            if not is_robot:
+                continue
+        elif dataset_name == "human":
+            if is_robot:
+                continue
+
+        # Get all trajectory directories (traj0, traj1, traj2, etc.)
+        traj_dirs = [p for p in task_dir.iterdir() if p.is_dir() and p.name.startswith("traj")]
+        print(f"Found {len(traj_dirs)} trajectory directories")
+        for traj_dir in traj_dirs:
+            print(f"Processing trajectory: {traj_dir.name}")
+            # Process each camera view
+            for camera_view in CAMERA_VIEWS:
+                print(f"Processing camera view: {camera_view}")
+                camera_dir = traj_dir / camera_view
+                if not camera_dir.exists():
+                    continue
+
+                # Get sorted list of JPG images
+                image_paths = _sorted_jpgs(camera_dir)
+                if not image_paths:
+                    continue
+
+                # Create trajectory
+                traj = _make_traj(image_paths, task_name, is_robot)
+                task_data.setdefault(task_name, []).append(traj)
+
+    print(f"Loaded {len(task_data)} unique tasks from HAND paired {dataset_name} dataset")
+    for task, trajs in task_data.items():
+        robot_count = sum(1 for t in trajs if t["is_robot"])
+        human_count = sum(1 for t in trajs if not t["is_robot"])
+        print(f"  {task}: {robot_count} robot, {human_count} human trajectories")
+
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/humanoid_everyday_loader.py b/robometer/dataset_upload/dataset_loaders/humanoid_everyday_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..fde874e6bfd80081547e0415bc61e35a3ef7ce4e
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/humanoid_everyday_loader.py
@@ -0,0 +1,346 @@
+import os
+import json
+import shutil
+import gc
+import glob
+from re import T
+import zipfile
+import tempfile
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import pandas as pd
+from datasets import Dataset
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+from tqdm import tqdm
+
+# Disable GPUs for TensorFlow in this loader to avoid CUDA context issues in workers
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+
+to_skip = set(["pull_out_tissue_from_tissue_box_h1.zip"])  # skip because it's incorrect videos
+
+# Google Sheet with task descriptions
+GOOGLE_SHEET_URL = "https://docs.google.com/spreadsheets/d/158Wzf8Xywky3aHJSCfp3OZxf4bkhzAJdcG94eHf8gVc/export?format=csv&gid=1307250382"
+
+
+def _load_google_sheet_tasks() -> dict[str, str]:
+    """Load task descriptions from Google Sheet.
+
+    Returns:
+        Dictionary mapping task names (from zip filenames) to task descriptions.
+    """
+    try:
+        # Read the Google Sheet as CSV, skipping the first 2 rows and using row 3 as header
+        df = pd.read_csv(GOOGLE_SHEET_URL, header=2)
+
+        # Create a mapping from task name to description
+        task_map = {}
+        for _, row in df.iterrows():
+            # Check if we have valid task name and description
+            if pd.notna(row.get("Task Name")) and pd.notna(row.get("Task Description")):
+                task_name = row["Task Name"]
+                description = row["Task Description"]
+
+                # Create mapping for both with and without .zip extension
+                task_map[f"{task_name}.zip"] = description
+                task_map[task_name] = description
+
+        print(f"Loaded {len(task_map) // 2} task descriptions from Google Sheet")
+        return task_map
+    except Exception as e:
+        print(f"Warning: Failed to load Google Sheet: {e}")
+        import traceback
+
+        traceback.print_exc()
+        return {}
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    """Generate stable shard directory name for trajectory indexing."""
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _build_humanoid_video_paths(
+    output_dir: str,
+    dataset_label: str,
+    episode_idx: int,
+    zip_file: str,
+) -> tuple[str, str]:
+    """Build video paths for humanoid everyday dataset."""
+    shard_dir = _stable_shard_for_index(episode_idx)
+    task_prefix = zip_file.split("/")[-2]
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), task_prefix, shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    full_path = os.path.join(episode_dir, f"clip.mp4")
+    rel_path = os.path.join(dataset_label.lower(), task_prefix, shard_dir, f"episode_{episode_idx:06d}", f"clip.mp4")
+    return full_path, rel_path
+
+
+def _process_single_humanoid_episode(args):
+    """Process a single episode from humanoid everyday dataset."""
+    episode_data, ep_idx, task, lang_vec, output_dir, dataset_name, max_frames, fps, zip_file = args
+
+    episode_entries = []
+
+    try:
+        # Extract frames from episode data
+        frames = []
+        for step_data in episode_data:
+            if "image" in step_data:
+                # Convert numpy array to uint8 if needed
+                img = step_data["image"]
+                if img.dtype != np.uint8:
+                    img = img.astype(np.uint8)
+                frames.append(img)
+
+        if not frames:
+            return episode_entries
+
+        full_path, rel_path = _build_humanoid_video_paths(
+            output_dir=output_dir,
+            dataset_label=dataset_name,
+            episode_idx=ep_idx,
+            zip_file=zip_file,
+        )
+
+        # Create trajectory dictionary
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,
+            "task": task,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+
+        try:
+            entry = create_hf_trajectory(
+                traj_dict=traj_dict,
+                video_path=full_path,
+                lang_vector=lang_vec,
+                max_frames=max_frames,
+                dataset_name=dataset_name,
+                use_video=True,
+                fps=fps,
+            )
+        except Exception as e:
+            print(f"Warning: Failed to create HF trajectory for ep {ep_idx}: {e}")
+            return episode_entries
+
+        if entry:
+            entry["frames"] = rel_path
+            episode_entries.append(entry)
+
+    except Exception as e:
+        print(f"Warning: Failed to process episode {ep_idx}: {e}")
+        return episode_entries
+
+    return episode_entries
+
+
+def _create_humanoid_dataloader(zip_path: str):
+    """Create a humanoid everyday dataloader for a zip file."""
+    try:
+        # Import humanoid_everyday dataloader
+        from humanoid_everyday import Dataloader
+
+        # Load dataset from zip file
+        ds = Dataloader(zip_path)
+        return ds
+
+    except ImportError:
+        print(f"Warning: humanoid_everyday package not found. Please install it with: pip install humanoid_everyday")
+        return None
+    except Exception as e:
+        print(f"Warning: Failed to create dataloader from {zip_path}: {e}")
+        return None
+
+
+def _load_single_humanoid_episode(ds, episode_idx: int):
+    """Load a single episode from an existing humanoid everyday dataloader."""
+    try:
+        # Get the specific episode
+        episode = ds[episode_idx]
+
+        # Convert episode to list of step dictionaries
+        episode_data = []
+        for step in episode:
+            episode_data.append(step)
+
+        return episode_data
+
+    except Exception as e:
+        print(f"Warning: Failed to load episode {episode_idx}: {e}")
+        return None
+
+
+def convert_humanoid_everyday_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert Humanoid Everyday datasets to HF format by writing videos directly.
+
+    Args:
+        dataset_path: Root path that contains zip files with humanoid everyday datasets.
+        dataset_name: Name to tag the resulting dataset (e.g., 'humanoid_everyday').
+        output_dir: Where to write video files and dataset.
+        max_trajectories: Limit number of produced trajectories (None/-1 for all).
+        max_frames: Max frames per video.
+        fps: Video fps.
+        num_workers: Number of workers for parallel processing.
+    """
+
+    # Normalize and checks
+    if dataset_name is None:
+        raise ValueError("dataset_name is required")
+
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    # Find all zip files in the dataset path
+    zip_files = glob.glob(os.path.join(root, "**/*.zip"), recursive=True)
+    if not zip_files:
+        raise FileNotFoundError(f"No .zip files found in {dataset_path}")
+
+    print(f"Found {len(zip_files)} zip files to process")
+
+    # Determine workers
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Process all zip files
+    all_entries: list[dict[str, Any]] = []
+    produced = 0
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+    print("Loading task descriptions from Google Sheet...")
+    google_sheet_tasks = _load_google_sheet_tasks()
+
+    for zip_file in tqdm(zip_files, desc="Processing zip files"):
+        print(f"Processing zip file: {zip_file}")
+
+        if zip_file.split("/")[-1] in to_skip:
+            print(f"Skipping zip file: {zip_file}")
+            continue
+
+        # Create dataloader once for this zip file
+        ds = _create_humanoid_dataloader(zip_file)
+        if ds is None:
+            print(f"Failed to create dataloader for {zip_file}")
+            continue
+
+        # Try to find task in Google Sheet
+        zip_filename = zip_file.split("/")[-1]
+        if zip_filename in google_sheet_tasks:
+            task_name = google_sheet_tasks[zip_filename]
+            print(f"Found task description from Google Sheet: {task_name}")
+        else:
+            try:
+                # Get the metadata.json for getting task description
+                # find metadata.json in the unzipped directory using correct glob pattern
+                metadata_paths = glob.glob(
+                    os.path.join(zip_file.replace(".zip", ""), "**", "metadata.json"), recursive=True
+                )
+                if not metadata_paths:
+                    print(f"metadata.json not found in extracted directory for {zip_file}")
+
+                else:
+                    metadata_path = metadata_paths[0]
+                    with open(metadata_path, "r") as f:
+                        metadata = json.load(f)
+                    task_name = metadata["description"]
+                    print(f"Found task description from metadata.json: {task_name}")
+            except Exception as e:
+                print(f"Warning: Failed to load metadata.json for {zip_file}: {e}")
+            print(f"Warning: No task description found for {zip_filename} in Google Sheet, skipping")
+            shutil.rmtree(zip_file.replace(".zip", ""))
+            continue
+
+        # Precompute embedding for this task
+        if task_name not in lang_cache:
+            lang_cache[task_name] = lang_model.encode(task_name)
+        lang_vec = lang_cache[task_name]
+
+        episode_count = len(ds)
+        if episode_count == 0:
+            print(f"No episodes found in {zip_file}")
+            shutil.rmtree(zip_file.replace(".zip", ""))
+            continue
+
+        print(f"Found {episode_count} episodes in {zip_file}")
+
+        # Process episodes one at a time to save memory
+        for ep_idx in tqdm(range(episode_count), desc=f"Processing episodes in {zip_file}"):
+            # Load single episode using the existing dataloader
+            episode_data = _load_single_humanoid_episode(ds, ep_idx)
+            if episode_data is None:
+                print(f"Failed to load episode {ep_idx} from {zip_file}")
+                continue
+
+            # Process single episode
+            episode_entries = _process_single_humanoid_episode((
+                episode_data,
+                ep_idx,
+                task_name,
+                lang_vec,
+                output_dir,
+                dataset_name,
+                max_frames,
+                fps,
+                zip_file,
+            ))
+
+            all_entries.extend(episode_entries)
+            produced += len(episode_entries)
+
+            # Clean up episode data to free memory
+            del episode_data
+
+            if produced >= max_limit:
+                break
+
+        # remove the unzipped file after done since humanoid loader unzips it
+        shutil.rmtree(zip_file.replace(".zip", ""))
+
+        if produced >= max_limit:
+            break
+
+    if not all_entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(all_entries)
diff --git a/robometer/dataset_upload/dataset_loaders/libero_loader.py b/robometer/dataset_upload/dataset_loaders/libero_loader.py
new file mode 100755
index 0000000000000000000000000000000000000000..d3474c83023eef8197621fb557f208120d926cfd
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/libero_loader.py
@@ -0,0 +1,158 @@
+#!/usr/bin/env python3
+"""
+LIBERO dataset loader for the generic dataset converter for Robometer model training.
+This module contains LIBERO-specific logic for loading and processing HDF5 files.
+"""
+
+import os
+from pathlib import Path
+
+import h5py
+import numpy as np
+from dataset_upload.helpers import generate_unique_id
+from tqdm import tqdm
+
+
+class LiberoFrameLoader:
+    """Pickle-able loader that reads LIBERO frames from an HDF5 dataset on demand.
+
+    Stores only simple fields so it can be safely passed across processes.
+    """
+
+    def __init__(self, hdf5_path: str, dataset_path: str, rotate_180: bool = True):
+        self.hdf5_path = hdf5_path
+        self.dataset_path = dataset_path  # e.g., "data/<trajectory_key>/obs/agentview_rgb"
+        self.rotate_180 = rotate_180
+
+    def __call__(self) -> np.ndarray:
+        """Load frames from HDF5 when called.
+
+        Returns:
+            np.ndarray of shape (T, H, W, 3), dtype uint8
+        """
+        with h5py.File(self.hdf5_path, "r") as f:
+            if self.dataset_path not in f:
+                raise KeyError(f"Dataset path '{self.dataset_path}' not found in {self.hdf5_path}")
+
+            frames = f[self.dataset_path][:]
+
+        # Ensure shape and dtype sanity
+        if not isinstance(frames, np.ndarray) or frames.ndim != 4 or frames.shape[-1] != 3:
+            raise ValueError(
+                f"Unexpected frames shape for {self.dataset_path} in {self.hdf5_path}: {getattr(frames, 'shape', None)}"
+            )
+
+        # Match existing behavior: flip vertically (previous code called this 180-degree rotate)
+        if self.rotate_180:
+            frames = frames[:, ::-1, :, :].copy()
+
+        # Ensure uint8
+        if frames.dtype != np.uint8:
+            frames = frames.astype(np.uint8, copy=False)
+
+        return frames
+
+
+def load_libero_dataset(base_path: str) -> dict[str, list[dict]]:
+    """Load LIBERO dataset from HDF5 files and organize by task.
+
+    Args:
+        base_path: Path to the LIBERO dataset directory containing HDF5 files
+
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+
+    print(f"Loading LIBERO dataset from: {base_path}")
+
+    task_data = {}
+
+    # Find all HDF5 files in the base path
+    base_path = Path(base_path)
+    if not base_path.exists():
+        raise FileNotFoundError(f"LIBERO dataset path not found: {base_path}")
+
+    hdf5_files = list(base_path.glob("*.hdf5"))
+    print("=" * 100)
+    print("LOADING LIBERO DATASET")
+    print("=" * 100)
+
+    print(f"Found {len(hdf5_files)} HDF5 files")
+
+    for file_path in tqdm(hdf5_files, desc=f"Processing LIBERO dataset, {len(hdf5_files)} files"):
+        task_name = file_path.stem  # Remove .hdf5 extension
+        # print(f"Loading task: {task_name}")
+
+        with h5py.File(file_path, "r") as f:
+            if "data" not in f:
+                print(f"No 'data' group in {task_name}")
+                continue
+
+            data_group = f["data"]
+            trajectories = []
+
+            for trajectory_key in data_group.keys():
+                trajectory = data_group[trajectory_key]
+                if isinstance(trajectory, h5py.Group):
+                    # Extract trajectory data
+                    trajectory_info = {"frames": [], "actions": []}
+
+                    # Set up lazy frame loader to avoid loading frames into memory up front
+                    if "obs" in trajectory and "agentview_rgb" in trajectory["obs"]:
+                        dataset_path = f"data/{trajectory_key}/obs/agentview_rgb"
+                        trajectory_info["frames"] = LiberoFrameLoader(
+                            hdf5_path=str(file_path),
+                            dataset_path=dataset_path,
+                            rotate_180=True,
+                        )
+
+                    # Get actions if available
+                    if "actions" in trajectory:
+                        trajectory_info["actions"] = trajectory["actions"][:]
+
+                    # Core attributes
+                    trajectory_info["is_robot"] = True
+                    trajectory_info["quality_label"] = "successful"
+                    trajectory_info["preference_group_id"] = None
+                    trajectory_info["preference_rank"] = None
+
+                    # Parse the original file path to extract scene and task info
+                    file_name = os.path.basename(file_path).replace(".hdf5", "")
+
+                    # Extract scene and task from the file name
+                    # Example: LIVING_ROOM_SCENE4_stack_the_right_bowl_on_the_left_bowl_and_place_them_in_the_tray
+                    parts = file_name.split("_")
+
+                    # Find the scene part (contains "SCENE")
+                    scene_part = None
+                    task_parts = []
+
+                    for i, part in enumerate(parts):
+                        if "SCENE" in part:
+                            scene_part = part
+                            # Everything after the scene is the task
+                            task_parts = parts[i + 1 :]
+                            break
+
+                    # If no scene found, then don't use a scene
+                    if scene_part is None:
+                        scene_part = "UNKNOWN_SCENE"
+                        task_parts = parts
+
+                    # Convert task parts to readable string
+                    task_string = " ".join(task_parts).replace("_", " ")
+                    task_string = task_string.replace("demo", "")
+
+                    # Add parsed information to trajectory
+                    trajectory_info["task"] = task_string.strip()
+                    # Assign unique UUID id
+                    trajectory_info["id"] = generate_unique_id()
+                    trajectories.append(trajectory_info)
+
+            task_data[task_name] = trajectories
+            # print(f"  Loaded {len(trajectories)} trajectories for {task_name}")
+
+    print(
+        f"Loaded {sum(len(trajectories) for trajectories in task_data.values())} trajectories from {len(task_data)} tasks"
+    )
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/mit_franka_prank_loader.py b/robometer/dataset_upload/dataset_loaders/mit_franka_prank_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..4fa1a33a14cc28f19bfd2516f68549aa06c47a96
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/mit_franka_prank_loader.py
@@ -0,0 +1,216 @@
+import json
+import os
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+from datasets import Dataset
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+
+def _load_video_frames(video_path: str) -> list[np.ndarray]:
+    """Load frames from an MP4 video file."""
+    if not os.path.exists(video_path):
+        raise FileNotFoundError(f"Video file not found: {video_path}")
+
+    frames = []
+    cap = cv2.VideoCapture(video_path)
+
+    try:
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+    finally:
+        cap.release()
+
+    return frames
+
+
+def _build_video_paths(output_dir: str, dataset_label: str, episode_idx: int, task_name: str) -> tuple[str, str]:
+    """Build output video paths with shard structure."""
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"{task_name}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_episode(args):
+    """Process a single episode."""
+    episode_meta, video_dir, lang_vec, output_dir, dataset_label, max_frames, fps = args
+
+    # Load video frames
+    video_path = os.path.join(video_dir, episode_meta["filename"])
+    try:
+        frames = _load_video_frames(video_path)
+    except Exception as e:
+        print(f"Error loading video {video_path}: {e}")
+        return None
+
+    if not frames:
+        return None
+
+    # Normalize quality label
+    quality_label = episode_meta["success"]
+    if quality_label == "success":
+        quality_label = "successful"
+    elif quality_label == "fail":
+        quality_label = "failure"
+    # "suboptimal" stays as is
+
+    # Build output video path
+    full_path, rel_path = _build_video_paths(
+        output_dir, dataset_label, episode_meta["episode_idx"], episode_meta["task_name"]
+    )
+
+    traj_dict = {
+        "id": generate_unique_id(),
+        "frames": frames,
+        "task": episode_meta["instruction"],
+        "is_robot": True,
+        "quality_label": quality_label,
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    entry = create_hf_trajectory(
+        traj_dict=traj_dict,
+        video_path=full_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_label,
+        use_video=True,
+        fps=fps,
+    )
+
+    if entry:
+        entry["frames"] = rel_path
+        return entry
+    return None
+
+
+def convert_mit_franka_prank_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert MIT-Franka-Prank dataset to HF format.
+
+    Args:
+        dataset_path: Path to the dataset directory containing metadata files and videos
+        dataset_name: Name for the dataset
+        output_dir: Output directory for processed videos
+        max_trajectories: Maximum number of trajectories to process (None for all)
+        max_frames: Maximum frames per trajectory
+        fps: Frames per second for output videos
+        num_workers: Number of worker processes (-1 for auto, 0 for sequential)
+
+    Returns:
+        HuggingFace Dataset
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset directory not found: {dataset_path}")
+
+    # Find the subdirectory with videos (20251210)
+    video_dir = None
+    for subdir in root.iterdir():
+        if subdir.is_dir() and (subdir / "foldtowel_metadata.json").exists():
+            video_dir = subdir
+            break
+
+    if video_dir is None:
+        raise FileNotFoundError(f"Could not find video directory with metadata in: {dataset_path}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Load sentence transformer model
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Load all metadata files
+    all_episodes = []
+    metadata_files = list(video_dir.glob("*_metadata.json"))
+
+    for metadata_file in metadata_files:
+        with open(metadata_file, "r") as f:
+            episodes = json.load(f)
+            all_episodes.extend(episodes)
+
+    print(f"Found {len(all_episodes)} episodes across {len(metadata_files)} metadata files")
+
+    # Limit trajectories if specified
+    if max_trajectories is not None and max_trajectories > 0:
+        all_episodes = all_episodes[:max_trajectories]
+
+    # Pre-compute language embeddings
+    print("Computing language embeddings...")
+    for episode in tqdm(all_episodes, desc="Computing embeddings"):
+        instruction = episode["instruction"]
+        if instruction not in lang_cache:
+            lang_cache[instruction] = lang_model.encode(instruction)
+
+    # Process episodes
+    entries = []
+    if num_workers == 1:
+        # Sequential processing
+        for episode in tqdm(all_episodes, desc="Processing episodes"):
+            lang_vec = lang_cache[episode["instruction"]]
+            result = _process_episode((episode, str(video_dir), lang_vec, output_dir, dataset_name, max_frames, fps))
+            if result:
+                entries.append(result)
+    else:
+        # Parallel processing
+        from multiprocessing import Pool
+
+        args_list = [
+            (episode, str(video_dir), lang_cache[episode["instruction"]], output_dir, dataset_name, max_frames, fps)
+            for episode in all_episodes
+        ]
+
+        with Pool(processes=num_workers) as pool:
+            results = list(
+                tqdm(pool.imap_unordered(_process_episode, args_list), total=len(args_list), desc="Processing episodes")
+            )
+
+        entries = [r for r in results if r is not None]
+
+    print(f"Successfully processed {len(entries)} episodes")
+
+    # Create HuggingFace dataset
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/molmoact_loader.py b/robometer/dataset_upload/dataset_loaders/molmoact_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c1e0e80a003e3bc08be75e65f80e7a1c638fc44
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/molmoact_loader.py
@@ -0,0 +1,244 @@
+import io
+import os
+from pathlib import Path
+from typing import Any, Optional
+
+import numpy as np
+from datasets import Dataset, load_dataset
+from PIL import Image
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _build_molmo_video_paths(
+    output_dir: str,
+    dataset_label: str,
+    episode_idx: int,
+    view_key: str,
+) -> tuple[str, str]:
+    shard_dir = _stable_shard_for_index(episode_idx)
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip@{view_key}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _to_rgb_numpy(img_cell: Any) -> Optional[np.ndarray]:
+    """Convert a datasets Image cell (dict with bytes/path, PIL.Image, or np.ndarray) to RGB uint8 ndarray."""
+    if img_cell is None:
+        return None
+    # Already numpy HxWxC
+    if isinstance(img_cell, np.ndarray):
+        if img_cell.ndim == 3 and img_cell.shape[-1] in (1, 3, 4):
+            if img_cell.shape[-1] == 1:
+                img_cell = np.repeat(img_cell, 3, axis=-1)
+            elif img_cell.shape[-1] == 4:
+                img_cell = img_cell[..., :3]
+            if img_cell.dtype != np.uint8:
+                img_cell = img_cell.astype(np.uint8, copy=False)
+            return img_cell
+        return None
+    # PIL
+    if isinstance(img_cell, Image.Image):
+        return np.asarray(img_cell.convert("RGB"), dtype=np.uint8)
+    # dict with bytes
+    if isinstance(img_cell, dict):
+        data = img_cell.get("bytes")
+        if data is None:
+            path = img_cell.get("path")
+            if path and os.path.exists(path):
+                with Image.open(path) as im:
+                    return np.asarray(im.convert("RGB"), dtype=np.uint8)
+            return None
+        with Image.open(io.BytesIO(data)) as im:
+            return np.asarray(im.convert("RGB"), dtype=np.uint8)
+    # Unknown
+    return None
+
+
+def convert_molmoact_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+) -> Dataset:
+    """Stream MolmoAct LeRobot (parquet) and convert to HF, using episodes.jsonl for task text.
+
+    Assumes dataset_path contains one or more subdirectories, each with parquet files and an
+    associated episodes.jsonl. We iterate per subdirectory to avoid episode_index collisions,
+    grouping rows by `episode_index` and writing videos for `first_view`, `second_view`, and `wrist_image`.
+    """
+
+    root = Path(os.path.expanduser(dataset_path)) / dataset_name
+    if not root.exists():
+        raise FileNotFoundError(f"MolmoAct dataset path not found: {root}")
+
+    # Discover dataset subdirectories that have episodes.jsonl; if none, fallback to root
+    assert (root / "train" / "meta" / "episodes.jsonl").exists(), "episodes.jsonl not found"
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    entries: list[dict] = []
+    produced = 0
+    max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+    def load_episode_text_map(ds_dir: Path) -> dict[int, str]:
+        mapping: dict[int, str] = {}
+        jsonl_path = ds_dir / "train" / "meta" / "episodes.jsonl"
+        if not jsonl_path.exists():
+            return mapping
+        try:
+            import json
+
+            with open(jsonl_path, "r") as f:
+                for line in f:
+                    line = line.strip()
+                    if not line:
+                        continue
+                    try:
+                        obj = json.loads(line)
+                    except Exception:
+                        continue
+                    ep_idx = obj.get("episode_index")
+                    if ep_idx is None:
+                        ep_idx = obj.get("index")
+                    if ep_idx is None:
+                        continue
+                    text = (obj.get("tasks"))[0]
+                    if isinstance(text, str) and text.strip():
+                        mapping[int(ep_idx)] = text.strip()
+        except Exception:
+            pass
+        return mapping
+
+    def flush_episode(ep_idx: int, task_text: str, label: str, frames_by_view: dict[str, list[np.ndarray]]) -> None:
+        nonlocal produced, entries
+        if not frames_by_view:
+            return
+        if task_text not in lang_cache:
+            lang_cache[task_text] = lang_model.encode(task_text)
+        lang_vec = lang_cache[task_text]
+
+        for view_key, frames in frames_by_view.items():
+            if not frames:
+                continue
+            if isinstance(frames[0], np.ndarray) and np.all(frames[0] == 0):
+                continue
+
+            full_path, rel_path = _build_molmo_video_paths(
+                output_dir=output_dir,
+                dataset_label=label,
+                episode_idx=ep_idx,
+                view_key=view_key,
+            )
+
+            traj_dict = {
+                "id": generate_unique_id(),
+                "frames": frames,
+                "task": task_text,
+                "is_robot": True,
+                "quality_label": "successful",
+                "preference_group_id": None,
+                "preference_rank": None,
+            }
+
+            entry = create_hf_trajectory(
+                traj_dict=traj_dict,
+                video_path=full_path,
+                lang_vector=lang_vec,
+                max_frames=max_frames,
+                dataset_name=dataset_name,
+                use_video=True,
+                fps=fps,
+            )
+            if entry:
+                entry["frames"] = rel_path
+                entries.append(entry)
+                produced += 1
+
+    # Process each dataset directory independently to avoid ep-index collisions
+    ep_text_map = load_episode_text_map(root)
+
+    # Discover parquet files in ds_dir
+    data_files: list[str] = []
+    for pat in ("**/*.parquet", "*.parquet"):
+        data_files.extend([str(p) for p in root.glob(pat)])
+    if not data_files:
+        raise ValueError("No parquet files found")
+
+    ds_iter = load_dataset(
+        "parquet",
+        data_files={"train": data_files},
+        split="train",
+        streaming=True,
+    )
+
+    current_ep: Optional[int] = None
+    frames_by_view: dict[str, list[np.ndarray]] = {}
+    label = f"{dataset_name}"
+
+    for row in tqdm(ds_iter, desc=f"MolmoAct rows ({dataset_name})"):
+        if produced >= max_limit:
+            break
+        ep_idx = int(row.get("episode_index", -1))
+        if ep_idx < 0:
+            continue
+
+        if current_ep is None:
+            current_ep = ep_idx
+            frames_by_view = {"first_view": [], "second_view": []}
+        elif ep_idx != current_ep:
+            task_text = ep_text_map.get(current_ep)
+            print(f"{task_text} episode loaded")
+            flush_episode(current_ep, task_text, label, frames_by_view)
+            current_ep = ep_idx
+            frames_by_view = {"first_view": [], "second_view": []}
+
+        for view_key in ("first_view", "second_view"):
+            cell = row.get(view_key)
+            img = _to_rgb_numpy(cell)
+            if img is not None:
+                frames_by_view[view_key].append(img)
+
+        if produced >= max_limit:
+            break
+
+    if current_ep is not None and produced < max_limit:
+        task_text = ep_text_map.get(current_ep)
+        print(f"{task_text} episode loaded")
+        flush_episode(current_ep, task_text, label, frames_by_view)
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/motif_loader.py b/robometer/dataset_upload/dataset_loaders/motif_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..b3ef423288bd8aa068fbcf52fbe500115c4ff5e0
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/motif_loader.py
@@ -0,0 +1,131 @@
+import os
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+from dataset_upload.helpers import generate_unique_id
+
+
+class MotifFrameLoader:
+    """Pickle-able loader that reads frames for a single trajectory on demand.
+
+    Supports two backing sources:
+    - A video file path (e.g., .mp4)
+    - A directory of image frames (sorted by filename)
+    """
+
+    def __init__(self, source_path: str) -> None:
+        self.source_path = source_path
+
+    def _load_from_video(self) -> np.ndarray:
+        cap = cv2.VideoCapture(self.source_path)
+        frames = []
+        while True:
+            ok, frame_bgr = cap.read()
+            if not ok:
+                break
+            frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+        cap.release()
+
+        frames_np = np.asarray(frames)
+        if frames_np.ndim != 4 or frames_np.shape[-1] != 3:
+            raise ValueError(
+                f"Unexpected frames shape from video {self.source_path}: {getattr(frames_np, 'shape', None)}"
+            )
+        if frames_np.dtype != np.uint8:
+            frames_np = frames_np.astype(np.uint8, copy=False)
+        return frames_np
+
+    def __call__(self) -> np.ndarray:
+        p = Path(self.source_path)
+        if p.is_file():
+            return self._load_from_video()
+        raise FileNotFoundError(f"Source path not found: {self.source_path}")
+
+
+def _infer_is_robot_from_path(path: Path) -> bool:
+    parts = [s.lower() for s in path.parts]
+    # MotIF repo mentions 'human_motion' and 'stretch_motion'
+    if any("stretch" in s for s in parts):
+        return True
+    elif any("human" in s for s in parts):
+        return False
+    else:
+        raise ValueError(f"Unknown robot/human: {path}")
+
+
+def _make_traj(source_path: Path, task_text: str) -> dict:
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    traj["task"] = task_text
+    traj["frames"] = MotifFrameLoader(str(source_path))
+    traj["is_robot"] = _infer_is_robot_from_path(source_path)
+    traj["quality_label"] = "successful"
+    # traj["partial_success"] = 1
+    traj["data_source"] = "motif"
+    return traj
+
+
+def load_motif_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    """Load MoTiF dataset using FrameLoader without HF conversion.
+    Returns mapping: task -> list of trajectory dicts.
+    """
+    import json
+
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"MoTiF dataset path not found: {root}")
+
+    task_to_trajs: dict[str, list[dict]] = {}
+
+    # Annotations
+    ann_dir = root / "annotations"
+
+    all_human_trajs = {}
+    path_precursor = "human_motion/videos_raw"
+    json_data = json.load(open(ann_dir / "human_motion_data_info.json"))
+    for item in json_data:
+        src = item["video_path"].split("/")[-1]
+        full_vid_path = root / path_precursor / src
+        # assert the path exists
+        if not full_vid_path.exists():
+            print(f"Human video path not found: {full_vid_path}")
+            continue
+        instruction = item.get("task_instruction") + ": " + item.get("motion_description")
+        all_human_trajs.setdefault(instruction, []).append(full_vid_path)
+
+    all_stretch_trajs = {}
+    path_precursor = "stretch_motion/videos_raw"
+    json_data = json.load(open(ann_dir / "stretch_motion_data_info.json"))
+    for item in json_data:
+        src = item["video_path"].split("/")[-1]
+        full_vid_path = root / path_precursor / src
+        # assert the path exists
+        if not full_vid_path.exists():
+            print(f"Stretch video path not found: {full_vid_path}")
+            continue
+        instruction = item.get("task_instruction") + ": " + item.get("motion_description")
+        all_stretch_trajs.setdefault(instruction, []).append(full_vid_path)
+
+    # get the keys in both
+    common_keys = set(all_human_trajs.keys()) & set(all_stretch_trajs.keys())
+    all_stretch_trajs = {k: v for k, v in all_stretch_trajs.items() if k in common_keys}
+    all_human_trajs = {k: v for k, v in all_human_trajs.items() if k in common_keys}
+
+    print(f"Number of human tasks: {len(all_human_trajs)}")
+    print(f"Number of stretch tasks: {len(all_stretch_trajs)}")
+
+    for instruction, paths in all_human_trajs.items():
+        for path in paths:
+            traj = _make_traj(path, instruction)
+            task_to_trajs.setdefault(instruction, []).append(traj)
+
+    for instruction, paths in all_stretch_trajs.items():
+        for path in paths:
+            traj = _make_traj(path, instruction)
+            task_to_trajs.setdefault(instruction, []).append(traj)
+
+    return task_to_trajs
diff --git a/robometer/dataset_upload/dataset_loaders/mw_collected_loader.py b/robometer/dataset_upload/dataset_loaders/mw_collected_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..6ad021d5884f7e9d2e7b14aeb3a1fdf5e122850a
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/mw_collected_loader.py
@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+"""
+Metaworld dataset loader for the generic dataset converter for Robometer model training.
+This module contains logic for loading metaworld data organized by task and quality.
+
+uv run python dataset_upload/generate_hf_dataset.py \
+    --config_path=dataset_upload/configs/data_gen_configs/metaworld.yaml
+"""
+
+import collections
+from pathlib import Path
+import h5py
+import numpy as np
+from PIL import Image
+from torchvision import transforms
+from tqdm import tqdm
+
+from dataset_upload.video_helpers import load_video_frames
+from dataset_upload.dataset_loaders.mw_task_annotations import TRAIN_GT_ANN, EVAL_GT_ANN
+
+
+def apply_center_crop_to_frames(frames: np.ndarray) -> np.ndarray:
+    """Apply center crop (224, 224) to video frames using torchvision transforms.
+
+    Args:
+        frames: numpy array of shape (T, H, W, 3) in RGB order
+
+    Returns:
+        numpy array of shape (T, 224, 224, 3) with center cropped frames
+    """
+    # Define the center crop transform
+    center_crop = transforms.CenterCrop(224)
+
+    cropped_frames = []
+    for frame in frames:
+        # Convert numpy array to PIL Image
+        pil_frame = Image.fromarray(frame.astype(np.uint8))
+
+        # Apply center crop
+        cropped_pil = center_crop(pil_frame)
+
+        # Convert back to numpy array
+        cropped_frame = np.array(cropped_pil)
+        cropped_frames.append(cropped_frame)
+
+    return np.array(cropped_frames)
+
+
+def map_quality_label(original_label: str) -> str:
+    """Map original quality labels to standardized Robometer labels."""
+    label_mapping = {"GT": "successful", "success": "successful", "all_fail": "failure", "close_succ": "suboptimal"}
+    return label_mapping.get(original_label, original_label)
+
+
+def load_metaworld_dataset(base_path: str, dataset_name: str) -> dict[str, list[dict]]:
+    """Load metaworld dataset and organize by task.
+
+    Args:
+        base_path: Path to the metaworld dataset directory
+
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+
+    print(f"Loading metaworld dataset from: {base_path}")
+
+    task_data = collections.defaultdict(list)
+    base_path = Path(base_path)
+
+    if not base_path.exists():
+        raise FileNotFoundError(f"Metaworld dataset path not found: {base_path}")
+
+    if "train" in dataset_name:
+        tasks = TRAIN_GT_ANN.keys()
+        anns = TRAIN_GT_ANN
+    elif "eval" in dataset_name:
+        tasks = EVAL_GT_ANN.keys()
+        anns = EVAL_GT_ANN
+
+    print("number of tasks: ", len(tasks))
+
+    scucessful_trajs_file = "downloaded_data/metaworld_video.h5"
+
+    with h5py.File(scucessful_trajs_file, "r") as f:
+        print("Available tasks: ", f.keys())
+        print("number of tasks: ", len(f.keys()))
+        task_names = list(f.keys())
+
+        for task_name in tqdm(task_names, desc="Loading successful trajectories"):
+            if task_name not in tasks:
+                continue
+
+            for traj_name in ["0", "1", "10", "11", "12"]:  # not sure why we use these, but we use these 5
+                traj = f[task_name][traj_name]
+                frames = np.array(traj)
+
+                cropped_frames = apply_center_crop_to_frames(frames)
+
+                trajectory = {
+                    "frames": cropped_frames,
+                    "task": anns.get(task_name),
+                    "quality_label": "successful",
+                    "is_robot": True,
+                    "partial_success": 0,
+                }
+                task_data[task_name].append(trajectory)
+
+    if "eval" in dataset_name:
+        task_dirs = [d for d in base_path.iterdir() if d.is_dir() and not d.name.startswith(".")]
+        for task_dir in tqdm(task_dirs, desc="Processing tasks"):
+            task_name = task_dir.name
+
+            if task_name in [".DS_Store"]:
+                continue
+
+            # Find all quality label directories within this task
+            quality_dirs = [d for d in task_dir.iterdir() if d.is_dir() and not d.name.startswith(".")]
+
+            for quality_dir in quality_dirs:
+                original_quality_label = quality_dir.name
+
+                # Map quality label to standardized format
+                quality_label = map_quality_label(original_quality_label)
+
+                # Find all video files in this quality directory
+                video_files = list(quality_dir.glob("*.mp4")) + list(quality_dir.glob("*.gif"))
+
+                # if len(video_files) != 2:
+                #     import ipdb; ipdb.set_trace()
+
+                for video_file in video_files:
+                    # # Extract index from filename (e.g., "1.mp4" -> 1)
+                    # try:
+                    #     int(video_file.stem)
+                    # except ValueError:
+                    #     print(f"Warning: Could not parse index from filename: {video_file.name}")
+
+                    # Load frames and apply center crop
+                    original_frames = load_video_frames(video_file)
+                    cropped_frames = apply_center_crop_to_frames(original_frames)
+
+                    nl_name = anns.get(task_name)
+
+                    # Create trajectory entry
+                    trajectory = {
+                        "frames": cropped_frames,
+                        "task": nl_name,  # Natural language task
+                        "quality_label": quality_label,  # Mapped quality label
+                        "is_robot": True,
+                        "partial_success": 0,
+                    }
+
+                    task_data[task_name].append(trajectory)
+
+    for task_name, trajectories in task_data.items():
+        print(f"Task {task_name}: {len(trajectories)} trajectories")
+
+    total_trajectories = sum(len(trajectories) for trajectories in task_data.values())
+    print(f"\nLoaded {total_trajectories} trajectories from {len(task_data)} tasks")
+
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/mw_task_annotations.py b/robometer/dataset_upload/dataset_loaders/mw_task_annotations.py
new file mode 100644
index 0000000000000000000000000000000000000000..44b1523bbd0a7be6c6834e7c9edd6040387f2f32
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/mw_task_annotations.py
@@ -0,0 +1,205 @@
+TRAIN_GT_ANN = {
+    "button-press-v2": "Press the button from side",
+    "button-press-topdown-wall-v2": "Press the button from top",
+    "coffee-pull-v2": "Pull the coffee cup",
+    "dial-turn-v2": "Turn the dial",
+    "door-open-v2": "Open the door",
+    "door-unlock-v2": "Turn door lock clockwise",
+    "drawer-close-v2": "Close the drawer",
+    "faucet-open-v2": "Open the faucet",
+    "handle-press-v2": "Press the handle",
+    "handle-pull-side-v2": "Pull the handle up from the side",
+    "peg-insert-side-v2": "Insert the peg from the side",
+    "pick-place-v2": "Pick up the block and placing it to the goal position",
+    "plate-slide-v2": "Slide the plate into the gate",
+    "plate-slide-back-side-v2": "Slide the plate out of the gate from the side",
+    "push-v2": "Push the block to the goal",
+    "reach-v2": "Reach the goal",
+    "stick-push-v2": "Push the stick",
+    "stick-pull-v2": "Pull the stick",
+    "window-open-v2": "Open the window",
+    "hand-insert-v2": "Pick up the block and insert it into the hole",
+}
+
+EVAL_GT_ANN = {
+    "button-press-wall-v2": "Press the button from side",
+    "button-press-topdown-v2": "Press the button from top",
+    "coffee-push-v2": "Push the coffee cup",
+    "coffee-button-v2": "Press the coffee button",
+    "door-close-v2": "Close the door",
+    "door-lock-v2": "Turn door lock counter-clockwise",
+    "faucet-close-v2": "Close the faucet",
+    "handle-press-side-v2": "Press the handle from side",
+    "handle-pull-v2": "Pull the handle",
+    "pick-place-wall-v2": "Pick up the block and place it to the goal position",
+    "plate-slide-back-v2": "Slide the plate out of the gate",
+    "plate-slide-side-v2": "Slide the plate into the gate from the side",
+    "push-back-v2": "Push the block back to the goal",
+    "reach-wall-v2": "Reach the goal",
+    "soccer-v2": "Slide the ball into the gate",
+    "window-close-v2": "Close the window",
+    "sweep-into-v2": "Sweep the block into the hole",
+}
+
+GENERATE_TRAIN_ANN = {
+    "button-press-v2": [
+        "Pressing the button from the side",
+        "Pressing the red button",
+        "Pressing the button on the yellow box",
+        "Pushing the button inside",
+    ],
+    "button-press-topdown-wall-v2": [
+        "Press the button from top",
+        "Press the red button",
+        "Press the red button on the yellow box",
+        "Push the button downward",
+    ],
+    "coffee-pull-v2": [
+        "Pull the coffee cup from the coffee machine",
+        "Extract the coffee cup",
+        "Remove the coffee cup from the coffee machine",
+        "Drag the coffee cup",
+    ],
+    "dial-turn-v2": [
+        "Turn the dial counterclockwise",
+        "Rotate the dial to the left",
+        "Twist the dial",
+        "Adjust the dial counterclockwise",
+    ],
+    "door-open-v2": [
+        "Rotate the door counterclockwise",
+        "Swing the door open",
+        "Draw the door open",
+        "Draw the door counterclockwise",
+    ],
+    "door-unlock-v2": [
+        "Turn door lock clockwise",
+        "Rotate the door lock",
+        "Twist the door lock",
+        "Adjust the door lock",
+    ],
+    "drawer-close-v2": ["Push the drawer closed", "Shut the drawer", "Slide the drawer shut", "Close the green drawer"],
+    "faucet-open-v2": [
+        "Rotate the faucet clockwise",
+        "Twist the faucet in a clockwise direction",
+        "Turn the faucet handle to the left",
+        "Adjust the faucet clockwise",
+    ],
+    "handle-press-v2": [
+        "Press the handle downward",
+        "Depress the handle",
+        "Lower the handle",
+        "Apply downward pressure on the handle",
+    ],
+    "handle-pull-side-v2": [
+        "Pull the handle upward",
+        "Lift the handle",
+        "Raise the handle",
+        "Apply upward force to the handle",
+    ],
+    "peg-insert-side-v2": [
+        "Push the peg into the hole from the side",
+        "Move the peg into the box",
+        "Insert the peg into the hole",
+        "Guide the peg into the hole",
+    ],
+    "pick-place-v2": [
+        "Lift the block to the goal position",
+        "Pick up the block and setting it at the goal position",
+        "Grab the block and positioning it at the goal",
+        "Move the box to the goal position",
+    ],
+    "plate-slide-v2": [
+        "Push the plate into the gate",
+        "Slide the plate into the gate",
+        "Move the plate into the gate",
+        "Shift the plate into the gate",
+    ],
+    "plate-slide-back-side-v2": [
+        "Pull the plate out of the gate",
+        "Slide the plate out of the gate",
+        "Move the plate out of the gate",
+        "Shift the plate out of the gate",
+    ],
+    "push-v2": [
+        "Move the block to the goal",
+        "Slide the block toward the goal",
+        "Propel the block to the goal",
+        "Drive the block toward the goal",
+    ],
+    "reach-v2": ["Extend to the goal", "Stretch toward the goal", "Move to the goal", "Direct toward the goal"],
+    "stick-push-v2": ["Move the stick", "Slide the stick", "Shift the stick", "Propel the stick"],
+    "stick-pull-v2": ["Pull the stick", "Extract the stick", "Move the stick to the bottle", "Drag the stick"],
+    "window-open-v2": [
+        "Pull the window open",
+        "Slide the window open",
+        "Move the window open",
+        "Shift the window open",
+    ],
+    "hand-insert-v2": [
+        "Move up the block and put it into the hole",
+        "Insert the block into the hole",
+        "Place the block into the hole",
+        "Guide the block into the hole",
+    ],
+}
+
+EVAL_ANN_1 = {
+    "button-press-wall-v2": "Press the button near the wall",
+    "button-press-topdown-v2": "Press the button down from above",
+    "coffee-push-v2": "Slide the coffee cup forward",
+    "coffee-button-v2": "Press the coffee button to activate",
+    "door-close-v2": "Shut the door",
+    "door-lock-v2": "Turn the door lock to the left",
+    "faucet-close-v2": "Turn the faucet off",
+    "handle-press-side-v2": "Press the handle down from side",
+    "handle-pull-v2": "Pull the handle up",
+    "pick-place-wall-v2": "Pick up the block and set it in the target spot bypassing the wall",
+    "plate-slide-back-v2": "Pull the plate backward from the gate",
+    "plate-slide-side-v2": "Guid the plate into the gate from the side",
+    "push-back-v2": "Push the block backward into place",
+    "reach-wall-v2": "Get to the goal position",
+    "soccer-v2": "Put the ball into the goal",
+    "window-close-v2": "Slide the window to close",
+    "sweep-into-v2": "Sweep the item into the hole",
+}
+
+EVAL_ANN_2 = {
+    "button-press-wall-v2": "Press the button behind the wall",
+    "button-press-topdown-v2": "Press the button downward",
+    "coffee-push-v2": "Move the coffee cup forward",
+    "coffee-button-v2": "Activate the coffee button by pressing",
+    "door-close-v2": "Bring the door to close",
+    "door-lock-v2": "Rotate the door lock counter-clockwise",
+    "faucet-close-v2": "Shut off the faucet",
+    "handle-press-side-v2": "Engage the handle with a press",
+    "handle-pull-v2": "Activate the handle by pulling",
+    "pick-place-wall-v2": "Lift the block and depositing it at the goal location over the wall",
+    "plate-slide-back-v2": "Slide the plate away from the gate",
+    "plate-slide-side-v2": "Slide the plate into the gate from a sideways approach",
+    "push-back-v2": "Take the block back to goal place",
+    "reach-wall-v2": "Move towards the goal position",
+    "soccer-v2": "Direct the ball into the gate",
+    "window-close-v2": "Move the window to closed position",
+    "sweep-into-v2": "Shovel the block into the opening",
+}
+
+EVAL_ANN_3 = {
+    "button-press-wall-v2": "Press the side-mounted button on the wall",
+    "button-press-topdown-v2": "Press the button firmly from above",
+    "coffee-push-v2": "Push the coffee cup along the table",
+    "coffee-button-v2": "Press the coffee maker's button to start",
+    "door-close-v2": "Close the door until it latches",
+    "door-lock-v2": "Turn the door lock left to secure",
+    "faucet-close-v2": "Rotate the faucet off to stop the flow",
+    "handle-press-side-v2": "Press the handle from a side angle",
+    "handle-pull-v2": "Pull the handle outward to engage",
+    "pick-place-wall-v2": "Pick up the block and placing it over the wall at the goal",
+    "plate-slide-back-v2": "Move the plate backward from the gate",
+    "plate-slide-side-v2": "Move the plate sideways into the gate",
+    "push-back-v2": "Push the block back to its final position",
+    "reach-wall-v2": "Reach the designated location near the wall",
+    "soccer-v2": "Slide the ball into the goal area",
+    "window-close-v2": "Draw the window shut",
+    "sweep-into-v2": "Sweep the block into the hole",
+}
diff --git a/robometer/dataset_upload/dataset_loaders/new_mit_franka_loader.py b/robometer/dataset_upload/dataset_loaders/new_mit_franka_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..3929138442da2442aa3b622cc1262f44f8bbd4d2
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/new_mit_franka_loader.py
@@ -0,0 +1,262 @@
+import json
+import os
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+from datasets import Dataset
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+
+# Task name to language instruction mapping
+TASK_INSTRUCTIONS = {
+    "foldtowel": "Fold the towel in half",
+    "movebanana": "Pick up the banana from the blue plate and place it on the green plate",
+    "movemouse": "Pick up the mouse and place it right next to the laptop, while avoiding spilling coffee",
+    "pourpebble": "Pour the pebbles from the cup onto the plate",
+    "pulltissue": "Pull a tissue from the tissue box",
+    "putspoon": "Pick up the spoon and place it inside the cup",
+    "stirpot": "Pick up the spatula and stir the beans in the pot",
+}
+
+
+def _load_video_frames(video_path: str) -> list[np.ndarray]:
+    """Load frames from an MP4 video file."""
+    if not os.path.exists(video_path):
+        raise FileNotFoundError(f"Video file not found: {video_path}")
+
+    frames = []
+    cap = cv2.VideoCapture(video_path)
+
+    try:
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+    finally:
+        cap.release()
+
+    return frames
+
+
+def _build_video_paths(
+    output_dir: str, dataset_label: str, episode_idx: int, task_name: str, view: str
+) -> tuple[str, str]:
+    """Build output video paths with shard structure."""
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"{task_name}_{view}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_episode(args):
+    """Process a single episode."""
+    episode_meta, lang_vec, output_dir, dataset_label, max_frames, fps = args
+
+    # Load video frames
+    video_path = episode_meta["video_path"]
+    try:
+        frames = _load_video_frames(video_path)
+    except Exception as e:
+        print(f"Error loading video {video_path}: {e}")
+        return None
+
+    if not frames:
+        return None
+
+    # Normalize quality label
+    quality_label = episode_meta["quality_label"]
+    if quality_label == "success":
+        quality_label = "successful"
+    elif quality_label == "fail":
+        quality_label = "failure"
+    # "suboptimal" stays as is
+
+    # Build output video path
+    full_path, rel_path = _build_video_paths(
+        output_dir, dataset_label, episode_meta["episode_idx"], episode_meta["task_name"], episode_meta["view"]
+    )
+
+    traj_dict = {
+        "id": generate_unique_id(),
+        "frames": frames,
+        "task": episode_meta["instruction"],
+        "is_robot": True,
+        "quality_label": quality_label,
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    entry = create_hf_trajectory(
+        traj_dict=traj_dict,
+        video_path=full_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_label,
+        use_video=True,
+        fps=fps,
+    )
+
+    if entry:
+        entry["frames"] = rel_path
+        return entry
+    return None
+
+
+def convert_new_mit_franka_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+    exclude_wrist_cam: bool = False,
+) -> Dataset:
+    """Convert New MIT Franka dataset to HF format.
+
+    Args:
+        dataset_path: Path to the dataset directory containing task folders
+        dataset_name: Name for the dataset
+        output_dir: Output directory for processed videos
+        max_trajectories: Maximum number of trajectories to process (None for all)
+        max_frames: Maximum frames per trajectory
+        fps: Frames per second for output videos
+        num_workers: Number of worker processes (-1 for auto, 0 for sequential)
+        exclude_wrist_cam: If True, only process external camera views and skip wrist camera
+
+    Returns:
+        HuggingFace Dataset
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset directory not found: {dataset_path}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Load sentence transformer model
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Collect all episodes
+    all_episodes = []
+    episode_idx = 0
+
+    # Iterate through task folders
+    task_folders = [d for d in root.iterdir() if d.is_dir()]
+
+    for task_folder in sorted(task_folders):
+        task_name = task_folder.name
+        if task_name not in TASK_INSTRUCTIONS:
+            print(f"Warning: Unknown task '{task_name}', skipping...")
+            continue
+
+        instruction = TASK_INSTRUCTIONS[task_name]
+
+        # Iterate through quality folders (suboptimal, failure, success)
+        for quality_folder in ["suboptimal", "failure", "success"]:
+            quality_dir = task_folder / quality_folder
+            if not quality_dir.exists():
+                continue
+
+            # Get all video files
+            video_files = sorted(quality_dir.glob("*.mp4"))
+
+            for video_file in video_files:
+                # Determine view type (ext or wrist)
+                if "_ext.mp4" in video_file.name:
+                    view = "ext"
+                elif "_wrist.mp4" in video_file.name:
+                    view = "wrist"
+                else:
+                    print(f"Warning: Unknown view type in {video_file.name}, skipping...")
+                    continue
+
+                # Skip wrist camera if exclude_wrist_cam is True
+                if exclude_wrist_cam and view == "wrist":
+                    continue
+
+                episode_meta = {
+                    "episode_idx": episode_idx,
+                    "task_name": task_name,
+                    "video_path": str(video_file),
+                    "instruction": instruction,
+                    "quality_label": quality_folder,  # "success", "failure", or "suboptimal"
+                    "view": view,
+                }
+                all_episodes.append(episode_meta)
+                episode_idx += 1
+
+    print(f"Found {len(all_episodes)} episodes across {len(task_folders)} tasks")
+
+    # Limit trajectories if specified
+    if max_trajectories is not None and max_trajectories > 0:
+        all_episodes = all_episodes[:max_trajectories]
+
+    # Pre-compute language embeddings
+    print("Computing language embeddings...")
+    for episode in tqdm(all_episodes, desc="Computing embeddings"):
+        instruction = episode["instruction"]
+        if instruction not in lang_cache:
+            lang_cache[instruction] = lang_model.encode(instruction)
+
+    # Process episodes
+    entries = []
+    if num_workers == 1:
+        # Sequential processing
+        for episode in tqdm(all_episodes, desc="Processing episodes"):
+            lang_vec = lang_cache[episode["instruction"]]
+            result = _process_episode((episode, lang_vec, output_dir, dataset_name, max_frames, fps))
+            if result:
+                entries.append(result)
+    else:
+        # Parallel processing
+        from multiprocessing import Pool
+
+        args_list = [
+            (episode, lang_cache[episode["instruction"]], output_dir, dataset_name, max_frames, fps)
+            for episode in all_episodes
+        ]
+
+        with Pool(processes=num_workers) as pool:
+            results = list(
+                tqdm(pool.imap_unordered(_process_episode, args_list), total=len(args_list), desc="Processing episodes")
+            )
+
+        entries = [r for r in results if r is not None]
+
+    print(f"Successfully processed {len(entries)} episodes")
+
+    # Create HuggingFace dataset
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/oxe_loader.py b/robometer/dataset_upload/dataset_loaders/oxe_loader.py
new file mode 100755
index 0000000000000000000000000000000000000000..f4e23687b97ec4435bcf3eb8094dc01d5d11d4f1
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/oxe_loader.py
@@ -0,0 +1,457 @@
+import os
+import cv2
+import gc
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from dataset_upload.dataset_helpers.oxe_helper import OXE_DATASET_CONFIGS
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+from tqdm import tqdm
+
+from datasets import Dataset
+
+# Disable GPUs for TensorFlow in this loader to avoid CUDA context issues in workers
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+DEBUG_MODE = False
+
+import tensorflow_datasets as tfds
+
+OXE_VALID_DATASETS = [
+    "austin_buds_dataset_converted_externally_to_rlds",
+    "austin_sirius_dataset_converted_externally_to_rlds",
+    "bc_z",
+    "berkeley_cable_routing",
+    "berkeley_fanuc_manipulation",
+    "bridge_v2",
+    "dlr_edan_shared_control_converted_externally_to_rlds",
+    "droid",
+    "fmb",
+    "fractal20220817_data",
+    "furniture_bench_dataset_converted_externally_to_rlds",
+    "iamlab_cmu_pickup_insert_converted_externally_to_rlds",
+    "jaco_play",
+    "language_table",
+    "stanford_hydra_dataset_converted_externally_to_rlds",
+    "taco_play",
+    "toto",
+    "ucsd_kitchen_dataset_converted_externally_to_rlds",
+    "utaustin_mutex",
+    "viola",
+    # not in original
+    "robo_set",
+    "aloha_mobile",
+    "imperialcollege_sawyer_wrist_cam",
+    "kaist_nonprehensile_converted_externally_to_rlds",
+    "berkeley_mvp_converted_externally_to_rlds",
+    "berkeley_rpt_converted_externally_to_rlds",
+    "nyu_rot_dataset_converted_externally_to_rlds",
+    "tokyo_u_lsmo_converted_externally_to_rlds",
+]
+POSSIBLE_LANG_INSTRUCTION_KEYS = [  # valid keys for language instruction in OXE
+    "natural_language_instruction",
+    "language_instruction",
+    "instruction",
+    "language_instruction1",
+    "language_instruction2",
+    "language_instruction3",
+]
+MAX_LANGTABLE_EPISODES = (
+    50_000  # for language table, we only want to label the first 50k episodes b/c it's way too many
+)
+possible_valid_keys = ["primary", "secondary", "tertiary"]
+
+
+def _stable_shard_for_index(index: int, shard_modulus: int = 1000) -> str:
+    try:
+        idx = int(index)
+    except Exception:
+        idx = abs(hash(str(index)))
+    shard_index = idx // shard_modulus
+    return f"shard_{shard_index:04d}"
+
+
+def _build_oxe_video_paths(
+    output_dir: str,
+    dataset_label: str,
+    episode_idx: int,
+    view_key: str,
+) -> tuple[str, str]:
+    shard_dir = _stable_shard_for_index(episode_idx)
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip@{view_key}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_single_oxe_episode(args):
+    """Worker function to process a single OXE episode.
+
+    This function must be defined at module level to be picklable for multiprocessing.
+    """
+    episode, ep_idx, task, lang_vec, output_dir, dataset_name, max_frames, fps, valid_img_keys = args
+
+    episode_entries = []
+
+    # Episode is already converted to numpy format
+
+    for img_key in valid_img_keys:
+        # Check first frame for all-black to prune
+        if img_key not in episode[0]["observation"]:
+            continue
+        if np.all(episode[0]["observation"][img_key] == 0):
+            continue
+
+        frames = [s["observation"][img_key] for s in episode if img_key in s["observation"]]
+
+        if not frames:
+            continue
+
+        if "nyu_rot_dataset_converted_externally_to_rlds" in dataset_name:
+            # convert each frame from bgr to rgb
+            frames = [cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) for frame in frames]
+
+        full_path, rel_path = _build_oxe_video_paths(
+            output_dir=output_dir,
+            dataset_label=dataset_name,
+            episode_idx=ep_idx,
+            view_key=img_key,
+        )
+
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,
+            "task": task,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+
+        entry = create_hf_trajectory(
+            traj_dict=traj_dict,
+            video_path=full_path,
+            lang_vector=lang_vec,
+            max_frames=max_frames,
+            dataset_name=dataset_name,
+            use_video=True,
+            fps=fps,
+        )
+
+        # Clear frames from memory immediately after video creation
+        del frames
+
+        if entry:
+            entry["frames"] = rel_path
+            episode_entries.append(entry)
+
+    # Clear frames from memory after processing
+    del episode
+
+    return episode_entries
+
+
+def convert_oxe_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert a single OXE TFDS dataset to HF format by writing videos directly.
+
+    Args:
+        dataset_path: Root path containing TFDS builder directories
+        dataset_name: Name prefixed with 'oxe_', e.g., 'oxe_language_table'
+        output_dir: Where to write video files and dataset
+        max_trajectories: Limit number of produced trajectories (None for all)
+        max_frames: Max frames per video
+        fps: Video fps
+
+    Returns:
+        datasets.Dataset with entries containing relative video paths.
+    """
+
+    # Normalize name and basic checks
+    if dataset_name is None:
+        raise ValueError("dataset_name is required")
+
+    base_ds_name = dataset_name.replace("oxe_", "")
+
+    if base_ds_name.endswith("_eval"):
+        base_ds_name = base_ds_name[:-5]
+        EVAL_MODE = True
+        # use eval/val/test
+    else:
+        EVAL_MODE = False
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {root}")
+
+    # Find builder directory/version
+    versions = os.listdir(f"{root}/{base_ds_name}")
+    if len(versions) == 0:
+        raise ValueError(f"No versions found for {base_ds_name} in {root}")
+
+    builder = None
+    for version in versions:
+        if "incomplete" in version:
+            continue
+        builder = tfds.builder_from_directory(f"{root}/{base_ds_name}/{version}")
+        break
+    if builder is None:
+        raise ValueError(f"No valid builder found for {base_ds_name} in {root}")
+
+    if EVAL_MODE:
+        ds_all_dict = builder.as_dataset()
+        splits = list(ds_all_dict.keys())
+        splits.remove("train")
+        if len(splits) == 0:
+            raise ValueError(f"No valid EVAL dataset found for {base_ds_name} in {root}")
+        elif len(splits) == 1:
+            dataset = builder.as_dataset(split=splits[0], shuffle_files=False)
+        else:
+            raise ValueError(f"Multiple EVAL splits found for {base_ds_name} in {root}: {splits}")
+        print(f"Loaded EVAL dataset for {base_ds_name} in {root}")
+        # splits = ["val", "test"]
+        # for split in splits:
+        #    try:
+        #        dataset = builder.as_dataset(split=split, shuffle_files=False)
+        #        break
+        #    except Exception as e:
+        #        print(f"Error loading {split} split: {e}")
+        #        dataset = None
+        #        continue
+        # if dataset is None:
+        #    raise ValueError(f"No valid {EVAL_MODE} dataset found for {base_ds_name} in {root}")
+    else:
+        dataset = builder.as_dataset(split="train", shuffle_files=False)
+
+    # Determine valid image observation keys
+    img_key_to_name = OXE_DATASET_CONFIGS[base_ds_name]["image_obs_keys"]
+    if "droid" not in base_ds_name:  # make sure to use DROID's wrist cam
+        img_key_to_name = {k: v for k, v in img_key_to_name.items() if k != "wrist"}
+    valid_img_keys = list(img_key_to_name.values())
+
+    # Determine number of workers
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)  # or else ram usage will blow up
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Language model and cache
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    entries: list[dict[str, Any]] = []
+    produced = 0
+    if DEBUG_MODE:
+        max_limit = 100
+    else:
+        max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+    if "language_table" in base_ds_name:
+        max_limit = MAX_LANGTABLE_EPISODES
+
+        # Process episodes in batches to avoid OOM
+    batch_size = 32  # Process episodes in smaller batches
+    entries = []
+    produced = 0
+
+    print(f"Processing episodes in batches of {batch_size} with {num_workers} workers...")
+
+    # Process episodes in batches to manage memory
+    episode_batch = []
+    episode_info_batch = []
+
+    for ep_idx, episode in enumerate(tqdm(dataset, desc=f"Processing {base_ds_name} episodes")):
+        if ep_idx >= max_limit:
+            break
+
+        # Materialize first step for language and sanity checks
+        try:
+            first_step = next(iter(tfds.as_numpy(episode["steps"])))
+        except StopIteration:
+            continue
+
+        # Extract task/instruction
+        task: str | None = None
+        for key in POSSIBLE_LANG_INSTRUCTION_KEYS:
+            if key in first_step.get("observation", {}):
+                if base_ds_name == "language_table":
+                    t = first_step["observation"][key]
+                    task = bytes(t[np.where(t != 0)].tolist()).decode("utf-8")
+                else:
+                    task = first_step["observation"][key].decode()
+                break
+            elif key in first_step:
+                task = first_step[key].decode()
+                break
+        if not task:
+            continue
+
+        # Precompute embedding
+        if task not in lang_cache:
+            lang_cache[task] = lang_model.encode(task)
+        lang_vec = lang_cache[task]
+
+        # Convert TensorFlow objects to numpy for pickling
+        try:
+            # Convert episode to numpy format for multiprocessing
+            episode_np = tfds.as_numpy(episode)
+
+            # iterate through all steps and just store as a list
+            episode_np = list(episode_np["steps"])
+
+            episode_batch.append(episode_np)
+            episode_info_batch.append((ep_idx, task, lang_vec))
+
+        except Exception as e:
+            print(f"Warning: Failed to convert episode {ep_idx} to numpy: {e}")
+            continue
+
+        # Process batch when it's full or we've reached the limit
+        if len(episode_batch) >= batch_size or ep_idx + 1 >= max_limit:
+            print(f"Processing batch of {len(episode_batch)} episodes...")
+
+            if num_workers == 1:
+                # Sequential processing
+                for args in zip(
+                    episode_batch,
+                    [info[0] for info in episode_info_batch],
+                    [info[1] for info in episode_info_batch],
+                    [info[2] for info in episode_info_batch],
+                    [output_dir] * len(episode_batch),
+                    [dataset_name] * len(episode_batch),
+                    [max_frames] * len(episode_batch),
+                    [fps] * len(episode_batch),
+                    [valid_img_keys] * len(episode_batch),
+                    strict=False,
+                ):
+                    episode_entries = _process_single_oxe_episode(args)
+                    entries.extend(episode_entries)
+                    produced += len(episode_entries)
+            else:
+                # Parallel processing
+                from multiprocessing import Pool
+
+                # Prepare arguments for workers
+                worker_args = list(
+                    zip(
+                        episode_batch,
+                        [info[0] for info in episode_info_batch],
+                        [info[1] for info in episode_info_batch],
+                        [info[2] for info in episode_info_batch],
+                        [output_dir] * len(episode_batch),
+                        [dataset_name] * len(episode_batch),
+                        [max_frames] * len(episode_batch),
+                        [fps] * len(episode_batch),
+                        [valid_img_keys] * len(episode_batch),
+                        strict=False,
+                    )
+                )
+
+                with Pool(processes=num_workers) as pool:
+                    results = list(
+                        tqdm(
+                            pool.imap_unordered(_process_single_oxe_episode, worker_args),
+                            total=len(worker_args),
+                            desc=f"Processing batch (workers={num_workers})",
+                        )
+                    )
+
+                # Collect all results
+                for episode_entries in results:
+                    entries.extend(episode_entries)
+                    produced += len(episode_entries)
+
+            # Clear batch for next iteration
+            episode_batch = []
+            episode_info_batch = []
+
+            # Force garbage collection to free memory
+            gc.collect()
+
+            # Check if we've reached the limit
+            if produced >= max_limit:
+                break
+
+        # For language_table, cap the number of episodes considered
+        if base_ds_name == "language_table" and ep_idx + 1 >= MAX_LANGTABLE_EPISODES:
+            break
+
+    # After iterating all episodes, process any remaining batch
+    if episode_batch:
+        if num_workers == 1:
+            for args in zip(
+                episode_batch,
+                [info[0] for info in episode_info_batch],
+                [info[1] for info in episode_info_batch],
+                [info[2] for info in episode_info_batch],
+                [output_dir] * len(episode_batch),
+                [dataset_name] * len(episode_batch),
+                [max_frames] * len(episode_batch),
+                [fps] * len(episode_batch),
+                [valid_img_keys] * len(episode_batch),
+                strict=False,
+            ):
+                episode_entries = _process_single_oxe_episode(args)
+                entries.extend(episode_entries)
+                produced += len(episode_entries)
+        else:
+            from multiprocessing import Pool
+
+            worker_args = list(
+                zip(
+                    episode_batch,
+                    [info[0] for info in episode_info_batch],
+                    [info[1] for info in episode_info_batch],
+                    [info[2] for info in episode_info_batch],
+                    [output_dir] * len(episode_batch),
+                    [dataset_name] * len(episode_batch),
+                    [max_frames] * len(episode_batch),
+                    [fps] * len(episode_batch),
+                    [valid_img_keys] * len(episode_batch),
+                    strict=False,
+                )
+            )
+            with Pool(processes=num_workers) as pool:
+                results = list(
+                    tqdm(
+                        pool.imap_unordered(_process_single_oxe_episode, worker_args),
+                        total=len(worker_args),
+                        desc=f"Processing batch (workers={num_workers})",
+                    )
+                )
+            for episode_entries in results:
+                entries.extend(episode_entries)
+                produced += len(episode_entries)
+                if produced >= max_limit:
+                    break
+
+            # Force garbage collection after final batch
+            gc.collect()
+
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/ph2d_loader.py b/robometer/dataset_upload/dataset_loaders/ph2d_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..8b243d6c83f44e033a805d772b89759a8d94f962
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/ph2d_loader.py
@@ -0,0 +1,104 @@
+import os
+import json
+from collections import defaultdict
+
+import cv2
+import h5py
+import numpy as np
+from dataset_upload.helpers import generate_unique_id
+
+
+DEBUG = False
+
+
+class Ph2dFrameloader:
+    """Pickle-able loader to read frames from an HDF5 file on demand.
+
+    Reads frames from either 'observation.image.right' or 'observation.image.left'.
+    Each stored frame is expected to be an encoded image buffer which is decoded
+    with cv2.imdecode and converted to RGB.
+    """
+
+    def __init__(self, hdf5_path: str, camera: str = "right") -> None:
+        self.hdf5_path = hdf5_path
+        self.camera = camera.lower()
+
+    def __call__(self) -> np.ndarray:
+        key = "observation.image.right" if self.camera == "right" else "observation.image.left"
+
+        with h5py.File(self.hdf5_path, "r") as f:
+            if key not in f:
+                raise KeyError(f"Key '{key}' not found in {self.hdf5_path}")
+
+            data = f[key]
+            encoded_frames = data[()]
+            frames = []
+            for encoded_frame in encoded_frames:
+                frame = cv2.imdecode(encoded_frame, cv2.IMREAD_COLOR)
+                frames.append(frame)
+
+        frames_np = np.asarray(frames, dtype=np.uint8)
+        if frames_np.ndim != 4 or frames_np.shape[-1] != 3:
+            raise ValueError(f"Unexpected frames shape for {self.hdf5_path}: {frames_np.shape} (expected (T,H,W,3))")
+        return frames_np
+
+
+def create_new_trajectory(hdf5_path: str, is_robot: bool, caption: str, camera: str) -> dict:
+    trajectory_info = {}
+    trajectory_info["id"] = generate_unique_id()
+    trajectory_info["task"] = caption
+    trajectory_info["frames"] = Ph2dFrameloader(hdf5_path, camera=camera)
+    trajectory_info["is_robot"] = is_robot
+    trajectory_info["quality_label"] = "successful"
+    trajectory_info["partial_success"] = 1
+    trajectory_info["data_source"] = "ph2d"
+    return trajectory_info
+
+
+def load_ph2d_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    """Load Ph2d dataset organized as folders with HDF5 files and optional metadata.json.
+
+    Expected layout:
+        <dataset_path>/
+            ph2d_metadata.json   # Optional (dataset-specific). TODO: parse to captions/tasks
+            <sequence_1>/
+                *.hdf5
+            <sequence_2>/
+                ...
+
+    Args:
+        dataset_path: Root directory containing sequence folders of HDF5 files.
+
+    Returns:
+        Mapping: task/caption -> list of trajectory dicts.
+    """
+
+    task_data: dict[str, list[dict]] = defaultdict(list)
+
+    # load metadata.json
+    all_task_attributes = json.load(open(os.path.join(dataset_path, "ph2d_metadata.json")))["per_task_attributes"]
+    for task_name, task_attributes in all_task_attributes.items():
+        # load all sequences
+        embodiment_type = task_attributes["embodiment_type"]
+        if "human" in embodiment_type:
+            is_robot = False
+        else:
+            is_robot = True
+
+        for h5_file in os.listdir(os.path.join(dataset_path, task_name)):
+            if not h5_file.endswith(".hdf5"):
+                continue
+
+            h5_path = os.path.join(dataset_path, task_name, h5_file)
+            loaded_data = h5py.File(h5_path, "r")
+            task_description = loaded_data.attrs["description"]
+
+            if "observation.image.right" in loaded_data:
+                task_data[task_name].append(create_new_trajectory(h5_path, is_robot, task_description, camera="right"))
+            if "observation.image.left" in loaded_data:
+                task_data[task_name].append(create_new_trajectory(h5_path, is_robot, task_description, camera="left"))
+
+        if DEBUG:
+            break
+    print(f"Loaded {sum(len(task_list) for task_list in task_data.values())} trajectories from {len(task_data)} tasks")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/racer_loader.py b/robometer/dataset_upload/dataset_loaders/racer_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d7d06e885700d9db74a422a0f362d4113e93885
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/racer_loader.py
@@ -0,0 +1,189 @@
+import json
+import os
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+from dataset_upload.helpers import generate_unique_id
+
+
+CAMERA_DIR_CANDIDATES = [
+    "front_rgb",
+    "left_shoulder_rgb",
+    "right_shoulder_rgb",
+    # "right_shoudler_rgb",  # sometimes misspelled in datasets
+    # "wrist_rgb",
+]
+
+
+class RacerFrameListLoader:
+    """Pickle-able loader that reads a list of image paths on demand (RGB, uint8)."""
+
+    def __init__(self, image_paths: list[str]) -> None:
+        if not image_paths:
+            raise ValueError("image_paths must be non-empty")
+        self.image_paths = image_paths
+
+    def __call__(self) -> np.ndarray:
+        frames: list[np.ndarray] = []
+        for p in self.image_paths:
+            img_bgr = cv2.imread(p, cv2.IMREAD_COLOR)
+            if img_bgr is None:
+                continue
+            img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+            frames.append(img_rgb)
+        if not frames:
+            return np.empty((0, 0, 0, 3), dtype=np.uint8)
+        frames_np = np.asarray(frames, dtype=np.uint8)
+        return frames_np
+
+
+def _sorted_pngs(dir_path: Path) -> list[str]:
+    paths = [p for p in dir_path.glob("*.png")]
+    paths.sort(key=lambda x: int(x.stem.split("_")[0]))
+    return [str(p) for p in paths]
+
+
+def _make_traj(image_paths: list[str], task_text: str, is_success: bool) -> dict[str, Any]:
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    traj["task"] = task_text
+    traj["frames"] = RacerFrameListLoader(image_paths)
+    traj["is_robot"] = True
+    traj["quality_label"] = "successful" if is_success else "failure"
+    traj["data_source"] = "racer"
+    traj["preference_group_id"] = None
+    traj["preference_rank"] = None
+    return traj
+
+
+def _collect_camera_views(sample_dir: Path) -> dict[str, list[str]]:
+    views: dict[str, list[str]] = {}
+    for cam in CAMERA_DIR_CANDIDATES:
+        d = sample_dir / cam
+        if d.exists() and d.is_dir():
+            imgs = _sorted_pngs(d)
+            if imgs:
+                views[cam] = imgs
+    return views
+
+
+def load_racer_dataset(dataset_path: str, dataset_name: str) -> dict[str, list[dict]]:
+    """Load RACER-augmented_rlbench dataset.
+
+    Args:
+        dataset_path: Path to dataset root containing 'train' and/or 'val' folders.
+        dataset_name: Use to pick split: 'racer_train' -> train, 'racer_val' -> val.
+
+    Behavior:
+        - Uses task_goal from language_description.json as language instruction.
+        - Creates success trajectories (full expert episode) per camera view.
+        - For each expert subgoal frame that contains heuristic failures in 'augmentation',
+          creates failure trajectories up to that expert frame index (inclusive), per camera view.
+
+    Returns:
+        Mapping: task_goal -> list of trajectory dicts.
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"RACER dataset path not found: {root}")
+
+    split = "val" if ("val" in dataset_name.lower()) else "train"
+
+    # Some distributions include an extra 'samples' segment
+    split_dir = root / split
+    alt_split_dir = split_dir / "samples"
+    if alt_split_dir.exists():
+        split_dir = alt_split_dir
+
+    if not split_dir.exists():
+        raise FileNotFoundError(f"Split directory not found: {split_dir}")
+
+    # Tasks are subdirectories under split_dir
+    task_dirs = [p for p in split_dir.iterdir() if p.is_dir()]
+
+    task_data: dict[str, list[dict]] = {}
+
+    for task_dir in task_dirs:
+        # Episodes are numeric directories under each task
+        episode_dirs = [p for p in task_dir.iterdir() if p.is_dir()]
+        for ep_dir in episode_dirs:
+            json_path = ep_dir / "language_description.json"
+            if not json_path.exists():
+                continue
+
+            try:
+                with open(json_path, "r") as f:
+                    desc = json.load(f)
+            except Exception:
+                continue
+
+            task_goal: str = desc.get("task_goal", "").strip() or task_dir.name
+            subgoal_dict: dict[str, Any] = desc.get("subgoal", {}) or {}
+
+            # Gather camera views for this episode once
+            views = _collect_camera_views(ep_dir)
+            if not views:
+                continue
+
+            # Success: use full length per view
+            for cam, img_list in views.items():
+                if not img_list:
+                    continue
+                expert_img_list = [p for p in img_list if "expert" in p]
+                traj = _make_traj(expert_img_list, task_goal, is_success=True)
+                task_data.setdefault(task_goal, []).append(traj)
+
+            # Failures: for each expert key that contains heuristic augmentations
+            for key, content in subgoal_dict.items():
+                # Expect keys like '0_expert', '48_expert', ...
+                if not isinstance(key, str) or "expert" not in key:
+                    continue
+                try:
+                    expert_frame_idx = int(key.split("_")[0])
+                except Exception:
+                    continue
+
+                aug = content.get("augmentation", {}) if isinstance(content, dict) else {}
+                if not isinstance(aug, dict) or not aug:
+                    continue
+
+                # Enumerate augmentations; select those labeled as heuristic failures
+                has_failure = False
+                for aug_image_name, aug_content in aug.items():
+                    if not isinstance(aug_content, dict):
+                        continue
+                    label = str(aug_content.get("label", "")).lower()
+                    if "failure" in label:  # e.g., 'recoverable_failure'
+                        has_failure = True
+                        break
+
+                if not has_failure:
+                    continue
+
+                # Build failure trajectories by truncating expert frames up to expert_frame_idx
+                for cam, img_list in views.items():
+                    if not img_list:
+                        continue
+
+                    # Find frames with numeric names and truncate accordingly
+                    def _frame_num(p: str) -> int:
+                        try:
+                            return int(Path(p).stem.split("_")[0])
+                        except Exception:
+                            return 1_000_000_000
+
+                    # Keep frames with index < expert_frame_idx
+                    subset = [p for p in img_list if _frame_num(p) < expert_frame_idx and "expert" in p]
+                    # add the augmented failure frame
+                    for img_name in img_list:
+                        if aug_image_name in img_name:
+                            subset.append(img_name)
+                            break
+                    if not subset:
+                        continue
+                    traj = _make_traj(subset, task_goal, is_success=False)
+                    task_data.setdefault(task_goal, []).append(traj)
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/roboarena_loader.py b/robometer/dataset_upload/dataset_loaders/roboarena_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..23bdb834eed8a440cb314f5664f945d03d4585fb
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/roboarena_loader.py
@@ -0,0 +1,123 @@
+import os
+from collections import defaultdict
+
+import cv2
+import numpy as np
+import yaml
+from dataset_upload.helpers import generate_unique_id
+
+trajectory_info_template = {
+    "id": [],
+    "task": [],
+    # "lang_vector": [],
+    "data_source": None,
+    "frames": None,
+    "is_robot": None,
+    "quality_label": None,
+    "partial_success": None,  # in [0, 1]
+}
+
+
+class RoboarenaFrameloader:
+    """Pickle-able loader that reads Roboarena frames from disk on demand.
+
+    Stores only simple fields so it can be safely passed across processes.
+    """
+
+    def __init__(self, video_path: str) -> None:
+        self.video_path = video_path
+
+    def __call__(self) -> np.ndarray:
+        """Load frames from disk when called.
+
+        Returns:
+            np.ndarray of shape (T, H, W, 3), dtype uint8
+        """
+        cap = cv2.VideoCapture(self.video_path)
+        frames = []
+        while True:
+            ret, frame = cap.read()
+
+            if not ret:
+                break
+
+            frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
+        cap.release()
+
+        frames = np.array(frames)
+
+        # Ensure shape and dtype sanity
+        if not isinstance(frames, np.ndarray) or frames.ndim != 4 or frames.shape[-1] != 3:
+            raise ValueError(
+                f"Unexpected frames shape for {self.video_path} in {self.video_path}: {getattr(frames, 'shape', None)}"
+            )
+
+        # Ensure uint8
+        if frames.dtype != np.uint8:
+            frames = frames.astype(np.uint8, copy=False)
+
+        return frames
+
+
+def create_new_trajectory(video_path: str, partial_success: int, task_name: str) -> dict:
+    trajectory_info = {}
+    trajectory_info["id"] = generate_unique_id()
+    trajectory_info["task"] = task_name
+    trajectory_info["frames"] = RoboarenaFrameloader(video_path)
+    trajectory_info["is_robot"] = True
+    trajectory_info["quality_label"] = "successful" if partial_success == 1.0 else "failure"
+    trajectory_info["partial_success"] = partial_success
+    trajectory_info["data_source"] = "roboarena"
+    return trajectory_info
+
+
+def load_roboarena_dataset(dataset_path: str) -> dict[str, list[dict]]:
+    eval_folder = os.path.join(dataset_path, "evaluation_sessions")
+    eval_sessions = os.listdir(eval_folder)
+
+    # tasks_to_videos = dict()
+    # task : {video_path: ..., partial_success}
+    task_data = defaultdict(list)
+
+    for eval_session in eval_sessions:
+        eval_session_path = os.path.join(eval_folder, eval_session)
+        metadata_path = os.path.join(eval_session_path, "metadata.yaml")
+        # load metadata
+        with open(metadata_path) as f:
+            metadata = yaml.load(f, Loader=yaml.FullLoader)
+        task = metadata["language_instruction"]
+        # if task in tasks_to_videos:
+        #    print(f"Task {task} already in tasks")
+        # tasks_to_videos[task] = []
+        # pprint.pprint(metadata)
+        policies = metadata["policies"]
+        for policy_id, policy_info in policies.items():
+            # get the partial success
+            partial_success = policy_info["partial_success"]
+            policy_name = policy_info["policy_name"]
+            policy_folder_name = f"{policy_id}_{policy_name}"
+            # get the videos of _left and _right
+            policy_folder_path = os.path.join(eval_session_path, policy_folder_name)
+            files_in_policy_folder = os.listdir(policy_folder_path)
+            video_left = [f for f in files_in_policy_folder if f.endswith("_left.mp4")]
+            video_right = [f for f in files_in_policy_folder if f.endswith("_right.mp4")]
+            # video_wrist = [f for f in files_in_policy_folder if f.endswith("_wrist.mp4")]
+            if len(video_left) > 0:
+                video_path = os.path.join(policy_folder_path, video_left[0])
+                task_data[task].append(
+                    create_new_trajectory(video_path, partial_success=partial_success, task_name=task)
+                )
+            if len(video_right) > 0:
+                video_path = os.path.join(policy_folder_path, video_right[0])
+                task_data[task].append(
+                    create_new_trajectory(video_path, partial_success=partial_success, task_name=task)
+                )
+            # if len(video_wrist) > 0:
+            #    video_path = os.path.join(policy_folder_path, video_wrist[0])
+            #    task_data[task].append(
+            #        create_new_trajectory(video_path, partial_success=partial_success, task_name=task)
+            #    )
+    print(
+        f"Loaded {sum([len(task_list) for task_list in task_data.values()])} trajectories from {len(task_data)} tasks"
+    )
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/robofac_loader.py b/robometer/dataset_upload/dataset_loaders/robofac_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..dc543be8553e7c40c4e91f6874f14ef97e803ba2
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/robofac_loader.py
@@ -0,0 +1,452 @@
+#!/usr/bin/env python3
+"""
+RoboFAC dataset loader for the generic dataset converter for Robometer model training.
+Loads MINT-SJTU/RoboFAC-dataset structure: realworld_data/<task>/videos/*.mp4 and simulation_data.
+
+Task descriptions from the dataset card: https://huggingface.co/datasets/MINT-SJTU/RoboFAC-dataset
+"""
+
+import json
+import re
+from pathlib import Path
+
+from dataset_upload.helpers import generate_unique_id, load_sentence_transformer_model
+from dataset_upload.video_helpers import load_video_frames
+from tqdm import tqdm
+
+# Official task name -> language description (from MINT-SJTU/RoboFAC-dataset dataset card).
+# Task names match the card exactly (including typos: UprightStask, PegInsetionSide).
+TASK_NAME_TO_DESCRIPTION: dict[str, str] = {
+    "SpinStack": "Pick up the cube on the spinning disc and stack it on another cube on the disc.",
+    "SpinPullStack": "Pull out the cube on the spinning disc and stack it on another cube on the disc.",
+    "MicrowaveTask": "Put the spoon on the table into the cup. Open the door of microwave, put the cup into the microwave and close the door.",
+    "SafeTask": "Put the gold bar into the safe, close the door of the safe and rotate the cross knob on the door to lock it.",
+    "ToolsTask": "Choose the correct (L-shaped) tools, grasp it to pull the correct (2-pins) charger and plug it.",
+    "UprightStask": "Upright the peg and stack it on the cube.",
+    "PegInsetionSide": "Insert the peg into the hole on the side of the block.",
+    "PullCubeTool": "Grasp the L-shaped tool and pull the cube by it.",
+    "PlugCharger": "Grasp the charger and plug it into the receptacle.",
+    "InsertCylinder": "Upright the cylinder and insert it into the middle hole on the shelf.",
+    "PlaceCube": "Pick up the cube and place it into the box.",
+    "LiftPegUpright": "Lift the peg and upright it.",
+    "PickCube": "Pick the cube to the target position.",
+    "PullCube": "Pull the cube to the red and white target.",
+    "PushCube": "Push the cube to the red and white target.",
+    "StackCube": "Pick up the cube and stack it on another cube.",
+}
+
+# Simulation path task names (e.g. UprightStack-v1) may differ from card names; map to card key.
+_SIMULATION_PATH_TO_TASK_KEY: dict[str, str] = {
+    "UprightStack": "UprightStask",
+    "PegInsertionSide": "PegInsetionSide",
+}
+
+
+class RoboFACFrameLoader:
+    """Pickle-able loader that reads RoboFAC video files on demand."""
+
+    def __init__(self, file_path: str):
+        self.file_path = file_path
+
+    def __call__(self):
+        """Load frames from video file. Returns np.ndarray (T, H, W, 3) uint8."""
+        return load_video_frames(Path(self.file_path))
+
+
+def _snake_to_camel(snake: str) -> str:
+    """Convert snake_case to CamelCase (e.g. insert_cylinder -> InsertCylinder)."""
+    return "".join(word.capitalize() for word in snake.split("_") if word)
+
+
+def _realworld_folder_to_task_description(folder_name: str) -> str:
+    """Convert realworld folder name to task description using dataset-card mapping.
+
+    E.g. so100_insert_cylinder_error -> InsertCylinder -> full description.
+    Falls back to title-cased name if not in TASK_NAME_TO_DESCRIPTION.
+    """
+    name = folder_name.replace("so100_", "").replace("_error", "").strip("_")
+    task_key = _snake_to_camel(name)
+    return TASK_NAME_TO_DESCRIPTION.get(task_key) or name.replace("_", " ").strip().title()
+
+
+def _find_mp4_under(path: Path) -> list[Path]:
+    """Find all .mp4 files under path (recursive). Handles videos/ or videos/chunk-000/ etc."""
+    if not path.exists():
+        return []
+    return sorted(path.rglob("*.mp4"))
+
+
+def _simulation_quality_from_folder(folder_name: str) -> str:
+    """Map simulation_data subfolder to quality_label (successful / failure)."""
+    name_lower = folder_name.lower()
+    if "success" in name_lower and "fail" not in name_lower:
+        return "successful"
+    if "fail" in name_lower or "error" in name_lower:
+        return "failure"
+    return "failure"  # default for unknown
+
+
+def _simulation_path_task_to_description(path_task_name: str) -> str:
+    """Map simulation path task name to dataset-card description.
+
+    success_data/ and failure_data/ have one subfolder per task variant, e.g.:
+    UprightStack-v1, LiftPegUpright-box, PickCube-apple, MicrowaveTask-fork.
+    Strip the trailing -<suffix> to get the base task name for lookup.
+    """
+    # Strip variant suffix: -v1, -box, -apple, -fork, -gen1, etc.
+    base = re.sub(r"-[a-zA-Z0-9_]+$", "", path_task_name)
+    task_key = _SIMULATION_PATH_TO_TASK_KEY.get(base, base)
+    return TASK_NAME_TO_DESCRIPTION.get(task_key) or path_task_name.replace("-", " ").replace("_", " ").strip().title()
+
+
+TASK_IDENTIFICATION_KEY = "Task identification"
+
+
+def _extract_task_identification_desc(annos: dict) -> str:
+    """Extract assistant 'value' from annos['Task identification'] conversation only.
+
+    We use only the "Task identification" section (not other anno keys). The value is
+    the assistant's reply in that conversation, e.g. "Insert the cylinder into the middle hole of the shelf."
+    """
+    # Prefer exact key, then case-insensitive match for "Task identification"
+    task_id = annos.get(TASK_IDENTIFICATION_KEY)
+    if task_id is None:
+        for k, v in annos.items():
+            if k.strip().lower() == TASK_IDENTIFICATION_KEY.lower():
+                task_id = v
+                break
+    if not isinstance(task_id, list):
+        return ""
+    for turn in task_id:
+        if turn.get("from") == "assistant" and "value" in turn:
+            return turn["value"].strip()
+    return ""
+
+
+def _load_test_qa_annos(root: Path) -> tuple[dict[str, str], dict[str, str]]:
+    """Load test_qa_realworld and test_qa_sim annos_per_video_split*.json and build video_id -> description.
+
+    Each JSON is a dict: { "video_id": { "video": "path", "task": "InsertCylinder", "annos": { "Task identification": [ { "from": "human", "value": "..." }, { "from": "assistant", "value": "Insert the cylinder into the middle hole of the shelf." } ] } } }.
+    We use the "Task identification" assistant value as the language description.
+    Tries root/test_qa_realworld, root/test_qa_sim, and root/main/... for each.
+    """
+    video_id_to_desc: dict[str, str] = {}
+    task_folder_to_desc: dict[str, str] = {}
+    dirs_to_try = [
+        root / "test_qa_realworld",
+        root / "test_qa_sim",
+        root / "main" / "test_qa_realworld",
+        root / "main" / "test_qa_sim",
+    ]
+    for annos_dir in dirs_to_try:
+        if not annos_dir.is_dir():
+            continue
+        for path in sorted(annos_dir.glob("annos_per_video_split*.json")):
+            with open(path, encoding="utf-8") as f:
+                data = json.load(f)
+            for video_id, entry in data.items():
+                if not isinstance(entry, dict):
+                    continue
+                annos = entry.get("annos", {})
+                desc = _extract_task_identification_desc(annos)
+                if not desc:
+                    continue
+                video_id_to_desc[video_id] = desc
+                video_path_str = entry.get("video", "")
+                if video_path_str:
+                    stem = Path(video_path_str).stem
+                    video_id_to_desc[stem] = desc
+                    task_folder_to_desc[video_path_str.split("/")[0]] = desc
+                task = entry.get("task", "")
+                if task:
+                    task_folder_to_desc[task] = desc
+    if video_id_to_desc:
+        print(f"  Loaded {len(video_id_to_desc)} video_id->description from test_qa_* annos")
+    
+    return video_id_to_desc, task_folder_to_desc
+
+
+# Human prompt that indicates task-identification QA in training_qa.json (we only extract from these).
+_TRAINING_QA_TASK_IDENTIFICATION_PROMPT = "identify what task the robot is doing"
+
+
+def _extract_task_desc_from_training_qa_conversations(convs: list) -> str:
+    """Extract language description from training_qa conversations.
+
+    Only use the assistant reply when the human asked the task-identification question
+    (e.g. 'Can you identify what task the robot is doing in the provided video?').
+    """
+    if not isinstance(convs, list):
+        return ""
+    for i, turn in enumerate(convs):
+        if turn.get("from") != "human" or "value" not in turn:
+            continue
+        human_val = (turn.get("value") or "").lower()
+        if _TRAINING_QA_TASK_IDENTIFICATION_PROMPT not in human_val:
+            continue
+        # Next turn should be assistant with the task description
+        if i + 1 < len(convs) and convs[i + 1].get("from") == "assistant" and "value" in convs[i + 1]:
+            return (convs[i + 1].get("value") or "").strip()
+    return ""
+
+
+def _load_training_qa(root: Path) -> tuple[dict[str, str], dict[str, str]]:
+    """Load training_qa.json and build video_id -> description and task_folder -> description.
+
+    training_qa.json is a list of {
+        "id": "09936392-adb5-4f34-9410-7c7305d9c76b",
+        "video": "dataset_success_cleaned/MicrowaveTask-fork/stack_error/09936392-adb5-4f34-9410-7c7305d9c76b.mp4",
+        "conversations": [
+            { "from": "human", "value": "<video>\\nCan you identify what task the robot is doing in the provided video?" },
+            { "from": "assistant", "value": "Put the fork in the cup and put them in the microwave" }
+        ]
+    }.
+    We only extract the assistant reply when the human asked the task-identification question
+    (contains 'identify what task the robot is doing'). That reply is the language description.
+    Tries root/training_qa.json and root/main/training_qa.json.
+    """
+    video_id_to_desc: dict[str, str] = {}
+    task_folder_to_desc: dict[str, str] = {}
+    for candidate in (root / "training_qa.json", root / "main" / "training_qa.json"):
+        if not candidate.exists():
+            continue
+        with open(candidate, encoding="utf-8") as f:
+            data = json.load(f)
+        for item in data:
+            video = item.get("video", "")
+            video_id = item.get("id") or (Path(video).stem if video else "")
+            if not video_id:
+                continue
+            convs = item.get("conversations", [])
+            desc = _extract_task_desc_from_training_qa_conversations(convs)
+            if not desc:
+                continue
+            video_id_to_desc[video_id] = desc
+            if video:
+                task_folder_to_desc[video.split("/")[0]] = desc
+        if video_id_to_desc:
+            print(f"  Loaded {len(video_id_to_desc)} video_id->description from {candidate}")
+            return video_id_to_desc, task_folder_to_desc
+            
+    return video_id_to_desc, task_folder_to_desc
+
+
+def _get_realworld_task_description(
+    folder_name: str,
+    task_folder_to_desc: dict[str, str] | None,
+) -> str:
+    """Get language description for realworld folder: prefer training_qa (match by base task name), else TASK_NAME_TO_DESCRIPTION. ipdb if no match."""
+    name = folder_name.replace("so100_", "").replace("_error", "").strip("_")
+    task_key = _snake_to_camel(name)
+    if task_folder_to_desc:
+        for qa_key, qa_desc in task_folder_to_desc.items():
+            base = re.sub(r"-[a-zA-Z0-9_]+$", "", qa_key)
+            card_key = _SIMULATION_PATH_TO_TASK_KEY.get(base, base)
+            if card_key == task_key:
+                return qa_desc
+    task_desc = _realworld_folder_to_task_description(folder_name)
+    if task_key not in TASK_NAME_TO_DESCRIPTION:
+        import ipdb
+        ipdb.set_trace()
+    return task_desc
+
+
+def _get_simulation_task_description(
+    path_task_name: str,
+    task_folder_to_desc: dict[str, str] | None,
+) -> str:
+    """Get language description for simulation task: prefer task_folder from training_qa, else TASK_NAME_TO_DESCRIPTION. ipdb if no match."""
+    if task_folder_to_desc and path_task_name in task_folder_to_desc:
+        return task_folder_to_desc[path_task_name]
+    desc = _simulation_path_task_to_description(path_task_name)
+    base = re.sub(r"-[a-zA-Z0-9_]+$", "", path_task_name)
+    task_key = _SIMULATION_PATH_TO_TASK_KEY.get(base, base)
+    if task_key not in TASK_NAME_TO_DESCRIPTION:
+        import ipdb
+        ipdb.set_trace()
+    return desc
+
+
+def _parse_simulation_video_path(
+    video_path: Path,
+    root: Path,
+    video_id_to_desc: dict[str, str] | None = None,
+    task_folder_to_desc: dict[str, str] | None = None,
+) -> tuple[str, str, str]:
+    """From a video under simulation_data/, derive (task_description, quality_label, data_source).
+
+    Supports two directory layouts:
+    - simulation_data/success_data|failure_data/<task_folder>/.../<video>.mp4
+      -> quality_folder = success_data|failure_data, path_task_name = task_folder (parts[2])
+    - simulation_data/<task_folder>/view0|.../<video>.mp4  (no success_data/failure_data)
+      -> path_task_name = task_folder (parts[1]), quality_folder = simulation_data
+    """
+    try:
+        rel = video_path.relative_to(root)
+    except ValueError:
+        return "Simulation", "failure", "simulation_data"
+    parts = rel.parts
+    if len(parts) < 3:
+        return "Simulation", "failure", "simulation_data"
+    if parts[0] != "simulation_data":
+        return "Simulation", "failure", "simulation_data"
+    # If second component is success_data/failure_data, task folder is parts[2]; else it's parts[1] (e.g. LiftPegUpright-box/view0/...)
+    if parts[1] in ("success_data", "failure_data"):
+        quality_folder = parts[1]
+        path_task_name = parts[2]
+    else:
+        quality_folder = "simulation_data"
+        path_task_name = parts[1]
+    quality_label = _simulation_quality_from_folder(quality_folder)
+    # Prefer description by video id (stem of .mp4), then by task folder
+    video_id = video_path.stem
+    if video_id_to_desc and video_id in video_id_to_desc:
+        task_desc = video_id_to_desc[video_id]
+    else:
+        task_desc = _get_simulation_task_description(path_task_name, task_folder_to_desc)
+    data_source = f"simulation_data/{quality_folder}"
+    return task_desc, quality_label, data_source
+
+
+def _discover_robofac_trajectories(
+    dataset_path: Path,
+    *,
+    realworld: bool = True,
+    simulation: bool = True,
+    video_id_to_desc: dict[str, str] | None = None,
+    task_folder_to_desc: dict[str, str] | None = None,
+) -> list[tuple[Path, str, str, str]]:
+    """Discover all video files in RoboFAC dataset structure.
+
+    Expected structure (from MINT-SJTU/RoboFAC-dataset):
+        realworld_data/
+            so100_insert_cylinder_error/
+                videos/
+                    *.mp4   OR  videos/chunk-000/*.mp4  (recursive)
+            ...
+        simulation_data/
+            success_data/   or  failure_data/
+                <TaskFolder>/   one subfolder per task variant (e.g. UprightStack-v1, LiftPegUpright-box, PickCube-apple)
+                    (optional subdirs)
+                        *.mp4
+
+    If realworld_data is not found at dataset_path, also tries dataset_path / "main"
+    (some download methods put repo content under a main/ subfolder).
+
+    Returns:
+        List of (video_path, task, quality_label, data_source) for each trajectory.
+    """
+    out: list[tuple[Path, str, str, str]] = []
+
+    # Resolve root: support both /path/to/RoboFAC-dataset and /path/to/RoboFAC-dataset/main
+    root = dataset_path
+    realworld_path = root / "realworld_data"
+    if not realworld_path.exists() and (root / "main").is_dir():
+        root = root / "main"
+        realworld_path = root / "realworld_data"
+    if not realworld_path.exists():
+        print(f"Warning: realworld_data not found at {dataset_path} or {dataset_path}/main")
+    else:
+        if realworld:
+            for task_dir in sorted(realworld_path.iterdir()):
+                if not task_dir.is_dir() or task_dir.name.startswith("."):
+                    continue
+                task_name = task_dir.name
+                videos = _find_mp4_under(task_dir)
+                # Realworld videos have generic stems (e.g. episode_000000) that repeat across tasks;
+                # use task name only for description, not video_id lookup.
+                for vid in videos:
+                    task_desc = _get_realworld_task_description(task_name, task_folder_to_desc)
+                    out.append((vid, task_desc, "failure", f"realworld_data/{task_name}"))
+                if videos:
+                    print(f"  realworld_data/{task_name}: {len(videos)} videos")
+
+    if simulation:
+        sim_path = root / "simulation_data"
+        if sim_path.exists():
+            # Discover all mp4s under simulation_data and parse path for task + quality
+            videos = _find_mp4_under(sim_path)
+
+            for vid in videos:
+                task_desc, quality_label, data_source = _parse_simulation_video_path(
+                    vid, root,
+                    video_id_to_desc=video_id_to_desc,
+                    task_folder_to_desc=task_folder_to_desc,
+                )
+                out.append((vid, task_desc, quality_label, data_source))
+            if videos:
+                print(f"  simulation_data: {len(videos)} videos (task/quality from path)")
+        else:
+            print("Warning: simulation_data not found")
+
+    return out
+
+
+def load_robofac_dataset(
+    dataset_path: str,
+    max_trajectories: int | None = None,
+    realworld: bool = True,
+    simulation: bool = True,
+) -> dict[str, list[dict]]:
+    """Load RoboFAC dataset and organize by task.
+
+    Args:
+        dataset_path: Path to the RoboFAC dataset root (e.g. .../RoboFAC-dataset or .../RoboFAC-dataset/main).
+        max_trajectories: Maximum number of trajectories to load (None for all).
+        realworld: Include realworld_data subfolders.
+        simulation: Include simulation_data.
+
+    Returns:
+        Dictionary mapping task names to lists of trajectory dicts (frames, task, quality_label, etc.).
+    """
+    print("Loading RoboFAC dataset from:", dataset_path)
+    dataset_path = Path(dataset_path).resolve()
+    if not dataset_path.exists():
+        raise FileNotFoundError(f"RoboFAC dataset path not found: {dataset_path}")
+
+    # Build video_id -> description from test_qa_* annos (Task identification) and training_qa.json (human asks "identify what task the robot is doing")
+    video_id_to_desc, task_folder_to_desc = _load_training_qa(dataset_path)
+    test_vid, test_task = _load_test_qa_annos(dataset_path)
+    video_id_to_desc.update(test_vid)
+    task_folder_to_desc.update(test_task)
+
+    print("Discovering videos (realworld_data/ and simulation_data/)...")
+    traj_list = _discover_robofac_trajectories(
+        dataset_path,
+        realworld=realworld,
+        simulation=simulation,
+        video_id_to_desc=video_id_to_desc or None,
+        task_folder_to_desc=task_folder_to_desc or None,
+    )
+    if not traj_list:
+        raise FileNotFoundError(
+            f"No .mp4 videos found under {dataset_path}. "
+            "Check that the path points to the RoboFAC-dataset root (containing realworld_data/ and optionally simulation_data/). "
+            "If you downloaded with Hugging Face CLI, the root may be under a 'main' subfolder; the loader will try that automatically."
+        )
+    if max_trajectories is not None and max_trajectories != -1:
+        traj_list = traj_list[:max_trajectories]
+
+    print(f"Found {len(traj_list)} trajectory videos total")
+
+    task_data: dict[str, list[dict]] = {}
+    for video_path, task_desc, quality_label, data_source in tqdm(
+        traj_list, desc="Building RoboFAC trajectories"
+    ):
+        frame_loader = RoboFACFrameLoader(str(video_path))
+        partial = 1.0 if quality_label == "successful" else 0.0
+        trajectory = {
+            "frames": frame_loader,
+            "actions": None,
+            "is_robot": True,
+            "task": task_desc,
+            "quality_label": quality_label,
+            "data_source": data_source,
+            "partial_success": partial,
+            "id": generate_unique_id(),
+        }
+        task_data.setdefault(task_desc, []).append(trajectory)
+
+    total = sum(len(v) for v in task_data.values())
+    print(f"Loaded {total} trajectories from {len(task_data)} tasks")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/robofail_loader.py b/robometer/dataset_upload/dataset_loaders/robofail_loader.py
new file mode 100755
index 0000000000000000000000000000000000000000..633a62712615f923030b21bfa4589827f49feb1c
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/robofail_loader.py
@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+"""
+RoboFail dataset loader for the generic dataset converter for Robometer model training.
+This module contains RoboFail-specific logic for loading and processing data files.
+"""
+
+from pathlib import Path
+
+import numpy as np
+from dataset_upload.helpers import generate_unique_id
+from robometer.data.video_helpers import load_video_frames
+from tqdm import tqdm
+
+
+class RoboFailFrameLoader:
+    """Pickle-able loader that reads RoboFail frames from files on demand.
+
+    Stores only simple fields so it can be safely passed across processes.
+    Supports both HDF5 and video file formats.
+    """
+
+    def __init__(self, file_path: str):
+        self.file_path = file_path
+
+    def __call__(self) -> np.ndarray:
+        """Load frames from file when called.
+
+        Returns:
+            np.ndarray of shape (T, H, W, 3), dtype uint8
+        """
+        return load_video_frames(Path(self.file_path))
+
+
+# Task mapping from folder names to task descriptions
+FOLDER_TO_TASK_NAME = {
+    "putAppleBowl1": "put apple in bowl",
+    "putAppleBowl2": "put apple in bowl",
+    "putAppleBowl3": "put apple in bowl",
+    "boilWater1": "boil water in a pot",
+    "boilWater2": "boil water in a pot",
+    "boilWater3": "boil water in a pot",
+    "boilWater4": "boil water in a pot",
+    "cutCarrot1": "cut carrot",
+    "cutCarrot2": "cut carrot",
+    "putPearDrawer1": "put pear in drawer",
+    "putPearDrawer2": "put pear in drawer",
+    "putPearDrawer3": "put pear in drawer",
+    "sauteeCarrot1": "cook carrot",
+    "sauteeCarrot2": "cook carrot",
+    "sauteeCarrot3": "cook carrot",
+    "sauteeCarrot4": "cook carrot",
+    "secureObjects1": "secure objects",
+    "putFruitsBowl1": "put all visible fruits in bowl",
+    "putFruitsBowl2": "put all fruits in bowl",
+    "makeCoffee1": "make coffee",
+    "makeCoffee2": "make coffee",
+    "makeCoffee3": "make coffee",
+    "heatPot1": "pre-heat pot",
+    "heatPot2": "pre-heat pot",
+    "appleInFridge1": "store apple in a bowl and put in a fridge",
+    "appleInFridge2": "store apple in a bowl and put in a fridge",
+    "appleInFridge3": "store apple in a bowl and put in a fridge",
+    "appleInFridge4": "store apple in a bowl and put in a fridge",
+    "heatPotato1": "serve a bowl of potato on table that was heated using a microwave",
+    "heatPotato2": "serve a bowl of potato on table that was heated using a microwave",
+}
+
+
+def _get_task_name_from_folder(folder_name: str) -> str:
+    """Convert folder name to task name using the mapping."""
+    # First try to find exact match
+    if folder_name in FOLDER_TO_TASK_NAME:
+        return FOLDER_TO_TASK_NAME[folder_name]
+
+    # If no exact match, try partial matching
+    for folder_key, task_name in FOLDER_TO_TASK_NAME.items():
+        if folder_key in folder_name or folder_name in folder_key:
+            return task_name
+
+    # If no mapping found, convert folder name to readable task
+    task = folder_name.replace("_", " ").replace("-", " ")
+    return task.strip()
+
+
+def _discover_robofail_files(dataset_path: Path) -> list[tuple[Path, str]]:
+    """Discover all video files in the RoboFail dataset structure.
+
+    Expected structure:
+    robofail/real_data/
+        folder_name_1/
+            videos/
+                color.mp4
+        folder_name_2/
+            videos/
+                color.mp4
+        ...
+
+    Returns:
+        List of tuples: (video_file_path, task_name)
+    """
+    trajectory_files: list[tuple[Path, str]] = []
+
+    # Look for folders under robofail/real_data
+    real_data_path = dataset_path / "real_data"
+    if not real_data_path.exists():
+        print(f"Warning: real_data directory not found at {real_data_path}")
+        return trajectory_files
+
+    for folder in real_data_path.iterdir():
+        if not folder.is_dir():
+            continue
+
+        folder_name = folder.name
+        task_name = _get_task_name_from_folder(folder_name)
+
+        # Look for videos/color.mp4
+        video_path = folder / "videos" / "color.mp4"
+        if video_path.exists():
+            trajectory_files.append((video_path, task_name))
+        else:
+            print(f"Warning: No color.mp4 found in {folder}/videos/")
+
+    return trajectory_files
+
+
+def load_robofail_dataset(dataset_path: str, max_trajectories: int | None = None) -> dict[str, list[dict]]:
+    """Load RoboFail dataset and organize by task.
+
+    Args:
+        dataset_path: Path to the RoboFail dataset directory (should contain real_data/)
+        max_trajectories: Maximum number of trajectories to load (None for all)
+
+    Returns:
+        Dictionary mapping task names to lists of trajectory dictionaries
+    """
+
+    print(f"Loading RoboFail dataset from: {dataset_path}")
+    print("=" * 100)
+    print("LOADING ROBOFAIL DATASET")
+    print("=" * 100)
+
+    dataset_path = Path(dataset_path)
+    if not dataset_path.exists():
+        raise FileNotFoundError(f"RoboFail dataset path not found: {dataset_path}")
+
+    traj_files = _discover_robofail_files(dataset_path)
+    print(f"Found {len(traj_files)} trajectory files")
+
+    task_data: dict[str, list[dict]] = {}
+    loaded_count = 0
+
+    for video_file, task_name in tqdm(traj_files, desc="Processing RoboFail trajectories"):
+        if max_trajectories is not None and loaded_count >= max_trajectories and max_trajectories != -1:
+            break
+
+        try:
+            # Set up frame loader for video file
+            frame_loader = RoboFailFrameLoader(
+                file_path=str(video_file),
+            )
+
+            # For now, we don't have action data from the video files
+            # This could be extended if action data is stored elsewhere
+            actions = None
+
+            # Use task name from folder mapping
+            task_description = task_name
+
+            # Create trajectory info
+            trajectory = {
+                "frames": frame_loader,
+                "actions": actions,
+                "is_robot": True,  # RoboFail is typically robot data
+                "quality_label": "failure",  # Default assumption
+                "preference_group_id": None,
+                "preference_rank": None,
+                "task": task_description,
+                "id": generate_unique_id(),
+            }
+
+            task_data.setdefault(task_name, []).append(trajectory)
+            loaded_count += 1
+
+        except Exception as e:
+            print(f"Error loading trajectory {video_file}: {e}")
+            continue
+
+    total_trajectories = sum(len(v) for v in task_data.values())
+    print(f"Loaded {total_trajectories} trajectories from {len(task_data)} tasks")
+
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/roboreward_loader.py b/robometer/dataset_upload/dataset_loaders/roboreward_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..dca111bd82b323786b6e764bee82811d0385eadc
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/roboreward_loader.py
@@ -0,0 +1,170 @@
+"""
+Loader for RoboReward dataset - a dataset for training vision-language reward models for robotics.
+
+Paper: https://arxiv.org/abs/2601.00675
+Dataset: https://huggingface.co/datasets/teetone/RoboReward
+"""
+
+import json
+import os
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+from dataset_upload.helpers import generate_unique_id
+
+
+class RoboRewardVideoLoader:
+    """Pickle-able loader that reads frames from an existing MP4 video file."""
+
+    def __init__(self, video_path: str) -> None:
+        self.video_path = video_path
+
+    def __call__(self) -> np.ndarray:
+        """Load all frames from the video file."""
+        cap = cv2.VideoCapture(self.video_path)
+        frames = []
+
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+
+        cap.release()
+
+        if not frames:
+            return np.empty((0, 0, 0, 3), dtype=np.uint8)
+
+        return np.asarray(frames, dtype=np.uint8)
+
+
+def _reward_to_partial_success(reward: int) -> float:
+    """Convert RoboReward score (1-5) to partial_success (0.0-1.0).
+
+    Reward scale:
+        1: No success -> 0.0
+        2: Minimal progress -> 0.25
+        3: Partial completion -> 0.5
+        4: Near completion -> 0.75
+        5: Perfect completion -> 1.0
+    """
+    return (reward - 1) / 4.0
+
+
+def _make_traj(video_path: str, task: str, reward: int, dataset_name: str) -> dict[str, Any]:
+    """Create a trajectory dictionary from RoboReward metadata."""
+    partial_success = _reward_to_partial_success(reward)
+
+    traj: dict[str, Any] = {}
+    traj["id"] = generate_unique_id()
+    traj["task"] = task
+    traj["frames"] = RoboRewardVideoLoader(video_path)  # Lazy loader for existing MP4
+    traj["is_robot"] = True
+    traj["quality_label"] = "successful" if partial_success == 1.0 else "failure"
+    traj["partial_success"] = partial_success
+    traj["data_source"] = f"roboreward_{dataset_name}"
+    traj["preference_group_id"] = None
+    traj["preference_rank"] = None
+    return traj
+
+
+def load_roboreward_dataset(dataset_path: str, dataset_name: str) -> dict[str, list[dict]]:
+    """Load RoboReward dataset from local folders.
+
+    Args:
+        dataset_path: Root directory containing train/, val/, test/ folders
+        dataset_name: Dataset name to determine split (e.g., 'roboreward_train', 'roboreward_val', 'roboreward_test')
+
+    Structure:
+        dataset_path/
+            train/
+                metadata.jsonl
+                [video folders with MP4s]
+            val/
+                metadata.jsonl
+                [video folders with MP4s]
+            test/
+                metadata.jsonl
+                [video folders with MP4s]
+
+    Returns:
+        Mapping: task instruction -> list of trajectory dicts
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"RoboReward dataset path not found: {root}")
+
+    # Determine split from dataset_name
+    if "train" in dataset_name.lower():
+        split = "train"
+    elif "val" in dataset_name.lower():
+        split = "val"
+    elif "test" in dataset_name.lower():
+        split = "test"
+    else:
+        raise ValueError(f"Dataset name must specify split (train/val/test): {dataset_name}")
+
+    split_dir = root / split
+    if not split_dir.exists():
+        raise FileNotFoundError(f"Split directory not found: {split_dir}")
+
+    metadata_file = split_dir / "metadata.jsonl"
+    if not metadata_file.exists():
+        raise FileNotFoundError(f"Metadata file not found: {metadata_file}")
+
+    task_data: dict[str, list[dict]] = {}
+
+    # Read metadata.jsonl
+    print(f"Loading RoboReward {split} split from {metadata_file}")
+    with open(metadata_file, "r") as f:
+        for line_idx, line in enumerate(f):
+            try:
+                entry = json.loads(line.strip())
+            except json.JSONDecodeError:
+                print(f"Warning: Could not parse line {line_idx} in metadata.jsonl")
+                continue
+
+            file_name = entry.get("file_name")
+            task = entry.get("task")
+            reward = entry.get("reward")
+
+            if not file_name or not task or reward is None:
+                print(f"Warning: Missing required fields in line {line_idx}")
+                continue
+
+            # Construct full video path
+            video_path = split_dir / file_name
+            if not video_path.exists():
+                print(f"Warning: Video file not found: {video_path}")
+                continue
+
+            dataset_name = file_name.split("/")[0]
+
+            if dataset_name == "robo_arena":
+                dataset_name = "roboarena"
+
+            # Create trajectory
+            traj = _make_traj(str(video_path), task, reward, dataset_name)
+            task_data.setdefault(task, []).append(traj)
+
+    print(f"Loaded {len(task_data)} unique tasks from RoboReward {split} split")
+
+    # Print reward distribution
+    all_trajs = [t for trajs in task_data.values() for t in trajs]
+    reward_counts = {i: 0 for i in range(1, 6)}
+    for traj in all_trajs:
+        # Reverse conversion to get original reward
+        reward = int(traj["partial_success"] * 4 + 1)
+        reward_counts[reward] += 1
+
+    print(f"Reward distribution:")
+    for reward, count in sorted(reward_counts.items()):
+        print(f"  Reward {reward}: {count} trajectories")
+    print(f"Total trajectories: {len(all_trajs)}")
+
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/soar_loader.py b/robometer/dataset_upload/dataset_loaders/soar_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..564aced863030bc15dc249844b7c4d4c1938d7b3
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/soar_loader.py
@@ -0,0 +1,235 @@
+import os
+import json
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from datasets import Dataset, concatenate_datasets
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+# We do not stream; assume RLDS TFDS builders are already downloaded locally.
+import tensorflow_datasets as tfds
+
+# soar_new_success_labels_path = "dataset_upload/dataset_helpers/soar_vlm_labels_checkpoint.json"
+soar_new_success_labels_path = "dataset_upload/dataset_helpers/soar_label_corrections_full.json"
+
+
+def _build_video_paths(output_dir: str, dataset_label: str, episode_idx: int, view_key: str) -> tuple[str, str]:
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"clip@{view_key}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_episode(args):
+    episode_steps, ep_idx, task, lang_vec, output_dir, dataset_label, max_frames, fps, img_key, quality_label = args
+
+    # Collect frames for the given image key
+    frames: list[np.ndarray] = []
+    for step in episode_steps:
+        obs = step.get("observation", {}) if isinstance(step, dict) else {}
+        if img_key not in obs:
+            continue
+        frame = obs[img_key]
+        if isinstance(frame, np.ndarray):
+            if frame.ndim == 3 and frame.shape[-1] in (1, 3, 4):
+                if frame.shape[-1] == 1:
+                    frame = np.repeat(frame, 3, axis=-1)
+                elif frame.shape[-1] == 4:
+                    frame = frame[..., :3]
+                if frame.dtype != np.uint8:
+                    frame = frame.astype(np.uint8, copy=False)
+                frames.append(frame)
+
+    if not frames:
+        return []
+
+    full_path, rel_path = _build_video_paths(output_dir, dataset_label, ep_idx, img_key)
+
+    traj_dict = {
+        "id": generate_unique_id(),
+        "frames": frames,
+        "task": task,
+        "is_robot": True,
+        "quality_label": quality_label,
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    entry = create_hf_trajectory(
+        traj_dict=traj_dict,
+        video_path=full_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_label,
+        use_video=True,
+        fps=fps,
+    )
+    if entry:
+        entry["frames"] = rel_path
+        return [entry]
+    return []
+
+
+def convert_soar_dataset_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert SOAR RLDS (local TFDS) to HF dataset. Non-streaming, local builders only.
+
+    Expects directory structure:
+      <dataset_path>/rlds/<split>/<version>/ (TFDS builder dir)
+    """
+
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"'rlds' directory not found under: {dataset_path}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    label_corrections = {}
+    if os.path.exists(soar_new_success_labels_path):
+        with open(soar_new_success_labels_path) as f:
+            data = json.load(f)
+        # Keys are strings in JSON, convert to int
+        label_corrections = {int(k): v for k, v in data.get("label_corrections", {}).items()}
+        print(f"Loaded label corrections for {len(label_corrections)} trajectories")
+
+    datasets_list: list[Dataset] = []
+
+    builder = tfds.builder_from_directory(root)
+    success_episode_instructions = set()  # to only upload failures that have a corresponding success episode
+    for split_name in ["success", "failure"]:
+        ds = builder.as_dataset(split=split_name, shuffle_files=False)
+
+        if split_name == "success":
+            with open(soar_new_success_labels_path, "r") as f:
+                # new_success_labels = json.load(f)["results"]
+                new_success_labels = json.load(f)["label_corrections"]
+                # episodes where qwen-3-vl predicted success
+                # new_success_labels = [
+                #    result["predicted_label"] for result in new_success_labels if result["original_label"] == "success"
+                # ]
+
+                # convert to int keys
+                new_success_labels = {int(k): v for k, v in new_success_labels.items()}
+
+        entries: list[dict] = []
+        produced = 0
+        max_limit = float("inf") if (max_trajectories is None or max_trajectories == -1) else int(max_trajectories)
+
+        for ep_idx, episode in enumerate(tqdm(ds, desc=f"SOAR {split_name} episodes")):
+            if split_name == "success":
+                if new_success_labels[ep_idx] != "successful":
+                    # disagree with qwen-3-vl's prediction, skip this episode
+                    continue
+            if produced >= max_limit:
+                break
+
+            # Convert to numpy steps list
+            try:
+                steps_np = list(tfds.as_numpy(episode["steps"]))
+            except Exception:
+                continue
+
+            # Extract language instruction from first step
+            task_text: str | None = None
+            first = steps_np[0] if steps_np else None
+            if first is not None:
+                # First try step-level keys
+                if "language_instruction" in first:
+                    val = first["language_instruction"]
+                    task_text = val.decode() if isinstance(val, (bytes, bytearray)) else str(val)
+
+            if not task_text:
+                continue
+            elif split_name == "failure":
+                if new_success_labels[ep_idx] != "failure":  # skip if the label is not correct
+                    continue
+                elif task_text not in success_episode_instructions:
+                    # no corresponding success episode, skip this failure
+                    print(f"No corresponding success episode for failure {ep_idx}, skipping")
+                    continue
+
+            if task_text not in lang_cache:
+                lang_cache[task_text] = lang_model.encode(task_text)
+            lang_vec = lang_cache[task_text]
+
+            # Choose a valid image key
+            valid_img_key: str | None = None
+            valid_img_key = "image_0"
+
+            # Determine quality label
+            quality_label = "successful" if split_name.lower().startswith("success") else "failure"
+
+            # Build entry for this view
+            episode_entries = _process_episode((
+                steps_np,
+                ep_idx,
+                task_text,
+                lang_vec,
+                output_dir,
+                dataset_name,
+                max_frames,
+                fps,
+                valid_img_key,
+                quality_label,
+            ))
+            entries.extend(episode_entries)
+            produced += len(episode_entries)
+
+        if not entries:
+            ds_out = Dataset.from_dict({
+                "id": [],
+                "task": [],
+                "lang_vector": [],
+                "data_source": [],
+                "frames": [],
+                "is_robot": [],
+                "quality_label": [],
+                "preference_group_id": [],
+                "preference_rank": [],
+            })
+        else:
+            ds_out = Dataset.from_list(entries)
+
+        datasets_list.append(ds_out)
+
+    if not datasets_list:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    if len(datasets_list) == 1:
+        return datasets_list[0]
+    return concatenate_datasets(datasets_list)
diff --git a/robometer/dataset_upload/dataset_loaders/usc_franka_policy_ranking_loader.py b/robometer/dataset_upload/dataset_loaders/usc_franka_policy_ranking_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..23cc2ebf0bad74a3c39a5a4c96684c9e4c536999
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/usc_franka_policy_ranking_loader.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+from collections import defaultdict
+from pathlib import Path
+from typing import Iterable
+
+import re
+import numpy as np
+import cv2
+from PIL import Image
+
+from dataset_upload.helpers import generate_unique_id
+
+TASK_DESCRIPTION_MAP = {
+    "plate_in_sink": "Put the yellow plate in the sink",
+    "fold_towel": "Fold the towel in half",
+    "red_cube_on_gray_cube": "Stack the red cube on the gray cube",
+    "put_banana_on_red_plate": "Put the banana on the red plate",
+}
+
+QUALITY_LABEL_MAP = {
+    "success": "successful",
+    "successful": "successful",
+    "subopt": "suboptimal",
+    "suboptimal": "suboptimal",
+    "fail": "failure",
+    "failure": "failure",
+}
+
+CAMERA_PATTERN = "camera_south_color*.png"
+
+
+def _extract_frame_number(path: str) -> int:
+    """Extract the numeric frame number from a filename like 'camera_south_color_1.png'."""
+    # Extract the number before .png
+    match = re.search(r"_(\d+)\.png$", path)
+    if match:
+        return int(match.group(1))
+    # Fallback: try to extract any number from the filename
+    match = re.search(r"(\d+)", Path(path).stem)
+    if match:
+        return int(match.group(1))
+    return 0  # Default to 0 if no number found
+
+
+class USCFrankaFrameLoader:
+    """Lazy loader that stitches RGB frames from PNG files."""
+
+    def __init__(self, image_paths: Iterable[str]) -> None:
+        image_paths = list(image_paths)
+        if not image_paths:
+            raise ValueError("USCFrankaFrameLoader requires at least one image path")
+        # Sort by frame number to ensure sequential ordering
+        self.image_paths = sorted(image_paths, key=_extract_frame_number)
+
+    def _load_frame(self, path: str) -> np.ndarray:
+        with Image.open(path) as img:
+            # Explicitly convert to RGB to ensure correct color format
+            img_rgb = img.convert("RGB")
+            frame = np.array(img_rgb, dtype=np.uint8)
+        if frame.ndim != 3:
+            raise ValueError(f"Expected 3D RGB frame, got shape {frame.shape} for {path}")
+        if frame.shape[2] != 3:
+            raise ValueError(f"Expected RGB frame with 3 channels, got {frame.shape[2]} channels for {path}")
+        # Ensure frame is in HWC format (height, width, channels) with RGB order
+        if frame.shape[0] == 3 and frame.shape[2] != 3:
+            # If channel-first, convert to channel-last
+            frame = np.transpose(frame, (1, 2, 0))
+        # Rotate 90 degrees clockwise
+        frame = np.rot90(frame, k=-1)
+        # Ensure the array is contiguous for efficient processing
+        if not frame.flags["C_CONTIGUOUS"]:
+            frame = np.ascontiguousarray(frame)
+        return frame
+
+    def __call__(self) -> np.ndarray:
+        frames = [self._load_frame(path) for path in self.image_paths]
+        if not frames:
+            raise ValueError("No RGB frames were loaded for this trajectory")
+        return np.stack(frames, axis=0)
+
+
+def _default_task_description(task_key: str) -> str:
+    if task_key in TASK_DESCRIPTION_MAP:
+        return TASK_DESCRIPTION_MAP[task_key]
+    return task_key.replace("_", " ").capitalize()
+
+
+def _parse_folder_metadata(folder_name: str) -> tuple[str, str, str]:
+    parts = folder_name.split("_")
+    if len(parts) < 3:
+        raise ValueError(f"Unexpected folder name format: {folder_name}")
+    attempt_id = parts[-1]
+    optimality_key = parts[-2].lower()
+    task_key = "_".join(parts[:-2])
+    return task_key, optimality_key, attempt_id
+
+
+def load_usc_franka_policy_ranking_dataset(
+    dataset_path: str, max_trajectories: int | None = None
+) -> dict[str, list[dict]]:
+    """Load USC Franka policy ranking trajectories grouped by language task."""
+
+    root = Path(dataset_path).expanduser()
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    folders = sorted([path for path in root.iterdir() if path.is_dir()])
+    if not folders:
+        raise ValueError(f"No task folders found in {root}")
+
+    limit = None if max_trajectories is None or max_trajectories < 0 else int(max_trajectories)
+    task_data: dict[str, list[dict]] = defaultdict(list)
+    total = 0
+
+    for folder in folders:
+        if limit is not None and total >= limit:
+            break
+
+        task_key, optimality_key, attempt_id = _parse_folder_metadata(folder.name)
+        if optimality_key not in QUALITY_LABEL_MAP:
+            raise ValueError(f"Unknown optimality label '{optimality_key}' in folder {folder.name}")
+
+        images_dir = folder / "images"
+        image_paths = [str(p) for p in images_dir.glob(CAMERA_PATTERN)]
+        # Sort by frame number to ensure sequential ordering
+        image_paths = sorted(image_paths, key=_extract_frame_number)
+        if not image_paths:
+            print(f"⚠️  Skipping {folder} (no matching {CAMERA_PATTERN} files found)")
+            continue
+
+        frame_loader = USCFrankaFrameLoader(image_paths)
+        task_description = _default_task_description(task_key)
+
+        trajectory = {
+            "id": generate_unique_id(),
+            "task": task_description,
+            "frames": frame_loader(),
+            "is_robot": True,
+            "quality_label": QUALITY_LABEL_MAP[optimality_key],
+            "data_source": "usc_franka_policy_ranking",
+        }
+
+        task_data[task_description].append(trajectory)
+        total += 1
+
+    print(f"Loaded {total} trajectories from {len(task_data)} tasks in USC Franka Policy Ranking dataset")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/usc_koch_human_robot_paired_loader.py b/robometer/dataset_upload/dataset_loaders/usc_koch_human_robot_paired_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..43d2d913da522af8c39fdeb49f89119a3682b23d
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/usc_koch_human_robot_paired_loader.py
@@ -0,0 +1,466 @@
+import json
+import os
+from difflib import SequenceMatcher
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any, Callable
+
+import cv2
+import numpy as np
+import pandas as pd
+from datasets import Dataset
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+MAX_EPISODES = 10  # at most 10 episodes per task, some have more but it's a mistake
+
+
+def _load_video_frames(video_path: str) -> list[np.ndarray]:
+    """Load frames from an MP4 video file."""
+    if not os.path.exists(video_path):
+        raise FileNotFoundError(f"Video file not found: {video_path}")
+
+    frames = []
+    cap = cv2.VideoCapture(video_path)
+
+    try:
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+    finally:
+        cap.release()
+
+    return frames
+
+
+def _string_similarity(s1: str, s2: str) -> float:
+    """Calculate similarity between two strings (0-1)."""
+    return SequenceMatcher(None, s1.lower(), s2.lower()).ratio()
+
+
+def _find_best_match(query: str, options: list[str], threshold: float = 0.6) -> str | None:
+    """Find the best matching string from options, or None if below threshold."""
+    best_match = None
+    best_score = threshold
+
+    for option in options:
+        score = _string_similarity(query, option)
+        if score > best_score:
+            best_score = score
+            best_match = option
+
+    return best_match
+
+
+def _require_lerobot_dataset_class():
+    """
+    Import LeRobotDataset lazily so this module can be imported without lerobot installed.
+
+    Upstream: https://github.com/huggingface/lerobot
+    """
+
+    try:
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset  # type: ignore
+    except Exception as e:  # pragma: no cover - depends on local environment
+        raise ImportError(
+            "Robot demos require Hugging Face LeRobot. Install it in your env, e.g.\n"
+            "  pip install lerobot\n"
+            "Then re-run dataset conversion."
+        ) from e
+    return LeRobotDataset
+
+
+def _load_lerobot_dataset(local_dataset_dir: Path):
+    """Load a local LeRobot dataset directory (downloaded to disk)."""
+
+    LeRobotDataset = _require_lerobot_dataset_class()
+
+    # LeRobot's API has evolved; try a few common constructor styles.
+    constructors: list[Callable[[], Any]] = [
+        lambda: LeRobotDataset(str(local_dataset_dir)),
+        lambda: LeRobotDataset(repo_id=str(local_dataset_dir)),
+        lambda: LeRobotDataset(path=str(local_dataset_dir)),
+        lambda: LeRobotDataset(dataset_path=str(local_dataset_dir)),
+    ]
+
+    last_err: Exception | None = None
+    for ctor in constructors:
+        try:
+            ds = ctor()
+            _ = len(ds)  # force __len__
+            return ds
+        except TypeError as e:
+            last_err = e
+            continue
+        except Exception as e:
+            last_err = e
+            break
+
+    raise RuntimeError(
+        f"Failed to construct LeRobotDataset from local path: {local_dataset_dir}\nLast error: {last_err}"
+    )
+
+
+def _build_video_paths(output_dir: str, dataset_label: str, episode_idx: int, view: str) -> tuple[str, str]:
+    """Build output video paths with shard structure."""
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"{view}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_human_episode(args):
+    """Process a single human episode."""
+    video_path, json_path, episode_idx, lang_vec, output_dir, dataset_label, max_frames, fps = args
+
+    # Load video frames
+    try:
+        frames = _load_video_frames(video_path)
+    except Exception as e:
+        print(f"Error loading video {video_path}: {e}")
+        return None
+
+    if not frames:
+        return None
+
+    # Load metadata
+    with open(json_path, "r") as f:
+        metadata = json.load(f)
+
+    instruction = metadata.get("notes", "")
+    if not instruction:
+        return None
+
+    # Build output video path
+    full_path, rel_path = _build_video_paths(output_dir, dataset_label, episode_idx, "human")
+
+    traj_dict = {
+        "id": generate_unique_id(),
+        "frames": frames,
+        "task": instruction,
+        "is_robot": False,  # Human demonstration
+        "quality_label": "successful",  # All human demos are successful
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    entry = create_hf_trajectory(
+        traj_dict=traj_dict,
+        video_path=full_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_label,
+        use_video=True,
+        fps=fps,
+    )
+
+    if entry:
+        entry["frames"] = rel_path
+        return entry
+    return None
+
+
+def _process_robot_episode_from_lerobot(
+    lerobot_dataset: Any,
+    episode_idx: int,
+    task_instruction: str,
+    lang_vec: Any,
+    output_dir: str,
+    dataset_label: str,
+    max_frames: int,
+    fps: int,
+    global_episode_idx: int,
+    preferred_camera: str = "observation.images.top",
+):
+    """Process a single robot episode from a loaded LeRobotDataset."""
+    try:
+        episode_meta = lerobot_dataset.meta.episodes[episode_idx]
+        ep_start_idx = episode_meta["dataset_from_index"]
+        ep_end_idx = episode_meta["dataset_to_index"]
+        frames = []
+        for frame_idx in range(ep_start_idx, ep_end_idx):
+            frames.append(
+                (lerobot_dataset[frame_idx][preferred_camera].numpy() * 255).astype(np.uint8).transpose(1, 2, 0)
+            )
+
+        assert len(frames) > 0, (
+            f"No frames found for episode {episode_idx} with preferred camera {preferred_camera} and instruction {task_instruction}"
+        )
+
+        full_path, rel_path = _build_video_paths(output_dir, dataset_label, global_episode_idx, "robot")
+        traj_dict = {
+            "id": generate_unique_id(),
+            "frames": frames,
+            "task": task_instruction,
+            "is_robot": True,
+            "quality_label": "successful",
+            "preference_group_id": None,
+            "preference_rank": None,
+        }
+
+        entry = create_hf_trajectory(
+            traj_dict=traj_dict,
+            video_path=full_path,
+            lang_vector=lang_vec,
+            max_frames=max_frames,
+            dataset_name=dataset_label,
+            use_video=True,
+            fps=fps,
+        )
+
+        if entry:
+            entry["frames"] = rel_path
+            return entry
+        return None
+
+    except Exception as e:
+        print(f"Error processing robot episode {episode_idx}: {e}")
+        return None
+
+
+def convert_usc_koch_human_robot_paired_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    trajectory_type: str = "human",  # "human", "robot"
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert USC Koch Human-Robot Paired dataset to HF format.
+
+    Args:
+        dataset_path: Path to the dataset directory containing human/ and robot/ folders
+        dataset_name: Name for the dataset
+        output_dir: Output directory for processed videos
+        trajectory_type: Type of trajectories to include ("human", "robot")
+        max_trajectories: Maximum number of trajectories to process per type (None for all)
+        max_frames: Maximum frames per trajectory
+        fps: Frames per second for output videos
+        num_workers: Number of worker processes (-1 for auto, 0 for sequential)
+
+    Returns:
+        HuggingFace Dataset
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset directory not found: {dataset_path}")
+
+    human_dir = root / "human" / "recordings"
+    robot_dir = root / "robot"
+
+    if not human_dir.exists():
+        raise FileNotFoundError(f"Human recordings directory not found: {human_dir}")
+
+    if not robot_dir.exists():
+        raise FileNotFoundError(f"Robot datasets directory not found: {robot_dir}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Load sentence transformer model
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    all_entries = []
+    global_episode_idx = 0
+
+    # Process human demonstrations
+    if trajectory_type in ["human"]:
+        print("Processing human demonstrations...")
+        human_videos = sorted(human_dir.glob("*.mp4"))
+
+        if max_trajectories is not None and max_trajectories > 0:
+            human_videos = human_videos[:max_trajectories]
+
+        # Pre-compute language embeddings
+        human_tasks = []
+        for video_path in tqdm(human_videos, desc="Loading human metadata"):
+            json_path = video_path.with_suffix(".json")
+            if not json_path.exists():
+                continue
+
+            with open(json_path, "r") as f:
+                metadata = json.load(f)
+            instruction = metadata.get("notes", "")
+            if instruction:
+                human_tasks.append(instruction)
+                if instruction not in lang_cache:
+                    lang_cache[instruction] = lang_model.encode(instruction)
+
+        print(f"Found {len(human_videos)} human videos with {len(set(human_tasks))} unique tasks")
+
+        # Process human episodes
+        human_entries = []
+        if num_workers == 1:
+            for video_path in tqdm(human_videos, desc="Processing human videos"):
+                json_path = video_path.with_suffix(".json")
+                if not json_path.exists():
+                    continue
+
+                with open(json_path, "r") as f:
+                    metadata = json.load(f)
+                instruction = metadata.get("notes", "")
+                if not instruction:
+                    continue
+
+                result = _process_human_episode((
+                    str(video_path),
+                    str(json_path),
+                    global_episode_idx,
+                    lang_cache[instruction],
+                    output_dir,
+                    dataset_name,
+                    max_frames,
+                    fps,
+                ))
+                if result:
+                    human_entries.append(result)
+                    global_episode_idx += 1
+        else:
+            from multiprocessing import Pool
+
+            args_list = []
+            for video_path in human_videos:
+                json_path = video_path.with_suffix(".json")
+                if not json_path.exists():
+                    continue
+
+                with open(json_path, "r") as f:
+                    metadata = json.load(f)
+                instruction = metadata.get("notes", "")
+                if not instruction:
+                    continue
+
+                args_list.append((
+                    str(video_path),
+                    str(json_path),
+                    global_episode_idx,
+                    lang_cache[instruction],
+                    output_dir,
+                    dataset_name,
+                    max_frames,
+                    fps,
+                ))
+                global_episode_idx += 1
+
+            with Pool(processes=num_workers) as pool:
+                results = list(
+                    tqdm(
+                        pool.imap_unordered(_process_human_episode, args_list),
+                        total=len(args_list),
+                        desc="Processing human videos",
+                    )
+                )
+
+            human_entries = [r for r in results if r is not None]
+
+        all_entries.extend(human_entries)
+        print(f"Successfully processed {len(human_entries)} human demonstrations")
+
+    # Process robot demonstrations
+    if trajectory_type in ["robot"]:
+        print("Processing robot demonstrations via LeRobotDataset (top view)...")
+
+        # Find all robot dataset directories that do not hae "suboptimal" or "failure" in the name
+        robot_datasets = [
+            d
+            for d in robot_dir.iterdir()
+            if d.is_dir()
+            and (d / "meta" / "info.json").exists()
+            and "suboptimal" not in d.name
+            and "failure" not in d.name
+        ]
+
+        print(f"Found {len(robot_datasets)} robot datasets")
+
+        robot_entries = []
+        for robot_dataset_path in robot_datasets:
+            # Load task information
+            tasks_path = robot_dataset_path / "meta" / "tasks.parquet"
+            if not tasks_path.exists():
+                print(f"Warning: tasks.parquet not found in {robot_dataset_path}, skipping")
+                continue
+
+            tasks_df = pd.read_parquet(tasks_path)
+            # tasks.parquet sometimes stores the task as index, sometimes as a column
+            task_name = None
+            if len(tasks_df) > 0:
+                if tasks_df.index is not None and len(tasks_df.index) > 0 and isinstance(tasks_df.index[0], str):
+                    task_name = tasks_df.index[0]
+                else:
+                    first_row = tasks_df.iloc[0].to_dict()
+                    for v in first_row.values():
+                        if isinstance(v, str) and v.strip():
+                            task_name = v
+                            break
+
+            if not task_name:
+                print(f"Warning: No task found in {robot_dataset_path}, skipping")
+                continue
+
+            # Pre-compute language embedding
+            if task_name not in lang_cache:
+                lang_cache[task_name] = lang_model.encode(task_name)
+
+            # Load LeRobot dataset from disk once per task dataset.
+            lerobot_ds = _load_lerobot_dataset(robot_dataset_path)
+            total_episodes = min(MAX_EPISODES, len(lerobot_ds.meta.episodes))
+
+            print(f"  Processing {total_episodes} episodes from: {robot_dataset_path.name}")
+
+            # Process each episode
+            for ep_idx in range(total_episodes):
+                result = _process_robot_episode_from_lerobot(
+                    lerobot_dataset=lerobot_ds,
+                    episode_idx=ep_idx,
+                    task_instruction=task_name,
+                    lang_vec=lang_cache[task_name],
+                    output_dir=output_dir,
+                    dataset_label=dataset_name,
+                    max_frames=max_frames,
+                    fps=fps,
+                    global_episode_idx=global_episode_idx,
+                    preferred_camera="observation.images.top",
+                )
+                if result:
+                    robot_entries.append(result)
+                    global_episode_idx += 1
+
+        all_entries.extend(robot_entries)
+        print(f"Successfully processed {len(robot_entries)} robot demonstrations")
+
+    print(f"Total entries: {len(all_entries)}")
+    print(f"Unique instructions: {len(lang_cache)}")
+
+    # Create HuggingFace dataset
+    if not all_entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(all_entries)
diff --git a/robometer/dataset_upload/dataset_loaders/usc_koch_p_ranking_loader.py b/robometer/dataset_upload/dataset_loaders/usc_koch_p_ranking_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..2e7ca7baba6e964a48512ee6ea6f38857c59e2de
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/usc_koch_p_ranking_loader.py
@@ -0,0 +1,274 @@
+import os
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any, Callable
+from uuid import uuid4
+
+import numpy as np
+import pandas as pd
+from datasets import Dataset
+from tqdm import tqdm
+
+from dataset_upload.helpers import create_hf_trajectory, load_sentence_transformer_model
+
+MAX_EPISODES_PER_DATASET = 5  # 5 episodes per dataset; any extra is a mistake
+
+
+def _require_lerobot_dataset_class():
+    """
+    Import LeRobotDataset lazily so this module can be imported without lerobot installed.
+
+    Upstream: https://github.com/huggingface/lerobot
+    """
+
+    try:
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset  # type: ignore
+    except Exception as e:  # pragma: no cover - depends on local environment
+        raise ImportError(
+            "USC Koch P-Ranking robot demos require Hugging Face LeRobot. Install it in your env, e.g.\n"
+            "  pip install lerobot\n"
+            "Then re-run dataset conversion."
+        ) from e
+    return LeRobotDataset
+
+
+def _load_lerobot_dataset(local_dataset_dir: Path):
+    """Load a local LeRobot dataset directory (downloaded to disk)."""
+    LeRobotDataset = _require_lerobot_dataset_class()
+
+    constructors: list[Callable[[], Any]] = [
+        lambda: LeRobotDataset(str(local_dataset_dir)),
+        lambda: LeRobotDataset(repo_id=str(local_dataset_dir)),
+        lambda: LeRobotDataset(path=str(local_dataset_dir)),
+        lambda: LeRobotDataset(dataset_path=str(local_dataset_dir)),
+    ]
+
+    last_err: Exception | None = None
+    for ctor in constructors:
+        try:
+            ds = ctor()
+            _ = len(ds)  # force __len__
+            return ds
+        except TypeError as e:
+            last_err = e
+            continue
+        except Exception as e:
+            last_err = e
+            break
+
+    raise RuntimeError(
+        f"Failed to construct LeRobotDataset from local path: {local_dataset_dir}\nLast error: {last_err}"
+    )
+
+
+def _infer_quality_from_dataset_dirname(dirname: str) -> str:
+    name = dirname.lower()
+    if "suboptimal" in name:
+        return "suboptimal"
+    if "failure" in name:
+        return "failure"
+    return "successful"
+
+
+def _read_task_from_tasks_parquet(tasks_path: Path) -> str | None:
+    tasks_df = pd.read_parquet(tasks_path)
+    if len(tasks_df) == 0:
+        return None
+    # Most of these store the task as the index.
+    if tasks_df.index is not None and len(tasks_df.index) > 0 and isinstance(tasks_df.index[0], str):
+        return str(tasks_df.index[0])
+
+    # Fallback: scan first row for a string value.
+    first_row = tasks_df.iloc[0].to_dict()
+    for v in first_row.values():
+        if isinstance(v, str) and v.strip():
+            return v
+    return None
+
+
+def _build_video_paths(output_dir: str, dataset_label: str, episode_idx: int, view: str) -> tuple[str, str]:
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"{view}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _to_rgb_uint8_hwc(frame: Any) -> np.ndarray:
+    """
+    Convert a single frame to RGB uint8 HWC.
+    Expected LeRobot image tensors are often float in [0,1] shaped (C,H,W).
+    """
+    if hasattr(frame, "detach") and hasattr(frame, "cpu") and hasattr(frame, "numpy"):
+        frame = frame.detach().cpu().numpy()
+    frame = np.asarray(frame)
+
+    # CHW -> HWC
+    if frame.ndim == 3 and frame.shape[0] in (1, 3, 4) and frame.shape[-1] not in (1, 3, 4):
+        frame = np.transpose(frame, (1, 2, 0))
+
+    if frame.dtype != np.uint8:
+        frame = np.clip(frame, 0.0, 1.0)
+        frame = (frame * 255).astype(np.uint8)
+
+    if frame.ndim == 3 and frame.shape[-1] == 4:
+        frame = frame[..., :3]
+    if frame.ndim == 3 and frame.shape[-1] == 1:
+        frame = np.repeat(frame, 3, axis=-1)
+
+    return frame
+
+
+def _extract_episode_frames_from_lerobot(
+    lerobot_dataset: Any,
+    episode_idx: int,
+    preferred_camera: str = "observation.images.top",
+) -> list[np.ndarray]:
+    """
+    Extract full episode frames using LeRobot's episode metadata boundaries.
+    """
+    episode_meta = lerobot_dataset.meta.episodes[episode_idx]
+    ep_start_idx = int(episode_meta["dataset_from_index"])
+    ep_end_idx = int(episode_meta["dataset_to_index"])
+
+    frames: list[np.ndarray] = []
+    for frame_idx in range(ep_start_idx, ep_end_idx):
+        item = lerobot_dataset[frame_idx]
+        if preferred_camera not in item:
+            raise KeyError(
+                f"Preferred camera '{preferred_camera}' not found in dataset item keys: {sorted(item.keys())}"
+            )
+        frames.append(_to_rgb_uint8_hwc(item[preferred_camera]))
+    return frames
+
+
+def convert_usc_koch_p_ranking_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+    max_episodes_per_dataset: int = 10,
+    preferred_camera: str = "observation.images.top",
+) -> Dataset:
+    """
+    Convert USC Koch datasets (robot-only) in LeRobot format to HF format for reward training.
+
+    Folder layout expected (see download script):
+      <dataset_path>/robot/usc_koch_*_{task}/
+        meta/info.json, meta/tasks.parquet, meta/episodes/... (LeRobot)
+        videos/observation.images.top/... (LeRobot)
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset directory not found: {dataset_path}")
+
+    robot_dir = root / "robot"
+    if not robot_dir.exists():
+        raise FileNotFoundError(f"Robot datasets directory not found: {robot_dir}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Load sentence transformer model once; cache task embeddings
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    robot_dataset_dirs = [d for d in robot_dir.iterdir() if d.is_dir() and (d / "meta" / "info.json").exists()]
+    robot_dataset_dirs = sorted(robot_dataset_dirs, key=lambda p: p.name)
+
+    all_entries: list[dict[str, Any]] = []
+    global_episode_idx = 0
+
+    print("Processing USC Koch P-Ranking robot demonstrations via LeRobotDataset...")
+    print(f"Found {len(robot_dataset_dirs)} robot dataset directories")
+
+    for robot_dataset_path in robot_dataset_dirs:
+        quality = _infer_quality_from_dataset_dirname(robot_dataset_path.name)
+
+        tasks_path = robot_dataset_path / "meta" / "tasks.parquet"
+        if not tasks_path.exists():
+            print(f"Warning: tasks.parquet not found in {robot_dataset_path}, skipping")
+            continue
+
+        task_instruction = _read_task_from_tasks_parquet(tasks_path)
+        if not task_instruction:
+            print(f"Warning: No task found in {robot_dataset_path}, skipping")
+            continue
+
+        if task_instruction not in lang_cache:
+            lang_cache[task_instruction] = lang_model.encode(task_instruction)
+
+        # Load LeRobot dataset
+        lerobot_ds = _load_lerobot_dataset(robot_dataset_path)
+        total_episodes = min(len(lerobot_ds.meta.episodes), MAX_EPISODES_PER_DATASET)
+
+        print(f"  {robot_dataset_path.name}: quality={quality} | episodes={total_episodes}")
+
+        for ep_idx in range(total_episodes):
+            if max_trajectories is not None and max_trajectories > 0 and len(all_entries) >= max_trajectories:
+                break
+
+            try:
+                frames = _extract_episode_frames_from_lerobot(
+                    lerobot_dataset=lerobot_ds,
+                    episode_idx=ep_idx,
+                    preferred_camera=preferred_camera,
+                )
+                if not frames:
+                    continue
+
+                full_path, rel_path = _build_video_paths(output_dir, dataset_name, global_episode_idx, "robot")
+                traj_dict = {
+                    "id": str(uuid4()),
+                    "frames": frames,
+                    "task": task_instruction,
+                    "is_robot": True,
+                    "quality_label": quality,
+                    "preference_group_id": None,
+                    "preference_rank": None,
+                }
+                entry = create_hf_trajectory(
+                    traj_dict=traj_dict,
+                    video_path=full_path,
+                    lang_vector=lang_cache[task_instruction],
+                    max_frames=max_frames,
+                    dataset_name=dataset_name,
+                    use_video=True,
+                    fps=fps,
+                )
+                if entry:
+                    entry["frames"] = rel_path
+                    all_entries.append(entry)
+                    global_episode_idx += 1
+            except Exception as e:
+                print(f"Error processing {robot_dataset_path.name} episode {ep_idx}: {e}")
+                continue
+
+        if max_trajectories is not None and max_trajectories > 0 and len(all_entries) >= max_trajectories:
+            break
+
+    print(f"Total entries: {len(all_entries)}")
+    print(f"Unique instructions: {len(lang_cache)}")
+
+    if not all_entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(all_entries)
diff --git a/robometer/dataset_upload/dataset_loaders/usc_xarm_policy_ranking_loader.py b/robometer/dataset_upload/dataset_loaders/usc_xarm_policy_ranking_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..d16b5047fa993d4a6134020f7d23efce430597cd
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/usc_xarm_policy_ranking_loader.py
@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+import pickle
+from collections import defaultdict
+from pathlib import Path
+from typing import Iterable
+
+import numpy as np
+import cv2
+
+from dataset_upload.helpers import generate_unique_id
+
+RGB_CAMERA_PRIORITY = (
+    "f1421749_rgb",
+    "f1421370_rgb",
+    "front_rgb",
+    "hand_rgb",
+    "rgb",
+)
+
+TASK_DESCRIPTION_MAP = {
+    "close_drawer": "Close the drawer",
+    "fold_towel": "Fold the towel in half",
+    "pick_blue_cup": "Put blue cup in sink",
+    "put_banana_on_red_plate": "Put the banana on the red plate",
+    "put_eggplant_black_bowl": "Put the eggplant in the black bowl",
+    "stack_green_on_red": "Stack the green shape on the red cylinder",
+}
+
+QUALITY_LABEL_MAP = {
+    "success": "successful",
+    "successful": "successful",
+    "subopt": "suboptimal",
+    "suboptimal": "suboptimal",
+    "fail": "failure",
+    "failure": "failure",
+}
+
+
+class USCXArmFrameLoader:
+    """Lazy loader that stitches RGB frames from timestep pickle files."""
+
+    def __init__(self, pickle_paths: Iterable[str], rgb_keys: tuple[str, ...] = RGB_CAMERA_PRIORITY) -> None:
+        self.pickle_paths = list(pickle_paths)
+        self.rgb_keys = rgb_keys
+
+    def _extract_rgb_frame(self, step_data: dict):
+        for key in self.rgb_keys:
+            if key in step_data and step_data[key] is not None:
+                return step_data[key]
+        return None
+
+    def _prepare_frame(self, frame_array: np.ndarray) -> np.ndarray:
+        frame = np.asarray(frame_array)
+        if frame.ndim != 3:
+            raise ValueError(f"Expected 3D RGB frame, got shape {frame.shape}")
+
+        # Convert channel-first -> channel-last if needed
+        if frame.shape[0] in (3, 4) and frame.shape[-1] not in (3, 4):
+            frame = frame[:3, ...]
+            frame = np.transpose(frame, (1, 2, 0))
+
+        if frame.dtype != np.uint8:
+            if np.issubdtype(frame.dtype, np.floating):
+                # Assume [0, 1] range and scale, otherwise clip to [0, 255]
+                frame = np.clip(frame, 0.0, 1.0) if frame.max() <= 1.0 else np.clip(frame, 0.0, 255.0)
+                scale = 255.0 if frame.max() <= 1.0 else 1.0
+                frame = (frame * scale).astype(np.uint8)
+            else:
+                frame = frame.astype(np.uint8, copy=False)
+
+        # Convert BGR to RGB (frames from pickle files are stored as BGR)
+        if frame.shape[2] == 3:
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+        # Ensure the array is contiguous for efficient processing
+        if not frame.flags["C_CONTIGUOUS"]:
+            frame = np.ascontiguousarray(frame)
+
+        return frame
+
+    def __call__(self) -> np.ndarray:
+        frames: list[np.ndarray] = []
+        for path in self.pickle_paths:
+            with open(path, "rb") as f:
+                step_data = pickle.load(f)
+            frame_array = self._extract_rgb_frame(step_data)
+            if frame_array is None:
+                continue
+            frame = self._prepare_frame(frame_array)
+            frames.append(frame)
+
+        if not frames:
+            raise ValueError(f"No RGB frames found in {self.pickle_paths[0] if self.pickle_paths else 'unknown path'}")
+
+        return np.stack(frames, axis=0)
+
+
+def _default_task_description(task_key: str) -> str:
+    if task_key in TASK_DESCRIPTION_MAP:
+        return TASK_DESCRIPTION_MAP[task_key]
+    return task_key.replace("_", " ").capitalize()
+
+
+def _parse_folder_metadata(folder_name: str) -> tuple[str, str, str]:
+    parts = folder_name.split("_")
+    if len(parts) < 3:
+        raise ValueError(f"Unexpected folder name format: {folder_name}")
+    attempt_id = parts[-1]
+    optimality_key = parts[-2].lower()
+    task_key = "_".join(parts[:-2])
+    return task_key, optimality_key, attempt_id
+
+
+def load_usc_xarm_policy_ranking_dataset(
+    dataset_path: str, max_trajectories: int | None = None
+) -> dict[str, list[dict]]:
+    """Load USC xArm policy ranking trajectories grouped by language task."""
+
+    root = Path(dataset_path).expanduser()
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    if (root / "xarm_policy_ranking").exists():
+        root = root / "xarm_policy_ranking"
+
+    folders = sorted([path for path in root.iterdir() if path.is_dir()])
+    if not folders:
+        raise ValueError(f"No task folders found in {root}")
+
+    limit = None if max_trajectories is None or max_trajectories < 0 else int(max_trajectories)
+    task_data: dict[str, list[dict]] = defaultdict(list)
+    total = 0
+
+    for folder in folders:
+        if limit is not None and total >= limit:
+            break
+
+        task_key, optimality_key, attempt_id = _parse_folder_metadata(folder.name)
+        if optimality_key not in QUALITY_LABEL_MAP:
+            raise ValueError(f"Unknown optimality label '{optimality_key}' in folder {folder.name}")
+
+        pickle_paths = sorted(str(p) for p in folder.glob("*.pkl"))
+        if not pickle_paths:
+            print(f"⚠️  Skipping {folder} (no pickle files found)")
+            continue
+
+        frame_loader = USCXArmFrameLoader(pickle_paths)
+        task_description = _default_task_description(task_key)
+
+        trajectory = {
+            "id": generate_unique_id(),
+            "task": task_description,
+            "frames": frame_loader(),
+            "is_robot": True,
+            "quality_label": QUALITY_LABEL_MAP[optimality_key],
+            "data_source": "usc_xarm_policy_ranking",
+        }
+
+        task_data[task_description].append(trajectory)
+        total += 1
+
+    print(f"Loaded {total} trajectories from {len(task_data)} tasks in USC xArm Policy Ranking dataset")
+    return task_data
diff --git a/robometer/dataset_upload/dataset_loaders/utd_so101_clean_policy_ranking_loader.py b/robometer/dataset_upload/dataset_loaders/utd_so101_clean_policy_ranking_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..8308acd8171aa75a055421fd1cadd83bc523d4d3
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/utd_so101_clean_policy_ranking_loader.py
@@ -0,0 +1,276 @@
+import json
+import os
+from multiprocessing import cpu_count
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from datasets import Dataset, concatenate_datasets
+from PIL import Image
+from tqdm import tqdm
+
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    generate_unique_id,
+    load_sentence_transformer_model,
+)
+
+
+def _load_gif_frames(gif_path: str) -> list[np.ndarray]:
+    """Load frames from a GIF file."""
+    if not os.path.exists(gif_path):
+        raise FileNotFoundError(f"GIF file not found: {gif_path}")
+
+    frames = []
+    with Image.open(gif_path) as img:
+        try:
+            for frame_idx in range(img.n_frames):
+                img.seek(frame_idx)
+                # Convert to RGB and then to numpy array
+                frame = img.convert("RGB")
+                frames.append(np.array(frame))
+        except EOFError:
+            pass  # End of frames
+
+    return frames
+
+
+def _parse_quality_label(folder_name: str) -> str:
+    """Parse quality label from folder name."""
+    if "closs_succ" in folder_name.lower():
+        return "suboptimal"
+    elif "fail" in folder_name.lower():
+        return "failure"
+    elif "succ" in folder_name.lower():
+        return "successful"
+    else:
+        raise ValueError(f"Unknown quality label in folder: {folder_name}")
+
+
+def _build_video_paths(output_dir: str, dataset_label: str, episode_idx: int, view: str) -> tuple[str, str]:
+    """Build output video paths with shard structure."""
+    shard_index = episode_idx // 1000
+    shard_dir = f"shard_{shard_index:04d}"
+    episode_dir = os.path.join(output_dir, dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}")
+    os.makedirs(episode_dir, exist_ok=True)
+    filename = f"{view}.mp4"
+    full_path = os.path.join(episode_dir, filename)
+    rel_path = os.path.join(dataset_label.lower(), shard_dir, f"episode_{episode_idx:06d}", filename)
+    return full_path, rel_path
+
+
+def _process_episode(args):
+    """Process a single episode for a specific view."""
+    gif_path, episode_idx, instruction, quality_label, view, lang_vec, output_dir, dataset_label, max_frames, fps = args
+
+    # Load GIF frames
+    try:
+        frames = _load_gif_frames(gif_path)
+    except Exception as e:
+        print(f"Error loading GIF {gif_path}: {e}")
+        return None
+
+    if not frames:
+        return None
+
+    # Build output video path
+    full_path, rel_path = _build_video_paths(output_dir, dataset_label, episode_idx, view)
+
+    traj_dict = {
+        "id": generate_unique_id(),
+        "frames": frames,
+        "task": instruction,
+        "is_robot": True,
+        "quality_label": quality_label,
+        "preference_group_id": None,
+        "preference_rank": None,
+    }
+
+    entry = create_hf_trajectory(
+        traj_dict=traj_dict,
+        video_path=full_path,
+        lang_vector=lang_vec,
+        max_frames=max_frames,
+        dataset_name=dataset_label,
+        use_video=True,
+        fps=fps,
+    )
+
+    if entry:
+        entry["frames"] = rel_path
+        return entry
+    return None
+
+
+def convert_utd_so101_clean_policy_ranking_to_hf(
+    dataset_path: str,
+    dataset_name: str,
+    output_dir: str,
+    view: str = "wrist",  # "wrist" or "top"
+    max_trajectories: int | None = None,
+    max_frames: int = 64,
+    fps: int = 10,
+    num_workers: int = -1,
+) -> Dataset:
+    """Convert UTD SO101 Clean Policy Ranking dataset to HF format.
+
+    Args:
+        dataset_path: Path to the dataset directory containing quality-labeled subfolders
+        dataset_name: Name for the dataset
+        output_dir: Output directory for processed videos
+        view: Camera view to use ("wrist" or "top")
+        max_trajectories: Maximum number of trajectories to process per quality label (None for all)
+        max_frames: Maximum frames per trajectory
+        fps: Frames per second for output videos
+        num_workers: Number of worker processes (-1 for auto, 0 for sequential)
+
+    Returns:
+        HuggingFace Dataset
+    """
+    root = Path(os.path.expanduser(dataset_path))
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset directory not found: {dataset_path}")
+
+    if num_workers == -1:
+        num_workers = min(cpu_count(), 8)
+    elif num_workers == 0:
+        num_workers = 1
+
+    # Load sentence transformer model
+    lang_model = load_sentence_transformer_model()
+    lang_cache: dict[str, Any] = {}
+
+    # Find all quality-labeled subdirectories
+    all_episodes = []
+    quality_dirs = []
+
+    for item in root.iterdir():
+        if item.is_dir() and any(label in item.name.lower() for label in ["succ", "fail", "closs_succ"]):
+            # Find the inner directory containing the actual data
+            inner_dirs = [d for d in item.iterdir() if d.is_dir()]
+            if inner_dirs:
+                quality_dirs.append((inner_dirs[0], _parse_quality_label(item.name)))
+
+    if not quality_dirs:
+        raise FileNotFoundError(f"No quality-labeled directories found in: {dataset_path}")
+
+    print(f"Found {len(quality_dirs)} quality-labeled directories")
+
+    # Process each quality directory
+    global_episode_idx = 0
+    for data_dir, quality_label in quality_dirs:
+        print(f"Processing {quality_label} episodes from: {data_dir}")
+
+        # Load task instructions
+        vla_task_path = data_dir / "vla_task.json"
+        if not vla_task_path.exists():
+            print(f"Warning: vla_task.json not found in {data_dir}, skipping")
+            continue
+
+        with open(vla_task_path, "r") as f:
+            instructions = json.load(f)
+
+        # Find all episode GIFs for the specified view
+        gif_files = sorted(data_dir.glob(f"episode_*_{view}.gif"))
+
+        print(f"  Found {len(gif_files)} episodes for {view} view with quality: {quality_label}")
+
+        # Limit trajectories if specified
+        if max_trajectories is not None and max_trajectories > 0:
+            gif_files = gif_files[:max_trajectories]
+
+        # Match episodes to instructions
+        for gif_file in gif_files:
+            # Extract episode number from filename (e.g., episode_0_wrist.gif -> 0)
+            episode_num = int(gif_file.stem.split("_")[1])
+
+            if episode_num >= len(instructions):
+                print(f"Warning: Episode {episode_num} has no instruction, skipping")
+                continue
+
+            instruction = instructions[episode_num]
+
+            # Pre-compute language embedding
+            if instruction not in lang_cache:
+                lang_cache[instruction] = lang_model.encode(instruction)
+
+            all_episodes.append({
+                "gif_path": str(gif_file),
+                "episode_idx": global_episode_idx,
+                "instruction": instruction,
+                "quality_label": quality_label,
+                "view": view,
+                "lang_vec": lang_cache[instruction],
+            })
+            global_episode_idx += 1
+
+    print(f"Total episodes to process: {len(all_episodes)}")
+    print(f"Unique instructions: {len(lang_cache)}")
+
+    # Process episodes
+    entries = []
+    if num_workers == 1:
+        # Sequential processing
+        for episode in tqdm(all_episodes, desc=f"Processing {view} episodes"):
+            result = _process_episode((
+                episode["gif_path"],
+                episode["episode_idx"],
+                episode["instruction"],
+                episode["quality_label"],
+                episode["view"],
+                episode["lang_vec"],
+                output_dir,
+                dataset_name,
+                max_frames,
+                fps,
+            ))
+            if result:
+                entries.append(result)
+    else:
+        # Parallel processing
+        from multiprocessing import Pool
+
+        args_list = [
+            (
+                episode["gif_path"],
+                episode["episode_idx"],
+                episode["instruction"],
+                episode["quality_label"],
+                episode["view"],
+                episode["lang_vec"],
+                output_dir,
+                dataset_name,
+                max_frames,
+                fps,
+            )
+            for episode in all_episodes
+        ]
+
+        with Pool(processes=num_workers) as pool:
+            results = list(
+                tqdm(
+                    pool.imap_unordered(_process_episode, args_list),
+                    total=len(args_list),
+                    desc=f"Processing {view} episodes",
+                )
+            )
+
+        entries = [r for r in results if r is not None]
+
+    print(f"Successfully processed {len(entries)} episodes for {view} view")
+
+    # Create HuggingFace dataset
+    if not entries:
+        return Dataset.from_dict({
+            "id": [],
+            "task": [],
+            "lang_vector": [],
+            "data_source": [],
+            "frames": [],
+            "is_robot": [],
+            "quality_label": [],
+            "preference_group_id": [],
+            "preference_rank": [],
+        })
+
+    return Dataset.from_list(entries)
diff --git a/robometer/dataset_upload/dataset_loaders/utd_so101_loader.py b/robometer/dataset_upload/dataset_loaders/utd_so101_loader.py
new file mode 100644
index 0000000000000000000000000000000000000000..ae7a724e1fdfb3fecb308e58f1cab537561d5793
--- /dev/null
+++ b/robometer/dataset_upload/dataset_loaders/utd_so101_loader.py
@@ -0,0 +1,158 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+from collections import defaultdict
+from pathlib import Path
+from typing import Iterable
+
+import numpy as np
+import cv2
+
+from dataset_upload.helpers import generate_unique_id
+
+TASK_DESCRIPTION_MAP = {
+    "Add_pepper_to_the_green_bowl": "Add pepper to the green bowl",
+    "Collect_the_fork_to_the_yellow_box": "Collect the fork to the yellow box",
+    "Press_the_button": "Press the button",
+    "Put_the_bread_in_the_oven": "Put the bread in the oven",
+    "Put_the_fruit_in_the_yellow_plate": "Put the apple on the yellow plate",
+    "Put_the_marker_into_the_pen_cup": "Put the marker into the pen cup",
+    "Put_the_red_bowl_on_the_blue_plate": "Put the red bowl on the blue plate",
+    "Put_the_red_cup_on_the_purple_coaster": "Put the red cup on the purple coaster",
+    "Put_the_rubber_to_the_blue_pencil_box": "Put the eraser in the blue pencil box",
+    "Stack_the_green_block_on_the_red_block": "Stack the green block on the red block",
+}
+
+QUALITY_LABEL_MAP = {
+    "succ": "successful",
+    "success": "successful",
+    "successful": "successful",
+    "subopt": "suboptimal",
+    "suboptimal": "suboptimal",
+    "fail": "failure",
+    "failure": "failure",
+}
+
+
+class UTDSO101FrameLoader:
+    """Lazy loader that extracts RGB frames from MP4 video files."""
+
+    def __init__(self, video_path: str) -> None:
+        if not Path(video_path).exists():
+            raise FileNotFoundError(f"Video file not found: {video_path}")
+        self.video_path = video_path
+
+    def __call__(self) -> np.ndarray:
+        """Load all frames from the MP4 video file."""
+        cap = cv2.VideoCapture(self.video_path)
+        if not cap.isOpened():
+            raise ValueError(f"Could not open video file: {self.video_path}")
+
+        frames: list[np.ndarray] = []
+        while True:
+            ret, frame_bgr = cap.read()
+            if not ret:
+                break
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+
+        cap.release()
+
+        if not frames:
+            raise ValueError(f"No frames found in video: {self.video_path}")
+
+        return np.stack(frames, axis=0).astype(np.uint8)
+
+
+def _default_task_description(task_key: str) -> str:
+    """Convert task key to a readable description."""
+    if task_key in TASK_DESCRIPTION_MAP:
+        return TASK_DESCRIPTION_MAP[task_key]
+    return task_key.replace("_", " ").capitalize()
+
+
+def _parse_video_metadata(video_filename: str) -> tuple[str, str, str]:
+    """Parse video filename to extract task, optimality, and demo_idx.
+
+    Expected format: {task}_{optimality}_{demo_idx}.mp4
+    Example: pick_blue_cup_success_1.mp4
+    """
+    # Remove .mp4 extension
+    name_without_ext = video_filename.replace(".mp4", "")
+    parts = name_without_ext.split("_")
+
+    if len(parts) < 3:
+        raise ValueError(f"Unexpected video filename format: {video_filename}")
+
+    demo_idx = parts[-1]
+    optimality_key = parts[-2].lower()
+    task_key = "_".join(parts[:-2])
+
+    return task_key, optimality_key, demo_idx
+
+
+def load_utd_so101_dataset(
+    dataset_path: str, max_trajectories: int | None = None, is_robot: bool = True, data_source: str = "utd_so101"
+) -> dict[str, list[dict]]:
+    """Load UTD SO101 trajectories grouped by language task.
+
+    Args:
+        dataset_path: Path to the dataset directory
+        max_trajectories: Maximum number of trajectories to load
+        is_robot: Whether trajectories are robot (True) or human (False)
+        data_source: Data source identifier for the dataset
+    """
+
+    root = Path(dataset_path).expanduser()
+    if not root.exists():
+        raise FileNotFoundError(f"Dataset path not found: {dataset_path}")
+
+    if (root / "koch_arm_ut_dallas").exists():
+        root = root / "koch_arm_ut_dallas"
+
+    # Find all MP4 files
+    video_files = sorted([p for p in root.glob("*.mp4")])
+    if not video_files:
+        raise ValueError(f"No MP4 files found in {root}")
+
+    limit = None if max_trajectories is None or max_trajectories < 0 else int(max_trajectories)
+    task_data: dict[str, list[dict]] = defaultdict(list)
+    total = 0
+
+    for video_path in video_files:
+        if limit is not None and total >= limit:
+            break
+
+        try:
+            task_key, optimality_key, demo_idx = _parse_video_metadata(video_path.name)
+        except ValueError as e:
+            print(f"⚠️  Skipping {video_path.name}: {e}")
+            continue
+
+        # For human videos, always set quality_label to "successful"
+        if not is_robot:
+            quality_label = "successful"
+        elif optimality_key not in QUALITY_LABEL_MAP:
+            print(f"⚠️  Skipping {video_path.name}: Unknown optimality label '{optimality_key}'")
+            continue
+        else:
+            quality_label = QUALITY_LABEL_MAP[optimality_key]
+
+        frame_loader = UTDSO101FrameLoader(str(video_path))
+        task_description = _default_task_description(task_key)
+
+        trajectory = {
+            "id": generate_unique_id(),
+            "task": task_description,
+            "frames": frame_loader,
+            "is_robot": is_robot,
+            "quality_label": quality_label,
+            "data_source": data_source,
+        }
+
+        task_data[task_description].append(trajectory)
+        total += 1
+
+    dataset_type = "robot" if is_robot else "human"
+    print(f"Loaded {total} {dataset_type} trajectories from {len(task_data)} tasks in UTD SO101 dataset")
+    return task_data
diff --git a/robometer/dataset_upload/generate_hf_dataset.py b/robometer/dataset_upload/generate_hf_dataset.py
new file mode 100755
index 0000000000000000000000000000000000000000..50446be784cacfbb48aaa2d678841a9ef25de836
--- /dev/null
+++ b/robometer/dataset_upload/generate_hf_dataset.py
@@ -0,0 +1,1083 @@
+#!/usr/bin/env python3
+"""
+Main dataset converter that can convert any dataset to HuggingFace format for Robometer model training.
+This is a generic converter that works with any dataset-specific loader.
+"""
+
+import os
+
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"  # hide INFO/WARN/ERROR; only FATAL remains
+import multiprocessing as mp
+
+import numpy as np
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from functools import partial
+from multiprocessing import Pool, cpu_count
+from typing import Any, Optional
+
+from pyrallis import wrap
+from tqdm import tqdm
+
+import datasets
+from datasets import Dataset
+
+# from robometer.data.dataset_types import Trajectory  # not needed, just type hint
+from dataset_upload.helpers import (
+    create_hf_trajectory,
+    create_output_directory,
+    flatten_task_data,
+    load_sentence_transformer_model,
+)
+from huggingface_hub import HfApi
+
+# make sure these come after importing torch. otherwise something breaks...
+try:
+    import absl.logging as absl_logging
+
+    absl_logging.set_verbosity(absl_logging.ERROR)
+except Exception:
+    pass
+try:
+    import tensorflow as tf
+
+    tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
+except Exception:
+    pass
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+
+def push_hf_dataset_and_video_files_to_hub(dataset, hub_repo_id, hub_token, dataset_name, output_dir):
+    print(f"Pushing dataset to HuggingFace Hub: {hub_repo_id}")
+    dataset.push_to_hub(
+        hub_repo_id,
+        config_name=dataset_name.lower(),
+        token=hub_token,
+        private=False,
+        commit_message=f"Add {dataset_name} dataset for Robometer training",
+    )
+    print(f"✅ Successfully pushed dataset to: https://huggingface.co/datasets/{hub_repo_id}")
+    api = HfApi(token=hub_token)
+    api.upload_large_folder(
+        folder_path=output_dir,
+        repo_id=hub_repo_id,
+        repo_type="dataset",
+        num_workers=min(4, cpu_count()),
+    )
+    print(f"✅ Successfully pushed video files for {dataset_name} to: https://huggingface.co/datasets/{hub_repo_id}")
+
+
+def get_trajectory_subdir_path(trajectory_idx: int, files_per_subdir: int = 1000) -> str:
+    """
+    Generate subdirectory path for a trajectory to avoid too many files per directory.
+
+    Args:
+        trajectory_idx: Index of the trajectory
+        files_per_subdir: Maximum files per subdirectory (default: 1000)
+
+    Returns:
+        str: Subdirectory name like 'batch_0000'
+    """
+    subdir_index = trajectory_idx // files_per_subdir
+    return f"batch_{subdir_index:04d}"
+
+
+# Global dataset features definition
+BASE_FEATURES = {
+    "id": datasets.Value("string"),
+    "task": datasets.Value("string"),
+    "lang_vector": datasets.Sequence(datasets.Value("float32")),
+    "data_source": datasets.Value("string"),
+    "frames": None,  # Will be set based on use_video parameter
+    "is_robot": datasets.Value("bool"),
+    "quality_label": datasets.Value("string"),
+    # "preference_group_id": datasets.Value("string"),
+    # "preference_rank": datasets.Value("int32"),
+    "partial_success": datasets.Value("float32"),  # in [0, 1]
+}
+
+
+@dataclass
+class DatasetConfig:
+    """Config for dataset settings"""
+
+    dataset_path: str = field(default="", metadata={"help": "Path to the dataset"})
+    dataset_name: str = field(default=None, metadata={"help": "Name of the dataset (defaults to dataset_type)"})
+    exclude_wrist_cam: bool = field(default=False, metadata={"help": "Exclude wrist camera views (MIT Franka only)"})
+
+
+@dataclass
+class OutputConfig:
+    """Config for output settings"""
+
+    output_dir: str = field(default="robometer_dataset", metadata={"help": "Output directory for the dataset"})
+    max_trajectories: Optional[int] = field(
+        default=None, metadata={"help": "Maximum number of trajectories to process (None for all)"}
+    )
+    max_frames: int = field(
+        default=64, metadata={"help": "Maximum number of frames per trajectory (-1 for no downsampling)"}
+    )
+    use_video: bool = field(default=True, metadata={"help": "Use MP4 videos instead of individual frame images"})
+    shortest_edge_size: Optional[int] = field(default=240, metadata={"help": "Shortest edge size for video resizing"})
+    center_crop: bool = field(
+        default=False,
+        metadata={"help": "Center crop the video to the target size. Defaults to False, which means no cropping."},
+    )
+    fps: int = field(default=10, metadata={"help": "Frames per second for video creation"})
+    num_workers: int = field(
+        default=-1, metadata={"help": "Number of parallel workers for processing (-1 for auto, 0 for sequential)"}
+    )
+
+
+@dataclass
+class HubConfig:
+    """Config for HuggingFace Hub settings"""
+
+    push_to_hub: bool = field(default=False, metadata={"help": "Push dataset to HuggingFace Hub"})
+    hub_repo_id: str = field(default=None, metadata={"help": "HuggingFace Hub repository ID"})
+    hub_token: str = field(
+        default=None, metadata={"help": "HuggingFace Hub token (or set HF_TOKEN environment variable)"}
+    )
+
+
+@dataclass
+class GenerateConfig:
+    """Main configuration for dataset generation"""
+
+    dataset: DatasetConfig = field(default_factory=DatasetConfig)
+    output: OutputConfig = field(default_factory=OutputConfig)
+    hub: HubConfig = field(default_factory=HubConfig)
+
+
+def process_single_trajectory(args):
+    """
+    Worker function to process a single trajectory.
+
+    Args:
+        args: Tuple containing (trajectory_idx, trajectory, lang_vector, hf_creator_fn, output_dir, dataset_name, max_frames, use_video, fps)
+
+    Returns:
+        Dict: Processed trajectory data or None if failed
+    """
+    trajectory_idx, trajectory, lang_vector, hf_creator_fn, output_dir, dataset_name, max_frames, use_video, fps = args
+
+    try:
+        # Create output directory for this trajectory with subdirectory structure
+        subdir_name = get_trajectory_subdir_path(trajectory_idx)
+        full_video_path = os.path.join(
+            output_dir, dataset_name.lower(), subdir_name, f"trajectory_{trajectory_idx:04d}.mp4"
+        )
+        relative_video_path = os.path.join(dataset_name.lower(), subdir_name, f"trajectory_{trajectory_idx:04d}.mp4")
+        os.makedirs(os.path.dirname(full_video_path), exist_ok=True)
+
+        # Process trajectory (lang_vector is already computed)
+        processed_trajectory = hf_creator_fn(
+            traj_dict=trajectory,
+            video_path=full_video_path,
+            lang_vector=lang_vector,  # Pre-computed language vector
+            max_frames=max_frames,
+            dataset_name=dataset_name,
+            use_video=use_video,
+            fps=fps,
+        )
+
+        if processed_trajectory is None:
+            return None
+
+        # Replace the full path with relative path in the processed trajectory
+        if processed_trajectory and "frames" in processed_trajectory:
+            processed_trajectory["frames"] = relative_video_path
+
+        return processed_trajectory
+
+    except Exception as e:
+        print(f"❌ Error processing trajectory {trajectory_idx}: {e}")
+        return None
+
+
+def convert_dataset_to_hf_format(
+    trajectories: list[dict],
+    hf_creator_fn: Callable[[dict, str, str, int, Any, int, str], Any],
+    output_dir: str = "robometer_dataset",
+    dataset_name: str = "",
+    max_trajectories: int | None = None,
+    max_frames: int = -1,
+    use_video: bool = True,
+    fps: int = 10,
+    num_workers: int = -1,
+    push_to_hub: bool = False,
+    hub_repo_id: str | None = None,
+    hub_token: str | None = None,
+) -> Dataset:
+    """Convert a list of trajectories to HuggingFace format."""
+
+    print(f"Converting {dataset_name} dataset to HuggingFace format...")
+
+    # Create output directory
+    create_output_directory(output_dir)
+
+    # Validate input
+    if not trajectories:
+        raise ValueError(f"No trajectories provided for {dataset_name} dataset.")
+
+    print(f"Processing {len(trajectories)} trajectories")
+
+    # Limit trajectories if specified
+    if max_trajectories != -1:
+        trajectories = trajectories[:max_trajectories]
+
+    # Determine number of workers
+    if num_workers == -1:
+        num_workers = min(cpu_count(), len(trajectories))
+    elif num_workers == 0:
+        num_workers = 1  # Sequential processing
+
+    print(f"Using {num_workers} worker(s) for parallel processing")
+
+    # Pre-compute language embeddings to avoid loading sentence transformer in each worker
+    print("Pre-computing language embeddings...")
+    lang_model = load_sentence_transformer_model()
+
+    lang_vectors = []
+    unique_tasks = {}  # Cache for identical task descriptions
+
+    for trajectory in tqdm(trajectories, desc="Computing language embeddings"):
+        task_description = trajectory["task"]
+
+        # Use cache to avoid recomputing identical task descriptions
+        if task_description not in unique_tasks:
+            unique_tasks[task_description] = lang_model.encode(task_description)
+
+        lang_vectors.append(unique_tasks[task_description])
+
+    print(f"Computed embeddings for {len(unique_tasks)} unique task descriptions")
+
+    # Process trajectories
+    all_entries = []
+
+    if num_workers == 1:
+        # Sequential processing (using pre-computed embeddings)
+        for trajectory_idx, (trajectory, lang_vector) in enumerate(
+            tqdm(zip(trajectories, lang_vectors, strict=False), desc="Processing trajectories")
+        ):
+            # Create output directory for this trajectory with subdirectory structure
+            subdir_name = get_trajectory_subdir_path(trajectory_idx)
+            trajectory_dir = os.path.join(
+                output_dir, dataset_name.lower(), subdir_name, f"trajectory_{trajectory_idx:04d}.mp4"
+            )
+            os.makedirs(os.path.dirname(trajectory_dir), exist_ok=True)
+
+            processed_trajectory = hf_creator_fn(
+                traj_dict=trajectory,
+                video_path=trajectory_dir,
+                lang_vector=lang_vector,  # Pre-computed language vector
+                max_frames=max_frames,
+                dataset_name=dataset_name,
+                use_video=use_video,
+                fps=fps,
+            )
+            if processed_trajectory is None:
+                continue
+            all_entries.append(processed_trajectory)
+    else:
+        # Parallel processing
+        all_entries = []  # ensure defined if Pool raises before we filter results
+        print(f"Preparing {len(trajectories)} trajectories for parallel processing...")
+
+        # Prepare arguments for worker processes
+        worker_args = []
+        for trajectory_idx, (trajectory, lang_vector) in enumerate(zip(trajectories, lang_vectors, strict=False)):
+            args = (
+                trajectory_idx,
+                trajectory,
+                lang_vector,  # Pre-computed language vector
+                hf_creator_fn,
+                output_dir,
+                dataset_name,
+                max_frames,
+                use_video,
+                fps,
+            )
+            worker_args.append(args)
+
+        # Use spawn to avoid CUDA context issues from forking after TF import
+        try:
+            mp.set_start_method("spawn", force=True)
+        except RuntimeError:
+            pass
+
+        # Process trajectories in parallel
+        with Pool(processes=num_workers) as pool:
+            results = list(
+                tqdm(
+                    pool.imap_unordered(process_single_trajectory, worker_args),
+                    total=len(worker_args),
+                    desc="Processing trajectories",
+                )
+            )
+
+        # Filter out failed trajectories (None results)
+        all_entries = [result for result in results if result is not None]
+
+        if len(all_entries) < len(trajectories):
+            failed_count = len(trajectories) - len(all_entries)
+            print(f"⚠️  {failed_count} trajectories failed to process and were skipped")
+
+    # Create HuggingFace dataset with proper features
+    print(f"Creating HuggingFace dataset with {len(all_entries)} entries...")
+
+    # Convert list of entries to dictionary format for from_dict()
+    data_dict = {
+        "id": [entry["id"] for entry in all_entries],
+        "task": [entry["task"] for entry in all_entries],
+        "lang_vector": [entry["lang_vector"] for entry in all_entries],
+        "data_source": [entry["data_source"] for entry in all_entries],
+        "frames": [entry["frames"] for entry in all_entries],
+        "is_robot": [entry["is_robot"] for entry in all_entries],
+        "quality_label": [entry.get("quality_label") for entry in all_entries],
+        "partial_success": [entry.get("partial_success") for entry in all_entries],
+        # "preference_group_id": [entry.get("preference_group_id") for entry in all_entries],
+        # "preference_rank": [entry.get("preference_rank") for entry in all_entries],
+    }
+
+    # Set frames feature based on video mode
+    features_dict = BASE_FEATURES.copy()
+    if use_video:
+        features_dict["frames"] = datasets.Value("string")  # Video file paths as strings
+    else:
+        features_dict["frames"] = datasets.Sequence(datasets.Image())
+
+    features = datasets.Features(features_dict)
+    dataset = Dataset.from_dict(data_dict, features=features)
+
+    print(f"{dataset_name} HuggingFace dataset created successfully!")
+    print(f"Total entries: {len(all_entries)}")
+
+    # Push to HuggingFace Hub if requested
+    if push_to_hub and hub_repo_id:
+        print(f"\nPushing dataset to HuggingFace Hub: {hub_repo_id}")
+        try:
+            # Push the dataset to the hub with dataset name as config name
+            dataset.push_to_hub(
+                hub_repo_id,
+                config_name=dataset_name.lower(),  # Use dataset name as config name
+                token=hub_token,
+                private=False,
+                commit_message=f"Add {dataset_name} dataset for Robometer training",
+            )
+            print(f"✅ Successfully pushed dataset to: https://huggingface.co/datasets/{hub_repo_id}")
+            print(f"📁 Dataset available as config: {dataset_name.lower()}")
+
+            # Also push the video files folder to the hub
+            print("\nPushing video files to HuggingFace Hub...")
+            from huggingface_hub import HfApi
+
+            api = HfApi(token=hub_token)
+
+            # Upload the entire output directory (which contains all the video files)
+            api.upload_large_folder(
+                folder_path=output_dir,
+                repo_id=hub_repo_id,
+                repo_type="dataset",
+                # commit_message=f"Add video files for {dataset_name} dataset"
+            )
+            print(f"✅ Successfully pushed video files to: https://huggingface.co/datasets/{hub_repo_id}")
+
+        except Exception as e:
+            print(f"❌ Error pushing to hub: {e}")
+            print("Dataset was created locally but failed to push to hub")
+    elif push_to_hub and not hub_repo_id:
+        print("❌ push_to_hub=True but no hub_repo_id provided")
+    else:
+        # Only save locally if not pushing to hub (to avoid redundant Arrow files)
+        dataset_path = os.path.join(output_dir, dataset_name.lower())
+        dataset.save_to_disk(dataset_path)
+        print(f"Dataset saved locally to: {dataset_path}")
+
+    return dataset
+
+
+@wrap()
+def main(cfg: GenerateConfig):
+    """Main function to convert any dataset to HuggingFace format."""
+
+    # Get hub token from environment if not provided
+    if cfg.hub.hub_token is None:
+        cfg.hub.hub_token = os.getenv("HF_TOKEN")
+
+    # Only require HF_USERNAME if pushing to hub
+    if cfg.hub.push_to_hub:
+        username = os.getenv("HF_USERNAME")
+        if not username:
+            raise ValueError(
+                "HF_USERNAME is not set. Please export it to push to the Hub, or set hub.push_to_hub=false."
+            )
+        if cfg.hub.hub_repo_id:
+            cfg.hub.hub_repo_id = username + "/" + cfg.hub.hub_repo_id
+
+    # Import the appropriate dataset loader and trajectory creator
+    if "libero" in cfg.dataset.dataset_name:
+        from dataset_upload.dataset_loaders.libero_loader import load_libero_dataset
+
+        # Load the trajectories using the loader
+        task_data = load_libero_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "agibotworld" in (cfg.dataset.dataset_name or "").lower():
+        # Stream + convert directly inside the AgiBotWorld loader
+        from dataset_upload.dataset_loaders.agibotworld_loader import (
+            convert_agibotworld_streaming_to_hf,
+        )
+
+        dataset = convert_agibotworld_streaming_to_hf(
+            dataset_name=cfg.dataset.dataset_path,
+            output_dir=cfg.output.output_dir,
+            dataset_label=cfg.dataset.dataset_name or "agibotworld",
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                # Push the arrow table
+                dataset.push_to_hub(
+                    cfg.hub.hub_repo_id,
+                    config_name=(cfg.dataset.dataset_name or "agibotworld").lower(),
+                    token=cfg.hub.hub_token,
+                    private=False,
+                    commit_message=f"Add {cfg.dataset.dataset_name} dataset for Robometer training",
+                )
+                print(f"✅ Successfully pushed dataset to: https://huggingface.co/datasets/{cfg.hub.hub_repo_id}")
+
+                # Push the large video folder(s)
+                print("\nPushing video files to HuggingFace Hub...")
+                from huggingface_hub import HfApi
+
+                api = HfApi(token=cfg.hub.hub_token)
+                api.upload_large_folder(
+                    folder_path=cfg.output.output_dir,
+                    repo_id=cfg.hub.hub_repo_id,
+                    repo_type="dataset",
+                )
+                print(f"✅ Successfully pushed video files to: https://huggingface.co/datasets/{cfg.hub.hub_repo_id}")
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name or "agibotworld").lower())
+            dataset.save_to_disk(dataset_path)
+            print(f"Dataset saved locally to: {dataset_path}")
+        print("Dataset conversion complete!")
+        return
+
+    elif "egodex" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.egodex_loader import load_egodex_dataset
+
+        # Load the trajectories using the loader with max_trajectories limit
+        print(f"Loading EgoDex dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_egodex_dataset(
+            cfg.dataset.dataset_path,
+            cfg.output.max_trajectories,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif cfg.dataset.dataset_name.lower().startswith("oxe_"):
+        # Treat OXE like AgiBotWorld: create videos and HF entries directly in the loader
+        os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+        from dataset_upload.dataset_loaders.oxe_loader import convert_oxe_dataset_to_hf
+
+        print(f"Converting OXE dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_oxe_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path)
+            print(f"Dataset saved locally to: {dataset_path}")
+        print("Dataset conversion complete!")
+        return
+    elif "robofail" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.robofail_loader import load_robofail_dataset
+
+        # Load the trajectories using the loader with max_trajectories limit
+        print(f"Loading RoboFail dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_robofail_dataset(
+            cfg.dataset.dataset_path,
+            cfg.output.max_trajectories,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "metaworld" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.mw_collected_loader import load_metaworld_dataset
+
+        # Load the trajectories using the loader with max_trajectories limit
+        print(f"Loading metaworld dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_metaworld_dataset(
+            cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "h2r" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly inside the H2R loader (OXE-style)
+        from dataset_upload.dataset_loaders.h2r_loader import convert_h2r_dataset_to_hf
+
+        print(f"Converting H2R dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_h2r_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path)
+            print(f"Dataset saved locally to: {dataset_path}")
+        print("Dataset conversion complete!")
+        return
+    elif "fino_net" in cfg.dataset.dataset_name.lower() or "fino-net" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly inside the FinoNet loader (H2R/OXE-style)
+        from dataset_upload.dataset_loaders.fino_net_loader import convert_fino_net_dataset_to_hf
+
+        print(f"Converting FinoNet dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_fino_net_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path)
+            print(f"Dataset saved locally to: {dataset_path}")
+        print("Dataset conversion complete!")
+        return
+    elif "epic" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly (H2R/OXE-style)
+        from dataset_upload.dataset_loaders.epic_loader import convert_epic_dataset_to_hf
+
+        print(f"Converting EPIC-KITCHENS dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_epic_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+            shortest_edge_size=cfg.output.shortest_edge_size,
+            center_crop=cfg.output.center_crop,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "roboarena" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.roboarena_loader import load_roboarena_dataset
+
+        # Load the trajectories using the loader with max_trajectories limit
+        print(f"Loading RoboArena dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_roboarena_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "ph2d" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.ph2d_loader import load_ph2d_dataset
+
+        print(f"Loading Ph2d dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_ph2d_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "galaxea" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly (OXE-style, multi-dataset)
+        from dataset_upload.dataset_loaders.galaxea_loader import convert_galaxea_dataset_to_hf
+
+        rlds_datasets = getattr(cfg.dataset, "rlds_datasets", []) or []
+        print(f"Converting Galaxea RLDS to HF from: {cfg.dataset.dataset_path} | datasets={rlds_datasets}")
+        dataset = convert_galaxea_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                # remove the galaxea_rfm prefix from the dataset name because otherwise it won't match the video folder name
+                # don't need to do this for OXE or others because I processed it in their loaders but forgot for this.
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "molmoact" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly (LeRobot parquet)
+        from dataset_upload.dataset_loaders.molmoact_loader import convert_molmoact_dataset_to_hf
+
+        print(f"Converting MolmoAct dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_molmoact_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+        )
+
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "auto_eval" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.autoeval_loader import load_autoeval_dataset
+
+        print(f"Loading AutoEval dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_autoeval_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "usc_xarm_policy_ranking" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.usc_xarm_policy_ranking_loader import (
+            load_usc_xarm_policy_ranking_dataset,
+        )
+
+        print(f"Loading USC xArm Policy Ranking dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_usc_xarm_policy_ranking_dataset(
+            cfg.dataset.dataset_path,
+            max_trajectories=cfg.output.max_trajectories,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "usc_franka_policy_ranking" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.usc_franka_policy_ranking_loader import (
+            load_usc_franka_policy_ranking_dataset,
+        )
+
+        print(f"Loading USC Franka Policy Ranking dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_usc_franka_policy_ranking_dataset(
+            cfg.dataset.dataset_path,
+            max_trajectories=cfg.output.max_trajectories,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "utd_so101_policy_ranking" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.utd_so101_loader import (
+            load_utd_so101_dataset,
+        )
+
+        print(f"Loading UTD SO101 robot dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_utd_so101_dataset(
+            cfg.dataset.dataset_path,
+            max_trajectories=cfg.output.max_trajectories,
+            is_robot=True,
+            data_source="utd_so101",
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "utd_so101_human" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.utd_so101_loader import (
+            load_utd_so101_dataset,
+        )
+
+        print(f"Loading UTD SO101 human dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_utd_so101_dataset(
+            cfg.dataset.dataset_path,
+            max_trajectories=cfg.output.max_trajectories,
+            is_robot=False,
+            data_source="utd_so101_human",
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "soar" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.soar_loader import convert_soar_dataset_to_hf
+
+        print(f"Converting SOAR RLDS (local) to HF from: {cfg.dataset.dataset_path} ")
+        dataset = convert_soar_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "mit_franka_p-rank" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.mit_franka_prank_loader import convert_mit_franka_prank_dataset_to_hf
+
+        print(f"Converting MIT-Franka-Prank dataset to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_mit_franka_prank_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "rfm_new_mit_franka" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.new_mit_franka_loader import convert_new_mit_franka_dataset_to_hf
+
+        print(f"Converting New MIT Franka dataset to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_new_mit_franka_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+            exclude_wrist_cam=cfg.dataset.exclude_wrist_cam,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "utd_so101_clean_policy_ranking" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.utd_so101_clean_policy_ranking_loader import (
+            convert_utd_so101_clean_policy_ranking_to_hf,
+        )
+
+        # Determine view from dataset name
+        if "wrist" in cfg.dataset.dataset_name.lower():
+            view = "wrist"
+        elif "top" in cfg.dataset.dataset_name.lower():
+            view = "top"
+        else:
+            raise ValueError(f"Dataset name must specify view (wrist or top): {cfg.dataset.dataset_name}")
+
+        print(f"Converting UTD SO101 Clean Policy Ranking ({view} view) to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_utd_so101_clean_policy_ranking_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            view=view,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "usc_koch_human_robot_paired" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.usc_koch_human_robot_paired_loader import (
+            convert_usc_koch_human_robot_paired_to_hf,
+        )
+
+        # Determine trajectory type from dataset name
+        if "usc_koch_human_robot_paired_human" in cfg.dataset.dataset_name.lower():
+            trajectory_type = "human"
+        elif "usc_koch_human_robot_paired_robot" in cfg.dataset.dataset_name.lower():
+            trajectory_type = "robot"
+        else:
+            raise ValueError(
+                f"Dataset name must specify either 'usc_koch_human_robot_paired_human' or 'usc_koch_human_robot_paired_robot': {cfg.dataset.dataset_name}. "
+            )
+
+        print(f"Converting USC Koch Human-Robot Paired ({trajectory_type}) to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_usc_koch_human_robot_paired_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            trajectory_type=trajectory_type,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            updated_repo_id = cfg.hub.hub_repo_id.replace("usc_koch_human_robot_paired_", "")
+            print(f"\nPushing dataset to HuggingFace Hub: {updated_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, updated_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(output_dir_override, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "usc_koch_p_ranking" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.usc_koch_p_ranking_loader import (  # type: ignore
+            convert_usc_koch_p_ranking_to_hf,
+        )
+
+        output_dir_override = os.path.join(os.path.dirname(cfg.output.output_dir), cfg.dataset.dataset_name.lower())
+
+        print(f"Converting USC Koch P-Ranking to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_usc_koch_p_ranking_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=output_dir_override,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            updated_repo_id = cfg.hub.hub_repo_id.replace("usc_koch_p_ranking_rfm", "")
+            print(f"\nPushing dataset to HuggingFace Hub: {updated_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, updated_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, output_dir_override
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push metadata to hub")
+        else:
+            dataset_path_local = os.path.join(output_dir_override, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path_local)
+            print(f"Dataset saved locally to: {dataset_path_local}")
+        print("Dataset conversion complete!")
+        return
+    elif "egocot" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.egocot_loader import load_egocot_dataset
+
+        # Load the trajectories using the loader
+        print(f"Loading EgoCoT dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_egocot_dataset(
+            cfg.dataset.dataset_path,
+        )
+        trajectories = flatten_task_data(task_data)
+    elif "humanoid_everyday" in cfg.dataset.dataset_name.lower():
+        # Stream + convert directly (OXE-style)
+        from dataset_upload.dataset_loaders.humanoid_everyday_loader import convert_humanoid_everyday_dataset_to_hf
+
+        print(f"Converting Humanoid Everyday dataset directly to HF from: {cfg.dataset.dataset_path}")
+        dataset = convert_humanoid_everyday_dataset_to_hf(
+            dataset_path=cfg.dataset.dataset_path,
+            dataset_name=cfg.dataset.dataset_name,
+            output_dir=cfg.output.output_dir,
+            max_trajectories=cfg.output.max_trajectories,
+            max_frames=cfg.output.max_frames,
+            fps=cfg.output.fps,
+            num_workers=cfg.output.num_workers,
+        )
+
+        # Handle pushing/saving consistently
+        if cfg.hub.push_to_hub and cfg.hub.hub_repo_id:
+            print(f"\nPushing dataset to HuggingFace Hub: {cfg.hub.hub_repo_id}")
+            try:
+                push_hf_dataset_and_video_files_to_hub(
+                    dataset, cfg.hub.hub_repo_id, cfg.hub.hub_token, cfg.dataset.dataset_name, cfg.output.output_dir
+                )
+            except Exception as e:
+                print(f"❌ Error pushing to hub: {e}")
+                print("Dataset was created locally but failed to push videos and/or metadata to hub")
+        else:
+            dataset_path = os.path.join(cfg.output.output_dir, (cfg.dataset.dataset_name).lower())
+            dataset.save_to_disk(dataset_path)
+            print(f"Dataset saved locally to: {dataset_path}")
+        print("Dataset conversion complete!")
+        return
+    elif "motif" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.motif_loader import load_motif_dataset
+
+        print(f"Loading MotIF dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_motif_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "failsafe" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.failsafe_loader import load_failsafe_dataset
+
+        print(f"Loading FailSafe dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_failsafe_dataset(cfg.dataset.dataset_path)
+        trajectories = flatten_task_data(task_data)
+    elif "racer" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.racer_loader import load_racer_dataset
+
+        print(f"Loading RACER dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_racer_dataset(cfg.dataset.dataset_path, cfg.dataset.dataset_name)
+        trajectories = flatten_task_data(task_data)
+    elif "hand_paired" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.hand_paired_loader import load_hand_paired_dataset
+
+        print(f"Loading HAND_paired dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_hand_paired_dataset(cfg.dataset.dataset_path, cfg.dataset.dataset_name)
+        trajectories = flatten_task_data(task_data)
+    elif "roboreward" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.roboreward_loader import load_roboreward_dataset
+
+        print(f"Loading RoboReward dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_roboreward_dataset(cfg.dataset.dataset_path, cfg.dataset.dataset_name)
+        trajectories = flatten_task_data(task_data)
+    elif "robofac" in cfg.dataset.dataset_name.lower():
+        from dataset_upload.dataset_loaders.robofac_loader import load_robofac_dataset
+
+        print(f"Loading RoboFAC dataset from: {cfg.dataset.dataset_path}")
+        task_data = load_robofac_dataset(
+            cfg.dataset.dataset_path,
+            max_trajectories=cfg.output.max_trajectories,
+        )
+        trajectories = flatten_task_data(task_data)
+    else:
+        raise ValueError(f"Unknown dataset type: {cfg.dataset.dataset_name}")
+
+    # Convert dataset (non-streaming datasets)
+    convert_dataset_to_hf_format(
+        trajectories=trajectories,
+        hf_creator_fn=partial(
+            create_hf_trajectory,
+            dataset_name=cfg.dataset.dataset_name,
+            use_video=cfg.output.use_video,
+            fps=cfg.output.fps,
+            shortest_edge_size=cfg.output.shortest_edge_size,
+            center_crop=cfg.output.center_crop,
+            hub_repo_id=cfg.hub.hub_repo_id,
+        ),
+        output_dir=cfg.output.output_dir,
+        dataset_name=cfg.dataset.dataset_name,
+        max_trajectories=cfg.output.max_trajectories,
+        max_frames=cfg.output.max_frames,
+        use_video=cfg.output.use_video,
+        fps=cfg.output.fps,
+        num_workers=cfg.output.num_workers,
+        push_to_hub=cfg.hub.push_to_hub,
+        hub_repo_id=cfg.hub.hub_repo_id,
+        hub_token=cfg.hub.hub_token,
+    )
+
+    print("Dataset conversion complete!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/dataset_upload/helpers.py b/robometer/dataset_upload/helpers.py
new file mode 100755
index 0000000000000000000000000000000000000000..94b9103863d6cd4c3ce8fb2ee2f043caa332dd11
--- /dev/null
+++ b/robometer/dataset_upload/helpers.py
@@ -0,0 +1,404 @@
+#!/usr/bin/env python3
+"""
+Helper functions for Robometer model dataset conversion.
+Contains utility functions for processing frames, saving images, and managing data.
+"""
+
+import os
+import subprocess as sp
+import uuid
+
+import cv2
+import numpy as np
+from PIL import Image
+from sentence_transformers import SentenceTransformer
+
+
+def save_frame_as_image(frame_data: np.ndarray, output_path: str) -> str:
+    """Save a frame as a JPG image."""
+    # Convert from HDF5 format to PIL Image
+    if frame_data.dtype != np.uint8:
+        frame_data = (frame_data * 255).astype(np.uint8)
+
+    image = Image.fromarray(frame_data)
+    image.save(output_path, "JPEG", quality=95)
+    return output_path
+
+
+def downsample_frames(frames: np.ndarray | list, max_frames: int = 32) -> np.ndarray | list:
+    """Downsample frames to at most max_frames using linear interpolation."""
+    # If max_frames is -1, don't downsample
+    if max_frames == -1:
+        return frames
+
+    if len(frames) <= max_frames:
+        return frames
+
+    # Use linear interpolation to downsample
+    indices = np.linspace(0, len(frames) - 1, max_frames, dtype=int)
+
+    # keep unique frames
+    unique_indices = np.unique(indices)
+
+    # Handle both list and numpy array inputs
+    if isinstance(frames, list):
+        return [frames[i] for i in unique_indices]
+    else:
+        return frames[unique_indices]
+
+
+def motion_aware_downsample(frames: np.ndarray, max_frames: int = 32) -> np.ndarray:
+    if len(frames) <= max_frames:
+        return frames
+    T = len(frames)
+    resize_long_side = 256
+    min_gap = 1
+
+    def _prep(f):
+        if resize_long_side:
+            h, w = f.shape[:2]
+            scale = resize_long_side / max(h, w)
+            if scale < 1.0:
+                f = cv2.resize(f, (int(w * scale), int(h * scale)), interpolation=cv2.INTER_AREA)
+        return cv2.cvtColor(f, cv2.COLOR_BGR2GRAY).astype(np.float32)
+
+    gray = [_prep(f) for f in frames]
+
+    scores = np.zeros(T, dtype=np.float32)
+    fb_args = {
+        "pyr_scale": 0.5,
+        "levels": 3,
+        "winsize": 15,
+        "iterations": 3,
+        "poly_n": 5,
+        "poly_sigma": 1.2,
+        "flags": 0,
+    }
+    for i in range(T - 1):
+        flow = cv2.calcOpticalFlowFarneback(gray[i], gray[i + 1], None, **fb_args)
+        scores[i + 1] = np.linalg.norm(flow, axis=-1).mean()
+
+    keep = {0, T - 1}
+    if max_frames > 2:
+        for idx in np.argsort(scores)[::-1]:
+            if len(keep) >= max_frames:
+                break
+            if all(abs(idx - k) >= min_gap for k in keep):
+                keep.add(int(idx))
+
+    return frames[sorted(keep)]
+
+
+def create_trajectory_video(
+    frames,
+    output_dir: str,
+    max_frames: int = -1,
+    fps: int = 10,
+    shortest_edge_size: int = 240,
+    center_crop: bool = False,
+) -> str:
+    """Create a trajectory video from frames and save as MP4 file."""
+    # Handle numpy array of frames (traditional case)
+    if not isinstance(frames, np.ndarray):
+        frames = np.array(frames)
+
+    # Downsample frames
+    frames = downsample_frames(frames, max_frames)
+
+    # Get video dimensions from first frame
+    if len(frames) == 0:
+        raise ValueError("No frames provided for video creation")
+
+    height, width = frames[0].shape[:2]
+
+    # First, optionally center crop to min(height, width)
+    if center_crop:
+        # Calculate crop coordinates for center crop
+        crop_h = min(height, width)
+        y_start = max((height - crop_h) // 2, 0)
+        x_start = max((width - crop_h) // 2, 0)
+        frames = frames[y_start : y_start + crop_h, x_start : x_start + crop_h]
+        height, width = frames[0].shape[:2]
+
+    # Figure out target dimensions for all frames
+    if height != width:
+        scale_factor = shortest_edge_size / min(height, width)
+        target_height = int(height * scale_factor)
+        target_width = int(width * scale_factor)
+    else:
+        target_height = height
+        target_width = width
+
+    # Create sequence directory and video file path
+    video_path = os.path.join(output_dir, "trajectory.mp4")
+    print(f"Saving video to: {video_path}")
+
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+    video_writer = cv2.VideoWriter(video_path, fourcc, fps, (target_width, target_height))
+
+    if not video_writer.isOpened():
+        raise Exception("Could not create video writer with any codec")
+
+    # Write frames to video
+    for frame in frames:
+        # Ensure frame is in uint8 format
+        if frame.dtype != np.uint8:
+            frame = (frame * 255).astype(np.uint8)
+
+        # Resize frame to target dimensions if needed
+        if frame.shape[:2] != (target_height, target_width):
+            frame = cv2.resize(frame, (target_width, target_height), interpolation=cv2.INTER_AREA)
+
+        # Convert RGB to BGR for OpenCV
+        if len(frame.shape) == 3 and frame.shape[2] == 3:
+            frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+
+        video_writer.write(frame)
+
+    # Release video writer
+    video_writer.release()
+
+    return video_path
+
+
+def create_trajectory_video_optimized(
+    frames,
+    video_path: str,
+    max_frames: int = -1,
+    fps: int = 10,
+    shortest_edge_size: int = 240,
+    center_crop: bool = False,
+) -> str:
+    """
+    Creates a web-optimized trajectory video using a memory-efficient FFmpeg pipe.
+
+    Args:
+        frames (list or np.ndarray): A list or array of frames (as RGB, uint8 arrays).
+        output_dir (str): Directory to save the video.
+        max_frames (int): Maximum number of frames to include in the video.
+        fps (int): Frames per second for the output video.
+        shortest_edge_size (int): The target size for the shortest edge of the video.
+        center_crop (bool): If True, center crop frames to a square before resizing.
+
+    Returns:
+        str: The path to the created video file.
+    """
+    # print(f"Saving optimized video to: {video_path}")
+    if os.path.exists(video_path):
+        # print(f"Video already exists at: {video_path}, skipping video creation")
+        return video_path
+
+    # If frames is callable, call it to get the actual frames
+    if callable(frames):
+        frames = frames()  # Load frames on-demand
+    else:
+        frames = frames  # Already loaded frames (legacy datasets)
+
+    if frames is None:
+        return None
+    if len(frames) == 0:
+        raise ValueError("No frames provided for video creation")
+
+    # Downsample frames by selecting indices, which is memory-cheap
+    processed_frames = downsample_frames(frames, max_frames)
+
+    # Get dimensions from the first frame
+    first_frame = processed_frames[0]
+    height, width = first_frame.shape[:2]
+
+    # Determine crop and target dimensions before starting the loop
+    if center_crop:
+        crop_size = min(height, width)
+        y_start = max((height - crop_size) // 2, 0)
+        x_start = max((width - crop_size) // 2, 0)
+        # After cropping, the frame is a square
+        height, width = crop_size, crop_size
+
+    if shortest_edge_size is not None:
+        scale_factor = shortest_edge_size / min(height, width)
+        target_width = int(width * scale_factor)
+        target_height = int(height * scale_factor)
+
+        # Ensure dimensions are even, as required by some codecs like H.264
+        target_width = target_width if target_width % 2 == 0 else target_width + 1
+        target_height = target_height if target_height % 2 == 0 else target_height + 1
+    else:
+        target_height, target_width = height, width
+
+    # FFmpeg command for creating a web-optimized H.264 video
+    # This pipes raw video frames from stdin
+    command = [
+        "ffmpeg",
+        "-y",  # Overwrite output file if it exists
+        "-f",
+        "rawvideo",
+        "-vcodec",
+        "rawvideo",
+        "-s",
+        f"{target_width}x{target_height}",  # Final size of frames sent to pipe
+        "-pix_fmt",
+        "bgr24",  # OpenCV provides BGR frames
+        "-r",
+        str(fps),
+        "-i",
+        "-",  # Input comes from stdin
+        "-an",  # No audio
+        "-c:v",
+        "libx264",  # Use the H.264 codec
+        "-profile:v",
+        "high",
+        "-pix_fmt",
+        "yuv420p",  # Pixel format for maximum web compatibility
+        "-movflags",
+        "+faststart",  # CRITICAL: For web streaming
+        video_path,
+    ]
+
+    # Start the FFmpeg subprocess
+    process = sp.Popen(command, stdin=sp.PIPE, stdout=sp.PIPE, stderr=sp.PIPE)
+
+    # Check if process started successfully
+    if process.poll() is not None:
+        stderr = process.stderr.read().decode()
+        print(f"FFmpeg failed to start. Command: {' '.join(command)}")
+        print(f"Error: {stderr}")
+        raise RuntimeError("FFmpeg process failed to start")
+
+    for i, frame in enumerate(processed_frames):
+        # Ensure frame is in uint8 format
+        if frame.dtype != np.uint8:
+            frame = (frame * 255).astype(np.uint8)
+
+        # Apply transformations one frame at a time
+        if center_crop:
+            frame = frame[y_start : y_start + crop_size, x_start : x_start + crop_size]
+
+        # Resize frame to target dimensions
+        if frame.shape[0] != target_height or frame.shape[1] != target_width:
+            frame = cv2.resize(frame, (target_width, target_height), interpolation=cv2.INTER_AREA)
+
+        # Convert RGB to BGR for FFmpeg pipe
+        if len(frame.shape) == 3 and frame.shape[2] == 3:
+            frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+
+        # Write the raw frame data to the process's stdin
+        try:
+            process.stdin.write(frame.tobytes())
+        except BrokenPipeError as e:
+            stderr = process.stderr.read().decode()
+            print(f"BrokenPipeError writing frame. FFmpeg stderr: {stderr}")
+            print(f"Frame shape: {frame.shape}, dtype: {frame.dtype}")
+            raise RuntimeError(f"Failed to write frame to FFmpeg: {e}")
+
+    # Close the pipe and finish the process
+    process.stdin.close()
+    process.wait()
+
+    # Check for errors
+    stderr = process.stderr.read().decode()
+    if process.returncode != 0:
+        print("FFmpeg Error:")
+        print(stderr)
+        raise RuntimeError("FFmpeg process failed to encode the video.")
+
+    # print("Video created successfully.")
+    return video_path
+
+
+def create_trajectory_sequence(
+    frames: list[str], output_dir: str, sequence_name: str, max_frames: int = -1
+) -> list[str]:
+    """Create a trajectory sequence from frames and save as images."""
+
+    sequence_dir = os.path.join(output_dir, sequence_name)
+    os.makedirs(sequence_dir, exist_ok=True)
+
+    # Downsample frames
+    frames = downsample_frames(frames, max_frames)
+
+    frame_paths = []
+    for i, frame in enumerate(frames):
+        frame_path = os.path.join(sequence_dir, f"frame_{i:02d}.jpg")
+        saved_path = save_frame_as_image(frame, frame_path)
+        frame_paths.append(saved_path)
+
+    return frame_paths
+
+
+def generate_unique_id() -> str:
+    """Generate a unique UUID for dataset entries."""
+    return str(uuid.uuid4())
+
+
+def create_hf_trajectory(
+    traj_dict: dict,
+    video_path: str,
+    lang_vector: np.ndarray,
+    max_frames: int = -1,
+    dataset_name: str = "",
+    use_video: bool = True,
+    fps: int = 10,
+    shortest_edge_size: int = 240,
+    center_crop: bool = False,
+    hub_repo_id: str | None = None,
+) -> dict:
+    """Create a HuggingFace dataset trajectory with unified frame loading."""
+
+    # Handle frames - could be np.array, callable, or missing
+    frames_data = traj_dict.get("frames")
+    if frames_data is None:
+        raise ValueError("Trajectory must contain 'frames'")
+
+    video_path = create_trajectory_video_optimized(
+        frames_data, video_path, max_frames, fps, shortest_edge_size, center_crop
+    )
+
+    if video_path is None:
+        print(f"Skipping trajectory {traj_dict.get('id', 'UNKNOWN')} because frames are None")
+        return None
+
+    # Get identifiers and fields
+    id = traj_dict.get("id", generate_unique_id())
+    task_description = traj_dict["task"]
+    is_robot: bool = bool(traj_dict.get("is_robot", False))
+    quality_label: str = str(traj_dict.get("quality_label", "successful"))
+    preference_group_id = traj_dict.get("preference_group_id", None)
+    preference_rank = traj_dict.get("preference_rank", None)
+    partial_success = traj_dict.get("partial_success", None)
+    data_source = traj_dict.get("data_source", dataset_name)
+
+    # Create dataset trajectory
+    trajectory = {
+        "id": id,
+        "task": task_description,
+        "lang_vector": lang_vector,  # Pre-computed language vector
+        "data_source": data_source,
+        "frames": video_path,
+        "is_robot": is_robot,
+        "quality_label": quality_label,
+        "preference_group_id": preference_group_id,
+        "preference_rank": preference_rank,
+        "partial_success": partial_success,
+    }
+
+    return trajectory
+
+
+def load_sentence_transformer_model() -> SentenceTransformer:
+    """Load the sentence transformer model for language embeddings."""
+    return SentenceTransformer("all-MiniLM-L6-v2")
+
+
+def create_output_directory(output_dir: str) -> None:
+    """Create the output directory if it doesn't exist."""
+    os.makedirs(output_dir, exist_ok=True)
+
+
+def flatten_task_data(task_data: dict[str, list[dict]]) -> list[dict]:
+    """Flatten task data into a list of trajectories."""
+    all_trajectories = []
+    for task_name, trajectories in task_data.items():
+        for trajectory in trajectories:
+            trajectory["task_name"] = task_name
+            all_trajectories.append(trajectory)
+    return all_trajectories
diff --git a/robometer/dataset_upload/validate_dataset.py b/robometer/dataset_upload/validate_dataset.py
new file mode 100755
index 0000000000000000000000000000000000000000..b3a1c5cc46ae76e93a183400c2985a20c03465da
--- /dev/null
+++ b/robometer/dataset_upload/validate_dataset.py
@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+"""
+Simple validation script for the Robometer dataset format.
+Checks fields and data types only.
+"""
+
+import argparse
+from typing import Any
+
+import numpy as np
+
+from datasets import Dataset, load_from_disk
+
+
+def validate_dataset_fields_and_types(dataset: Dataset, sample_size: int = 10) -> dict[str, Any]:
+    """Validate dataset fields and data types."""
+
+    print(f"Validating dataset fields and data types on {sample_size} sample entries...")
+
+    validation_results = {
+        "valid": True,
+        "errors": [],
+        "warnings": [],
+        "stats": {"dataset_size": len(dataset), "samples_checked": 0},
+    }
+
+    # Expected schema for the new format
+    expected_fields = [
+        "id",
+        "task",
+        "lang_vector",
+        "data_source",
+        "frames",
+        "is_robot",
+        "quality_label",
+        "preference_group_id",
+        "preference_rank",
+    ]
+
+    # Check if dataset has features
+    if not hasattr(dataset, "features") or dataset.features is None:
+        validation_results["valid"] = False
+        validation_results["errors"].append("Dataset has no features defined")
+        return validation_results
+
+    print(f"Dataset size: {len(dataset)} entries")
+    print(f"Dataset features: {list(dataset.features.keys())}")
+
+    # Check required fields
+    for field_name in expected_fields:
+        if field_name not in dataset.features:
+            validation_results["valid"] = False
+            validation_results["errors"].append(f"Missing required field: {field_name}")
+        else:
+            print(f"✓ Field '{field_name}' present")
+
+    # Sample entries for validation
+    sample_indices = np.random.choice(len(dataset), min(sample_size, len(dataset)), replace=False)
+    validation_results["stats"]["samples_checked"] = len(sample_indices)
+
+    for idx in sample_indices:
+        trajectory = dataset[idx]
+
+        try:
+            # Validate each field
+            if not isinstance(trajectory["id"], str):
+                validation_results["errors"].append(f"Trajectory {idx}: 'id' is not a string")
+
+            if not isinstance(trajectory["task"], str):
+                validation_results["errors"].append(f"Trajectory {idx}: 'task' is not a string")
+
+            # lang_vector should be length-384 sequence
+            lv = trajectory["lang_vector"]
+            if isinstance(lv, np.ndarray):
+                if lv.shape != (384,):
+                    validation_results["errors"].append(
+                        f"Trajectory {idx}: 'lang_vector' shape is {lv.shape}, expected (384,)"
+                    )
+            elif isinstance(lv, list):
+                if len(lv) != 384:
+                    validation_results["errors"].append(
+                        f"Trajectory {idx}: 'lang_vector' length is {len(lv)}, expected 384"
+                    )
+                else:
+                    # check element types
+                    if not all(isinstance(x, (int, float, np.floating, np.integer)) for x in lv):
+                        validation_results["warnings"].append(
+                            f"Trajectory {idx}: 'lang_vector' contains non-numeric elements"
+                        )
+            else:
+                validation_results["errors"].append(f"Trajectory {idx}: 'lang_vector' has unexpected type {type(lv)}")
+
+            if not isinstance(trajectory["data_source"], str):
+                validation_results["errors"].append(f"Trajectory {idx}: 'data_source' is not a string")
+
+            if not isinstance(trajectory["frames"], str):
+                validation_results["errors"].append(f"Trajectory {idx}: 'frames' is not a string path")
+
+            if not isinstance(trajectory["is_robot"], bool):
+                validation_results["errors"].append(f"Trajectory {idx}: 'is_robot' is not a boolean")
+
+            if not isinstance(trajectory["quality_label"], str):
+                validation_results["errors"].append(f"Trajectory {idx}: 'quality_label' is not a string")
+            else:
+                if trajectory["quality_label"] not in {"successful", "failure", "suboptimal"}:
+                    validation_results["warnings"].append(
+                        f"Trajectory {idx}: 'quality_label' has unexpected value '{trajectory['quality_label']}'"
+                    )
+
+            # preference fields can be None
+            if trajectory.get("preference_group_id") is not None and not isinstance(
+                trajectory["preference_group_id"], str
+            ):
+                validation_results["errors"].append(
+                    f"Trajectory {idx}: 'preference_group_id' is neither None nor string"
+                )
+            if trajectory.get("preference_rank") is not None and not isinstance(trajectory["preference_rank"], int):
+                validation_results["errors"].append(f"Trajectory {idx}: 'preference_rank' is neither None nor int")
+
+            # Print sample task for first trajectory
+            if idx == sample_indices[0]:
+                print("\nSample task from first trajectory:")
+                print(f"  Task: {trajectory['task']}")
+                print(f"  ID: {trajectory['id']}")
+
+        except Exception as e:
+            validation_results["errors"].append(f"Trajectory {idx}: Error during validation: {e}")
+
+    if validation_results["errors"]:
+        validation_results["valid"] = False
+
+    return validation_results
+
+
+def print_validation_summary(validation_results: dict[str, Any]):
+    """Print validation summary."""
+
+    print("\n" + "=" * 50)
+    print("VALIDATION SUMMARY")
+    print("=" * 50)
+
+    status = "✅ PASS" if validation_results["valid"] else "❌ FAIL"
+    print(f"Status: {status}")
+
+    print(f"Dataset size: {validation_results['stats']['dataset_size']}")
+    print(f"Samples checked: {validation_results['stats']['samples_checked']}")
+
+    if validation_results.get("errors"):
+        print(f"\nErrors ({len(validation_results['errors'])}):")
+        for error in validation_results["errors"][:10]:  # Show first 10 errors
+            print(f"  - {error}")
+        if len(validation_results["errors"]) > 10:
+            print(f"  ... and {len(validation_results['errors']) - 10} more errors")
+
+    if validation_results.get("warnings"):
+        print(f"\nWarnings ({len(validation_results['warnings'])}):")
+        for warning in validation_results["warnings"][:5]:  # Show first 5 warnings
+            print(f"  - {warning}")
+        if len(validation_results["warnings"]) > 5:
+            print(f"  ... and {len(validation_results['warnings']) - 5} more warnings")
+
+    print("=" * 50)
+
+
+def main():
+    """Main validation function."""
+
+    parser = argparse.ArgumentParser(description="Validate dataset fields and data types")
+    parser.add_argument("dataset_path", help="Path to the HuggingFace dataset")
+    parser.add_argument("--sample-size", type=int, default=10, help="Number of samples to check")
+
+    args = parser.parse_args()
+
+    # Load dataset
+    print(f"Loading dataset from: {args.dataset_path}")
+    try:
+        dataset = load_from_disk(args.dataset_path)
+    except Exception as e:
+        print(f"Error loading dataset: {e}")
+        return
+
+    print("Dataset loaded successfully.")
+
+    # Run validation
+    validation_results = validate_dataset_fields_and_types(dataset, args.sample_size)
+
+    # Print summary
+    print_validation_summary(validation_results)
+
+    # Exit with error code if validation failed
+    if not validation_results["valid"]:
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/dataset_upload/video_helpers.py b/robometer/dataset_upload/video_helpers.py
new file mode 100755
index 0000000000000000000000000000000000000000..e090b4ddbbd1d40327de3163297a30d72e9e046c
--- /dev/null
+++ b/robometer/dataset_upload/video_helpers.py
@@ -0,0 +1,238 @@
+#!/usr/bin/env python3
+"""
+Video helper utilities for robust loading of videos with codec handling and
+backend fallbacks. Centralized here to be reused across dataset loaders.
+"""
+
+import gc
+import os
+import shutil
+import subprocess
+import tempfile
+import time
+from signal import SIGTERM
+from typing import Optional
+
+import cv2
+import numpy as np
+
+# Configuration constants
+DEFAULT_SUBPROCESS_TIMEOUT = 30  # seconds for ffprobe/ffmpeg operations
+DEFAULT_FRAME_READ_TIMEOUT = 60  # seconds for reading all frames from a video
+MAX_FRAMES_SANITY_CHECK = 10000  # maximum number of frames to prevent infinite loops
+
+
+def _ffprobe_codec_name(path: str, timeout: int = DEFAULT_SUBPROCESS_TIMEOUT) -> str | None:
+    """Return codec_name for the first video stream using ffprobe, or None on failure."""
+    if shutil.which("ffprobe") is None:
+        return None
+    try:
+        completed = subprocess.run(
+            [
+                "ffprobe",
+                "-v",
+                "error",
+                "-select_streams",
+                "v:0",
+                "-show_entries",
+                "stream=codec_name",
+                "-of",
+                "default=nk=1:nw=1",
+                path,
+            ],
+            capture_output=True,
+            check=True,
+            text=True,
+            timeout=timeout,
+        )
+        codec_name = completed.stdout.strip().splitlines()[0] if completed.stdout.strip() else None
+        return codec_name
+    except subprocess.TimeoutExpired:
+        print(f"Warning: ffprobe timed out after {timeout}s for {path}")
+        return None
+    except Exception as e:
+        print(f"Warning: ffprobe failed for {path}: {e}")
+        return None
+
+
+def _reencode_to_h264(input_path: str, timeout: int = DEFAULT_SUBPROCESS_TIMEOUT) -> str | None:
+    """Re-encode input video to H.264 yuv420p if ffmpeg is available. Returns output path or None."""
+    if shutil.which("ffmpeg") is None:
+        return None
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_out:
+        output_path = tmp_out.name
+    try:
+        subprocess.run(
+            [
+                "ffmpeg",
+                "-y",
+                "-loglevel",
+                "error",
+                "-i",
+                input_path,
+                "-c:v",
+                "libx264",
+                "-pix_fmt",
+                "yuv420p",
+                "-movflags",
+                "+faststart",
+                output_path,
+            ],
+            check=True,
+            timeout=timeout,
+        )
+        return output_path
+    except subprocess.TimeoutExpired:
+        print(f"Warning: ffmpeg re-encode timed out after {timeout}s for {input_path}")
+        try:
+            os.unlink(output_path)
+        except Exception:
+            pass
+        return None
+    except Exception as e:
+        print(f"Warning: ffmpeg re-encode failed for {input_path}: {e}")
+        try:
+            os.unlink(output_path)
+        except Exception:
+            pass
+        return None
+
+
+def _open_with_best_backend(path: str) -> cv2.VideoCapture | None:
+    """Try multiple OpenCV backends and return an opened capture or None."""
+    backends: list[int] = [getattr(cv2, "CAP_FFMPEG", cv2.CAP_ANY), cv2.CAP_ANY]
+    for backend in backends:
+        try:
+            cap_try = cv2.VideoCapture(path, backend)
+            if cap_try.isOpened():
+                # Validate we can read at least one frame
+                ret, test_frame = cap_try.read()
+                if ret and test_frame is not None:
+                    cap_try.set(cv2.CAP_PROP_POS_FRAMES, 0)
+                    return cap_try
+            cap_try.release()
+        except Exception:
+            try:
+                cap_try.release()
+            except Exception:
+                pass
+    return None
+
+
+def load_video_frames(
+    video_input,
+    timeout: Optional[int] = DEFAULT_FRAME_READ_TIMEOUT,
+    max_frames: int = MAX_FRAMES_SANITY_CHECK,
+) -> np.ndarray:
+    """Load video frames (RGB uint8) from a file path (str/Path) or video bytes.
+
+    - For byte inputs, detect AV1 and re-encode to H.264 for compatibility
+    - Uses OpenCV with FFMPEG backend when available
+    - Returns numpy array of shape (T, H, W, 3) in RGB order
+    - Protected by timeout and maximum frame count to prevent hangs
+
+    Args:
+        video_input: Path to video file or video bytes
+        timeout: Maximum time in seconds to spend reading frames (None to disable)
+        max_frames: Maximum number of frames to read (sanity check against infinite loops)
+
+    Raises:
+        TimeoutError: If frame reading exceeds the timeout
+        ValueError: If video cannot be opened or no frames extracted
+    """
+    temp_files_to_cleanup: list[str] = []
+    cap: cv2.VideoCapture | None = None
+    video_path_str = str(video_input) if isinstance(video_input, (str, os.PathLike)) else "bytes"
+
+    if isinstance(video_input, (str, os.PathLike)):
+        video_path = str(video_input)
+        decodable_path = video_path
+        # If file is AV1, re-encode to H.264 so OpenCV can decode (many platforms lack AV1)
+        codec_name = _ffprobe_codec_name(video_path)
+        if codec_name == "av1":
+            h264_path = _reencode_to_h264(video_path, timeout=120)
+            if h264_path is not None:
+                temp_files_to_cleanup.append(h264_path)
+                decodable_path = h264_path
+        cap = _open_with_best_backend(decodable_path)
+    else:
+        # Save bytes to temp file first
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
+            temp_file.write(video_input)
+            temp_input_path = temp_file.name
+        temp_files_to_cleanup.append(temp_input_path)
+
+        # If the input is AV1, transcode to H.264 for compatibility
+        codec_name = _ffprobe_codec_name(temp_input_path)
+        decodable_path = temp_input_path
+        if codec_name == "av1":
+            h264_path = _reencode_to_h264(temp_input_path)
+            if h264_path is not None:
+                temp_files_to_cleanup.append(h264_path)
+                decodable_path = h264_path
+        cap = _open_with_best_backend(decodable_path)
+
+    try:
+        frames: list[np.ndarray] = []
+        if cap is None or not cap.isOpened():
+            raise ValueError(
+                f"Could not open video file {video_path_str} with available backends. "
+                "If the source is AV1, install AV1 support or enable ffmpeg re-encode."
+            )
+
+        start_time = time.time()
+        frame_count = 0
+
+        while True:
+            # Check timeout
+            if timeout is not None and (time.time() - start_time) > timeout:
+                raise TimeoutError(
+                    f"Video frame reading timed out after {timeout}s for {video_path_str}. "
+                    f"Read {frame_count} frames before timeout."
+                )
+
+            # Check max frames sanity limit
+            if frame_count >= max_frames:
+                print(
+                    f"Warning: Reached maximum frame limit ({max_frames}) for {video_path_str}. "
+                    f"This might indicate a corrupted or unusually long video."
+                )
+                frames = []
+                gc.collect()
+                break
+
+            ret, frame = cap.read()
+            if not ret:
+                break
+
+            # Convert BGR to RGB for consistency
+            rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(rgb_frame)
+            frame_count += 1
+        cap.release()
+
+        # Clean up temp file(s) if we created any
+        for path in temp_files_to_cleanup:
+            try:
+                os.unlink(path)
+            except Exception:
+                pass
+
+        if len(frames) == 0:
+            raise ValueError(f"No frames could be extracted from video {video_path_str}")
+
+        return np.array(frames)
+
+    except Exception as e:
+        if cap is not None:
+            cap.release()
+        # Clean up temp files in case of error
+        for path in temp_files_to_cleanup:
+            try:
+                os.unlink(path)
+            except Exception:
+                pass
+        # Add more context to the error
+        if isinstance(e, (TimeoutError, ValueError)):
+            raise
+        raise RuntimeError(f"Error loading video {video_path_str}: {e}") from e
diff --git a/robometer/dataset_upload/visualize_dataset.py b/robometer/dataset_upload/visualize_dataset.py
new file mode 100755
index 0000000000000000000000000000000000000000..ea5ed5b2c83a34fb10b2db7b0b0aece59c671a11
--- /dev/null
+++ b/robometer/dataset_upload/visualize_dataset.py
@@ -0,0 +1,273 @@
+#!/usr/bin/env python3
+"""
+Dataset Visualization Script for AgiBotWorld datasets.
+Loads and displays information about the converted HuggingFace datasets.
+Saves both video frame visualizations (PNG) and actual video files (MP4).
+"""
+
+import argparse
+import os
+import tempfile
+from pathlib import Path
+
+import cv2
+import matplotlib.pyplot as plt
+import numpy as np
+
+from datasets import load_from_disk
+
+
+def _reencode_video_for_compatibility(video_bytes: bytes) -> bytes:
+    """Re-encode video with optimal settings for player compatibility."""
+    # Save video bytes to temporary file
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_input:
+        temp_input.write(video_bytes)
+        temp_input_path = temp_input.name
+
+    try:
+        # Open video with OpenCV
+        cap = cv2.VideoCapture(temp_input_path)
+
+        # Get video properties
+        fps = cap.get(cv2.CAP_PROP_FPS)
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+        # Create output file with compatibility-focused encoding
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_output:
+            temp_output_path = temp_output.name
+
+        # Try multiple H.264 codec options for maximum compatibility
+        codecs_to_try = ["avc1", "h264", "x264", "mp4v"]  # Order of preference
+        video_writer = None
+
+        for codec in codecs_to_try:
+            try:
+                fourcc = cv2.VideoWriter_fourcc(*codec)
+                video_writer = cv2.VideoWriter(temp_output_path, fourcc, fps, (width, height))
+                if video_writer.isOpened():
+                    break
+                video_writer.release()
+            except:
+                continue
+
+        if not video_writer or not video_writer.isOpened():
+            raise Exception("Could not initialize video writer with any codec")
+
+        # Copy all frames
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            video_writer.write(frame)
+
+        cap.release()
+        video_writer.release()
+
+        # Read the re-encoded video
+        with open(temp_output_path, "rb") as f:
+            compatible_bytes = f.read()
+
+        return compatible_bytes
+
+    finally:
+        # Clean up temporary files
+        try:
+            os.unlink(temp_input_path)
+            os.unlink(temp_output_path)
+        except:
+            pass
+
+
+def extract_video_frames(video_bytes: bytes, num_frames: int = 5) -> list:
+    """Extract frames from video bytes for visualization."""
+    # Save video bytes to temporary file
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
+        temp_file.write(video_bytes)
+        temp_path = temp_file.name
+
+    try:
+        # Open video with OpenCV
+        cap = cv2.VideoCapture(temp_path)
+        frames = []
+        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+        if total_frames == 0:
+            print("⚠️ Warning: Could not read frames from video")
+            return []
+
+        # Extract evenly spaced frames
+        frame_indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
+
+        for frame_idx in frame_indices:
+            cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
+            ret, frame = cap.read()
+            if ret:
+                # Convert BGR to RGB
+                frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+                frames.append(frame_rgb)
+
+        cap.release()
+        return frames
+
+    finally:
+        # Clean up temporary file
+        os.unlink(temp_path)
+
+
+def visualize_dataset(dataset_path: str, max_samples: int = 3):
+    """Visualize a HuggingFace dataset saved to disk."""
+    dataset_path = Path(dataset_path)
+
+    if not dataset_path.exists():
+        print(f"❌ Dataset not found: {dataset_path}")
+        return
+
+    print(f"📁 Loading dataset from: {dataset_path}")
+
+    try:
+        # Load dataset from disk
+        dataset = load_from_disk(str(dataset_path))
+        print(f"✅ Successfully loaded dataset with {len(dataset)} entries")
+
+        # Display dataset info
+        print("\n📊 Dataset Information:")
+        print(f"  - Number of samples: {len(dataset)}")
+        print(f"  - Features: {list(dataset.features.keys())}")
+
+        # Show sample data
+        print(f"\n🔍 Sample Data (showing up to {max_samples} samples):")
+
+        for i, sample in enumerate(dataset):
+            if i >= max_samples:
+                break
+
+            task_name = sample.get("task", "N/A")
+            print(f"\n--- Sample {i + 1} ---")
+            print(f"🎯 Task: {task_name}")
+            print(f"🤖 Is Robot: {sample.get('is_robot', 'N/A')}")
+            print(f"⭐ Optimal: {sample.get('optimal', 'N/A')}")
+
+            # Video info (check both 'video' and 'frames' fields)
+            video_data = sample.get("video", b"") or sample.get("frames", [])
+
+            # Handle case where frames is a list containing video bytes
+            if isinstance(video_data, list) and len(video_data) > 0:
+                video_bytes = video_data[0]  # Take first video from list
+            elif isinstance(video_data, bytes):
+                video_bytes = video_data
+            else:
+                video_bytes = b""
+
+            if video_bytes and isinstance(video_bytes, bytes):
+                print(f"Video size: {len(video_bytes):,} bytes ({len(video_bytes) / 1024 / 1024:.1f} MB)")
+
+                # Extract and display frames
+                frames = extract_video_frames(video_bytes, num_frames=3)
+                if frames:
+                    print(f"Extracted {len(frames)} frames from video")
+
+                    # Create visualization
+                    _fig, axes = plt.subplots(1, len(frames), figsize=(15, 5))
+                    if len(frames) == 1:
+                        axes = [axes]
+
+                    for j, (frame, ax) in enumerate(zip(frames, axes, strict=False)):
+                        ax.imshow(frame)
+                        ax.set_title(f"Frame {j + 1}")
+                        ax.axis("off")
+
+                    plt.suptitle(f"Sample {i + 1}: {task_name}", fontsize=14, wrap=True)
+                    plt.tight_layout()
+
+                    # Save visualization
+                    output_path = dataset_path.parent / f"sample_{i + 1}_frames.png"
+                    plt.savefig(output_path, dpi=150, bbox_inches="tight")
+                    print(f"🖼️ Saved frame visualization: {output_path}")
+                    plt.close()
+                else:
+                    print("⚠️ Could not extract frames from video")
+
+                # Save the actual video file with re-encoding for better compatibility
+                # Create safe filename from task name
+                safe_task_name = "".join(c if c.isalnum() or c in (" ", "-", "_") else "_" for c in task_name)
+                safe_task_name = safe_task_name.replace(" ", "_")[:50]  # Limit length
+                video_output_path = dataset_path.parent / f"sample_{i + 1}_{safe_task_name}.mp4"
+
+                # Re-encode video for maximum compatibility with video players/previews
+                try:
+                    re_encoded_bytes = _reencode_video_for_compatibility(video_bytes)
+                    with open(video_output_path, "wb") as f:
+                        f.write(re_encoded_bytes)
+                    print(f"🎬 Saved compatible video file: {video_output_path}")
+                except Exception as e:
+                    # Fallback: save original bytes
+                    print(f"⚠️ Re-encoding failed ({e}), saving original video bytes")
+                    with open(video_output_path, "wb") as f:
+                        f.write(video_bytes)
+                    print(f"🎬 Saved video file: {video_output_path}")
+            else:
+                print("❌ No video data found")
+
+            # Actions info
+            actions = sample.get("actions", [])
+            if actions:
+                actions_array = np.array(actions)
+                print(f"Actions shape: {actions_array.shape}")
+                print(f"Actions range: [{actions_array.min():.3f}, {actions_array.max():.3f}]")
+
+            # Text embedding info
+            text_embedding = sample.get("text_embedding", [])
+            if text_embedding:
+                embedding_array = np.array(text_embedding)
+                print(f"Text embedding shape: {embedding_array.shape}")
+
+        # Dataset statistics
+        print("\n📈 Dataset Statistics:")
+        tasks = [sample.get("task", "Unknown") for sample in dataset]
+        unique_tasks = list(set(tasks))
+        print(f"  - Unique tasks: {len(unique_tasks)}")
+        for task in unique_tasks:
+            count = tasks.count(task)
+            print(f"    • {task}: {count} samples")
+
+        # Video size statistics
+        video_sizes = []
+        for sample in dataset:
+            video_data = sample.get("video", b"") or sample.get("frames", [])
+
+            # Handle case where frames is a list containing video bytes
+            if isinstance(video_data, list) and len(video_data) > 0 and isinstance(video_data[0], bytes):
+                video_sizes.append(len(video_data[0]))
+            elif isinstance(video_data, bytes):
+                video_sizes.append(len(video_data))
+            else:
+                video_sizes.append(0)
+        if video_sizes:
+            total_size = sum(video_sizes)
+            avg_size = total_size / len(video_sizes)
+            print(f"  - Total video data: {total_size / 1024 / 1024:.1f} MB")
+            print(f"  - Average video size: {avg_size / 1024 / 1024:.1f} MB")
+            print(
+                f"  - Video size range: {min(video_sizes) / 1024 / 1024:.1f} - {max(video_sizes) / 1024 / 1024:.1f} MB"
+            )
+
+    except Exception as e:
+        print(f"❌ Error loading dataset: {e}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Visualize AgiBotWorld datasets")
+    parser.add_argument("dataset_path", help="Path to the saved dataset directory")
+    parser.add_argument("--max_samples", type=int, default=3, help="Maximum number of samples to visualize")
+
+    args = parser.parse_args()
+
+    print("🎬 AgiBotWorld Dataset Visualizer")
+    print("=" * 50)
+
+    visualize_dataset(args.dataset_path, args.max_samples)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/eval_commands/confusion_matrix.sh b/robometer/eval_commands/confusion_matrix.sh
new file mode 100644
index 0000000000000000000000000000000000000000..6d906758ecff1c3ef3edc478cfb78fcf2b944fb4
--- /dev/null
+++ b/robometer/eval_commands/confusion_matrix.sh
@@ -0,0 +1,80 @@
+# GVL
+export GEMINI_API_KEY="your-api-key-here"
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=gvl \
+    model_config.model_name=gemini-2.5-flash-lite \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=8
+
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=gvl \
+    model_config.provider=openai \
+    model_config.model_name=gpt-4o-mini \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=8
+
+# ReWIND
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rewind \
+    model_path=rewardfm/rewind-scale-rfm1M-32layers-8frame-20260118-180522 \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    custom_eval.use_frame_steps=false \
+    max_frames=8 \
+    model_config.batch_size=64
+
+# Robo-Dopamine (run with venv Python so vLLM is found; do not use uv run)
+.venv-robodopamine/bin/python robometer/evals/run_baseline_eval.py \
+    reward_model=robodopamine \
+    model_path=tanhuajie2001/Robo-Dopamine-GRM-3B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=64 \
+    model_config.batch_size=1
+
+# VLAC
+uv run --extra vlac --python .venv-vlac/bin/python python robometer/evals/run_baseline_eval.py \
+    reward_model=vlac \
+    model_path=InternRobotics/VLAC \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    custom_eval.pad_frames=false \
+    max_frames=64
+
+# RoboReward-8B
+# without koch
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=roboreward \
+    model_path=teetone/RoboReward-8B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=64
+
+# on all
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=roboreward \
+    model_path=teetone/RoboReward-8B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking,jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all]] \
+    max_frames=64
+
+# Robometer-4B
+# without koch
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking]] \
+    max_frames=8 \
+    model_config.batch_size=32
+
+# on all
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[confusion_matrix] \
+    custom_eval.confusion_matrix=[[aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking,jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all]] \
+    max_frames=8 \
+    model_config.batch_size=32
\ No newline at end of file
diff --git a/robometer/eval_commands/policy_ranking.sh b/robometer/eval_commands/policy_ranking.sh
new file mode 100644
index 0000000000000000000000000000000000000000..635bdebcf45b9ff748fa113be0951cf7f84938a2
--- /dev/null
+++ b/robometer/eval_commands/policy_ranking.sh
@@ -0,0 +1,65 @@
+# ReWIND
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rewind \
+    model_path=rewardfm/rewind-scale-rfm1M-32layers-8frame-20260118-180522 \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.num_examples_per_quality_pr=1000 \
+    max_frames=8 \
+    model_config.batch_size=64
+
+# Robo-Dopamine (run with venv Python so vLLM is found; do not use uv run)
+.venv-robodopamine/bin/python robometer/evals/run_baseline_eval.py \
+    reward_model=robodopamine \
+    model_path=tanhuajie2001/Robo-Dopamine-GRM-3B \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.num_examples_per_quality_pr=1000 \
+    max_frames=64 \
+    model_config.batch_size=1
+
+# VlAC
+uv run --extra vlac --python .venv-vlac/bin/python robometer/evals/run_baseline_eval.py \
+    reward_model=vlac \
+    model_path=InternRobotics/VLAC \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.pad_frames=false \
+    max_frames=64 \
+    custom_eval.num_examples_per_quality_pr=1000
+
+# RoboReward-4B
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=roboreward \
+    model_path=teetone/RoboReward-8B \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.pad_frames=false \
+    custom_eval.num_examples_per_quality_pr=1000 \
+    max_frames=64
+
+# Robometer-4B
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.num_examples_per_quality_pr=1000 \
+    max_frames=8 \
+    model_config.batch_size=32
+
+# Robometer-4B Libero Ablation
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B-LIBERO \
+    custom_eval.eval_types=[policy_ranking] \
+    custom_eval.policy_ranking=[libero_pi0] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.num_examples_per_quality_pr=20 \
+    max_frames=4 \
+    model_config.batch_size=32
\ No newline at end of file
diff --git a/robometer/eval_commands/reward_alignment.sh b/robometer/eval_commands/reward_alignment.sh
new file mode 100644
index 0000000000000000000000000000000000000000..33f27fd41576999aabbe9d918455ef24c9600e66
--- /dev/null
+++ b/robometer/eval_commands/reward_alignment.sh
@@ -0,0 +1,68 @@
+# ReWIND
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rewind \
+    model_path=rewardfm/rewind-scale-rfm1M-32layers-8frame-20260118-180522 \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=true \
+    custom_eval.subsample_n_frames=5 \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    max_frames=8 \
+    model_config.batch_size=64
+
+# Robo-Dopamine (run with venv Python so vLLM is found; do not use uv run)
+.venv-robodopamine/bin/python robometer/evals/run_baseline_eval.py \
+    reward_model=robodopamine \
+    model_path=tanhuajie2001/Robo-Dopamine-GRM-3B \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    max_frames=64 \
+    model_config.batch_size=1
+
+# VLAC
+uv run --extra vlac --python .venv-vlac/bin/python  robometer/evals/run_baseline_eval.py \
+    reward_model=vlac \
+    model_path=InternRobotics/VLAC \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    custom_eval.pad_frames=false \
+    max_frames=64
+
+# RoboReward-4B
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=roboreward \
+    model_path=teetone/RoboReward-4B \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=true \
+    custom_eval.subsample_n_frames=5 \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    max_frames=64 \
+    model_config.batch_size=32
+
+# Robometer-4B
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[rbm-1m-id,rbm-1m-ood] \
+    custom_eval.use_frame_steps=true \
+    custom_eval.subsample_n_frames=5 \
+    custom_eval.reward_alignment_max_trajectories=30 \
+    max_frames=4 \
+    model_config.batch_size=32
+
+# Robometer-4B Libero Ablation
+uv run python robometer/evals/run_baseline_eval.py \
+    reward_model=rbm \
+    model_path=aliangdw/Robometer-4B-LIBERO \
+    custom_eval.eval_types=[reward_alignment] \
+    custom_eval.reward_alignment=[libero_pi0] \
+    custom_eval.use_frame_steps=false \
+    custom_eval.reward_alignment_max_trajectories=100 \
+    max_frames=4 \
+    model_config.batch_size=32
\ No newline at end of file
diff --git a/robometer/outputs/2026-03-12/18-56-20/.hydra/config.yaml b/robometer/outputs/2026-03-12/18-56-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..97e25040d6d8a81c21ba5b1a749e2dcf5b62d7b4
--- /dev/null
+++ b/robometer/outputs/2026-03-12/18-56-20/.hydra/config.yaml
@@ -0,0 +1,35 @@
+reward_model: rbm
+model_config:
+  batch_size: 8
+model_path: aliangdw/Robometer-4B
+max_frames: 4
+custom_eval:
+  eval_types:
+  - reward_alignment
+  policy_ranking:
+  - aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking
+  - aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking
+  - aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking
+  - jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm
+  - jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top
+  - jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist
+  - jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all
+  confusion_matrix:
+  - aliangdw_metaworld_metaworld_eval
+  reward_alignment:
+  - rbm-1m-id
+  quality_preference:
+  - aliangdw_metaworld_metaworld_eval
+  comparisons_per_task: null
+  max_comparisons: 100
+  num_examples_per_quality_pr: null
+  num_partial_successes: 5
+  policy_ranking_max_tasks: 20
+  custom_eval_random_seed: 42
+  reward_alignment_max_trajectories: 3
+  use_frame_steps: true
+  subsample_n_frames: 5
+  pad_frames: true
+output_dir: ./baseline_eval_output/rbm_smoke
+server_url: 0.0.0.0
+server_port: 8001
diff --git a/robometer/outputs/2026-03-12/18-56-20/.hydra/hydra.yaml b/robometer/outputs/2026-03-12/18-56-20/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..af737013180db3c23771ec86b4d6b5ad8b8f35c2
--- /dev/null
+++ b/robometer/outputs/2026-03-12/18-56-20/.hydra/hydra.yaml
@@ -0,0 +1,165 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - reward_model=rbm
+    - model_path=aliangdw/Robometer-4B
+    - custom_eval.eval_types=[reward_alignment]
+    - custom_eval.reward_alignment=[rbm-1m-id]
+    - custom_eval.use_frame_steps=true
+    - custom_eval.subsample_n_frames=5
+    - custom_eval.reward_alignment_max_trajectories=3
+    - max_frames=4
+    - model_config.batch_size=8
+    - output_dir=./baseline_eval_output/rbm_smoke
+  job:
+    name: run_baseline_eval
+    chdir: null
+    override_dirname: custom_eval.eval_types=[reward_alignment],custom_eval.reward_alignment=[rbm-1m-id],custom_eval.reward_alignment_max_trajectories=3,custom_eval.subsample_n_frames=5,custom_eval.use_frame_steps=true,max_frames=4,model_config.batch_size=8,model_path=aliangdw/Robometer-4B,output_dir=./baseline_eval_output/rbm_smoke,reward_model=rbm
+    id: ???
+    num: ???
+    config_name: baseline_eval_config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer/robometer/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer/outputs/2026-03-12/18-56-20
+    choices:
+      reward_model: rbm
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/robometer/outputs/2026-03-12/18-56-20/.hydra/overrides.yaml b/robometer/outputs/2026-03-12/18-56-20/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..131505127941fe58b500111ebbe5c31070586210
--- /dev/null
+++ b/robometer/outputs/2026-03-12/18-56-20/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- reward_model=rbm
+- model_path=aliangdw/Robometer-4B
+- custom_eval.eval_types=[reward_alignment]
+- custom_eval.reward_alignment=[rbm-1m-id]
+- custom_eval.use_frame_steps=true
+- custom_eval.subsample_n_frames=5
+- custom_eval.reward_alignment_max_trajectories=3
+- max_frames=4
+- model_config.batch_size=8
+- output_dir=./baseline_eval_output/rbm_smoke
diff --git a/robometer/outputs/2026-03-12/18-56-20/run_baseline_eval.log b/robometer/outputs/2026-03-12/18-56-20/run_baseline_eval.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/robometer/outputs/2026-03-12/19-21-52/.hydra/config.yaml b/robometer/outputs/2026-03-12/19-21-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5b51e4f4a4181a1613a04226111c60047863b51d
--- /dev/null
+++ b/robometer/outputs/2026-03-12/19-21-52/.hydra/config.yaml
@@ -0,0 +1,35 @@
+reward_model: rbm
+model_config:
+  batch_size: 8
+model_path: aliangdw/Robometer-4B
+max_frames: 4
+custom_eval:
+  eval_types:
+  - reward_alignment
+  policy_ranking:
+  - aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking
+  - aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking
+  - aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking
+  - jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm
+  - jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top
+  - jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist
+  - jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all
+  confusion_matrix:
+  - aliangdw_metaworld_metaworld_eval
+  reward_alignment:
+  - jesbu1_racer_rfm_racer_val
+  quality_preference:
+  - aliangdw_metaworld_metaworld_eval
+  comparisons_per_task: null
+  max_comparisons: 100
+  num_examples_per_quality_pr: null
+  num_partial_successes: 5
+  policy_ranking_max_tasks: 20
+  custom_eval_random_seed: 42
+  reward_alignment_max_trajectories: 10
+  use_frame_steps: true
+  subsample_n_frames: 5
+  pad_frames: true
+output_dir: ./baseline_eval_output/rbm_smoke_10
+server_url: 0.0.0.0
+server_port: 8001
diff --git a/robometer/outputs/2026-03-12/19-21-52/.hydra/hydra.yaml b/robometer/outputs/2026-03-12/19-21-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e0e4b9ffb60d316696d251549f2cbc33f16c325f
--- /dev/null
+++ b/robometer/outputs/2026-03-12/19-21-52/.hydra/hydra.yaml
@@ -0,0 +1,165 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - reward_model=rbm
+    - model_path=aliangdw/Robometer-4B
+    - custom_eval.eval_types=[reward_alignment]
+    - custom_eval.reward_alignment=[jesbu1_racer_rfm_racer_val]
+    - custom_eval.use_frame_steps=true
+    - custom_eval.subsample_n_frames=5
+    - custom_eval.reward_alignment_max_trajectories=10
+    - max_frames=4
+    - model_config.batch_size=8
+    - output_dir=./baseline_eval_output/rbm_smoke_10
+  job:
+    name: run_baseline_eval
+    chdir: null
+    override_dirname: custom_eval.eval_types=[reward_alignment],custom_eval.reward_alignment=[jesbu1_racer_rfm_racer_val],custom_eval.reward_alignment_max_trajectories=10,custom_eval.subsample_n_frames=5,custom_eval.use_frame_steps=true,max_frames=4,model_config.batch_size=8,model_path=aliangdw/Robometer-4B,output_dir=./baseline_eval_output/rbm_smoke_10,reward_model=rbm
+    id: ???
+    num: ???
+    config_name: baseline_eval_config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer/robometer/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /wang_ssd/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer/outputs/2026-03-12/19-21-52
+    choices:
+      reward_model: rbm
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/robometer/outputs/2026-03-12/19-21-52/.hydra/overrides.yaml b/robometer/outputs/2026-03-12/19-21-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45c915ba1aacbbfae550e2bb58a7b08316012dd9
--- /dev/null
+++ b/robometer/outputs/2026-03-12/19-21-52/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- reward_model=rbm
+- model_path=aliangdw/Robometer-4B
+- custom_eval.eval_types=[reward_alignment]
+- custom_eval.reward_alignment=[jesbu1_racer_rfm_racer_val]
+- custom_eval.use_frame_steps=true
+- custom_eval.subsample_n_frames=5
+- custom_eval.reward_alignment_max_trajectories=10
+- max_frames=4
+- model_config.batch_size=8
+- output_dir=./baseline_eval_output/rbm_smoke_10
diff --git a/robometer/outputs/2026-03-12/19-21-52/run_baseline_eval.log b/robometer/outputs/2026-03-12/19-21-52/run_baseline_eval.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/robometer/processed_datasets/.gitattributes b/robometer/processed_datasets/.gitattributes
new file mode 100644
index 0000000000000000000000000000000000000000..1ef325f1b111266a6b26e0196871bd78baa8c2f3
--- /dev/null
+++ b/robometer/processed_datasets/.gitattributes
@@ -0,0 +1,59 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text
diff --git a/robometer/pyproject.toml b/robometer/pyproject.toml
new file mode 100755
index 0000000000000000000000000000000000000000..18e2f930f51e284c8a2a7fd97cabeb3a38781ea0
--- /dev/null
+++ b/robometer/pyproject.toml
@@ -0,0 +1,184 @@
+# ==================================================================================
+# Project Metadata
+# Defines core information about the package.
+# ==================================================================================
+[project]
+name = "robometer"
+version = "0.0.1" # You can update this as you release new versions
+description = "Robometer: scalable reward modeling with progress and preference"
+readme = "README.md"
+requires-python = "==3.10.*" # Based on your GitHub repo's setup
+authors = [
+    { name = "Anthony Liang", email = "aliang80@usc.edu" }
+]
+
+# ==================================================================================
+# Dependencies
+# This section lists the direct dependencies of your project.
+# It is best practice to list only the packages your code directly imports,
+# not every package from your environment. Tools like 'uv' or 'pip' will
+# resolve and install the dependencies of these packages automatically.
+# ==================================================================================
+dependencies = [
+    "accelerate>=1.9.0",
+    "datasets==4.1.1",
+    "einops>=0.8.0",
+    "evaluate>=0.4.3",
+    "h5py>=3.10.0",
+    "huggingface-hub[hf-xet]>0.35",
+    "numpy>=1.21.0", # Compatible with modern PyTorch
+    "matplotlib>=3.5.0",
+    "peft",
+    "pyrallis>=0.3.0",
+    "safetensors>=0.6.0rc0",
+    "sentence-transformers>=2.0.0",
+    "tqdm>=4.67.1",
+    "wandb>=0.20.1",
+    "pyyaml>=6.0.0",
+    "rich>=14.0.0",
+    "soundfile>=0.13.1",
+    "ipdb>=0.13.13",
+    "hf_transfer>=0.1.0",
+    "gradio>=4.0.0",
+    "pillow>=9.0.0",
+    "ffmpeg",
+    "termcolor>=3.1.0",
+    "qwen-vl-utils[decord]==0.0.14",
+    "fastapi>=0.116.1",
+    "uvicorn>=0.35.0",
+    "requests>=2.32.4",
+    "google-generativeai",  # For RL-VLM-F baseline (Gemini API)
+    "scipy>=1.15.3",
+    "codetiming>=1.4.0",
+    "robosuite==1.4.0",
+    "imageio>=2.37.0",
+    "omegaconf>=2.3.0",
+    "opencv-python-headless>=4.8.0",
+    "opencv-contrib-python-headless>=4.8.0", # Additional OpenCV modules
+    "einops>=0.8.0",
+    "num2words",
+    "seaborn",
+    "bitsandbytes",
+    "torch==2.8.0",  # Required for Qwen models, xformers, and modern transformers. Use CUDA build on Linux via tool.uv.sources
+    "torchvision",
+    "moviepy",
+    "wandb[media]",
+    "torchcodec",
+    "decord>=0.6.0",
+    "unsloth>=2025.10",
+    "gymnasium>=1.2.2",
+    "metaworld>=3.0.0",
+    "loguru>=0.7.3",
+    "xformers==0.0.32.post2",
+    "tensorboard",
+    "requests",
+    "hydra-core",
+    "omegaconf",
+    "tensorflow-datasets==4.9.2", # just for oxe and other tfds dataset uploading
+    "tensorflow==2.19", # just for oxe and other tfds dataset uploading
+    "tf-keras>=2.19.0", # just for oxe and other tfds dataset uploading
+    "protobuf<3.21", # just for oxe and other tfds dataset uploading
+
+]
+
+[project.urls]
+Repository = "https://github.com/aliang8/robometer"
+
+# ==================================================================================
+# Optional Dependencies
+# These are groups of dependencies for specific purposes, like development or
+# running optional features. Install with `uv pip install .[dev]`
+# ==================================================================================
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0.0",
+    "ruff>=0.8.0",
+    "pre-commit>=4.0.0",
+    "ipykernel",
+    "ipywidgets",
+]
+quantization = [
+    "bitsandbytes>=0.46.1",  # Linux/Windows only - for model quantization
+]
+# Standard Robometer/training installation (includes transformers and trl)
+robometer = [
+    "transformers>=4.57",
+    "trl==0.20.0",
+]
+# VLAC baseline installation (conflicts with robometer versions - use separate environment)
+vlac = [
+    "transformers<4.52",  # Required by evo-vlac
+    "trl>=0.13,<0.17",  # Required by evo-vlac
+    "evo-vlac @ git+https://github.com/InternRobotics/VLAC.git",
+]
+
+# ==================================================================================
+# Build System
+# Specifies how your project should be packaged. Hatchling is a modern,
+# fast, and recommended build backend.
+# ==================================================================================
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["robometer"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+
+# ==================================================================================
+# Tool Configuration
+# Settings for various development tools like 'uv', 'ruff' (linter), etc.
+# ==================================================================================
+
+# Configuration for the 'uv' package manager
+[tool.uv]
+# Declare that robometer and vlac optional dependencies are incompatible
+# They have conflicting trl and transformers version requirements
+conflicts = [
+    [
+        { extra = "robometer" },
+        { extra = "vlac" },
+    ],
+]
+# This section could be used for advanced features like overriding dependencies
+# from your environment if needed, similar to the 'openpi' example.
+#override-dependencies = ["av<16.0.0"]
+
+# Ensure flash-attn builds against the same Torch that's in your env
+[tool.uv.extra-build-dependencies]
+flash-attn = [{ requirement = "torch", match-runtime = true }]
+
+# Tell flash-attn to use a prebuilt wheel instead of compiling CUDA kernels
+[tool.uv.extra-build-variables]
+flash-attn = { FLASH_ATTENTION_SKIP_CUDA_BUILD = "TRUE" }
+
+# Specify PyTorch CUDA index for Linux/Windows (CUDA 12.8 builds)
+[tool.uv.sources]
+torch = { index = "pytorch-cu128", marker = "sys_platform == 'linux' or sys_platform == 'win32'" }
+torchvision = { index = "pytorch-cu128", marker = "sys_platform == 'linux' or sys_platform == 'win32'" }
+
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128"
+explicit = true
+
+# Tool configuration for the 'ruff' linter and formatter
+[tool.ruff]
+line-length = 120
+target-version = "py310"
+fix = true
+unsafe-fixes = true  # <- allow deletions that might have side effects; optional, use with care
+
+[tool.ruff.lint]
+select = ["E", "W", "F", "I", "B", "C4", "UP", "RUF"]
+ignore = ["T201"]
+
+[tool.ruff.lint.isort]
+combine-as-imports = true
+split-on-trailing-comma = true
+
+[tool.ruff.format]
+preview = true
\ No newline at end of file
diff --git a/robometer/requirements-robodopamine.txt b/robometer/requirements-robodopamine.txt
new file mode 100644
index 0000000000000000000000000000000000000000..76799d041fbcdb05e005799568aedb0d3418881d
--- /dev/null
+++ b/robometer/requirements-robodopamine.txt
@@ -0,0 +1,54 @@
+# Reward-fm dependencies for Robo-Dopamine venv.
+# Install this file first: vLLM brings torch 2.9.x; we omit torch/torchvision/xformers here.
+vllm
+accelerate>=1.9.0
+datasets==4.1.1
+einops>=0.8.0
+evaluate>=0.4.3
+h5py>=3.10.0
+huggingface-hub[hf-xet]>0.35
+numpy>=1.21.0
+matplotlib>=3.5.0
+peft
+pyrallis>=0.3.0
+safetensors>=0.6.0rc0
+sentence-transformers>=2.0.0
+tqdm>=4.67.1
+wandb>=0.20.1
+pyyaml>=6.0.0
+rich>=14.0.0
+soundfile>=0.13.1
+ipdb>=0.13.13
+hf_transfer>=0.1.0
+gradio>=4.0.0
+pillow>=9.0.0
+termcolor>=3.1.0
+qwen-vl-utils[decord]==0.0.14
+fastapi>=0.116.1
+uvicorn>=0.35.0
+requests>=2.32.4
+google-generativeai
+scipy>=1.15.3
+codetiming>=1.4.0
+robosuite==1.4.0
+imageio>=2.37.0
+omegaconf>=2.3.0
+opencv-python-headless>=4.8.0
+opencv-contrib-python-headless>=4.8.0
+num2words
+seaborn
+bitsandbytes
+moviepy
+wandb[media]
+torchcodec
+decord>=0.6.0
+unsloth>=2025.10
+gymnasium>=1.2.2
+metaworld>=3.0.0
+loguru>=0.7.3
+tensorboard
+hydra-core
+tensorflow-datasets==4.9.2
+tensorflow==2.19
+tf-keras>=2.19.0
+protobuf<3.21
diff --git a/robometer/robometer.txt b/robometer/robometer.txt
new file mode 100644
index 0000000000000000000000000000000000000000..13672d2d89cee4b830055e03e9b656854e6dd673
--- /dev/null
+++ b/robometer/robometer.txt
@@ -0,0 +1,25 @@
+==================== INFERENCE COMMAND ====================
+cd /home/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer
+source .venv/bin/activate
+CUDA_VISIBLE_DEVICES=1 uv run python scripts/example_inference_local.py \
+  --model-path aliangdw/Robometer-4B \
+  --video scripts/example_videos/jaco_play_pick_up_green_cup.mp4 \
+  --task "Pick up green cup" \
+  --fps 3 \
+  --max-frames 512
+
+===================== BASELINE COMMAND ====================
+cd /home/vcj9002/jianshu/workspace/code_keliang/Robometer/robometer
+source .venv/bin/activate
+export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/processed_datasets
+CUDA_VISIBLE_DEVICES=1 uv run python robometer/evals/run_baseline_eval.py \
+  reward_model=rbm \
+  model_path=aliangdw/Robometer-4B \
+  custom_eval.eval_types=[reward_alignment] \
+  custom_eval.reward_alignment=[rbm-1m-id] \
+  custom_eval.use_frame_steps=true \
+  custom_eval.subsample_n_frames=5 \
+  custom_eval.reward_alignment_max_trajectories=3 \
+  max_frames=4 \
+  model_config.batch_size=8 \
+  output_dir=./baseline_eval_output/rbm_smoke
\ No newline at end of file
diff --git a/robometer/robometer/__init__.py b/robometer/robometer/__init__.py
new file mode 100755
index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc
--- /dev/null
+++ b/robometer/robometer/__init__.py
@@ -0,0 +1 @@
+
diff --git a/robometer/robometer/configs/baseline_eval_config.yaml b/robometer/robometer/configs/baseline_eval_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dda911b57dfa967a790175c5b784bba542b744a9
--- /dev/null
+++ b/robometer/robometer/configs/baseline_eval_config.yaml
@@ -0,0 +1,38 @@
+# @package _global_
+
+# Baseline Evaluation Configuration for RBM
+# This config is used for running baseline evaluations (GVL, RL-VLM-F) on datasets
+# Used by run_baseline_eval.py
+
+defaults:
+  - _self_
+  - reward_model: rlvlmf  # Options: rlvlmf, gvl, vlac, rbm, rewind, roboreward, robodopamine
+
+# Shared configuration
+# model_path is set in reward_model/*.yaml files, but can be overridden here
+# model_path: null  # Override model path if needed (HuggingFace repo ID or local path)
+max_frames: 32  # Maximum frames for models (used by GVL and potentially others)
+
+# Custom evaluation configuration
+# This specifies which evaluation types and datasets to run
+custom_eval:
+  eval_types: ["quality_preference"]  # Options: ["quality_preference", "reward_alignment", "policy_ranking", "confusion_matrix"]
+  quality_preference: ["aliangdw_metaworld_metaworld_eval"]  # Datasets for preference evaluation
+  reward_alignment: ["aliangdw_metaworld_metaworld_eval"]  # Datasets for progress evaluation
+  policy_ranking: [aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking,aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking,aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking,jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm,jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top,jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist,jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all]  # Datasets for progress evaluation
+  confusion_matrix: ["aliangdw_metaworld_metaworld_eval"]  # Datasets for progress evaluation
+  comparisons_per_task: null # Limit number of comparisons per task (None = use all)
+  max_comparisons: 100  # Limit total number of comparisons across all tasks (None = use all)
+  reward_alignment_max_trajectories: 10  # Max trajectories for reward alignment
+  policy_ranking_max_tasks: 20  # Max tasks for policy ranking
+  use_frame_steps: false  # Whether to use frame steps (subsequences) for reward_alignment and policy_ranking. true = generate subsequences, false = use whole trajectory.
+  num_examples_per_quality_pr: null  # Examples per quality label for policy ranking
+  num_partial_successes: 5  # For RoboArena: Total trajectories to sample via circular sampling (None = use num_examples_per_quality_pr per group)
+  custom_eval_random_seed: 42
+
+# Output directory for evaluation results
+output_dir: null  # Defaults to ./baseline_eval_output/<reward_model>_<timestamp>
+
+# Server settings (for baseline_eval_server.py)
+server_url: "0.0.0.0"  # Evaluation server URL
+server_port: 8001  # Evaluation server port
diff --git a/robometer/robometer/configs/config.yaml b/robometer/robometer/configs/config.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..d906b23a481f4e43dd992f1c79b1047d2aec8eea
--- /dev/null
+++ b/robometer/robometer/configs/config.yaml
@@ -0,0 +1,244 @@
+# @package _global_
+
+# Robometer (RBM) Configuration
+# Configuration file for training and evaluating the RBM model with three objectives
+
+defaults:
+  - base_config
+  - _self_
+
+# =====================================================================================
+# Experiment Mode
+# =====================================================================================
+mode: "train"  # "train" or "evaluate"
+debug: false  # Set to true for debugging (reduces dataset size, enables debug features)
+trainer_cls: "rbm_heads" # "rewind_transformer"
+
+# =====================================================================================
+# Model Configuration
+# =====================================================================================
+model:
+  model_type: default
+  base_model_id: Qwen/Qwen3-VL-4B-Instruct
+  torch_dtype: bfloat16
+  # torch_dtype: float16
+  trust_remote_code: true
+  # Additional options for more comprehensive training (when use_peft: false)
+  train_vision_encoder: false   # Whether to train the vision encoder
+  train_language_model: true    # Whether to train the language model
+  train_progress_head: true     # Whether to train the progress prediction head
+  train_preference_head: true   # Whether to train the preference prediction head
+  train_success_head: false     # Whether to train the success prediction head
+  average_temporal_patches: false # Whether to average the patches for video frames when multi_image=False
+  use_per_frame_progress_token: false
+  # Pooling strategy used when extracting per-frame embeddings from vision token spans (multi-image mode)
+  # - mean: average over patch tokens for each frame (current default)
+  # - boundary: use the last patch token in the frame span
+  # - attention: learned attention pooling over patch tokens in the frame span
+  frame_pooling: "mean"
+  # Softmax temperature for attention pooling (only used when frame_pooling="attention")
+  frame_pooling_attn_temperature: 1.0
+  use_multi_image: ${data.use_multi_image}  # Reference from data config - If True, feed frames as a list of images instead of converting to video (avoids encoding overhead)
+
+  use_peft: false
+  peft_vision_encoder: false
+  quantization: false
+  use_unsloth: true  # Set to true to use unsloth for faster training with Qwen models
+  progress_loss_type: ${loss.progress_loss_type}  # Type of progress loss: "l1", "l2", or "discrete"
+  progress_discrete_bins: ${loss.progress_discrete_bins}  # Number of discrete bins for progress (synced from loss config)
+
+# =====================================================================================
+# PEFT Configuration
+# =====================================================================================
+peft:
+  r: 32
+  lora_alpha: 64
+  target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
+  lora_dropout: 0.05
+  bias: "none"
+
+
+# =====================================================================================
+# Data Configuration
+# =====================================================================================
+data:
+  # Dataset paths and sources
+  train_datasets: 
+    - mw
+  
+  # Evaluation dataset settings
+  eval_datasets:
+    - mw
+
+  eval_subset_size: null  # Number of evaluation examples
+
+  # Video processing settings
+  max_frames_after_preprocessing: 64
+  max_frames: 16  # Maximum frames per trajectory
+  resized_height: null # Height to resize images/videos to
+  resized_width: null   # Width to resize images/videos to
+  load_embeddings: false
+  min_frames_per_trajectory: 5 # Minimum number of frames required per trajectory (trajectories with fewer frames will be filtered out)
+  use_multi_image: false  # If True, feed frames as a list of images instead of converting to video (avoids encoding overhead)
+  shuffle_progress_frames: false  # If True, shuffle progress frames (except the first frame) and labels during training
+  traj_same_source_prob: 0.5  # Probability of sampling alternative task instruction from same data source
+
+  # Data generation settings
+  sample_type_ratio: [1, 0, 0]  # Ratio of preference, progress (progress absorbed in preference if predict_pref_progress is true)
+  dataset_preference_ratio: 0.7  # Ratio of preferences from dataset vs generated
+  # Tunable strategy ratios for preference negative generation: [rewind, suboptimal_same_task, different_task, reverse_progress]
+  preference_strategy_ratio: [1, 1, 1, 1]
+  # Tunable strategy ratios for progress generation: [different_task, forward_progress, reverse_progress, rewind]
+  progress_strategy_ratio: [1, 1, 1, 1]  # [different_task, forward_progress, reverse_progress, rewind]
+  shuffle: true
+  seed: 42
+
+  dataset_type: rbm
+
+  # Data loading settings
+  dataloader_pin_memory: True
+  dataloader_num_workers: 8
+  dataloader_persistent_workers: True
+
+  progress_pred_type: "absolute_first_frame" # "absolute_first_frame", "relative_first_frame", or "absolute_wrt_total_frames"
+
+  # Partial success threshold (for datasets with partial_success like RoboArena and RoboReward)
+  partial_success_threshold: 0.2  # Minimum difference in partial_success between chosen and rejected trajectories
+
+  # Success label parameters
+  min_success: 0.5  # Progress threshold below which success label pred is applied, anything in between min and max is not used for success pred loss unless succ = 1 for those frames.
+  max_success: 1.0  # Progress threshold above which success label is 1. Most sim datasets don't have a special cutoff number so we use 1.0 as default as sims will succeed at 1.0 progress.
+  # Dataset-specific success thresholds (CSV format: dataset_name,success_percentage)
+  dataset_success_cutoff_file: "robometer/data/dataset_success_cutoff.txt"
+  
+  progress_loss_type: ${loss.progress_loss_type}  # Type of progress loss: "l1", "l2", or "discrete"
+  progress_discrete_bins: ${loss.progress_discrete_bins}  # Number of discrete bins for progress when using discrete loss
+
+  use_per_frame_progress_token: ${model.use_per_frame_progress_token}
+  predict_last_frame_partial_progress: False 
+  
+# =====================================================================================
+# Custom Evaluation Configuration
+# =====================================================================================
+custom_eval:
+  # this section is for adding additional evals like policy ranking, confusion matrix, and alignment etc
+  # during training 
+  eval_types: ["policy_ranking", "confusion_matrix", "reward_alignment"]
+  policy_ranking: 
+    - mw
+  confusion_matrix: 
+    - mw
+  reward_alignment: 
+    - mw
+  quality_preference:
+    - mw
+  comparisons_per_task: 5  # Limit number of quality preference comparisons per task. null = use all comparisons. Uniformly samples if limit is set.
+  max_comparisons: null  # Limit total number of quality preference comparisons across all tasks. null = use all comparisons. Uniformly samples if limit is set.
+  num_examples_per_quality_pr: 5  # Number of trajectories to sample per quality label for policy ranking evaluation. Only tasks with multiple quality labels are used.
+  reward_alignment_max_trajectories: 10  # Maximum number of trajectories to use for reward alignment evaluation. null = use all trajectories.
+  custom_eval_random_seed: 42  # Random seed for custom eval samplers to ensure deterministic sampling across ranks.
+  policy_ranking_max_tasks: 100  # Maximum number of tasks to use for policy ranking evaluation. null = use all tasks with multiple quality labels.
+  use_frame_steps: true  # Whether to use frame steps (subsequences) for reward_alignment and policy_ranking. true = generate subsequences, false = use whole trajectory.
+  num_partial_successes: 5  # Number of partial successes to sample for policy ranking evaluation. null = use all partial successes.
+  
+# =====================================================================================
+# Training Configuration
+# =====================================================================================
+training:  
+  # Output and logging
+  output_dir: "./logs"
+  overwrite_output_dir: false  # If true, overwrite output directory if it exists (works with accelerate/distributed training)
+  
+  # Training parameters
+  per_device_train_batch_size: 16
+  gradient_accumulation_steps: 1
+  learning_rate: 2e-5
+  num_train_epochs: -1  # Number of epochs to train for
+  save_strategy: "no"
+  logging_steps: 1
+  save_steps: 200
+  max_steps: 20000  # -1 means no limit, use num_train_epochs instead
+  run_default_eval: false
+  # weight_decay: 0.1
+  weight_decay: 0.01
+  
+  # Evaluation settings
+  evaluation_strategy: "steps"  # Options: "no", "steps", "epoch"
+  eval_steps: 250  
+  custom_eval_steps: 250
+  # eval_epochs: 100  
+  per_device_eval_batch_size: 16  # Evaluation batch size per device
+  do_eval: true  # Enable evaluation during training
+  prediction_loss_only: true # Only compute loss for the prediction head
+  dataloader_pin_memory: ${data.dataloader_pin_memory}
+  dataloader_num_workers: ${data.dataloader_num_workers}
+  dataloader_persistent_workers: ${data.dataloader_persistent_workers}
+  
+  # Model settings
+  max_seq_length: 1024 # currently not used
+  beta: 0.1  # DPO beta parameter
+  
+  # Checkpoint loading (setup_utils loads weights; optimizer/step not restored)
+  load_from_checkpoint: null  # Hub path or local path to load model weights at setup
+  # Full resume (restore optimizer state and step counter)
+  resume_from_checkpoint: null  # Set to checkpoint path for full resume, null otherwise
+  
+  # Mixed precision and optimization
+  bf16: true
+  fp16: false
+  remove_unused_columns: false
+  gradient_checkpointing: true  # Re-enabled since FSDP is handled by accelerate config
+    
+  # Distributed training settings
+  ddp_find_unused_parameters: false
+  ddp_bucket_cap_mb: 25
+
+  # Optimizer settings
+  lr_scheduler_type: "cosine"
+  warmup_steps: 0
+  warmup_ratio: 0.1
+  max_grad_norm: 10.0
+  
+  # Vision encoder fine-tuning settings
+  vision_encoder_lr: 1e-5  # Learning rate for last N vision encoder layers. If null, uses the same LR as other parameters.
+  vision_encoder_num_layers: 3  # Number of last vision encoder layers to fine-tune with vision_encoder_lr. Only used if vision_encoder_lr is set.
+
+  # RBM-specific settings
+  predict_pref_progress: true
+
+# =====================================================================================
+# Loss Configuration
+# =====================================================================================
+loss:
+  # Success loss weight and pos_weight are computed dynamically from class imbalance
+  # to weight positive and negative classes equally (balanced accuracy)
+  predict_last_frame_progress: false  # If true, only compute progress loss for the last frame
+  progress_loss_type: "l2"  # Options: "l1", "l2", or "discrete" (default: "l2")
+  progress_discrete_bins: 32  # Number of discrete bins for progress when using discrete loss (default: 10)
+
+# =====================================================================================
+# Logging and Output Configuration
+# =====================================================================================
+logging:
+  save_model: true
+  save_processor: true
+  log_to: []
+  wandb_project: "robometer"
+  wandb_entity: clvr
+  wandb_notes: "training Robometer"
+  
+  # SaveBestCallback configuration
+  save_best:
+    metric_names: 
+      - eval_rew_align/pearson_mw_eval
+      - eval_p_rank/spearman_mw_eval
+
+    greater_is_better: [true, true]  # Whether higher values are better for each metric (must match metric_names length)
+    keep_top_k: 5  # Number of best checkpoints and uploads to keep
+    save_every: 1000  # Save 'latest' checkpoint every N steps (should be multiple of custom_eval_steps). Set to null to disable.
+    
+    # Hub upload configuration (optional)
+    upload_to_hub: false   # Whether to upload best models to HuggingFace Hub
+    hub_save_every: 1000  # Frequency (in steps) to upload to Hub. null = upload every checkpoint. Local saves always happen regardless.
+    hub_token: null       # Set to your HF token or use HF_TOKEN environment variable
+    hub_private: false    # Whether to make the uploaded model private
diff --git a/robometer/robometer/configs/distributed/ddp.yaml b/robometer/robometer/configs/distributed/ddp.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5ff9f3fc71134eaab82807e5d023b731ba3eb89b
--- /dev/null
+++ b/robometer/robometer/configs/distributed/ddp.yaml
@@ -0,0 +1,13 @@
+# accelerate_config.yaml — DDP for LLM-only fine-tuning
+
+compute_environment: LOCAL_MACHINE
+distributed_type: MULTI_GPU        # ← DDP
+num_processes: 2                   # ← set to #GPUs (example: 2)
+num_machines: 1
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+use_cpu: false
+
+# (Optional) leave dynamo off unless you’ve profiled it
+dynamo_backend: "no"
\ No newline at end of file
diff --git a/robometer/robometer/configs/distributed/fsdp.yaml b/robometer/robometer/configs/distributed/fsdp.yaml
new file mode 100755
index 0000000000000000000000000000000000000000..91ff4120cbae54e603ec92a92e3c7388fc925fe5
--- /dev/null
+++ b/robometer/robometer/configs/distributed/fsdp.yaml
@@ -0,0 +1,36 @@
+# Configuration for running full-parameter RBM (Robometer) training on multiple GPUs
+# using FSDP (Fully Sharded Data Parallelism).
+
+# Environment settings
+compute_environment: LOCAL_MACHINE
+distributed_type: FSDP
+num_processes: 1  # The number of GPUs you want to use
+machine_rank: 0
+num_machines: 1
+main_training_function: main
+mixed_precision: bf16 # Use bfloat16 for training
+use_cpu: false
+
+# FSDP specific configurations
+fsdp_config:
+  # Use transformer-based auto wrap policy with correct class names
+  fsdp_auto_wrap_policy: "TRANSFORMER_BASED_WRAP"
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  # fsdp_transformer_layer_cls_to_wrap: "Qwen2_5_VLVisionBlock,Qwen2_5_VLDecoderLayer"
+  fsdp_transformer_layer_cls_to_wrap: "Qwen3VLVisionBlock,Qwen3VLTextDecoderLayer"
+  # fsdp_transformer_layer_cls_to_wrap: "Qwen3VLTextDecoderLayer"
+  
+  # Sharding strategy: FULL_SHARD is the most memory-efficient
+  fsdp_sharding_strategy: FULL_SHARD # 1 for FULL_SHARD, 2 for SHARD_GRAD_OP
+
+  # CPU Offloading: Offload parameters and gradients to CPU RAM to save VRAM.
+  # This is crucial for fitting large models.
+  fsdp_offload_params: false
+
+  fsdp_use_orig_params: true
+
+  # Set the state dict type for saving sharded model checkpoints
+  fsdp_state_dict_type: FULL_STATE_DICT
+  # fsdp_state_dict_type: SHARDED_STATE_DICT
\ No newline at end of file
diff --git a/robometer/robometer/configs/eval_config_server.yaml b/robometer/robometer/configs/eval_config_server.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ea56c61ec235f8f657f4c4eb111a5c2254a2d56e
--- /dev/null
+++ b/robometer/robometer/configs/eval_config_server.yaml
@@ -0,0 +1,14 @@
+# Evaluation Configuration for RBM (Robometer)
+# This config is used specifically for evaluation runs
+
+# Model path to load training config from
+model_path: "aliangdw/Robometer-4B"  # or your RBM/ReWiND checkpoint
+
+# GPU Pool settings
+num_gpus: 1  # Number of GPUs to use (null for all available)
+max_workers: 8 # Max worker threads (null for auto, typically same as num_gpus)
+
+# Evaluation parameters
+batch_size: 4  # Batch size for evaluation
+server_url: "0.0.0.0"  # Evaluation server URL
+server_port: 8000  # Evaluation server port
diff --git a/robometer/robometer/configs/eval_configs.py b/robometer/robometer/configs/eval_configs.py
new file mode 100644
index 0000000000000000000000000000000000000000..d7aeb792ba66827fc4172135642108289c6c21f4
--- /dev/null
+++ b/robometer/robometer/configs/eval_configs.py
@@ -0,0 +1,274 @@
+#!/usr/bin/env python3
+"""
+Evaluation configuration for RBM.
+This file contains evaluation configuration classes:
+- EvalServerConfig: For evaluation server runs (eval_server.py)
+- OfflineEvalConfig: For standalone evaluation runs (run_eval_only.py)
+"""
+
+from dataclasses import dataclass, field, asdict
+from typing import Optional, Dict, Any
+
+from hydra.core.config_store import ConfigStore
+
+from robometer.configs.experiment_configs import CustomEvaluationConfig, DataConfig
+
+
+@dataclass
+class EvalServerConfig:
+    """Configuration for evaluation server runs (eval_server.py)."""
+
+    # Model path to load training config from
+    model_path: str = field(
+        default="./rbm_model_output/checkpoint-1000",
+        metadata={"help": "Path to the trained model checkpoint (will load training_config.yaml from here)"},
+    )
+
+    # GPU Pool settings
+    num_gpus: int = field(default=1, metadata={"help": "Number of GPUs to use (None for all available)"})
+    max_workers: int = field(
+        default=1, metadata={"help": "Max worker threads (None for auto, typically same as num_gpus)"}
+    )
+
+    # Evaluation parameters
+    batch_size: int = field(default=4, metadata={"help": "Batch size for evaluation"})
+    server_url: str = field(default="0.0.0.0", metadata={"help": "Evaluation server URL"})
+    server_port: int = field(default=8000, metadata={"help": "Evaluation server port"})
+
+
+@dataclass
+class OfflineEvalConfig:
+    """Configuration for standalone evaluation runs (run_eval_only.py)."""
+
+    # Model path (HuggingFace model ID or local checkpoint path)
+    model_path: str = field(
+        default="",
+        metadata={"help": "HuggingFace model ID (e.g., 'aliangdw/rbm_model') or local checkpoint path"},
+    )
+
+    # Output directory for evaluation results
+    output_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Output directory for evaluation results (defaults to checkpoint_path/eval_output)"},
+    )
+
+    # Custom evaluation configuration
+    custom_eval: CustomEvaluationConfig = field(
+        default_factory=CustomEvaluationConfig,
+        metadata={"help": "Custom evaluation configuration (reused from experiment_configs)"},
+    )
+
+    def __post_init__(self):
+        """Convert nested dict configs to dataclass instances."""
+        if isinstance(self.custom_eval, dict):
+            self.custom_eval = CustomEvaluationConfig(**self.custom_eval)
+
+
+# Model-specific configuration classes
+@dataclass
+class RLVLMFConfig:
+    """Configuration for RL-VLM-F baseline model."""
+
+    vlm_provider: str = field(
+        default="gemini",
+        metadata={"help": "VLM provider for RL-VLM-F: 'gemini' or 'openai'"},
+    )
+    temperature: float = field(
+        default=0.0,
+        metadata={"help": "Temperature for RL-VLM-F"},
+    )
+
+
+@dataclass
+class GVLConfig:
+    """Configuration for GVL baseline model."""
+
+    api_key: Optional[str] = field(
+        default=None,
+        metadata={"help": "API key (defaults to GEMINI_API_KEY or OPENAI_API_KEY env var based on provider)"},
+    )
+    provider: str = field(
+        default="gemini",
+        metadata={"help": "API provider: 'gemini' or 'openai'"},
+    )
+    model_name: str = field(
+        default="gemini-2.5-flash-lite",
+        metadata={"help": "Model name (e.g., 'gemini-2.0-flash' for Gemini, 'gpt-4o' for OpenAI)"},
+    )
+    offset: float = field(
+        default=0.5,
+        metadata={"help": "Frame offset for GVL"},
+    )
+    # Retry settings for API throttling
+    max_retries: int = field(
+        default=5,
+        metadata={"help": "Max retry attempts on 429/5xx errors"},
+    )
+    base_delay: float = field(
+        default=1.0,
+        metadata={"help": "Base delay in seconds for exponential backoff"},
+    )
+    max_delay: float = field(
+        default=60.0,
+        metadata={"help": "Maximum delay between retries"},
+    )
+
+
+@dataclass
+class VLACConfig:
+    """Configuration for VLAC baseline model."""
+
+    device: str = field(
+        default="cuda:0",
+        metadata={"help": "Device for VLAC model"},
+    )
+    model_type: str = field(
+        default="internvl2",
+        metadata={"help": "VLAC model type"},
+    )
+    temperature: float = field(
+        default=0.5,
+        metadata={"help": "Temperature for VLAC"},
+    )
+    batch_size: int = field(
+        default=5,
+        metadata={"help": "Batch size for VLAC processing"},
+    )
+    skip: int = field(
+        default=5,
+        metadata={"help": "Pair-wise step size for VLAC"},
+    )
+    frame_skip: bool = field(
+        default=True,
+        metadata={"help": "Whether to skip frames for VLAC efficiency"},
+    )
+    use_images: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, use image mode (get_trajectory_critic). If False, use video mode (web_trajectory_critic)"
+        },
+    )
+
+
+@dataclass
+class RBMConfig:
+    """Configuration for RBM/ReWiND baseline model."""
+
+    batch_size: int = field(
+        default=32,
+        metadata={"help": "Batch size for RBM/ReWiND model inference"},
+    )
+
+
+@dataclass
+class RoboRewardConfig:
+    """Configuration for RoboReward baseline model."""
+
+    max_new_tokens: int = field(
+        default=128,
+        metadata={"help": "Maximum number of tokens to generate for RoboReward"},
+    )
+
+
+@dataclass
+class RoboDopamineConfig:
+    """Configuration for Robo-Dopamine GRM baseline model."""
+
+    frame_interval: int = field(
+        default=1,
+        metadata={"help": "Step between sampled frames for before/after pairs (1 = every frame)"},
+    )
+    batch_size: int = field(
+        default=1,
+        metadata={"help": "Batch size for vLLM inference"},
+    )
+    eval_mode: str = field(
+        default="incremental",
+        metadata={"help": "Evaluation mode: 'incremental', 'forward', or 'backward'"},
+    )
+
+
+@dataclass
+class BaselineEvalConfig:
+    """Configuration for baseline evaluation runs (run_baseline_eval.py)."""
+
+    # Reward model discriminator: "gvl", "vlac", "rlvlmf", "rbm", "rewind", "roboreward", or "robodopamine"
+    reward_model: str = field(
+        default="rlvlmf",
+        metadata={
+            "help": "Reward model: 'gvl', 'vlac', 'robodopamine' for progress; 'rlvlmf' for preference; 'rbm', 'rewind' for trained models; 'roboreward' for RoboReward baseline"
+        },
+    )
+
+    model_config: Any = field(
+        default=None,
+        metadata={"help": "Model-specific configuration (automatically selected based on reward_model)"},
+    )
+
+    # Shared configuration
+    model_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Path to model checkpoint (HuggingFace repo ID or local path). Used by vlac, rbm, rewind, and roboreward models."
+        },
+    )
+    max_frames: int = field(
+        default=15,
+        metadata={"help": "Maximum frames for models (used by GVL and potentially others)"},
+    )
+
+    # Custom evaluation configuration
+    custom_eval: CustomEvaluationConfig = field(
+        default_factory=CustomEvaluationConfig,
+        metadata={"help": "Custom evaluation configuration"},
+    )
+
+    # Output directory for evaluation results
+    output_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Output directory for evaluation results"},
+    )
+
+    # Server settings (for baseline_eval_server.py)
+    server_url: str = field(
+        default="0.0.0.0",
+        metadata={"help": "Evaluation server URL"},
+    )
+    server_port: int = field(
+        default=8001,
+        metadata={"help": "Evaluation server port"},
+    )
+
+    def __post_init__(self):
+        """Convert nested dict configs to dataclass instances and initialize model_config."""
+        # Handle custom_eval conversion
+        if isinstance(self.custom_eval, dict):
+            self.custom_eval = CustomEvaluationConfig(**self.custom_eval)
+
+        # Handle model_config conversion based on reward_model
+        # If model_config is None or a dict, convert to appropriate dataclass
+        if self.model_config is None or isinstance(self.model_config, dict):
+            if self.reward_model == "rlvlmf":
+                self.model_config = RLVLMFConfig(**(self.model_config or {}))
+            elif self.reward_model == "gvl":
+                self.model_config = GVLConfig(**(self.model_config or {}))
+            elif self.reward_model == "vlac":
+                self.model_config = VLACConfig(**(self.model_config or {}))
+            elif self.reward_model in ["rewind", "rbm"]:
+                self.model_config = RBMConfig(**(self.model_config or {}))
+            elif self.reward_model == "roboreward":
+                self.model_config = RoboRewardConfig(**(self.model_config or {}))
+            elif self.reward_model == "robodopamine":
+                self.model_config = RoboDopamineConfig(**(self.model_config or {}))
+            else:
+                raise ValueError(
+                    f"Unknown reward_model: {self.reward_model}. "
+                    f"Must be 'rlvlmf', 'gvl', 'vlac', 'rbm', 'rewind', 'roboreward', or 'robodopamine'"
+                )
+
+
+# Register structured configs with Hydra
+cs = ConfigStore.instance()
+cs.store(name="eval_server_config", node=EvalServerConfig)
+cs.store(name="eval_only_config", node=OfflineEvalConfig)
+cs.store(name="baseline_eval_config", node=BaselineEvalConfig)
diff --git a/robometer/robometer/configs/experiment_configs.py b/robometer/robometer/configs/experiment_configs.py
new file mode 100755
index 0000000000000000000000000000000000000000..3a06b9b5ec98a9d51207ffe37b3702de17ce5021
--- /dev/null
+++ b/robometer/robometer/configs/experiment_configs.py
@@ -0,0 +1,562 @@
+#!/usr/bin/env python3
+"""
+Experiment configurations for RBM training.
+This file contains all the dataclass configurations that can be reused
+across different training scripts.
+"""
+
+from dataclasses import dataclass, field
+from typing import Optional, Dict, Any, List
+from transformers import PretrainedConfig
+
+
+@dataclass
+class ModelConfig(PretrainedConfig):
+    """Config for model settings"""
+
+    base_model_id: str = field(default="Qwen/Qwen2.5-VL-3B-Instruct")
+    model_type: str = field(default="default")
+    torch_dtype: str = field(default="bfloat16")
+    trust_remote_code: bool = field(default=True)
+    train_vision_encoder: bool = field(default=False, metadata={"help": "Whether to train the vision encoder"})
+    train_language_model: bool = field(default=True, metadata={"help": "Whether to train the language model"})
+
+    train_progress_head: bool = field(default=False, metadata={"help": "Whether to train the progress prediction head"})
+    train_preference_head: bool = field(
+        default=False, metadata={"help": "Whether to train the preference prediction head"}
+    )
+    train_success_head: bool = field(default=False, metadata={"help": "Whether to train the success prediction head"})
+
+    average_temporal_patches: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, average all tokens within each temporal patch group for progress prediction. If False, use the last token (boundary) of each temporal patch group."
+        },
+    )
+
+    frame_pooling: str = field(
+        default="mean",
+        metadata={
+            "help": "How to pool vision tokens into a per-frame embedding for progress/success heads in multi-image mode. "
+            "Options: 'mean' (average over patch tokens), 'boundary' (use last patch token), 'attention' (learned attention pooling)."
+        },
+    )
+    frame_pooling_attn_temperature: float = field(
+        default=1.0,
+        metadata={
+            "help": "Softmax temperature for attention pooling over patch tokens (only used when frame_pooling='attention')."
+        },
+    )
+
+    use_per_frame_progress_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, add a <|prog_token|> after each frame and use those token embeddings "
+            "for per-frame progress prediction. Requires use_multi_image=True."
+        },
+    )
+
+    use_multi_image: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, feed frames as a list of images instead of converting to video (avoids encoding overhead). "
+            "If False, detect automatically based on number of vision_start tokens in the sequence."
+        },
+    )
+
+    use_peft: bool = field(default=False, metadata={"help": "Whether to use PEFT/LoRA or train full model"})
+    peft_vision_encoder: bool = field(default=False, metadata={"help": "Whether to attach LoRA to the vision encoder"})
+
+    # use bitsandbytes for quantization
+    quantization: bool = field(default=False, metadata={"help": "Whether to use bitsandbytes for quantization"})
+
+    # use unsloth for faster training
+    use_unsloth: bool = field(
+        default=False, metadata={"help": "Whether to use unsloth for faster vision model training"}
+    )
+
+    progress_loss_type: str = field(
+        default="l2",
+        metadata={"help": "Type of progress loss: 'l1', 'l2', or 'discrete'"},
+    )
+    progress_discrete_bins: Optional[int] = field(
+        default=None,
+        metadata={"help": "Number of discrete bins for progress when using discrete loss (None for continuous)"},
+    )
+    # rewind sub-config
+    rewind: Optional[Dict[str, Any]] = field(default=None)
+
+    def __post_init__(self):
+        from robometer.models.rewind_transformer import ReWINDTransformerConfig
+
+        if self.rewind is not None and isinstance(self.rewind, dict):
+            # Pass progress_loss_type and progress_discrete_bins from parent config if not set in rewind dict
+            if "progress_loss_type" not in self.rewind:
+                self.rewind["progress_loss_type"] = self.progress_loss_type
+            if "progress_discrete_bins" not in self.rewind:
+                self.rewind["progress_discrete_bins"] = self.progress_discrete_bins or 10
+            if "use_per_frame_progress_token" not in self.rewind:
+                self.rewind["use_per_frame_progress_token"] = self.use_per_frame_progress_token
+
+            self.rewind = ReWINDTransformerConfig(**self.rewind)
+
+
+@dataclass
+class PEFTConfig:
+    """Config for PEFT/LoRA settings"""
+
+    r: int = field(default=32)
+    lora_alpha: int = field(default=64)
+    lora_dropout: float = field(default=0.05)
+    bias: str = field(default="none")
+    target_modules: List[str] = field(
+        default_factory=lambda: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
+    )
+    peft_vision_encoder: bool = field(default=False, metadata={"help": "Whether to attach LoRA to the vision encoder"})
+
+
+@dataclass
+class DataConfig:
+    """Configuration for data loading and processing."""
+
+    # Dataset paths
+    train_datasets: List[str] = field(
+        default_factory=lambda: ["abraranwar/libero_rfm"], metadata={"help": "List of training dataset names"}
+    )
+    eval_datasets: List[str] = field(
+        default_factory=lambda: ["abraranwar/libero_rfm"], metadata={"help": "List of evaluation dataset names"}
+    )
+
+    # Dataset type and configuration
+    dataset_type: str = field(
+        default="default",
+        metadata={"help": "Dataset type: 'default', 'rewound', 'success_failure'"},
+    )
+
+    max_frames_after_preprocessing: int = field(
+        default=64, metadata={"help": "Maximum number of frames to extract from videos after preprocessing"}
+    )
+    max_frames: int = field(default=8, metadata={"help": "Maximum number of frames to extract from videos"})
+    min_frames_per_trajectory: int = field(
+        default=5,
+        metadata={
+            "help": "Minimum number of frames required per trajectory (trajectories with fewer frames will be filtered out)"
+        },
+    )
+    resized_height: Optional[int] = field(default=None, metadata={"help": "Height to resize video frames to"})
+    resized_width: Optional[int] = field(default=None, metadata={"help": "Width to resize video frames to"})
+
+    # Video/image processing mode
+    use_multi_image: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, feed frames as a list of images instead of converting to video. "
+            "This avoids video encoding overhead and works for both SmolVLM and Qwen models."
+        },
+    )
+    traj_same_source_prob: float = field(
+        default=0.5,
+        metadata={
+            "help": "Probability of sampling a different task instruction from the same data source "
+            "when generating negative instructions. Remaining probability samples across all sources."
+        },
+    )
+    shuffle_progress_frames: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, shuffle progress trajectory frames (except the first frame) "
+            "and their corresponding target progress labels during training for RBM heads."
+        },
+    )
+
+    use_per_frame_progress_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, add a <|prog_token|> after each frame for per-frame progress prediction. "
+            "Requires use_multi_image=True."
+        },
+    )
+
+    # Data generation parameters
+    sample_type_ratio: List[float] = field(
+        default_factory=lambda: [1, 1, 1], metadata={"help": "Ratio of pref and progress samples"}
+    )
+    dataset_preference_ratio: float = field(
+        default=0.8, metadata={"help": "Ratio of dataset preference samples to generated preference samples"}
+    )
+    # [rewind, suboptimal_same_task, different_task, reverse_progress]
+    preference_strategy_ratio: List[float] = field(default_factory=lambda: [1, 1, 1, 1])
+    # [different_task, forward_progress, reverse_progress, rewind]
+    progress_strategy_ratio: List[float] = field(default_factory=lambda: [1, 1, 1, 1])
+
+    data_source_weights: Optional[Dict[str, float]] = field(
+        default=None,
+        metadata={
+            "help": "Dictionary mapping data source names to sampling weights (e.g., {'metaworld': 0.2, 'libero': 0.8})"
+        },
+    )
+
+    # Processing parameters
+    shuffle: bool = field(default=True, metadata={"help": "Whether to shuffle the dataset"})
+    seed: int = field(default=42, metadata={"help": "Random seed for reproducibility"})
+
+    # Evaluation parameters
+    eval_subset_size: Optional[int] = field(default=None, metadata={"help": "Number of samples to use for evaluation"})
+
+    # Dataloader parameters
+    dataloader_pin_memory: bool = field(default=True, metadata={"help": "Whether to pin memory in dataloader"})
+    dataloader_num_workers: int = field(default=0, metadata={"help": "Number of worker processes for dataloader"})
+    dataloader_persistent_workers: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, the data loader will not shut down the worker processes after a dataset has been consumed once. This allows to maintain the workers dataset instances alive."
+        },
+    )
+
+    max_trajectories: int = field(default=-1, metadata={"help": "Maximum number of trajectories to use for dataset"})
+
+    # Embedding loading parameters
+    load_embeddings: bool = field(
+        default=False,
+        metadata={"help": "Whether to load precomputed embeddings instead of processing frames (ReWiND only)"},
+    )
+
+    progress_pred_type: str = field(
+        default="absolute_wrt_total_frames", metadata={"help": "Type of progress prediction: 'absolute' or 'relative'"}
+    )
+
+    # Success prediction thresholds
+    min_success: float = field(
+        default=0.0, metadata={"help": "Minimum progress threshold for success prediction (label=0, failure)"}
+    )
+    max_success: float = field(
+        default=1.0, metadata={"help": "Maximum progress threshold for success prediction (label=1, success)"}
+    )
+    dataset_success_cutoff_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to dataset-specific success cutoff file (CSV format: dataset_name,success_percentage)"},
+    )
+
+    # Partial success threshold (for datasets with partial_success like RoboArena and RoboReward)
+    partial_success_threshold: float = field(
+        default=0.2,
+        metadata={
+            "help": "Minimum difference in partial_success required between chosen and rejected trajectories for preference sampling"
+        },
+    )
+
+    # Predict last frame partial progress mask
+    predict_last_frame_partial_progress: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, create a mask that marks the last frame for partial_success trajectories (< 1.0). "
+            "If False, the mask is all 1.0s (no masking)."
+        },
+    )
+
+    # Progress loss configuration
+    progress_loss_type: str = field(
+        default="l2",
+        metadata={"help": "Type of progress loss: 'l1', 'l2', or 'discrete' (synced from loss config)"},
+    )
+    progress_discrete_bins: int = field(
+        default=10,
+        metadata={"help": "Number of discrete bins for progress when using discrete loss (synced from loss config)"},
+    )
+
+
+@dataclass
+class CustomEvaluationConfig:
+    """Config for custom evaluation settings"""
+
+    eval_types: List[str] = field(default_factory=lambda: ["policy_ranking", "confusion_matrix", "reward_alignment"])
+    policy_ranking: List[str] = field(default_factory=lambda: ["aliangdw_metaworld_metaworld_eval"])
+    confusion_matrix: List[str] = field(default_factory=lambda: ["aliangdw_metaworld_metaworld_eval"])
+    reward_alignment: List[str] = field(default_factory=lambda: ["aliangdw_metaworld_metaworld_eval"])
+    quality_preference: List[str] = field(default_factory=lambda: ["aliangdw_metaworld_metaworld_eval"])
+    comparisons_per_task: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Limit number of quality preference comparisons per task. None = use all comparisons. Uniformly samples if limit is set."
+        },
+    )
+    max_comparisons: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Limit total number of quality preference comparisons across all tasks. None = use all comparisons. Uniformly samples if limit is set."
+        },
+    )
+    num_examples_per_quality_pr: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Number of trajectories to sample per quality label for policy ranking evaluation. Only tasks with multiple quality labels are used. If None = use all."
+        },
+    )
+    num_partial_successes: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For RoboArena datasets: Number of total trajectories to sample using circular sampling across partial_success values. None = use num_examples_per_quality_pr per partial_success group."
+        },
+    )
+    policy_ranking_max_tasks: Optional[int] = field(
+        default=100,
+        metadata={
+            "help": "Maximum number of tasks to use for policy ranking evaluation. None = use all tasks with multiple quality labels."
+        },
+    )
+    custom_eval_random_seed: int = field(
+        default=42,
+        metadata={
+            "help": "Random seed for sampling trajectories in custom evaluation samplers. Ensures all ranks sample the same trajectories to prevent hangs when dataloaders have different lengths."
+        },
+    )
+    reward_alignment_max_trajectories: Optional[int] = field(
+        default=10,
+        metadata={
+            "help": "Maximum number of trajectories to use for reward alignment evaluation. None = use all trajectories."
+        },
+    )
+    use_frame_steps: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to use frame steps (subsequences) for reward_alignment and policy_ranking evaluations. True = generate subsequences (0:frame_step, 0:2*frame_step, etc.), False = use whole trajectory."
+        },
+    )
+    subsample_n_frames: Optional[int] = field(
+        default=None,
+        metadata={"help": "Number of frames to subsample for reward alignment evaluation. null = use all frames."},
+    )
+    pad_frames: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to pad frames for reward alignment and policy ranking evaluations. True = pad frames, False = use original frames."
+        },
+    )
+
+
+@dataclass
+class TrainingConfig:
+    """Config for training settings"""
+
+    # Hardware settings
+    num_gpus: int = field(default=2, metadata={"help": "Number of GPUs to use for training"})
+
+    # Output and logging
+    output_dir: str = field(default="./logs")
+    exp_name: str = field(default="rbm")
+    max_seq_length: int = field(default=1024)
+    beta: float = field(default=0.1)
+    resume_from_checkpoint: Optional[str] = field(default=None)
+    load_from_checkpoint: Optional[str] = field(default=None)
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, overwrite the output directory if it exists. If False, raise an error if it exists."
+        },
+    )
+
+    # Training arguments
+    per_device_train_batch_size: int = field(default=1)
+    gradient_accumulation_steps: int = field(default=16)
+    learning_rate: float = field(default=5e-7)
+    num_train_epochs: Optional[int] = field(default=1)  # Default to 1 epoch if not specified
+    save_strategy: str = field(default="steps")
+    logging_steps: int = field(default=10)
+    bf16: bool = field(default=False)
+    fp16: bool = field(default=True)
+    remove_unused_columns: bool = field(default=False)
+    gradient_checkpointing: bool = field(default=True)
+    ddp_find_unused_parameters: bool = field(default=False)
+    ddp_bucket_cap_mb: int = field(default=25)
+    max_steps: Optional[int] = field(default=-1)  # -1 means no limit, use num_train_epochs instead
+    save_steps: int = field(default=100)
+    dataloader_pin_memory: bool = field(default=True)
+    dataloader_num_workers: int = field(default=0)
+    dataloader_persistent_workers: bool = field(default=False)
+
+    # Evaluation settings
+    evaluation_strategy: str = field(default="no", metadata={"help": "Evaluation strategy: 'no', 'steps', 'epoch'"})
+    eval_steps: Optional[int] = field(
+        default=None, metadata={"help": "Number of steps between evaluations (required if evaluation_strategy='steps')"}
+    )
+    run_default_eval: bool = field(
+        default=False, metadata={"help": "Whether to run default evaluation during training"}
+    )
+    custom_eval_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "Number of steps between custom evaluations (required if evaluation_strategy='steps')"},
+    )
+    per_device_eval_batch_size: int = field(default=1, metadata={"help": "Batch size for evaluation"})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run evaluation during training"})
+    prediction_loss_only: bool = field(default=True, metadata={"help": "Only compute loss for the prediction head"})
+
+    # Optimizer settings
+    lr_scheduler_type: str = field(default="cosine")
+    warmup_steps: int = field(default=0)
+    warmup_ratio: float = field(default=0.1)
+    max_grad_norm: float = field(default=1.0)
+    weight_decay: float = field(default=0.01, metadata={"help": "Weight decay for optimizer"})
+
+    # Vision encoder fine-tuning settings
+    vision_encoder_lr: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "Learning rate for last N vision encoder layers. If None, uses the same LR as other parameters."
+        },
+    )
+    vision_encoder_num_layers: int = field(
+        default=2,
+        metadata={
+            "help": "Number of last vision encoder layers to fine-tune with vision_encoder_lr. Only used if vision_encoder_lr is set."
+        },
+    )
+
+    # RBM-specific settings
+    predict_pref_progress: bool = field(
+        default=False, metadata={"help": "Whether to predict progress for preference samples"}
+    )
+
+
+@dataclass
+class LossConfig:
+    """Config for loss computation settings"""
+
+    success_positive_weight: float = field(
+        default=1.0,
+        metadata={"help": "Positive class weight for BCEWithLogits loss in success prediction (pos_weight)."},
+    )
+    predict_last_frame_progress: bool = field(
+        default=False,
+        metadata={"help": "If True, only compute progress loss for the last frame in the sequence"},
+    )
+    progress_loss_type: str = field(
+        default="l2",
+        metadata={"help": "Type of progress loss: 'l1', 'l2', or 'discrete'"},
+    )
+    progress_discrete_bins: int = field(
+        default=10,
+        metadata={"help": "Number of discrete bins for progress when using discrete loss (default: 10)"},
+    )
+
+
+@dataclass
+class SaveBestConfig:
+    """Configuration for SaveBestCallback"""
+
+    # Metric monitoring
+    metric_names: List[str] = field(
+        default_factory=lambda: ["custom_eval/p_rank_spearman_mw"],
+        metadata={"help": "List of metric names to monitor for saving best models (will be averaged)"},
+    )
+    greater_is_better: List[bool] = field(
+        default_factory=lambda: [True],
+        metadata={"help": "Whether higher values are better for each metric (must match length of metric_names)"},
+    )
+    keep_top_k: int = field(default=1, metadata={"help": "Number of best checkpoints/uploads to keep"})
+    save_every: Optional[int] = field(
+        default=None,
+        metadata={"help": "Save 'latest' checkpoint every N steps (should be multiple of eval_steps). None disables."},
+    )
+
+    # Hub upload configuration
+    upload_to_hub: bool = field(default=False, metadata={"help": "Whether to upload best models to HuggingFace Hub"})
+    hub_save_every: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Frequency (in steps) to upload to Hub. None = upload every checkpoint. Local saves always happen regardless."
+        },
+    )
+    hub_token: Optional[str] = field(default=None, metadata={"help": "HuggingFace token (or set HF_TOKEN env var)"})
+    hub_private: bool = field(default=False, metadata={"help": "Whether to make the Hub model private"})
+
+    def __post_init__(self):
+        """Validate that metric_names and greater_is_better have the same length"""
+        if len(self.metric_names) != len(self.greater_is_better):
+            raise ValueError(
+                f"metric_names ({len(self.metric_names)}) and greater_is_better ({len(self.greater_is_better)}) must have the same length"
+            )
+
+
+@dataclass
+class LoggingConfig:
+    """Config for logging settings"""
+
+    save_model: bool = field(default=True)
+    save_processor: bool = field(default=True)
+    # Logging backends
+    log_to: List[str] = field(
+        default_factory=list,
+        metadata={"help": "List of logging backends to use, e.g., ['wandb', 'tensorboard']"},
+    )
+    # Wandb configuration
+    wandb_project: str = field(default="rbm-model", metadata={"help": "Wandb project name"})
+    wandb_entity: Optional[str] = field(default=None, metadata={"help": "Wandb entity/username"})
+    wandb_notes: Optional[str] = field(
+        default=None, metadata={"help": "Optional notes/comment to add to the wandb run"}
+    )
+    wandb_mode: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Wandb mode: 'online' (default), 'offline' (local files only, no network I/O), or 'disabled'. "
+            "Offline mode prevents network I/O blocking that can cause deadlocks in distributed training."
+        },
+    )
+    # Log level: "TRACE", "DEBUG2", "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"
+    log_level: str = field(
+        default="INFO",
+        metadata={
+            "help": "Logging level for console output. Options: TRACE (most verbose), DEBUG2, DEBUG, INFO, WARNING, ERROR, CRITICAL"
+        },
+    )
+
+    # SaveBest configuration
+    save_best: Optional[SaveBestConfig] = field(default=None, metadata={"help": "SaveBestCallback configuration"})
+
+
+@dataclass
+class ExperimentConfig:
+    """Main experiment configuration"""
+
+    mode: str = field(default="train", metadata={"help": "Mode: 'train' or 'evaluate'"})
+    debug: bool = field(default=False, metadata={"help": "Whether to run in debug mode"})
+    trainer_cls: str = field(
+        default="rbm_heads",
+        metadata={"help": "Trainer class: 'rbm_heads', 'rewind_transformer', 'rewind_scale_transformer'"},
+    )
+    model: ModelConfig = field(default_factory=ModelConfig)
+    peft: PEFTConfig = field(default_factory=PEFTConfig)
+    data: DataConfig = field(default_factory=DataConfig)
+    training: TrainingConfig = field(default_factory=TrainingConfig)
+    loss: LossConfig = field(default_factory=LossConfig)
+    logging: LoggingConfig = field(default_factory=LoggingConfig)
+    custom_eval: CustomEvaluationConfig = field(default_factory=CustomEvaluationConfig)
+
+    def __post_init__(self):
+        if isinstance(self.loss, dict):
+            self.loss = LossConfig(**self.loss)
+
+        if isinstance(self.model, dict):
+            self.model = ModelConfig(**self.model)
+
+        if isinstance(self.peft, dict):
+            self.peft = PEFTConfig(**self.peft)
+
+        if isinstance(self.data, dict):
+            self.data.pop("roboarena_partial_success_threshold", None)
+            self.data = DataConfig(**self.data)
+
+        if isinstance(self.training, dict):
+            self.training = TrainingConfig(**self.training)
+
+        if isinstance(self.logging, dict):
+            self.logging = LoggingConfig(**self.logging)
+
+        # Handle nested SaveBestConfig in LoggingConfig
+        if hasattr(self.logging, "save_best") and isinstance(self.logging.save_best, dict):
+            self.logging.save_best = SaveBestConfig(**self.logging.save_best)
+        elif self.logging.save_best is None:
+            # Set default SaveBestConfig if not provided
+            self.logging.save_best = SaveBestConfig()
+
+        if isinstance(self.custom_eval, dict):
+            self.custom_eval = CustomEvaluationConfig(**self.custom_eval)
diff --git a/robometer/robometer/configs/preprocess.yaml b/robometer/robometer/configs/preprocess.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b6ebf723f4bcebe6ed315e4fc1b81fad815f5ee1
--- /dev/null
+++ b/robometer/robometer/configs/preprocess.yaml
@@ -0,0 +1,105 @@
+train_datasets:
+  - "aliangdw/metaworld"
+  - "jesbu1/molmoact_rfm" 
+  - "abraranwar/libero_rfm"
+  - "abraranwar/agibotworld_alpha_headcam_rfm"
+  - "ykorkmaz/libero_failure_rfm"
+  - "jesbu1/ph2d_rfm"
+  - "jesbu1/h2r_rfm"
+  - "anqil/rh20t_subset_rfm"
+  - "jesbu1/roboarena_0825_rfm"
+  - "jesbu1/oxe_rfm"
+  - "jesbu1/galaxea_rfm"
+  - "jesbu1/epic_rfm"
+  - "jesbu1/failsafe_rfm"
+  - "jesbu1/humanoid_everyday_rfm"
+  - "jesbu1/motif_rfm"
+  - "jesbu1/auto_eval_rfm"
+  - "jesbu1/soar_rfm"
+  - "jesbu1/fino_net_rfm"
+  - "jesbu1/racer_rfm"
+  - "jesbu1/roboreward_rfm"
+  - "jesbu1/roboreward_rfm_high_res"
+
+train_subsets:
+  - ["metaworld_train"]
+  - ["molmoact_dataset_household", "molmoact_dataset_tabletop"]
+  - ["libero256_90"]
+  - ["agibotworld"]
+  - ["libero_90_failure"]
+  - ["ph2d"]
+  - ["h2r"]
+  - ["rh20t_human", "rh20t_robot"]
+  - ["roboarena"]
+  - ["oxe_aloha_mobile", "oxe_austin_buds_dataset_converted_externally_to_rlds", "oxe_bc_z", "oxe_berkeley_cable_routing", "oxe_berkeley_fanuc_manipulation", "oxe_berkeley_mvp_converted_externally_to_rlds", "oxe_berkeley_rpt_converted_externally_to_rlds", "oxe_bridge_v2", "oxe_dlr_edan_shared_control_converted_externally_to_rlds", "oxe_droid", "oxe_fractal20220817_data", "oxe_furniture_bench_dataset_converted_externally_to_rlds", "oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds", "oxe_imperialcollege_sawyer_wrist_cam", "oxe_jaco_play", "oxe_language_table", "oxe_nyu_rot_dataset_converted_externally_to_rlds", "oxe_robo_set", "oxe_stanford_hydra_dataset_converted_externally_to_rlds", "oxe_tokyo_u_lsmo_converted_externally_to_rlds", "oxe_toto", "oxe_ucsd_kitchen_dataset_converted_externally_to_rlds", "oxe_utaustin_mutex"]
+  - ["galaxea_part1_r1_lite", "galaxea_part2_r1_lite", "galaxea_part3_r1_lite", "galaxea_part4_r1_lite", "galaxea_part5_r1_lite"]
+  - ["epic"]
+  - ["failsafe"]
+  - ["humanoid_everyday_rfm"]
+  - ["motif_rfm"]
+  - ["auto_eval_rfm"]
+  - ["soar_rfm"]
+  - ["fino_net"]
+  - ["racer_train"]
+  - ["roboreward_train"]
+  - ["roboreward_train"]
+
+eval_datasets:
+  - "abraranwar/libero_rfm"
+  - "ykorkmaz/libero_failure_rfm"
+  - "jesbu1/oxe_rfm_eval"
+  - "aliangdw/metaworld"
+  - "jesbu1/racer_rfm"
+  - "abraranwar/usc_koch_rewind_rfm"
+  - "aliangdw/usc_xarm_policy_ranking"
+  - "aliangdw/usc_franka_policy_ranking"
+  - "aliangdw/utd_so101_policy_ranking"
+  - "aliangdw/utd_so101_human"
+  #- "jesbu1/roboarena_eval_debug_nowrist"
+  - "jesbu1/hand_paired_rfm"
+  - "jesbu1/mit_franka_p-rank_rfm"
+  - "jesbu1/utd_so101_clean_policy_ranking_top"
+  - "jesbu1/utd_so101_clean_policy_ranking_wrist"
+  - "jesbu1/usc_koch_human_robot_paired"
+  - "jesbu1/usc_koch_p_ranking_rfm"
+  - "jesbu1/roboreward_rfm"
+  - "jesbu1/roboreward_rfm_high_res"
+  - "jesbu1/rfm_new_mit_franka_rfm_nowrist"
+  - "ykorkmaz/usc_trossen_rfm"
+
+eval_subsets:
+  - ["libero256_10", "libero256_goal", "libero256_object", "libero256_spatial"]
+  - ["libero_10_failure", "libero_goal_failure", "libero_object_failure", "libero_spatial_failure"]
+  - ["oxe_bc_z_eval", "oxe_berkeley_cable_routing_eval", "oxe_bridge_v2_eval", "oxe_jaco_play_eval", "oxe_toto_eval", "oxe_viola_eval"]
+  - ["metaworld_eval"]
+  - ["racer_val"]
+  - ["usc_koch_rewind"]
+  - ["usc_xarm_policy_ranking"]
+  - ["usc_franka_policy_ranking"]
+  - ["utd_so101_policy_ranking"]
+  - ["utd_so101_human"]
+  #- ["roboarena_eval_debug_nowrist"]
+  - ["hand_paired_robot", "hand_paired_human"]
+  - ["mit_franka_p-rank_rfm"]
+  - ["utd_so101_clean_policy_ranking_top"]
+  - ["utd_so101_clean_policy_ranking_wrist"]
+  - ["usc_koch_human_robot_paired_human", "usc_koch_human_robot_paired_robot"]
+  - ["usc_koch_p_ranking_all"]
+  - ["roboreward_val", "roboreward_test"]
+  - ["roboreward_val", "roboreward_test"]
+  - ["rfm_new_mit_franka_rfm_nowrist"]
+  - ["usc_trossen"]
+
+max_frames_for_preprocessing: 32
+video_frame_sampling: "uniform"
+num_proc: 1
+num_threads: 8
+force_reprocess: false
+cache_dir: /path/to/robometer/processed_datasets
+
+# Embedding preprocessing parameters
+precompute_embeddings: true
+embeddings_cache_dir: "embeddings"
+dinov2_model: "facebook/dinov2-base"
+sentence_model: "sentence-transformers/all-MiniLM-L12-v2"
+embedding_batch_size: 64
diff --git a/robometer/robometer/configs/preprocess_finetune.yaml b/robometer/robometer/configs/preprocess_finetune.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d806bf840042caf99d36adcfbe200a5b8dd7f51
--- /dev/null
+++ b/robometer/robometer/configs/preprocess_finetune.yaml
@@ -0,0 +1,39 @@
+# Preprocess config for LoRA fine-tuning on RoboFAC after running generate_hf_dataset with robofac.yaml.
+# Use with: uv run python -m robometer.data.scripts.preprocess_datasets --config robometer/configs/preprocess_finetune.yaml
+#
+# Order: 1) generate_hf_dataset (robofac.yaml) → 2) this preprocess → 3) train.
+# Run from repo root so the path below resolves, or use an absolute path for train_datasets/eval_datasets.
+#
+# Before running:
+#   export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/save/processed_datasets  # for training
+#
+# For training, set data.train_datasets and data.eval_datasets to the cache key: "robometer_dataset/robofac/robofac/robofac"
+# (dataset_path/subset). See FINETUNE_ROBOMETER.md.
+
+# Path to the converted dataset (output_dir/dataset_name from robofac.yaml = robometer_dataset/robofac/robofac)
+train_datasets:
+  - "aliangdw/robofac_rbm"
+
+# Single subset matching generate_hf_dataset output (config name = dataset_name = robofac)
+train_subsets:
+  - ["robofac"]
+
+eval_datasets:
+  - "aliangdw/robofac_rbm"
+
+eval_subsets:
+  - ["robofac"]
+
+max_frames_for_preprocessing: 32
+video_frame_sampling: "uniform"
+num_proc: 1
+num_threads: 4
+force_reprocess: false
+cache_dir: /path/to/robometer/processed_datasets
+
+# Embedding preprocessing (optional; set false for faster preprocessing without embeddings)
+precompute_embeddings: false
+embeddings_cache_dir: "embeddings"
+dinov2_model: "facebook/dinov2-base"
+sentence_model: "sentence-transformers/all-MiniLM-L12-v2"
+embedding_batch_size: 64
diff --git a/robometer/robometer/configs/reward_model/gvl.yaml b/robometer/robometer/configs/reward_model/gvl.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6c62129939c182a5681fa040dca0cce66ffd80ad
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/gvl.yaml
@@ -0,0 +1,14 @@
+# @package _global_
+
+# GVL model configuration
+reward_model: "gvl"
+
+model_config:
+  # api_key: null  # Will use GEMINI_API_KEY or OPENAI_API_KEY env var based on provider
+  provider: "gemini"  # Options: "gemini" or "openai"
+  model_name: "gemini-2.5-flash-lite"  # Gemini: gemini-2.0-flash, gemini-1.5-pro | OpenAI: gpt-4o, gpt-4o-mini
+  offset: 0.5
+  # Retry settings for API throttling
+  max_retries: 5       # Max retry attempts on 429/5xx errors
+  base_delay: 1.0      # Base delay in seconds for exponential backoff
+  max_delay: 60.0      # Maximum delay between retries
diff --git a/robometer/robometer/configs/reward_model/rbm.yaml b/robometer/robometer/configs/reward_model/rbm.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..50ebd63ef5e02bd3f2f7fd6f839c150cef5f77a1
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/rbm.yaml
@@ -0,0 +1,9 @@
+# @package _global_
+
+# RBM model configuration
+reward_model: "rbm"
+
+model_path: null  # Path to RBM model checkpoint (HuggingFace repo ID or local path, required)
+
+model_config:
+  batch_size: 32
diff --git a/robometer/robometer/configs/reward_model/rewind.yaml b/robometer/robometer/configs/reward_model/rewind.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a3696716060b00c396eca9d4eb9c0f715cabc2d0
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/rewind.yaml
@@ -0,0 +1,10 @@
+# @package _global_
+
+# ReWiND model configuration
+reward_model: "rewind"
+
+model_path: null  # Path to ReWiND model checkpoint (HuggingFace repo ID or local path, required)
+
+model_config:
+  batch_size: 32
+
diff --git a/robometer/robometer/configs/reward_model/rfm.yaml b/robometer/robometer/configs/reward_model/rfm.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9a918e5dbad79d47b91de91661f029f836a94f74
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/rfm.yaml
@@ -0,0 +1,8 @@
+# @package _global_
+# Alias for rbm (legacy config name; use reward_model/rbm.yaml)
+reward_model: "rbm"
+model_path: null  # Path to RBM model checkpoint (HuggingFace repo ID or local path, required)
+
+model_config:
+  batch_size: 32
+
diff --git a/robometer/robometer/configs/reward_model/rlvlmf.yaml b/robometer/robometer/configs/reward_model/rlvlmf.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bf96408e7f64aa6e7528c9f29c0580362329d3a0
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/rlvlmf.yaml
@@ -0,0 +1,9 @@
+# @package _global_
+
+# RL-VLM-F model configuration
+reward_model: "rlvlmf"
+
+model_config:
+  vlm_provider: "gemini"  # Options: "gemini", "openai"
+  temperature: 0.0
+
diff --git a/robometer/robometer/configs/reward_model/robodopamine.yaml b/robometer/robometer/configs/reward_model/robodopamine.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..81f74aef537b0beb5801e77fc8a48fc5ac631994
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/robodopamine.yaml
@@ -0,0 +1,11 @@
+# @package _global_
+
+# Robo-Dopamine GRM model configuration
+reward_model: "robodopamine"
+
+model_path: "tanhuajie2001/Robo-Dopamine-GRM-3B"
+
+model_config:
+  frame_interval: 1
+  batch_size: 1
+  eval_mode: "incremental"
diff --git a/robometer/robometer/configs/reward_model/roboreward.yaml b/robometer/robometer/configs/reward_model/roboreward.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bf32c9e94ebba4e85718f5072c9ddfd0580facba
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/roboreward.yaml
@@ -0,0 +1,10 @@
+# @package _global_
+
+# RoboReward model configuration
+reward_model: "roboreward"
+
+model_path: "teetone/RoboReward-4B"  # Path to RoboReward model (HuggingFace repo ID, e.g., 'teetone/RoboReward-8B' or 'teetone/RoboReward-4B')
+
+model_config:
+  max_new_tokens: 128
+
diff --git a/robometer/robometer/configs/reward_model/vlac.yaml b/robometer/robometer/configs/reward_model/vlac.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..252b2f062c28a008c4469a5275426efeac635955
--- /dev/null
+++ b/robometer/robometer/configs/reward_model/vlac.yaml
@@ -0,0 +1,17 @@
+# @package _global_
+
+# VLAC model configuration
+reward_model: "vlac"
+
+model_path: "InternRobotics/VLAC"  # Path to VLAC model checkpoint (HuggingFace repo ID or local path)
+
+model_config:
+  device: "cuda:0"
+  model_type: "internvl2"
+  temperature: 0.5
+  # top_k: 1
+  batch_size: 32
+  skip: 1 
+  frame_skip: true
+  use_images: false  # If true, use image mode; if false, use video mode
+
diff --git a/robometer/robometer/configs/rewind_transformer.yaml b/robometer/robometer/configs/rewind_transformer.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..59c76bf7c4fc8f43c964673449fea6b3146c9aa6
--- /dev/null
+++ b/robometer/robometer/configs/rewind_transformer.yaml
@@ -0,0 +1,50 @@
+# @package _global_
+
+# ReWIND Transformer Configuration
+# Inherits from config.yaml and overrides specific settings
+defaults:
+  - config 
+  - _self_
+
+trainer_cls: rewind_transformer
+
+# =====================================================================================
+# Model Configuration
+# =====================================================================================
+model:
+  base_model_id: rewind_transformer  # custom transformer model
+
+  # set rewind specific parameters
+  rewind:
+    video_feature_dim: 768
+    text_feature_dim: 384
+    hidden_dim: 512
+    num_layers: 4
+    num_attention_heads: 8
+    dropout: 0.1  
+    max_len: 16
+    causal_mask: true
+    use_per_frame_progress_token: true
+    progress_loss_type: l2
+    progress_discrete_bins: 32
+
+
+training:
+  gradient_checkpointing: false
+
+  bf16: false
+  fp16: true
+
+  per_device_train_batch_size: 1024
+  per_device_eval_batch_size: 512
+  learning_rate: 1e-4
+  weight_decay: 0.01
+  warmup_ratio: 0.01
+  
+  eval_steps: 200
+  custom_eval_steps: 200
+  max_steps: 20000
+
+data:  
+  load_embeddings: true
+  progress_pred_type: absolute 
\ No newline at end of file
diff --git a/robometer/robometer/configs/rewind_transformer_scaled.yaml b/robometer/robometer/configs/rewind_transformer_scaled.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2f9edbd3c93d05c3d3e1b60cc1eefc7a1c30e7be
--- /dev/null
+++ b/robometer/robometer/configs/rewind_transformer_scaled.yaml
@@ -0,0 +1,49 @@
+# @package _global_
+
+# ReWIND Transformer Configuration
+# Inherits from config.yaml and overrides specific settings
+defaults:
+  - config 
+  - _self_
+
+trainer_cls: rewind_transformer
+
+# =====================================================================================
+# Model Configuration
+# =====================================================================================
+model:
+  base_model_id: rewind_transformer  # custom transformer model
+
+  # set rewind specific parameters
+  rewind:
+    video_feature_dim: 768
+    text_feature_dim: 384
+    hidden_dim: 1024
+    num_layers: 32
+    num_attention_heads: 16
+    dropout: 0.1  
+    max_len: 16
+    causal_mask: true
+    use_per_frame_progress_token: true
+    progress_loss_type: discrete
+    progress_discrete_bins: 32
+
+training:
+  gradient_checkpointing: false
+
+  bf16: false
+  fp16: true
+
+  per_device_train_batch_size: 1024
+  per_device_eval_batch_size: 512
+  learning_rate: 1e-4
+  weight_decay: 0.01
+  warmup_ratio: 0.01
+  
+  eval_steps: 200
+  custom_eval_steps: 200
+  max_steps: 20000
+
+data:  
+  load_embeddings: true
+  progress_pred_type: absolute 
\ No newline at end of file
diff --git a/robometer/robometer/data/collators/__init__.py b/robometer/robometer/data/collators/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..90167b195d356b39bafd5deea1f89d7cffdf57cc
--- /dev/null
+++ b/robometer/robometer/data/collators/__init__.py
@@ -0,0 +1,6 @@
+from .base import BaseCollator
+from .rewind import ReWiNDBatchCollator
+from .rbm_heads import RBMBatchCollator
+from .utils import convert_frames_to_pil_images, pad_list_to_max
+
+__all__ = ["BaseCollator", "RBMBatchCollator", "ReWiNDBatchCollator"]
diff --git a/robometer/robometer/data/collators/base.py b/robometer/robometer/data/collators/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..13c9d3569f582bd7a8867888a18c10e63376d961
--- /dev/null
+++ b/robometer/robometer/data/collators/base.py
@@ -0,0 +1,86 @@
+import torch
+from transformers import AutoProcessor, AutoTokenizer
+
+from robometer.data.dataset_types import PreferenceSample, ProgressSample, SampleType
+
+
+class BaseCollator:
+    def __init__(
+        self,
+        processor: AutoProcessor,
+        tokenizer: AutoTokenizer = None,
+        max_length: int = 1024,
+        resized_height: int = 128,
+        resized_width: int = 128,
+        base_model_id: str = None,
+        load_embeddings: bool = False,
+        **kwargs,
+    ):
+        self.processor = processor
+        self.tokenizer = tokenizer
+        self.max_length = max_length
+        self.resized_height = resized_height
+        self.resized_width = resized_width
+        self.tokenizer = tokenizer
+        self.base_model_id = base_model_id
+        self.load_embeddings = load_embeddings
+
+        # Update processor based on base model id
+        # if "SmolVLM" in self.base_model_id:
+        #     # For image processor
+        #     self.processor.image_processor.max_image_size = {"longest_edge": self.resized_height}
+        #     self.processor.image_processor.size = {"longest_edge": self.resized_height}
+        #     self.processor.image_processor.video_sampling["video_size"] = {"longest_edge": self.resized_height}
+
+        #     # for video processor
+        #     self.processor.video_processor.max_image_size = {"longest_edge": self.resized_height}
+        #     self.processor.video_processor.size = {"longest_edge": self.resized_height}
+        #     self.processor.video_processor.video_sampling["video_size"] = {"longest_edge": self.resized_height}
+
+    def __call__(
+        self,
+        samples: list[SampleType],
+    ) -> dict[str, torch.Tensor]:
+        """
+        Collate a list of samples into separate batches for preferences and progress.
+
+        Args:
+            samples: List of Sample objects or dictionaries that can be converted to Sample objects
+
+        Returns:
+            Dictionary containing separate batches for preferences and progress
+        """
+        # Convert dictionaries to Sample objects if needed
+        sample_objects = []
+        for sample in samples:
+            if isinstance(sample, dict):
+                sample_type = sample.get("sample_type", "unknown")
+                if sample_type == "preference":
+                    sample_obj = PreferenceSample(**sample)
+                elif sample_type == "progress":
+                    sample_obj = ProgressSample(**sample)
+                else:
+                    raise ValueError(f"Unknown sample_type: {sample_type}. Must be 'preference' or 'progress'")
+                sample_objects.append(sample_obj)
+            elif isinstance(sample, (PreferenceSample, ProgressSample)):
+                sample_objects.append(sample)
+            else:
+                raise ValueError(f"Expected Sample object or dict, got {type(sample)}")
+
+        preference_samples = [s for s in sample_objects if s.sample_type == "preference"]
+        progress_samples = [s for s in sample_objects if s.sample_type == "progress"]
+
+        preference_inputs = {}
+        if preference_samples:
+            preference_inputs = self._process_preference_batch(preference_samples)
+
+        progress_inputs = {}
+        if progress_samples:
+            progress_inputs = self._process_progress_batch(progress_samples)
+
+        return {
+            "preference_inputs": preference_inputs,
+            "progress_inputs": progress_inputs,
+            "num_preferences": len(preference_samples),
+            "num_progress": len(progress_samples),
+        }
diff --git a/robometer/robometer/data/collators/rbm_heads.py b/robometer/robometer/data/collators/rbm_heads.py
new file mode 100755
index 0000000000000000000000000000000000000000..cadf0c2a2095c307c93229b1c1d1a07766b95ffa
--- /dev/null
+++ b/robometer/robometer/data/collators/rbm_heads.py
@@ -0,0 +1,694 @@
+#!/usr/bin/env python3
+"""
+Batch collator for processing list of samples.
+"""
+
+import tempfile
+import uuid
+from pathlib import Path
+import numpy as np
+import torch
+from qwen_vl_utils import process_vision_info
+
+from .base import BaseCollator
+from .utils import convert_frames_to_pil_images, pad_list_to_max, write_mp4
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+from robometer.data.dataset_category import is_preference_only_ds
+from robometer.data.datasets.helpers import DataGenStrat
+from typing import List, Dict, Union
+from robometer.models.utils import convert_discrete_target_to_continuous
+from PIL import Image
+
+MAX_IMAGE_SIDE = 480  # bigger side
+MAX_IMAGE_PIXELS = 1024 * 1024  # safety cap (1.0 MP). raise to 1.5MP if stable
+
+
+def _resize_pil(pil: Image.Image, max_side: int = MAX_IMAGE_SIDE, max_pixels: int = MAX_IMAGE_PIXELS) -> Image.Image:
+    pil = pil.convert("RGB")
+    w, h = pil.size
+
+    # Scale down if max side too large
+    scale_side = min(1.0, max_side / float(max(w, h)))
+
+    # Scale down if too many pixels (area cap)
+    scale_area = (max_pixels / float(w * h)) ** 0.5 if (w * h) > max_pixels else 1.0
+
+    scale = min(scale_side, scale_area)
+
+    if scale < 1.0:
+        nw, nh = max(1, int(w * scale)), max(1, int(h * scale))
+        pil = pil.resize((nw, nh), resample=Image.BICUBIC)
+
+    return pil
+
+
+def should_compute_progress(
+    quality_label: str,
+    data_gen_strategy: str,
+    data_source: str = None,
+    is_chosen: bool = False,
+    partial_success: float | None = None,
+) -> float:
+    """
+    Check if progress should be computed for a trajectory.
+
+    Includes if it is successful or rewound
+    but NOT suboptimal or failure. Also masks out progress if data_source is in preference_only category,
+    except when the strategy is DIFFERENT_TASK and it's the rejected trajectory (where progress should still be computed, but will be 0.0).
+    Also includes trajectories with partial_success.
+
+    Args:
+        quality_label: The quality label of the trajectory
+        data_gen_strategy: The data generation strategy
+        data_source: The data source name (optional)
+        is_chosen: Whether this is the chosen trajectory (traj A) in a preference sample
+        partial_success: Partial success value (0-1) or None. If present, progress should be computed.
+
+    Returns:
+        1.0 if progress should be computed, 0.0 otherwise
+    """
+
+    # Mask out progress if data_source is in preference_only category
+    if data_source is not None and is_preference_only_ds(data_source):
+        # For preference_only datasets:
+        # - If it's the chosen trajectory, always mask out (don't compute)
+        # - If it's the rejected trajectory with DIFFERENT_TASK strategy, still compute (it will be 0.0)
+        return 0.0
+        # if is_chosen:
+        #    return 0.0
+        # elif data_gen_strategy == DataGenStrat.DIFFERENT_TASK.value:
+        #    return 1.0
+        # else:
+        #    return 0.0
+    # If partial_success and not is_preference_only_ds, always compute progress
+    # predict partial success for roboreward trajectories not preference only
+    elif partial_success is not None and "roboreward" in data_source:
+        return 1.0
+    elif quality_label in ["suboptimal", "failure", "failed"]:
+        return 0.0
+    elif quality_label == "successful" or data_gen_strategy == DataGenStrat.REWIND.value:
+        return 1.0
+
+    return 0.0
+
+
+def create_padding_mask(frames_shapes: torch.Tensor, max_length: int = None) -> torch.Tensor:
+    """
+    Create padding mask based on frames_shape.
+
+    Args:
+        frames_shapes: Tensor of shape (batch_size, ...) where first dim of each row is num_frames
+        max_length: Maximum length for padding. If None, uses max of first dim in frames_shapes
+
+    Returns:
+        Tensor of shape (batch_size, max_length) with 1.0 for valid frames, 0.0 for padding
+    """
+    # Extract num_frames from first dimension of each shape
+    if frames_shapes.dim() > 1:
+        num_frames = frames_shapes[:, 0].float()
+    else:
+        num_frames = frames_shapes.float()
+
+    if max_length is None:
+        max_length = int(num_frames.max().item())
+
+    # Create range tensor: [0, 1, 2, ..., max_length-1]
+    range_tensor = torch.arange(max_length, dtype=torch.float32, device=frames_shapes.device)
+
+    # Broadcast comparison: (batch_size, 1) vs (1, max_length) -> (batch_size, max_length)
+    # For each sample, positions < num_frames are valid (1.0), others are padding (0.0)
+    masks = (range_tensor.unsqueeze(0) < num_frames.unsqueeze(1)).float()
+
+    return masks
+
+
+class RBMBatchCollator(BaseCollator):
+    def __init__(
+        self,
+        processor,
+        tokenizer=None,
+        max_length: int = 1024,
+        resized_height: int = 128,
+        resized_width: int = 128,
+        base_model_id: str = None,
+        load_embeddings: bool = False,
+        use_multi_image: bool = False,
+        prog_pref: bool = False,
+        use_per_frame_progress_token: bool = False,
+        shuffle_progress_frames: bool = False,
+        inference: bool = False,
+        **kwargs,
+    ):
+        super().__init__(
+            processor=processor,
+            tokenizer=tokenizer,
+            max_length=max_length,
+            resized_height=resized_height,
+            resized_width=resized_width,
+            base_model_id=base_model_id,
+            load_embeddings=load_embeddings,
+            **kwargs,
+        )
+        self.use_multi_image = use_multi_image
+        self.prog_pref = prog_pref
+
+        # Molmo2 only supports multi-image mode, not video
+        if "Molmo" in self.base_model_id and not self.use_multi_image:
+            raise ValueError(
+                "Molmo2 does not support video mode (use_multi_image=False). "
+                "Please set data.use_multi_image=True to use Molmo2 with multi-image input."
+            )
+        self.use_per_frame_progress_token = use_per_frame_progress_token
+        # Validate that use_per_frame_progress_token requires use_multi_image
+        if self.use_per_frame_progress_token and not self.use_multi_image:
+            raise ValueError(
+                "use_per_frame_progress_token=True requires use_multi_image=True. "
+                "Per-frame progress tokens can only be added in multi-image mode."
+            )
+        self.shuffle_progress_frames = shuffle_progress_frames
+        self.inference = inference
+
+    def _prepare_frames_for_conversation(self, frames: List, prefix: str = "tmp") -> tuple[Union[List, str], dict]:
+        """
+        Prepare frames for conversation based on use_multi_image flag.
+
+        Args:
+            frames: List of PIL Images
+            prefix: Prefix for temporary video file (if needed)
+
+        Returns:
+            tuple: (video_field, content_extras)
+                - video_field: Either list of PIL Images (if use_multi_image) or video file path (str)
+                - content_extras: Dictionary with resized_height/width or empty dict
+        """
+        if self.use_multi_image:
+            # # Use images directly - return list of PIL Images
+            # if self.resized_height is not None and self.resized_width is not None:
+            #     content_extras = {
+            #         "resized_height": self.resized_height,
+            #         "resized_width": self.resized_width,
+            #     }
+            # else:
+            #     frames = [_resize_pil(frame) for frame in frames]
+            #     content_extras = {}
+            content_extras = {}
+            return frames, content_extras
+        elif "Qwen" in self.base_model_id or "Molmo" in self.base_model_id:
+            # Qwen and Molmo accept list of PIL Images directly
+            if self.resized_height is not None and self.resized_width is not None:
+                content_extras = {
+                    "resized_height": self.resized_height,
+                    "resized_width": self.resized_width,
+                }
+            else:
+                frames = [_resize_pil(frame) for frame in frames]
+                content_extras = {}
+            return frames, content_extras
+        elif "SmolVLM" in self.base_model_id:
+            frames = [_resize_pil(frame) for frame in frames]
+            # Convert to video file for SmolVLM
+            unique_id = uuid.uuid4().hex
+            tmp = Path(tempfile.gettempdir()) / f"{prefix}_{unique_id}.mp4"
+            write_mp4(frames, tmp, fps=1)
+            return str(tmp), {}
+        else:
+            # Default: return frames as-is
+            return frames, {}
+
+    def _add_vision_content_to_list(
+        self, content_list: List[Dict], frames_or_video: Union[List, str], content_extras: dict
+    ) -> None:
+        """
+        Add vision content (images or video) to a content list.
+
+        Args:
+            content_list: List to append vision content to
+            frames_or_video: Either list of PIL Images (if use_multi_image) or video file path (str)
+            content_extras: Dictionary with additional content parameters
+        """
+        if self.use_multi_image:
+            # Add each image as a separate entry
+            for img in frames_or_video:
+                content_list.append({
+                    "type": "image",
+                    "image": img,
+                    **content_extras,
+                })
+                # Add per-frame progress token after each frame if enabled
+                if self.use_per_frame_progress_token:
+                    content_list.append({"type": "text", "text": "<|prog_token|>"})
+        else:
+            # Add video entry
+            content_list.append({
+                "type": "video",
+                "video": frames_or_video,
+                "sample_fps": 1.0,
+                **content_extras,
+            })
+
+    def _process_conversation(
+        self, conversations: List[List[Dict]], add_generation_prompt: bool = False
+    ) -> Dict[str, torch.Tensor]:
+        """
+        Process a list of conversations into a batch of inputs.
+
+        Args:
+            conversations: List of conversations
+            add_generation_prompt: Whether to add generation prompt (otherwise conversation will be closed with </im_end> token)
+        Returns:
+            Batch of inputs
+        """
+        if "Qwen" in self.base_model_id or "Molmo" in self.base_model_id:
+            # Process all messages in one batch
+            texts = [
+                self.processor.apply_chat_template(
+                    msg,
+                    tokenize=False,
+                    add_generation_prompt=add_generation_prompt,
+                    add_vision_id=True,
+                    enable_thinking=False,
+                    fps=1,
+                )
+                for msg in conversations
+            ]
+
+            is_qwen3 = "Qwen3" in self.base_model_id or "Molmo2" in self.base_model_id
+
+            # For Qwen3, pass image_patch_size to process_vision_info
+            process_kwargs = {
+                "return_video_kwargs": True,
+                "return_video_metadata": is_qwen3,
+            }
+            if (
+                is_qwen3
+                and hasattr(self.processor, "image_processor")
+                and hasattr(self.processor.image_processor, "patch_size")
+            ):
+                process_kwargs["image_patch_size"] = self.processor.image_processor.patch_size
+
+            image_inputs, video_inputs, video_kwargs = process_vision_info(conversations, **process_kwargs)
+
+            # For Qwen3, video_inputs is a list of (video, video_metadata) tuples
+            # that need to be split before passing to processor
+            if is_qwen3 and video_inputs is not None and len(video_inputs) > 0:
+                # Check if video_inputs contains tuples (Qwen3 format) or is already split
+                if isinstance(video_inputs[0], tuple) and len(video_inputs[0]) == 2:
+                    videos, video_metadatas = zip(*video_inputs)
+                    videos, video_metadatas = list(videos), list(video_metadatas)
+                else:
+                    # Already in the correct format
+                    videos = video_inputs
+                    video_metadatas = None
+            else:
+                videos = video_inputs if video_inputs else None
+                video_metadatas = None
+
+            # Process through the processor in one batch
+            processor_kwargs = {
+                "text": texts,
+                "images": image_inputs,
+                "padding": True,
+                "truncation": False,
+                "max_length": self.max_length,
+                "return_tensors": "pt",
+                "do_resize": False,
+            }
+
+            # Only add videos if they exist
+            if videos is not None:
+                processor_kwargs["videos"] = videos
+
+            # Add video_metadata and video_kwargs for Qwen3
+            # Note: video_kwargs may contain keys like 'videos_kwargs' that need special handling
+            if is_qwen3:
+                if video_metadatas is not None:
+                    processor_kwargs["video_metadata"] = video_metadatas
+                # Pass video_kwargs - these contain important metadata for video processing
+                # The processor will handle 'videos_kwargs' internally if present
+                if video_kwargs:
+                    processor_kwargs.update(video_kwargs)
+
+            batch_inputs = self.processor(**processor_kwargs)
+        elif "SmolVLM" in self.base_model_id:
+            batch_inputs = self.processor.apply_chat_template(
+                conversations,
+                add_generation_prompt=add_generation_prompt,
+                tokenize=True,
+                padding=True,
+                truncation=False,
+                max_length=self.max_length,
+                return_dict=True,
+                return_tensors="pt",
+                fps=1,  # this should be same as fps for write_mp4
+            )
+        else:
+            raise ValueError(f"Invalid base model id: {self.base_model_id}")
+
+        return batch_inputs
+
+    def _process_progress_batch(self, progress_samples: list[ProgressSample]) -> dict[str, torch.Tensor]:
+        """Process a batch of progress samples."""
+        # Collect all messages for batch processing
+        all_messages = []
+
+        target_progress_overrides: list[list[float] | None] = []
+
+        for sample in progress_samples:
+            # Convert frames to appropriate format using stored shapes
+            frames = convert_frames_to_pil_images(sample.trajectory.frames, sample.trajectory.frames_shape)
+            target_progress = sample.trajectory.target_progress
+
+            # Optionally shuffle frames (except the first frame) and keep target progress aligned
+            if self.shuffle_progress_frames and target_progress is not None and not self.inference:
+                if len(target_progress) > 1 and len(target_progress) == len(frames):
+                    shuffle_indices = np.random.permutation(range(1, len(frames)))
+                    frames = [frames[0]] + [frames[idx] for idx in shuffle_indices]
+                    target_progress = [target_progress[0]] + [target_progress[idx] for idx in shuffle_indices]
+                else:
+                    raise ValueError(
+                        "Target progress must be a list with at least 1 element and match the number of frames "
+                        f"for shuffling, got {len(target_progress)} entries for {len(frames)} frames"
+                    )
+
+            target_progress_overrides.append(target_progress)
+
+            video_field, content_extras = self._prepare_frames_for_conversation(frames, prefix="tmp_progress")
+
+            # Create conversation for progress evaluation
+            prompt = f"The task for the robot is '{sample.trajectory.task}'. Given the trajectory video, predict the task progress at each frame, how far along the robot is towards completing the task, a float between 0 and 1, where 0 is the starting state and 1 is when the task is completed. If the robot is not performing the same task, predict 0 progress."
+
+            # Build content list
+            content_list = [{"type": "text", "text": prompt}]
+            self._add_vision_content_to_list(content_list, video_field, content_extras)
+
+            conversation = [
+                {
+                    "role": "user",
+                    "content": content_list,
+                }
+            ]
+
+            all_messages.append(conversation)
+
+        batch_inputs = self._process_conversation(all_messages)
+        all_have_target_progress = all(tp is not None for tp in target_progress_overrides)
+        if all_have_target_progress:
+            batch_inputs = self._add_progress_meta(
+                batch_inputs,
+                progress_samples,
+                target_progress_override=target_progress_overrides,
+            )
+        batch_inputs["resample_attempts"] = [sample.resample_attempts for sample in progress_samples]
+        return batch_inputs
+
+    def _add_progress_meta(
+        self,
+        batch_inputs: dict[str, torch.Tensor],
+        progress_samples: list[ProgressSample],
+        target_progress_override: list[list[float] | None] | None = None,
+    ) -> dict[str, torch.Tensor]:
+        batch_inputs["sample_type"] = ["progress"] * len(progress_samples)
+        batch_inputs["task"] = [sample.trajectory.task for sample in progress_samples]
+        batch_inputs["metadata"] = [sample.trajectory.metadata for sample in progress_samples]
+
+        # Pad target progress tensors to max length in last dimension
+        if target_progress_override is not None:
+            target_progress_list = target_progress_override
+        else:
+            target_progress_list = [sample.trajectory.target_progress for sample in progress_samples]
+
+        batch_inputs["target_progress"] = pad_list_to_max(target_progress_list)
+        batch_inputs["quality_labels"] = [sample.trajectory.quality_label for sample in progress_samples]
+
+        frames_shape_list = [sample.trajectory.frames_shape for sample in progress_samples]
+        batch_inputs["frames_shape"] = torch.tensor(frames_shape_list, dtype=torch.int32)
+
+        max_length = batch_inputs["target_progress"].shape[-1]
+        batch_inputs["padding_mask"] = create_padding_mask(batch_inputs["frames_shape"], max_length)
+
+        batch_inputs["data_source"] = [sample.trajectory.data_source for sample in progress_samples]
+        batch_inputs["partial_success"] = [sample.trajectory.partial_success for sample in progress_samples]
+        batch_inputs["data_gen_strategy"] = [sample.data_gen_strategy for sample in progress_samples]
+        target_progress_mask = [
+            should_compute_progress(
+                sample.trajectory.quality_label,
+                sample.data_gen_strategy,
+                data_source=sample.trajectory.data_source,
+                partial_success=sample.trajectory.partial_success,
+            )
+            for sample in progress_samples
+        ]
+        batch_inputs["target_progress_mask"] = torch.tensor(target_progress_mask, dtype=torch.float32)
+
+        # Create predict_last_frame masks for trajectories with partial_success
+        predict_last_frame_mask_list = [sample.trajectory.predict_last_frame_mask for sample in progress_samples]
+
+        # Add predict_last_frame_mask (padded to max_length)
+        batch_inputs["predict_last_frame_mask"] = pad_list_to_max(predict_last_frame_mask_list)
+
+        success_label_list = [sample.trajectory.success_label for sample in progress_samples]
+        batch_inputs["success_labels"] = pad_list_to_max(success_label_list)
+
+        return batch_inputs
+
+    def _process_preference_batch(self, preference_samples: list[PreferenceSample]) -> dict[str, torch.Tensor]:
+        """Process a batch of preference samples."""
+        # Collect all messages for batch processing
+        all_messages = []
+
+        # During inference, keep original order (chosen=A, rejected=B)
+        # During training, randomly decide whether chosen trajectory goes first or second
+        if self.inference:
+            # Keep original order: chosen is always A (preference_label=1.0)
+            preference_labels = np.ones(len(preference_samples), dtype=np.int32)
+        else:
+            preference_labels = np.random.randint(0, 2, len(preference_samples))
+
+        # Build batch of conversations
+        for i, sample in enumerate(preference_samples):
+            # Convert frames to appropriate format using stored shapes
+            chosen_frames = convert_frames_to_pil_images(
+                sample.chosen_trajectory.frames, sample.chosen_trajectory.frames_shape
+            )
+            rejected_frames = convert_frames_to_pil_images(
+                sample.rejected_trajectory.frames, sample.rejected_trajectory.frames_shape
+            )
+
+            chosen_video_field, content_extras = self._prepare_frames_for_conversation(
+                chosen_frames, prefix="tmp_chosen"
+            )
+            rejected_video_field, _ = self._prepare_frames_for_conversation(rejected_frames, prefix="tmp_rejected")
+
+            prompt = f"Given these two trajectories for the task '{sample.chosen_trajectory.task}', evaluate which one makes more progress towards the task. Return A for the first trajectory and B for the second trajectory."
+
+            if self.prog_pref:
+                # We ask the model to predict both of the task progress and preference
+                task_prompt = f" Also predict the task progress at each frame of the first trajectory, how far along the robot is towards completing the task, a float between 0 and 1, where 0 is the starting state and 1 is when the task is completed. If the robot is not performing the same task, predict 0 progress."
+                prompt += task_prompt
+
+            # Determine which trajectory is A and which is B based on preference label
+            if preference_labels[i] == 1.0:
+                # Chosen trajectory first: task + video A (chosen) + <|split_token|> + video B (rejected) + <|pref_token|>
+                traj_a_field = chosen_video_field
+                traj_b_field = rejected_video_field
+            else:
+                # Chosen trajectory second: task + video A (rejected) + <|split_token|> + video B (chosen) + <|pref_token|>
+                traj_a_field = rejected_video_field
+                traj_b_field = chosen_video_field
+
+            # Build content list
+            content_list = [
+                {"type": "text", "text": prompt},
+                {"type": "text", "text": "This is Trajectory A. "},
+            ]
+            self._add_vision_content_to_list(content_list, traj_a_field, content_extras)
+
+            content_list.extend([
+                {"type": "text", "text": "<|split_token|>"},
+                {"type": "text", "text": "This is Trajectory B. "},
+            ])
+            self._add_vision_content_to_list(content_list, traj_b_field, content_extras)
+
+            content_list.append({"type": "text", "text": "Now predict the preference between the two trajectories."})
+            content_list.append({"type": "text", "text": "<|pref_token|>"})
+
+            conversation = [
+                {
+                    "role": "user",
+                    "content": content_list,
+                }
+            ]
+            all_messages.append(conversation)
+
+        batch_inputs = self._process_conversation(all_messages)
+        # Use the dynamically generated preference labels based on trajectory order
+        batch_inputs["preference_labels"] = torch.tensor(preference_labels, dtype=torch.float32)
+        batch_inputs = self._add_preference_meta(batch_inputs, preference_samples)
+        return batch_inputs
+
+    def _add_preference_meta(
+        self, batch_inputs: dict[str, torch.Tensor], preference_samples: list[PreferenceSample]
+    ) -> dict[str, torch.Tensor]:
+        batch_inputs["data_source"] = [sample.chosen_trajectory.data_source for sample in preference_samples]
+        batch_inputs["sample_type"] = ["preference"] * len(preference_samples)
+        batch_inputs["task"] = [sample.chosen_trajectory.task for sample in preference_samples]
+        batch_inputs["data_gen_strategy"] = [sample.data_gen_strategy for sample in preference_samples]
+
+        # Determine which trajectory is A and which is B based on preference_label
+        # Trajectory A is chosen if preference_label==1.0, otherwise rejected is A
+        trajectory_A_list = [
+            sample.chosen_trajectory
+            if batch_inputs["preference_labels"][i].item() == 1.0
+            else sample.rejected_trajectory
+            for i, sample in enumerate(preference_samples)
+        ]
+        trajectory_B_list = [
+            sample.rejected_trajectory
+            if batch_inputs["preference_labels"][i].item() == 1.0
+            else sample.chosen_trajectory
+            for i, sample in enumerate(preference_samples)
+        ]
+
+        batch_inputs["trajectory_A_quality_label"] = [traj.quality_label for traj in trajectory_A_list]
+        batch_inputs["trajectory_A_data_source"] = [traj.data_source for traj in trajectory_A_list]
+
+        trajectory_A_data_gen_strategy = []
+        trajectory_B_data_gen_strategy = []
+        for i, sample in enumerate(preference_samples):
+            if batch_inputs["preference_labels"][i].item() == 1.0:
+                trajectory_A_data_gen_strategy.append("subsample_task")
+                trajectory_B_data_gen_strategy.append(sample.data_gen_strategy)
+            else:
+                trajectory_A_data_gen_strategy.append(sample.data_gen_strategy)
+                trajectory_B_data_gen_strategy.append("subsample_task")
+
+        batch_inputs["trajectory_A_data_gen_strategy"] = trajectory_A_data_gen_strategy
+
+        # Add target progress for both trajectories using list comprehensions
+        target_progress_A = [traj.target_progress for traj in trajectory_A_list]
+        # Check if any of the progresses in target_progress_A is None
+        if any(p is None for p in target_progress_A):
+            return batch_inputs
+        target_progress_B = [traj.target_progress for traj in trajectory_B_list]
+        target_progress_A_mask = [
+            should_compute_progress(
+                traj.quality_label,
+                strategy,
+                data_source=traj.data_source,
+                partial_success=traj.partial_success,
+            )
+            for traj, strategy in zip(trajectory_A_list, trajectory_A_data_gen_strategy)
+        ]
+        target_progress_B_mask = [
+            should_compute_progress(
+                traj.quality_label,
+                strategy,
+                data_source=traj.data_source,
+                partial_success=traj.partial_success,
+            )
+            for traj, strategy in zip(trajectory_B_list, trajectory_B_data_gen_strategy)
+        ]
+
+        frames_shape_A = [traj.frames_shape for traj in trajectory_A_list]
+        frames_shape_B = [traj.frames_shape for traj in trajectory_B_list]
+        batch_inputs["frames_shape_A"] = torch.tensor(frames_shape_A, dtype=torch.int32)
+        batch_inputs["frames_shape_B"] = torch.tensor(frames_shape_B, dtype=torch.int32)
+
+        # Create predict_last_frame masks for trajectories with partial_success
+        predict_last_frame_mask_A_list = [traj.predict_last_frame_mask for traj in trajectory_A_list]
+        predict_last_frame_mask_B_list = [traj.predict_last_frame_mask for traj in trajectory_B_list]
+
+        batch_inputs["target_progress_A"] = pad_list_to_max(target_progress_A)
+        batch_inputs["target_progress_B"] = pad_list_to_max(target_progress_B)
+        batch_inputs["target_progress_A_mask"] = torch.tensor(target_progress_A_mask, dtype=torch.float32)
+        batch_inputs["target_progress_B_mask"] = torch.tensor(target_progress_B_mask, dtype=torch.float32)
+        batch_inputs["predict_last_frame_mask_A"] = pad_list_to_max(predict_last_frame_mask_A_list)
+        batch_inputs["predict_last_frame_mask_B"] = pad_list_to_max(predict_last_frame_mask_B_list)
+
+        max_length_A = batch_inputs["target_progress_A"].shape[-1]
+        max_length_B = batch_inputs["target_progress_B"].shape[-1]
+        batch_inputs["padding_mask_A"] = create_padding_mask(batch_inputs["frames_shape_A"], max_length_A)
+        batch_inputs["padding_mask_B"] = create_padding_mask(batch_inputs["frames_shape_B"], max_length_B)
+
+        batch_inputs["chosen_data_gen_strategy"] = [DataGenStrat.FORWARD_PROGRESS.value] * len(preference_samples)
+        batch_inputs["rejected_data_gen_strategy"] = [sample.data_gen_strategy for sample in preference_samples]
+        batch_inputs["chosen_quality_label"] = [sample.chosen_trajectory.quality_label for sample in preference_samples]
+
+        target_progress_chosen = [sample.chosen_trajectory.target_progress for sample in preference_samples]
+        target_progress_rejected = [sample.rejected_trajectory.target_progress for sample in preference_samples]
+        target_progress_chosen_mask = [
+            should_compute_progress(
+                sample.chosen_trajectory.quality_label,
+                DataGenStrat.FORWARD_PROGRESS.value,
+                data_source=sample.chosen_trajectory.data_source,
+                is_chosen=True,
+                partial_success=sample.chosen_trajectory.partial_success,
+            )
+            for sample in preference_samples
+        ]
+        target_progress_rejected_mask = [
+            should_compute_progress(
+                sample.rejected_trajectory.quality_label,
+                sample.data_gen_strategy,
+                data_source=sample.rejected_trajectory.data_source,
+                is_chosen=False,
+                partial_success=sample.rejected_trajectory.partial_success,
+            )
+            for sample in preference_samples
+        ]
+
+        # Create predict_last_frame masks for chosen/rejected trajectories with partial_success
+        predict_last_frame_mask_chosen_list = [
+            sample.chosen_trajectory.predict_last_frame_mask for sample in preference_samples
+        ]
+        predict_last_frame_mask_rejected_list = [
+            sample.rejected_trajectory.predict_last_frame_mask for sample in preference_samples
+        ]
+
+        # Pad target progress tensors to max length in last dimension
+        batch_inputs["target_progress_chosen"] = pad_list_to_max(target_progress_chosen)
+        batch_inputs["target_progress_rejected"] = pad_list_to_max(target_progress_rejected)
+        batch_inputs["target_progress_chosen_mask"] = torch.tensor(target_progress_chosen_mask, dtype=torch.float32)
+        batch_inputs["target_progress_rejected_mask"] = torch.tensor(target_progress_rejected_mask, dtype=torch.float32)
+        batch_inputs["predict_last_frame_mask_chosen"] = pad_list_to_max(predict_last_frame_mask_chosen_list)
+        batch_inputs["predict_last_frame_mask_rejected"] = pad_list_to_max(predict_last_frame_mask_rejected_list)
+
+        batch_inputs["chosen_frames_shape"] = torch.tensor(
+            [sample.chosen_trajectory.frames_shape for sample in preference_samples], dtype=torch.int32
+        )
+        batch_inputs["rejected_frames_shape"] = torch.tensor(
+            [sample.rejected_trajectory.frames_shape for sample in preference_samples], dtype=torch.int32
+        )
+        batch_inputs["resample_attempts"] = [sample.resample_attempts for sample in preference_samples]
+
+        # Aggregate success labels for trajectory A from trajectories
+        success_label_A_list = [traj.success_label for traj in trajectory_A_list]
+        batch_inputs["success_labels_A"] = pad_list_to_max(success_label_A_list)
+
+        # Add metadata structure for evaluation
+        metadata_list = []
+        for sample in preference_samples:
+            metadata = {
+                "chosen_metadata": {
+                    "quality_label": sample.chosen_trajectory.quality_label,
+                    "data_gen_strategy": "subsample_task",
+                    "id": sample.chosen_trajectory.id,
+                    "video_path": sample.chosen_trajectory.metadata.get("video_path")
+                    if sample.chosen_trajectory.metadata
+                    else None,
+                    "partial_success": sample.chosen_trajectory.partial_success,
+                },
+                "rejected_metadata": {
+                    "quality_label": sample.rejected_trajectory.quality_label,
+                    "data_gen_strategy": sample.data_gen_strategy,
+                    "id": sample.rejected_trajectory.id,
+                    "video_path": sample.rejected_trajectory.metadata.get("video_path")
+                    if sample.rejected_trajectory.metadata
+                    else None,
+                    "partial_success": sample.rejected_trajectory.partial_success,
+                },
+            }
+            metadata_list.append(metadata)
+        batch_inputs["metadata"] = metadata_list
+
+        return batch_inputs
diff --git a/robometer/robometer/data/collators/rewind.py b/robometer/robometer/data/collators/rewind.py
new file mode 100644
index 0000000000000000000000000000000000000000..de29b2e592398377c3068169b2e4ab12cd62b332
--- /dev/null
+++ b/robometer/robometer/data/collators/rewind.py
@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import torch
+
+from .rbm_heads import RBMBatchCollator
+from .utils import convert_frames_to_pil_images
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+
+
+class ReWiNDBatchCollator(RBMBatchCollator):
+    def _process_preference_batch(self, preference_samples: list[PreferenceSample]) -> dict[str, torch.Tensor]:
+        """Process a batch of preference samples."""
+        # During inference, keep original order (chosen=A, rejected=B)
+        # During training, randomly decide whether chosen trajectory goes first or second
+        if self.inference:
+            # Keep original order: chosen is always A (preference_label=1.0)
+            preference_labels = np.ones(len(preference_samples), dtype=np.int32)
+        else:
+            preference_labels = np.random.randint(0, 2, len(preference_samples))
+
+        if self.load_embeddings:
+            # Use embeddings directly from trajectories (already loaded by dataset)
+            all_chosen_video_embeddings = []
+            all_rejected_video_embeddings = []
+            all_chosen_text_embeddings = []
+            all_rejected_text_embeddings = []
+
+            for sample in preference_samples:
+                # Get embeddings directly from trajectories
+                chosen_video_emb = sample.chosen_trajectory.video_embeddings
+                chosen_text_emb = sample.chosen_trajectory.text_embedding
+                rejected_video_emb = sample.rejected_trajectory.video_embeddings
+                rejected_text_emb = sample.rejected_trajectory.text_embedding
+
+                if any(
+                    emb is None for emb in [chosen_video_emb, chosen_text_emb, rejected_video_emb, rejected_text_emb]
+                ):
+                    import ipdb
+
+                    ipdb.set_trace()
+                    raise ValueError("Sample trajectories are missing embeddings")
+
+                all_chosen_video_embeddings.append(chosen_video_emb)
+                all_chosen_text_embeddings.append(chosen_text_emb)
+                all_rejected_video_embeddings.append(rejected_video_emb)
+                all_rejected_text_embeddings.append(rejected_text_emb)
+
+            # Stack embeddings into batches
+            chosen_video_embeddings = torch.stack(all_chosen_video_embeddings)  # [B, T, D]
+            rejected_video_embeddings = torch.stack(all_rejected_video_embeddings)  # [B, T, D]
+            chosen_text_embeddings = torch.stack(all_chosen_text_embeddings)  # [B, D]
+            rejected_text_embeddings = torch.stack(all_rejected_text_embeddings)  # [B, D]
+
+            # Interleave embeddings based on preference_labels
+            frame_len = chosen_video_embeddings.shape[1]
+            video_embeddings = torch.empty(len(preference_samples), frame_len * 2, chosen_video_embeddings.shape[2])
+            text_embeddings = torch.empty(len(preference_samples), chosen_text_embeddings.shape[1])
+
+            for i in range(len(preference_samples)):
+                if preference_labels[i] == 1:  # chosen first
+                    video_embeddings[i] = torch.cat([chosen_video_embeddings[i], rejected_video_embeddings[i]], dim=0)
+                    text_embeddings[i] = chosen_text_embeddings[i]  # Use chosen text embedding
+                else:
+                    video_embeddings[i] = torch.cat([rejected_video_embeddings[i], chosen_video_embeddings[i]], dim=0)
+                    text_embeddings[i] = chosen_text_embeddings[i]  # Use chosen text embedding
+
+            batch_inputs = {
+                "video_embeddings": video_embeddings,  # [B, T*2, D]
+                "text_embeddings": text_embeddings,  # [B, D]
+            }
+        else:
+            all_chosen_frames = []
+            all_rejected_frames = []
+            all_tasks = []
+
+            for i, sample in enumerate(preference_samples):
+                # Convert frames to appropriate format using stored shapes
+                # NOTE: these should already be padded to max_frames by the data generator
+                chosen_frames = convert_frames_to_pil_images(
+                    sample.chosen_trajectory.frames, sample.chosen_trajectory.frames_shape
+                )
+                rejected_frames = convert_frames_to_pil_images(
+                    sample.rejected_trajectory.frames, sample.rejected_trajectory.frames_shape
+                )
+                all_chosen_frames.append(chosen_frames)
+                all_rejected_frames.append(rejected_frames)
+                all_tasks.append(sample.chosen_trajectory.task)
+
+            frame_len = len(all_chosen_frames[0])
+            # [(B*T), C, H, W]
+            chosen_video_inputs = self.processor(images=all_chosen_frames, return_tensors="pt")["pixel_values"]
+            _, C, H, W = chosen_video_inputs.shape
+            chosen_video_inputs = chosen_video_inputs.view(len(preference_samples), frame_len, C, H, W)
+            rejected_video_inputs = self.processor(images=all_rejected_frames, return_tensors="pt")["pixel_values"]
+            rejected_video_inputs = rejected_video_inputs.view(len(preference_samples), frame_len, C, H, W)
+
+            # interleave them based on preference_labels
+            video_inputs = torch.empty(len(preference_samples), frame_len * 2, C, H, W)
+            for i in range(len(preference_samples)):
+                if preference_labels[i] == 1:  # means chosen first
+                    video_inputs[i] = torch.cat([chosen_video_inputs[i], rejected_video_inputs[i]], dim=0)
+                else:
+                    video_inputs[i] = torch.cat([rejected_video_inputs[i], chosen_video_inputs[i]], dim=0)
+
+            encodings = self.tokenizer(
+                all_tasks,
+                padding=True,
+                truncation=True,
+                return_tensors="pt",
+            )
+            batch_inputs = {
+                "input_ids": encodings["input_ids"],
+                "attention_mask": encodings["attention_mask"],
+                "pixel_values_videos": video_inputs,
+            }
+
+        batch_inputs["preference_labels"] = torch.tensor(preference_labels, dtype=torch.float32)
+        batch_inputs = self._add_preference_meta(batch_inputs, preference_samples)
+        batch_inputs["resample_attempts"] = [sample.resample_attempts for sample in preference_samples]
+        return batch_inputs
+
+    def _process_progress_batch(self, progress_samples: list[ProgressSample]) -> dict[str, torch.Tensor]:
+        if self.load_embeddings:
+            all_video_embeddings = [sample.trajectory.video_embeddings for sample in progress_samples]
+            all_text_embeddings = [sample.trajectory.text_embedding for sample in progress_samples]
+            video_embeddings = torch.stack(all_video_embeddings)  # [B, T, D]
+            text_embeddings = torch.stack(all_text_embeddings)  # [B, D]
+
+            batch_inputs = {
+                "video_embeddings": video_embeddings,  # [B, T, D]
+                "text_embeddings": text_embeddings,  # [B, D]
+            }
+        else:
+            all_frames = [
+                convert_frames_to_pil_images(sample.trajectory.frames, sample.trajectory.frames_shape)
+                for sample in progress_samples
+            ]
+
+            # here we directly use dino processor process the images and videos to tensors
+            video_inputs = self.processor(images=all_frames, return_tensors="pt")["pixel_values"]
+            frame_len = len(all_frames[0])
+            _, C, H, W = video_inputs.shape
+            video_inputs = video_inputs.view(len(progress_samples), frame_len, C, H, W)
+
+            # here we directly use the tokenizer to process the texts to input_ids and attention_mask
+            texts = [sample.trajectory.task for sample in progress_samples]
+            encodings = self.tokenizer(
+                texts,
+                padding=True,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            batch_inputs = {
+                "input_ids": encodings["input_ids"],
+                "attention_mask": encodings["attention_mask"],
+                "pixel_values_videos": video_inputs,
+            }
+
+        if progress_samples[0].trajectory.target_progress is not None:
+            batch_inputs = self._add_progress_meta(batch_inputs, progress_samples)
+        batch_inputs["resample_attempts"] = [sample.resample_attempts for sample in progress_samples]
+        return batch_inputs
diff --git a/robometer/robometer/data/collators/utils.py b/robometer/robometer/data/collators/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..1350cb967b51e9bec168a3069d3e8c412a222c56
--- /dev/null
+++ b/robometer/robometer/data/collators/utils.py
@@ -0,0 +1,197 @@
+import numpy as np
+import torch
+import av
+
+
+def write_mp4(frames, out_path, fps=4):
+    w, h = frames[0].size
+    c = av.open(str(out_path), mode="w")
+    s = c.add_stream("libx264", rate=fps)
+    s.width, s.height = w, h
+    s.pix_fmt = "yuv420p"
+    s.options = {
+        "preset": "ultrafast",
+        "tune": "zerolatency",
+        "crf": "28",
+        "x264-params": "keyint=1:min-keyint=1:scenecut=0",
+    }
+    for img in frames:
+        frame = av.VideoFrame.from_ndarray(np.array(img), format="rgb24")
+        for pkt in s.encode(frame):
+            c.mux(pkt)
+    for pkt in s.encode(None):
+        c.mux(pkt)
+    c.close()
+
+
+def pad_list_to_max(progress_list):
+    """Helper function to pad lists of sequences to max length.
+
+    Args:
+        progress_list: List of sequences (each sequence is list of floats or tensors)
+
+    Returns:
+        - Continuous mode: tensor of shape (batch, seq_len)
+        - Discrete C51 mode: tensor of shape (batch, seq_len, num_bins)
+    """
+    if not progress_list:
+        return None
+
+    max_length = max(len(progress) for progress in progress_list)
+
+    # Check first non-empty sequence's first element to determine mode
+    first_elem = None
+    for seq in progress_list:
+        if len(seq) > 0:
+            first_elem = seq[0]
+            break
+
+    if first_elem is None:
+        return None
+
+    is_discrete = isinstance(first_elem, torch.Tensor) and first_elem.dim() > 0
+
+    if is_discrete:
+        # Discrete C51 mode: each element is a tensor of shape (num_bins,)
+        num_bins = first_elem.shape[0]
+        pad_tensor = torch.zeros(num_bins, dtype=torch.float32)
+        padded_list = []
+        for progress in progress_list:
+            tensor_list = list(progress)
+            while len(tensor_list) < max_length:
+                tensor_list.append(pad_tensor.clone())
+            padded_list.append(torch.stack(tensor_list))
+        return torch.stack(padded_list)
+    else:
+        # Continuous mode: each element is a float
+        padded_list = []
+        for progress in progress_list:
+            float_list = [float(x) for x in progress]
+            while len(float_list) < max_length:
+                float_list.append(0.0)
+            padded_list.append(float_list)
+        return torch.tensor(padded_list, dtype=torch.float32)
+
+
+def convert_frames_to_pil_images(frames, frames_shape=None):
+    """Convert frames to PIL images if they are numpy arrays or serialized bytes.
+
+    Handles:
+    - Bytes with shape: deserializes to numpy array then converts
+    - Numpy arrays (TxHxWxC or HxWxC): converts each frame to PIL Image
+    - List of numpy arrays: converts each to PIL Image
+    - List of PIL Images: returns as-is
+    - List of mixed types (strings, PIL Images, numpy arrays): converts appropriately
+    """
+    from PIL import Image
+
+    # If frames are serialized bytes, deserialize first
+    if isinstance(frames, bytes):
+        # Deserialize bytes to numpy array (TxHxWxC) using provided shape
+        if frames_shape is not None:
+            # Convert to tuple if it's a list
+            if isinstance(frames_shape, list):
+                frames_shape = tuple(frames_shape)
+            try:
+                frames = np.frombuffer(frames, dtype=np.uint8).reshape(frames_shape)
+            except Exception as e:
+                print(f"Warning: Failed to reshape with provided shape {frames_shape}: {e}")
+                # Fall back to 1D array
+                frames = np.frombuffer(frames, dtype=np.uint8)
+        else:
+            # No shape provided, try to infer
+            frames = np.frombuffer(frames, dtype=np.uint8)
+
+    # If frames are numpy array (TxHxWxC), convert to list of PIL images
+    if isinstance(frames, np.ndarray):
+        pil_images = []
+
+        # Handle different array shapes
+        if len(frames.shape) == 4:  # TxHxWxC
+            for i in range(frames.shape[0]):  # Iterate over time dimension
+                frame = frames[i]  # HxWxC
+                # Ensure uint8 dtype
+                if frame.dtype != np.uint8:
+                    frame = np.clip(frame, 0, 255).astype(np.uint8)
+                pil_image = Image.fromarray(frame)
+                pil_images.append(pil_image)
+        elif len(frames.shape) == 3:  # HxWxC (single frame)
+            frame = frames
+            if frame.dtype != np.uint8:
+                frame = np.clip(frame, 0, 255).astype(np.uint8)
+            pil_image = Image.fromarray(frame)
+            pil_images.append(pil_image)
+        else:
+            raise ValueError(f"Unexpected frames shape {frames.shape}. Expected 3D (HxWxC) or 4D (TxHxWxC) array.")
+
+        return pil_images
+
+    # If frames are list, handle each element
+    if isinstance(frames, list):
+        pil_images = []
+        for frame in frames:
+            if isinstance(frame, str):
+                # File path - open it
+                pil_images.append(Image.open(frame))
+            elif isinstance(frame, Image.Image):
+                # Already PIL Image
+                pil_images.append(frame)
+            elif isinstance(frame, np.ndarray):
+                # Numpy array - convert to PIL
+                if frame.dtype != np.uint8:
+                    frame = np.clip(frame, 0, 255).astype(np.uint8)
+                pil_image = Image.fromarray(frame)
+                pil_images.append(pil_image)
+            else:
+                # Try to convert to numpy array first
+                try:
+                    frame_array = np.array(frame)
+                    if frame_array.dtype != np.uint8:
+                        frame_array = np.clip(frame_array, 0, 255).astype(np.uint8)
+                    pil_images.append(Image.fromarray(frame_array))
+                except Exception as e:
+                    print(f"Warning: Could not convert frame to PIL Image: {e}")
+                    continue
+        return pil_images
+
+    raise ValueError(f"Unsupported frames type: {type(frames)}")
+
+
+def frames_to_numpy_array(frames):
+    """Convert frames to a numpy array of shape (T, H, W, C).
+
+    Handles:
+    - None: returns None
+    - Numpy arrays: returns as-is (ensures uint8 dtype)
+    - List of strings (file paths): opens each path and converts to array
+    - List of PIL Images: converts each to array
+    - List of numpy arrays: stacks them
+    """
+    from PIL import Image
+
+    if frames is None:
+        return None
+
+    if isinstance(frames, np.ndarray):
+        # Already numpy array - ensure uint8 dtype
+        frames_array = frames
+        if frames_array.dtype != np.uint8:
+            frames_array = np.clip(frames_array, 0, 255).astype(np.uint8)
+        return frames_array
+    elif isinstance(frames, list):
+        # Convert list of images/paths to array
+        frame_list = []
+        for frame in frames:
+            if isinstance(frame, str):
+                img = np.array(Image.open(frame))
+            elif isinstance(frame, Image.Image):
+                img = np.array(frame)
+            elif isinstance(frame, np.ndarray):
+                img = frame
+            else:
+                img = np.array(frame)
+            frame_list.append(img)
+        if frame_list:
+            return np.stack(frame_list)
+
+    return None
diff --git a/robometer/robometer/data/dataset_category.py b/robometer/robometer/data/dataset_category.py
new file mode 100644
index 0000000000000000000000000000000000000000..817ed54042df8176971bf0d9483ca741b08f001d
--- /dev/null
+++ b/robometer/robometer/data/dataset_category.py
@@ -0,0 +1,558 @@
+#!/usr/bin/env python3
+"""
+Dataset categorization for RBM training.
+
+This module defines categories for datasets to control sampling behavior:
+- success: Datasets containing primarily successful trajectories
+- failure: Datasets containing failure trajectories
+- preference_only: Datasets that should only generate preference samples (no progress)
+- paired: Datasets containing paired human/robot trajectories
+"""
+
+from typing import List
+
+ALL_DATASOURCES = [
+    "auto_eval_rfm",
+    "failsafe",
+    "fino_net",
+    "h2r",
+    "metaworld_train",
+    "molmoact_dataset_household",
+    "molmoact_dataset_tabletop",
+    "motif_rfm",
+    "oxe_aloha_mobile",
+    "oxe_austin_buds_dataset_converted_externally_to_rlds",
+    "oxe_bc_z",
+    "oxe_berkeley_cable_routing",
+    "oxe_berkeley_fanuc_manipulation",
+    "oxe_berkeley_mvp_converted_externally_to_rlds",
+    "oxe_berkeley_rpt_converted_externally_to_rlds",
+    "oxe_bridge_v2",
+    "oxe_dlr_edan_shared_control_converted_externally_to_rlds",
+    "oxe_droid",
+    "oxe_fractal20220817_data",
+    "oxe_furniture_bench_dataset_converted_externally_to_rlds",
+    "oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds",
+    "oxe_imperialcollege_sawyer_wrist_cam",
+    "oxe_jaco_play",
+    "oxe_language_table",
+    "oxe_nyu_rot_dataset_converted_externally_to_rlds",
+    "oxe_robo_set",
+    "oxe_stanford_hydra_dataset_converted_externally_to_rlds",
+    "oxe_tokyo_u_lsmo_converted_externally_to_rlds",
+    "oxe_toto",
+    "oxe_ucsd_kitchen_dataset_converted_externally_to_rlds",
+    "oxe_utaustin_mutex",
+    "ph2d",
+    "racer_train",
+    "rh20t_human",
+    "rh20t_robot",
+    "roboarena",
+    "soar_rfm",
+    "usc_koch_human_robot_paired_human",
+    "usc_koch_human_robot_paired_robot",
+    "hand_paired_human",
+    "hand_paired_robot",
+]
+
+DATASET_CATEGORY = {
+    "success": [
+        # Most datasets are success by default, so this list is intentionally minimal
+        # Only explicitly list if needed for special handling
+    ],
+    "failure": [
+        "ykorkmaz_libero_failure_rfm_libero_90_failure",
+        "ykorkmaz_libero_failure_rfm_libero_10_failure",
+    ],
+    "preference_only": [
+        "jesbu1_oxe_rfm_oxe_bc_z",
+        "jesbu1_oxe_rfm_oxe_dlr_edan_shared_control_converted_externally_to_rlds",
+        "jesbu1_roboarena_0825_rfm_roboarena",
+        "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+        # "roboarena",
+        "auto_eval_rfm",
+        "jesbu1_epic_rfm_epic",
+    ],
+    "paired": [
+        "jesbu1_h2r_rfm_h2r",
+        "jesbu1_motif_rfm_motif_rfm",
+        "anqil_rh20t_subset_rfm_rh20t_human",
+        "anqil_rh20t_subset_rfm_rh20t_robot",
+        "jesbu1_ph2d_rfm_ph2d",
+        "jesbu1_hand_paired_rfm_hand_paired_human",
+        "jesbu1_hand_paired_rfm_hand_paired_robot",
+        "jesbu1_usc_koch_human_robot_paired_human",
+        "jesbu1_usc_koch_human_robot_paired_robot",
+    ],
+}
+
+DATASET_MAP = {
+    "others": {
+        "train": [
+            "jesbu1_molmoact_rfm_molmoact_dataset_household",
+            "jesbu1_molmoact_rfm_molmoact_dataset_tabletop",
+            # "jesbu1_egodex_rfm_egodex_part1",
+            # "jesbu1_egodex_rfm_egodex_part2",
+            # "jesbu1_egodex_rfm_egodex_part3",
+            # "jesbu1_egodex_rfm_egodex_part4",
+            # "jesbu1_egodex_rfm_egodex_part5",
+            "abraranwar_agibotworld_alpha_headcam_rfm_agibotworld",
+            "jesbu1_humanoid_everyday_rfm_humanoid_everyday_rfm",
+            "jesbu1_epic_rfm_epic",
+            "jesbu1_galaxea_rfm_galaxea_part1_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part2_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part3_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part4_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part5_r1_lite",
+        ],
+        # "eval": [
+        #    #"jesbu1_egodex_rfm_egodex_test",
+        # ],
+    },
+    "oxe": {
+        "train": [
+            "jesbu1_oxe_rfm_oxe_aloha_mobile",
+            "jesbu1_oxe_rfm_oxe_austin_buds_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_bc_z",
+            "jesbu1_oxe_rfm_oxe_berkeley_cable_routing",
+            "jesbu1_oxe_rfm_oxe_berkeley_fanuc_manipulation",
+            "jesbu1_oxe_rfm_oxe_berkeley_mvp_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_berkeley_rpt_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_bridge_v2",
+            "jesbu1_oxe_rfm_oxe_dlr_edan_shared_control_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_droid",
+            "jesbu1_oxe_rfm_oxe_fractal20220817_data",
+            "jesbu1_oxe_rfm_oxe_furniture_bench_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_imperialcollege_sawyer_wrist_cam",
+            "jesbu1_oxe_rfm_oxe_jaco_play",
+            "jesbu1_oxe_rfm_oxe_language_table",
+            "jesbu1_oxe_rfm_oxe_nyu_rot_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_robo_set",
+            "jesbu1_oxe_rfm_oxe_stanford_hydra_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_tokyo_u_lsmo_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_toto",
+            "jesbu1_oxe_rfm_oxe_ucsd_kitchen_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_utaustin_mutex",
+        ],
+        "eval": [
+            "jesbu1_oxe_rfm_eval_oxe_bc_z_eval",
+            "jesbu1_oxe_rfm_eval_oxe_berkeley_cable_routing_eval",
+            "jesbu1_oxe_rfm_eval_oxe_bridge_v2_eval",
+            "jesbu1_oxe_rfm_eval_oxe_jaco_play_eval",
+            "jesbu1_oxe_rfm_eval_oxe_toto_eval",
+            "jesbu1_oxe_rfm_eval_oxe_viola_eval",
+        ],
+    },
+    "mw": {
+        "train": [
+            "aliangdw_metaworld_metaworld_train",
+        ],
+        "eval": [
+            "aliangdw_metaworld_metaworld_eval",
+        ],
+    },
+    "reward_alignment": {
+        "eval": [
+            # "aliangdw_metaworld_metaworld_eval",
+            "abraranwar_usc_koch_rewind_rfm_usc_koch_rewind",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            "jesbu1_soar_rfm_soar_rfm",
+            # "jesbu1_egodex_rfm_egodex_test",
+        ]
+    },
+    "quality_preference": {
+        "eval": [
+            # "aliangdw_metaworld_metaworld_eval",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            # "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking",
+            # "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking",
+            # "aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking",
+            "jesbu1_soar_rfm_soar_rfm",
+        ]
+    },
+    "policy_ranking": {
+        "eval": [
+            # "aliangdw_metaworld_metaworld_eval",
+            # "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            # "jesbu1_roboarena_0825_rfm_roboarena",
+            "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking",
+            "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking",
+            "aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking",
+            "jesbu1_soar_rfm_soar_rfm",
+            "jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm",
+            "jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top",
+            "jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist",
+            "jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all",
+            # "jesbu1_roboreward_rfm_roboreward_test",
+            "jesbu1_roboreward_rfm_roboreward_high_res_test",
+            # ["abraranwar_libero_rfm_libero256_10", "ykorkmaz_libero_failure_rfm_libero_10_failure"],
+        ]
+    },
+    "paired": {
+        "train": [
+            "jesbu1_h2r_rfm_h2r",
+            "jesbu1_motif_rfm_motif_rfm",
+            "anqil_rh20t_subset_rfm_rh20t_human",
+            "anqil_rh20t_subset_rfm_rh20t_robot",
+            "jesbu1_ph2d_rfm_ph2d",
+        ],
+        "eval": [
+            # "jesbu1_egodex_rfm_egodex_test",
+            ["jesbu1_hand_paired_rfm_hand_paired_human", "jesbu1_hand_paired_rfm_hand_paired_robot"],
+            [
+                "jesbu1_usc_koch_human_robot_paired_usc_koch_human_robot_paired_human",
+                "jesbu1_usc_koch_human_robot_paired_usc_koch_human_robot_paired_robot",
+            ],
+            ["anqil_rh20t_subset_rfm_rh20t_human", "anqil_rh20t_subset_rfm_rh20t_robot"],
+            # "aliangdw_utd_so101_human_utd_so101_human",
+            # "aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking",
+        ],
+    },
+    "suboptimal_fail": {
+        "train": [
+            # "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            "jesbu1_fino_net_rfm_fino_net",
+            "jesbu1_failsafe_rfm_failsafe",
+            "jesbu1_soar_rfm_soar_rfm",
+            "jesbu1_auto_eval_rfm_auto_eval_rfm",
+            "jesbu1_racer_rfm_racer_train",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_roboreward_rfm_roboreward_high_res_train",
+            # "jesbu1_roboreward_rfm_roboreward_train",
+        ],
+        "eval": [
+            "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            "jesbu1_racer_rfm_racer_val",
+        ],
+    },
+    "franka": {
+        "train": [
+            "jesbu1_oxe_rfm_oxe_droid",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_molmoact_rfm_molmoact_dataset_household",
+            "jesbu1_molmoact_rfm_molmoact_dataset_tabletop",
+        ],
+        "eval": [
+            "jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm",
+            "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            # "jesbu1_roboarena_0825_rfm_roboarena",
+            "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking",
+            "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking",
+        ],
+    },
+    "libero": {
+        "train": [
+            "abraranwar_libero_rfm_libero256_90",
+            "ykorkmaz_libero_failure_rfm_libero_90_failure",
+        ],
+        "eval": [
+            "abraranwar_libero_rfm_libero256_10",
+            "abraranwar_libero_rfm_libero256_object",
+            "abraranwar_libero_rfm_libero256_spatial",
+            "abraranwar_libero_rfm_libero256_goal",
+            # "ykorkmaz_libero_failure_rfm_libero_90_failure",
+            # "ykorkmaz_libero_failure_rfm_libero_10_failure",
+        ],
+    },
+    "libero_pi0": {
+        "train": [
+            "abraranwar_libero_rfm_libero256_10",
+            "abraranwar_libero_rfm_libero256_object",
+            "abraranwar_libero_rfm_libero256_spatial",
+            "abraranwar_libero_rfm_libero256_goal",
+            "ykorkmaz_libero_failure_rfm_libero_10_failure",
+            "ykorkmaz_libero_failure_rfm_libero_object_failure",
+            "ykorkmaz_libero_failure_rfm_libero_spatial_failure",
+            "ykorkmaz_libero_failure_rfm_libero_goal_failure",
+        ],
+        "eval": [
+            [
+                "abraranwar_libero_rfm_libero256_90",
+                "ykorkmaz_libero_failure_rfm_libero_90_failure",
+            ],
+            [
+                "abraranwar_libero_rfm_libero256_10",
+                "ykorkmaz_libero_failure_rfm_libero_10_failure",
+            ],
+        ],
+    },
+    "rbm-1m-id": {
+        "train": [
+            # franka
+            "jesbu1_oxe_rfm_oxe_droid",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_molmoact_rfm_molmoact_dataset_household",
+            "jesbu1_molmoact_rfm_molmoact_dataset_tabletop",
+            # oxe
+            "jesbu1_oxe_rfm_oxe_aloha_mobile",
+            "jesbu1_oxe_rfm_oxe_austin_buds_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_bc_z",
+            "jesbu1_oxe_rfm_oxe_berkeley_cable_routing",
+            "jesbu1_oxe_rfm_oxe_berkeley_fanuc_manipulation",
+            "jesbu1_oxe_rfm_oxe_berkeley_mvp_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_berkeley_rpt_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_bridge_v2",
+            "jesbu1_oxe_rfm_oxe_dlr_edan_shared_control_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_droid",
+            "jesbu1_oxe_rfm_oxe_fractal20220817_data",
+            "jesbu1_oxe_rfm_oxe_furniture_bench_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_imperialcollege_sawyer_wrist_cam",
+            "jesbu1_oxe_rfm_oxe_jaco_play",
+            "jesbu1_oxe_rfm_oxe_language_table",
+            "jesbu1_oxe_rfm_oxe_nyu_rot_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_robo_set",
+            "jesbu1_oxe_rfm_oxe_stanford_hydra_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_tokyo_u_lsmo_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_toto",
+            "jesbu1_oxe_rfm_oxe_ucsd_kitchen_dataset_converted_externally_to_rlds",
+            "jesbu1_oxe_rfm_oxe_utaustin_mutex",
+            # others
+            "jesbu1_molmoact_rfm_molmoact_dataset_household",
+            "jesbu1_molmoact_rfm_molmoact_dataset_tabletop",
+            "abraranwar_agibotworld_alpha_headcam_rfm_agibotworld",
+            "jesbu1_humanoid_everyday_rfm_humanoid_everyday_rfm",
+            "jesbu1_epic_rfm_epic",
+            "jesbu1_galaxea_rfm_galaxea_part1_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part2_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part3_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part4_r1_lite",
+            "jesbu1_galaxea_rfm_galaxea_part5_r1_lite",
+            # libero
+            "abraranwar_libero_rfm_libero256_10",
+            "abraranwar_libero_rfm_libero256_object",
+            "abraranwar_libero_rfm_libero256_spatial",
+            "abraranwar_libero_rfm_libero256_goal",
+            "ykorkmaz_libero_failure_rfm_libero_10_failure",
+            "ykorkmaz_libero_failure_rfm_libero_object_failure",
+            "ykorkmaz_libero_failure_rfm_libero_spatial_failure",
+            # suboptimal_fail
+            "jesbu1_fino_net_rfm_fino_net",
+            "jesbu1_failsafe_rfm_failsafe",
+            "jesbu1_soar_rfm_soar_rfm",
+            "jesbu1_auto_eval_rfm_auto_eval_rfm",
+            "jesbu1_racer_rfm_racer_train",
+            "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_roboreward_rfm_roboreward_high_res_train",
+            # paired
+            "jesbu1_h2r_rfm_h2r",
+            "jesbu1_motif_rfm_motif_rfm",
+            "anqil_rh20t_subset_rfm_rh20t_human",
+            "anqil_rh20t_subset_rfm_rh20t_robot",
+            "jesbu1_ph2d_rfm_ph2d",
+            # mw
+            "aliangdw_metaworld_metaworld_train",
+        ],
+        "eval": [
+            # "abraranwar_libero_rfm_libero256_10",
+            # "abraranwar_libero_rfm_libero256_object",
+            # "abraranwar_libero_rfm_libero256_spatial",
+            # "abraranwar_libero_rfm_libero256_goal",
+            # "ykorkmaz_libero_failure_rfm_libero_10_failure",
+            # "ykorkmaz_libero_failure_rfm_libero_object_failure",
+            # "ykorkmaz_libero_failure_rfm_libero_spatial_failure",
+            # "ykorkmaz_libero_failure_rfm_libero_goal_failure",
+            # "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",
+            # "jesbu1_roboarena_0825_rfm_roboarena",
+            "jesbu1_racer_rfm_racer_val",
+            # "jesbu1_soar_rfm_soar_rfm",
+            # "jesbu1_auto_eval_rfm_auto_eval_rfm",
+            # "jesbu1_roboreward_rfm_roboreward_test",
+            "jesbu1_oxe_rfm_eval_oxe_bc_z_eval",
+            "jesbu1_oxe_rfm_eval_oxe_berkeley_cable_routing_eval",
+            "jesbu1_oxe_rfm_eval_oxe_bridge_v2_eval",
+            "jesbu1_oxe_rfm_eval_oxe_jaco_play_eval",
+            "jesbu1_oxe_rfm_eval_oxe_toto_eval",
+            "jesbu1_oxe_rfm_eval_oxe_viola_eval",
+            "aliangdw_metaworld_metaworld_eval",
+            "abraranwar_libero_rfm_libero256_90",
+        ],
+    },
+    "rbm-1m-ood": {
+        "eval": [
+            # "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist",  # putting this here for now
+            # "jesbu1_roboreward_rfm_roboreward_test",
+            # "jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm",
+            "ykorkmaz_usc_trossen_rfm_usc_trossen",
+            "jesbu1_rfm_new_mit_franka_rfm_rfm_new_mit_franka_rfm",
+            "jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top",
+            "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking",
+            "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking",
+            "jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all",
+        ]
+    },
+    "libero_pi0_no_fail": {
+        "train": [
+            "abraranwar_libero_rfm_libero256_10",
+            "abraranwar_libero_rfm_libero256_object",
+            "abraranwar_libero_rfm_libero256_spatial",
+            "abraranwar_libero_rfm_libero256_goal",
+        ],
+        "eval": [
+            [
+                "abraranwar_libero_rfm_libero256_90",
+                "ykorkmaz_libero_failure_rfm_libero_90_failure",
+            ],
+            [
+                "abraranwar_libero_rfm_libero256_10",
+                "ykorkmaz_libero_failure_rfm_libero_10_failure",
+            ],
+        ],
+    },
+}
+
+
+DATA_SOURCE_CATEGORY = {
+    "success": [
+        # All datasets are success by default, so this list is intentionally minimal
+        # Only explicitly list if needed for special handling
+    ],
+    "failure": [
+        "libero_90_failure",
+        "libero_10_failure",
+        "libero_object_failure",
+        "libero_spatial_failure",
+        "libero_goal_failure",
+    ],
+    "preference_only": [
+        "oxe_bc_z",
+        "oxe_dlr_edan_shared_control_converted_externally_to_rlds",
+        "auto_eval_rfm",
+        # "humanoid_everyday_rfm",
+        "galaxea_part1_r1_lite",
+        "galaxea_part2_r1_lite",
+        "galaxea_part3_r1_lite",
+        "galaxea_part4_r1_lite",
+        "galaxea_part5_r1_lite",
+        "epic",
+        "roboreward_roboarena",
+        "roboreward_dlr_edan_shared_control_converted_externally_to_rldsroboarena",
+        "roboarena",
+        "roboarena_eval_debug_nowrist",
+    ],
+    "paired": [
+        "h2r",
+        "motif_rfm",
+        "rh20t_human",
+        "rh20t_robot",
+        "ph2d",
+        "hand_paired_human",
+        "hand_paired_robot",
+        "usc_koch_human_robot_paired_human",
+        "usc_koch_human_robot_paired_robot",
+    ],
+    "suboptimal_fail": [
+        "roboarena",
+        "fino_net",
+        "failsafe",
+        "soar_rfm",
+        "auto_eval_rfm",
+        "racer_train",
+        "racer_val",
+        "roboarena_eval_debug_nowrist",
+    ],
+    "franka": [
+        "oxe_droid",
+        "roboarena",
+        "molmoact_dataset_household",
+        "molmoact_dataset_tabletop",
+    ],
+}
+
+# DATASET helper functions
+
+
+def get_preference_only_datasets() -> List[str]:
+    """Get list of datasets that should only generate preference samples."""
+    return DATASET_CATEGORY.get("preference_only", [])
+
+
+def get_paired_datasets() -> List[str]:
+    """Get list of datasets containing paired human/robot trajectories."""
+    return DATASET_CATEGORY.get("paired", [])
+
+
+def get_failure_datasets() -> List[str]:
+    """Get list of datasets containing failure trajectories."""
+    return DATASET_CATEGORY.get("failure", [])
+
+
+def get_success_datasets() -> List[str]:
+    """Get list of datasets containing successful trajectories."""
+    return DATASET_CATEGORY.get("success", [])
+
+
+def is_preference_only(dataset_name: str) -> bool:
+    """Check if a dataset should only generate preference samples."""
+    return dataset_name in DATASET_CATEGORY.get("preference_only", [])
+
+
+def is_paired(dataset_name: str) -> bool:
+    """Check if a dataset contains paired human/robot trajectories."""
+    return dataset_name in DATASET_CATEGORY.get("paired", [])
+
+
+def is_failure(dataset_name: str) -> bool:
+    """Check if a dataset contains failure trajectories."""
+    return dataset_name in DATASET_CATEGORY.get("failure", [])
+
+
+def is_success(dataset_name: str) -> bool:
+    """Check if a dataset contains successful trajectories."""
+    return dataset_name in DATASET_CATEGORY.get("success", [])
+
+
+# DATA_SOURCE helper functions
+
+
+def get_preference_only_ds() -> List[str]:
+    """Get list of data sources that should only generate preference samples."""
+    return DATA_SOURCE_CATEGORY.get("preference_only", [])
+
+
+def get_paired_ds() -> List[str]:
+    """Get list of data sources containing paired human/robot trajectories."""
+    return DATA_SOURCE_CATEGORY.get("paired", [])
+
+
+def get_failure_ds() -> List[str]:
+    """Get list of data sources containing failure trajectories."""
+    return DATA_SOURCE_CATEGORY.get("failure", [])
+
+
+def get_success_ds() -> List[str]:
+    """Get list of data sources containing successful trajectories."""
+    return DATA_SOURCE_CATEGORY.get("success", [])
+
+
+def is_preference_only_ds(data_source_name: str) -> bool:
+    """Check if a data source should only generate preference samples."""
+    return data_source_name in DATA_SOURCE_CATEGORY.get("preference_only", [])
+
+
+def is_paired_ds(data_source_name: str) -> bool:
+    """Check if a data source contains paired human/robot trajectories."""
+    return data_source_name in DATA_SOURCE_CATEGORY.get("paired", [])
+
+
+def is_failure_ds(data_source_name: str) -> bool:
+    """Check if a data source contains failure trajectories."""
+    return data_source_name in DATA_SOURCE_CATEGORY.get("failure", [])
+
+
+def is_success_ds(data_source_name: str) -> bool:
+    """Check if a data source contains successful trajectories."""
+    return data_source_name in DATA_SOURCE_CATEGORY.get("success", [])
+
+
+def get_suboptimal_fail_ds() -> List[str]:
+    """Get list of data sources containing suboptimal/failure trajectories."""
+    return DATA_SOURCE_CATEGORY.get("suboptimal_fail", [])
+
+
+def is_suboptimal_fail_ds(data_source_name: str) -> bool:
+    """Check if a data source contains suboptimal/failure trajectories."""
+    return data_source_name in DATA_SOURCE_CATEGORY.get("suboptimal_fail", [])
diff --git a/robometer/robometer/data/dataset_success_cutoff.txt b/robometer/robometer/data/dataset_success_cutoff.txt
new file mode 100644
index 0000000000000000000000000000000000000000..e9a1caf44ff8c48d40f37f15df6c8df6e4495dd1
--- /dev/null
+++ b/robometer/robometer/data/dataset_success_cutoff.txt
@@ -0,0 +1,65 @@
+oxe_aloha_mobile,0.95
+oxe_austin_buds_dataset_converted_externally_to_rlds,0.95
+oxe_berkeley_cable_routing,0.95 
+oxe_berkeley_fanuc_manipulation,0.98
+oxe_bridge_v2,0.95
+oxe_dlr_edan_shared_control_converted_externally_to_rlds,0.95
+oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds,0.95
+oxe_ucsd_kitchen_dataset_converted_externally_to_rlds,0.95
+oxe_utaustin_mutex,0.95
+oxe_bc_z,0.95
+oxe_berkeley_mvp_converted_externally_to_rlds,1.0
+oxe_berkeley_rpt_converted_externally_to_rlds,0.76
+oxe_fractal20220817_data,1.0
+oxe_furniture_bench_dataset_converted_externally_to_rlds,1.0
+oxe_droid,0.95
+oxe_imperialcollege_sawyer_wrist_cam,0.9
+oxe_language_table,1.0
+oxe_nyu_rot_dataset_converted_externally_to_rlds,0.7
+oxe_robo_set,0.85
+oxe_stanford_hydra_dataset_converted_externally_to_rlds,1.0
+oxe_tokyo_u_lsmo_converted_externally_to_rlds,1.0
+oxe_toto,1.0
+roboreward_austin_buds_dataset_converted_externally_to_rlds,0.95
+roboreward_austin_sirius_dataset_converted_externally_to_rlds,0.98
+roboreward_berkeley_autolab_ur5,0.90
+roboreward_berkeley_cable_routing,0.95 
+roboreward_berkeley_fanuc_manipulation,0.98
+roboreward_bridge,0.95
+roboreward_dlr_edan_shared_control_converted_externally_to_rlds,0.95
+roboreward_iamlab_cmu_pickup_insert_converted_externally_to_rlds,0.95
+roboreward_ucsd_kitchen_dataset_converted_externally_to_rlds,0.95
+roboreward_utaustin_mutex,0.95
+roboreward_berkeley_mvp_converted_externally_to_rlds,1.0
+roboreward_berkeley_rpt_converted_externally_to_rlds,0.76
+roboreward_fractal20220817_data,1.0
+roboreward_droid,0.95
+roboreward_imperialcollege_sawyer_wrist_cam,0.9
+roboreward_language_table,1.0
+roboreward_nyu_rot_dataset_converted_externally_to_rlds,0.7
+roboreward_robo_set,0.85
+roboreward_ucsd_pick_and_place_dataset_converted_externally_to_rlds,0.78
+roboreward_roboturk,1.0
+roboreward_stanford_hydra_dataset_converted_externally_to_rlds,1.0
+roboreward_tokyo_u_lsmo_converted_externally_to_rlds,1.0
+roboreward_kaist_nonprehensile_converted_externally_to_rlds,1.0
+roboreward_utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds,0.78
+roboreward_utokyo_xarm_bimanual_converted_externally_to_rlds,1.0
+roboreward_cmu_play_fusion,0.97
+roboreward_toto,1.0
+metaworld_train,1.0
+molmoact_dataset_household,0.94
+molmoact_dataset_tabletop,0.94
+agibotworld,0.95
+ph2d,0.95
+soar_rfm,0.95
+auto_eval_rfm,0.94
+rh20t_human,0.92
+rh20t_robot,0.94
+galaxea_rfm,0.8
+fino_net_rfm,0.75
+h2r,0.9
+humanoid_everyday_rfm,0.80
+motif_rfm,0.95
+roboarena_eval_debug_nowrist,0.90
+roboarena,0.90
\ No newline at end of file
diff --git a/robometer/robometer/data/dataset_types.py b/robometer/robometer/data/dataset_types.py
new file mode 100755
index 0000000000000000000000000000000000000000..79f7c897e821dc25dfcb7bf8bb06615a45f15f74
--- /dev/null
+++ b/robometer/robometer/data/dataset_types.py
@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+"""
+Dataclasses for RBM model dataset trajectory structures.
+Defines the standard format for HuggingFace dataset trajectories.
+"""
+
+from typing import Any, Union, List, Dict, Tuple, Optional
+
+import numpy as np
+from pydantic import BaseModel, ConfigDict
+import torch
+
+
+class Trajectory(BaseModel):
+    """Trajectory structure containing frames, metadata, and progress information."""
+
+    # Core trajectory fields
+    frames: Union[List[str], np.ndarray, None] = None
+    frames_shape: Optional[Tuple] = None
+
+    # If embeddings are precomputed
+    embeddings_path: Optional[str] = None
+    video_embeddings: Union[torch.Tensor, np.ndarray, None] = None
+    text_embedding: Union[torch.Tensor, np.ndarray, None] = None
+
+    id: Optional[str] = None
+    task: Optional[str] = None
+    lang_vector: Union[np.ndarray, List[float], None] = None
+    data_source: Optional[str] = None
+    quality_label: Optional[str] = None
+    is_robot: Optional[bool] = None
+
+    # Progress and metadata
+    # Can be List[float] for continuous progress, np.ndarray, or List[np.ndarray] for C51 discrete distributions
+    target_progress: Optional[Union[List[float], List[torch.Tensor], torch.Tensor, None]] = None
+    partial_success: Optional[Union[float, torch.Tensor]] = None  # float for continuous, Tensor for C51 discrete
+    success_label: Optional[List[float]] = None  # Success labels for each frame (1.0 for success, 0.0 for failure)
+    predict_last_frame_mask: Optional[List[float]] = (
+        None  # Mask for partial_success: 1.0 for last frame if partial_success < 1.0, otherwise all 1.0s
+    )
+    metadata: Optional[Dict[str, Any]] = None
+    data_gen_strategy: Optional[str] = None
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+
+class ProgressSample(BaseModel):
+    """Sample structure for progress evaluation."""
+
+    trajectory: Trajectory
+    sample_type: str = "progress"
+    data_gen_strategy: Optional[str] = None
+    resample_attempts: int = 1
+
+
+class PreferenceSample(BaseModel):
+    """Sample structure for preference prediction: chosen vs rejected where chosen is preferred."""
+
+    # Trajectories
+    chosen_trajectory: Trajectory
+    rejected_trajectory: Trajectory
+
+    sample_type: str = "preference"
+    data_gen_strategy: Optional[str] = None
+    resample_attempts: int = 1
+
+
+SampleType = Union[PreferenceSample, ProgressSample]
diff --git a/robometer/robometer/data/datasets/README.md b/robometer/robometer/data/datasets/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..c59bbc16a9a2f03d2e2eeaf247366fe995d9b678
--- /dev/null
+++ b/robometer/robometer/data/datasets/README.md
@@ -0,0 +1,153 @@
+# StrategyFirstDataset Documentation
+
+## Overview
+
+The `StrategyFirstDataset` is a dataset class that first selects a sample type (preference/progress), then a strategy, then picks a data source uniformly. This is different from other dataset classes that select trajectories first.
+
+**Sampling Flow:**
+1. Select sample type (preference/progress) based on `sample_type_ratio`
+2. Select strategy for that sample type based on strategy ratios
+3. Filter indices based on strategy requirements
+4. Select data source uniformly from available data sources
+5. Sample trajectory from selected data source and generate sample
+
+## Strategy Filtering
+
+### REWIND Strategy
+- **Requirement**: Only uses **successful trajectories** (`quality_label == "successful"`)
+- **Implementation**: Filters indices to only include trajectories from `quality_indices["successful"]`
+- **Reason**: We should not rewind suboptimal trajectories; only rewind successful ones to create suboptimal examples
+
+### SUBOPTIMAL Strategy
+- **Requirement**: Filters to tasks that have both optimal and suboptimal trajectories (unless RoboArena/RoboReward)
+- **RoboArena/RoboReward Exception**: For datasets with `partial_success` (RoboArena/RoboReward), task-level filtering is skipped since they use `partial_success` logic
+
+## Preference Sampling Details
+
+### Chosen Trajectory (Trajectory A)
+
+The chosen trajectory in a preference sample must be:
+- **Successful trajectories only** (`quality_label == "successful"`) OR
+- **Trajectories with `partial_success = 1.0`** (full success)
+
+**Subsequence Strategy**: Uses `subsample_forward` by default (normal forward sampling)
+
+**Progress & Success Prediction**:
+- **Always predict progress and success** if it's trajectory A (chosen trajectory)
+- **Exception**: If the chosen trajectory has `partial_success < 1.0`, do **NOT** predict success (only predict progress)
+
+**Partial Success Handling**:
+- Can be a `partial_success` trajectory with value < 1.0
+- In this case, progress is predicted but success is NOT predicted
+
+### Rejected Trajectory
+
+The rejected trajectory depends on the strategy used:
+
+#### 1. REWIND Strategy
+- **Chosen trajectory requirement**: Must be successful (`quality_label == "successful"`)
+  - Enforced by `StrategyFirstDataset._filter_indices_by_strategy()` which filters REWIND to only use successful indices
+- **Subsequence strategy**: Uses `subsample_rewind` (rewound version of the chosen trajectory)
+- **Progress prediction**: ✅ Yes (progress is computed)
+- **Success prediction**: ✅ Yes (success labels are computed from progress)
+
+#### 2. DIFFERENT_TASK Strategy
+- **Chosen trajectory requirement**: No specific requirement (can be any trajectory)
+- **Subsequence strategy**: Uses `subsample_forward` (normal forward sampling)
+- **Rejected trajectory**: Any trajectory from a completely different task
+- **Progress prediction**: ✅ Yes (but progress is set to `[0.0]` for all timesteps)
+- **Success prediction**: ✅ Yes (but success labels are `[0.0]` for all timesteps, computed from progress=0.0)
+  - Implementation: `target_progress = [0.0]` → `success_labels = [0.0]` (via `compute_success_labels`)
+
+#### 3. REVERSE_PROGRESS Strategy
+- **Chosen trajectory requirement**: Must be successful (`quality_label == "successful"`)
+  - **Enforced by**: `StrategyFirstDataset._filter_indices_by_strategy()` filters REVERSE_PROGRESS (for preference samples) to only use successful indices
+- **Subsequence strategy**: Uses `subsample_reverse` (reverse version of the chosen trajectory)
+- **Progress prediction**: ✅ Yes (progress is computed, but in reverse order)
+- **Success prediction**: ✅ Yes (success labels are computed from progress)
+
+#### 4. SUBOPTIMAL Strategy
+- **Chosen trajectory requirement**: Must be from a task that has suboptimal trajectories
+- **Subsequence strategy**: Uses `subsample_forward` (normal forward sampling)
+- **Rejected trajectory**: Suboptimal/failure trajectory from the same task
+- **Partial Success Handling**: 
+  - For trajectories with `partial_success`, chosen trajectory must have higher `partial_success` than rejected
+  - If found trajectory has higher `partial_success`, trajectories are swapped
+- **Progress prediction**: ❌ **No** (progress is **masked out** during training)
+  - Implementation: `should_compute_progress()` returns `0.0` for trajectories with `quality_label in ["suboptimal", "failure", "failed"]`
+  - Progress loss is not computed for suboptimal trajectories (masked via `target_progress_rejected_mask`)
+- **Success prediction**: ❌ No (success labels would be `[0.0]` anyway for suboptimal trajectories via `compute_success_labels`)
+
+## Progress Masking
+
+Progress prediction is controlled by the `should_compute_progress()` function in `robometer/data/collators/rbm_heads.py`:
+
+- **Suboptimal/Failure trajectories**: Progress is masked out (`returns 0.0`)
+- **Successful trajectories**: Progress is computed (`returns 1.0`)
+- **REWIND strategy**: Progress is computed (`returns 1.0`)
+- **DIFFERENT_TASK strategy**: Progress is computed (even though it's 0.0)
+- **Trajectories with `partial_success`**: Progress is always computed (`returns 1.0`)
+
+The mask is applied in the trainer via `target_progress_chosen_mask` and `target_progress_rejected_mask` tensors.
+
+## Success Label Computation
+
+Success labels are computed from `target_progress` using `compute_success_labels()`:
+
+- If `quality_label in ["failure", "suboptimal", "failed"]`: All success labels are `[0.0]`
+- Otherwise: Success labels are `1.0` if `progress >= threshold`, else `0.0`
+- For `DIFFERENT_TASK` strategy: Since `target_progress = [0.0]`, all success labels are `[0.0]`
+
+## Strategy Filtering in PrefSampler
+
+The `PrefSampler` handles strategy selection and trajectory pairing:
+
+1. **Non-successful trajectories (non-partial_success)**: Automatically used as rejected with optimal from same task as chosen
+2. **Partial success trajectories**: Handled via `_get_different_partial_success_traj()` with swapping logic
+3. **Strategy execution**: Handled via `_execute_strategy()` which returns rejected trajectory and subsample strategy
+
+## Progress Sampling Details
+
+### Strategy Requirements
+
+#### 1. FORWARD_PROGRESS (Subsequence)
+- **Requirement**: Only uses **successful trajectories** (`quality_label == "successful"`)
+- **Enforced by**: `StrategyFirstDataset._filter_indices_by_strategy()` filters FORWARD_PROGRESS (for progress samples) to only use successful indices
+- **Progress prediction**: ✅ Yes
+- **Success prediction**: ✅ Yes
+
+#### 2. REWIND
+- **Requirement**: Only uses **successful trajectories** (`quality_label == "successful"`)
+- **Enforced by**: `StrategyFirstDataset._filter_indices_by_strategy()` filters REWIND (for all sample types) to only use successful indices
+- **Progress prediction**: ✅ Yes
+- **Success prediction**: ✅ Yes
+
+#### 3. DIFFERENT_TASK_INSTRUCTION
+- **Requirement**: Can use any trajectory
+- **Progress prediction**: ✅ Yes (but progress is set to `[0.0]` for all timesteps)
+- **Success prediction**: ✅ Yes (but success labels are set to `[0.0]` for all timesteps)
+  - Implementation: `target_progress = [0.0]` → `success_label = [0.0]` (explicitly set in `progress.py`)
+
+#### 4. REVERSE_PROGRESS
+- **Requirement**: Only uses **successful trajectories** (`quality_label == "successful"`)
+- **Enforced by**: `StrategyFirstDataset._filter_indices_by_strategy()` filters REVERSE_PROGRESS (for progress samples) to only use successful indices
+- **Progress prediction**: ✅ Yes (progress computed in reverse order)
+- **Success prediction**: ✅ Yes
+
+## Verification Checklist
+
+### Preference Sampling
+- ✅ REWIND strategy only uses successful trajectories (enforced in `StrategyFirstDataset`)
+- ✅ REVERSE_PROGRESS strategy only uses successful trajectories for preference samples (enforced in `StrategyFirstDataset`)
+- ✅ DIFFERENT_TASK sets rejected progress to `[0.0]` (line 351 in `pref.py`)
+- ✅ DIFFERENT_TASK success labels are correctly set to `[0.0]` after setting progress to 0.0
+- ✅ SUBOPTIMAL progress is masked out (`should_compute_progress` returns 0.0)
+- ✅ Chosen trajectory always predicts progress (unless preference-only dataset)
+- ✅ Chosen trajectory with `partial_success < 1.0` does NOT predict success
+
+### Progress Sampling
+- ✅ FORWARD_PROGRESS strategy only uses successful trajectories (enforced in `StrategyFirstDataset` for progress samples)
+- ✅ REWIND strategy only uses successful trajectories (enforced in `StrategyFirstDataset` for all sample types)
+- ✅ REVERSE_PROGRESS strategy only uses successful trajectories (enforced in `StrategyFirstDataset` for progress samples)
+- ✅ DIFFERENT_TASK_INSTRUCTION sets progress to `[0.0]` (line 157 in `progress.py`)
+- ✅ DIFFERENT_TASK_INSTRUCTION success labels are correctly set to `[0.0]` after setting progress to 0.0
diff --git a/robometer/robometer/data/datasets/__init__.py b/robometer/robometer/data/datasets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..f12caaab9b28bbb2c73072bc12371e0f3b65ac5a
--- /dev/null
+++ b/robometer/robometer/data/datasets/__init__.py
@@ -0,0 +1,13 @@
+from robometer.data.datasets.rbm_data import RBMDataset
+from robometer.data.datasets.strategy_first_dataset import StrategyFirstDataset
+from robometer.data.datasets.base import BaseDataset
+from robometer.data.datasets.custom_eval import CustomEvalDataset
+from robometer.data.datasets.repeated_dataset import RepeatedDataset
+
+__all__ = [
+    "RBMDataset",
+    "StrategyFirstDataset",
+    "BaseDataset",
+    "CustomEvalDataset",
+    "RepeatedDataset",
+]
diff --git a/robometer/robometer/data/datasets/base.py b/robometer/robometer/data/datasets/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..87180545d88991d92eec44031a883e855300d493
--- /dev/null
+++ b/robometer/robometer/data/datasets/base.py
@@ -0,0 +1,774 @@
+#!/usr/bin/env python3
+import collections
+import json
+import os
+from typing import Any, Dict, List, Tuple, Set, Optional, Union
+import torch
+
+from datasets import Dataset, concatenate_datasets
+from robometer.configs.experiment_configs import DataConfig
+from robometer.data.datasets.helpers import load_dataset_success_percent
+from robometer.data.dataset_category import DATASET_MAP, DATA_SOURCE_CATEGORY, get_paired_ds
+from robometer.utils.distributed import banner
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+def resolve_dataset_keys(
+    dataset_keys: Union[List[str], List[List[str]]], split: str
+) -> Union[List[str], List[List[str]]]:
+    resolved_datasets = []
+    for key in dataset_keys:
+        if key in DATASET_MAP:
+            if split in DATASET_MAP[key]:
+                resolved_datasets.extend(DATASET_MAP[key][split])
+            else:
+                logger.warning(f"Key '{key}' found in DATASET_MAP but no '{split}' split defined. Skipping.")
+        else:
+            # Not a key, assume it's already a dataset name
+            resolved_datasets.append(key)
+
+    # Deduplicate: handle both strings and lists
+    seen = set()
+    seen_lists = []
+    deduplicated_datasets = []
+    for item in resolved_datasets:
+        if isinstance(item, list):
+            # For lists, convert to tuple for hashing, but store the original list
+            item_tuple = tuple(item)
+            if item_tuple not in seen:
+                seen.add(item_tuple)
+                seen_lists.append(item)
+                deduplicated_datasets.append(item)
+        else:
+            # For strings, use normal set membership
+            if item not in seen:
+                seen.add(item)
+                deduplicated_datasets.append(item)
+    return deduplicated_datasets
+
+
+class BaseDataset(torch.utils.data.Dataset):
+    def __init__(
+        self, config: DataConfig, is_evaluation: bool = False, filter_quality_labels: Optional[List[str]] = None
+    ):
+        """Initialize BaseDataset.
+
+        Args:
+            config: Data configuration
+            is_evaluation: Whether this is an evaluation dataset
+            filter_quality_labels: List of quality labels to keep (e.g., ["successful"], ["successful", "suboptimal"]).
+                                   If None, no quality label filtering is applied.
+        """
+        self.config = config
+        self.is_evaluation = is_evaluation
+
+        if is_evaluation and config.eval_datasets:
+            self.datasets = config.eval_datasets
+        else:
+            self.datasets = config.train_datasets
+
+        # Load dataset-specific success cutoff map if available
+        self.dataset_success_cutoff_map = {}
+        if hasattr(config, "dataset_success_cutoff_file") and config.dataset_success_cutoff_file:
+            self.dataset_success_cutoff_map = load_dataset_success_percent(config.dataset_success_cutoff_file)
+
+        # Load trajectory dataset
+        self.dataset, self._combined_indices = self._load_all_datasets()
+
+        # Apply all filters simultaneously
+        # excluded_keywords = ["rings", "flick"]
+        excluded_keywords = []
+        min_frames = config.min_frames_per_trajectory
+
+        # Check if we're in progress_only mode (sample_type_ratio == [0, 1, 0])
+        # In progress_only mode, filter to only include successful trajectories
+        if config.sample_type_ratio == [0, 1, 0] and not is_evaluation:
+            filter_quality_labels = ["successful"]
+            logger.info(
+                "Progress-only mode detected (sample_type_ratio=[0, 1, 0]), filtering to only successful trajectories"
+            )
+
+        dataset_type = "evaluation" if is_evaluation else "training"
+        logger.info(f"Filtering {dataset_type} dataset with {len(self.dataset)} total trajectories")
+        if self.is_evaluation and filter_quality_labels is None:
+            logger.info(f"SKIPPING FILTERING for {dataset_type} dataset BECAUSE IT'S EVALUATION")
+        else:
+            self.dataset, self._combined_indices = self._filter_dataset(
+                excluded_keywords=excluded_keywords,
+                min_frames=min_frames,
+                dataset=self.dataset,
+                combined_indices=self._combined_indices,
+                filter_quality_labels=filter_quality_labels,
+            )
+        if filter_quality_labels:
+            logger.info(
+                f"{dataset_type.capitalize()} dataset filtered with {len(self.dataset)} total trajectories (filtered for quality_labels: {filter_quality_labels})"
+            )
+        else:
+            logger.info(
+                f"{dataset_type.capitalize()} dataset filtered with {len(self.dataset)} total trajectories (excluded keywords and min_frames only, no quality label filtering)"
+            )
+
+        # Filter out trajectories based on multiple criteria (build indices first, then filter once)
+        self.dataset, self._combined_indices = self._filter_task_based_criteria(
+            dataset=self.dataset,
+            combined_indices=self._combined_indices,
+        )
+
+        # Set cached fields after filtering
+        self._cached_ids = self.dataset["id"]
+        self._cached_is_robot = self.dataset["is_robot"]
+
+        logger.info(f"Dataset loaded with {len(self.dataset)} total trajectories")
+        # Log trajectory counts per quality label
+        if "quality_label" in self.dataset.column_names:
+            quality_counts = collections.Counter(
+                str(q) if q is not None else "null" for q in self.dataset["quality_label"]
+            )
+            for label in sorted(quality_counts.keys()):
+                logger.info(f"  quality_label {label}: {quality_counts[label]} trajectories")
+
+        # Initialize resampling stats containers shared by subclasses
+        self._resample_attempt_stats: Dict[str, collections.defaultdict[str, List[int]]] = {
+            "preference": collections.defaultdict(list),
+            "progress": collections.defaultdict(list),
+        }
+        self._resample_dataset_attempt_stats: Dict[str, collections.defaultdict[str, List[int]]] = {
+            "preference": collections.defaultdict(list),
+            "progress": collections.defaultdict(list),
+        }
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def _load_all_datasets(self) -> Tuple[Dataset, Dict[str, Any]]:
+        """Load trajectory dataset using preprocessed index-based cache.
+
+        Returns:
+            tuple: (dataset, combined_indices)
+                - dataset: The loaded and concatenated dataset
+                - combined_indices: Dictionary of combined indices
+        """
+        cache_dir = os.environ.get("ROBOMETER_PROCESSED_DATASETS_PATH", "")
+        if not cache_dir:
+            raise ValueError(
+                "ROBOMETER_PROCESSED_DATASETS_PATH not set. Set it to the directory containing your processed datasets."
+            )
+        cache_type = "evaluation" if self.is_evaluation else "training"
+
+        # Check if preprocessed cache exists
+        if os.path.exists(cache_dir):
+            logger.debug(f"Found preprocessed cache at {cache_dir}, loading {cache_type} datasets...")
+
+            dataset, combined_indices = self._load_preprocessed_cache(cache_dir, is_training=not self.is_evaluation)
+
+            logger.debug(
+                f"Successfully loaded preprocessed {cache_type} datasets with {len(dataset)} trajectory indices"
+            )
+
+            return dataset, combined_indices
+        else:
+            # If no cache exists, we need to run the preprocessor first
+            logger.warning("No preprocessed cache found. Please run preprocess_datasets.py first to create the cache.")
+            raise RuntimeError(
+                "Dataset preprocessing required. Please run:\n"
+                "uv run scripts/preprocess_datasets.py\n"
+                "This will create the necessary index-based cache for efficient data loading."
+            )
+
+    def _get_available_datasets(self, cache_dir: str) -> Tuple[List[Tuple[str, str]], List[str]]:
+        """Check which datasets are available in the cache.
+
+        Returns:
+            tuple: (available_datasets, missing_datasets)
+                - available_datasets: List of (dataset_path, individual_cache_dir) tuples
+                - missing_datasets: List of dataset_path strings that are missing
+        """
+        available_datasets = []
+        missing_datasets = []
+
+        for dataset_path in self.datasets:
+            # The preprocessing script creates individual cache directories for each dataset
+            individual_cache_dir = os.path.join(cache_dir, dataset_path.replace("/", "_").replace(":", "_"))
+
+            if os.path.exists(individual_cache_dir):
+                info_file = os.path.join(individual_cache_dir, "dataset_info.json")
+                if os.path.exists(info_file):
+                    try:
+                        with open(info_file) as f:
+                            json.load(f)
+
+                        available_datasets.append((dataset_path, individual_cache_dir))
+                        logger.debug(f"Found cache: {individual_cache_dir}")
+                    except:
+                        logger.warning(f"Cache info file corrupted, skipping: {individual_cache_dir}")
+                        continue
+                else:
+                    logger.debug(f"No info file found, skipping: {individual_cache_dir}")
+                    continue
+            else:
+                missing_datasets.append(dataset_path)
+                logger.debug(f"Missing cache: {individual_cache_dir}")
+
+        # Warn about missing datasets
+        if missing_datasets:
+            logger.warning("⚠️  Warning: The following configured datasets are not available in the cache:")
+            for dataset_path in missing_datasets:
+                logger.warning(f"    ❌ {dataset_path}")
+            logger.warning("  Available datasets will be loaded, but some configured data may be missing.")
+
+        if not available_datasets:
+            raise RuntimeError(
+                f"No configured datasets are available in the cache. "
+                f"Please run preprocess_datasets.py to create the cache for: {self.datasets}"
+            )
+
+        logger.debug(f"Summary: {len(available_datasets)} available, {len(missing_datasets)} missing")
+
+        return available_datasets, missing_datasets
+
+    def _load_datasets(self, available_datasets: List[Tuple[str, str]]) -> Tuple[List[Dataset], List[Dict[str, Any]]]:
+        """Load datasets from cache and return them along with per-dataset index mappings.
+
+        Args:
+            available_datasets: List of (dataset_path, individual_cache_dir) tuples
+
+        Returns:
+            tuple: (loaded_datasets, dataset_indices_list)
+                - loaded_datasets: List of loaded Dataset objects
+                - dataset_indices_list: List of index dictionaries, one per dataset (with original indices, not offset)
+        """
+        loaded_datasets = []
+        dataset_indices_list = []
+
+        for dataset_path, individual_cache_dir in available_datasets:
+            # Load the processed dataset
+            dataset_cache_dir = os.path.join(individual_cache_dir, "processed_dataset")
+            if not os.path.exists(dataset_cache_dir):
+                logger.warning(f"Processed dataset not found at {dataset_cache_dir}, skipping...")
+                continue
+
+            dataset = Dataset.load_from_disk(dataset_cache_dir, keep_in_memory=True)
+            loaded_datasets.append(dataset)
+
+            # Load index mappings
+            indices = {}
+            mappings_file = os.path.join(individual_cache_dir, "index_mappings.json")
+            if os.path.exists(mappings_file):
+                with open(mappings_file) as f:
+                    indices = json.load(f)
+
+            dataset_indices_list.append(indices)
+
+            logger.debug(f"Loaded {len(dataset)} trajectories from {dataset_path}")
+
+        if not loaded_datasets:
+            raise RuntimeError("No datasets could be loaded from the cache")
+
+        return loaded_datasets, dataset_indices_list
+
+    def _build_indices(
+        self, loaded_datasets: List[Dataset], dataset_indices_list: List[Dict[str, Any]], cached_is_robot: List[bool]
+    ) -> Dict[str, Any]:
+        """Build combined indices from loaded datasets and their index mappings.
+
+        Args:
+            loaded_datasets: List of loaded Dataset objects
+            dataset_indices_list: List of index dictionaries, one per dataset (with original indices, not offset)
+            cached_is_robot: List of is_robot flags for the concatenated dataset
+
+        Returns:
+            dict: Combined indices dictionary with all index mappings
+        """
+        combined_indices = {
+            "robot_trajectories": [],
+            "human_trajectories": [],
+            "optimal_by_task": {},
+            "suboptimal_by_task": {},
+            "quality_indices": {},
+            "task_indices": {},
+            "source_indices": {},
+            "partial_success_indices": {},
+        }
+
+        offset = 0
+
+        for dataset, indices in zip(loaded_datasets, dataset_indices_list):
+            # Adjust indices by adding offset and combine
+            for key in combined_indices:
+                if key in indices:
+                    if isinstance(indices[key], list):
+                        # For list indices, add offset
+                        combined_indices[key].extend([idx + offset for idx in indices[key]])
+                    elif isinstance(indices[key], dict):
+                        # For regular dict indices, add offset to values
+                        for subkey, subindices in indices[key].items():
+                            if subkey not in combined_indices[key]:
+                                combined_indices[key][subkey] = []
+                            combined_indices[key][subkey].extend([idx + offset for idx in subindices])
+
+            offset += len(dataset)
+
+        # Build paired human robot index (needs cached_is_robot)
+        combined_indices["paired_human_robot_by_task"] = self._build_paired_human_robot_index(
+            combined_indices, cached_is_robot
+        )
+
+        # Find tasks that have both optimal and suboptimal trajectories
+        tasks_with_multiple_quality_labels = set(combined_indices["optimal_by_task"].keys()) & set(
+            combined_indices["suboptimal_by_task"].keys()
+        )
+        # Add it to combined_indices so samplers can access it
+        combined_indices["tasks_with_multiple_quality_labels"] = list(tasks_with_multiple_quality_labels)
+
+        return combined_indices
+
+    def _load_preprocessed_cache(self, cache_dir: str, is_training: bool = True) -> Tuple[Dataset, Dict[str, Any]]:
+        """Load the preprocessed cache with index mappings for datasets.
+
+        Returns:
+            tuple: (dataset, combined_indices)
+                - dataset: The loaded and concatenated dataset
+                - combined_indices: Dictionary of combined indices
+        """
+        # Get available datasets
+        available_datasets, missing_datasets = self._get_available_datasets(cache_dir)
+
+        # Load datasets
+        loaded_datasets, dataset_indices_list = self._load_datasets(available_datasets)
+
+        # Concatenate datasets if multiple
+        if len(loaded_datasets) == 1:
+            dataset = loaded_datasets[0]
+        else:
+            dataset = concatenate_datasets(loaded_datasets)
+
+        # Build combined indices
+        combined_indices = self._build_indices(loaded_datasets, dataset_indices_list, dataset["is_robot"])
+
+        dataset_type = "training" if is_training else "evaluation"
+        logger.info(f"✅ Loaded {len(dataset)} total trajectories from preprocessed {dataset_type} datasets")
+        logger.debug(f"Available datasets: {len(available_datasets)}/{len(missing_datasets) + len(available_datasets)}")
+        logger.debug(f"Missing datasets: {len(missing_datasets)}")
+        banner(f"{dataset_type.capitalize()} dataset statistics", inner_padding=1)
+        logger.debug(f"[{dataset_type.upper()}] Robot trajectories: {len(combined_indices['robot_trajectories'])}")
+        logger.debug(f"[{dataset_type.upper()}] Human trajectories: {len(combined_indices['human_trajectories'])}")
+        logger.debug(f"[{dataset_type.upper()}] Number of different tasks: {len(combined_indices['task_indices'])}")
+        logger.debug(f"[{dataset_type.upper()}] Data sources: {len(combined_indices['source_indices'])}")
+        logger.debug(
+            f"[{dataset_type.upper()}] Tasks available: {list(combined_indices['task_indices'].keys())[:10]} ..."
+        )
+        logger.debug(f"[{dataset_type.upper()}] Number of quality labels: {len(combined_indices['quality_indices'])}")
+        for quality_label in combined_indices["quality_indices"]:
+            logger.debug(
+                f"[{dataset_type.upper()}]   {quality_label}: {len(combined_indices['quality_indices'][quality_label])}"
+            )
+        logger.debug(f"[{dataset_type.upper()}] Data sources available: {combined_indices['source_indices'].keys()}")
+        logger.debug(
+            f"[{dataset_type.upper()}] Number of paired tasks: {len(combined_indices['paired_human_robot_by_task'])}"
+        )
+        logger.debug(
+            f"[{dataset_type.upper()}] Number of tasks with both multiple quality labels: {len(combined_indices['tasks_with_multiple_quality_labels'])}"
+        )
+
+        return dataset, combined_indices
+
+    def _filter_dataset(
+        self,
+        excluded_keywords: List[str],
+        min_frames: int,
+        dataset: Dataset,
+        combined_indices: Dict[str, Any],
+        filter_quality_labels: Optional[List[str]] = None,
+    ) -> Tuple[Dataset, Dict[str, Any]]:
+        """Filter dataset based on multiple criteria simultaneously.
+
+        Filters out trajectories that:
+        - Have tasks containing excluded keywords
+        - Have <= min_frames frames
+        - (If filter_quality_labels is set) Have quality_label not in the specified list
+        - RoboArena trajectories from tasks with only one partial_success category
+
+        Uses batched map operations for efficient parallel processing.
+
+        Args:
+            excluded_keywords: List of keywords to exclude from task names (case-insensitive)
+            min_frames: Minimum number of frames required (trajectories with frames > min_frames are kept)
+            dataset: The dataset to filter
+            combined_indices: Dictionary of combined indices to update after filtering
+            filter_quality_labels: List of quality labels to keep. If None, no quality filtering is applied.
+
+        Returns:
+            tuple: (filtered_dataset, filtered_combined_indices)
+                - filtered_dataset: The filtered dataset
+                - filtered_combined_indices: The filtered combined indices
+        """
+        excluded_keywords_lower = [kw.lower() for kw in excluded_keywords]
+
+        # Pre-compute tasks with only one partial_success category (for datasets with partial_success)
+        all_tasks = dataset["task"]
+        data_sources = dataset["data_source"]
+        # Handle case where partial_success column might not exist
+        if "partial_success" in dataset.column_names:
+            partial_successes = dataset["partial_success"]
+        else:
+            partial_successes = [None] * len(dataset)
+
+        # Group trajectories with partial_success by task
+        tasks_to_partial_success = collections.defaultdict(set)
+
+        for task, partial_success in zip(all_tasks, partial_successes):
+            if task is None:
+                continue
+            # Check if this trajectory has partial_success
+            if partial_success is not None:
+                tasks_to_partial_success[task].add(partial_success)
+
+        # # Find tasks with only one unique partial_success category
+        # tasks_with_single_partial_success = {
+        #     task
+        #     for task, partial_success_set in tasks_to_partial_success.items()
+        #     if len(partial_success_set) == 1
+        # }
+
+        def add_filter_flags(batch):
+            """Add filter flags to batch for efficient filtering."""
+            tasks = batch["task"]
+            frames_shapes = batch["frames_shape"]
+            quality_labels = batch["quality_label"]
+            data_sources_batch = batch.get("data_source", [None] * len(tasks))
+            # Handle case where partial_success column might not exist
+            batch_partial_success = batch.get("partial_success", [None] * len(tasks))
+
+            drop_kw = []
+            drop_frames = []
+            drop_quality = []
+            drop_roboarena = []
+
+            for idx, (task, fs, quality_label, data_source, partial_success) in enumerate(
+                zip(tasks, frames_shapes, quality_labels, data_sources_batch, batch_partial_success)
+            ):
+                dkw = False
+                dfr = False
+                dq = False
+                dr = False
+
+                # Check excluded keywords
+                if task is not None:
+                    t = task.lower()
+                    if any(kw in t for kw in excluded_keywords_lower):
+                        dkw = True
+
+                # Check minimum frames (only if not dropped by keywords)
+                if not dkw and fs is not None and not (isinstance(fs, (list, tuple)) and len(fs) == 0):
+                    if isinstance(fs, (list, tuple)):
+                        num_frames = fs[0]
+                    else:
+                        num_frames = fs
+                    if num_frames <= min_frames:
+                        dfr = True
+
+                # Check quality_label filter (only if not dropped by other filters)
+                if filter_quality_labels and not dkw and not dfr:
+                    if quality_label not in filter_quality_labels:
+                        dq = True
+
+                # # Check tasks with single partial_success (only if not dropped by other filters)
+                # if not dkw and not dfr and not dq:
+                #     if task is not None and task in tasks_with_single_partial_success:
+                #         # Check if this trajectory has partial_success
+                #         if partial_success is not None:
+                #             dr = True
+
+                drop_kw.append(dkw)
+                drop_frames.append(dfr)
+                drop_quality.append(dq)
+                # drop_roboarena.append(dr)
+
+            return {
+                "drop_by_keywords": drop_kw,
+                "drop_by_frames": drop_frames,
+                "drop_by_quality": drop_quality,
+                # "drop_by_roboarena": drop_roboarena,
+            }
+
+        # 1) Compute flags in a single batched pass
+        dataset_with_flags = dataset.map(
+            add_filter_flags,
+            batched=True,
+            num_proc=8,  # Can be increased for parallel processing if needed
+            desc="Computing filter flags",
+        )
+
+        # 2) Compute counts and build keep_indices from flags
+        drop_kw_list = dataset_with_flags["drop_by_keywords"]
+        drop_frames_list = dataset_with_flags["drop_by_frames"]
+        drop_quality_list = dataset_with_flags["drop_by_quality"]
+        # drop_roboarena_list = dataset_with_flags["drop_by_roboarena"]
+
+        filtered_by_keywords = int(sum(drop_kw_list))
+        filtered_by_frames = int(sum(drop_frames_list))
+        filtered_by_quality = int(sum(drop_quality_list))
+        # filtered_by_roboarena = int(sum(drop_roboarena_list))
+        total_filtered = filtered_by_keywords + filtered_by_frames + filtered_by_quality  # + filtered_by_roboarena
+
+        # 3) Filter using precomputed flags (efficient)
+        if total_filtered > 0:
+            filter_messages = []
+            if filtered_by_keywords > 0:
+                filter_messages.append(f"{filtered_by_keywords} with excluded task keywords: {excluded_keywords}")
+            if filtered_by_frames > 0:
+                filter_messages.append(f"{filtered_by_frames} with <= {min_frames} frames")
+            if filtered_by_quality > 0:
+                filter_messages.append(f"{filtered_by_quality} with quality_label != 'successful'")
+            # if filtered_by_roboarena > 0:
+            #     filter_messages.append(
+            #         f"{filtered_by_roboarena} trajectories with partial_success from {len(tasks_with_single_partial_success)} tasks with only one partial_success category"
+            #     )
+
+            logger.info(f"Filtering out {total_filtered} trajectories ({', '.join(filter_messages)})")
+
+            # Build keep_indices from flags (before filtering)
+            keep_indices = [
+                i
+                for i, (dkw, dfr, dq) in enumerate(
+                    zip(drop_kw_list, drop_frames_list, drop_quality_list)  # , drop_roboarena_list)
+                )
+                if not (dkw or dfr or dq)  # or dr)
+            ]
+
+            removed_indices = set(range(len(dataset))) - set(keep_indices)
+
+            logger.trace(f"Removed indices: {removed_indices}")
+
+            # Filter dataset using select with keep_indices (more efficient than filter)
+            filtered_dataset = dataset.select(keep_indices)
+
+            # Update combined_indices using the shared helper
+            filtered_combined_indices = self._update_indices_after_filtering(combined_indices, keep_indices)
+        else:
+            # No filtering needed, return original dataset and indices
+            filtered_dataset = dataset
+            filtered_combined_indices = combined_indices
+
+        return filtered_dataset, filtered_combined_indices
+
+    def _filter_task_based_criteria(
+        self,
+        dataset: Dataset,
+        combined_indices: Dict[str, Any],
+    ) -> Tuple[Dataset, Dict[str, Any]]:
+        """Filter out suboptimal/failed trajectories that don't have optimal counterparts with the same task name.
+        Also filter out tasks that only have failed/suboptimal trajectories.
+
+        This filtering is skipped for datasets with partial_success.
+
+        Args:
+            dataset: The dataset to filter
+            combined_indices: Dictionary of combined indices to update after filtering
+
+        Returns:
+            tuple: (filtered_dataset, filtered_combined_indices)
+                - filtered_dataset: The filtered dataset
+                - filtered_combined_indices: The filtered combined indices
+        """
+        # Get tasks that have optimal trajectories
+        tasks_with_optimal = set(combined_indices.get("optimal_by_task", {}).keys())
+
+        # Get all tasks in the dataset
+        all_tasks = dataset["task"]
+        quality_labels = dataset["quality_label"]
+        # Handle case where partial_success column might not exist
+        if "partial_success" in dataset.column_names:
+            partial_successes = dataset["partial_success"]
+        else:
+            partial_successes = [None] * len(dataset)
+
+        # Identify trajectories to remove:
+        # All trajectories from tasks that have no optimal trajectories
+        indices_to_remove = set()
+        tasks_removed = set()
+
+        for idx, (task, quality_label, partial_success) in enumerate(zip(all_tasks, quality_labels, partial_successes)):
+            if task is None:
+                # Skip trajectories with None task
+                continue
+
+            # Skip filtering for trajectories with partial_success
+            if partial_success is not None:
+                continue
+
+            if task not in tasks_with_optimal:
+                # This task has no optimal trajectories
+                # Remove all trajectories from this task (they're all suboptimal/failed with no optimal counterparts)
+                indices_to_remove.add(idx)
+                tasks_removed.add(task)
+
+        if indices_to_remove:
+            # Build keep_indices
+            keep_indices = [i for i in range(len(dataset)) if i not in indices_to_remove]
+
+            filtered_by_no_optimal = len(indices_to_remove)
+            tasks_removed_count = len(tasks_removed)
+
+            logger.info(
+                f"Filtering out {filtered_by_no_optimal} trajectories from {tasks_removed_count} tasks "
+                f"that have no optimal trajectories (only suboptimal/failed)"
+            )
+
+            # Filter dataset using select with keep_indices
+            filtered_dataset = dataset.select(keep_indices)
+
+            # Update combined_indices using the shared helper
+            filtered_combined_indices = self._update_indices_after_filtering(combined_indices, keep_indices)
+        else:
+            # No filtering needed
+            filtered_dataset = dataset
+            filtered_combined_indices = combined_indices
+
+        return filtered_dataset, filtered_combined_indices
+
+    def _update_indices_after_filtering(
+        self, combined_indices: Dict[str, Any], keep_indices: List[int]
+    ) -> Dict[str, Any]:
+        """Update combined_indices after filtering the dataset.
+
+        Args:
+            combined_indices: Dictionary of combined indices before filtering
+            keep_indices: List of indices to keep (from the original dataset)
+
+        Returns:
+            dict: Updated combined_indices with remapped indices
+        """
+        # Create a mapping from old index to new index
+        old_to_new = {old_idx: new_idx for new_idx, old_idx in enumerate(keep_indices)}
+
+        # Create a copy of combined_indices to avoid mutating the input
+        filtered_combined_indices = {}
+
+        # Filter all index lists and dicts
+        for key in combined_indices:
+            if isinstance(combined_indices[key], list):
+                # Filter list indices
+                filtered_combined_indices[key] = [old_to_new[idx] for idx in combined_indices[key] if idx in old_to_new]
+            elif isinstance(combined_indices[key], dict):
+                # Filter dict indices
+                filtered_dict = {}
+                for subkey, subindices in combined_indices[key].items():
+                    # Handle nested dict structure (e.g., paired_human_robot_by_task has {"robot": [...], "human": [...]})
+                    if isinstance(subindices, dict):
+                        # This is a nested dict like {"robot": [...], "human": [...]}
+                        filtered_nested_dict = {}
+                        for nested_key, nested_list in subindices.items():
+                            if isinstance(nested_list, list):
+                                filtered_nested_list = [old_to_new[idx] for idx in nested_list if idx in old_to_new]
+                                if filtered_nested_list:  # Only keep if there are remaining indices
+                                    filtered_nested_dict[nested_key] = filtered_nested_list
+                        # Only keep the task if it has at least one non-empty list (robot or human)
+                        if filtered_nested_dict:
+                            filtered_dict[subkey] = filtered_nested_dict
+                    elif isinstance(subindices, list):
+                        # Regular dict with list values (e.g., task_indices, source_indices)
+                        filtered_indices = [old_to_new[idx] for idx in subindices if idx in old_to_new]
+                        if filtered_indices:  # Only keep keys with remaining indices
+                            filtered_dict[subkey] = filtered_indices
+                filtered_combined_indices[key] = filtered_dict
+            elif isinstance(combined_indices[key], set):
+                # Filter set indices
+                filtered_combined_indices[key] = {old_to_new[idx] for idx in combined_indices[key] if idx in old_to_new}
+            else:
+                # Keep other types as-is (e.g., strings, numbers)
+                filtered_combined_indices[key] = combined_indices[key]
+
+        return filtered_combined_indices
+
+    # ------------------------------------------------------------------
+    # Shared resample helpers for subclasses
+    # ------------------------------------------------------------------
+    def _record_resample_attempt(
+        self, sample_type: str, strategy: str, sample_attempts: int, dataset_attempts: int
+    ) -> None:
+        if sample_type not in self._resample_attempt_stats:
+            return
+
+        self._resample_attempt_stats[sample_type][strategy].append(sample_attempts)
+        self._resample_dataset_attempt_stats[sample_type][strategy].append(dataset_attempts)
+
+    def _set_resample_attempts(self, sample, dataset_attempts: int):
+        if sample is None:
+            return None
+
+        dataset_attempts = max(1, int(dataset_attempts))
+        sample_attempts = int(getattr(sample, "resample_attempts", dataset_attempts))
+        sample_attempts = max(1, sample_attempts)
+        sample.resample_attempts = sample_attempts
+
+        sample_type = getattr(sample, "sample_type", "unknown")
+        strategy = str(getattr(sample, "data_gen_strategy", "unknown"))
+        self._record_resample_attempt(sample_type, strategy, sample_attempts, dataset_attempts)
+
+        return sample
+
+    def get_resample_attempt_stats(self):
+        return self._resample_attempt_stats
+
+    def get_resample_dataset_attempt_stats(self):
+        return self._resample_dataset_attempt_stats
+
+    def _build_paired_human_robot_index(
+        self, combined_indices: Dict[str, Any], cached_is_robot: List[bool]
+    ) -> Dict[str, Dict[str, List[int]]]:
+        """Build paired_human_robot_by_task index from task_indices by checking is_robot field.
+
+        This builds the index after concatenation by iterating through task_indices
+        and checking the is_robot field for each trajectory. Only includes trajectories
+        from PAIRED data sources.
+
+        Args:
+            combined_indices: Dictionary of combined indices
+            cached_is_robot: List of is_robot flags for the concatenated dataset
+
+        Returns:
+            dict: paired_human_robot_by_task dictionary
+        """
+        paired_human_robot_by_task = {}
+
+        # Filter indices for paired data sources
+        paired_data_source_indices = set()
+        for data_source in get_paired_ds():
+            if data_source in combined_indices["source_indices"]:
+                paired_data_source_indices.update(combined_indices["source_indices"][data_source])
+
+        if not paired_data_source_indices:
+            logger.debug("No paired data sources found, skipping paired index building")
+            return {}
+
+        # Build index from task_indices using cached is_robot field, but only for paired data sources
+        for task, indices in combined_indices["task_indices"].items():
+            # Filter to only paired data sources
+            task_indices_paired = [idx for idx in indices if idx in paired_data_source_indices]
+
+            if not task_indices_paired:
+                continue
+
+            paired_human_robot_by_task[task] = {"robot": [], "human": []}
+
+            for idx in task_indices_paired:
+                is_robot = cached_is_robot[idx] if idx < len(cached_is_robot) else True
+                if is_robot:
+                    paired_human_robot_by_task[task]["robot"].append(idx)
+                else:
+                    paired_human_robot_by_task[task]["human"].append(idx)
+
+        # Count tasks with both robot and human trajectories
+        tasks_with_pairs = [
+            task for task, task_dict in paired_human_robot_by_task.items() if task_dict["robot"] and task_dict["human"]
+        ]
+        num_tasks_with_pairs = len(tasks_with_pairs)
+        logger.debug(
+            f"Built paired_human_robot_by_task index: {num_tasks_with_pairs} tasks with both robot and human trajectories (from paired data sources only)"
+        )
+
+        return paired_human_robot_by_task
diff --git a/robometer/robometer/data/datasets/custom_eval.py b/robometer/robometer/data/datasets/custom_eval.py
new file mode 100644
index 0000000000000000000000000000000000000000..444edf2963e9352e90b5fba818cf5cee485129e2
--- /dev/null
+++ b/robometer/robometer/data/datasets/custom_eval.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+import torch
+from typing import List, Optional
+
+from robometer.data.datasets.base import BaseDataset
+from robometer.data.samplers import *
+from robometer.configs.experiment_configs import DataConfig
+
+
+class CustomEvalDataset(BaseDataset):
+    """Dataset that wraps a sampler for custom evaluation purposes."""
+
+    def __init__(
+        self,
+        sampler_type: str,
+        config: DataConfig,
+        verbose: bool = True,
+        sampler_kwargs: dict = None,
+    ):
+        """Initialize custom eval dataset with a sampler type.
+
+        Args:
+            sampler_type: Type of sampler to create (e.g., "confusion_matrix", "reward_alignment", "policy_ranking", "success_failure")
+            config: Configuration object
+            verbose: Verbose flag
+            sampler_kwargs: Additional keyword arguments for the sampler
+        """
+        filter_quality_labels: Optional[List[str]] = None
+
+        if sampler_type == "reward_alignment":
+            filter_quality_labels = ["successful"]
+        elif sampler_type == "confusion_matrix":
+            filter_quality_labels = ["successful", "suboptimal"]
+
+        # Special case: roboreward datasets should not filter by quality
+        if len(config.eval_datasets) == 1 and "roboreward" in config.eval_datasets[0]:
+            filter_quality_labels = None
+
+        super().__init__(config=config, is_evaluation=True, filter_quality_labels=filter_quality_labels)
+
+        sampler_cls = {
+            "confusion_matrix": ConfusionMatrixSampler,
+            "reward_alignment": RewardAlignmentSampler,
+            "policy_ranking": ProgressPolicyRankingSampler,
+            "quality_preference": QualityPreferenceSampler,
+        }
+
+        if "roboarena" in self.config.eval_datasets:
+            sampler_cls["quality_preference"] = RoboArenaQualityPreferenceSampler
+
+        if sampler_type not in sampler_cls:
+            raise ValueError(f"Unknown sampler type: {sampler_type}. Available: {list(sampler_cls.keys())}")
+
+        self.sampler = sampler_cls[sampler_type](
+            config=config,
+            dataset=self.dataset,
+            combined_indices=self._combined_indices,
+            dataset_success_cutoff_map=self.dataset_success_cutoff_map,
+            verbose=verbose,
+            **sampler_kwargs,
+        )
+
+    def __len__(self):
+        return len(self.sampler)
+
+    def __getitem__(self, idx):
+        return self.sampler[idx]
diff --git a/robometer/robometer/data/datasets/helpers.py b/robometer/robometer/data/datasets/helpers.py
new file mode 100644
index 0000000000000000000000000000000000000000..085794239406d3b8c23c588dd560f1182a5a4a07
--- /dev/null
+++ b/robometer/robometer/data/datasets/helpers.py
@@ -0,0 +1,703 @@
+from math import e, floor, ceil
+from typing import List, Dict, Tuple, Optional, Union, Any
+import os
+import random
+from enum import Enum
+
+import numpy as np
+import json
+import torch
+from robometer.utils.distributed import rank_0_print
+from robometer.data.dataset_types import Trajectory
+
+
+class DataGenStrat(Enum):
+    """Enum for different data generation strategies used in preference generation."""
+
+    REVERSE_PROGRESS = "reverse_progress"
+    UNIFORM_SAMPLE = "uniform_sample"
+    FORWARD_PROGRESS = "forward_progress"
+    REWIND = "rewind"
+    SUBOPTIMAL = "suboptimal"
+    DIFFERENT_TASK = "different_task"
+    DIFFERENT_TASK_INSTRUCTION = "different_task_instruction"
+    PAIRED_HUMAN_ROBOT = "paired_human_robot"
+
+
+def load_dataset_success_percent(cutoff_file_path):
+    """Load dataset-specific success percentage from file."""
+    success_percent_dict = {}
+    try:
+        with open(cutoff_file_path, "r") as f:
+            for line in f:
+                line = line.strip()
+                if line and "," in line:
+                    dataset_name, success_percent = line.split(",")
+                    success_percent_dict[dataset_name.strip()] = float(success_percent.strip())
+    except FileNotFoundError:
+        print(f"Warning: Success cutoff file {cutoff_file_path} not found. Using default thresholds.")
+    return success_percent_dict
+
+
+def convert_continuous_to_discrete_bin(value: float, num_bins: int) -> int:
+    """Convert a single continuous progress value in [0, 1] to a discrete bin [0, num_bins-1].
+
+    Args:
+        value: Single continuous progress value in [0, 1]
+        num_bins: Number of discrete bins to use
+
+    Returns:
+        Discrete bin index in [0, num_bins-1]
+    """
+    return round(min(max(value, 0.0), 1.0) * (num_bins - 1))
+
+
+def convert_continuous_to_discrete_bin_c51(x, num_bins):
+    """
+    x: (...,) tensor with values in [0, 1]
+    returns: (..., num_bins) probability distribution
+    """
+    b = x * (num_bins - 1)  # fractional bin index
+    l = floor(b)  # left bin
+    u = ceil(b)  # right bin
+
+    probs = torch.zeros((num_bins,), dtype=torch.float32)
+
+    probs[l] += u - b
+    probs[u] += b - l
+
+    # special case: l == u then assign all probability to the left bin
+    if l == u:
+        probs[l] = 1.0
+
+    return probs
+
+
+def convert_continuous_to_discrete_bins(
+    progress_values: Union[List[float], np.ndarray], num_bins: int
+) -> List[torch.Tensor]:
+    """Convert continuous progress values in [0, 1] to C51-style discrete probability distributions.
+
+    Args:
+        progress_values: List or array of continuous progress values in [0, 1]
+        num_bins: Number of discrete bins to use
+
+    Returns:
+        List of probability distributions as torch tensors, each of shape (num_bins,)
+    """
+    if isinstance(progress_values, np.ndarray):
+        progress_values = progress_values.tolist()
+    return [convert_continuous_to_discrete_bin_c51(p, num_bins) for p in progress_values]
+
+
+def compute_success_labels(
+    target_progress: List[float],
+    data_source: Optional[str],
+    dataset_success_percent: Optional[Dict[str, float]] = None,
+    max_success: float = 1.0,
+    quality_label: Optional[str] = None,
+) -> List[float]:
+    """
+    Compute success labels from target_progress.
+
+    Args:
+        target_progress: List of progress values (floats between 0 and 1)
+        data_source: Data source name (used to look up dataset-specific threshold)
+        dataset_success_percent: Dictionary mapping data source names to max_success thresholds
+        max_success: Default max_success threshold if data_source not in dataset_success_percent
+        quality_label: Quality label of the trajectory ("failure", "suboptimal", "successful", etc.)
+
+    Returns:
+        List of success labels (1.0 for success, 0.0 for failure) for each frame
+    """
+    if target_progress is None or len(target_progress) == 0:
+        return []
+
+    # If trajectory is failure or suboptimal, return all 0s
+    if quality_label is not None and quality_label.lower() in ("failure", "suboptimal", "failed"):
+        return [0.0] * len(target_progress)
+
+    # Get the threshold for this data source
+    if data_source is not None and dataset_success_percent is not None:
+        threshold = dataset_success_percent.get(
+            data_source if isinstance(data_source, str) else str(data_source), max_success
+        )
+    else:
+        threshold = max_success
+
+    # Generate success labels: 1.0 for success (progress >= threshold), 0.0 for failure
+    success_labels = [1.0 if prog >= threshold else 0.0 for prog in target_progress]
+    return success_labels
+
+
+def load_frames_from_npz(npz_filepath: str) -> np.ndarray:
+    """Load frames on-demand from npz file.
+
+    Args:
+        npz_filepath: Path to the .npz file containing frames
+
+    Returns:
+        numpy array with shape (T, H, W, C) containing the video frames
+    """
+    if not npz_filepath:
+        raise ValueError("npz_filepath is None or empty")
+
+    # If path is relative, prepend ROBOMETER_PROCESSED_DATASETS_PATH
+    if not os.path.isabs(npz_filepath):
+        processed_root = os.environ.get("ROBOMETER_PROCESSED_DATASETS_PATH", "")
+        # Normalize: strip processed_datasets when env points to repo root
+        processed_root = processed_root.replace("processed_datasets", "")
+        if processed_root:
+            npz_filepath = os.path.join(processed_root, npz_filepath)
+
+    if not os.path.exists(npz_filepath):
+        raise ValueError(f"NPZ file not found: {npz_filepath}")
+
+    try:
+        # Load frames from npz file
+        with np.load(npz_filepath) as data:
+            frames = data["frames"]
+            # Verify the data structure
+            if "shape" in data:
+                expected_shape = tuple(data["shape"])
+                if frames.shape != expected_shape:
+                    rank_0_print(f"Warning: Loaded frames shape {frames.shape} doesn't match expected {expected_shape}")
+
+            return frames
+    except Exception as e:
+        rank_0_print(f"Error loading frames from {npz_filepath}: {e}")
+        raise RuntimeError(f"Failed to load frames from {npz_filepath}: {e}")
+
+
+def load_embeddings_from_path(embeddings_path: str) -> torch.Tensor:
+    """Load video embeddings from .pt file and return just the video embeddings."""
+    if not embeddings_path:
+        import ipdb
+
+        ipdb.set_trace()
+        raise ValueError(f"embeddings_path: {embeddings_path} is None or empty")
+
+    # If path is relative, prepend ROBOMETER_PROCESSED_DATASETS_PATH
+    if not os.path.isabs(embeddings_path):
+        processed_root = os.environ.get("ROBOMETER_PROCESSED_DATASETS_PATH", "")
+        # Normalize: strip processed_datasets when env points to repo root
+        processed_root = processed_root.replace("processed_datasets/", "")
+        processed_root = processed_root.replace("processed_datasets", "")
+        if processed_root:
+            embeddings_path = os.path.join(processed_root, embeddings_path)
+
+    with open(embeddings_path, "rb") as f:
+        embeddings_data = torch.load(f, map_location="cpu")
+    return embeddings_data
+
+
+def pad_trajectory_to_max_frames_np(
+    frames: np.ndarray, progress: List[float], max_frames: int, pad_from: str = "right"
+) -> Tuple[np.ndarray, List[float]]:
+    """Pad trajectory frames and progress to max_frames by repeating the first frame/progress if needed.
+
+    Args:
+        frames: Trajectory frames (numpy array)
+        progress: Progress values (list of floats)
+        max_frames: Target number of frames
+
+    Returns:
+        Tuple[np.ndarray, List[float]: (padded_frames, padded_progress)
+    """
+    current_frames = frames.shape[0]
+
+    if current_frames >= max_frames:
+        # No padding needed
+        return frames, progress
+
+    if pad_from == "left":
+        pad_frame = frames[0:1]  # Keep the batch dimension
+        pad_progress = progress[0]
+    else:
+        pad_frame = frames[-1:]
+        pad_progress = progress[-1]
+
+    # Calculate how many frames to pad
+    frames_to_pad = max_frames - current_frames
+
+    # Pad frames by repeating the first frame
+    if pad_from == "left":
+        padded_frames = np.concatenate([np.repeat(pad_frame, frames_to_pad, axis=0), frames], axis=0)
+        padded_progress = [pad_progress] * frames_to_pad + progress
+    else:
+        padded_frames = np.concatenate([frames, np.repeat(pad_frame, frames_to_pad, axis=0)], axis=0)
+        padded_progress = progress + [pad_progress] * frames_to_pad
+
+    return padded_frames, padded_progress
+
+
+def pad_trajectory_to_max_frames_torch(
+    frames, progress: List[float], max_frames: int, pad_from: str = "right"
+) -> Tuple[torch.Tensor, List[float]]:
+    """Pad trajectory frames and progress to max_frames by repeating the first frame/progress if needed.
+
+    Args:
+        frames: PyTorch tensor
+        progress: Progress values (list of floats)
+        max_frames: Target number of frames
+
+    Returns:
+        Tuple[torch.Tensor, List[float]: (padded_frames, padded_progress)
+    """
+    current_frames = frames.shape[0]
+
+    if current_frames >= max_frames:
+        # No padding needed
+        return frames, progress
+
+    # Need to pad - repeat the first frame and first progress
+    if pad_from == "left":
+        pad_frame = frames[0:1]  # Keep the batch dimension
+        pad_progress = progress[0]
+    else:
+        pad_frame = frames[-1:]
+        pad_progress = progress[-1]
+
+    # Calculate how many frames to pad
+    frames_to_pad = max_frames - current_frames
+
+    # Pad frames by repeating the first frame
+    padding = pad_frame.repeat(frames_to_pad, 1)
+
+    if pad_from == "left":
+        padded_frames = torch.cat([padding, frames], dim=0)
+        padded_progress = [pad_progress] * frames_to_pad + progress
+    else:
+        padded_frames = torch.cat([frames, padding], dim=0)
+        padded_progress = progress + [pad_progress] * frames_to_pad
+
+    return padded_frames, padded_progress
+
+
+def linspace_subsample_frames(
+    frames: np.ndarray, num_frames: int = 8, end_idx: Optional[int] = None
+) -> Tuple[np.ndarray, List[int]]:
+    """Uniformly subsample frames from a trajectory and return the indices.
+
+    This method takes the full trajectory (e.g., 64 frames) and uniformly subsamples
+    num_frames from it. The first and last frames are always included.
+
+    Args:
+        frames: Full trajectory frames (N frames)
+        num_frames: Number of frames to subsample (default: 8)
+        end_idx: Optional end index to subsample up to (if None, uses total_frames - 1)
+
+    Returns:
+        Tuple[np.ndarray, List[int]: (subsampled_frames, subsampled_indices)
+    """
+    if hasattr(frames, "shape"):
+        total_frames = frames.shape[0]
+    else:
+        total_frames = len(frames)
+
+    if total_frames <= 0:
+        return frames, []
+
+    # Use end_idx if provided, otherwise use full trajectory
+    if end_idx is not None:
+        end_idx = min(end_idx, total_frames - 1)
+        frames_to_subsample = frames[: end_idx + 1]
+        effective_total = end_idx + 1
+    else:
+        frames_to_subsample = frames
+        effective_total = total_frames
+
+    if effective_total <= num_frames:
+        # If we have fewer (or equal) frames than requested, return all frames
+        indices = list(range(effective_total))
+        return frames_to_subsample, indices
+
+    # Special case: if num_frames == 1, always take the last frame
+    if num_frames == 1:
+        indices = [effective_total - 1]
+        subsampled_frames = frames_to_subsample[indices]
+        return subsampled_frames, indices
+
+    # Evenly spaced indices from 0 to effective_total-1, inclusive
+    indices_np = np.linspace(0, effective_total - 1, num_frames)
+    indices = np.rint(indices_np).astype(int).tolist()
+
+    # Enforce first and last explicitly
+    indices[0] = 0
+    indices[-1] = effective_total - 1
+
+    # Ensure indices are strictly non-decreasing and within bounds
+    for k in range(1, len(indices)):
+        if indices[k] < indices[k - 1]:
+            indices[k] = indices[k - 1]
+        if indices[k] >= effective_total:
+            indices[k] = effective_total - 1
+
+    # Subsample frames
+    subsampled_frames = frames_to_subsample[indices]
+
+    return subsampled_frames, indices
+
+
+def randomly_subsample_frames(
+    frames: np.ndarray, num_frames: int = 8, seed: Optional[int] = None
+) -> Tuple[np.ndarray, List[int]]:
+    """Randomly subsample frames from a trajectory and return the indices.
+
+    This method takes the full trajectory and randomly selects num_frames from it.
+    This is useful for creating diverse trajectory samples and avoiding bias
+    towards specific frame patterns.
+
+    Args:
+        frames: Full trajectory frames
+        num_frames: Number of frames to subsample (default: 8)
+        seed: Random seed for reproducible sampling (default: None)
+
+    Returns:
+        Tuple[np.ndarray, List[int]: (subsampled_frames, subsampled_indices)
+    """
+    if hasattr(frames, "shape"):
+        total_frames = frames.shape[0]
+    else:
+        total_frames = len(frames)
+
+    if total_frames < num_frames:
+        # If we have fewer frames than requested, return all frames
+        indices = list(range(total_frames))
+        return frames, indices
+
+    # Set random seed if provided
+    if seed is not None:
+        random.seed(seed)
+        np.random.seed(seed)
+
+    # Randomly sample indices without replacement
+    indices = sorted(random.sample(range(total_frames), num_frames))
+
+    # Subsample frames
+    subsampled_frames = frames[indices]
+
+    return subsampled_frames, indices
+
+
+def get_segment_indices_with_middle(
+    num_frames_total: int,
+    start_idx: int,
+    end_idx: Optional[int],
+    middle_idx: Optional[int] = None,
+    max_frames: int = 8,
+) -> list[int]:
+    """Get frame indices following the path: start -> middle -> end, constrained by max_frames.
+
+    Logic:
+    1. Handle edge cases for max_frames == 1 (only start_idx) or max_frames == 2 (start_idx and end_idx)
+    2. Handle edge cases for num_frames_total == 1 or 2
+    3. If middle_idx is None, set it to end_idx (simple start -> end path)
+    4. Enumerate all frames from start -> middle
+    5. Enumerate all frames from middle -> end
+    6. Join the two segments together, removing duplicate middle
+    7. Subsample to max_frames using linspace
+
+    Args:
+        num_frames_total: Total number of frames in the trajectory
+        start_idx: Start index
+        end_idx: End index (can be None for max_frames == 1 case)
+        middle_idx: Optional middle index. If None, set to end_idx (simple start -> end path)
+        max_frames: Maximum number of indices to return
+
+    Returns:
+        List of indices into the original frames array, ordered as start -> middle -> end, constrained to max_frames
+    """
+    # Handle edge case: num_frames_total == 1
+    if num_frames_total == 1:
+        return [0]
+
+    # Handle edge case: num_frames_total == 2
+    if num_frames_total == 2:
+        if max_frames == 1:
+            return [0]
+        return [0, 1]
+
+    # Ensure indices are valid (allow end_idx < start_idx for reverse progress)
+    start_idx = max(0, min(start_idx, num_frames_total - 1))
+
+    # Handle edge case: max_frames == 1 (only start_idx, end_idx and middle_idx are None)
+    if end_idx is None or max_frames == 1:
+        return [start_idx]
+
+    end_idx = max(0, min(end_idx, num_frames_total - 1))
+
+    # Handle edge case: max_frames == 2 (start_idx and end_idx, no middle)
+    if middle_idx is None:
+        # Simple start -> end path
+        if start_idx <= end_idx:
+            indices = list(range(start_idx, end_idx + 1))
+        else:
+            # Reverse: end -> start
+            indices = list(range(end_idx, start_idx + 1))
+            indices.reverse()
+        return indices
+
+    middle_idx = max(0, min(middle_idx, num_frames_total - 1))
+
+    # Enumerate frames from start -> middle
+    if start_idx <= middle_idx:
+        # Forward: start -> middle
+        segment1_indices = list(range(start_idx, middle_idx + 1))
+    else:
+        # Backward: start -> middle (going backwards)
+        segment1_indices = list(range(middle_idx, start_idx + 1))
+        segment1_indices.reverse()  # Descending order: [start_idx, start_idx-1, ..., middle_idx]
+
+    # Enumerate frames from middle -> end
+    if middle_idx <= end_idx:
+        # Forward: middle -> end
+        segment2_indices = list(range(middle_idx, end_idx + 1))
+    else:
+        # Backward: middle -> end (going backwards)
+        segment2_indices = list(range(end_idx, middle_idx + 1))
+        segment2_indices.reverse()  # Descending order: [middle_idx, middle_idx-1, ..., end_idx]
+
+    # Join the two segments, removing duplicate middle
+    if segment1_indices and segment2_indices:
+        # Remove middle from segment2 since it's already in segment1
+        segment2_indices = [idx for idx in segment2_indices if idx != middle_idx]
+        combined_indices = segment1_indices + segment2_indices
+    elif segment1_indices:
+        combined_indices = segment1_indices
+    elif segment2_indices:
+        combined_indices = segment2_indices
+    else:
+        # All indices are the same
+        combined_indices = [start_idx]
+
+    return combined_indices
+
+
+def convert_absolute_to_relative_progress(absolute_progress: List[float]) -> List[float]:
+    """Convert absolute progress values to relative deltas.
+
+    Args:
+        absolute_progress: List of absolute progress values (cumulative)
+
+    Returns:
+        List of relative progress deltas where first element is 0.0
+        and subsequent elements are deltas: progress[i] - progress[i-1]
+    """
+    if not absolute_progress:
+        return []
+
+    relative_progress = [0.0]
+    for i in range(1, len(absolute_progress)):
+        relative_progress.append(absolute_progress[i] - absolute_progress[i - 1])
+
+    return relative_progress
+
+
+def _compute_absolute_wrt_total_frames_progress(
+    num_frames_total: int,
+    subsampled_indices: List[int],
+    success_cutoff: Optional[float] = None,
+) -> List[float]:
+    """Compute progress using absolute_wrt_total_frames method.
+
+    Args:
+        num_frames_total: Total number of frames in the original trajectory.
+        subsampled_indices: Absolute indices into the original trajectory.
+        success_cutoff: Optional success cutoff threshold.
+
+    Returns:
+        List of progress values: progress[i] = (subsampled_indices[i] + 1) / num_frames_total.
+    """
+    segment_progress: List[float] = []
+    cutoff_index: Optional[int] = None
+    if success_cutoff is not None and success_cutoff > 0:
+        # Index of the first frame where progress exceeds the cutoff
+        cutoff_index = int(success_cutoff * num_frames_total)
+
+    for abs_idx in subsampled_indices:
+        if cutoff_index is not None and abs_idx >= cutoff_index:
+            # All frames after cutoff get 1.0 progress
+            progress = 1.0
+        else:
+            progress = (abs_idx + 1) / num_frames_total
+        segment_progress.append(progress)
+    return segment_progress
+
+
+def _compute_absolute_first_frame_progress(
+    num_frames_total: int,
+    subsampled_indices: List[int],
+    success_cutoff: Optional[float] = None,
+) -> List[float]:
+    """Compute progress using absolute_first_frame method.
+
+    Args:
+        num_frames_total: Total number of frames in the original trajectory.
+        subsampled_indices: Absolute indices into the original trajectory.
+        success_cutoff: Optional success cutoff threshold.
+
+    Returns:
+        List of progress values: progress[i] = (subsampled_indices[i] - start_idx) / (num_frames_total - start_idx - 1),
+        where start_idx is the first index in subsampled_indices.
+    """
+    if not subsampled_indices:
+        return []
+
+    # Get the start index (first index in the subsampled indices)
+    start_idx = min(subsampled_indices)
+
+    cutoff_index: Optional[int] = None
+    if success_cutoff is not None and success_cutoff > 0:
+        # Index of the first frame where progress exceeds the cutoff
+        cutoff_index = int(success_cutoff * num_frames_total)
+
+    segment_progress: List[float] = []
+    for abs_idx in subsampled_indices:
+        # Calculate relative position from start
+        relative_pos = abs_idx - start_idx
+
+        if cutoff_index is not None:
+            # ensure denominator is at least 1 to avoid division by zero
+            denominator = max(1, cutoff_index - start_idx - 1)
+            # if it goes pass the cutoff, cap at 1.0
+            computed_progress = relative_pos / denominator
+            segment_progress.append(min(1.0, computed_progress))
+        else:
+            # ensure denominator is at least 1 to avoid division by zero
+            denominator = max(1, num_frames_total - start_idx - 1)
+            # Normal progress calculation
+            segment_progress.append(relative_pos / denominator)
+
+    return segment_progress
+
+
+def _compute_relative_first_frame_progress(
+    num_frames_total: int,
+    subsampled_indices: List[int],
+    success_cutoff: Optional[float] = None,
+) -> List[float]:
+    """Compute progress using relative_first_frame method.
+
+    Args:
+        num_frames_total: Total number of frames in the original trajectory.
+        subsampled_indices: Absolute indices into the original trajectory.
+        success_cutoff: Optional success cutoff threshold.
+
+    Returns:
+        List of relative progress deltas: progress[0] = 0.0;
+        progress[i] = (subsampled_indices[i] - subsampled_indices[i-1]) / (num_frames_total - start_idx).
+    """
+    # First compute absolute_first_frame progress
+    absolute_progress = _compute_absolute_first_frame_progress(num_frames_total, subsampled_indices, success_cutoff)
+    # Convert to relative deltas
+    return convert_absolute_to_relative_progress(absolute_progress)
+
+
+def compute_progress_from_segment(
+    num_frames_total: int,
+    frame_indices: List[int],
+    progress_pred_type: str = "absolute_first_frame",
+    success_cutoff: Optional[float] = None,
+    partial_success: Optional[float] = None,
+) -> List[float]:
+    """Compute progress values given total frames and subsampled indices.
+
+    Args:
+        num_frames_total: Total number of frames in the original trajectory (before segmenting).
+        frame_indices: Absolute indices into the original trajectory.
+        progress_pred_type: Type of progress calculation:
+            - "absolute_first_frame": progress[i] = (frame_indices[i] - start_idx) / (num_frames_total - start_idx - 1),
+              where start_idx is the minimum index in frame_indices.
+            - "relative_first_frame": progress[0] = 0.0; progress[i] = (frame_indices[i] - frame_indices[i-1]) / (num_frames_total - start_idx).
+            - "absolute_wrt_total_frames": progress[i] = (frame_indices[i] + 1) / num_frames_total.
+        success_cutoff: Optional success cutoff threshold.
+        partial_success: Optional partial success value to use for frames past cutoff instead of 1.0.
+
+    Returns:
+        List of progress values based on the specified progress_pred_type.
+    """
+    if progress_pred_type == "absolute_wrt_total_frames":
+        progress = _compute_absolute_wrt_total_frames_progress(num_frames_total, frame_indices, success_cutoff)
+    elif progress_pred_type == "relative_first_frame":
+        progress = _compute_relative_first_frame_progress(num_frames_total, frame_indices, success_cutoff)
+    else:  # default: "absolute_first_frame"
+        progress = _compute_absolute_first_frame_progress(num_frames_total, frame_indices, success_cutoff)
+
+    # If partial_success is present and not 1.0, find whichever frame corresponds to the last frame of the original trajectory
+    if partial_success is not None and abs(partial_success - 1.0) > 1e-6 and progress and len(frame_indices) > 0:
+        # Find the index in frame_indices that corresponds to num_frames_total - 1 (last frame of original trajectory)
+        last_frame_idx = num_frames_total - 1
+        frame_idx_in_subsampled = next(
+            (i for i, abs_idx in enumerate(frame_indices) if abs_idx == last_frame_idx), None
+        )
+
+        # If we found the last frame in the subsampled indices, set it to partial_success and others to 0.0
+        if frame_idx_in_subsampled is not None:
+            modified_progress = [0.0] * len(progress)
+            modified_progress[frame_idx_in_subsampled] = partial_success
+            return modified_progress
+
+    return progress
+
+
+def create_trajectory_from_dict(traj_dict: Dict[str, Any], overrides: Optional[Dict[str, Any]] = None) -> Trajectory:
+    """Create a Trajectory from a dictionary with optional field overrides.
+
+    This helper function simplifies Trajectory creation by extracting common fields
+    from a trajectory dictionary and allowing specific fields to be overridden.
+
+    Args:
+        traj_dict: Dictionary containing trajectory data (e.g., from dataset)
+        overrides: Dictionary of field values to override (e.g., frames, target_progress)
+
+    Returns:
+        A new Trajectory instance
+    """
+    traj_data = {
+        "id": traj_dict.get("id"),
+        "task": traj_dict.get("task"),
+        "lang_vector": traj_dict.get("lang_vector"),
+        "data_source": traj_dict.get("data_source"),
+        "quality_label": traj_dict.get("quality_label"),
+        "is_robot": traj_dict.get("is_robot"),
+        "partial_success": traj_dict.get("partial_success"),
+    }
+
+    if overrides:
+        traj_data.update(overrides)
+
+    return Trajectory.model_validate(traj_data)
+
+
+def show_available_datasets():
+    """Show which datasets are available in the cache."""
+    # The preprocessing script now creates individual cache directories for each dataset/subset pair
+    cache_dir = os.environ.get("ROBOMETER_PROCESSED_DATASETS_PATH", "")
+    if not cache_dir:
+        raise ValueError(
+            "ROBOMETER_PROCESSED_DATASETS_PATH not set. Set it to the directory containing your processed datasets."
+        )
+
+    print("=" * 100)
+    print("Available datasets:")
+
+    # List all subdirectories (individual dataset caches)
+    if os.path.exists(cache_dir):
+        subdirs = [d for d in os.listdir(cache_dir) if os.path.isdir(os.path.join(cache_dir, d))]
+        if subdirs:
+            for subdir in sorted(subdirs):
+                # Try to load dataset info
+                info_file = os.path.join(cache_dir, subdir, "dataset_info.json")
+                if os.path.exists(info_file):
+                    with open(info_file) as f:
+                        info = json.load(f)
+                    dataset_path = info.get("dataset_path", "unknown")
+                    subset = info.get("subset", "unknown")
+                    trajectories = info.get("total_trajectories", 0)
+                    print(f"   {dataset_path}/{subset}: {trajectories} trajectories")
+        else:
+            print("  ❌ No dataset caches found")
+    else:
+        print("  ❌ Cache directory does not exist")
+    print("=" * 100)
diff --git a/robometer/robometer/data/datasets/name_mapping.py b/robometer/robometer/data/datasets/name_mapping.py
new file mode 100644
index 0000000000000000000000000000000000000000..cbdaec3eeea876770d2301fc368b62d92ed3131a
--- /dev/null
+++ b/robometer/robometer/data/datasets/name_mapping.py
@@ -0,0 +1,96 @@
+DS_SHORT_NAME_MAPPING = {
+    # MetaWorld
+    "aliangdw_metaworld_metaworld_train": "mw_train",
+    "aliangdw_metaworld_metaworld_eval": "mw_eval",
+    # LIBERO
+    "abraranwar_libero_rfm_libero256_10": "libero_10",
+    "abraranwar_libero_rfm_libero256_90": "libero_90",
+    "abraranwar_libero_rfm_libero256_object": "libero_object",
+    "abraranwar_libero_rfm_libero256_spatial": "libero_spatial",
+    "abraranwar_libero_rfm_libero256_goal": "libero_goal",
+    # LIBERO Failure
+    "ykorkmaz_libero_failure_rfm_libero_10_failure": "libero_10_failure",
+    "ykorkmaz_libero_failure_rfm_libero_90_failure": "libero_90_failure",
+    "ykorkmaz_libero_failure_rfm_libero_object_failure": "libero_object_failure",
+    "ykorkmaz_libero_failure_rfm_libero_spatial_failure": "libero_spatial_failure",
+    "ykorkmaz_libero_failure_rfm_libero_goal_failure": "libero_goal_failure",
+    # OXE Train
+    "jesbu1_oxe_rfm_oxe_aloha_mobile": "oxe_aloha_mobile",
+    "jesbu1_oxe_rfm_oxe_austin_buds_dataset_converted_externally_to_rlds": "oxe_austin_buds",
+    "jesbu1_oxe_rfm_oxe_bc_z": "oxe_bc_z",
+    "jesbu1_oxe_rfm_oxe_berkeley_cable_routing": "oxe_berkeley_cable",
+    "jesbu1_oxe_rfm_oxe_berkeley_fanuc_manipulation": "oxe_berkeley_fanuc",
+    "jesbu1_oxe_rfm_oxe_berkeley_mvp_converted_externally_to_rlds": "oxe_berkeley_mvp",
+    "jesbu1_oxe_rfm_oxe_berkeley_rpt_converted_externally_to_rlds": "oxe_berkeley_rpt",
+    "jesbu1_oxe_rfm_oxe_bridge_v2": "oxe_bridge_v2",
+    "jesbu1_oxe_rfm_oxe_dlr_edan_shared_control_converted_externally_to_rlds": "oxe_dlr_edan",
+    "jesbu1_oxe_rfm_oxe_droid": "oxe_droid",
+    "jesbu1_oxe_rfm_oxe_fractal20220817_data": "oxe_fractal",
+    "jesbu1_oxe_rfm_oxe_furniture_bench_dataset_converted_externally_to_rlds": "oxe_furniture_bench",
+    "jesbu1_oxe_rfm_oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds": "oxe_cmu_pickup",
+    "jesbu1_oxe_rfm_oxe_imperialcollege_sawyer_wrist_cam": "oxe_imperial_sawyer",
+    "jesbu1_oxe_rfm_oxe_language_table": "oxe_language_table",
+    "jesbu1_oxe_rfm_oxe_jaco_play": "oxe_jaco",
+    "jesbu1_oxe_rfm_oxe_nyu_rot_dataset_converted_externally_to_rlds": "oxe_nyu_rot",
+    "jesbu1_oxe_rfm_oxe_robo_set": "oxe_robo_set",
+    "jesbu1_oxe_rfm_oxe_stanford_hydra_dataset_converted_externally_to_rlds": "oxe_stanford_hydra",
+    "jesbu1_oxe_rfm_oxe_tokyo_u_lsmo_converted_externally_to_rlds": "oxe_tokyo_lsmo",
+    "jesbu1_oxe_rfm_oxe_toto": "oxe_toto",
+    "jesbu1_oxe_rfm_oxe_ucsd_kitchen_dataset_converted_externally_to_rlds": "oxe_ucsd_kitchen",
+    "jesbu1_oxe_rfm_oxe_utaustin_mutex": "oxe_utaustin_mutex",
+    # OXE Eval
+    "jesbu1_oxe_rfm_eval_oxe_bc_z_eval": "oxe_bc_z_eval",
+    "jesbu1_oxe_rfm_eval_oxe_berkeley_cable_routing_eval": "oxe_berkeley_cable_eval",
+    "jesbu1_oxe_rfm_eval_oxe_bridge_v2_eval": "oxe_bridge_v2_eval",
+    "jesbu1_oxe_rfm_eval_oxe_jaco_play_eval": "oxe_jaco_eval",
+    "jesbu1_oxe_rfm_eval_oxe_toto_eval": "oxe_toto_eval",
+    "jesbu1_oxe_rfm_eval_oxe_viola_eval": "oxe_viola_eval",
+    # Others - Train
+    "jesbu1_molmoact_rfm_molmoact_dataset_household": "molmoact_household",
+    "jesbu1_molmoact_rfm_molmoact_dataset_tabletop": "molmoact_tabletop",
+    "jesbu1_egodex_rfm_egodex_part1": "egodex_part1",
+    "jesbu1_egodex_rfm_egodex_part2": "egodex_part2",
+    "jesbu1_egodex_rfm_egodex_part3": "egodex_part3",
+    "jesbu1_egodex_rfm_egodex_part4": "egodex_part4",
+    "jesbu1_egodex_rfm_egodex_part5": "egodex_part5",
+    "abraranwar_agibotworld_alpha_rfm_agibotworld": "agibotworld",
+    "jesbu1_epic_rfm_epic": "epic_kitchens",
+    "jesbu1_galaxea_rfm_galaxea_part1_r1_lite": "galaxea_part1",
+    "jesbu1_galaxea_rfm_galaxea_part2_r1_lite": "galaxea_part2",
+    "jesbu1_galaxea_rfm_galaxea_part3_r1_lite": "galaxea_part3",
+    "jesbu1_galaxea_rfm_galaxea_part4_r1_lite": "galaxea_part4",
+    "jesbu1_galaxea_rfm_galaxea_part5_r1_lite": "galaxea_part5",
+    "jesbu1_humanoid_everyday_rfm_humanoid_everyday_rfm": "humanoid_everyday",
+    # Others - Eval
+    "jesbu1_egodex_rfm_egodex_test": "egodex_test",
+    # Paired
+    "jesbu1_h2r_rfm_h2r": "h2r",
+    "jesbu1_motif_rfm_motif_rfm": "motif",
+    "anqil_rh20t_subset_rfm_rh20t_human": "rh20t",
+    "anqil_rh20t_subset_rfm_rh20t_robot": "rh20t",
+    "jesbu1_ph2d_rfm_ph2d": "ph2d",
+    "aliangdw_utd_so101_human_utd_so101_human": "utd_so101",
+    "aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking": "utd_so101",
+    # Suboptimal/Fail
+    "jesbu1_roboarena_0825_rfm_roboarena": "roboarena",
+    "jesbu1_fino_net_rfm_fino_net": "fino_net",
+    "jesbu1_failsafe_rfm_failsafe": "failsafe",
+    "jesbu1_soar_rfm_soar_rfm": "soar",
+    "jesbu1_auto_eval_rfm_auto_eval_rfm": "auto_eval",
+    "jesbu1_racer_rfm_racer_train": "racer_train",
+    "jesbu1_racer_rfm_racer_val": "racer_val",
+    "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist": "roboarena_eval",
+    "jesbu1_roboarena_0825_rfm_roboarena": "roboarena_w_wrist",
+    # Policy Ranking
+    "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking": "usc_xarm",
+    "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking": "usc_franka",
+    "jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm": "mit_franka",
+    "jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top": "utd_so101_clean_top",
+    "jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist": "utd_so101_clean_wrist",
+    "jesbu1_rfm_new_mit_franka_rfm_rfm_new_mit_franka_rfm": "rfm_new_mit_franka",
+    "ykorkmaz_usc_trossen_rfm_usc_trossen": "usc_trossen",
+    # Reward Alignment
+    "abraranwar_usc_koch_rewind_rfm_usc_koch_rewind": "usc_koch_rewind",
+    # RoboFAC
+    "aliangdw_robofac_rbm_robofac": "robofac",
+}
diff --git a/robometer/robometer/data/datasets/name_mapping_final.py b/robometer/robometer/data/datasets/name_mapping_final.py
new file mode 100644
index 0000000000000000000000000000000000000000..ee832291370d5e055aaa361d68d95891a78f9379
--- /dev/null
+++ b/robometer/robometer/data/datasets/name_mapping_final.py
@@ -0,0 +1,96 @@
+DS_SHORT_NAME_MAPPING = {
+    # MetaWorld
+    "aliangdw_metaworld_metaworld_train": "Metaworld (Train)",
+    "aliangdw_metaworld_metaworld_eval": "Metaworld (Eval)",
+    # LIBERO
+    "abraranwar_libero_rfm_libero256_10": "Libero (10)",
+    "abraranwar_libero_rfm_libero256_90": "Libero (90)",
+    "abraranwar_libero_rfm_libero256_object": "Libero (Object)",
+    "abraranwar_libero_rfm_libero256_spatial": "Libero (Spatial)",
+    "abraranwar_libero_rfm_libero256_goal": "Libero (Goal)",
+    # LIBERO Failure
+    "ykorkmaz_libero_failure_rfm_libero_10_failure": "Libero (10) Failure",
+    "ykorkmaz_libero_failure_rfm_libero_90_failure": "Libero (90) Failure",
+    "ykorkmaz_libero_failure_rfm_libero_object_failure": "Libero (Object) Failure",
+    "ykorkmaz_libero_failure_rfm_libero_spatial_failure": "Libero (Spatial) Failure",
+    "ykorkmaz_libero_failure_rfm_libero_goal_failure": "Libero (Goal) Failure",
+    # OXE Train
+    "jesbu1_oxe_rfm_oxe_aloha_mobile": "OXE (Aloha Mobile)",
+    "jesbu1_oxe_rfm_oxe_austin_buds_dataset_converted_externally_to_rlds": "oxe_austin_buds",
+    "jesbu1_oxe_rfm_oxe_bc_z": "OXE (BC-Z)",
+    "jesbu1_oxe_rfm_oxe_berkeley_cable_routing": "oxe_berkeley_cable",
+    "jesbu1_oxe_rfm_oxe_berkeley_fanuc_manipulation": "OXE (Berkeley Fanuc Manipulation)",
+    "jesbu1_oxe_rfm_oxe_berkeley_mvp_converted_externally_to_rlds": "OXE (Berkeley MVP)",
+    "jesbu1_oxe_rfm_oxe_berkeley_rpt_converted_externally_to_rlds": "OXE (Berkeley RPT)",
+    "jesbu1_oxe_rfm_oxe_bridge_v2": "OXE (Bridge V2)",
+    "jesbu1_oxe_rfm_oxe_dlr_edan_shared_control_converted_externally_to_rlds": "OXE (DLR EDAN Shared Control)",
+    "jesbu1_oxe_rfm_oxe_droid": "OXE (Droid)",
+    "jesbu1_oxe_rfm_oxe_fractal20220817_data": "OXE (Fractal 20220817 Data)",
+    "jesbu1_oxe_rfm_oxe_furniture_bench_dataset_converted_externally_to_rlds": "oxe_furniture_bench",
+    "jesbu1_oxe_rfm_oxe_iamlab_cmu_pickup_insert_converted_externally_to_rlds": "OXE (CMU Pickup Insert)",
+    "jesbu1_oxe_rfm_oxe_imperialcollege_sawyer_wrist_cam": "oxe_imperial_sawyer",
+    "jesbu1_oxe_rfm_oxe_language_table": "OXE (Language Table)",
+    "jesbu1_oxe_rfm_oxe_jaco_play": "OXE (Jaco Play)",
+    "jesbu1_oxe_rfm_oxe_nyu_rot_dataset_converted_externally_to_rlds": "OXE (NYU Rot Dataset)",
+    "jesbu1_oxe_rfm_oxe_robo_set": "OXE (Robo Set)",
+    "jesbu1_oxe_rfm_oxe_stanford_hydra_dataset_converted_externally_to_rlds": "OXE (Stanford Hydra)",
+    "jesbu1_oxe_rfm_oxe_tokyo_u_lsmo_converted_externally_to_rlds": "OXE (Tokyo U LSMo)",
+    "jesbu1_oxe_rfm_oxe_toto": "OXE (Toto)",
+    "jesbu1_oxe_rfm_oxe_ucsd_kitchen_dataset_converted_externally_to_rlds": "OXE (UCSD Kitchen Dataset)",
+    "jesbu1_oxe_rfm_oxe_utaustin_mutex": "OXE (UT Austin Mutex)",
+    # OXE Eval
+    "jesbu1_oxe_rfm_eval_oxe_bc_z_eval": "OXE (BC-Z Eval)",
+    "jesbu1_oxe_rfm_eval_oxe_berkeley_cable_routing_eval": "OXE (Berkeley Cable Routing Eval)",
+    "jesbu1_oxe_rfm_eval_oxe_bridge_v2_eval": "OXE (Bridge V2 Eval)",
+    "jesbu1_oxe_rfm_eval_oxe_jaco_play_eval": "OXE (Jaco Play Eval)",
+    "jesbu1_oxe_rfm_eval_oxe_toto_eval": "OXE (Toto Eval)",
+    "jesbu1_oxe_rfm_eval_oxe_viola_eval": "OXE (Viola Eval)",
+    # Others - Train
+    "jesbu1_molmoact_rfm_molmoact_dataset_household": "Molmoact (Household)",
+    "jesbu1_molmoact_rfm_molmoact_dataset_tabletop": "molmoact_tabletop",
+    "jesbu1_egodex_rfm_egodex_part1": "Egodex (Part 1)",
+    "jesbu1_egodex_rfm_egodex_part2": "Egodex (Part 2)",
+    "jesbu1_egodex_rfm_egodex_part3": "Egodex (Part 3)",
+    "jesbu1_egodex_rfm_egodex_part4": "Egodex (Part 4)",
+    "jesbu1_egodex_rfm_egodex_part5": "Egodex (Part 5)",
+    "abraranwar_agibotworld_alpha_rfm_agibotworld": "Agibotworld",
+    "jesbu1_epic_rfm_epic": "epic_kitchens",
+    "jesbu1_galaxea_rfm_galaxea_part1_r1_lite": "Galaxea (Part 1)",
+    "jesbu1_galaxea_rfm_galaxea_part2_r1_lite": "Galaxea (Part 2)",
+    "jesbu1_galaxea_rfm_galaxea_part3_r1_lite": "Galaxea (Part 3)",
+    "jesbu1_galaxea_rfm_galaxea_part4_r1_lite": "Galaxea (Part 4)",
+    "jesbu1_galaxea_rfm_galaxea_part5_r1_lite": "Galaxea (Part 5)",
+    "jesbu1_humanoid_everyday_rfm_humanoid_everyday_rfm": "Humanoid Everyday",
+    # Others - Eval
+    "jesbu1_egodex_rfm_egodex_test": "Egodex (Test)",
+    # Paired
+    "jesbu1_h2r_rfm_h2r": "H2R",
+    "jesbu1_motif_rfm_motif_rfm": "Motif",
+    "anqil_rh20t_subset_rfm_rh20t_human": "Rh20t (Human)",
+    "anqil_rh20t_subset_rfm_rh20t_robot": "Rh20t (Robot)",
+    "jesbu1_ph2d_rfm_ph2d": "Ph2d",
+    "aliangdw_utd_so101_human_utd_so101_human": "Utd SO101 (Human)",
+    "aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking": "Utd SO101 (Policy Ranking)",
+    # Suboptimal/Fail
+    "jesbu1_roboarena_0825_rfm_roboarena": "Roboarena",
+    "jesbu1_fino_net_rfm_fino_net": "Fino Net",
+    "jesbu1_failsafe_rfm_failsafe": "Failsafe",
+    "jesbu1_soar_rfm_soar_rfm": "Soar",
+    "jesbu1_auto_eval_rfm_auto_eval_rfm": "Auto Eval",
+    "jesbu1_racer_rfm_racer_train": "Racer (Train)",
+    "jesbu1_racer_rfm_racer_val": "Racer (Val)",
+    "jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist": "Roboarena (Eval Debug No Wrist)",
+    "jesbu1_roboarena_0825_rfm_roboarena": "Roboarena (W Wrist)",
+    # Policy Ranking
+    "aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking": "Usc Xarm (Policy Ranking)",
+    "aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking": "Usc Franka (Policy Ranking)",
+    "jesbu1_mit_franka_p-rank_rfm_mit_franka_p-rank_rfm": "Mit Franka (P-Rank)",
+    "jesbu1_utd_so101_clean_policy_ranking_top_utd_so101_clean_policy_ranking_top": "Utd SO101 (Clean Policy Ranking Top)",
+    "jesbu1_utd_so101_clean_policy_ranking_wrist_utd_so101_clean_policy_ranking_wrist": "Utd SO101 (Clean Policy Ranking Wrist)",
+    "jesbu1_usc_koch_p_ranking_rfm_usc_koch_p_ranking_all": "Usc Koch (P-Rank All)",
+    # Reward Alignment
+    "abraranwar_usc_koch_rewind_rfm_usc_koch_rewind": "Usc Koch Rewind",
+    # Extras
+    "jesbu1_roboreward_rfm_roboreward_test": "RoboReward (Test)",
+    "jesbu1_roboreward_rfm_roboreward_val": "RoboReward (Val)",
+}
diff --git a/robometer/robometer/data/datasets/rbm_data.py b/robometer/robometer/data/datasets/rbm_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..7b62bc18d7d9998be143c84e37d2557b4ca57d1d
--- /dev/null
+++ b/robometer/robometer/data/datasets/rbm_data.py
@@ -0,0 +1,253 @@
+import random
+
+from robometer.data.datasets.base import BaseDataset
+from robometer.data.samplers.pref import PrefSampler
+from robometer.data.samplers.progress import ProgressSampler
+from robometer.data.dataset_category import is_preference_only
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class RBMDataset(BaseDataset):
+    """Dataset that combines preference and progress generation."""
+
+    def __init__(self, config, is_evaluation=False, max_samples=None, sampler_kwargs=None, **kwargs):
+        super().__init__(config, is_evaluation, **kwargs)
+
+        self.pref_sampler = None
+        self.progress_sampler = None
+
+        if sampler_kwargs is None:
+            sampler_kwargs = {}
+
+        base_sampler_kwargs = {
+            "config": config,
+            "dataset": self.dataset,
+            "combined_indices": self._combined_indices,
+            "dataset_success_cutoff_map": self.dataset_success_cutoff_map,
+            "verbose": False,
+            **sampler_kwargs,
+        }
+
+        if self.config.sample_type_ratio[0] > 0:
+            self.pref_sampler = PrefSampler(is_evaluation=is_evaluation, **base_sampler_kwargs)
+        if self.config.sample_type_ratio[1] > 0:
+            self.progress_sampler = ProgressSampler(is_evaluation=is_evaluation, **base_sampler_kwargs)
+
+        self.sample_type_ratio = config.sample_type_ratio
+        self.max_samples = max_samples
+
+        self.data_len = len(self.dataset)
+
+    def get_random_state(self) -> dict:
+        """Get random state from all samplers for checkpointing.
+
+        Returns:
+            Dictionary containing random state for all samplers
+        """
+        state = {
+            "pref_sampler": self.pref_sampler._local_random.getstate() if self.pref_sampler else None,
+            "progress_sampler": self.progress_sampler._local_random.getstate() if self.progress_sampler else None,
+        }
+        return state
+
+    def set_random_state(self, state: dict):
+        """Restore random state from checkpoint.
+
+        Args:
+            state: Dictionary containing random state for all samplers
+        """
+        if state.get("pref_sampler") and self.pref_sampler:
+            self.pref_sampler._local_random.setstate(state["pref_sampler"])
+        if state.get("progress_sampler") and self.progress_sampler:
+            self.progress_sampler._local_random.setstate(state["progress_sampler"])
+
+    def get_resample_attempt_stats(self):
+        return self._resample_attempt_stats
+
+    def get_resample_dataset_attempt_stats(self):
+        return self._resample_dataset_attempt_stats
+
+    def __len__(self):
+        if self.max_samples is None:
+            return self.data_len
+        else:
+            return self.max_samples
+
+    def __getitem__(self, idx):
+        """Create a data sample from the dataset."""
+        idx = idx % self.data_len
+        logger.trace(f"[RBMDataset] __getitem__: Starting for idx={idx}")
+
+        # Get the item from the filtered dataset
+        item = self.dataset[idx]
+        traj_id = item.get("id", "unknown")
+        logger.trace(f"[RBMDataset] __getitem__: Got item with ID={traj_id}, calling _generate_sample_from_item")
+
+        sample = self._generate_sample_from_item(item)
+        logger.trace(f"[RBMDataset] __getitem__: Successfully generated sample for idx={idx}, ID={traj_id}")
+        return sample
+
+    def _generate_sample_from_item(self, item):
+        """Shared sampler logic that can be reused by balanced datasets."""
+        traj_id = item["id"]
+        data_source = item["data_source"]
+        quality_label = item["quality_label"]
+        task_name = item["task"]
+
+        logger.trace(
+            f"[RBMDataset] _generate_sample_from_item: Starting for ID={traj_id}, data_source={data_source}, quality={quality_label}"
+        )
+
+        # Force preference-only for non-successful trajectories
+        if quality_label != "successful" and self.pref_sampler is not None:
+            logger.trace(
+                f"[RBMDataset] _generate_sample_from_item: Non-successful quality detected for ID={traj_id}, forcing preference-only"
+            )
+            sample = self.pref_sampler._generate_sample(item)
+            if sample is not None:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: Preference sample generated successfully for non-successful traj ID={traj_id}"
+                )
+                return self._set_resample_attempts(sample, 1)
+            else:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: Preference sampler returned None for non-successful traj ID={traj_id}"
+                )
+                # If preference fails for non-successful traj, we can't use other samplers
+                raise ValueError(
+                    f"Preference sampler failed for non-successful trajectory ID={traj_id} and no fallback available, task={task_name}"
+                )
+
+        # Preference-only override by data_source using raw filtered dataset entry
+        if is_preference_only(data_source) and self.pref_sampler is not None:
+            logger.trace(f"[RBMDataset] _generate_sample_from_item: Using preference-only override for ID={traj_id}")
+            sample = self.pref_sampler._generate_sample(item)
+            if sample is not None:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: Preference-only sample generated successfully for ID={traj_id}"
+                )
+                return self._set_resample_attempts(sample, 1)
+            else:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: Preference-only sampler returned None for ID={traj_id}, task={task_name}"
+                )
+
+        # Available samplers with their probabilities
+        samplers = [
+            ("pref", self.sample_type_ratio[0], self.pref_sampler),
+            ("progress", self.sample_type_ratio[1], self.progress_sampler),
+        ]
+
+        # Remove samplers with zero probability or None samplers
+        available_samplers = [
+            (name, prob, sampler) for name, prob, sampler in samplers if prob > 0 and sampler is not None
+        ]
+
+        logger.trace(
+            f"[RBMDataset] _generate_sample_from_item: Available samplers for ID={traj_id}: {[name for name, _, _ in available_samplers]}"
+        )
+
+        # Fallback to progress sampler if no samplers have positive probability
+        if not available_samplers:
+            if self.progress_sampler is not None:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: No available samplers, using progress fallback for ID={traj_id}"
+                )
+                sample = self.progress_sampler._generate_sample(item)
+                if sample is not None:
+                    return self._set_resample_attempts(sample, 1)
+            raise ValueError("No samplers available")
+
+        # Try samplers until we get a non-None result
+        # Limit max_attempts to prevent hangs with num_workers=0 in distributed training
+        # With num_workers=0, blocking here blocks the entire rank's main thread
+        max_attempts = min(len(available_samplers) * 2, 10)  # Cap at 10 to prevent infinite loops
+        attempt = 0
+        tried_samplers = set()
+
+        logger.trace(
+            f"[RBMDataset] _generate_sample_from_item: Starting sampling loop for ID={traj_id}, max_attempts={max_attempts}"
+        )
+
+        while attempt < max_attempts:
+            attempt += 1
+
+            # If we've tried all samplers, reset and try again
+            if len(tried_samplers) >= len(available_samplers):
+                tried_samplers.clear()
+
+            # Filter out already tried samplers if we haven't exhausted all options
+            remaining_samplers = [
+                (name, prob, sampler) for name, prob, sampler in available_samplers if name not in tried_samplers
+            ]
+
+            # If no remaining samplers, reset and try all again
+            if not remaining_samplers:
+                tried_samplers.clear()
+                remaining_samplers = available_samplers
+
+            # Normalize probabilities for remaining samplers
+            total_prob = sum(prob for _, prob, _ in remaining_samplers)
+            if total_prob == 0:
+                # Reset and try all samplers again
+                tried_samplers.clear()
+                remaining_samplers = available_samplers
+                total_prob = sum(prob for _, prob, _ in remaining_samplers)
+
+            normalized_samplers = [(name, prob / total_prob, sampler) for name, prob, sampler in remaining_samplers]
+
+            # Select sampler based on normalized probabilities
+            prob = random.random()
+            cumulative_prob = 0.0
+            selected_sampler = None
+            selected_name = None
+
+            for name, normalized_prob, sampler in normalized_samplers:
+                cumulative_prob += normalized_prob
+                if prob <= cumulative_prob:
+                    selected_sampler = sampler
+                    selected_name = name
+                    break
+
+            # Fallback: select first sampler if selection failed
+            if selected_sampler is None:
+                selected_name, _, selected_sampler = remaining_samplers[0]
+
+            logger.trace(
+                f"[RBMDataset] _generate_sample_from_item: Attempt {attempt}/{max_attempts} for ID={traj_id}, trying sampler '{selected_name}'"
+            )
+
+            # Try the selected sampler
+            sample = selected_sampler._generate_sample(item)
+
+            # If sample is not None, return it
+            if sample is not None:
+                logger.trace(
+                    f"[RBMDataset] _generate_sample_from_item: Successfully generated sample on attempt {attempt} using '{selected_name}' for ID={traj_id}"
+                )
+                return self._set_resample_attempts(sample, attempt)
+
+            # Sample is None, mark this sampler as tried
+            logger.trace(
+                f"[RBMDataset] _generate_sample_from_item: Attempt {attempt} failed (sampler '{selected_name}' returned None) for ID={traj_id}"
+            )
+            tried_samplers.add(selected_name)
+
+        logger.trace(
+            f"[RBMDataset] _generate_sample_from_item: All {max_attempts} attempts exhausted for ID={traj_id}, trying progress fallback"
+        )
+
+        # All attempts failed, try progress sampler as final fallback
+        if self.progress_sampler is not None:
+            sample = self.progress_sampler._generate_sample(item)
+            if sample is not None:
+                logger.trace(f"[RBMDataset] _generate_sample_from_item: Progress fallback succeeded for ID={traj_id}")
+                return self._set_resample_attempts(sample, attempt)
+
+        # Final fallback: raise error if all samplers returned None
+        logger.error(
+            f"[RBMDataset] _generate_sample_from_item: ERROR - All samplers failed for ID={traj_id} after {max_attempts} attempts"
+        )
+        raise ValueError(f"All samplers failed to generate a sample after {max_attempts} attempts")
diff --git a/robometer/robometer/data/datasets/repeated_dataset.py b/robometer/robometer/data/datasets/repeated_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..95a2725288d96f92367469aae71c72a95d1a086e
--- /dev/null
+++ b/robometer/robometer/data/datasets/repeated_dataset.py
@@ -0,0 +1,24 @@
+"""Minimal repeat wrapper for training; exposes .dataset for checkpointing."""
+
+from torch.utils.data import Dataset
+
+
+class RepeatedDataset(Dataset):
+    """Wraps a dataset and repeats it for a fixed number of effective steps (or a large default)."""
+
+    def __init__(self, dataset: Dataset, num_repeats: int = 1000):
+        self.dataset = dataset
+        self._len = len(dataset)
+        self.num_repeats = num_repeats
+        if self._len == 0:
+            self._effective_len = 0
+        else:
+            self._effective_len = min(self._len * num_repeats, self._len * 10_000)
+
+    def __len__(self) -> int:
+        return self._effective_len if self._effective_len > 0 else 1
+
+    def __getitem__(self, index: int):
+        if self._len == 0:
+            raise IndexError("Underlying dataset is empty")
+        return self.dataset[index % self._len]
diff --git a/robometer/robometer/data/datasets/strategy_first_dataset.py b/robometer/robometer/data/datasets/strategy_first_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..e74c277e179aa8744e844df246321e059fa5d59b
--- /dev/null
+++ b/robometer/robometer/data/datasets/strategy_first_dataset.py
@@ -0,0 +1,657 @@
+from collections import defaultdict
+from typing import Dict, List, Optional, Any
+from random import Random
+
+from robometer.data.datasets.base import BaseDataset
+from robometer.data.samplers.pref import PrefSampler
+from robometer.data.samplers.progress import ProgressSampler
+from robometer.data.datasets.helpers import DataGenStrat
+from robometer.data.dataset_category import (
+    is_preference_only,
+    is_suboptimal_fail_ds,
+    is_paired_ds,
+)
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class StrategyFirstDataset(BaseDataset):
+    """
+    Dataset that first selects sample type, then strategy, then picks a data source uniformly.
+
+    This is different from RBMDataset which selects a trajectory first based on dataset iteration,
+    and from StrategyBalancedDataset which selects sample type then data source (with optional weights).
+
+    Sampling flow:
+    1. Select sample type (preference/progress) based on sample_type_ratio
+    2. Select strategy for that sample type based on strategy ratios
+    3. Select data source uniformly from all available data sources
+    4. Sample trajectory from selected data source and generate sample
+    """
+
+    def __init__(self, config, is_evaluation=False, max_samples=None, sampler_kwargs=None, random_seed=42, **kwargs):
+        super().__init__(config, is_evaluation, **kwargs)
+
+        # Initialize local random instance for deterministic sampling
+        self._local_random = Random(random_seed)
+
+        self.pref_sampler = None
+        self.progress_sampler = None
+
+        if sampler_kwargs is None:
+            sampler_kwargs = {}
+
+        base_sampler_kwargs = {
+            "config": config,
+            "dataset": self.dataset,
+            "combined_indices": self._combined_indices,
+            "dataset_success_cutoff_map": self.dataset_success_cutoff_map,
+            "verbose": False,
+            **sampler_kwargs,
+        }
+
+        if self.config.sample_type_ratio[0] > 0:
+            self.pref_sampler = PrefSampler(is_evaluation=is_evaluation, **base_sampler_kwargs)
+        if self.config.sample_type_ratio[1] > 0:
+            self.progress_sampler = ProgressSampler(is_evaluation=is_evaluation, **base_sampler_kwargs)
+
+        self.sample_type_ratio = config.sample_type_ratio
+        self.max_samples = max_samples
+        self.data_len = len(self.dataset)
+
+        # Build source indices for efficient uniform sampling
+        self.source_indices = defaultdict(list)
+        if "data_source" in self.dataset.column_names:
+            sources = self.dataset["data_source"]
+            for i, source in enumerate(sources):
+                self.source_indices[source].append(i)
+
+        # Handle data source weights if provided
+        self.data_source_weights = getattr(config, "data_source_weights", None)
+        if self.data_source_weights:
+            self._normalize_data_source_weights()
+        else:
+            self.normalized_weights = None
+
+        # Build set of tasks with optimal trajectories for efficient filtering
+        self.tasks_with_optimal = set(self._combined_indices.get("optimal_by_task", {}).keys())
+
+        # Build set of tasks with both optimal and suboptimal trajectories for SUBOPTIMAL strategy
+        suboptimal_by_task = self._combined_indices.get("suboptimal_by_task", {})
+        # Only include tasks that have non-empty suboptimal indices
+        tasks_with_suboptimal = {task for task, indices in suboptimal_by_task.items() if indices}
+        self.tasks_with_both = self.tasks_with_optimal & tasks_with_suboptimal
+
+        # Build set of all indices from tasks with optimal trajectories for efficient filtering
+        task_indices = self._combined_indices.get("task_indices", {})
+        self.optimal_task_indices = set()
+        for task in self.tasks_with_optimal:
+            if task in task_indices:
+                self.optimal_task_indices.update(task_indices[task])
+
+        # Build set of all indices from tasks with both optimal and suboptimal trajectories
+        self.tasks_with_both_indices = set()
+        for task in self.tasks_with_both:
+            if task in task_indices:
+                self.tasks_with_both_indices.update(task_indices[task])
+
+        # Build set of successful trajectory indices for REWIND strategy filtering
+        quality_indices = self._combined_indices.get("quality_indices", {})
+        self.successful_indices = set(quality_indices.get("successful", []))
+
+        logger.info(f"StrategyFirstDataset initialized with {len(self.dataset)} trajectories")
+        logger.info(f"Sample type ratios: pref={self.sample_type_ratio[0]}, progress={self.sample_type_ratio[1]}")
+        logger.info(f"Available data sources: {list(self.source_indices.keys())}")
+        if self.normalized_weights:
+            logger.info("Data source weights enabled:")
+            for source, weight in sorted(self.normalized_weights.items(), key=lambda x: x[1], reverse=True):
+                logger.info(f"  {source}: {weight:.3f}")
+        logger.info(f"Tasks with optimal trajectories: {len(self.tasks_with_optimal)}")
+        logger.info(f"Tasks with both optimal and suboptimal trajectories: {len(self.tasks_with_both)}")
+
+    def get_random_state(self) -> dict:
+        """Get random state from dataset and all samplers for checkpointing.
+
+        Returns:
+            Dictionary containing random state for dataset and all samplers
+        """
+        state = {
+            "dataset": self._local_random.getstate() if hasattr(self, "_local_random") else None,
+            "pref_sampler": self.pref_sampler._local_random.getstate() if self.pref_sampler else None,
+            "progress_sampler": self.progress_sampler._local_random.getstate() if self.progress_sampler else None,
+        }
+        return state
+
+    def set_random_state(self, state: dict):
+        """Restore random state from checkpoint.
+
+        Args:
+            state: Dictionary containing random state for dataset and all samplers
+        """
+        if state.get("dataset") and hasattr(self, "_local_random"):
+            self._local_random.setstate(state["dataset"])
+        if state.get("pref_sampler") and self.pref_sampler:
+            self.pref_sampler._local_random.setstate(state["pref_sampler"])
+        if state.get("progress_sampler") and self.progress_sampler:
+            self.progress_sampler._local_random.setstate(state["progress_sampler"])
+
+    def __len__(self):
+        if self.max_samples is None:
+            return self.data_len
+        else:
+            return self.max_samples
+
+    def __getitem__(self, idx):
+        """Create a sample by selecting sample type, then strategy, then data source uniformly."""
+        logger.trace(f"[StrategyFirstDataset] __getitem__: Starting for idx={idx}")
+
+        # Step 1: Select sample type based on ratios
+        sample_type = self._select_sample_type()
+        logger.trace(f"[StrategyFirstDataset] Selected sample type: {sample_type}")
+
+        # Step 2: Select strategy for this sample type
+        strategy = self._select_strategy(sample_type)
+        if strategy is None:
+            # Fallback: try to generate sample without specific strategy
+            logger.trace(f"[StrategyFirstDataset] No strategy selected, using sampler default")
+            return self._generate_without_specific_strategy(sample_type)
+
+        logger.trace(
+            f"[StrategyFirstDataset] Selected strategy: {strategy.value if hasattr(strategy, 'value') else strategy}"
+        )
+
+        # Step 3: Filter data sources based on strategy requirements
+        filtered_sources = self._filter_data_sources_by_strategy(strategy)
+        if not filtered_sources:
+            logger.trace(
+                f"[StrategyFirstDataset] No viable data sources for strategy {strategy.value if hasattr(strategy, 'value') else strategy}, retrying..."
+            )
+            # Retry by selecting a different strategy/sample type
+            return self._generate_without_specific_strategy(sample_type)
+
+        # Step 4-6: Sample and generate using helper method
+        # First try with preferred strategy
+        sample = self._try_generate_sample(
+            sample_type=sample_type,
+            filtered_sources=filtered_sources,
+            strategy=strategy,
+            preferred_strategy=strategy,
+        )
+        if sample is not None:
+            return sample
+
+        # If preferred strategy failed, try without strategy (let sampler choose its own)
+        logger.trace(
+            f"[StrategyFirstDataset] Failed to generate {sample_type} sample with preferred strategy {strategy.value if hasattr(strategy, 'value') else strategy}, "
+            f"trying without strategy..."
+        )
+        sample = self._try_generate_sample(
+            sample_type=sample_type,
+            filtered_sources=filtered_sources,
+            strategy=strategy,  # Keep strategy for filtering, but let sampler choose
+            preferred_strategy=None,  # Let sampler select its own strategy
+        )
+        if sample is not None:
+            return sample
+
+        # All attempts failed for the selected sample type, try other samplers as fallback
+        logger.trace(f"[StrategyFirstDataset] Failed to generate {sample_type} sample, trying other samplers...")
+        return self._try_other_samplers(sample_type)
+
+    def _select_sample_type(self) -> str:
+        """Select a sample type based on sample_type_ratio."""
+        available_types = []
+        available_probs = []
+
+        if self.sample_type_ratio[0] > 0 and self.pref_sampler is not None:
+            available_types.append("pref")
+            available_probs.append(self.sample_type_ratio[0])
+
+        if self.sample_type_ratio[1] > 0 and self.progress_sampler is not None:
+            available_types.append("progress")
+            available_probs.append(self.sample_type_ratio[1])
+
+        if not available_types:
+            raise ValueError("No available sample types (all ratios are 0 or samplers are None)")
+
+        # Normalize probabilities
+        total_prob = sum(available_probs)
+        normalized_probs = [p / total_prob for p in available_probs]
+
+        # Select based on weighted random sampling
+        prob = self._local_random.random()
+        cumulative_prob = 0.0
+
+        for sample_type, normalized_prob in zip(available_types, normalized_probs):
+            cumulative_prob += normalized_prob
+            if prob <= cumulative_prob:
+                return sample_type
+
+        # Fallback (should not reach here)
+        return available_types[-1]
+
+    def _select_strategy(self, sample_type: str) -> Optional[DataGenStrat]:
+        """Select a strategy for the given sample type based on strategy ratios."""
+        strategies = []
+        strategy_ratios = []
+
+        if sample_type == "pref":
+            strategy_ratios = self.config.preference_strategy_ratio
+            # Map ratios to strategies: [rewind, suboptimal_same_task, different_task, reverse_progress]
+            if len(strategy_ratios) > 0 and strategy_ratios[0] > 0:
+                strategies.append((DataGenStrat.REWIND, strategy_ratios[0]))
+            if len(strategy_ratios) > 1 and strategy_ratios[1] > 0:
+                strategies.append((DataGenStrat.SUBOPTIMAL, strategy_ratios[1]))
+            if len(strategy_ratios) > 2 and strategy_ratios[2] > 0:
+                strategies.append((DataGenStrat.DIFFERENT_TASK, strategy_ratios[2]))
+            if len(strategy_ratios) > 3 and strategy_ratios[3] > 0:
+                strategies.append((DataGenStrat.REVERSE_PROGRESS, strategy_ratios[3]))
+
+        elif sample_type == "progress":
+            strategy_ratios = self.config.progress_strategy_ratio
+            # Map ratios to strategies: [different_task_instruction, forward_progress, reverse_progress, rewind]
+            if len(strategy_ratios) > 0 and strategy_ratios[0] > 0:
+                strategies.append((DataGenStrat.DIFFERENT_TASK_INSTRUCTION, strategy_ratios[0]))
+            if len(strategy_ratios) > 1 and strategy_ratios[1] > 0:
+                strategies.append((DataGenStrat.FORWARD_PROGRESS, strategy_ratios[1]))
+            if len(strategy_ratios) > 2 and strategy_ratios[2] > 0:
+                strategies.append((DataGenStrat.REVERSE_PROGRESS, strategy_ratios[2]))
+            if len(strategy_ratios) > 3 and strategy_ratios[3] > 0:
+                strategies.append((DataGenStrat.REWIND, strategy_ratios[3]))
+
+        if not strategies:
+            return None
+
+        # Normalize probabilities
+        total_prob = sum(prob for _, prob in strategies)
+        if total_prob == 0:
+            return None
+
+        normalized_strategies = [(strat, prob / total_prob) for strat, prob in strategies]
+
+        # Select based on weighted random sampling
+        prob = self._local_random.random()
+        cumulative_prob = 0.0
+
+        for strategy, normalized_prob in normalized_strategies:
+            cumulative_prob += normalized_prob
+            if prob <= cumulative_prob:
+                return strategy
+
+        # Fallback
+        return strategies[0][0]
+
+    def _normalize_data_source_weights(self):
+        """Normalize data source weights across all available sources."""
+        available_sources = list(self.source_indices.keys())
+
+        if not available_sources:
+            self.normalized_weights = {}
+            return
+
+        # Get weights for available sources
+        weights = {}
+        total_weight = 0.0
+
+        for source in available_sources:
+            weight = self.data_source_weights.get(source, 1.0)
+            weights[source] = weight
+            total_weight += weight
+
+        # Normalize weights
+        if total_weight > 0:
+            self.normalized_weights = {source: weight / total_weight for source, weight in weights.items()}
+        else:
+            # Equal weights fallback
+            equal_weight = 1.0 / len(available_sources)
+            self.normalized_weights = {source: equal_weight for source in available_sources}
+
+    def _select_data_source(self, allowed_sources: Optional[List[str]] = None) -> str:
+        """Select a data source from allowed data sources, using weights if available.
+
+        Args:
+            allowed_sources: Optional list of allowed data source names. If None, uses all available sources.
+
+        Returns:
+            Selected data source name
+        """
+        if allowed_sources is None:
+            available_sources = list(self.source_indices.keys())
+        else:
+            # Filter to only include sources that exist in our source_indices
+            available_sources = [source for source in allowed_sources if source in self.source_indices]
+
+        if not available_sources:
+            raise ValueError("No available data sources")
+
+        if len(available_sources) == 1:
+            return available_sources[0]
+
+        # Use weighted selection if weights are provided
+        if self.normalized_weights:
+            # Re-normalize weights for filtered sources
+            filtered_weights = {}
+            total_weight = 0.0
+            for source in available_sources:
+                weight = self.normalized_weights.get(source, 0.0)
+                if weight > 0:
+                    filtered_weights[source] = weight
+                    total_weight += weight
+
+            if total_weight > 0:
+                # Use re-normalized weights for filtered sources
+                prob = self._local_random.random()
+                cumulative_prob = 0.0
+
+                for source in available_sources:
+                    weight = filtered_weights.get(source, 0.0)
+                    if weight <= 0:
+                        continue
+                    normalized_weight = weight / total_weight
+                    cumulative_prob += normalized_weight
+                    if prob <= cumulative_prob:
+                        return source
+
+                # Fallback: return last source with positive weight
+                for source in reversed(available_sources):
+                    if filtered_weights.get(source, 0.0) > 0:
+                        return source
+
+        # Uniform selection fallback
+        return self._local_random.choice(available_sources)
+
+    def _filter_data_sources_by_strategy(self, strategy: Optional[DataGenStrat]) -> List[str]:
+        """Filter data sources based on strategy requirements.
+
+        Args:
+            strategy: The selected strategy
+
+        Returns:
+            List of viable data source names for the strategy
+        """
+        all_sources = list(self.source_indices.keys())
+
+        if strategy is None:
+            return all_sources
+
+        # Filter based on strategy requirements
+        if strategy == DataGenStrat.SUBOPTIMAL:
+            # SUBOPTIMAL strategy needs data sources with suboptimal/failure trajectories
+            filtered = [source for source in all_sources if is_suboptimal_fail_ds(source)]
+            logger.trace(
+                f"[StrategyFirstDataset] Filtered {len(filtered)}/{len(all_sources)} data sources for SUBOPTIMAL strategy"
+            )
+            return filtered if filtered else all_sources
+
+        elif strategy == DataGenStrat.PAIRED_HUMAN_ROBOT:
+            # PAIRED_HUMAN_ROBOT strategy needs paired data sources
+            filtered = [source for source in all_sources if is_paired_ds(source)]
+            logger.trace(
+                f"[StrategyFirstDataset] Filtered {len(filtered)}/{len(all_sources)} data sources for PAIRED_HUMAN_ROBOT strategy"
+            )
+            return filtered if filtered else all_sources
+
+        # Other strategies (REWIND, DIFFERENT_TASK, REVERSE_PROGRESS, etc.) can work with any data source
+        return all_sources
+
+    def _filter_indices_by_strategy(
+        self, indices: List[int], data_source: str, sample_type: str, strategy: Optional[DataGenStrat]
+    ) -> List[int]:
+        """Filter indices based on strategy requirements.
+
+        For SUBOPTIMAL strategy (preference), filters to only include indices from tasks
+        that have optimal trajectories (unless RoboArena).
+
+        Args:
+            indices: List of trajectory indices to filter
+            data_source: The data source name
+            sample_type: The sample type (pref/progress)
+            strategy: The selected strategy
+
+        Returns:
+            Filtered list of viable indices for the strategy
+        """
+        if strategy is None:
+            return indices
+
+        # For strategies that require successful trajectories only
+        # REWIND: all sample types
+        # REVERSE_PROGRESS: preference and progress samples
+        # FORWARD_PROGRESS: progress samples only
+        requires_successful = (
+            strategy == DataGenStrat.REWIND
+            or (strategy == DataGenStrat.REVERSE_PROGRESS and sample_type in ["pref", "progress"])
+            or (strategy == DataGenStrat.FORWARD_PROGRESS and sample_type == "progress")
+        )
+
+        if requires_successful:
+            indices_set = set(indices)
+            filtered = indices_set & self.successful_indices
+
+            if not filtered:
+                logger.trace(
+                    f"[StrategyFirstDataset] No successful trajectories available for {strategy.value} strategy in source {data_source}"
+                )
+                return []
+
+            logger.trace(
+                f"[StrategyFirstDataset] Filtered {len(filtered)}/{len(indices)} indices for {strategy.value} strategy "
+                f"(keeping only successful trajectories)"
+            )
+            return list(filtered)
+
+        # Check if this is RoboArena or RoboReward (skip task filtering for these datasets)
+        is_roboarena = data_source and "roboarena" in str(data_source).lower()
+        is_roboreward = data_source and "roboreward" in str(data_source).lower()
+
+        # For SUBOPTIMAL strategy (preference), filter to tasks with both optimal and suboptimal trajectories
+        if strategy == DataGenStrat.SUBOPTIMAL and sample_type == "pref":
+            # Skip task filtering for RoboArena and RoboReward (they use partial_success logic)
+            if is_roboarena or is_roboreward:
+                return indices
+
+            if not self.tasks_with_both:
+                # No tasks with both optimal and suboptimal trajectories, return empty list
+                logger.trace(
+                    f"[StrategyFirstDataset] No tasks with both optimal and suboptimal trajectories available for SUBOPTIMAL strategy"
+                )
+                return []
+
+            # Use pre-computed tasks_with_both_indices and intersect with our current indices
+            indices_set = set(indices)
+            filtered = self.tasks_with_both_indices & indices_set
+
+            if not filtered:
+                logger.trace(f"[StrategyFirstDataset] No viable indices after filtering for SUBOPTIMAL strategy")
+                return []
+
+            logger.trace(
+                f"[StrategyFirstDataset] Filtered {len(filtered)}/{len(indices)} indices for SUBOPTIMAL strategy "
+                f"(keeping only tasks with both optimal and suboptimal trajectories)"
+            )
+            return list(filtered)
+
+        # Other strategies don't require task-level filtering
+        return indices
+
+    def _generate_sample_for_type(
+        self, sample_type: str, item: Dict[str, Any], preferred_strategy: Optional[DataGenStrat] = None
+    ):
+        """Generate a sample using the appropriate sampler for the sample type.
+
+        Args:
+            sample_type: The sample type (pref/progress)
+            item: The trajectory item
+            preferred_strategy: Optional strategy to use (if None, sampler will select its own)
+        """
+        data_source = item["data_source"]
+        quality_label = item["quality_label"]
+
+        # Get the appropriate sampler
+        if sample_type == "pref":
+            sampler = self.pref_sampler
+        elif sample_type == "progress":
+            sampler = self.progress_sampler
+        else:
+            return None
+
+        if sampler is None:
+            return None
+
+        # Handle non-successful trajectories: force preference-only
+        if quality_label != "successful" and sample_type != "pref":
+            if self.pref_sampler is not None:
+                logger.trace(f"[StrategyFirstDataset] Non-successful quality detected, switching to preference sampler")
+                return self.pref_sampler._generate_sample(item, preferred_strategy=preferred_strategy)
+            else:
+                return None
+
+        # Handle preference-only data sources
+        if is_preference_only(data_source) and sample_type != "pref":
+            if self.pref_sampler is not None:
+                logger.trace(
+                    f"[StrategyFirstDataset] Preference-only data source detected, switching to preference sampler"
+                )
+                return self.pref_sampler._generate_sample(item, preferred_strategy=preferred_strategy)
+            else:
+                return None
+
+        # Generate sample using the selected sampler with preferred strategy
+        return sampler._generate_sample(item, preferred_strategy=preferred_strategy)
+
+    def _try_generate_sample(
+        self,
+        sample_type: str,
+        filtered_sources: Optional[List[str]] = None,
+        strategy: Optional[DataGenStrat] = None,
+        preferred_strategy: Optional[DataGenStrat] = None,
+        max_attempts: int = 10,
+    ):
+        """Helper method to try generating a sample with retry logic.
+
+        Args:
+            sample_type: The sample type to generate (pref/progress)
+            filtered_sources: Optional list of allowed data sources. If None, uses all sources.
+            strategy: Optional strategy for filtering indices. If None, no index filtering.
+            preferred_strategy: Optional strategy to pass to sampler. If None, sampler selects its own.
+            max_attempts: Maximum number of attempts before giving up.
+
+        Returns:
+            Generated sample if successful, None otherwise.
+        """
+        for attempt in range(max_attempts):
+            selected_source = self._select_data_source(filtered_sources)
+            source_indices = self.source_indices.get(selected_source)
+
+            if not source_indices:
+                logger.trace(f"[StrategyFirstDataset] No indices for source {selected_source}, retrying...")
+                continue
+
+            # Filter indices based on strategy requirements if strategy is provided
+            if strategy is not None:
+                filtered_indices = self._filter_indices_by_strategy(
+                    source_indices, selected_source, sample_type, strategy
+                )
+                if not filtered_indices:
+                    logger.trace(
+                        f"[StrategyFirstDataset] No viable indices after strategy filtering for source {selected_source}, retrying..."
+                    )
+                    continue
+            else:
+                filtered_indices = source_indices
+
+            # Select a trajectory from filtered indices
+            # For progress: trajectory for progress prediction; for preference: chosen trajectory
+            selected_traj_idx = self._local_random.choice(filtered_indices)
+            item = self.dataset[selected_traj_idx]
+
+            traj_id = item["id"]
+            data_source = item["data_source"]
+            quality_label = item["quality_label"]
+
+            strategy_str = strategy.value if strategy and hasattr(strategy, "value") else strategy
+            logger.trace(
+                f"[StrategyFirstDataset] Attempt {attempt + 1}/{max_attempts}: "
+                f"Selected traj ID={traj_id}, source={data_source}, quality={quality_label}, "
+                f"sample_type={sample_type}, strategy={strategy_str}"
+            )
+
+            # Generate sample using the selected sampler with the preferred strategy
+            sample = self._generate_sample_for_type(sample_type, item, preferred_strategy=preferred_strategy)
+            if sample is not None:
+                # Check if the generated sample matches our preferred strategy (if available)
+                generated_strategy = getattr(sample, "data_gen_strategy", None)
+                if generated_strategy and preferred_strategy:
+                    logger.trace(
+                        f"[StrategyFirstDataset] Generated sample with strategy {generated_strategy} "
+                        f"(preferred: {preferred_strategy.value if hasattr(preferred_strategy, 'value') else preferred_strategy})"
+                    )
+                logger.trace(f"[StrategyFirstDataset] Successfully generated {sample_type} sample for ID={traj_id}")
+                return self._set_resample_attempts(sample, attempt + 1)
+
+            logger.trace(f"[StrategyFirstDataset] Sampler returned None for ID={traj_id}, retrying...")
+
+        return None
+
+    def _try_other_samplers(self, failed_sample_type: str):
+        """Try other available samplers when the selected one fails.
+
+        Args:
+            failed_sample_type: The sample type that failed
+
+        Returns:
+            A sample from one of the other samplers, or raises ValueError if all fail
+        """
+        # Get list of available samplers excluding the one that failed
+        available_samplers = []
+        if failed_sample_type != "pref" and self.pref_sampler is not None:
+            available_samplers.append("pref")
+        if failed_sample_type != "progress" and self.progress_sampler is not None:
+            available_samplers.append("progress")
+
+        if not available_samplers:
+            logger.error(f"[StrategyFirstDataset] No other samplers available after {failed_sample_type} failed")
+            raise ValueError(f"Failed to generate {failed_sample_type} sample and no other samplers available")
+
+        # Try each available sampler
+        for fallback_sample_type in available_samplers:
+            logger.trace(f"[StrategyFirstDataset] Trying fallback sampler: {fallback_sample_type}")
+            sample = self._try_generate_sample(
+                sample_type=fallback_sample_type,
+                filtered_sources=None,
+                strategy=None,
+                preferred_strategy=None,
+            )
+            if sample is not None:
+                logger.trace(f"[StrategyFirstDataset] Fallback sampler {fallback_sample_type} succeeded")
+                return sample
+
+            logger.trace(f"[StrategyFirstDataset] Fallback sampler {fallback_sample_type} failed")
+
+        # All fallback samplers failed
+        logger.error(f"[StrategyFirstDataset] All samplers (including fallbacks) failed")
+        raise ValueError(f"Failed to generate {failed_sample_type} sample and all fallback samplers also failed")
+
+    def _generate_without_specific_strategy(self, sample_type: str):
+        """Fallback method to generate sample without specific strategy selection."""
+        sample = self._try_generate_sample(
+            sample_type=sample_type,
+            filtered_sources=None,
+            strategy=None,
+            preferred_strategy=None,
+        )
+        if sample is not None:
+            return sample
+
+        # If this also fails, try other samplers
+        logger.trace(
+            f"[StrategyFirstDataset] _generate_without_specific_strategy failed for {sample_type}, trying other samplers..."
+        )
+        return self._try_other_samplers(sample_type)
+
+    def get_resample_attempt_stats(self):
+        return self._resample_attempt_stats
+
+    def get_resample_dataset_attempt_stats(self):
+        return self._resample_dataset_attempt_stats
diff --git a/robometer/robometer/data/samplers/README.md b/robometer/robometer/data/samplers/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..931e3726fca692ebe8c80603e602547ce29f2d3c
--- /dev/null
+++ b/robometer/robometer/data/samplers/README.md
@@ -0,0 +1,135 @@
+# Sampler Strategies Documentation
+
+This document summarizes the data generation strategies used by each sampler type in the RBM data pipeline.
+
+## Overview
+
+The codebase contains two main sampler types:
+- **PrefSampler**: Generates preference prediction samples
+- **ProgressSampler**: Generates progress prediction samples
+
+Each sampler implements multiple strategies for generating training data, with automatic retry logic and strategy rebalancing on failure.
+
+---
+
+## PrefSampler (Preference Prediction)
+
+The `PrefSampler` creates preference prediction samples with a chosen (preferred) trajectory and a rejected (suboptimal) trajectory.
+
+### Strategies
+
+#### 1. **REWIND**
+- **Description**: Uses the same optimal trajectory for both chosen and rejected, but applies rewind subsampling to the rejected trajectory.
+- **Purpose**: Learn that full trajectories are preferred over truncated/rewound versions.
+- **Implementation**:
+  - `chosen_trajectory`: Original optimal trajectory (forward subsampling)
+  - `rejected_trajectory`: Same trajectory with `subsample_rewind` strategy
+
+#### 2. **SUBOPTIMAL**
+- **Description**: Uses an optimal trajectory as chosen and a suboptimal/failure trajectory from the same task as rejected.
+- **Purpose**: Learn to prefer optimal trajectories over suboptimal ones from the same task.
+- **Conditions**: Only available when suboptimal trajectories exist for the task
+- **Implementation**:
+  - `chosen_trajectory`: Optimal trajectory
+  - `rejected_trajectory`: Suboptimal trajectory from same task (via `_get_same_task_suboptimal`)
+
+#### 3. **DIFFERENT_TASK**
+- **Description**: Uses an optimal trajectory as chosen and a trajectory from a completely different task as rejected.
+- **Purpose**: Learn that trajectories from the same task are preferred over trajectories from different tasks.
+- **Implementation**:
+  - `chosen_trajectory`: Optimal trajectory
+  - `rejected_trajectory`: Trajectory from different task (via `_get_different_video_traj`)
+  - **Note**: Rejected trajectory's `target_progress` is set to `[0.0]` for all timesteps
+
+#### 4. **REVERSE_PROGRESS**
+- **Description**: Uses the same optimal trajectory for both chosen and rejected, but applies reverse uniform sampling to the rejected trajectory.
+- **Purpose**: Learn that forward progress is preferred over reverse progress.
+- **Implementation**:
+  - `chosen_trajectory`: Original optimal trajectory (forward subsampling)
+  - `rejected_trajectory`: Same trajectory with `subsample_reverse` strategy
+
+#### 5. **ROBOARENA_PARTIAL_SUCCESS**
+- **Description**: Uses two trajectories from the same task with different `partial_success` values. The trajectory with higher `partial_success` becomes chosen, and the one with lower `partial_success` becomes rejected.
+- **Purpose**: Learn to prefer trajectories with higher partial success scores (RoboArena-specific).
+- **Conditions**: Only available for RoboArena trajectories (has `partial_success` field and data_source contains "roboarena")
+- **Implementation**:
+  - Finds a different trajectory from same task (via `_get_different_partial_success_traj`)
+  - Swaps trajectories if found trajectory has higher `partial_success`
+  - `chosen_trajectory`: Trajectory with higher `partial_success`
+  - `rejected_trajectory`: Trajectory with lower `partial_success`
+
+### Special Handling
+- **Non-successful trajectories**: If a trajectory has `quality_label != "successful"` (and is not RoboArena), it is automatically used as the rejected trajectory, with an optimal trajectory from the same task as the chosen trajectory.
+
+### Strategy Selection
+- Strategies are selected probabilistically based on `preference_strategy_ratio` configuration
+- Probabilities are rebalanced when strategies fail
+- Strategies are removed after 3 consecutive failures
+- Maximum 10 total attempts per sample generation
+
+---
+
+## ProgressSampler (Progress Prediction)
+
+The `ProgressSampler` creates progress prediction samples from a single trajectory, applying different subsampling strategies to create training data.
+
+### Strategies
+
+#### 1. **DIFFERENT_TASK_INSTRUCTION**
+- **Description**: Uses a trajectory from a different task, but keeps the original task's embeddings and instruction.
+- **Purpose**: Learn that progress should be 0.0 when the trajectory doesn't match the task instruction.
+- **Implementation**:
+  - Gets trajectory from different task (via `_get_different_task_instruction`)
+  - Replaces embeddings with original task's embeddings
+  - Sets `target_progress = [0.0]` for all timesteps
+  - Uses forward subsampling
+
+#### 2. **FORWARD_PROGRESS**
+- **Description**: Samples the same trajectory with forward direction (start < middle < end).
+- **Purpose**: Learn normal forward progress patterns.
+- **Implementation**:
+  - Uses same trajectory with `subsample_forward` strategy
+  - Progress increases from start to end
+
+#### 3. **REVERSE_PROGRESS**
+- **Description**: Samples the same trajectory with reverse direction (end < middle < start).
+- **Purpose**: Learn to handle reverse progress scenarios.
+- **Implementation**:
+  - Uses same trajectory with `subsample_reverse` strategy
+  - Progress decreases from start to end
+
+#### 4. **REWIND**
+- **Description**: Samples the same trajectory with rewind direction (start < end < middle).
+- **Purpose**: Learn to handle non-monotonic progress patterns.
+- **Implementation**:
+  - Uses same trajectory with `subsample_rewind` strategy
+  - Progress pattern: increases, then decreases
+
+### Strategy Selection
+- Strategies are selected probabilistically based on `progress_strategy_ratio` configuration
+- Probabilities are rebalanced when strategies fail
+- Failed strategies are immediately removed (no retry count threshold)
+- Maximum 10 total attempts per sample generation
+
+---
+
+## Common Features
+
+### Retry Logic
+All samplers implement retry logic with:
+- Maximum attempt limits (typically 10 attempts)
+- Strategy-specific retry counts (3-4 attempts per strategy)
+- Automatic strategy removal after consecutive failures
+- Probability rebalancing when strategies are removed
+
+### Subsample Strategies
+Common subsampling strategies used across samplers:
+- `subsample_forward`: Normal forward sampling (start → end)
+- `subsample_reverse`: Reverse sampling (end → start)
+- `subsample_rewind`: Rewind sampling (start → end → start)
+
+### Data Source Filtering
+- Strategies may be filtered or boosted based on data source categories:
+  - **Failure category**: Boosts SUBOPTIMAL strategy probability by 2x
+  - **Paired category**: Boosts PAIRED_HUMAN_ROBOT strategy probability by 2x
+  - **RoboArena**: Special handling for `partial_success` field
diff --git a/robometer/robometer/data/samplers/__init__.py b/robometer/robometer/data/samplers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3df21bb034284c945610217af457bbc6e0863d33
--- /dev/null
+++ b/robometer/robometer/data/samplers/__init__.py
@@ -0,0 +1,19 @@
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.data.samplers.pref import PrefSampler
+from robometer.data.samplers.progress import ProgressSampler
+from robometer.data.samplers.eval.confusion_matrix import ConfusionMatrixSampler
+from robometer.data.samplers.eval.progress_policy_ranking import ProgressPolicyRankingSampler
+from robometer.data.samplers.eval.reward_alignment import RewardAlignmentSampler
+from robometer.data.samplers.eval.quality_preference import QualityPreferenceSampler
+from robometer.data.samplers.eval.roboarena_quality_preference import RoboArenaQualityPreferenceSampler
+
+__all__ = [
+    "RBMBaseSampler",
+    "PrefSampler",
+    "ProgressSampler",
+    "ConfusionMatrixSampler",
+    "ProgressPolicyRankingSampler",
+    "RewardAlignmentSampler",
+    "QualityPreferenceSampler",
+    "RoboArenaQualityPreferenceSampler",
+]
diff --git a/robometer/robometer/data/samplers/base.py b/robometer/robometer/data/samplers/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..b193396a7f29059560280d86137982fafcd60419
--- /dev/null
+++ b/robometer/robometer/data/samplers/base.py
@@ -0,0 +1,782 @@
+#!/usr/bin/env python3
+from typing import Optional, Dict, Any, List, Set, Tuple, Union
+
+import numpy as np
+import random
+import torch
+from random import Random
+from datasets import Dataset
+
+from robometer.configs.experiment_configs import DataConfig
+from robometer.data.datasets.helpers import (
+    load_frames_from_npz,
+    get_segment_indices_with_middle,
+    compute_progress_from_segment,
+    pad_trajectory_to_max_frames_torch,
+    pad_trajectory_to_max_frames_np,
+    compute_success_labels,
+    create_trajectory_from_dict,
+    load_embeddings_from_path,
+    linspace_subsample_frames,
+    convert_continuous_to_discrete_bins,
+)
+from robometer.data.dataset_types import Trajectory
+from robometer.utils.logger import get_logger
+from robometer.data.dataset_category import is_preference_only_ds
+
+logger = get_logger()
+
+
+class RBMBaseSampler:
+    """Base sampler class that provides trajectory retrieval functions for generating samples."""
+
+    def __init__(
+        self,
+        config: DataConfig,
+        dataset: Dataset,
+        combined_indices: Dict[str, Any],
+        dataset_success_cutoff_map: Optional[Dict[str, float]] = None,
+        verbose: bool = True,
+        random_seed: int = 42,
+        pad_frames: bool = True,
+    ):
+        """Initialize sampler with dataset and indices.
+
+        Args:
+            config: Configuration object
+            dataset: The loaded dataset
+            combined_indices: Dictionary of combined indices from dataset loading
+            dataset_success_cutoff_map: Dictionary mapping dataset names to success cutoff percentages
+            verbose: Verbose flag
+            random_seed: Random seed for deterministic sampling. Creates a local Random instance to avoid affecting global random state.
+        """
+        self.config = config
+        self.dataset = dataset
+        self.verbose = verbose
+        self.dataset_success_cutoff_map = dataset_success_cutoff_map or {}
+        self._local_random = Random(random_seed)
+        self.pad_frames = pad_frames
+        self._cached_ids = self.dataset["id"]
+        self._cached_is_robot = self.dataset["is_robot"]
+
+        # Build indices from combined_indices
+        self._build_indices(combined_indices)
+
+    def _build_indices(self, combined_indices):
+        """Build all index mappings from combined_indices.
+
+        Args:
+            combined_indices: Dictionary of combined indices from dataset loading
+        """
+        # Initialize index mappings from the loaded indices
+        self.robot_trajectories = combined_indices["robot_trajectories"]
+        self.human_trajectories = combined_indices["human_trajectories"]
+        self.optimal_by_task = combined_indices["optimal_by_task"]
+        self.suboptimal_by_task = combined_indices["suboptimal_by_task"]
+        self.quality_indices = combined_indices["quality_indices"]
+        self.task_indices = combined_indices["task_indices"]
+        self.source_indices = combined_indices["source_indices"]
+        self.partial_success_indices = combined_indices["partial_success_indices"]
+        self.paired_human_robot_by_task = combined_indices["paired_human_robot_by_task"]
+        self.tasks_with_multiple_quality_labels = combined_indices["tasks_with_multiple_quality_labels"]
+
+        # Build mapping from data source -> available task instructions
+        self._build_tasks_by_data_source()
+
+    def _build_tasks_by_data_source(self):
+        """Cache mapping from data_source to available task instructions."""
+        self.tasks_by_data_source: Dict[str, List[str]] = {}
+
+        all_tasks = self.dataset["task"]
+        all_sources = self.dataset["data_source"]
+
+        source_to_tasks: Dict[str, Set[str]] = {}
+        for task, source in zip(all_tasks, all_sources):
+            if task is None or source is None:
+                continue
+            if source not in source_to_tasks:
+                source_to_tasks[source] = set()
+            source_to_tasks[source].add(task)
+
+        self.tasks_by_data_source = {source: list(tasks) for source, tasks in source_to_tasks.items()}
+
+    def _generate_sample(self, item):
+        """Generate a sample from an item.
+
+        This method should be overridden by subclasses to implement their specific
+        sample generation logic.
+
+        Args:
+            item: An item from the dataset (typically a trajectory dict)
+
+        Returns:
+            A sample object (e.g., PreferenceSample, ProgressSample)
+        """
+        raise NotImplementedError("Subclasses must implement _generate_sample")
+
+    def _get_same_task_optimal(self, ref_traj: dict) -> dict | None:
+        """Get optimal trajectory from same task (different from ref).
+
+        Args:
+            ref_traj: Reference trajectory
+
+        Returns:
+            Same task optimal trajectory dict or None if not available
+        """
+        task_name = ref_traj["task"]
+        same_task_optimal_indices = self.optimal_by_task.get(task_name, [])
+        if not same_task_optimal_indices:
+            logger.trace(f"[BASE SAMPLER] _get_same_task_optimal: No optimal indices for task '{task_name}'")
+            return None
+
+        # Use cached IDs to check without loading full trajectories
+        chosen_id = ref_traj["id"]
+        random_idx = random.choice(same_task_optimal_indices)
+
+        # Retry if the selected trajectory has the same ID as ref
+        max_retries = min(10, len(same_task_optimal_indices))
+        retries = 0
+        while self._cached_ids[random_idx] == chosen_id and retries < max_retries:
+            random_idx = random.choice(same_task_optimal_indices)
+            retries += 1
+
+        # If still matches after retries, fall back to filtering
+        if self._cached_ids[random_idx] == chosen_id:
+            filtered_indices = [idx for idx in same_task_optimal_indices if self._cached_ids[idx] != chosen_id]
+            if filtered_indices:
+                random_idx = random.choice(filtered_indices)
+            else:
+                # No other trajectories available
+                logger.trace(
+                    f"[BASE SAMPLER] _get_same_task_optimal: All trajectories have same ID '{chosen_id}' for task '{task_name}'"
+                )
+                return None
+
+        result = self.dataset[random_idx]
+        logger.trace(
+            f"[BASE SAMPLER] _get_same_task_optimal: Found trajectory {result.get('id', 'unknown')} for task '{task_name}'"
+        )
+        return result
+
+    def _get_same_task_suboptimal(self, ref_traj: dict) -> dict | None:
+        """Get suboptimal trajectory from same task.
+
+        For trajectories with partial_success, uses partial_success logic instead of quality_label logic.
+
+        Args:
+            ref_traj: Reference trajectory
+
+        Returns:
+            Suboptimal trajectory dict or None if not available
+        """
+        # Check if this trajectory uses partial_success
+        use_partial_success = ref_traj.get("partial_success") is not None
+
+        if use_partial_success:
+            # For trajectories with partial_success, use partial_success logic
+            return self._get_different_partial_success_traj(ref_traj)
+
+        # For trajectories without partial_success, use the standard suboptimal logic
+        task_name = ref_traj["task"]
+        same_task_suboptimal_indices = self.suboptimal_by_task.get(task_name, [])
+        if not same_task_suboptimal_indices:
+            logger.trace(f"[BASE SAMPLER] _get_same_task_suboptimal: No suboptimal indices for task '{task_name}'")
+            return None
+
+        # Use cached IDs to check without loading full trajectories
+        chosen_id = ref_traj["id"]
+        random_idx = random.choice(same_task_suboptimal_indices)
+
+        # Retry if the selected trajectory has the same ID as ref
+        max_retries = min(10, len(same_task_suboptimal_indices))
+        retries = 0
+        while self._cached_ids[random_idx] == chosen_id and retries < max_retries:
+            random_idx = random.choice(same_task_suboptimal_indices)
+            retries += 1
+
+        # If still matches after retries, fall back to filtering
+        if self._cached_ids[random_idx] == chosen_id:
+            filtered_indices = [idx for idx in same_task_suboptimal_indices if self._cached_ids[idx] != chosen_id]
+            if filtered_indices:
+                random_idx = random.choice(filtered_indices)
+            else:
+                # No other trajectories available
+                logger.trace(
+                    f"[BASE SAMPLER] _get_same_task_suboptimal: All trajectories have same ID '{chosen_id}' for task '{task_name}'"
+                )
+                return None
+
+        result = self.dataset[random_idx]
+        logger.trace(
+            f"[BASE SAMPLER] _get_same_task_suboptimal: Found trajectory {result.get('id', 'unknown')} for task '{task_name}'"
+        )
+        return result
+
+    def _get_different_video_traj(self, ref_traj: dict) -> dict | None:
+        """Get trajectory from different task.
+
+        Args:
+            ref_traj: Reference trajectory
+
+        Returns:
+            Different task trajectory dict or None if not available
+        """
+        same_source_prob = self.config.traj_same_source_prob
+        data_source = ref_traj.get("data_source")
+        other_tasks = []
+
+        if data_source and data_source in self.tasks_by_data_source and random.random() < same_source_prob:
+            other_tasks = [task for task in self.tasks_by_data_source[data_source] if task != ref_traj["task"]]
+
+        if not other_tasks:
+            other_tasks = [task for task in self.optimal_by_task.keys() if task != ref_traj["task"]]
+
+        if not other_tasks:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_video_traj: No other tasks available (ref task: '{ref_traj['task']}')"
+            )
+            return None
+
+        # Try up to 2 times to find a valid task
+        max_retries = 2
+        other_task_indices = None
+        other_task = None
+
+        for attempt in range(max_retries):
+            other_task = random.choice(other_tasks)
+            if other_task not in self.optimal_by_task:
+                logger.trace(
+                    f"[BASE SAMPLER] _get_different_video_traj: Attempt {attempt + 1}/{max_retries}: Task '{other_task}' not found in optimal_by_task"
+                )
+                continue
+
+            other_task_indices = self.optimal_by_task[other_task]
+            if not other_task_indices:
+                logger.trace(
+                    f"[BASE SAMPLER] _get_different_video_traj: Attempt {attempt + 1}/{max_retries}: Task '{other_task}' has no optimal indices"
+                )
+                continue
+
+            # Found a valid task with indices
+            break
+
+        if other_task_indices is None or not other_task_indices:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_video_traj: Failed to find valid task after {max_retries} attempts"
+            )
+            return None
+
+        other_idx = random.choice(other_task_indices)
+        result = self.dataset[other_idx]
+        logger.trace(
+            f"[BASE SAMPLER] _get_different_video_traj: Found trajectory {result.get('id', 'unknown')} from task '{other_task}'"
+        )
+        return result
+
+    def _get_different_task_instruction(self, ref_traj: dict) -> dict | None:
+        """Get the same trajectory but with a different task instruction.
+
+        Args:
+            ref_traj: Reference trajectory
+
+        Returns:
+            Trajectory dict with different task instruction or None if not available
+        """
+        same_source_prob = self.config.traj_same_source_prob
+        data_source = ref_traj.get("data_source")
+        candidate_tasks = []
+
+        if data_source and data_source in self.tasks_by_data_source and random.random() < same_source_prob:
+            candidate_tasks = [task for task in self.tasks_by_data_source[data_source] if task != ref_traj["task"]]
+
+        if not candidate_tasks:
+            candidate_tasks = [task for task in self.optimal_by_task.keys() if task != ref_traj["task"]]
+
+        if not candidate_tasks:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_task_instruction: No candidate tasks available (ref task: '{ref_traj['task']}')"
+            )
+            return None
+
+        other_task = random.choice(candidate_tasks)
+
+        # Get embeddings_path and lang_vector from a random trajectory with the other_task
+        other_task_indices = self.optimal_by_task.get(other_task, [])
+        if not other_task_indices:
+            logger.trace(f"[BASE SAMPLER] _get_different_task_instruction: Task '{other_task}' has no optimal indices")
+            return None
+
+        other_task_idx = random.choice(other_task_indices)
+        other_task_traj = self.dataset[other_task_idx]
+
+        # Create a copy of the trajectory with the task changed
+        # Use embeddings_path and lang_vector from the other_task trajectory
+        new_traj = ref_traj.copy()
+        new_traj["task"] = other_task
+        # Get embeddings_path and lang_vector from a random trajectory with the other_task
+        if "embeddings_path" in other_task_traj:
+            new_traj["embeddings_path"] = other_task_traj["embeddings_path"]
+        if "lang_vector" in other_task_traj:
+            new_traj["lang_vector"] = other_task_traj["lang_vector"]
+        return new_traj
+
+    def _get_paired_human_robot_traj(self, ref_traj: dict) -> dict | None:
+        """Get paired human/robot trajectory for the same task.
+
+        Given a reference trajectory, if it's a robot trajectory, returns a human trajectory
+        from the same task. If it's a human trajectory, returns a robot trajectory from the
+        same task.
+
+        Args:
+            ref_traj: Reference trajectory (can be robot or human)
+
+        Returns:
+            Paired trajectory dict (opposite type) or None if not available
+        """
+        task = ref_traj["task"]
+        is_robot = ref_traj.get("is_robot", True)
+
+        if task not in self.paired_human_robot_by_task:
+            logger.trace(
+                f"[BASE SAMPLER] _get_paired_human_robot_traj: Task '{task}' not in paired_human_robot_by_task"
+            )
+            return None
+
+        task_pairs = self.paired_human_robot_by_task[task]
+
+        # Get opposite type
+        opposite_key = "human" if is_robot else "robot"
+        opposite_indices = task_pairs.get(opposite_key, [])
+
+        if not opposite_indices:
+            logger.trace(f"[BASE SAMPLER] _get_paired_human_robot_traj: No {opposite_key} indices for task '{task}'")
+            return None
+
+        # Sample a paired trajectory and verify it's different from reference
+        chosen_id = ref_traj["id"]
+        available_indices = opposite_indices.copy()
+        paired_traj = None
+
+        # Add retry limit to prevent infinite loops
+        max_retries = min(len(available_indices), 10)
+        retries = 0
+
+        logger.trace(
+            f"[BASE SAMPLER] _get_paired_human_robot_traj: Looking for {opposite_key} trajectory (chosen_id: {chosen_id}, available: {len(available_indices)})"
+        )
+
+        while (paired_traj is None or paired_traj.get("id") == chosen_id) and retries < max_retries:
+            retries += 1
+
+            if not available_indices:
+                logger.trace(
+                    f"[BASE SAMPLER] _get_paired_human_robot_traj: No more available indices after {retries} retries"
+                )
+                return None
+
+            paired_idx = random.choice(available_indices)
+            paired_traj = self.dataset[paired_idx]
+
+            # If it matches, remove this index and try again
+            if paired_traj.get("id") == chosen_id:
+                available_indices = [idx for idx in available_indices if idx != paired_idx]
+                paired_traj = None
+                continue
+
+        # If we exhausted retries without finding a valid trajectory, return None
+        if paired_traj is None or paired_traj.get("id") == chosen_id:
+            logger.trace(
+                f"[BASE SAMPLER] _get_paired_human_robot_traj: Failed to find valid paired trajectory after {max_retries} retries"
+            )
+            return None
+
+        logger.trace(
+            f"[BASE SAMPLER] _get_paired_human_robot_traj: Found paired trajectory {paired_traj.get('id', 'unknown')} on retry {retries}"
+        )
+        return paired_traj
+
+    def _get_different_partial_success_traj(self, ref_traj: dict) -> dict | None:
+        """Get trajectory from same task with different partial_success.
+
+        Finds trajectories with either higher or lower partial_success than the reference,
+        using absolute difference for threshold checking.
+
+        Args:
+            ref_traj: Reference trajectory
+
+        Returns:
+            Trajectory dict with different partial_success from same task or None if not available
+        """
+        task_name = ref_traj["task"]
+        ref_partial_success = ref_traj.get("partial_success")
+
+        # Check if partial_success is available
+        if ref_partial_success is None:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_partial_success_traj: No partial_success for trajectory {ref_traj.get('id', 'unknown')}"
+            )
+            return None
+
+        # Get minimum threshold from config
+        min_threshold = getattr(self.config, "partial_success_threshold", 0.2)
+
+        # Get all trajectories from the same task
+        same_task_indices = self.task_indices.get(task_name, [])
+        if not same_task_indices:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_partial_success_traj: No trajectories found for task '{task_name}'"
+            )
+            return None
+
+        # Filter to trajectories with different partial_success that meet the threshold requirement
+        # Uses absolute difference to allow both higher and lower partial_success
+        chosen_id = ref_traj["id"]
+        candidate_indices = []
+
+        for idx in same_task_indices:
+            # Skip if same trajectory
+            if self._cached_ids[idx] == chosen_id:
+                continue
+
+            # Get partial_success for this trajectory
+            traj_dict = self.dataset[idx]
+            traj_partial_success = traj_dict.get("partial_success", None)
+
+            if traj_partial_success is None:
+                logger.trace(
+                    f"[BASE SAMPLER] _get_different_partial_success_traj: No partial_success for trajectory {traj_dict.get('id', 'unknown')}, task '{task_name}'"
+                )
+                continue
+
+            # Include if partial_success differs from reference by at least the threshold (using abs)
+            partial_success_diff = abs(ref_partial_success - traj_partial_success)
+            if partial_success_diff >= min_threshold:
+                candidate_indices.append(idx)
+
+        if not candidate_indices:
+            logger.trace(
+                f"[BASE SAMPLER] _get_different_partial_success_traj: No trajectories with different partial_success (threshold: {min_threshold}) for task '{task_name}' (ref: {ref_partial_success})"
+            )
+            return None
+
+        # Randomly select from candidates
+        selected_idx = random.choice(candidate_indices)
+        result = self.dataset[selected_idx]
+        result_partial_success = result.get("partial_success")
+        # If ref_partial_success is 1.0, direction is always "lower" since 1.0 is the maximum
+        if ref_partial_success == 1.0:
+            direction = "lower"
+        else:
+            direction = "higher" if result_partial_success > ref_partial_success else "lower"
+        logger.trace(
+            f"[BASE SAMPLER] _get_different_partial_success_traj: Found trajectory {result.get('id', 'unknown')} with partial_success {result_partial_success} ({direction} than {ref_partial_success}, abs diff: {abs(ref_partial_success - result_partial_success):.3f}, threshold: {min_threshold})"
+        )
+        return result
+
+    def _get_subsample_indices(
+        self, data, direction: str = "bidirectional", max_frames: int = None
+    ) -> Optional[Tuple[int, int, int]]:
+        """Get start, middle, and end indices for subsample strategy.
+
+        Samples three random frames from the trajectory. The relationship between indices
+        follows three main scenarios:
+        1. start < middle < end: forward progress - normal forward progression through trajectory
+        2. start < end < middle: rewind progress - forward from start to end, then continues to middle (simulating rewind/backtrack)
+        3. end < middle < start: reverse progress - backward from start through middle to end (full backward traversal)
+
+        Args:
+            data: Trajectory data (frames or embeddings) to sample from
+            direction: Sampling direction - "forward" (start < middle < end),
+                      "reverse" (end < middle < start),
+                      "rewind" (start < end < middle),
+                      or "bidirectional" (any of the 3 orderings)
+            max_frames: Maximum number of frames to subsample. If 1, returns only start. If 2, returns start and end.
+
+        Returns:
+            Tuple of (start_idx, middle_idx, end_idx), or None if insufficient frames
+            For max_frames == 1: returns (start_idx, None, None)
+            For max_frames == 2: returns (start_idx, None, end_idx)
+        """
+        num_frames_total = len(data) if hasattr(data, "__len__") else data.shape[0]
+
+        # Handle edge cases for max_frames == 1 or 2
+        if max_frames == 1:
+            # Randomly sample 1 frame
+            random_idx = random.randint(0, num_frames_total - 1)
+            logger.trace(f"[BASE SAMPLER] _get_subsample_indices: max_frames=1, randomly sampled idx={random_idx}")
+            return (random_idx, None, None)
+
+        if max_frames == 2:
+            # Sample 2 frames: either forward (start < end) or reverse (end < start)
+            # No rewind possible with only 2 frames
+            if direction == "reverse":
+                # Reverse: sample end first, then start (end < start)
+                end_idx = random.randint(0, num_frames_total - 2)
+                start_idx = random.randint(end_idx + 1, num_frames_total - 1)
+            else:
+                # Forward: sample start first, then end (start < end)
+                start_idx = random.randint(0, num_frames_total - 2)
+                end_idx = random.randint(start_idx + 1, num_frames_total - 1)
+            logger.trace(
+                f"[BASE SAMPLER] _get_subsample_indices: max_frames=2, start_idx={start_idx}, end_idx={end_idx}, direction={direction}"
+            )
+            return (start_idx, None, end_idx)
+
+        if num_frames_total < 3:
+            logger.trace(f"[BASE SAMPLER] _get_subsample_indices: Not enough frames ({num_frames_total})")
+            return None
+
+        # Sample three random distinct frames
+        frame_indices = sorted(random.sample(range(num_frames_total), 3))
+        frame1_idx, frame2_idx, frame3_idx = frame_indices
+
+        # Determine start, middle, and end based on direction
+        # We only care about 3 cases:
+        # 1. start < middle < end: forward progress
+        # 2. start < end < middle: rewind progress
+        # 3. end < middle < start: reverse progress
+
+        if direction == "forward":
+            # Case 1: start < middle < end
+            start_idx = frame1_idx
+            middle_idx = frame2_idx
+            end_idx = frame3_idx
+        elif direction == "reverse":
+            # Case 3: end < middle < start
+            end_idx = frame1_idx
+            middle_idx = frame2_idx
+            start_idx = frame3_idx
+        elif direction == "rewind":
+            # Case 2: start < end < middle
+            start_idx = frame1_idx
+            end_idx = frame2_idx
+            middle_idx = frame3_idx
+        else:  # bidirectional (default)
+            # Randomly choose from the 3 cases
+            pattern = random.choice([1, 2, 3])
+            if pattern == 1:  # start < middle < end: forward progress
+                start_idx = frame1_idx
+                middle_idx = frame2_idx
+                end_idx = frame3_idx
+            elif pattern == 2:  # start < end < middle: rewind progress
+                start_idx = frame1_idx
+                end_idx = frame2_idx
+                middle_idx = frame3_idx
+            else:  # pattern == 3: end < middle < start: reverse progress
+                end_idx = frame1_idx
+                middle_idx = frame2_idx
+                start_idx = frame3_idx
+
+        logger.trace(
+            f"[BASE SAMPLER] _get_subsample_indices: Selected indices start={start_idx}, middle={middle_idx}, end={end_idx} "
+            f"from {num_frames_total} total frames (direction: {direction})"
+        )
+        return start_idx, middle_idx, end_idx
+
+    def _get_traj_from_data(
+        self,
+        traj: dict | Trajectory,
+        subsample_strategy: str | None = None,
+        frame_indices: List[int] | None = None,
+        metadata: Dict[str, Any] | None = None,
+        pad_frames: bool = True,
+    ) -> Trajectory:
+        """Load, subsample, and optionally pad trajectory data and create a Trajectory object.
+
+        Args:
+            traj: Trajectory dict or Trajectory object
+            subsample_strategy: Optional strategy for subsampling ("subsample_forward", "subsample_reverse", "subsample_rewind", or None for default/bidirectional). Ignored if frame_indices is provided.
+            frame_indices: Optional list of specific frame indices to use. If provided, subsample_strategy is ignored.
+            metadata: Optional metadata dict to merge into trajectory metadata.
+            pad_frames: Whether to pad the trajectory data to max_frames.
+
+        Returns:
+            Trajectory object with loaded and subsampled data (padded)
+        """
+        # Initialize variables
+        frames = None
+        video_embeddings = None
+        text_embedding = None
+        data = None
+
+        if isinstance(traj, Trajectory):
+            # If already a Trajectory, just return it
+            return traj
+
+        # Load from dict
+        # Check if text_embedding is already provided in the dict (for samplers that need to override it)
+        if "text_embedding" in traj and traj["text_embedding"] is not None:
+            text_embedding = traj["text_embedding"]
+
+        if self.config.load_embeddings and traj.get("embeddings_path"):
+            embeddings = load_embeddings_from_path(traj["embeddings_path"])
+            video_embeddings = embeddings["video_embeddings"]
+            # Only use loaded text_embedding if not already provided in dict
+            if text_embedding is None:
+                text_embedding = embeddings["text_embedding"]
+            data = video_embeddings
+        else:
+            if isinstance(traj["frames"], str):
+                frames = load_frames_from_npz(traj["frames"])
+            else:
+                frames = traj["frames"]
+            data = frames
+
+        # Get total frames for progress computation
+        if hasattr(data, "shape"):
+            num_frames_total = data.shape[0]
+        else:
+            num_frames_total = len(data)
+
+        ds_key = traj["data_source"]
+        success_cutoff = self.dataset_success_cutoff_map.get(ds_key, self.config.max_success)
+
+        # Determine which indices to use (construct indices first, then subsample uniformly)
+        if frame_indices is not None:
+            # Use provided frame indices directly
+            indices = frame_indices
+        elif subsample_strategy is not None:
+            # Use subsampling strategy
+            # Get subsample indices (handles edge cases for max_frames == 1 or 2)
+            if subsample_strategy == "subsample_forward":
+                strategy_indices = self._get_subsample_indices(
+                    data, direction="forward", max_frames=self.config.max_frames
+                )
+            elif subsample_strategy == "subsample_reverse":
+                strategy_indices = self._get_subsample_indices(
+                    data, direction="reverse", max_frames=self.config.max_frames
+                )
+            elif subsample_strategy == "subsample_rewind":
+                strategy_indices = self._get_subsample_indices(
+                    data, direction="rewind", max_frames=self.config.max_frames
+                )
+            else:
+                strategy_indices = self._get_subsample_indices(
+                    data, direction="bidirectional", max_frames=self.config.max_frames
+                )
+
+            if strategy_indices is None:
+                logger.trace("[BASE SAMPLER] _get_traj_from_data: Failed to get uniform sample indices")
+                return None
+
+            start_idx, middle_idx, end_idx = strategy_indices
+
+            logger.trace(
+                f"[BASE SAMPLER] _get_traj_from_data: Subsampling trajectory with strategy: {subsample_strategy}, start_idx: {start_idx}, middle_idx: {middle_idx}, end_idx: {end_idx}"
+            )
+
+            # Use middle_idx only for rewind strategy (requires at least 3 frames)
+            use_middle = subsample_strategy == "subsample_rewind" and middle_idx is not None and num_frames_total >= 3
+
+            # Use get_segment_indices_with_middle to construct indices
+            indices = get_segment_indices_with_middle(
+                num_frames_total=num_frames_total,
+                start_idx=start_idx,
+                end_idx=end_idx,
+                middle_idx=middle_idx if use_middle else None,
+                max_frames=self.config.max_frames,
+            )
+        else:
+            # No subsampling strategy or indices provided - use all frames
+            indices = list(range(num_frames_total))
+
+        # Extract data using indices
+        subsampled = data[indices]
+
+        # Get partial_success early to pass to compute_progress_from_segment
+        partial_success = traj.get("partial_success")
+
+        # Compute progress
+        target_progress = compute_progress_from_segment(
+            num_frames_total=num_frames_total,
+            frame_indices=indices,
+            progress_pred_type=self.config.progress_pred_type,
+            success_cutoff=success_cutoff,
+            partial_success=partial_success,
+        )
+
+        # Subsample uniformly if needed (if we have more frames than max_frames)
+        current_frame_count = len(subsampled) if hasattr(subsampled, "__len__") else subsampled.shape[0]
+        if current_frame_count > self.config.max_frames:
+            subsampled, frame_indices_subsample = linspace_subsample_frames(subsampled, self.config.max_frames)
+            # Update indices and target_progress
+            if target_progress and len(target_progress) == current_frame_count:
+                target_progress = [target_progress[idx] for idx in frame_indices_subsample]
+            indices = [indices[idx] for idx in frame_indices_subsample] if isinstance(indices, list) else indices
+
+        # Pad if needed
+        if target_progress and pad_frames:
+            if self.config.load_embeddings:
+                subsampled, target_progress = pad_trajectory_to_max_frames_torch(
+                    subsampled, target_progress, self.config.max_frames
+                )
+            else:
+                subsampled, target_progress = pad_trajectory_to_max_frames_np(
+                    subsampled, target_progress, self.config.max_frames
+                )
+
+        # Create predict_last_frame_mask: mark the last frame if partial_success < 1.0
+        # If predict_last_frame_partial_progress is True and partial_success < 1.0 and the last original frame is in the subsampled indices,
+        # mark all positions where it appears with 1.0, all others 0.0. Otherwise, all 1.0s.
+        final_frame_count = len(subsampled)
+        predict_last_frame_mask = [1.0] * final_frame_count  # Default: all 1.0s (no masking)
+
+        if self.config.predict_last_frame_partial_progress and partial_success is not None:
+            if partial_success == 1.0 and not is_preference_only_ds(traj["data_source"]):
+                pass
+            else:
+                last_original_frame_idx = num_frames_total - 1
+                if isinstance(indices, list) and last_original_frame_idx in indices:
+                    # Find all positions where the last frame index appears
+                    last_frame_positions = [
+                        i for i, idx in enumerate(indices) if idx == last_original_frame_idx and i < final_frame_count
+                    ]
+                    if last_frame_positions:
+                        # Mark all positions where the last frame appears with 1.0, all others 0.0
+                        predict_last_frame_mask = [0.0] * final_frame_count
+                        for pos in last_frame_positions:
+                            predict_last_frame_mask[pos] = 1.0
+                else:
+                    predict_last_frame_mask = [0.0] * final_frame_count
+
+        # Update frames_shape
+        frames_shape = subsampled.shape if hasattr(subsampled, "shape") else tuple()
+
+        # Set frames or video_embeddings
+        if self.config.load_embeddings:
+            video_embeddings = subsampled
+        else:
+            frames = subsampled
+
+        # Compute success labels
+        success_label = compute_success_labels(
+            target_progress=target_progress,
+            data_source=traj["data_source"],
+            dataset_success_percent=self.dataset_success_cutoff_map,
+            max_success=self.config.max_success,
+            quality_label=traj.get("quality_label"),
+        )
+
+        # Convert partial_success and target_progress to discrete bins if in discrete mode
+        if self.config.progress_loss_type.lower() == "discrete":
+            if partial_success is not None:
+                partial_success = convert_continuous_to_discrete_bins(
+                    [partial_success], self.config.progress_discrete_bins
+                )[0]
+            target_progress = convert_continuous_to_discrete_bins(target_progress, self.config.progress_discrete_bins)
+
+        trajectory = create_trajectory_from_dict(
+            traj,
+            overrides={
+                "frames": frames,
+                "frames_shape": frames_shape,
+                "video_embeddings": video_embeddings,
+                "text_embedding": text_embedding,
+                "target_progress": target_progress,
+                "success_label": success_label,
+                "partial_success": partial_success,
+                "predict_last_frame_mask": predict_last_frame_mask,
+                "metadata": metadata,
+            },
+        )
+        return trajectory
diff --git a/robometer/robometer/data/samplers/eval/base_pref.py b/robometer/robometer/data/samplers/eval/base_pref.py
new file mode 100644
index 0000000000000000000000000000000000000000..4b8e66ac16c1a62c5e9a4ffa0eb92c2ec5d82fa2
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/base_pref.py
@@ -0,0 +1,73 @@
+from typing import Dict, Any
+
+import numpy as np
+
+from robometer.data.dataset_types import PreferenceSample, Trajectory
+from robometer.data.samplers.base import RBMBaseSampler
+
+
+class BaseQualityPreferenceSampler(RBMBaseSampler):
+    """Base class for quality preference samplers.
+
+    Subclasses should implement `_generate_all_sample_indices` to define how
+    trajectories are paired. This base class provides the common `_generate_sample_from_indices`
+    method that loads and processes the trajectories.
+    """
+
+    def _generate_sample_from_indices(self, sample_idx_info: Dict[str, Any]) -> PreferenceSample:
+        """Generate a single sample from stored indices."""
+        chosen_idx = sample_idx_info["chosen_traj_idx"]
+        rejected_idx = sample_idx_info["rejected_traj_idx"]
+
+        # Get the trajectories
+        chosen_traj = self.dataset[chosen_idx]
+        rejected_traj = self.dataset[rejected_idx]
+
+        chosen_metadata = {
+            "quality_label": chosen_traj["quality_label"],
+            "data_source": chosen_traj["data_source"],
+            "task": chosen_traj["task"],
+            "id": chosen_traj["id"],
+            "video_path": chosen_traj["frames"],
+        }
+        # Add partial_success if available
+        if chosen_traj.get("partial_success") is not None:
+            chosen_metadata["partial_success"] = chosen_traj.get("partial_success")
+
+        chosen_trajectory = self._get_traj_from_data(
+            traj=chosen_traj,
+            metadata=chosen_metadata,
+        )
+
+        rejected_metadata = {
+            "quality_label": rejected_traj["quality_label"],
+            "data_source": rejected_traj["data_source"],
+            "task": rejected_traj["task"],
+            "id": rejected_traj["id"],
+            "video_path": rejected_traj["frames"],
+        }
+        # Add partial_success if available
+        if rejected_traj.get("partial_success") is not None:
+            rejected_metadata["partial_success"] = rejected_traj.get("partial_success")
+
+        rejected_trajectory = self._get_traj_from_data(
+            traj=rejected_traj,
+            metadata=rejected_metadata,
+        )
+
+        data_gen_strategy = getattr(self, "data_gen_strategy", "quality_preference")
+
+        # Create preference sample
+        sample = PreferenceSample(
+            chosen_trajectory=chosen_trajectory,
+            rejected_trajectory=rejected_trajectory,
+            data_gen_strategy=data_gen_strategy,
+        )
+
+        return sample
+
+    def __len__(self):
+        return len(self.sample_indices)
+
+    def __getitem__(self, idx):
+        return self._generate_sample_from_indices(self.sample_indices[idx])
diff --git a/robometer/robometer/data/samplers/eval/confusion_matrix.py b/robometer/robometer/data/samplers/eval/confusion_matrix.py
new file mode 100755
index 0000000000000000000000000000000000000000..c139bfd12f0b6aa4206b1ab95fb666698ce75005
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/confusion_matrix.py
@@ -0,0 +1,300 @@
+#!/usr/bin/env python3
+"""
+Data generator for confusion matrix analysis.
+"""
+
+import random
+import torch
+from collections import Counter, defaultdict
+from typing import Tuple
+
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.utils.distributed import rank_0_print
+from sentence_transformers import SentenceTransformer
+
+
+class ConfusionMatrixSampler(RBMBaseSampler):
+    """
+    Data generator that creates task-trajectory pairs for confusion matrix analysis.
+
+    For each unique task, creates samples with each trajectory to analyze
+    how well the model can distinguish between different tasks.
+
+    If multiple data sources are present, samples N random trajectories from each data source
+    and prioritizes different language instructions by randomizing the pairing order.
+    """
+
+    def __init__(self, n_trajectories_per_source: int = None, **kwargs):
+        """Initialize confusion matrix sampler.
+
+        Args:
+            n_trajectories_per_source: Number of trajectories to sample from each data source.
+                If None, uses all available trajectories.
+            **kwargs: Additional arguments passed to parent class.
+        """
+        super().__init__(**kwargs)
+        self.n_trajectories_per_source = n_trajectories_per_source
+
+        # Load sentence transformer model and precompute embeddings for all unique tasks
+        self.sentence_model = SentenceTransformer("sentence-transformers/all-MiniLM-L12-v2")
+        self.sentence_model.eval()
+
+        # Precompute language embeddings for all unique tasks
+        unique_tasks = list(self.task_indices.keys())
+        rank_0_print(f"Precomputing language embeddings for {len(unique_tasks)} unique tasks", verbose=self.verbose)
+        self.task_embeddings = {}
+        for task in unique_tasks:
+            embedding = self.sentence_model.encode(task)
+            self.task_embeddings[task] = torch.tensor(embedding)
+        rank_0_print(f"Precomputed {len(self.task_embeddings)} language embeddings", verbose=self.verbose)
+
+        # Free up the model after precomputation (no longer needed)
+        del self.sentence_model
+
+        self.sample_indices = self._generate_all_sample_indices()
+
+        rank_0_print(
+            f"Generated {len(self.sample_indices)} confusion matrix sample indices from {len(self.robot_trajectories)} trajectories and {len(self.task_indices)} tasks"
+        )
+
+    def _generate_all_sample_indices(self) -> list[dict]:
+        """Generate all possible task-trajectory pair sample indices.
+
+        If multiple data sources exist, samples N random trajectories from each data source.
+        Prioritizes different video tasks first, then prioritizes different language instructions
+        when creating pairs.
+        """
+        sample_indices = []
+
+        # Get unique tasks (these will be the language instructions)
+        unique_lang_tasks = list(self.task_indices.keys())
+        rank_0_print(f"Found {len(unique_lang_tasks)} unique language tasks: {unique_lang_tasks}", verbose=self.verbose)
+
+        # Sample trajectories per data source (prioritizing different video tasks)
+        sampled_trajectories, stats = self._sample_trajectories_by_data_source()
+
+        rank_0_print(
+            f"Processing {len(sampled_trajectories)} trajectories for confusion matrix analysis",
+            verbose=self.verbose,
+        )
+
+        # Print statistics about sampled trajectories
+        self._print_sampling_stats(stats)
+
+        # Shuffle language tasks once for round-robin pairing
+        shuffled_lang_tasks = unique_lang_tasks.copy()
+        self._local_random.shuffle(shuffled_lang_tasks)
+
+        # Create task-trajectory pairs with prioritized language instruction pairing
+        video_task_count = Counter()
+
+        for traj_idx in sampled_trajectories:
+            traj = self.dataset[traj_idx]
+            video_task = traj["task"]
+
+            # # Limit the number of video trajectories for each task to 5
+            # if video_task_count[video_task] >= 5:
+            #     continue
+
+            video_task_count[video_task] += 1
+
+            # Pair this trajectory with all language tasks (shuffled for variety)
+            traj_id = traj.get("id", str(traj_idx))
+            for lang_task in shuffled_lang_tasks:
+                sample_indices.append({
+                    "traj_idx": traj_idx,
+                    "lang_task": lang_task,
+                    "video_task": video_task,
+                    "video_path": traj["frames"],
+                    "id": traj_id,
+                })
+
+        # Shuffle final sample indices to further randomize the order
+        self._local_random.shuffle(sample_indices)
+
+        # Print statistics about pairs created
+        rank_0_print(f"Generated {len(sample_indices)} task-trajectory pairs", verbose=self.verbose)
+        rank_0_print(f"  Video tasks sampled: {dict(video_task_count)}", verbose=self.verbose)
+        rank_0_print(f"  Trajectories per video task: {dict(sorted(video_task_count.items()))}", verbose=self.verbose)
+
+        return sample_indices
+
+    def _sample_trajectories_by_data_source(self) -> Tuple[list[int], dict]:
+        """Sample N random trajectories from each data source, prioritizing different video tasks.
+
+        When sampling N trajectories, first selects one trajectory from each unique video task,
+        then repeats in round-robin fashion until N trajectories are sampled.
+
+        Returns:
+            Tuple of (list of sampled trajectory indices, stats dictionary)
+        """
+        sampled_indices = []
+        stats = {
+            "by_source": {},
+            "by_task": Counter(),
+            "traj_to_task": {},
+        }
+
+        # Group robot trajectories by data source, then by video task
+        trajectories_by_source_and_task = defaultdict(lambda: defaultdict(list))
+        for traj_idx in self.robot_trajectories:
+            traj = self.dataset[traj_idx]
+            data_source = traj.get("data_source", "unknown")
+            video_task = traj.get("task", "unknown")
+            trajectories_by_source_and_task[data_source][video_task].append(traj_idx)
+
+        rank_0_print(
+            f"Found {len(trajectories_by_source_and_task)} data sources: {list(trajectories_by_source_and_task.keys())}",
+            verbose=self.verbose,
+        )
+
+        # Sample N trajectories from each data source, prioritizing different tasks
+        for data_source, tasks_to_indices in trajectories_by_source_and_task.items():
+            # Shuffle trajectories within each task for randomization
+            for task in tasks_to_indices:
+                self._local_random.shuffle(tasks_to_indices[task])
+
+            # Get all unique tasks for this data source
+            all_tasks = list(tasks_to_indices.keys())
+            self._local_random.shuffle(all_tasks)  # Randomize task order too
+
+            source_stats = {
+                "total_available": sum(len(indices) for indices in tasks_to_indices.values()),
+                "tasks_available": {task: len(indices) for task, indices in tasks_to_indices.items()},
+                "tasks_sampled": Counter(),
+            }
+
+            if self.n_trajectories_per_source is None:
+                # Use all available trajectories
+                sampled_from_source = []
+                for task, indices in tasks_to_indices.items():
+                    sampled_from_source.extend(indices)
+                    source_stats["tasks_sampled"][task] = len(indices)
+                    stats["by_task"][task] += len(indices)
+
+                rank_0_print(
+                    f"  Data source '{data_source}': Using all {len(sampled_from_source)} trajectories",
+                    verbose=self.verbose,
+                )
+            else:
+                # Sample N trajectories using round-robin to prioritize different tasks
+                n_to_sample = min(self.n_trajectories_per_source, source_stats["total_available"])
+                sampled_from_source = []
+
+                # Round-robin sampling: first get one from each task, then repeat
+                task_iterators = {task: iter(indices) for task, indices in tasks_to_indices.items()}
+                task_list = all_tasks.copy()
+                round_idx = 0
+
+                while len(sampled_from_source) < n_to_sample:
+                    # If we've gone through all tasks once, reshuffle for next round
+                    if round_idx >= len(task_list):
+                        round_idx = 0
+                        self._local_random.shuffle(task_list)
+
+                    # Try to get one trajectory from current task
+                    task = task_list[round_idx]
+                    try:
+                        traj_idx = next(task_iterators[task])
+                        sampled_from_source.append(traj_idx)
+                        source_stats["tasks_sampled"][task] += 1
+                        stats["by_task"][task] += 1
+                    except StopIteration:
+                        # This task is exhausted, remove it from rotation
+                        task_list.pop(round_idx)
+                        if not task_list:
+                            break  # All tasks exhausted
+                        continue
+
+                    round_idx += 1
+
+                rank_0_print(
+                    f"  Data source '{data_source}': Sampled {len(sampled_from_source)} out of {source_stats['total_available']} trajectories",
+                    verbose=self.verbose,
+                )
+                rank_0_print(
+                    f"    Tasks sampled: {dict(sorted(source_stats['tasks_sampled'].items()))}",
+                    verbose=self.verbose,
+                )
+
+            # Track trajectory to task mapping for stats
+            for traj_idx in sampled_from_source:
+                traj = self.dataset[traj_idx]
+                traj_id = traj.get("id", str(traj_idx))
+                stats["traj_to_task"][traj_id] = traj.get("task", "unknown")
+
+            sampled_indices.extend(sampled_from_source)
+            stats["by_source"][data_source] = source_stats
+
+        return sampled_indices, stats
+
+    def _print_sampling_stats(self, stats: dict):
+        """Print detailed statistics about sampled trajectories.
+
+        Args:
+            stats: Statistics dictionary from _sample_trajectories_by_data_source
+        """
+        if not self.verbose:
+            return
+
+        rank_0_print("\n=== Confusion Matrix Sampling Statistics ===", verbose=self.verbose)
+
+        # Overall task statistics
+        rank_0_print(f"\nOverall trajectories per video task:", verbose=self.verbose)
+        for task, count in sorted(stats["by_task"].items()):
+            rank_0_print(f"  {task}: {count} trajectories", verbose=self.verbose)
+
+        # Per data source statistics
+        rank_0_print(f"\nPer data source breakdown:", verbose=self.verbose)
+        for data_source, source_stats in stats["by_source"].items():
+            rank_0_print(f"  Data source: {data_source}", verbose=self.verbose)
+            rank_0_print(f"    Total available: {source_stats['total_available']}", verbose=self.verbose)
+            rank_0_print(f"    Tasks available: {len(source_stats['tasks_available'])}", verbose=self.verbose)
+            for task, count in sorted(source_stats["tasks_available"].items()):
+                sampled_count = source_stats["tasks_sampled"].get(task, 0)
+                rank_0_print(
+                    f"      {task}: {sampled_count}/{count} trajectories sampled",
+                    verbose=self.verbose,
+                )
+
+        rank_0_print("=" * 50, verbose=self.verbose)
+
+    def _generate_sample_from_indices(self, sample_idx_info: dict) -> PreferenceSample:
+        """Generate a single task-trajectory sample from stored indices."""
+        traj_idx = sample_idx_info["traj_idx"]
+        lang_task = sample_idx_info["lang_task"]
+        video_task = sample_idx_info["video_task"]
+        video_path = sample_idx_info["video_path"]
+
+        video_traj = self.dataset[traj_idx]
+
+        # Look up precomputed embedding instead of encoding
+        text_embedding = self.task_embeddings[lang_task]
+
+        metadata = {
+            "id": video_traj["id"],
+            "lang_task": lang_task,
+            "video_task": video_task,
+            "video_path": video_path,
+        }
+
+        # Override task and text_embedding in the trajectory dict
+        video_traj_with_task = video_traj.copy()
+        video_traj_with_task["task"] = lang_task
+        video_traj_with_task["text_embedding"] = text_embedding
+
+        sample_trajectory = self._get_traj_from_data(
+            traj=video_traj_with_task,
+            metadata=metadata,
+            pad_frames=self.pad_frames,
+        )
+
+        sample = ProgressSample(trajectory=sample_trajectory)
+        return sample
+
+    def __len__(self):
+        return len(self.sample_indices)
+
+    def __getitem__(self, idx):
+        return self._generate_sample_from_indices(self.sample_indices[idx])
diff --git a/robometer/robometer/data/samplers/eval/progress_policy_ranking.py b/robometer/robometer/data/samplers/eval/progress_policy_ranking.py
new file mode 100644
index 0000000000000000000000000000000000000000..c0e5fc8a38c770b100e8e873aeac0ebc2ea038ed
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/progress_policy_ranking.py
@@ -0,0 +1,235 @@
+from typing import Dict, List, Any, Optional
+from itertools import cycle
+
+import numpy as np
+from collections import defaultdict
+from robometer.data.dataset_types import ProgressSample
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class ProgressPolicyRankingSampler(RBMBaseSampler):
+    """Dataset that generates progress samples for policy ranking by selecting N trajectories per quality label for tasks with multiple quality labels."""
+
+    def __init__(
+        self,
+        num_examples_per_quality_pr: int = 5,
+        num_partial_successes: Optional[int] = None,
+        frame_step: int = 1,
+        use_frame_steps: bool = True,
+        max_tasks: Optional[int] = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if num_examples_per_quality_pr is None:
+            num_examples_per_quality_pr = float("inf")
+        self.num_examples_per_quality_pr = num_examples_per_quality_pr
+        self.num_partial_successes = num_partial_successes
+        self.frame_step = frame_step
+        self.use_frame_steps = use_frame_steps
+        self.max_tasks = max_tasks
+        logger.info(f"ProgressPolicyRankingSampler initialized with {len(self.robot_trajectories)} trajectories")
+
+        self.sample_indices = self._generate_all_sample_indices()
+
+        logger.info(f"Generated {len(self.sample_indices)} sample indices")
+
+    def _generate_all_sample_indices(self) -> List[Dict[str, Any]]:
+        """Generate sample indices by selecting tasks with multiple quality labels/partial_success values and sampling N trajectories per group.
+
+        For non-RoboArena: Groups by task and quality_label.
+        For RoboArena: Groups by task and partial_success values.
+
+        If use_frame_steps=True, generates subsequence samples like reward_alignment (0:frame_step, 0:2*frame_step, etc.).
+        If use_frame_steps=False, generates one sample per trajectory (whole trajectory).
+        """
+
+        # Check if this is RoboArena (has partial_success)
+        is_roboarena = False
+        if self.robot_trajectories:
+            first_traj = self.dataset[self.robot_trajectories[0]]
+            is_roboarena = first_traj.get("partial_success") is not None
+
+        # Group trajectories by task and grouping key (quality_label or partial_success)
+        task_to_key_to_trajs = defaultdict(lambda: defaultdict(list))
+
+        for traj_idx in self.robot_trajectories:
+            traj = self.dataset[traj_idx]
+            task = traj["task"]
+
+            if is_roboarena:
+                # RoboArena: Use rounded partial_success as key to group similar values
+                partial_success_val = traj.get("partial_success")
+                if partial_success_val is not None:
+                    partial_success = round(float(partial_success_val), 2)
+                    task_to_key_to_trajs[task][partial_success].append(traj_idx)
+            else:
+                # Non-RoboArena: Use quality_label
+                quality = traj["quality_label"]
+                task_to_key_to_trajs[task][quality].append(traj_idx)
+
+        # Filter to tasks that have multiple grouping values
+        tasks_with_multiple_values = {
+            task: key_to_trajs for task, key_to_trajs in task_to_key_to_trajs.items() if len(key_to_trajs) > 1
+        }
+
+        dataset_type_str = "partial_success values" if is_roboarena else "quality labels"
+        logger.info(f"Found {len(tasks_with_multiple_values)} tasks with multiple {dataset_type_str}")
+
+        # Limit number of tasks if max_tasks is specified
+        if self.max_tasks is not None and self.max_tasks > 0:
+            # Convert to list, shuffle, and take first max_tasks
+            # Sort tasks first to ensure deterministic ordering before shuffling
+            tasks_list = sorted(tasks_with_multiple_values.items())
+            self._local_random.shuffle(tasks_list)
+            tasks_with_multiple_values = dict(tasks_list[: self.max_tasks])
+            logger.info(f"Limited to {len(tasks_with_multiple_values)} tasks (max_tasks={self.max_tasks})")
+
+        # Sample trajectories for each task
+        sample_indices = []
+        all_sampled_traj_indices = []
+        # Sort tasks to ensure deterministic processing order
+        for task, key_to_trajs in sorted(tasks_with_multiple_values.items()):
+            if is_roboarena:
+                # RoboArena: Use num_partial_successes for circular sampling
+                num_to_sample_total = self.num_partial_successes
+
+                # Build lists of available indices per partial_success (sorted for deterministic sampling)
+                available_lists = []
+                for partial_success in sorted(key_to_trajs.keys()):
+                    traj_indices = sorted(key_to_trajs[partial_success])
+                    if traj_indices:
+                        available_lists.append(traj_indices)
+
+                # Circular sampling: cycle through partial_success groups until we reach max
+                sampled_traj_indices = []
+                for available_indices in cycle(available_lists):
+                    if len(sampled_traj_indices) >= num_to_sample_total:
+                        break
+                    if not available_indices:
+                        # If all lists are empty, stop
+                        if all(not lst for lst in available_lists):
+                            break
+                        continue
+
+                    # Sample one trajectory from this group
+                    sampled_idx = self._local_random.choice(available_indices)
+                    sampled_traj_indices.append(sampled_idx)
+                    # Remove the sampled index from this list
+                    available_indices.remove(sampled_idx)
+
+                # Generate samples for all sampled trajectories
+                for traj_idx in sampled_traj_indices:
+                    traj = self.dataset[traj_idx]
+                    sample_indices.extend(self._generate_indices_for_trajectory(traj_idx, traj))
+                    all_sampled_traj_indices.append(traj_idx)
+            else:
+                # Non-RoboArena: Sample N trajectories per quality label
+                # Sort quality labels to ensure deterministic order
+                for quality in sorted(key_to_trajs.keys()):
+                    traj_indices = key_to_trajs[quality]
+                    # Sort trajectory indices to ensure deterministic sampling
+                    traj_indices = sorted(traj_indices)
+                    # Sample up to num_examples_per_quality_pr trajectories for this quality label
+                    num_to_sample = min(self.num_examples_per_quality_pr, len(traj_indices))
+                    sampled_traj_indices = self._local_random.sample(traj_indices, num_to_sample)
+                    for traj_idx in sampled_traj_indices:
+                        traj = self.dataset[traj_idx]
+                        sample_indices.extend(self._generate_indices_for_trajectory(traj_idx, traj))
+                        all_sampled_traj_indices.append(traj_idx)
+
+        logger.info(f"Sampled {len(sample_indices)} samples across {len(tasks_with_multiple_values)} tasks")
+        logger.info(f"Sampled trajectory indices: {all_sampled_traj_indices}")
+
+        return sample_indices
+
+    def _generate_indices_for_trajectory(self, traj_idx: int, traj: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """Generate sample indices for a single trajectory.
+
+        Args:
+            traj_idx: Index of the trajectory in the dataset
+            traj: Trajectory dictionary
+
+        Returns:
+            List of sample index dictionaries
+        """
+        num_frames = traj["num_frames"]
+        indices = []
+
+        if self.use_frame_steps:
+            # Generate subsequence indices like reward_alignment: 0:frame_step, 0:2*frame_step, etc.
+            for end_idx in range(self.frame_step, num_frames + 1, self.frame_step):
+                frame_indices = list(range(end_idx))
+                indices.append({
+                    "traj_idx": traj_idx,
+                    "frame_indices": frame_indices,
+                    "num_frames": num_frames,
+                    "video_path": traj["frames"],
+                    "id": traj["id"],
+                    "use_frame_steps": True,
+                })
+        else:
+            # Generate one sample per trajectory (whole trajectory)
+            indices.append({
+                "traj_idx": traj_idx,
+                "video_path": traj["frames"],
+                "id": traj["id"],
+                "use_frame_steps": False,
+            })
+
+        return indices
+
+    def _generate_sample_from_indices(self, sample_idx_info: dict) -> ProgressSample:
+        """Generate a single progress sample from trajectory index."""
+        traj_idx = sample_idx_info["traj_idx"]
+        use_frame_steps = sample_idx_info.get("use_frame_steps", True)
+
+        traj = self.dataset[traj_idx]
+
+        if use_frame_steps:
+            # Frame steps mode: create subsequence like reward_alignment
+            frame_indices = sample_idx_info["frame_indices"]
+            num_frames = sample_idx_info["num_frames"]
+
+            metadata = {
+                "quality_label": traj["quality_label"],
+                "data_source": traj["data_source"],
+                "task": traj["task"],
+                "id": traj["id"],
+                "video_path": sample_idx_info["video_path"],
+                "frame_step": frame_indices[-1] if frame_indices else 0,
+            }
+
+            trajectory = self._get_traj_from_data(
+                traj=traj,
+                frame_indices=frame_indices,
+                metadata=metadata,
+                pad_frames=self.pad_frames,
+            )
+        else:
+            # Whole trajectory mode
+            metadata = {
+                "quality_label": traj["quality_label"],
+                "data_source": traj["data_source"],
+                "task": traj["task"],
+                "id": traj["id"],
+                "video_path": sample_idx_info["video_path"],
+            }
+
+            trajectory = self._get_traj_from_data(
+                traj=traj,
+                metadata=metadata,
+                pad_frames=self.pad_frames,
+            )
+
+        sample = ProgressSample(trajectory=trajectory)
+        return sample
+
+    def __len__(self):
+        return len(self.sample_indices)
+
+    def __getitem__(self, idx):
+        return self._generate_sample_from_indices(self.sample_indices[idx])
diff --git a/robometer/robometer/data/samplers/eval/quality_preference.py b/robometer/robometer/data/samplers/eval/quality_preference.py
new file mode 100644
index 0000000000000000000000000000000000000000..d8d96048a551bac50f20700d6fbee6dbf6659dbb
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/quality_preference.py
@@ -0,0 +1,219 @@
+from typing import Dict, List, Any
+
+from itertools import combinations
+from tqdm import tqdm
+
+from robometer.data.samplers.eval.base_pref import BaseQualityPreferenceSampler
+from robometer.utils.distributed import rank_0_print
+
+
+class QualityPreferenceSampler(BaseQualityPreferenceSampler):
+    """Dataset that generates preference samples by pairing trajectories with different quality labels or partial_success values for the same task.
+
+    For non-RoboArena: Pairs trajectories with different quality labels (failure, suboptimal, successful).
+    For RoboArena: Pairs trajectories with different partial_success values (higher partial_success = chosen).
+    """
+
+    def __init__(
+        self,
+        comparisons_per_task=None,
+        max_comparisons=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        # Set data_gen_strategy for this sampler
+        self.data_gen_strategy = "quality_preference"
+        self.comparisons_per_task = comparisons_per_task
+        self.max_comparisons = max_comparisons
+
+        # Generate all possible sample indices upfront (not the actual samples)
+        self.sample_indices = self._generate_all_sample_indices()
+        rank_0_print(f"Generated {len(self.sample_indices)} quality preference sample indices", verbose=self.verbose)
+
+    def _generate_all_sample_indices(self) -> List[Dict[str, Any]]:
+        """Generate all possible quality preference sample indices (not the actual samples).
+
+        For non-RoboArena: Groups by task and quality_label, pairs trajectories with different quality labels.
+        For RoboArena: Groups by task and partial_success values, pairs trajectories with different partial_success.
+        """
+        sample_indices = []
+
+        # Check if this is RoboArena (has partial_success)
+        is_roboarena = False
+        if self.robot_trajectories:
+            first_traj = self.dataset[self.robot_trajectories[0]]
+            is_roboarena = first_traj.get("partial_success") is not None
+
+        rank_0_print(
+            f"Generating quality preference samples for {len(self.robot_trajectories)} trajectories "
+            f"({'RoboArena (partial_success)' if is_roboarena else 'non-RoboArena (quality_label)'})",
+            verbose=self.verbose,
+        )
+
+        if is_roboarena:
+            # RoboArena: Group by task and partial_success (rounded to 2 decimals)
+            task_to_partial_trajs = {}
+
+            for traj_idx in self.robot_trajectories:
+                traj = self.dataset[traj_idx]
+                task = traj["task"]
+                partial_success_val = traj.get("partial_success")
+
+                if partial_success_val is None:
+                    rank_0_print(
+                        f"Warning: Trajectory {traj_idx} (task: {task}) missing partial_success, skipping",
+                        verbose=self.verbose,
+                    )
+                    continue
+
+                # Round partial_success to 2 decimals for grouping
+                partial_success = round(float(partial_success_val), 2)
+
+                if task not in task_to_partial_trajs:
+                    task_to_partial_trajs[task] = {}
+
+                if partial_success not in task_to_partial_trajs[task]:
+                    task_to_partial_trajs[task][partial_success] = []
+
+                task_to_partial_trajs[task][partial_success].append(traj_idx)
+
+            # Generate pairs for each task
+            for task in tqdm(task_to_partial_trajs, desc="Generating RoboArena quality preference samples"):
+                partial_groups = task_to_partial_trajs[task]
+                partial_values = list(partial_groups.keys())
+
+                # Only create pairs if we have at least 2 different partial_success values
+                if len(partial_values) < 2:
+                    continue
+
+                # Collect all pairs for this task
+                task_pairs = []
+
+                # Create pairs of different partial_success values
+                for partial1, partial2 in combinations(partial_values, 2):
+                    trajs1 = partial_groups[partial1]
+                    trajs2 = partial_groups[partial2]
+
+                    if not trajs1 or not trajs2:
+                        continue
+
+                    # Determine which partial_success is higher (chosen)
+                    if partial1 > partial2:
+                        chosen_partial = partial1
+                        rejected_partial = partial2
+                        chosen_indices = trajs1
+                        rejected_indices = trajs2
+                    elif partial2 > partial1:
+                        chosen_partial = partial2
+                        rejected_partial = partial1
+                        chosen_indices = trajs2
+                        rejected_indices = trajs1
+                    else:
+                        # Same value, skip this pair
+                        continue
+
+                    # Create all possible pairs for this partial_success combination
+                    for chosen_idx in chosen_indices:
+                        for rejected_idx in rejected_indices:
+                            task_pairs.append({
+                                "chosen_traj_idx": chosen_idx,
+                                "rejected_traj_idx": rejected_idx,
+                                "task": task,
+                                "chosen_partial_success": chosen_partial,
+                                "rejected_partial_success": rejected_partial,
+                            })
+
+                # Apply comparisons_per_task limit if set (sample uniformly across all pairs for this task)
+                if self.comparisons_per_task is not None and len(task_pairs) > self.comparisons_per_task:
+                    # Uniformly sample comparisons for this task
+                    task_pairs = self._local_random.sample(task_pairs, self.comparisons_per_task)
+
+                sample_indices.extend(task_pairs)
+
+        else:
+            # Non-RoboArena: Group by task and quality label
+            task_to_quality_trajs = {}
+
+            for traj_idx in self.robot_trajectories:
+                traj = self.dataset[traj_idx]
+                task = traj["task"]
+                quality_label = traj["quality_label"]
+
+                if task not in task_to_quality_trajs:
+                    task_to_quality_trajs[task] = {}
+
+                if quality_label not in task_to_quality_trajs[task]:
+                    task_to_quality_trajs[task][quality_label] = []
+
+                task_to_quality_trajs[task][quality_label].append(traj_idx)
+
+            # Generate pairs for each task
+            quality_order = {"failure": 1, "suboptimal": 2, "successful": 3}
+
+            for task in tqdm(task_to_quality_trajs, desc="Generating quality preference samples"):
+                quality_groups = task_to_quality_trajs[task]
+                quality_labels = list(quality_groups.keys())
+
+                # Only create pairs if we have at least 2 different quality labels
+                if len(quality_labels) < 2:
+                    continue
+
+                # Collect all pairs for this task
+                task_pairs = []
+
+                # Create pairs of different quality labels
+                for quality1, quality2 in combinations(quality_labels, 2):
+                    trajs1 = quality_groups[quality1]
+                    trajs2 = quality_groups[quality2]
+
+                    if not trajs1 or not trajs2:
+                        continue
+
+                    # Determine which quality is better (chosen)
+                    order1 = quality_order.get(quality1, 0)
+                    order2 = quality_order.get(quality2, 0)
+
+                    # Only create pairs if one quality is strictly better than the other
+                    if order1 > order2:
+                        chosen_quality = quality1
+                        rejected_quality = quality2
+                        chosen_indices = trajs1
+                        rejected_indices = trajs2
+                    elif order2 > order1:
+                        chosen_quality = quality2
+                        rejected_quality = quality1
+                        chosen_indices = trajs2
+                        rejected_indices = trajs1
+                    else:
+                        # Same order, skip this pair as we can't reliably compare them
+                        continue
+
+                    # Create all possible pairs for this quality combination
+                    for chosen_idx in chosen_indices:
+                        for rejected_idx in rejected_indices:
+                            task_pairs.append({
+                                "chosen_traj_idx": chosen_idx,
+                                "rejected_traj_idx": rejected_idx,
+                                "task": task,
+                                "chosen_quality": chosen_quality,
+                                "rejected_quality": rejected_quality,
+                            })
+
+                # Apply comparisons_per_task limit if set (sample uniformly across all pairs for this task)
+                if self.comparisons_per_task is not None and len(task_pairs) > self.comparisons_per_task:
+                    # Uniformly sample comparisons for this task
+                    task_pairs = self._local_random.sample(task_pairs, self.comparisons_per_task)
+
+                sample_indices.extend(task_pairs)
+
+        # Apply max_comparisons limit if set (sample uniformly across all comparisons)
+        original_count = len(sample_indices)
+        if self.max_comparisons is not None and original_count > self.max_comparisons:
+            sample_indices = self._local_random.sample(sample_indices, self.max_comparisons)
+            rank_0_print(
+                f"Limited total comparisons to {self.max_comparisons} (from {original_count} total comparisons)",
+                verbose=self.verbose,
+            )
+
+        return sample_indices
diff --git a/robometer/robometer/data/samplers/eval/reward_alignment.py b/robometer/robometer/data/samplers/eval/reward_alignment.py
new file mode 100644
index 0000000000000000000000000000000000000000..e7c4877d3058922072f3a1678476035e38c01e37
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/reward_alignment.py
@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""
+Data generator for reward alignment evaluation.
+
+This generator creates subsequence samples from trajectories for progress prediction evaluation.
+For each trajectory, it creates multiple subsequences (0:2, 0:4, 0:6, etc.) and formats them
+as PreferenceSample objects that can be evaluated by the model.
+"""
+
+from typing import Dict, List, Any
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from robometer.data.dataset_types import ProgressSample, Trajectory
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.utils.distributed import rank_0_print
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class RewardAlignmentSampler(RBMBaseSampler):
+    """
+    Data generator that creates subsequence samples for reward alignment evaluation.
+
+    For each trajectory, creates subsequences of frames (0:2, 0:4, 0:6, etc.)
+    and formats them as PreferenceSample objects for evaluation.
+    """
+
+    def __init__(
+        self,
+        max_trajectories: int | None = None,
+        frame_step: int = 1,
+        use_frame_steps: bool = True,
+        subsample_n_frames: int = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.max_trajectories = max_trajectories
+        self.frame_step = frame_step
+        self.use_frame_steps = use_frame_steps
+        self.subsample_n_frames = subsample_n_frames
+        self.sample_indices = self._generate_all_sample_indices()
+
+        rank_0_print(
+            f"Generated {len(self.sample_indices)} reward alignment sample indices from {min(len(self.robot_trajectories), self.max_trajectories) if self.max_trajectories else len(self.robot_trajectories)} trajectories",
+            verbose=self.verbose,
+        )
+
+    def _generate_all_sample_indices(self) -> List[Dict[str, Any]]:
+        """Generate all possible subsequence sample indices (not the actual samples)."""
+        sample_indices = []
+
+        # Limit number of trajectories if specified
+        trajectories_to_process = self.robot_trajectories
+        if self.max_trajectories is not None and self.max_trajectories < len(self.robot_trajectories):
+            trajectories_to_process = self._local_random.sample(self.robot_trajectories, self.max_trajectories)
+
+        rank_0_print(
+            f"Generating subsequence samples for {len(trajectories_to_process)} trajectories", verbose=self.verbose
+        )
+
+        all_num_frames = []
+        for traj_idx in trajectories_to_process:
+            traj = self.dataset[traj_idx]
+            num_frames = traj["num_frames"]
+            all_num_frames.append(num_frames)
+            sample_indices.extend(self._generate_indices_for_trajectory(traj_idx, traj))
+
+        logger.info(f"All num frames: {all_num_frames}")
+        return sample_indices
+
+    def _generate_indices_for_trajectory(self, traj_idx: int, traj: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """Generate sample indices for a single trajectory.
+
+        Args:
+            traj_idx: Index of the trajectory in the dataseqt
+            traj: Trajectory dictionary
+
+        Returns:
+            List of sample index dictionaries
+        """
+        num_frames = traj["num_frames"]
+        indices = []
+
+        if self.use_frame_steps:
+            if self.subsample_n_frames:
+                if self.subsample_n_frames > num_frames:
+                    end_indices = list(range(num_frames))
+                else:
+                    end_indices = np.linspace(0, num_frames - 1, self.subsample_n_frames)
+                for end_idx in end_indices:
+                    frame_indices = list(range(int(end_idx) + 1))
+                    indices.append({
+                        "traj_idx": traj_idx,
+                        "frame_indices": frame_indices,
+                        "num_frames": num_frames,
+                        "video_path": traj["frames"],
+                        "id": traj["id"],
+                        "use_frame_steps": True,
+                    })
+            else:
+                # Generate subsequence indices like reward_alignment: 0:frame_step, 0:2*frame_step, etc.
+                for end_idx in range(self.frame_step, num_frames + 1, self.frame_step):
+                    frame_indices = list(range(end_idx))
+                    indices.append({
+                        "traj_idx": traj_idx,
+                        "frame_indices": frame_indices,
+                        "num_frames": num_frames,
+                        "video_path": traj["frames"],
+                        "id": traj["id"],
+                        "use_frame_steps": True,
+                    })
+        else:
+            # Generate one sample per trajectory (whole trajectory)
+            indices.append({
+                "traj_idx": traj_idx,
+                "video_path": traj["frames"],
+                "id": traj["id"],
+                "use_frame_steps": False,
+            })
+
+        return indices
+
+    def _generate_sample_from_indices(self, sample_idx_info: dict) -> ProgressSample:
+        """Generate a single subsequence sample from stored indices."""
+        traj_idx = sample_idx_info["traj_idx"]
+        use_frame_steps = sample_idx_info.get("use_frame_steps", True)
+
+        traj = self.dataset[traj_idx]
+
+        if use_frame_steps:
+            # Frame steps mode: create subsequence like reward_alignment
+            frame_indices = sample_idx_info["frame_indices"]
+            num_frames = sample_idx_info["num_frames"]
+
+            metadata = {
+                "data_gen_strategy": "reward_alignment",
+                "id": traj["id"],
+                "video_path": sample_idx_info["video_path"],
+                "frame_step": frame_indices[-1] if frame_indices else 0,
+                "num_frames": num_frames,
+            }
+
+            trajectory = self._get_traj_from_data(
+                traj=traj,
+                frame_indices=frame_indices,
+                metadata=metadata,
+                pad_frames=self.pad_frames,
+            )
+        else:
+            # Whole trajectory mode
+            metadata = {
+                "data_gen_strategy": "reward_alignment",
+                "id": traj["id"],
+                "video_path": sample_idx_info["video_path"],
+            }
+
+            trajectory = self._get_traj_from_data(
+                traj=traj,
+                metadata=metadata,
+                pad_frames=self.pad_frames,
+            )
+
+        sample = ProgressSample(trajectory=trajectory, sample_type="progress")
+        return sample
+
+    def __len__(self):
+        return len(self.sample_indices)
+
+    def __getitem__(self, idx):
+        return self._generate_sample_from_indices(self.sample_indices[idx])
diff --git a/robometer/robometer/data/samplers/eval/roboarena_quality_preference.py b/robometer/robometer/data/samplers/eval/roboarena_quality_preference.py
new file mode 100644
index 0000000000000000000000000000000000000000..9d2cc5a396dd09bb283ebe0b311d187e56dd8234
--- /dev/null
+++ b/robometer/robometer/data/samplers/eval/roboarena_quality_preference.py
@@ -0,0 +1,121 @@
+from typing import Dict, List, Any
+
+from tqdm import tqdm
+
+from robometer.data.samplers.eval.base_pref import BaseQualityPreferenceSampler
+from robometer.utils.distributed import rank_0_print
+
+
+class RoboArenaQualityPreferenceSampler(BaseQualityPreferenceSampler):
+    """Dataset that generates preference samples by pairing trajectories with different partial_rewards for the same task.
+
+    For RoboArena dataset, pairs trajectories from the same task where the chosen trajectory
+    has a higher partial_reward (partial_success) than the rejected trajectory.
+    """
+
+    def __init__(
+        self,
+        comparisons_per_task=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        # Set data_gen_strategy for this sampler
+        self.data_gen_strategy = "quality_preference_roboarena"
+
+        self._cached_tasks = self.dataset["task"]
+        self._cached_partial_success = self.dataset.get("partial_success")
+
+        self.comparisons_per_task = comparisons_per_task
+
+        # Generate all possible sample indices upfront (not the actual samples)
+        self.sample_indices = self._generate_all_sample_indices()
+        rank_0_print(
+            f"Generated {len(self.sample_indices)} RoboArena quality preference sample indices", verbose=self.verbose
+        )
+
+    def _generate_all_sample_indices(self) -> List[Dict[str, Any]]:
+        """Generate all possible quality preference sample indices based on partial_reward (partial_success)."""
+        sample_indices = []
+
+        # Group trajectories by task
+        task_to_trajs = {}
+
+        rank_0_print(
+            f"Generating RoboArena quality preference samples for {len(self.robot_trajectories)} trajectories",
+            verbose=self.verbose,
+        )
+
+        for traj_idx in self.robot_trajectories:
+            # Use cached arrays for efficient access
+            task = self._cached_tasks[traj_idx]
+            partial_success = (
+                self._cached_partial_success[traj_idx] if traj_idx < len(self._cached_partial_success) else None
+            )
+
+            # Ensure partial_success exists
+            if partial_success is None:
+                rank_0_print(
+                    f"Warning: Trajectory {traj_idx} (task: {task}) missing partial_success, skipping",
+                    verbose=self.verbose,
+                )
+                continue
+
+            if task not in task_to_trajs:
+                task_to_trajs[task] = []
+
+            task_to_trajs[task].append({
+                "traj_idx": traj_idx,
+                "partial_success": float(partial_success),
+            })
+
+        # Generate pairs for each task
+        for task in tqdm(task_to_trajs, desc="Generating RoboArena quality preference samples"):
+            trajs = task_to_trajs[task]
+
+            # Need at least 2 trajectories to create pairs
+            if len(trajs) < 2:
+                continue
+
+            # Create all pairs of trajectories
+            task_pairs = []
+            for i, traj1 in enumerate(trajs):
+                for j, traj2 in enumerate(trajs):
+                    if i >= j:  # Avoid duplicates and self-pairs
+                        continue
+
+                    partial1 = traj1["partial_success"]
+                    partial2 = traj2["partial_success"]
+
+                    # Skip if partial_success values are equal (can't determine preference)
+                    if partial1 == partial2:
+                        continue
+
+                    # Determine which trajectory is chosen (higher partial_success)
+                    if partial1 > partial2:
+                        chosen_traj_idx = traj1["traj_idx"]
+                        rejected_traj_idx = traj2["traj_idx"]
+                        chosen_partial = partial1
+                        rejected_partial = partial2
+                    else:
+                        chosen_traj_idx = traj2["traj_idx"]
+                        rejected_traj_idx = traj1["traj_idx"]
+                        chosen_partial = partial2
+                        rejected_partial = partial1
+
+                    task_pairs.append({
+                        "chosen_traj_idx": chosen_traj_idx,
+                        "rejected_traj_idx": rejected_traj_idx,
+                        "task": task,
+                        "chosen_partial_success": chosen_partial,
+                        "rejected_partial_success": rejected_partial,
+                    })
+
+            # Apply comparisons_per_task limit if set (sample uniformly across all pairs for this task)
+            if self.comparisons_per_task is not None and len(task_pairs) > self.comparisons_per_task:
+                # Uniformly sample comparisons for this task
+                task_pairs = self._local_random.sample(task_pairs, self.comparisons_per_task)
+
+            sample_indices.extend(task_pairs)
+
+        return sample_indices
diff --git a/robometer/robometer/data/samplers/pref.py b/robometer/robometer/data/samplers/pref.py
new file mode 100755
index 0000000000000000000000000000000000000000..41d7021db8df64b9db0d2502cf6552c4a036c748
--- /dev/null
+++ b/robometer/robometer/data/samplers/pref.py
@@ -0,0 +1,370 @@
+#!/usr/bin/env python3
+"""
+PrefSampler class for producing batches of preference data.
+"""
+
+from typing import Dict, List, Optional, Any
+
+import random
+
+from robometer.data.dataset_types import PreferenceSample, Trajectory
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.data.datasets.helpers import (
+    DataGenStrat,
+    convert_continuous_to_discrete_bins,
+)
+from robometer.utils.logger import get_logger, rank_0_info, trace
+from robometer.utils.timer import timer
+
+logger = get_logger()
+
+
+class PrefSampler(RBMBaseSampler):
+    """Data generator for producing batches of preference prediction data."""
+
+    def __init__(self, is_evaluation=False, **kwargs):
+        super().__init__(**kwargs)
+
+        self.dataset_preference_ratio = self.config.dataset_preference_ratio
+        self.preference_strategy_ratio: List[float] = self.config.preference_strategy_ratio
+        self._has_suboptimal = (
+            any(len(indices) > 0 for indices in self.suboptimal_by_task.values()) if self.suboptimal_by_task else False
+        )
+        rank_0_info(f"[PREF SAMPLER] Has suboptimal: {self._has_suboptimal}")
+
+        # Initialize preference dataset
+        self._load_preference_dataset()
+
+    def _generate_sample(self, item: dict, preferred_strategy: Optional[DataGenStrat] = None):
+        """Generate a preference sample from an item.
+
+        If the item has a non-successful quality label, it will be used as the rejected
+        trajectory and an optimal trajectory from the same task will be found as the chosen one.
+        Otherwise, normal preference sampling logic is used.
+
+        Args:
+            item: The trajectory item
+            preferred_strategy: Optional strategy to use (if None, will select strategy based on ratios)
+        """
+        quality_label = item["quality_label"]
+        use_partial_success = item.get("partial_success") is not None
+
+        # Handle non-successful trajectories: use as rejected, find optimal from same task as chosen
+        # skip this for trajectories with partial_success which we will handle with partial success logic
+        if quality_label != "successful" and not use_partial_success:
+            traj_id = item["id"]
+            task_name = item["task"]
+
+            logger.trace(
+                f"[PREF SAMPLER] Non-successful quality detected for ID={traj_id}, using as rejected trajectory, task={task_name}"
+            )
+
+            # Find optimal trajectories from the same task
+            same_task_optimal_indices = self.optimal_by_task.get(task_name, [])
+
+            if not same_task_optimal_indices:
+                logger.trace(
+                    f"[PREF SAMPLER] No optimal trajectories found for task '{task_name}', falling through to normal sampling"
+                )
+                return self._create_pref_sample(item, preferred_strategy=preferred_strategy)
+
+            # Select a random optimal trajectory from the same task as chosen
+            chosen_idx = random.choice(same_task_optimal_indices)
+            chosen_traj_dict = self.dataset[chosen_idx]
+
+            chosen_trajectory = self._get_traj_from_data(chosen_traj_dict)
+            rejected_trajectory = self._get_traj_from_data(item)
+
+            sample = PreferenceSample(
+                chosen_trajectory=chosen_trajectory,
+                rejected_trajectory=rejected_trajectory,
+                data_gen_strategy=DataGenStrat.SUBOPTIMAL.value,
+            )
+
+            logger.trace(
+                f"[PREF SAMPLER] Created preference sample for non-successful traj ID={traj_id} with optimal traj from same task"
+            )
+            return sample
+
+        return self._create_pref_sample(item, preferred_strategy=preferred_strategy)
+
+    def _execute_strategy(
+        self, strategy: DataGenStrat, chosen_traj: Dict[str, Any], use_partial_success: bool
+    ) -> tuple[Dict[str, Any], str, Dict[str, Any]] | None:
+        """Execute a strategy to get rejected trajectory.
+
+        Args:
+            strategy: The strategy to execute
+            chosen_traj: The chosen trajectory
+            use_partial_success: Whether this trajectory uses partial_success
+
+        Returns:
+            Tuple of (rejected_traj, rejected_subsample_strategy, chosen_traj) or None if failed
+            Note: chosen_traj may be swapped with rejected_traj for partial_success trajectories
+        """
+        max_retries = 3
+        rejected_subsample_strategy = None
+        rejected_traj = None
+
+        if strategy == DataGenStrat.REWIND:
+            rejected_traj = chosen_traj.copy()
+            rejected_subsample_strategy = "subsample_rewind"
+        elif strategy == DataGenStrat.SUBOPTIMAL:
+            for _ in range(max_retries):
+                rejected_traj = self._get_same_task_suboptimal(chosen_traj)
+                if rejected_traj is not None:
+                    # For trajectories with partial_success, if the returned trajectory has higher partial_success, swap them
+                    if use_partial_success:
+                        chosen_partial_success = chosen_traj.get("partial_success")
+                        rejected_partial_success = rejected_traj.get("partial_success")
+                        if rejected_partial_success is not None and chosen_partial_success is not None:
+                            if rejected_partial_success > chosen_partial_success:
+                                logger.trace(
+                                    f"[PREF SAMPLER] Swapping trajectories: found higher partial_success "
+                                    f"({rejected_partial_success} > {chosen_partial_success})"
+                                )
+                                rejected_traj, chosen_traj = chosen_traj, rejected_traj
+                    break
+            rejected_subsample_strategy = "subsample_forward"
+        elif strategy == DataGenStrat.DIFFERENT_TASK:
+            for _ in range(max_retries):
+                rejected_traj = self._get_different_video_traj(chosen_traj)
+                if rejected_traj is not None:
+                    break
+            rejected_subsample_strategy = "subsample_forward"
+        elif strategy == DataGenStrat.REVERSE_PROGRESS:
+            rejected_traj = chosen_traj.copy()
+            rejected_subsample_strategy = "subsample_reverse"
+        else:
+            return None
+
+        if rejected_traj is None:
+            return None
+
+        return (rejected_traj, rejected_subsample_strategy, chosen_traj)
+
+    def _create_pref_sample_from_dataset(self) -> PreferenceSample:
+        """Create a preference sample from the loaded preference dataset."""
+        if not self.preferences:
+            return None
+
+        # For now, return a simple preference sample
+        # This can be enhanced later when we have actual preference data
+        random.choice(self.preferences)
+
+        # This is a placeholder - would need to be implemented based on actual preference data structure
+        return None
+
+    def _load_preference_dataset(self):
+        """Load the preference dataset from disk or hub if provided."""
+        self.preferences = []
+
+        # For now, we'll use empty preferences since the config structure has changed
+        # This can be updated later if needed
+        rank_0_info("[PREF SAMPLER] No preference dataset provided, will use random sampling for preferences")
+        return
+
+    def _create_preference_sample(self) -> PreferenceSample:
+        """Create a preference prediction sample: chosen vs rejected where chosen is preferred.
+        Either from dataset or from generated trajectories.
+
+        Returns:
+            PreferenceSample: A preference sample with chosen (preferred) vs rejected
+            (suboptimal) trajectories and associated metadata
+        """
+
+        with timer("create_preference_sample", verbose=False):
+            if random.random() < self.dataset_preference_ratio and self.preferences:
+                # Use preference trajectories from dataset
+                return self._create_pref_sample_from_dataset()
+            else:
+                return self._create_pref_sample()
+
+    def _create_pref_sample(
+        self, chosen_traj: Optional[Dict[str, Any]] = None, preferred_strategy: Optional[DataGenStrat] = None
+    ) -> PreferenceSample:
+        """Create a preference prediction sample using various rejected trajectory generation strategies.
+
+        Rewind Same Task
+        - Creates a suboptimal trajectory by rewinding the chosen trajectory
+
+        Suboptimal/Failure Same Task
+        - Uses existing suboptimal/failure trajectories from the same task
+
+        Different Task
+        - Uses trajectories from completely different tasks
+
+        Returns:
+            PreferenceSample: A preference sample with chosen (preferred) vs rejected
+            (suboptimal) trajectories and associated metadata
+
+        Raises:
+            ValueError: If no chosen trajectories are available for preference generation
+            RuntimeError: If all strategies fail and fallback rewind also fails
+        """
+        # Log when preference sampler is called
+        traj_id = chosen_traj["id"] if chosen_traj is not None else "sampling_new"
+        logger.trace(f"[PREF SAMPLER] Creating preference sample for trajectory ID: {traj_id}")
+
+        # Use provided chosen trajectory if given; otherwise sample one
+        if chosen_traj is None:
+            # Use preprocessed chosen trajectories from index maps
+            if not self.optimal_by_task:
+                return None
+
+            # Filter out tasks with empty optimal_indices to avoid infinite loop
+            valid_tasks = {
+                task: indices
+                for task, indices in self.optimal_by_task.items()
+                if indices  # Only include tasks with non-empty indices
+            }
+
+            if not valid_tasks:
+                # No valid tasks with optimal trajectories available
+                return None
+
+            # Get a random task and chosen trajectory from it
+            task_name = random.choice(list(valid_tasks.keys()))
+            optimal_indices = valid_tasks[task_name]
+
+            # Double-check that we have valid indices (should always be true now)
+            if not optimal_indices:
+                return None
+
+            chosen_idx = random.choice(optimal_indices)
+            chosen_traj = self.dataset[chosen_idx]
+
+        # Initialize variables for strategy selection
+        rejected_traj = None
+        strategy_used = None
+        rejected_subsample_strategy = None
+
+        # Check if this trajectory uses partial_success
+        use_partial_success = chosen_traj.get("partial_success") is not None
+        if use_partial_success:
+            partial_success = chosen_traj.get("partial_success")
+            logger.trace(
+                f"[PREF SAMPLER] Trajectory with partial_success detected (ID: {chosen_traj.get('id', 'unknown')}, partial_success: {partial_success})"
+            )
+
+        # Strategy selection: use preferred_strategy if provided, otherwise select based on ratios
+        if preferred_strategy is not None:
+            # Use the preferred strategy directly
+            logger.trace(f"[PREF SAMPLER] Using preferred strategy: {preferred_strategy.value}")
+            result = self._execute_strategy(preferred_strategy, chosen_traj, use_partial_success)
+            if result is None:
+                logger.trace(f"[PREF SAMPLER] Preferred strategy {preferred_strategy.value} failed, returning None")
+                return None
+            rejected_traj, rejected_subsample_strategy, chosen_traj = result
+            strategy_used = preferred_strategy
+            attempt = 1  # Set attempt for preferred strategy path
+        else:
+            # Strategy selection with rebalancing on failure
+            strategies = []
+            if self.preference_strategy_ratio[0] > 0:
+                strategies.append((DataGenStrat.REWIND, self.preference_strategy_ratio[0]))
+            if self._has_suboptimal and self.preference_strategy_ratio[1] > 0:
+                strategies.append((DataGenStrat.SUBOPTIMAL, self.preference_strategy_ratio[1]))
+            if self.preference_strategy_ratio[2] > 0:
+                strategies.append((DataGenStrat.DIFFERENT_TASK, self.preference_strategy_ratio[2]))
+            if self.preference_strategy_ratio[3] > 0:
+                strategies.append((DataGenStrat.REVERSE_PROGRESS, self.preference_strategy_ratio[3]))
+
+            max_attempts = 10  # Limit retry attempts to prevent infinite loops
+            max_strategy_attempts = 3  # Maximum attempts per strategy before removing it
+            attempt = 0
+
+            # Track attempts per strategy
+            strategy_attempt_counts = {strat: 0 for strat, _ in strategies}
+
+            while rejected_traj is None and attempt < max_attempts:
+                attempt += 1
+
+                # Check if we have any strategies left
+                if not strategies:
+                    return None
+
+                # Rebalance probabilities based on remaining strategies
+                total_prob = sum(prob for _, prob in strategies)
+                if total_prob == 0:
+                    return None
+
+                # Normalize probabilities
+                normalized_strategies = [(strat, prob / total_prob) for strat, prob in strategies]
+
+                # Select strategy based on rebalanced probabilities
+                prob = random.random()
+                cumulative_prob = 0.0
+                selected_strategy = None
+
+                for strat, normalized_prob in normalized_strategies:
+                    cumulative_prob += normalized_prob
+                    if prob <= cumulative_prob:
+                        selected_strategy = strat
+                        break
+
+                # Log strategy attempt
+                logger.trace(
+                    f"[PREF SAMPLER] Attempt {attempt}/{max_attempts}: Trying strategy {selected_strategy.value if selected_strategy else 'None'}"
+                )
+
+                # Execute selected strategy
+                result = self._execute_strategy(selected_strategy, chosen_traj, use_partial_success)
+                if result is not None:
+                    rejected_traj, rejected_subsample_strategy, chosen_traj = result
+                    strategy_used = selected_strategy
+                    logger.trace(f"[PREF SAMPLER] Strategy {selected_strategy.value} succeeded on attempt {attempt}")
+                else:
+                    # Strategy failed - increment attempt count
+                    strategy_attempt_counts[selected_strategy] = strategy_attempt_counts.get(selected_strategy, 0) + 1
+                    failed_count = strategy_attempt_counts[selected_strategy]
+
+                    logger.trace(
+                        f"[PREF SAMPLER] Strategy {selected_strategy.value} failed (failure count: {failed_count}/{max_strategy_attempts})"
+                    )
+
+                    # Only remove strategy if it has failed max_strategy_attempts times
+                    if strategy_attempt_counts[selected_strategy] >= max_strategy_attempts:
+                        logger.trace(
+                            f"[PREF SAMPLER] Removing strategy {selected_strategy.value} after {max_strategy_attempts} consecutive failures"
+                        )
+                        strategies = [(strat, prob) for strat, prob in strategies if strat != selected_strategy]
+                        continue
+
+            # If we still don't have a sample after all attempts, return None
+            if rejected_traj is None:
+                logger.trace(
+                    f"[PREF SAMPLER] Failed to generate preference sample after {max_attempts} attempts - all strategies exhausted"
+                )
+                return None
+
+        chosen_subsample_strategy = "subsample_forward"
+        chosen_trajectory = self._get_traj_from_data(chosen_traj, subsample_strategy=chosen_subsample_strategy)
+
+        rejected_trajectory = self._get_traj_from_data(rejected_traj, subsample_strategy=rejected_subsample_strategy)
+
+        if rejected_trajectory is None or chosen_trajectory is None:
+            return None
+
+        # If our strategy is different task, make sure the rejected trajectory has 0 progress and 0 success labels
+        if strategy_used in [
+            DataGenStrat.DIFFERENT_TASK,
+            DataGenStrat.DIFFERENT_TASK_INSTRUCTION,
+        ]:
+            rejected_trajectory.target_progress = [0.0] * len(rejected_trajectory.target_progress)
+            if self.config.progress_loss_type.lower() == "discrete":
+                rejected_trajectory.target_progress = convert_continuous_to_discrete_bins(
+                    rejected_trajectory.target_progress, self.config.progress_discrete_bins
+                )
+            # Also set success labels to 0.0 (predict 0 success for different task trajectories)
+            if rejected_trajectory.success_label is not None:
+                rejected_trajectory.success_label = [0.0] * len(rejected_trajectory.success_label)
+
+        # Create preference sample structure
+        sample = PreferenceSample(
+            chosen_trajectory=chosen_trajectory,
+            rejected_trajectory=rejected_trajectory,
+            data_gen_strategy=strategy_used.value,
+        )
+        sample.resample_attempts = attempt
+        return sample
diff --git a/robometer/robometer/data/samplers/progress.py b/robometer/robometer/data/samplers/progress.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad0a0985513751cb7d0e9798a4fb3f6f4e26ab77
--- /dev/null
+++ b/robometer/robometer/data/samplers/progress.py
@@ -0,0 +1,176 @@
+from typing import Dict, Any, Optional
+
+import random
+import torch
+
+from robometer.data.dataset_types import ProgressSample, Trajectory
+from robometer.data.samplers.base import RBMBaseSampler
+from robometer.data.datasets.helpers import (
+    DataGenStrat,
+    load_embeddings_from_path,
+    convert_continuous_to_discrete_bins,
+)
+from robometer.utils.distributed import rank_0_print
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class ProgressSampler(RBMBaseSampler):
+    """Data generator for progress samples."""
+
+    def __init__(self, is_evaluation=False, **kwargs):
+        super().__init__(**kwargs)
+
+    def _generate_sample(self, item: Dict[str, Any], preferred_strategy: Optional[DataGenStrat] = None):
+        return self._create_progress_sample(item, preferred_strategy=preferred_strategy)
+
+    def _execute_strategy(self, strategy: DataGenStrat, traj: Dict[str, Any]) -> tuple[Dict[str, Any], str] | None:
+        """Execute a strategy to get processed trajectory.
+
+        Args:
+            strategy: The strategy to execute
+            traj: The trajectory to process
+
+        Returns:
+            Tuple of (processed_traj, subsample_strategy) or None if failed
+        """
+        if strategy == DataGenStrat.FORWARD_PROGRESS:
+            return (traj, "subsample_forward")
+        elif strategy == DataGenStrat.REVERSE_PROGRESS:
+            return (traj, "subsample_reverse")
+        elif strategy == DataGenStrat.REWIND:
+            return (traj, "subsample_rewind")
+        elif strategy == DataGenStrat.DIFFERENT_TASK_INSTRUCTION:
+            processed_traj = self._get_different_task_instruction(traj)
+            if processed_traj is None:
+                return None
+            return (processed_traj, "subsample_forward")
+        else:
+            return None
+
+    def _create_progress_sample(self, traj: Dict[str, Any], preferred_strategy: Optional[DataGenStrat] = None):
+        """Create a progress sample using normalized and rebalanced strategy selection.
+
+        Implements four strategies:
+        1. Different Task: Use trajectory from different task (progress set to 0.0)
+        2. Forward Progress: Sample with forward direction (start < middle < end)
+        3. Reverse Progress: Sample with reverse direction (end < middle < start)
+        4. Rewind: Sample with rewind direction (start < end < middle)
+        """
+        # Initialize variables for strategy selection
+        processed_traj = None
+        strategy_used = None
+        subsample_strategy = None
+
+        # Strategy selection: use preferred_strategy if provided, otherwise select based on ratios
+        if preferred_strategy is not None:
+            # Use the preferred strategy directly
+            logger.trace(f"[PROGRESS SAMPLER] Using preferred strategy: {preferred_strategy.value}")
+            result = self._execute_strategy(preferred_strategy, traj)
+            if result is None:
+                logger.trace(f"[PROGRESS SAMPLER] Preferred strategy {preferred_strategy.value} failed, returning None")
+                return None
+            processed_traj, subsample_strategy = result
+            strategy_used = preferred_strategy
+            attempt = 1  # Set attempt for preferred strategy path
+        else:
+            # Strategy setup with rebalancing on failure
+            # [different_task_instruction, forward_progress, reverse_progress, rewind]
+            strategies = [
+                (
+                    DataGenStrat.DIFFERENT_TASK_INSTRUCTION,
+                    self.config.progress_strategy_ratio[0] if len(self.config.progress_strategy_ratio) > 0 else 0.0,
+                ),
+                (
+                    DataGenStrat.FORWARD_PROGRESS,
+                    self.config.progress_strategy_ratio[1] if len(self.config.progress_strategy_ratio) > 1 else 0.0,
+                ),
+                (
+                    DataGenStrat.REVERSE_PROGRESS,
+                    self.config.progress_strategy_ratio[2] if len(self.config.progress_strategy_ratio) > 2 else 0.0,
+                ),
+                (
+                    DataGenStrat.REWIND,
+                    self.config.progress_strategy_ratio[3] if len(self.config.progress_strategy_ratio) > 3 else 0.0,
+                ),
+            ]
+
+            # Remove strategies with zero probability
+            strategies = [(strat, prob) for strat, prob in strategies if prob > 0]
+
+            max_attempts = 10  # Limit retry attempts to prevent infinite loops
+            attempt = 0
+
+            while processed_traj is None and attempt < max_attempts:
+                attempt += 1
+
+                # Check if we have any strategies left
+                if not strategies:
+                    return None
+
+                # Rebalance probabilities based on remaining strategies
+                total_prob = sum(prob for _, prob in strategies)
+                if total_prob == 0:
+                    return None
+
+                # Normalize probabilities
+                normalized_strategies = [(strat, prob / total_prob) for strat, prob in strategies]
+
+                # Select strategy based on rebalanced probabilities
+                prob = random.random()
+                cumulative_prob = 0.0
+                selected_strategy = None
+
+                for strat, normalized_prob in normalized_strategies:
+                    cumulative_prob += normalized_prob
+                    if prob <= cumulative_prob:
+                        selected_strategy = strat
+                        break
+
+                # Execute selected strategy
+                result = self._execute_strategy(selected_strategy, traj)
+                if result is not None:
+                    processed_traj, subsample_strategy = result
+                    strategy_used = selected_strategy
+                else:
+                    # Remove failed strategy and try again
+                    strategies = [(strat, prob) for strat, prob in strategies if strat != selected_strategy]
+                    continue
+
+            # If we still don't have a sample after all attempts, return None
+            if processed_traj is None:
+                logger.trace(
+                    f"[PROGRESS SAMPLER] Failed to generate progress sample after {max_attempts} attempts - all strategies exhausted"
+                )
+                return None
+
+        progress_traj = self._get_traj_from_data(processed_traj, subsample_strategy=subsample_strategy)
+
+        if progress_traj is None:
+            return None
+
+        # Handle special cases
+        if strategy_used in [DataGenStrat.DIFFERENT_TASK, DataGenStrat.DIFFERENT_TASK_INSTRUCTION]:
+            # We need to use the original task embeddings instead of the different task embeddings
+            if self.config.load_embeddings and traj.get("embeddings_path"):
+                progress_traj.text_embedding = load_embeddings_from_path(traj["embeddings_path"])["text_embedding"]
+            progress_traj.lang_vector = traj["lang_vector"]
+            progress_traj.task = traj["task"]
+            progress_traj.target_progress = [0.0] * len(progress_traj.target_progress)
+            if self.config.progress_loss_type.lower() == "discrete":
+                progress_traj.target_progress = convert_continuous_to_discrete_bins(
+                    progress_traj.target_progress, self.config.progress_discrete_bins
+                )
+            # Also set success labels to 0.0 (predict 0 success for different task trajectories)
+            if progress_traj.success_label is not None:
+                progress_traj.success_label = [0.0] * len(progress_traj.success_label)
+
+        strategy_value = strategy_used.value if isinstance(strategy_used, DataGenStrat) else strategy_used
+        sample = ProgressSample(
+            trajectory=progress_traj,
+            sample_type="progress",
+            data_gen_strategy=strategy_value,
+        )
+        sample.resample_attempts = attempt
+        return sample
diff --git a/robometer/robometer/data/scripts/preprocess_datasets.py b/robometer/robometer/data/scripts/preprocess_datasets.py
new file mode 100755
index 0000000000000000000000000000000000000000..ab6fea6a26284e1cd6b9c3a25fb3a5afd0e42419
--- /dev/null
+++ b/robometer/robometer/data/scripts/preprocess_datasets.py
@@ -0,0 +1,1091 @@
+#!/usr/bin/env python3
+"""
+Dataset preprocessing script that creates index-based caches for fast trajectory access.
+Uses HuggingFace's .map() for efficient processing and saves trajectory indices.
+Creates one unified cache for each dataset/subset split.
+"""
+
+import datetime
+import shutil
+import json
+import os
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+from dataclasses import dataclass, field
+from typing import Any
+
+import numpy as np
+import decord  # type: ignore
+import torch
+from pyrallis import wrap
+from tqdm import tqdm
+from sentence_transformers import SentenceTransformer
+from transformers import AutoModel, AutoImageProcessor
+from PIL import Image
+
+from datasets import Dataset, DatasetDict, Video, load_dataset
+from robometer.utils.distributed import rank_0_print
+from robometer.utils.embedding_utils import compute_video_embeddings, compute_text_embeddings
+
+# VIDEO_ERROR_PRINTED = False
+# maps subsets to functions that filter the dataset. If true, the example is dropped.
+soar_bad_trajectories = [
+    # success
+    "ca7fec9e-0c19-4e43-8503-ac6cf274a5f3",
+    "96913909-04c5-4b95-91b4-42f59dffa90c",
+    "521ea359-caf2-449d-b9c7-73574ce6de01",
+    "a0c119ab-2cd4-4f65-a0b4-8fb736322bf0",
+    "116411d4-fb67-4cde-ba17-e45edf60d922",
+    "db1582fc-76c9-4069-ac6d-a296e574900c",
+    # failure
+    "e0cd2f7e-cfc6-46fb-9880-d65318709642",
+    "2dfaf622-9aac-4965-b492-dcb5c6da8c62",
+    "88ea7b35-c9f2-494f-9cc4-4c9a5c8def6d",
+    "52cdf32b-702d-417c-b29f-9c1cbe022755",
+    "c7868d1d-6e99-4490-b9ad-41b0e9c89069",
+    "ee6735c6-9d11-4412-9b34-11a9e6d67646",
+    "fed7da83-a9d6-4c71-acd6-5b6a212b1e6a",
+    "d31ad630-14d2-47f2-b4a9-0579247b72b8",
+    "31f3679d-2242-4d98-89d9-6a05b06fc807",
+    "a49d8c5f-bf02-42f5-b130-1f3d53968120",
+]
+filters = {
+    "jesbu1/molmoact_rfm/molmoact_dataset_tabletop": lambda x: "load the bowl" in x["task"].lower(),
+    "jesbu1/galaxea_rfm/galaxea_part1_r1_lite": lambda x: all(
+        word in x["task"].lower() for word in ["return", "to", "initial", "position"]
+    ),
+    "jesbu1/galaxea_rfm/galaxea_part2_r1_lite": lambda x: all(
+        word in x["task"].lower() for word in ["return", "to", "initial", "position"]
+    ),
+    "jesbu1/galaxea_rfm/galaxea_part3_r1_lite": lambda x: all(
+        word in x["task"].lower() for word in ["return", "to", "initial", "position"]
+    ),
+    "jesbu1/galaxea_rfm/galaxea_part4_r1_lite": lambda x: all(
+        word in x["task"].lower() for word in ["return", "to", "initial", "position"]
+    ),
+    "jesbu1/galaxea_rfm/galaxea_part5_r1_lite": lambda x: all(
+        word in x["task"].lower() for word in ["return", "to", "initial", "position"]
+    ),
+    "jesbu1/soar_rfm/soar_rfm": lambda x: x["id"] in soar_bad_trajectories,
+    "jesbu1/auto_eval_rfm/auto_eval_rfm": lambda x: x["frames_shape"][0]
+    <= 5,  # some episodes are too short and are likely poor/faulty success detections
+    "anqil/rh20t_subset_rfm/rh20t_human": lambda x: x["frames_shape"][0] < 16,  # some episodes are too short
+    "anqil/rh20t_subset_rfm/rh20t_robot": lambda x: x["frames_shape"][0] < 16,  # some episodes are too short
+}
+
+
+@dataclass
+class DataPreprocessConfig:
+    """Configuration for data loading and processing."""
+
+    # Dataset paths and subsets
+    train_datasets: list[str] = field(default_factory=lambda: [], metadata={"help": "List of training dataset names"})
+    train_subsets: list[list[str]] = field(
+        default_factory=lambda: [[]], metadata={"help": "List of training dataset subsets"}
+    )
+    eval_datasets: list[str] = field(default_factory=lambda: [], metadata={"help": "List of evaluation dataset names"})
+    eval_subsets: list[list[str]] = field(
+        default_factory=lambda: [[]], metadata={"help": "List of evaluation dataset subsets"}
+    )
+
+    # Video processing parameters
+    max_frames_for_preprocessing: int = field(
+        default=64, metadata={"help": "Maximum number of frames to extract from videos for preprocessing"}
+    )
+    video_frame_sampling: str = field(
+        default="uniform", metadata={"help": "Frame sampling strategy: 'uniform', 'random', 'start', 'end'"}
+    )
+    num_proc: int = field(default=1, metadata={"help": "Number of processes for dataset processing"})
+    force_reprocess: bool = field(
+        default=False, metadata={"help": "Force reprocessing of datasets even if cache exists"}
+    )
+    num_threads: int = field(default=36, metadata={"help": "Number of threads for dataset processing"})
+    cache_dir: str = field(default="", metadata={"help": "Directory to store processed dataset caches"})
+
+    # Embedding preprocessing parameters
+    precompute_embeddings: bool = field(
+        default=False, metadata={"help": "Whether to precompute DINOv2 and sentence transformer embeddings"}
+    )
+    embeddings_cache_dir: str = field(
+        default="embeddings_cache", metadata={"help": "Directory to save precomputed embeddings"}
+    )
+    dinov2_model: str = field(default="facebook/dinov2-base", metadata={"help": "DINOv2 model for video embeddings"})
+    sentence_model: str = field(
+        default="sentence-transformers/all-MiniLM-L12-v2",
+        metadata={"help": "Sentence transformer model for text embeddings"},
+    )
+    embedding_batch_size: int = field(default=32, metadata={"help": "Batch size for embedding computation"})
+
+
+class DatasetPreprocessor:
+    """Unified preprocessor for all datasets with individual caching per dataset/subset."""
+
+    def __init__(self, config):
+        self.config = config
+
+        # Dataset storage - store individual datasets
+        self.datasets: dict[str, Dataset] = {}  # key: "dataset_path/subset"
+        self.dataset_indices: dict[str, dict] = {}  # key: "dataset_path/subset", value: index mappings
+
+        # Initialize embedding models if precompute_embeddings is enabled
+        self.dinov2_model = None
+        self.dinov2_processor = None
+        self.sentence_model = None
+
+        if config.precompute_embeddings:
+            rank_0_print("🚀 Initializing embedding models...")
+
+            # Initialize DINOv2 for video embeddings
+            self.dinov2_model = AutoModel.from_pretrained(config.dinov2_model)
+            self.dinov2_processor = AutoImageProcessor.from_pretrained(config.dinov2_model, use_fast=True)
+            self.dinov2_model.eval()
+
+            # Initialize Sentence Transformer for text embeddings
+            self.sentence_model = SentenceTransformer(config.sentence_model)
+
+            # Move to GPU if available
+            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+            self.dinov2_model = self.dinov2_model.to(device)
+            self.sentence_model = self.sentence_model.to(device)
+
+            rank_0_print(f"✅ Embedding models initialized on {device}")
+
+    def preprocess_datasets(self):
+        """Preprocess each dataset individually and create index-based caches."""
+        rank_0_print("\n🔧 Preprocessing all datasets...")
+
+        # Collect all dataset/subset combinations
+        all_datasets = []
+
+        # Add training datasets
+        for dataset_path, dataset_subsets in zip(self.config.train_datasets, self.config.train_subsets, strict=False):
+            if isinstance(dataset_subsets, str):
+                dataset_subsets = [dataset_subsets]
+            for subset in dataset_subsets:
+                all_datasets.append((dataset_path, subset))
+
+        # Add evaluation datasets
+        for dataset_path, dataset_subsets in zip(self.config.eval_datasets, self.config.eval_subsets, strict=False):
+            if isinstance(dataset_subsets, str):
+                dataset_subsets = [dataset_subsets]
+            for subset in dataset_subsets:
+                all_datasets.append((dataset_path, subset))
+
+        # Show which datasets are already preprocessed
+        self._show_preprocessed_datasets(all_datasets)
+
+        # Process each dataset and its associated subsets
+        for i, (dataset_path, subset) in enumerate(all_datasets):
+            rank_0_print(f"\n📚 Processing dataset {i + 1}/{len(all_datasets)}: {dataset_path}/{subset}")
+
+            # Create individual cache key
+            cache_key = f"{dataset_path}/{subset}"
+            individual_cache_dir = os.path.join(self.config.cache_dir, cache_key.replace("/", "_").replace(":", "_"))
+
+            # Check if already processed
+            if os.path.exists(individual_cache_dir) and not self.config.force_reprocess:
+                rank_0_print(f"    ✅ Cache already exists at {individual_cache_dir}, loading...")
+                self._load_individual_cache(individual_cache_dir, cache_key)
+                continue
+
+            if os.path.exists(individual_cache_dir) and self.config.force_reprocess:
+                rank_0_print(f"    ❌ Force reprocessing enabled, deleting cache at {individual_cache_dir}")
+                shutil.rmtree(individual_cache_dir)
+
+            # Load and process individual dataset
+            try:
+                dataset = self._load_dataset_from_path(dataset_path, subset)
+
+                # Handle DatasetDict
+                if isinstance(dataset, DatasetDict):
+                    if "train" in dataset:
+                        dataset = dataset["train"]
+                    else:
+                        rank_0_print(f"    ⚠️  Warning: No 'train' split found in {dataset_path}/{subset}")
+                        continue
+
+                rank_0_print(f"    📥 Loaded {len(dataset)} trajectories from {dataset_path}/{subset}")
+
+                # Process this individual dataset
+                processed_dataset, indices = self._process_individual_dataset(dataset, individual_cache_dir, cache_key)
+
+                # Store processed dataset and indices
+                self.datasets[cache_key] = processed_dataset
+                self.dataset_indices[cache_key] = indices
+
+                # Save individual cache
+                self._save_individual_cache(individual_cache_dir, processed_dataset, indices, dataset_path, subset)
+
+                rank_0_print(f"    ✅ Successfully processed and cached {dataset_path}")
+
+            except Exception as e:
+                rank_0_print(f"    ❌ Failed to process {dataset_path}: {e}")
+                continue
+
+        if not self.datasets:
+            raise ValueError("No datasets were successfully processed")
+
+        rank_0_print(f"✅ Successfully processed {len(self.datasets)} datasets")
+
+        # Log summary of processed datasets
+        total_trajectories = sum(len(dataset) for dataset in self.datasets.values())
+        rank_0_print(f"📊 Total trajectories across all datasets: {total_trajectories}")
+
+        for cache_key, dataset in self.datasets.items():
+            rank_0_print(f"  📚 {cache_key}: {len(dataset)} trajectories")
+
+        # Show final status summary
+        self._show_final_status_summary(all_datasets)
+
+    def _process_individual_dataset(self, dataset: Dataset, cache_dir: str, cache_key: str):
+        """Process a single dataset and build its index mappings."""
+        # Process videos and build indices
+        # Only cast to Video with decode=True for the .map path. The threaded path will
+        # open and close readers per-sample to avoid keeping many files open at once.
+        if self.config.num_threads > 1:
+            processed_dataset, indices = self._process_dataset_videos_threaded(dataset, cache_dir, cache_key)
+        else:
+            dataset = dataset.cast_column("frames_video", Video(decode=True))
+            processed_dataset, indices = self._process_dataset_videos_map(dataset, cache_dir, cache_key)
+
+        return processed_dataset, indices
+
+    def _compute_video_embeddings(self, frames_array: np.ndarray) -> torch.Tensor:
+        """
+        Compute DINOv2 embeddings for video frames.
+
+        Args:
+            frames_array: Video frames as numpy array (T, H, W, C)
+
+        Returns:
+            Video embeddings as torch tensor (T, D) where D is embedding dimension
+        """
+        return compute_video_embeddings(
+            frames_array,
+            self.dinov2_model,
+            self.dinov2_processor,
+            batch_size=self.config.embedding_batch_size,
+            use_autocast=True,
+        )
+
+    def _compute_text_embeddings(self, text: str) -> torch.Tensor:
+        """
+        Compute sentence transformer embeddings for text.
+
+        Args:
+            text: Text description
+
+        Returns:
+            Text embedding as torch tensor (D,) where D is embedding dimension
+        """
+        return compute_text_embeddings(
+            text,
+            self.sentence_model,
+            use_autocast=True,
+            show_progress_bar=False,
+        )
+
+    def _save_embeddings(self, video_embeddings: torch.Tensor, text_embedding: torch.Tensor, embeddings_path: str):
+        """
+        Save video and text embeddings to a .pt file.
+
+        Args:
+            video_embeddings: Video embeddings tensor (T, D)
+            text_embedding: Text embedding tensor (D,)
+            embeddings_path: Path to save the .pt file
+        """
+        os.makedirs(os.path.dirname(embeddings_path), exist_ok=True)
+
+        embeddings_data = {
+            "video_embeddings": video_embeddings,
+            "text_embedding": text_embedding,
+            "video_shape": video_embeddings.shape,
+            "text_shape": text_embedding.shape,
+        }
+
+        torch.save(embeddings_data, embeddings_path)
+
+    def _process_embeddings_for_example(
+        self, frames_array: np.ndarray, task_text: str, idx: int, example_id: str, embeddings_dir: str
+    ) -> tuple[str, tuple, tuple]:
+        """
+        Process embeddings for a single example.
+
+        Args:
+            frames_array: Video frames as numpy array
+            task_text: Task description text
+            idx: Example index
+            example_id: Example ID
+            embeddings_dir: Directory to save embeddings
+
+        Returns:
+            Tuple of (embeddings_filepath, video_embedding_shape, text_embedding_shape)
+        """
+        if not self.config.precompute_embeddings:
+            return None, None, None
+
+        # Create embeddings directory
+        os.makedirs(embeddings_dir, exist_ok=True)
+
+        # Compute video embeddings
+        video_embeddings = self._compute_video_embeddings(frames_array)
+
+        # Compute text embeddings
+        text_embedding = self._compute_text_embeddings(task_text)
+
+        # Save embeddings
+        embeddings_filename = f"trajectory_{idx:06d}_{example_id}_embeddings.pt"
+        embeddings_filepath = os.path.join(embeddings_dir, embeddings_filename)
+        self._save_embeddings(video_embeddings, text_embedding, embeddings_filepath)
+
+        return embeddings_filepath, video_embeddings.shape, text_embedding.shape
+
+    def _process_dataset_videos_map(self, dataset, cache_dir: str, cache_key: str):
+        """
+        Process dataset frames using .map() method for efficient on-the-fly processing.
+        Also builds index mappings during the same pass to avoid multiple iterations.
+        Frames are saved as .npz files and only file paths are stored in the dataset.
+
+        Args:
+            dataset: HuggingFace dataset containing trajectories
+
+        Returns:
+            Dataset with processed frame paths and metadata
+        """
+        raise NotImplementedError("This method is not properly implemented rn, use threaded version with 1 process.")
+        # Check if frames are already processed (npz file paths)
+        sample_item = dataset[0]
+        frames_data = sample_item.get("frames")
+
+        if isinstance(frames_data, str) and frames_data.endswith(".npz") and not self.config.force_reprocess:
+            rank_0_print("Frames already processed as npz file paths, skipping processing.")
+            return dataset, {}  # Return empty index mappings if already processed
+        elif self.config.force_reprocess and isinstance(frames_data, str) and frames_data.endswith(".npz"):
+            rank_0_print("Force reprocessing enabled. Reprocessing frames despite being already processed.")
+
+        rank_0_print("Processing video frames into npz files and building index mappings...")
+
+        # Debug: Check the dataset structure
+        sample_item = dataset[0]
+        rank_0_print(f"Sample dataset item keys: {list(sample_item.keys())}")
+        rank_0_print(f"Sample item structure: {sample_item}")
+
+        # Initialize index mappings
+        robot_trajectories = []
+        human_trajectories = []
+        optimal_by_task = {}
+        suboptimal_by_task = {}
+        quality_indices = {}
+        task_indices = {}
+        source_indices = {}
+        partial_success_indices = {}
+
+        frames_dir = os.path.join(cache_dir, "frames")
+        embeddings_dir = os.path.join(cache_dir, self.config.embeddings_cache_dir)
+        os.makedirs(frames_dir, exist_ok=True)
+        os.makedirs(embeddings_dir, exist_ok=True)
+
+        def process_videos_and_build_indices(example, idx):
+            """Process frames and build index mappings in a single pass."""
+            # Debug: Log what we're processing
+            if idx < 3:  # Only log first 5 examples to avoid spam
+                rank_0_print(f"Processing example {idx}: {example['id']} - {example['task']}")
+
+            # filter the example if needed
+            should_drop = filters.get(cache_key, lambda x: False)(example)
+            if should_drop:
+                rank_0_print(f"Dropping example {idx}: {example['id']} - {example['task']}")
+                return {"frames": None, "frames_processed": False}
+
+            # Get the video reader object from the Video feature
+            frames = example.get("frames_video")
+            if frames is None:
+                rank_0_print(f"Warning: No frames_video for example {idx}")
+                return {"frames": None, "frames_processed": False}
+
+            # Process video frames using the _preprocess_videos method
+            frames_array = self._preprocess_videos(frames, self.config.max_frames_for_preprocessing)
+
+            if frames_array.size == 0:
+                rank_0_print(f"Warning: No frames processed for example {idx}")
+                return {"frames": None, "frames_processed": False}
+
+            # Save frames as npz file
+            frames_filename = f"trajectory_{example['id']}.npz"
+            frames_filepath = os.path.join(frames_dir, frames_filename)
+
+            # Store file path and metadata in dataset (not the actual frames)
+            example["frames"] = frames_filepath  # Store path to npz file
+            example["frames_shape"] = frames_array.shape
+            example["num_frames"] = frames_array.shape[0] if len(frames_array.shape) > 0 else 0
+            example["frames_processed"] = True
+
+            # Compute and save embeddings if enabled
+            embeddings_filepath, video_emb_shape, text_emb_shape = self._process_embeddings_for_example(
+                frames_array, example.get("task", ""), idx, example["id"], embeddings_dir
+            )
+
+            if embeddings_filepath is not None:
+                # Store embeddings path in dataset
+                example["embeddings_path"] = embeddings_filepath
+                example["video_embedding_shape"] = video_emb_shape
+                example["text_embedding_shape"] = text_emb_shape
+
+            # BUILD INDEX MAPPINGS DURING THE SAME PASS
+            # Debug: Log the values we're extracting
+            if idx < 3:
+                rank_0_print(
+                    f"  Example {idx} - is_robot: {example.get('is_robot', True)}, task: {example.get('task', 'unknown')}, quality: {example.get('quality_label', 'successful')}"
+                )
+
+            # Robot/Human trajectories
+            if example.get("is_robot", True):
+                robot_trajectories.append(idx)
+            else:
+                human_trajectories.append(idx)
+
+            # Quality-based indices
+            quality = example["quality_label"]
+            if quality not in quality_indices:
+                quality_indices[quality] = []
+            quality_indices[quality].append(idx)
+
+            # Task-based indices
+            task = example["task"]
+            if task not in task_indices:
+                task_indices[task] = []
+            task_indices[task].append(idx)
+
+            # Source-based indices
+            source = example["data_source"]
+            if source not in source_indices:
+                source_indices[source] = []
+            source_indices[source].append(idx)
+
+            # Partial success-based indices
+            partial_success = example.get("partial_success", None)
+            if partial_success is not None and quality == "failure":  # only record partial success for failure cases
+                if partial_success not in partial_success_indices:
+                    partial_success_indices[partial_success] = []
+                partial_success_indices[partial_success].append(idx)
+
+            # Optimal/Suboptimal by task
+            if task not in optimal_by_task:
+                optimal_by_task[task] = []
+                suboptimal_by_task[task] = []
+
+            if "frames_video" in example:
+                del example["frames_video"]
+
+            if quality in ["successful", "optimal"]:
+                optimal_by_task[task].append(idx)
+            elif quality in ["suboptimal", "failed", "failure"]:
+                suboptimal_by_task[task].append(idx)
+
+            # Save frames with metadata
+            np.savez_compressed(
+                frames_filepath,
+                frames=frames_array,
+                shape=frames_array.shape,
+                num_frames=frames_array.shape[0] if len(frames_array.shape) > 0 else 0,
+            )
+
+            return example
+
+        # Apply the mapping function to the dataset
+        processed_dataset = dataset.map(
+            process_videos_and_build_indices,
+            with_indices=True,
+            desc="Processing video frames and building indices",
+            num_proc=self.config.num_proc,
+        )
+
+        # Filter out dropped examples (those with frames=None and frames_processed=False)
+        original_length = len(processed_dataset)
+
+        # Build a list of which indices were kept (had frames_processed=True)
+        kept_indices = []
+        for idx in range(original_length):
+            if processed_dataset[idx].get("frames_processed", False):
+                kept_indices.append(idx)
+
+        # Filter the dataset
+        processed_dataset = processed_dataset.filter(lambda x: x.get("frames_processed", False))
+        num_filtered = original_length - len(processed_dataset)
+
+        if num_filtered > 0:
+            rank_0_print(f"Filtered out {num_filtered} trajectories")
+
+            # Build mapping from old indices to new indices
+            old_to_new_idx = {old_idx: new_idx for new_idx, old_idx in enumerate(kept_indices)}
+
+            # Remap all indices
+            robot_trajectories = [old_to_new_idx[idx] for idx in robot_trajectories if idx in old_to_new_idx]
+            human_trajectories = [old_to_new_idx[idx] for idx in human_trajectories if idx in old_to_new_idx]
+
+            quality_indices = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in quality_indices.items()
+            }
+
+            task_indices = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in task_indices.items()
+            }
+
+            source_indices = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in source_indices.items()
+            }
+
+            partial_success_indices = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in partial_success_indices.items()
+            }
+
+            optimal_by_task = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in optimal_by_task.items()
+            }
+
+            suboptimal_by_task = {
+                key: [old_to_new_idx[idx] for idx in indices if idx in old_to_new_idx]
+                for key, indices in suboptimal_by_task.items()
+            }
+
+        # Log the built indices
+        rank_0_print(f"Built index mappings for {cache_key}:")
+        rank_0_print(f"  Robot trajectories: {len(robot_trajectories)}")
+        rank_0_print(f"  Human trajectories: {len(human_trajectories)}")
+        rank_0_print(f"  Tasks: {len(task_indices)}")
+        rank_0_print(f"  Quality labels: {len(quality_indices)}")
+        rank_0_print(f"  Data sources: {len(source_indices)}")
+        rank_0_print(f"  Partial success indices: {len(partial_success_indices)}")
+
+        return processed_dataset, {
+            "robot_trajectories": robot_trajectories,
+            "human_trajectories": human_trajectories,
+            "optimal_by_task": optimal_by_task,
+            "suboptimal_by_task": suboptimal_by_task,
+            "quality_indices": quality_indices,
+            "task_indices": task_indices,
+            "source_indices": source_indices,
+            "partial_success_indices": partial_success_indices,
+        }
+
+    def _process_dataset_videos_threaded(self, dataset, cache_dir: str, cache_key: str):
+        """
+        Threaded implementation to process frames and build indices, saving npz files concurrently.
+        """
+        from concurrent.futures import ThreadPoolExecutor, as_completed
+
+        from datasets import Sequence, Value
+
+        # Initialize index mappings
+        robot_trajectories = []
+        human_trajectories = []
+        optimal_by_task = {}
+        suboptimal_by_task = {}
+        quality_indices = {}
+        task_indices = {}
+        source_indices = {}
+        partial_success_indices = {}
+
+        frames_dir = os.path.join(cache_dir, "frames")
+        embeddings_dir = os.path.join(cache_dir, self.config.embeddings_cache_dir)
+        os.makedirs(frames_dir, exist_ok=True)
+        os.makedirs(embeddings_dir, exist_ok=True)
+
+        def process_one(idx: int):
+            global VIDEO_ERROR_PRINTED
+            # Fetch example inside the worker to avoid holding many decoded readers
+            example: dict[str, Any] = dataset[idx]
+            frames_src = example.get("frames_video")
+            if frames_src is None:
+                return idx, None, None, None, None, None
+
+            # If the source is a path string, open with a lightweight reader
+            try:
+                # Try decord first (fastest for video decoding)
+                vr = decord.VideoReader(frames_src, num_threads=1)
+                total_frames = len(vr)
+
+                # Sample frames efficiently
+                if total_frames <= self.config.max_frames_for_preprocessing:
+                    frame_indices = list(range(total_frames))
+                else:
+                    # Uniform sampling
+                    frame_indices = [
+                        int(i * total_frames / self.config.max_frames_for_preprocessing)
+                        for i in range(self.config.max_frames_for_preprocessing)
+                    ]
+
+                frames_array = vr.get_batch(frame_indices).asnumpy()
+
+                del vr
+
+            except Exception as e:
+                rank_0_print(f"Error in _process_one: {e}")
+                return idx, None, None, None, None, None
+
+            # Save frames as npz file
+            frames_filename = f"trajectory_{example['id']}.npz"
+            frames_filepath = os.path.join(frames_dir, frames_filename)
+            np.savez_compressed(
+                frames_filepath,
+                frames=frames_array,
+                shape=frames_array.shape,
+                num_frames=frames_array.shape[0] if len(frames_array.shape) > 0 else 0,
+            )
+
+            # Process embeddings if enabled
+            embeddings_filepath, video_emb_shape, text_emb_shape = self._process_embeddings_for_example(
+                frames_array, example.get("task", ""), idx, example["id"], embeddings_dir
+            )
+
+            # Return minimal info to update record and indices
+            return idx, frames_filepath, frames_array.shape, embeddings_filepath, video_emb_shape, text_emb_shape
+
+        # Prepare indices only; fetch rows inside workers to limit open file handles
+        indices_only = list(range(len(dataset)))
+        # indices_only = list(range(100))
+        _ = tqdm(indices_only, total=len(indices_only), desc=f"Indexing {cache_key}", unit="traj", leave=False)
+
+        # Concurrently convert and write npz for all examples, collecting results
+        idx_to_data = {}
+        with ThreadPoolExecutor(max_workers=self.config.num_threads) as executor:
+            futures = [executor.submit(process_one, i) for i in indices_only]
+            for fut in tqdm(
+                as_completed(futures), total=len(futures), desc=f"Processing {cache_key}", unit="traj", leave=False
+            ):
+                idx, frames_path, shape, embeddings_path, video_emb_shape, text_emb_shape = fut.result()
+
+                if frames_path is not None:
+                    idx_to_data[idx] = (frames_path, shape, embeddings_path, video_emb_shape, text_emb_shape)
+
+        # Update dataset with saved paths and build indices
+        updated_rows = []
+        kept_indices = []  # Track which original indices were kept
+
+        for i in tqdm(indices_only, total=len(indices_only), desc=f"Finalizing {cache_key}", unit="traj", leave=False):
+            ex = dataset[i]
+            if i in idx_to_data:
+                frames_path, shape, embeddings_path, video_emb_shape, text_emb_shape = idx_to_data[i]
+                ex = dict(ex)
+                ex["frames"] = frames_path
+                ex["frames_shape"] = shape
+                ex["num_frames"] = shape[0] if len(shape) > 0 else 0
+                ex["frames_processed"] = True
+
+                should_drop = filters.get(cache_key, lambda x: False)(ex)
+                if should_drop:
+                    # remove the frames_path and then continue processing others
+                    os.remove(frames_path)
+                    print(f"Removed frames_path {frames_path} for {ex['id']} because it was filtered out")
+                    continue
+
+                # Add embedding information if available
+                if embeddings_path is not None:
+                    ex["embeddings_path"] = embeddings_path
+                    ex["video_embedding_shape"] = video_emb_shape
+                    ex["text_embedding_shape"] = text_emb_shape
+
+                # Track that this index was kept
+                kept_indices.append(i)
+
+                # Build indices using the NEW index (position in updated_rows)
+                new_idx = len(updated_rows)
+
+                if ex.get("is_robot", True):
+                    robot_trajectories.append(new_idx)
+                else:
+                    human_trajectories.append(new_idx)
+
+                quality = ex.get("quality_label", "successful")
+                quality_indices.setdefault(quality, []).append(new_idx)
+
+                task = ex.get("task", "unknown")
+                task_indices.setdefault(task, []).append(new_idx)
+
+                source = ex.get("data_source", "unknown")
+                source_indices.setdefault(source, []).append(new_idx)
+
+                partial_success = ex.get("partial_success", None)
+                if (
+                    partial_success is not None and quality == "failure"
+                ):  # only record partial success for failure cases
+                    partial_success_indices.setdefault(partial_success, []).append(new_idx)
+
+                if task not in optimal_by_task:
+                    optimal_by_task[task] = []
+                    suboptimal_by_task[task] = []
+                if quality in ["successful", "optimal"]:
+                    optimal_by_task[task].append(new_idx)
+                elif quality in ["suboptimal", "failed", "failure"]:
+                    suboptimal_by_task[task].append(new_idx)
+
+                # Drop any lingering video readers from rows to free FDs
+                if isinstance(ex, dict):
+                    ex.pop("frames_video", None)
+
+                updated_rows.append(ex)
+
+        # Report filtering stats
+        num_filtered = len(indices_only) - len(kept_indices)
+        if num_filtered > 0:
+            rank_0_print(f"Filtered out {num_filtered} trajectories")
+
+        # Create a new Dataset from updated_rows
+        processed_dataset = Dataset.from_list(updated_rows)
+        processed_dataset = processed_dataset.cast_column("lang_vector", Sequence(feature=Value("float32")))
+
+        # Log the built indices
+        rank_0_print(f"Built index mappings for {cache_key} (threaded):")
+        rank_0_print(f"  Robot trajectories: {len(robot_trajectories)}")
+        rank_0_print(f"  Human trajectories: {len(human_trajectories)}")
+        rank_0_print(f"  Tasks: {len(task_indices)}")
+        rank_0_print(f"  Quality labels: {len(quality_indices)}")
+        rank_0_print(f"  Data sources: {len(source_indices)}")
+        rank_0_print(f"  Partial success indices: {len(partial_success_indices)}")
+
+        return processed_dataset, {
+            "robot_trajectories": robot_trajectories,
+            "human_trajectories": human_trajectories,
+            "optimal_by_task": optimal_by_task,
+            "suboptimal_by_task": suboptimal_by_task,
+            "quality_indices": quality_indices,
+            "task_indices": task_indices,
+            "source_indices": source_indices,
+            "partial_success_indices": partial_success_indices,
+        }
+
+    def _save_individual_cache(
+        self,
+        cache_dir: str,
+        processed_dataset: Dataset,
+        indices: dict,
+        dataset_path: str,
+        subset: str,
+    ):
+        """Save the processed dataset and index mappings for an individual dataset."""
+        # Create cache directory
+        os.makedirs(cache_dir, exist_ok=True)
+
+        # Save the processed dataset
+        dataset_cache_dir = os.path.join(cache_dir, "processed_dataset")
+        processed_dataset.save_to_disk(dataset_cache_dir)
+
+        # Save index mappings
+        index_mappings = indices
+
+        mappings_file = os.path.join(cache_dir, "index_mappings.json")
+        with open(mappings_file, "w") as f:
+            json.dump(index_mappings, f, indent=2)
+
+        # Save dataset info
+        dataset_info = {
+            "dataset_path": dataset_path,
+            "subset": subset,
+            "total_trajectories": len(processed_dataset),
+            "cache_timestamp": str(datetime.datetime.now()),
+            "config_hash": self._get_config_hash(),
+        }
+
+        info_file = os.path.join(cache_dir, "dataset_info.json")
+        with open(info_file, "w") as f:
+            json.dump(dataset_info, f, indent=2)
+
+        rank_0_print(f"Individual cache saved to {cache_dir}")
+
+    def _get_config_hash(self) -> str:
+        """Generate a hash of the relevant config parameters."""
+        import hashlib
+
+        config_str = str(self.config.max_frames_for_preprocessing)
+        return hashlib.md5(config_str.encode()).hexdigest()
+
+    def _load_individual_cache(self, cache_dir: str, cache_key: str):
+        """Load a pre-processed dataset and its index mappings from a cache directory."""
+        dataset_cache_dir = os.path.join(cache_dir, "processed_dataset")
+        if not os.path.exists(dataset_cache_dir):
+            raise FileNotFoundError(f"Processed dataset not found at {dataset_cache_dir}")
+
+        # Load the processed dataset
+        self.datasets[cache_key] = Dataset.load_from_disk(dataset_cache_dir)
+
+        # Load index mappings
+        mappings_file = os.path.join(cache_dir, "index_mappings.json")
+        if not os.path.exists(mappings_file):
+            raise FileNotFoundError(f"Index mappings not found at {mappings_file}")
+
+        with open(mappings_file) as f:
+            self.dataset_indices[cache_key] = json.load(f)
+
+        # Log loaded cache
+        rank_0_print(f"  📂 Loaded individual cache from {cache_dir}")
+
+    def get_combined_indices(self):
+        """Get combined index mappings from all individual datasets."""
+        if not self.dataset_indices:
+            return {}
+
+        # Combine indices from all datasets
+        combined_indices = {
+            "robot_trajectories": [],
+            "human_trajectories": [],
+            "optimal_by_task": {},
+            "suboptimal_by_task": {},
+            "quality_indices": {},
+            "task_indices": {},
+            "source_indices": {},
+            "partial_success_indices": {},
+        }
+
+        # Track offset for each dataset
+        offset = 0
+
+        for cache_key, indices in self.dataset_indices.items():
+            # Adjust indices by adding offset
+            for key in combined_indices:
+                if key in indices:
+                    if isinstance(indices[key], list):
+                        # For list indices, add offset
+                        combined_indices[key].extend([idx + offset for idx in indices[key]])
+                    elif isinstance(indices[key], dict):
+                        # For dict indices, add offset to values
+                        if key not in combined_indices[key]:
+                            combined_indices[key] = {}
+                        for subkey, subindices in indices[key].items():
+                            if subkey not in combined_indices[key]:
+                                combined_indices[key][subkey] = []
+                            combined_indices[key][subkey].extend([idx + offset for idx in subindices])
+
+            # Update offset for next dataset
+            if cache_key in self.datasets:
+                offset += len(self.datasets[cache_key])
+
+        return combined_indices
+
+    def _load_dataset_from_path(self, dataset_path: str, subset: str):
+        """Load dataset from path with proper video handling."""
+        if "/" in dataset_path and not os.path.exists(dataset_path):
+            # Loading from HuggingFace Hub - handle video paths
+            rank_0_print(f"Loading dataset: {dataset_path}")
+
+            dataset_root = os.environ.get("ROBOMETER_DATASET_PATH", "")
+            if not dataset_root:
+                raise ValueError(
+                    "ROBOMETER_DATASET_PATH not set. "
+                    "Set it to the directory containing your downloaded datasets. "
+                    "Example: export ROBOMETER_DATASET_PATH=/path/to/your/datasets"
+                )
+
+            dataset_name = dataset_path.split("/")[-1]
+
+            def patch_path(old_path):
+                root_dir = f"{dataset_root}/{dataset_name}"
+                return f"{root_dir}/{old_path}"  # e.g., "./videos/trajectory_0000.mp4"
+
+            # Load dataset
+            dataset = load_dataset(dataset_path, name=subset, split="train")
+
+            # dataset = dataset.select(range(100))
+
+            # Just patch the paths, don't decode videos yet
+            dataset = dataset.map(
+                lambda x: {"frames_video": patch_path(x["frames"]), "frames_path": patch_path(x["frames"])}
+            )
+            return dataset
+        else:
+            # Load from local disk
+            dataset = load_dataset(dataset_path)
+            return dataset
+
+    def _show_preprocessed_datasets(self, all_datasets: list[str]):
+        """
+        Show which datasets are already preprocessed and which are not.
+        This helps avoid re-processing already cached datasets.
+        """
+        rank_0_print("\n🔍 Checking for pre-existing caches...")
+
+        cached_count = 0
+        total_count = len(all_datasets)
+
+        for dataset_path, subset in all_datasets:
+            cache_key = f"{dataset_path}/{subset}"
+            individual_cache_dir = os.path.join(self.config.cache_dir, cache_key.replace("/", "_").replace(":", "_"))
+
+            if os.path.exists(individual_cache_dir):
+                cached_count += 1
+                # Try to load cache info to show details
+                info_file = os.path.join(individual_cache_dir, "dataset_info.json")
+                if os.path.exists(info_file):
+                    try:
+                        with open(info_file) as f:
+                            info = json.load(f)
+                        trajectories = info.get("total_trajectories", "unknown")
+                        timestamp = info.get("cache_timestamp", "unknown")
+                        rank_0_print(
+                            f"  ✅ {dataset_path}/{subset}: {trajectories} trajectories (cached at {timestamp})"
+                        )
+                    except:
+                        rank_0_print(f"  ✅ {dataset_path}/{subset}: Cache exists but info file corrupted")
+                else:
+                    rank_0_print(f"  ✅ {dataset_path}/{subset}: Cache exists (no info file)")
+            else:
+                rank_0_print(f"  ❌ {dataset_path}/{subset}: No cache found")
+
+        # Show summary
+        rank_0_print("\n📊 Cache Status Summary:")
+        rank_0_print(f"  ✅ Already cached: {cached_count}/{total_count} dataset/subset pairs")
+        rank_0_print(f"  🔄 Need processing: {total_count - cached_count}/{total_count} dataset/subset pairs")
+
+        if cached_count == total_count:
+            rank_0_print("  🎉 All dataset/subset pairs are already cached! Use --force-reprocess to reprocess.")
+        elif cached_count > 0:
+            rank_0_print("  💡 Some dataset/subset pairs are cached. Only uncached ones will be processed.")
+        else:
+            rank_0_print("  🚀 No dataset/subset pairs are cached. All will be processed.")
+
+    def _show_final_status_summary(self, all_datasets: list[tuple[str, str]]):
+        """
+        Show a summary of which datasets were processed and which were loaded from cache.
+        """
+        rank_0_print("\n📊 Final Status Summary for Dataset Preprocessing:")
+
+        processed_count = 0
+        loaded_count = 0
+        total_count = len(all_datasets)
+
+        for dataset_path, subset in all_datasets:
+            cache_key = f"{dataset_path}/{subset}"
+            individual_cache_dir = os.path.join(self.config.cache_dir, cache_key.replace("/", "_").replace(":", "_"))
+
+            if cache_key in self.datasets:
+                if os.path.exists(individual_cache_dir):
+                    loaded_count += 1
+                    rank_0_print(
+                        f"  ✅ {dataset_path}/{subset}: Loaded from cache ({len(self.datasets[cache_key])} trajectories)"
+                    )
+                else:
+                    processed_count += 1
+                    rank_0_print(
+                        f"  🔄 {dataset_path}/{subset}: Newly processed ({len(self.datasets[cache_key])} trajectories)"
+                    )
+            else:
+                rank_0_print(f"  ❌ {dataset_path}/{subset}: Failed to load/process")
+
+        # Show summary counts
+        rank_0_print("\n📈 Processing Summary:")
+        rank_0_print(f"  🔄 Newly processed: {processed_count} dataset/subset pairs")
+        rank_0_print(f"  ✅ Loaded from cache: {loaded_count} dataset/subset pairs")
+        rank_0_print(f"  ❌ Failed: {total_count - processed_count - loaded_count} dataset/subset pairs")
+        rank_0_print(f"  📊 Total available: {processed_count + loaded_count}/{total_count} dataset/subset pairs")
+
+
+@wrap()
+def main(config: DataPreprocessConfig):
+    """Main preprocessing function."""
+    # Try to increase the soft limit for open files to reduce 'Too many open files'
+    try:
+        import resource  # type: ignore
+
+        soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+        target_soft = min(hard, 65535)
+        if soft < target_soft:
+            resource.setrlimit(resource.RLIMIT_NOFILE, (target_soft, hard))
+    except Exception:
+        pass
+
+    # Show dataset structure info
+    print("\n🏗️  Dataset Configuration Structure:")
+    print("Training datasets:")
+    for dataset_path, dataset_subsets in zip(config.train_datasets, config.train_subsets, strict=False):
+        if isinstance(dataset_subsets, str):
+            dataset_subsets = [dataset_subsets]
+        print(f"  📚 {dataset_path}: {len(dataset_subsets)} subset(s)")
+        for subset in dataset_subsets:
+            print(f"    📂 {subset}")
+
+    print("Evaluation datasets:")
+    for dataset_path, dataset_subsets in zip(config.eval_datasets, config.eval_subsets, strict=False):
+        if isinstance(dataset_subsets, str):
+            dataset_subsets = [dataset_subsets]
+        print(f"  📚 {dataset_path}: {len(dataset_subsets)} subset(s)")
+        for subset in dataset_subsets:
+            print(f"    📂 {subset}")
+
+    print("\n💡 Note: Each dataset can now have multiple subsets!")
+    print("   - Single subset: ['subset1'] or 'subset1'")
+    print("   - Multiple subsets: ['subset1', 'subset2', 'subset3']")
+
+    # Create unified preprocessor for all datasets
+    preprocessor = DatasetPreprocessor(config)
+
+    # Preprocess all datasets
+    print("\n=== Processing All Datasets ===")
+
+    preprocessor.preprocess_datasets()
+
+    # Test the caches
+    print("\n=== Testing Caches ===")
+
+    if preprocessor.datasets:
+        print("\n📚 All Datasets:")
+        total_trajectories = sum(len(dataset) for dataset in preprocessor.datasets.values())
+        print(f"  Total trajectories: {total_trajectories}")
+
+        # Get combined indices
+        combined_indices = preprocessor.get_combined_indices()
+        if combined_indices:
+            print(f"  Robot trajectories: {len(combined_indices.get('robot_trajectories', []))}")
+            print(f"  Human trajectories: {len(combined_indices.get('human_trajectories', []))}")
+            print(f"  Tasks: {list(combined_indices.get('task_indices', {}).keys())}")
+
+        # Test direct access to first dataset
+        if preprocessor.datasets:
+            first_dataset = next(iter(preprocessor.datasets.values()))
+            if len(first_dataset) > 0:
+                test_traj = first_dataset[0]
+                print(f"  Sample trajectory: {test_traj['id']} - {test_traj['task']}")
+
+    # Show individual dataset info
+    print("\n📊 Individual Dataset Summary:")
+    print(f"Total datasets processed: {len(preprocessor.datasets)}")
+    for cache_key, dataset in preprocessor.datasets.items():
+        print(f"  ✅ {cache_key}: {len(dataset)} trajectories")
+
+    # Show dataset structure
+    print("\n🏗️  Dataset Structure:")
+    print("Training datasets:")
+    for dataset_path, dataset_subsets in zip(config.train_datasets, config.train_subsets, strict=False):
+        if isinstance(dataset_subsets, str):
+            dataset_subsets = [dataset_subsets]
+        print(f"  📚 {dataset_path}: {len(dataset_subsets)} subset(s)")
+        for subset in dataset_subsets:
+            cache_key = f"{dataset_path}/{subset}"
+            if cache_key in preprocessor.datasets:
+                print(f"    ✅ {subset}: {len(preprocessor.datasets[cache_key])} trajectories")
+            else:
+                print(f"    ❌ {subset}: Failed to load")
+
+    print("Evaluation datasets:")
+    for dataset_path, dataset_subsets in zip(config.eval_datasets, config.eval_subsets, strict=False):
+        if isinstance(dataset_subsets, str):
+            dataset_subsets = [dataset_subsets]
+        print(f"  📚 {dataset_path}: {len(dataset_subsets)} subset(s)")
+        for subset in dataset_subsets:
+            cache_key = f"{dataset_path}/{subset}"
+            if cache_key in preprocessor.datasets:
+                print(f"    ✅ {subset}: {len(preprocessor.datasets[cache_key])} trajectories")
+            else:
+                print(f"    ❌ {subset}: Failed to load")
+
+    print("\n✅ Dataset preprocessing complete!")
+    print(f"Unified cache: {config.cache_dir}")
+    print(f"Please set ROBOMETER_PROCESSED_DATASETS_PATH to {config.cache_dir} by running:")
+    print(f"export ROBOMETER_PROCESSED_DATASETS_PATH={config.cache_dir}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/robometer/evals/README.md b/robometer/robometer/evals/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..a39b1804769291f55cff45e2ea8409bdef4dbb41
--- /dev/null
+++ b/robometer/robometer/evals/README.md
@@ -0,0 +1,147 @@
+# Baseline Evaluation Guide
+
+This guide documents how to run baseline evaluations for GVL, RL-VLM-F, VLAC, and Robometer on all datasets.
+
+## Overview
+
+The baseline evaluation system supports:
+
+- **GVL** (Gemini Video Language): Progress prediction via Gemini API
+- **RL-VLM-F**: Preference comparison via Vision-Language Models (Gemini or OpenAI)
+- **VLAC**: Progress prediction using the VLAC model
+- **Robometer (RBM)**: Progress prediction and preference comparison using trained Robometer checkpoints:
+  - **aliangdw/Robometer-4B** – general-purpose model
+  - **aliangdw/Robometer-4B-LIBERO** – LIBERO-focused variant
+- **Robo-Dopamine (GRM)**: Progress prediction using Robo-Dopamine GRM (separate venv; see setup below)
+- **ReWiND**: Progress prediction and preference using ReWiND checkpoints (same interface as RBM)
+
+## Evaluation Types
+
+1. **Reward Alignment**: Evaluates progress prediction along trajectories
+2. **Policy Ranking**: Evaluates ability to rank trajectories by quality/partial_success
+3. **Quality Preference**: Evaluates preference comparison between trajectory pairs
+
+## Standard Evaluation Configuration
+
+For comprehensive evaluation across all datasets, use the following settings:
+
+- **Reward Alignment**: 20 trajectories per dataset
+- **Policy Ranking**: 20 tasks per dataset, 2 trajectories per quality label
+- **Quality Preference**: 100 comparisons per dataset
+
+## Prerequisites
+
+### Environment Variables
+
+For **GVL** and **RL-VLM-F** (Gemini):
+```bash
+export GEMINI_API_KEY="your-gemini-api-key"
+```
+
+For **RL-VLM-F** (OpenAI):
+```bash
+export OPENAI_API_KEY="your-openai-api-key"
+```
+
+### VLAC Model Setup
+
+For **VLAC**, you need to:
+
+1. **Create a separate virtual environment** (required because VLAC dependencies conflict with Robometer/RBM dependencies):
+   ```bash
+   uv venv .venv-vlac
+   uv pip install -e ".[vlac]" --python .venv-vlac/bin/python
+   ```
+   This creates a separate `.venv-vlac` environment with VLAC-compatible dependencies.
+
+2. **Download the model** from Hugging Face (auto-download if `model_path` is a Hugging Face repo ID) or provide a local path to the model checkpoint
+
+### Robo-Dopamine (GRM) setup
+
+The **Robo-Dopamine GRM** baseline is built on [Robo-Dopamine](https://github.com/FlagOpen/Robo-Dopamine). It uses vLLM, which requires **torch 2.9.x**. Reward-fm pins **torch 2.8.0** and **xformers**, so the resolver cannot satisfy both. Use a dedicated venv and install **vLLM first**, then the rest of the deps:
+
+1. **Create a separate virtual environment** (requires CUDA 12.x). Run from the **repo root**:
+   ```bash
+   cd /path/to/robometer   # or your repo root
+   uv venv .venv-robodopamine
+   uv pip install vllm --python .venv-robodopamine/bin/python
+   uv pip install -r requirements-robodopamine.txt --python .venv-robodopamine/bin/python
+   uv pip install -e . --no-deps --python .venv-robodopamine/bin/python
+   ```
+   Verify vLLM is installed:
+   ```bash
+   .venv-robodopamine/bin/python -c "import vllm; print('vLLM OK')"
+   ```
+   Use this venv only for the Robo-Dopamine baseline.
+
+2. **Run evals** using the venv’s Python directly (do **not** use `uv run`, or vLLM may not be found):
+   ```bash
+   .venv-robodopamine/bin/python robometer/evals/run_baseline_eval.py reward_model=robodopamine ...
+   ```
+   Run from the repo root. The config key is still `reward_model=robodopamine`.
+
+### Robometer (RBM) / ReWiND Model Setup
+
+For **Robometer** or **ReWiND** (`reward_model=rbm` or `reward_model=rewind`), provide:
+
+1. **`model_path`**: Hugging Face repo ID or local checkpoint path.
+   - **Robometer**: `aliangdw/Robometer-4B` (general) or `aliangdw/Robometer-4B-LIBERO` (LIBERO).
+   - Config is loaded automatically from the checkpoint when available.
+2. **Optional**: `model_config.batch_size` for inference batch size (default: 32).
+
+
+### Output
+
+- `output_dir`: Output directory (default: `./baseline_eval_output/<reward_model>_<timestamp>`)
+
+## Output Structure
+
+Each evaluation run creates an output directory with:
+
+```
+baseline_eval_output/
+└── <reward_model>_<timestamp>/
+    ├── baseline_metrics.json                    # Aggregated metrics
+    ├── <eval_type>_<dataset>_results.json       # Detailed results per eval type and dataset
+    ├── <eval_type>_<dataset>_task_groups.json  # Task-level groupings (for policy_ranking, quality_preference)
+    ├── <eval_type>_<dataset>_task_details.json # Task-level details (for policy_ranking, quality_preference)
+    └── <eval_type>_<dataset>_plots/             # Visualization plots (for reward_alignment)
+        └── trajectory_*.gif                     # Plot+video GIFs
+```
+
+## Dataset Names
+
+Common dataset names (use the full dataset key as it appears in the Hugging Face dataset):
+
+- `aliangdw_metaworld_metaworld_eval`
+- `jesbu1_roboarena_eval_debug_nowrist`
+- `jesbu1_oxe_rfm_eval`
+- `abraranwar_libero_rfm`
+- `ykorkmaz_libero_failure_rfm`
+- And many more...
+
+To see all available datasets, check the dataset loading code or Hugging Face.
+
+## Notes
+
+1. **RoboArena Datasets**: For RoboArena datasets, policy ranking uses `partial_success` values instead of quality labels. Use `num_partial_successes` parameter for circular sampling across partial_success values.
+
+2. **Frame Steps vs Whole Trajectories**: 
+   - `use_frame_steps=true`: Generates subsequences (0:frame_step, 0:2*frame_step, etc.)
+   - `use_frame_steps=false`: Uses whole trajectories
+
+3. **VLAC Dependencies**: VLAC requires a separate environment due to dependency conflicts. See `pyproject.toml` for optional dependency groups.
+
+4. **API Rate Limits**: Be aware of API rate limits when running RL-VLM-F or GVL evaluations on large datasets.
+
+5. **GPU Memory**: VLAC evaluations may require significant GPU memory depending on model size and batch settings.
+
+## Troubleshooting
+
+### Common Issues
+
+1. **API Key Not Found**: Ensure `GEMINI_API_KEY` or `OPENAI_API_KEY` is set in your environment
+2. **VLAC Model Not Found**: Check that `vlac_model_path` is correct or that the Hugging Face model can be downloaded.
+3. **Robometer/RBM Model Not Found**: Use `model_path="aliangdw/Robometer-4B"` or `aliangdw/Robometer-4B-LIBERO`, or ensure your local path is correct.
+4. **CUDA Out of Memory**: Reduce `vlac_batch_num` or `model_config.batch_size`, or use a smaller model.
+5. **Dataset Not Found**: Verify the dataset name matches exactly (including underscores and case).
\ No newline at end of file
diff --git a/robometer/robometer/evals/__init__.py b/robometer/robometer/evals/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/robometer/robometer/evals/baseline_eval_server.py b/robometer/robometer/evals/baseline_eval_server.py
new file mode 100644
index 0000000000000000000000000000000000000000..a743c3e712163156afb42a2d908c3a7697b9e84f
--- /dev/null
+++ b/robometer/robometer/evals/baseline_eval_server.py
@@ -0,0 +1,475 @@
+#!/usr/bin/env python3
+"""
+FastAPI server to evaluate baseline models (RL-VLM-F, GVL, VLAC, RoboReward) with batch processing.
+
+Usage examples:
+    # RL-VLM-F baseline server
+    uv run python robometer/evals/baseline_eval_server.py \
+        reward_model=rlvlmf \
+        reward_model.model_config.vlm_provider=gemini \
+        server_port=8001
+    
+    # VLAC baseline server
+    uv run --extra vlac --python .venv-vlac/bin/python robometer/evals/baseline_eval_server.py \
+        reward_model=vlac \
+        model_path=InternRobotics/VLAC \
+        server_port=8010
+    
+    # RoboReward baseline server
+    uv run python robometer/evals/baseline_eval_server.py \
+        reward_model=roboreward \
+        model_path=teetone/RoboReward-8B \
+        server_port=8003
+
+Endpoints:
+  POST /evaluate_batch        - JSON payload with samples
+  POST /evaluate_batch_npy    - multipart payload with .npy blobs for numpy arrays
+  GET /health                 - Health check
+  GET /model_info             - Model information
+
+Response payload per request contains predictions grouped by sample type:
+  {
+    "outputs_preference": [...],   # Preference predictions
+    "outputs_progress": [...],     # Progress predictions
+  }
+"""
+
+from __future__ import annotations
+
+import asyncio
+import copy
+import json
+import os
+import time
+from concurrent.futures import ThreadPoolExecutor
+from dataclasses import asdict
+from threading import Lock
+from typing import Any, Dict, List, Optional
+
+import uvicorn
+import numpy as np
+from omegaconf import DictConfig
+from hydra import main as hydra_main
+from fastapi import FastAPI, Request
+from fastapi.middleware.cors import CORSMiddleware
+
+from robometer.configs.eval_configs import BaselineEvalConfig
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+from robometer.evals.baselines.rlvlmf import RLVLMF
+from robometer.evals.baselines.gvl import GVL
+from robometer.evals.baselines.vlac import VLAC
+
+try:
+    from robometer.evals.baselines.roboreward import RoboReward
+except ImportError:
+    RoboReward = None
+
+from robometer.evals.run_baseline_eval import process_preference_sample, process_progress_sample
+from robometer.evals.eval_utils import parse_npy_form_data, reconstruct_payload_from_npy
+from robometer.utils.config_utils import display_config, convert_hydra_to_dataclass
+from robometer.utils.logger import get_logger, setup_loguru_logging
+
+LOG_LEVEL = "TRACE"
+setup_loguru_logging(log_level=LOG_LEVEL)
+logger = get_logger()
+logger.info(f"robometer.baseline_eval_server logger initialized at level {LOG_LEVEL}")
+
+
+def aggregate_frame_step_predictions_baseline(
+    progress_preds: List[List[float]],
+    sample_frame_counts: List[int],
+) -> List[List[float]]:
+    """
+    Aggregate frame-step predictions back into full sequences for baseline models.
+
+    Args:
+        progress_preds: List of progress predictions from sub-samples (each is a list of floats)
+        sample_frame_counts: List indicating how many frames each original sample had
+
+    Returns:
+        Aggregated progress predictions, one list per original sample
+    """
+    aggregated_progress = []
+    current_idx = 0
+
+    for num_frames in sample_frame_counts:
+        if num_frames == 1:
+            # Single-frame sample, pass through
+            if current_idx < len(progress_preds):
+                aggregated_progress.append(progress_preds[current_idx])
+                current_idx += 1
+            else:
+                aggregated_progress.append([])
+        else:
+            # Collect predictions from sub-samples and extract last prediction from each
+            sample_predictions = []
+            for i in range(num_frames):
+                if current_idx < len(progress_preds):
+                    sub_pred = progress_preds[current_idx]
+                    # Extract the last prediction from this sub-sample
+                    if isinstance(sub_pred, list) and len(sub_pred) > 0:
+                        sample_predictions.append(sub_pred[-1])
+                    current_idx += 1
+            aggregated_progress.append(sample_predictions)
+
+    return aggregated_progress
+
+
+def process_batch_helper(
+    model: Any,
+    reward_model: str,
+    batch_data: List[Dict[str, Any]],
+    job_id: int = 0,
+    use_frame_steps: bool = False,
+) -> Dict[str, Any]:
+    """Synchronous batch processing."""
+    if not batch_data:
+        raise ValueError("No samples found in batch data")
+
+    logger.debug(f"[job {job_id}] Processing {len(batch_data)} samples (use_frame_steps={use_frame_steps})")
+
+    input_samples: List[Any] = []
+    for sample in batch_data:
+        if isinstance(sample, (PreferenceSample, ProgressSample)):
+            input_samples.append(sample)
+        elif isinstance(sample, dict):
+            sample_type = sample.get("sample_type")
+            if sample_type == "preference":
+                input_samples.append(PreferenceSample(**sample))
+            elif sample_type == "progress":
+                input_samples.append(ProgressSample(**sample))
+            else:
+                raise ValueError(f"Unsupported sample_type: {sample_type}")
+        else:
+            raise ValueError(f"Unsupported sample object type: {type(sample)}")
+
+    # Handle frame steps for progress samples - expand into sub-samples with increasing frame counts
+    # For RoboReward, we use all frames (no subsampling) since it handles variable-length inputs
+    sample_frame_counts = None
+    if use_frame_steps and reward_model == "roboreward":
+        expanded_samples = []
+        sample_frame_counts = []
+
+        for sample in input_samples:
+            if isinstance(sample, ProgressSample):
+                # Get the frames from the trajectory
+                frames = sample.trajectory.frames
+                num_frames = frames.shape[0] if hasattr(frames, "shape") else len(frames)
+
+                # Create sub-samples with increasing frame counts: 0:1, 0:2, 0:3, ..., 0:T
+                # Use all frames in each sub-sample (no subsampling)
+                for i in range(1, num_frames + 1):
+                    sub_frames = frames[:i]
+
+                    sub_trajectory = copy.deepcopy(sample.trajectory)
+                    sub_trajectory.frames = sub_frames
+                    sub_trajectory.frames_shape = (
+                        sub_frames.shape if hasattr(sub_frames, "shape") else (len(sub_frames),)
+                    )
+
+                    # Create sub-sample
+                    sub_sample = ProgressSample(
+                        trajectory=sub_trajectory,
+                        data_gen_strategy=sample.data_gen_strategy,
+                    )
+                    expanded_samples.append(sub_sample)
+
+                sample_frame_counts.append(num_frames)
+            else:
+                # Non-progress samples are passed through unchanged
+                expanded_samples.append(sample)
+                sample_frame_counts.append(1)
+
+        input_samples = expanded_samples
+        logger.debug(f"[job {job_id}] Expanded samples into {len(input_samples)} sub-samples with frame steps")
+
+    outputs_preference = []
+    outputs_progress_list = []
+
+    # Check if we can use batched processing for RoboReward
+    if reward_model == "roboreward" and hasattr(model, "compute_progress_batched"):
+        # Collect all progress samples for batched processing
+        progress_samples = [s for s in input_samples if isinstance(s, ProgressSample)]
+        preference_samples = [s for s in input_samples if isinstance(s, PreferenceSample)]
+
+        if progress_samples:
+            # Prepare batched inputs
+            frames_list = [s.trajectory.frames for s in progress_samples]
+            task_descriptions = [s.trajectory.task for s in progress_samples]
+
+            # Process in batch
+            logger.debug(f"[job {job_id}] Processing {len(progress_samples)} progress samples in batch with RoboReward")
+            batched_results = model.compute_progress_batched(frames_list, task_descriptions)
+
+            # Convert to expected format
+            for result in batched_results:
+                outputs_progress_list.append({"progress_pred": result})
+
+        # Process preference samples (if any, though RoboReward doesn't support them)
+        for sample in preference_samples:
+            logger.warning(f"Preference samples not supported for roboreward")
+    else:
+        # Original sequential processing for other models
+        for sample in input_samples:
+            if isinstance(sample, PreferenceSample):
+                if reward_model != "rlvlmf":
+                    logger.warning(f"Preference samples only supported for rlvlmf, got {reward_model}")
+                    continue
+                result = process_preference_sample(sample, model)
+                if result:
+                    outputs_preference.append(result)
+            elif isinstance(sample, ProgressSample):
+                if reward_model not in ["gvl", "vlac", "roboreward"]:
+                    logger.warning(f"Progress samples only supported for gvl, vlac, roboreward, got {reward_model}")
+                    continue
+                result = process_progress_sample(sample, model)
+                if result:
+                    outputs_progress_list.append(result)
+
+    # Format outputs to match regular eval_server format for compatibility
+    # Regular eval_server returns: {"outputs_progress": {"progress_pred": [[...], [...]], ...}}
+    # Each inner list is the progress_pred for one sample
+    if outputs_progress_list:
+        # Extract progress_pred from each result dict - each result has "progress_pred": [list of floats]
+        progress_preds = [result.get("progress_pred", []) for result in outputs_progress_list]
+
+        # Aggregate frame-step predictions if needed
+        if use_frame_steps and sample_frame_counts is not None:
+            progress_preds = aggregate_frame_step_predictions_baseline(progress_preds, sample_frame_counts)
+
+        outputs_progress = {
+            "progress_pred": progress_preds,  # List of lists, one per sample
+        }
+        # Also include success_probs if available
+        success_probs_list = [
+            result.get("success_probs") for result in outputs_progress_list if result.get("success_probs") is not None
+        ]
+        if success_probs_list:
+            outputs_progress["success_probs"] = success_probs_list
+    else:
+        outputs_progress = {"progress_pred": []}  # Empty dict with empty list, not None
+
+    # Format preference outputs to match regular eval_server format
+    if outputs_preference:
+        predictions = [r.get("preference_pred") for r in outputs_preference if r.get("preference_pred") is not None]
+        prediction_probs = [
+            r.get("preference_pred") for r in outputs_preference if r.get("preference_pred") is not None
+        ]
+        outputs_preference_dict = {
+            "predictions": predictions,
+            "prediction_probs": prediction_probs,
+        }
+    else:
+        outputs_preference_dict = None
+
+    return {
+        "outputs_preference": outputs_preference_dict,
+        "outputs_progress": outputs_progress,
+    }
+
+
+class BaselineEvalServer:
+    """Baseline evaluation server that processes batches of samples."""
+
+    def __init__(self, cfg: BaselineEvalConfig):
+        self.cfg = cfg
+        self.reward_model = cfg.reward_model
+        self.model = None
+        self.executor = ThreadPoolExecutor(max_workers=1)
+        self._job_counter = 0
+        self._job_counter_lock = Lock()
+
+        logger.info(f"Initializing baseline eval server: reward_model={self.reward_model}")
+
+        # Initialize model
+        self._initialize_model()
+
+        logger.info("Baseline eval server initialized successfully")
+
+    def _initialize_model(self):
+        """Initialize the baseline model based on config."""
+        model_config_dict = (
+            asdict(self.cfg.model_config)
+            if hasattr(self.cfg.model_config, "__dataclass_fields__")
+            else self.cfg.model_config.__dict__
+        )
+
+        if self.reward_model == "rlvlmf":
+            self.model = RLVLMF(**model_config_dict)
+        elif self.reward_model == "gvl":
+            self.model = GVL(max_frames=self.cfg.max_frames, **model_config_dict)
+        elif self.reward_model == "vlac":
+            if not self.cfg.model_path:
+                raise ValueError("model_path is required for VLAC baseline")
+            self.model = VLAC(model_path=self.cfg.model_path, **model_config_dict)
+        elif self.reward_model == "roboreward":
+            self.model = RoboReward(model_path=self.cfg.model_path or "teetone/RoboReward-4B", **model_config_dict)
+        else:
+            raise ValueError(
+                f"Unknown reward_model: {self.reward_model}. Must be 'rlvlmf', 'gvl', 'vlac', or 'roboreward'"
+            )
+
+        logger.info(f"Loaded {self.reward_model} baseline model")
+
+    async def process_batch(self, batch_data: List[Dict[str, Any]], use_frame_steps: bool = False):
+        """Process a batch using the executor."""
+        loop = asyncio.get_event_loop()
+
+        with self._job_counter_lock:
+            self._job_counter += 1
+            job_id = self._job_counter
+
+        logger.debug(
+            f"[job {job_id}] Processing batch with {len(batch_data)} samples (use_frame_steps={use_frame_steps})"
+        )
+
+        start_time = time.time()
+
+        try:
+            # Process batch in thread pool
+            result = await loop.run_in_executor(
+                self.executor,
+                process_batch_helper,
+                self.model,
+                self.reward_model,
+                batch_data,
+                job_id,
+                use_frame_steps,
+            )
+
+            processing_time = time.time() - start_time
+            logger.debug(f"[job {job_id}] Completed in {processing_time:.3f}s")
+
+            return result
+
+        except Exception as e:
+            logger.error(f"[job {job_id}] Error processing batch: {e}", exc_info=True)
+            raise
+
+    def get_model_info(self) -> Dict[str, Any]:
+        """Get model information."""
+        return {
+            "reward_model": self.reward_model,
+            "config": asdict(self.cfg),
+        }
+
+    def shutdown(self):
+        """Shutdown the executor."""
+        logger.info("Shutting down baseline eval server...")
+        self.executor.shutdown(wait=True)
+        logger.info("Baseline eval server shutdown complete")
+
+
+def create_app(cfg: BaselineEvalConfig, baseline_server: BaselineEvalServer | None = None):
+    app = FastAPI(title="Baseline Evaluation Server")
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+    # Initialize baseline server
+    baseline_server = baseline_server or BaselineEvalServer(cfg)
+    logger.info(f"Baseline eval server initialized with model: {baseline_server.reward_model}")
+
+    @app.post("/evaluate_batch")
+    async def evaluate_batch(batch: Dict[str, Any]) -> Dict[str, Any]:
+        """Evaluate a batch of samples using the baseline server."""
+        logger.debug(f"Received /evaluate_batch request with keys: {list(batch.keys())}")
+
+        # Handle both list and dict formats
+        use_frame_steps = False
+        if isinstance(batch, list):
+            batch_data = batch
+        elif isinstance(batch, dict) and "samples" in batch:
+            batch_data = batch["samples"]
+            use_frame_steps = batch.get("use_frame_steps", False)
+        else:
+            # Assume it's a single sample wrapped in a dict
+            use_frame_steps = batch.pop("use_frame_steps", False)
+            batch_data = [batch]
+
+        return await baseline_server.process_batch(batch_data, use_frame_steps=use_frame_steps)
+
+    @app.post("/evaluate_batch_npy")
+    async def evaluate_batch_npy(request: Request) -> Dict[str, Any]:
+        """Evaluate a batch with .npy file support for numpy arrays.
+
+        This endpoint handles multipart form data where:
+        - numpy arrays are sent as .npy files
+        - other data is sent as form fields
+        """
+        # Parse form data
+        form_data = await request.form()
+
+        # Extract numpy arrays and other data using shared utility (await async function)
+        numpy_arrays, other_data = await parse_npy_form_data(form_data)
+
+        # Extract use_frame_steps flag from other_data (handle both bool and string)
+        use_frame_steps_value = other_data.pop("use_frame_steps", False)
+        if isinstance(use_frame_steps_value, bool):
+            use_frame_steps = use_frame_steps_value
+        else:
+            use_frame_steps = str(use_frame_steps_value).lower() == "true"
+
+        # Reconstruct the original payload structure (baselines don't need torch tensor conversion)
+        batch_data = reconstruct_payload_from_npy(
+            numpy_arrays,
+            other_data,
+            trajectory_keys=["chosen_trajectory", "rejected_trajectory", "trajectory"],
+            convert_embeddings_to_torch=False,
+        )
+
+        # Process the batch
+        logger.debug(
+            f"Received /evaluate_batch_npy request with {len(numpy_arrays)} numpy arrays "
+            f"and {len(other_data)} other fields, use_frame_steps={use_frame_steps}"
+        )
+        return await baseline_server.process_batch(batch_data, use_frame_steps=use_frame_steps)
+
+    @app.get("/health")
+    def health_check() -> Dict[str, Any]:
+        """Health check endpoint."""
+        return {"status": "healthy", "reward_model": baseline_server.reward_model}
+
+    @app.get("/model_info")
+    def get_model_info() -> Dict[str, Any]:
+        """Get model information."""
+        return baseline_server.get_model_info()
+
+    @app.on_event("shutdown")
+    async def shutdown_event():
+        """Cleanup on shutdown."""
+        baseline_server.shutdown()
+
+    return app
+
+
+@hydra_main(version_base=None, config_path="../configs", config_name="baseline_eval_config")
+def main(cfg: DictConfig):
+    """Main entry point for baseline evaluation server using Hydra configuration."""
+    # Convert Hydra config to dataclass
+    baseline_cfg = convert_hydra_to_dataclass(cfg, BaselineEvalConfig)
+
+    # Display the configuration
+    display_config(baseline_cfg)
+
+    # Validate reward model
+    if baseline_cfg.reward_model not in ["rlvlmf", "gvl", "vlac", "roboreward"]:
+        raise ValueError(
+            f"reward_model must be 'rlvlmf', 'gvl', 'vlac', or 'roboreward', got {baseline_cfg.reward_model}"
+        )
+
+    baseline_server = BaselineEvalServer(baseline_cfg)
+    app = create_app(baseline_cfg, baseline_server)
+
+    print(f"Running baseline eval server on {baseline_cfg.server_url}:{baseline_cfg.server_port}")
+    print(f"Using {baseline_cfg.reward_model} baseline model")
+    uvicorn.run(app, host=baseline_cfg.server_url, port=baseline_cfg.server_port)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/robometer/evals/baselines/gvl.py b/robometer/robometer/evals/baselines/gvl.py
new file mode 100644
index 0000000000000000000000000000000000000000..67b10831e5122a4a507686774c4d3abfa2cdfd9b
--- /dev/null
+++ b/robometer/robometer/evals/baselines/gvl.py
@@ -0,0 +1,446 @@
+import cv2
+import base64
+import random
+import re
+import requests
+import json
+import os
+import time
+from typing import List, Dict, Optional
+import numpy as np
+
+
+class GVL:
+    def __init__(
+        self,
+        api_key: str = None,
+        max_frames: int = 15,
+        offset: float = 0.5,
+        model_name: str = "gemini-2.0-flash",
+        provider: str = "gemini",
+        max_retries: int = 5,
+        base_delay: float = 1.0,
+        max_delay: float = 60.0,
+        **kwargs,
+    ):
+        """
+        :param api_key:          API key for the provider (defaults to env var based on provider)
+        :param max_frames:       If N > max_frames, sample exactly max_frames frames
+        :param offset:           Time offset used when sampling frames (seconds/frame). Keep consistent meaning with the frontend if applicable.
+        :param model_name:       Model name to use (e.g., "gemini-2.0-flash" for Gemini, "gpt-4o" for OpenAI)
+        :param provider:         API provider: "gemini" or "openai"
+        :param max_retries:      Maximum number of retries on API throttling/errors
+        :param base_delay:       Base delay in seconds for exponential backoff
+        :param max_delay:        Maximum delay in seconds between retries
+        """
+        self.provider = provider.lower()
+        if self.provider not in ["gemini", "openai"]:
+            raise ValueError(f"Unsupported provider: {provider}. Must be 'gemini' or 'openai'")
+
+        # Set API key based on provider
+        if api_key is None:
+            if self.provider == "gemini":
+                api_key = os.environ.get("GEMINI_API_KEY")
+                if not api_key:
+                    raise ValueError("GEMINI_API_KEY environment variable must be set")
+            else:  # openai
+                api_key = os.environ.get("OPENAI_API_KEY")
+                if not api_key:
+                    raise ValueError("OPENAI_API_KEY environment variable must be set")
+
+        self.api_key = api_key
+        self.max_frames = max_frames
+        self.offset = offset
+        self.model_name = model_name
+        self.max_retries = max_retries
+        self.base_delay = base_delay
+        self.max_delay = max_delay
+
+        # List to store frame info: each element contains
+        # {"gt_index": i, "shuffled_index": ..., "base64": "..."}
+        self.frames_info: List[Dict] = []
+
+    def extract_frames_from_memory(self, frames_array: np.ndarray) -> None:
+        """
+        Improved sampling logic:
+        - If N <= max_frames, keep previous behavior (typically capturing first and last frames)
+        - If N > max_frames:
+          1) Always include frame 0 and frame N-1
+          2) Uniformly sample (max_frames - 2) frames from the middle (N-2) range
+          3) Merge indices, deduplicate, sort, then encode
+        """
+        total_frames = frames_array.shape[0]
+        if total_frames == 0:
+            print("[!] frames_array is empty; cannot extract frames.")
+            self.frames_info = []
+            return
+
+        # If total_frames <= max_frames: keep original behavior
+        if total_frames <= self.max_frames:
+            frame_count = total_frames
+            frame_interval = 1.0
+            print(f"Extracting all {frame_count} frames (<= max_frames).")
+
+            temp_indices = []
+            for i in range(frame_count):
+                # Use offset + i*frame_interval then int(...) to compute index
+                sample_time = self.offset + i * frame_interval
+                frame_index = int(sample_time)
+                if 0 <= frame_index < total_frames:
+                    temp_indices.append(frame_index)
+
+        else:
+            # total_frames > max_frames
+            print(
+                f"Extracting exactly {self.max_frames} frames from total={total_frames}, ensuring first & last included."
+            )
+            # 1) Always include frame 0 and frame N-1
+            temp_indices = [0, total_frames - 1]
+
+            # 2) Uniformly sample (max_frames - 2) middle frames
+            #    Note: you can keep offset as offset + i*frame_interval if desired.
+            #    Whether offset is meaningful for the first frame=0 depends on your use case.
+            #    Here we keep offset so that middle frames also carry a 0.5 shift.
+            inner_count = self.max_frames - 2
+            if (total_frames - 2) <= 0:
+                # If there are only 1 or 2 frames, skip middle sampling
+                print("Warning: total_frames - 2 <= 0, cannot sample middle frames.")
+            else:
+                frame_interval = (total_frames - 2) / float(inner_count)
+                for i in range(inner_count):
+                    sample_time = self.offset + i * frame_interval
+                    # Middle frame index falls within [1, N-2]
+                    frame_index = int(1 + sample_time)
+                    if 1 <= frame_index < (total_frames - 1):
+                        temp_indices.append(frame_index)
+
+            # Deduplicate and sort
+            temp_indices = sorted(set(temp_indices))
+            print(f"Extracted {len(temp_indices)} frames: {temp_indices}")
+
+        # ============ Convert indices to JPEG + base64 ============
+        temp_frames_info = []
+        for idx in temp_indices:
+            frame = frames_array[idx]  # shape = (H, W, 3)
+            ret, buffer = cv2.imencode(".jpg", frame)
+            if not ret:
+                continue
+
+            frame_b64 = base64.b64encode(buffer).decode("utf-8")
+            temp_frames_info.append({"gt_index": len(temp_frames_info) + 1, "base64": frame_b64})
+
+        self.frames_info = temp_frames_info
+
+    def shuffle_frames(self) -> None:
+        """
+        Randomly shuffle frames_info. Similar to the previous shuffle_frames behavior.
+        """
+        indices = list(range(1, len(self.frames_info) + 1))
+        random.shuffle(indices)
+        for frame, new_idx in zip(self.frames_info, indices):
+            frame["shuffled_index"] = new_idx
+
+    def build_prompt_parts(self, task_description: str) -> List[Dict]:
+        """
+        Build the `parts` list for the request, following the previous logic:
+          - First find the frame with gt_index=1 as the initial scene
+          - Add prompt1 + that frame
+          - Add prompt2
+          - Then, sorted by shuffled_index, append "Frame i:" + the corresponding frame
+        """
+        # Find the frame with gt_index = 1
+        initial_frame = next((f for f in self.frames_info if f["gt_index"] == 1), None)
+        if not initial_frame:
+            # If not found (extreme case), use the first frame
+            if not self.frames_info:
+                print("[!] No available frames; cannot build prompt_parts")
+                return []
+            initial_frame = self.frames_info[0]
+
+        prompt1 = (
+            f"You are an expert roboticist tasked to predict task completion percentages "
+            f"for frames of a robot for the task of {task_description}. "
+            f"The task completion percentages are between 0 and 100, where 100 corresponds to full task completion. "
+            f"Note that these frames are in random order, so please pay attention to the individual frames. "
+            f"\nInitial robot scene:\nThis frame:"
+        )
+
+        prompt2 = (
+            f" shows the initial robot scene, where the task completion percentage is 0.\n\n"
+            f"Now, for the task of *{task_description}*, output the task completion percentage "
+            f"for the following frames that are presented in random order. "
+            f"Format your response in JSON as follows, making sure to include all frames:\n\n"
+            f"[\n"
+            f'  {{"frame_number": i, "frame_description": "...", "task_completion_percentage": 0-100}}\n'
+            f"]\n"
+        )
+
+        parts = []
+        # 1) prompt1
+        parts.append({"text": prompt1})
+        # 2) initial_frame inline
+        parts.append({"inline_data": {"mime_type": "image/jpeg", "data": initial_frame["base64"]}})
+        # 3) prompt2
+        parts.append({"text": prompt2})
+
+        # 4) "Frame X" + inline, sorted by shuffled_index
+        frames_sorted_by_shuffle = sorted(self.frames_info, key=lambda f: f["shuffled_index"])
+
+        for i, frame in enumerate(frames_sorted_by_shuffle, start=1):
+            parts.append({"text": f"Frame {i}:"})
+            parts.append({"inline_data": {"mime_type": "image/jpeg", "data": frame["base64"]}})
+
+        return parts
+
+    def _stream_inference_gemini(self, parts: List[Dict]) -> str:
+        """
+        Call the Gemini SSE endpoint and return the concatenated streamed text.
+        Implements exponential backoff retry on throttling (429) and server errors (5xx).
+        """
+        url = f"https://generativelanguage.googleapis.com/v1beta/models/{self.model_name}:streamGenerateContent?alt=sse&key={self.api_key}"
+        body = {"contents": [{"parts": parts}]}
+        headers = {"Content-Type": "application/json"}
+
+        last_exception = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                full_text = ""
+                with requests.post(url, headers=headers, json=body, stream=True) as resp:
+                    # Check for throttling or server errors
+                    if resp.status_code == 429 or resp.status_code >= 500:
+                        delay = min(self.base_delay * (2**attempt), self.max_delay)
+                        print(
+                            f"[GVL] API returned {resp.status_code}, retrying in {delay:.1f}s (attempt {attempt + 1}/{self.max_retries + 1})"
+                        )
+                        time.sleep(delay)
+                        continue
+
+                    resp.raise_for_status()
+
+                    for line in resp.iter_lines(decode_unicode=True):
+                        if not line:
+                            continue
+                        if line.startswith("data: "):
+                            data_str = line[len("data: ") :]
+                            if data_str == "[DONE]":
+                                break
+                            try:
+                                data_json = json.loads(data_str)
+                                candidates = data_json.get("candidates")
+                                if candidates and len(candidates) > 0:
+                                    content = candidates[0].get("content", {})
+                                    parts_list = content.get("parts", [])
+                                    if parts_list:
+                                        text_piece = parts_list[0].get("text", "")
+                                        full_text += text_piece
+                            except json.JSONDecodeError:
+                                continue
+
+                return full_text
+
+            except requests.exceptions.RequestException as e:
+                last_exception = e
+                delay = min(self.base_delay * (2**attempt), self.max_delay)
+                print(
+                    f"[GVL] Request failed: {e}, retrying in {delay:.1f}s (attempt {attempt + 1}/{self.max_retries + 1})"
+                )
+                time.sleep(delay)
+
+        # All retries exhausted
+        print(f"[GVL] All {self.max_retries + 1} attempts failed")
+        if last_exception:
+            raise last_exception
+        return ""
+
+    def _stream_inference_openai(self, parts: List[Dict]) -> str:
+        """
+        Call the OpenAI API with vision capabilities and return the response text.
+        Implements exponential backoff retry on throttling (429) and server errors (5xx).
+        """
+        url = "https://api.openai.com/v1/chat/completions"
+        headers = {"Content-Type": "application/json", "Authorization": f"Bearer {self.api_key}"}
+
+        # Convert parts to OpenAI message format
+        content = []
+        for part in parts:
+            if "text" in part:
+                content.append({"type": "text", "text": part["text"]})
+            elif "inline_data" in part:
+                content.append({
+                    "type": "image_url",
+                    "image_url": {
+                        "url": f"data:{part['inline_data']['mime_type']};base64,{part['inline_data']['data']}",
+                        "detail": "low",  # Use low detail for efficiency
+                    },
+                })
+
+        body = {
+            "model": self.model_name,
+            "messages": [{"role": "user", "content": content}],
+            "max_tokens": 4096,
+            "stream": True,
+        }
+
+        last_exception = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                full_text = ""
+                with requests.post(url, headers=headers, json=body, stream=True) as resp:
+                    # Check for throttling or server errors
+                    if resp.status_code == 429 or resp.status_code >= 500:
+                        delay = min(self.base_delay * (2**attempt), self.max_delay)
+                        print(
+                            f"[GVL-OpenAI] API returned {resp.status_code}, retrying in {delay:.1f}s (attempt {attempt + 1}/{self.max_retries + 1})"
+                        )
+                        time.sleep(delay)
+                        continue
+
+                    resp.raise_for_status()
+
+                    for line in resp.iter_lines(decode_unicode=True):
+                        if not line:
+                            continue
+                        if line.startswith("data: "):
+                            data_str = line[len("data: ") :]
+                            if data_str == "[DONE]":
+                                break
+                            try:
+                                data_json = json.loads(data_str)
+                                choices = data_json.get("choices", [])
+                                if choices:
+                                    delta = choices[0].get("delta", {})
+                                    text_piece = delta.get("content", "")
+                                    if text_piece:
+                                        full_text += text_piece
+                            except json.JSONDecodeError:
+                                continue
+
+                return full_text
+
+            except requests.exceptions.RequestException as e:
+                last_exception = e
+                delay = min(self.base_delay * (2**attempt), self.max_delay)
+                print(
+                    f"[GVL-OpenAI] Request failed: {e}, retrying in {delay:.1f}s (attempt {attempt + 1}/{self.max_retries + 1})"
+                )
+                time.sleep(delay)
+
+        # All retries exhausted
+        print(f"[GVL-OpenAI] All {self.max_retries + 1} attempts failed")
+        if last_exception:
+            raise last_exception
+        return ""
+
+    def stream_inference(self, parts: List[Dict]) -> str:
+        """
+        Call the appropriate API endpoint based on provider and return the response text.
+        """
+        if self.provider == "gemini":
+            return self._stream_inference_gemini(parts)
+        else:  # openai
+            return self._stream_inference_openai(parts)
+
+    @staticmethod
+    def extract_json_from_response(text: str) -> str:
+        """
+        Extract JSON from a long text similar to the frontend JS logic:
+        - Prefer a ```json ... ``` fenced code block
+        - Otherwise try to match an array like [... { ... }, ...]
+        If not found, return an empty string.
+        """
+        code_block_pattern = re.compile(r"```(?:json)?\s*([\s\S]*?)\s*```")
+        match = code_block_pattern.search(text)
+        if match:
+            return match.group(1).strip()
+
+        array_pattern = re.compile(r"\[\s*\{[\s\S]*?\}\s*\]")
+        match = array_pattern.search(text)
+        if match:
+            return match.group(0).strip()
+
+        return ""
+
+    @staticmethod
+    def parse_model_output(model_text: str) -> Optional[List[Dict]]:
+        """
+        Try to extract JSON from the model text and json.loads it into a Python list.
+        Return None on failure.
+        """
+        json_str = GVL.extract_json_from_response(model_text)
+        if not json_str:
+            return None
+        try:
+            data = json.loads(json_str)
+            if isinstance(data, list):
+                return data
+            return None
+        except (json.JSONDecodeError, TypeError):
+            return None
+
+    def compute_progress(self, frames_array: np.ndarray, task_description: str = "") -> List[Optional[float]]:
+        """
+        Compute progress predictions for frames using GVL baseline.
+
+        Main pipeline:
+          1) Extract and encode frames from the in-memory array
+          2) Randomly shuffle
+          3) Build prompt
+          4) Run SSE inference
+          5) Parse JSON
+          6) Return completion array aligned with original frame order (gt_index)
+
+        :param frames_array: (N, H, W, 3) uint8 array from HDF5 (per-frame video data)
+        :param task_description: Robot task description
+        :return: List of completion percentages (0-1, normalized from 0-100) aligned with (gt_index)
+        """
+        # 1) Extract frames
+        self.extract_frames_from_memory(frames_array)
+        if not self.frames_info:
+            print("[!] No frames extracted from memory array.")
+            return []
+
+        # 2) Randomly shuffle
+        self.shuffle_frames()
+
+        # 3) Build prompt
+        parts = self.build_prompt_parts(task_description)
+        if not parts:
+            print("[!] build_prompt_parts failed; possibly no frame data")
+            return []
+
+        # 4) SSE inference
+        model_output_text = self.stream_inference(parts)
+
+        # 5) Parse JSON
+        result_data = self.parse_model_output(model_output_text)
+        if result_data is None:
+            print("[!] Failed to extract valid JSON; please check the model output.")
+            print("Full model output text:", model_output_text)
+            return []
+
+        # 6) Map model output back to gt_index
+        mapped_by_shuffled = {}
+        for item in result_data:
+            sidx = item.get("frame_number")
+            if isinstance(sidx, int):
+                mapped_by_shuffled[sidx] = item
+
+        # Record model results in frames_info
+        for frame in self.frames_info:
+            sidx = frame.get("shuffled_index")
+            if sidx in mapped_by_shuffled:
+                frame["model_output"] = mapped_by_shuffled[sidx]
+            else:
+                frame["model_output"] = None
+
+        # Iterate in ascending gt_index order
+        frames_in_gt_order = sorted(self.frames_info, key=lambda f: f["gt_index"])
+        task_completion_list = []
+        for f in frames_in_gt_order:
+            if f["model_output"] is not None:
+                percentage = f["model_output"].get("task_completion_percentage")
+                # Normalize from 0-100 to 0-1
+                normalized = percentage / 100.0
+                task_completion_list.append(normalized)
+        print("Task completion list:", task_completion_list)
+        return task_completion_list
diff --git a/robometer/robometer/evals/baselines/rbd_inference.py b/robometer/robometer/evals/baselines/rbd_inference.py
new file mode 100644
index 0000000000000000000000000000000000000000..7a029bb5c4751fc58a10f226c9c4227a9b53237a
--- /dev/null
+++ b/robometer/robometer/evals/baselines/rbd_inference.py
@@ -0,0 +1,713 @@
+import os
+import json
+import shutil
+import re
+from pathlib import Path
+from datetime import datetime
+from typing import List, Tuple, Optional, Dict
+
+import cv2
+import numpy as np
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from tqdm import tqdm
+
+from PIL import Image
+
+# VLLM & Transformers
+try:
+    from vllm import LLM, SamplingParams
+except ImportError:
+    print("vllm not found, please install it with `pip install vllm`")
+    pass
+
+from transformers import AutoProcessor
+from moviepy.video.io.ImageSequenceClip import ImageSequenceClip
+
+# Global Environment Settings
+os.environ["LOCAL_RANK"] = "0"
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+
+
+# -----------------------------
+# Configuration & Prompt
+# -----------------------------
+
+SYSTEM_PROMPT = """
+You are a rigorous, impartial vision evaluator for robot task progress. Your job is to judge whether the AFTER image set moves closer to the task objective than the BEFORE image set, using the provided reference examples only as anchors.
+
+<Task>
+`{task}`
+
+REFERENCE EXAMPLES (for visual anchoring only; not necessarily this run's actual START/END):
+- REFERENCE START — Robot Front Image (task just starting): <image>
+- REFERENCE END — Robot Front Image (task fully completed): <image>
+</Task>
+
+BEFORE Robot Front Image: <image>
+BEFORE Robot Left Wrist Image: <image>
+BEFORE Robot Right Wrist Image: <image>
+
+AFTER Robot Front Image: <image>
+AFTER Robot Left Wrist Image: <image>
+AFTER Robot Right Wrist Image: <image>
+
+Goal
+Compare the BEFORE and AFTER three-view sets and judge whether AFTER moves closer to accomplishing the task than BEFORE, using the REFERENCE START/END images as conceptual anchors.
+
+Progress Estimation (no formulas)
+1) Calibrate using the references:
+   - REFERENCE START = “just beginning”; REFERENCE END = “fully completed.”
+   - Visually estimate how far BEFORE and AFTER are along this START→END continuum.
+2) Direction:
+   - AFTER better than BEFORE → positive score.
+   - AFTER worse than BEFORE → negative score.
+   - Essentially the same → 0.
+3) Normalize to an integer percentage in [-100%, +100%]:
+   - For improvements, scale the improvement relative to what remained from BEFORE to END.
+   - For regressions, scale the deterioration relative to how far BEFORE had progressed from START.
+   - Clip to [-100%, +100%] and round to the nearest integer percent.
+
+Evaluation Criteria (apply across all three views)
+1) Task Alignment: Evidence directly tied to `{task}`.
+2) Completeness & Accuracy: Correct pose, contact, placement, orientation, grasp quality, absence of collisions, stability, etc.
+3) View-Specific Evidence & Consistency:
+   - Use the **Front** view for global layout, object pose, approach path, end-state geometry, and scene-level constraints.
+   - Use the **Left/Right Wrist** views to inspect **fine-grained gripper state** (finger closure, contact location/area, slippage, wedge/misalignment, object deformation, cable/wire/cloth entanglement, unintended contact, occluded collisions).
+   - When views disagree, prioritize the view that provides **decisive cues** for the criterion at hand. In particular, wrist views often **override** for grasp/contact validity and safety.
+   - If any single view shows a failure that invalidates success (e.g., mis-grasp, collision, unsafe/unstable pose), let that override when judging progress.
+4) Ignore Irrelevant Factors: Lighting, color shifts, background clutter, or UI/watermarks that don't affect task success.
+5) Ambiguity: If evidence is genuinely inconclusive or conflicting without decisive cues, treat progress as unchanged → 0%.
+
+Output Format (STRICT)
+Return ONLY one line containing the score wrapped in <score> tags, as an integer percentage with a percent sign:
+<score>+NN%</score>  or  <score>-NN%</score>  or  <score>0%</score>
+"""
+
+# -----------------------------
+# File & Video Utilities
+# -----------------------------
+
+
+def ensure_dir(path: Path):
+    path.mkdir(parents=True, exist_ok=True)
+
+
+def list_pngs_sorted(dir_path: Path) -> List[Path]:
+    """Return lexicographically sorted .png files (case-insensitive) under dir_path."""
+    return sorted([p for p in dir_path.iterdir() if p.is_file() and p.suffix.lower() == ".png"])
+
+
+def get_frame_count(path: Path) -> Tuple[str, int]:
+    """
+    Detects if path is a video file or a directory of images.
+    Returns: (source_type, frame_count). source_type is 'dir' or 'video'.
+    """
+    if path.is_dir():
+        files = list_pngs_sorted(path)
+        if not files:
+            raise RuntimeError(f"No PNG frames found in directory: {path}")
+        return "dir", len(files)
+    else:
+        cap = cv2.VideoCapture(str(path))
+        if not cap.isOpened():
+            raise RuntimeError(f"Cannot open video: {path}")
+        n = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        cap.release()
+        if n <= 0:
+            raise RuntimeError(f"Invalid frame count from video: {path}")
+        return "video", n
+
+
+def make_sample_indices_by_interval(num_frames: int, interval: int) -> List[int]:
+    """
+    Generate indices based on a fixed frame interval.
+    Always includes the first frame (0) and ensures the very last frame (num_frames-1) is included.
+    """
+    if num_frames < 1:
+        return []
+
+    # Generate base steps: 0, interval, 2*interval, ...
+    indices = list(range(0, num_frames, interval))
+
+    # Ensure the absolute last frame is included to cover the full episode
+    last_idx = num_frames - 1
+    if not indices or indices[-1] != last_idx:
+        indices.append(last_idx)
+
+    return indices
+
+
+def save_frames(src_path: Path, out_dir: Path, indices: List[int], src_type: str) -> None:
+    """
+    Extracts and saves specific frames from a video or copies them from a directory.
+    Output format: frame_{:06d}.png
+    """
+    ensure_dir(out_dir)
+
+    if src_type == "video":
+        cap = cv2.VideoCapture(str(src_path))
+        if not cap.isOpened():
+            raise RuntimeError(f"Cannot open video: {src_path}")
+
+        for idx in indices:
+            cap.set(cv2.CAP_PROP_POS_FRAMES, int(idx))
+            ok, frame = cap.read()
+            if not ok or frame is None:
+                ok, frame = cap.read()  # Retry once
+            if not ok or frame is None:
+                print(f"[WARN] Failed to read frame {idx} from {src_path}")
+                continue
+
+            out_path = out_dir / f"frame_{idx:06d}.png"
+            cv2.imwrite(str(out_path), frame, [int(cv2.IMWRITE_PNG_COMPRESSION), 3])
+        cap.release()
+
+    elif src_type == "dir":
+        files = list_pngs_sorted(src_path)
+        n = len(files)
+        for idx in indices:
+            if not (0 <= idx < n):
+                continue
+            src = files[idx]
+            dst = out_dir / f"frame_{idx:06d}.png"
+            shutil.copyfile(src, dst)
+
+
+def build_samples_json(
+    run_root: Path, task: str, indices: List[int], ref_end_path: str, mode: str = "incremental"
+) -> List[Dict]:
+    """
+    Constructs the list of samples for VLLM inference.
+
+    Args:
+        mode: "incremental", "forward", or "backward"
+        ref_end_path: Absolute path to the Goal Image.
+    """
+    timestamp_name = run_root.name
+    cache_root = run_root / ".cache"
+    items = []
+
+    if len(indices) < 2:
+        return items
+
+    # For Forward mode, we need the Start Frame (Index 0)
+    # For Incremental, we use Index k
+    # For Backward, we use Goal Image
+
+    for k in range(len(indices) - 1):
+        af = indices[k + 1]  # The "After" frame (Current Step)
+
+        # Define "Before" images based on mode
+        bf_images = []
+        bf_id_str = ""
+
+        if mode == "incremental":
+            bf = indices[k]
+            bf_id_str = f"bf_{bf:06d}"
+            bf_images = [
+                str(cache_root / "cam_high" / f"frame_{bf:06d}.png"),
+                str(cache_root / "cam_left_wrist" / f"frame_{bf:06d}.png"),
+                str(cache_root / "cam_right_wrist" / f"frame_{bf:06d}.png"),
+            ]
+        elif mode == "forward":
+            # Compare Start(0) -> Current(af)
+            bf = indices[0]
+            bf_id_str = f"start_{bf:06d}"
+            bf_images = [
+                str(cache_root / "cam_high" / f"frame_{bf:06d}.png"),
+                str(cache_root / "cam_left_wrist" / f"frame_{bf:06d}.png"),
+                str(cache_root / "cam_right_wrist" / f"frame_{bf:06d}.png"),
+            ]
+        elif mode == "backward":
+            # Compare Goal -> Current(af)
+            # Use ref_end_path for all 3 views if wrist goals aren't explicit
+            bf_id_str = "goal"
+            bf_images = [ref_end_path, ref_end_path, ref_end_path]
+        else:
+            raise ValueError(f"Unknown mode: {mode}")
+
+        # Define "After" images (Always the current step)
+        af_images = [
+            str(cache_root / "cam_high" / f"frame_{af:06d}.png"),
+            str(cache_root / "cam_left_wrist" / f"frame_{af:06d}.png"),
+            str(cache_root / "cam_right_wrist" / f"frame_{af:06d}.png"),
+        ]
+
+        items.append({
+            "id": f"step-{timestamp_name}-{k:04d}-{bf_id_str}-af_{af:06d}",
+            "task": task,
+            "image": [
+                str(cache_root / "cam_high" / f"frame_{0:06d}.png"),  # 1. Ref Start
+                ref_end_path,  # 2. Ref End
+                bf_images[0],  # 3. Before High
+                bf_images[1],  # 4. Before Left
+                bf_images[2],  # 5. Before Right
+                af_images[0],  # 6. After High
+                af_images[1],  # 7. After Left
+                af_images[2],  # 8. After Right
+            ],
+        })
+    return items
+
+
+# -----------------------------
+# Visualization
+# -----------------------------
+
+
+def plot_video_reward(episode_root: Path):
+    """
+    Generates a visualization video (reward_vis.mp4).
+    Feature: Uses 'symlog' (Symmetric Log) scale for the Hop plot.
+    Modified: Uses moviepy instead of cv2.VideoWriter for better compatibility.
+    """
+    pred_path = episode_root / "pred_vllm.json"
+    if not pred_path.exists():
+        print(f"[WARN] No prediction file found at {pred_path}, skipping visualization.")
+        return
+
+    with open(pred_path, "r", encoding="utf-8") as f:
+        items = json.load(f)
+    if not items:
+        return
+
+    def get_fid(p):
+        m = re.search(r"frame_(\d+)\.png", os.path.basename(p))
+        return int(m.group(1)) if m else 0
+
+    # Output path
+    out_path = episode_root / "reward_vis.mp4"
+
+    # Layout dimensions
+    W_panel, H_panel = 384, 288
+    H_plot = 260
+    W_total = 3 * W_panel
+    H_total = H_panel + H_plot
+
+    # Data containers
+    xs_step, ys_hop, c_hop = [], [], []
+    xs_frame, ys_prog, c_prog = [], [], []
+
+    # Buffer to store frames for moviepy (Must be RGB)
+    frames_buffer = []
+
+    def draw_plots():
+        dpi = 100
+        fig = plt.figure(figsize=(W_total / dpi, H_plot / dpi), dpi=dpi)
+        ax1 = fig.add_subplot(1, 2, 1)
+        ax2 = fig.add_subplot(1, 2, 2)
+
+        # Plot 1: Step vs Hop (SymLog)
+        if xs_step:
+            if len(xs_step) > 1:
+                ax1.plot(xs_step, ys_hop, "-", color="#999999", alpha=0.6)
+            ax1.scatter(xs_step, ys_hop, c=c_hop, s=36, edgecolors="k", linewidths=0.5)
+
+        ax1.set_yscale("symlog", linthresh=0.05)
+        ax1.set_ylim(-1.5, 1.5)
+        ticks = [-1.0, -0.1, 0, 0.1, 1.0]
+        ax1.set_yticks(ticks)
+        ax1.set_yticklabels(["-100%", "-10%", "0", "10%", "100%"])
+
+        max_step = len(items) + 1
+        ax1.set_xlim(left=0, right=max(5, max_step))
+        ax1.set_title("Instant Hop/Reward (SymLog)")
+        ax1.grid(True, which="major", linestyle="-", alpha=0.4)
+
+        # Plot 2: Frame vs Progress (Linear)
+        if xs_frame:
+            if len(xs_frame) > 1:
+                ax2.plot(xs_frame, ys_prog, "-", color="#999999", alpha=0.6)
+            ax2.scatter(xs_frame, ys_prog, c=c_prog, s=36, edgecolors="k", linewidths=0.5)
+
+        ax2.set_ylim(-0.05, 1.05)
+        ax2.set_title("Accumulated Progress")
+        ax2.grid(True, linestyle="--", alpha=0.5)
+
+        fig.tight_layout(pad=1.0)
+
+        fig.canvas.draw()
+        img = np.asarray(fig.canvas.buffer_rgba())
+        plt.close(fig)
+
+        # NOTE: Convert to RGB for moviepy (Matplotlib RGBA -> RGB)
+        img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
+        return cv2.resize(img, (W_total, H_plot), interpolation=cv2.INTER_AREA)
+
+    def read_images_rgb(paths):
+        """Helper to read images and convert BGR to RGB immediately."""
+        imgs = []
+        for p in paths:
+            im = cv2.imread(p)
+            if im is None:
+                return None
+            im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+            imgs.append(im)
+        return imgs
+
+    try:
+        # --- Initial Frame ---
+        first = items[0]
+        # Images: [Goal, Start, Current] equivalent indices based on original logic
+        imgs_top = read_images_rgb([first["image"][i] for i in [3, 2, 4]])
+
+        if imgs_top:
+            top_row = np.hstack([cv2.resize(im, (W_panel, H_panel)) for im in imgs_top])
+
+            # Frame 0 setup
+            xs_step.append(0)
+            ys_hop.append(0.0)
+            c_hop.append("#808080")
+
+            xs_frame.append(0)
+            ys_prog.append(0.0)
+            c_prog.append("#00A000")
+
+            bottom_plot = draw_plots()
+            frame_comp = np.vstack([top_row, bottom_plot])
+
+            if frame_comp.shape[:2] != (H_total, W_total):
+                frame_comp = cv2.resize(frame_comp, (W_total, H_total))
+
+            # Text (White is (255,255,255) in RGB as well)
+            cv2.putText(frame_comp, "Start", (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2)
+
+            frames_buffer.append(frame_comp)
+
+        # --- Loop ---
+        prev_prog = 0.0
+        cum_frame = 0
+
+        for i, item in enumerate(items):
+            # Indices for standard display logic from original code
+            imgs_top = read_images_rgb([item["image"][idx] for idx in [6, 5, 7]])
+            if not imgs_top:
+                continue
+
+            top_row = np.hstack([cv2.resize(im, (W_panel, H_panel)) for im in imgs_top])
+
+            hop = float(item.get("hop", 0))
+            prog = float(item.get("progress", 0))
+
+            step_id = i + 1
+            cum_frame += 10
+            try:
+                fid_curr = get_fid(item["image"][5])
+                cum_frame = fid_curr
+            except Exception:
+                pass
+
+            xs_step.append(step_id)
+            ys_hop.append(hop)
+            c_hop.append("#00A000" if hop > 0 else ("#CC0000" if hop < 0 else "#808080"))
+
+            xs_frame.append(cum_frame)
+            ys_prog.append(prog)
+            c_prog.append("#00A000" if prog >= prev_prog else "#CC0000")
+            prev_prog = prog
+
+            bottom_plot = draw_plots()
+            frame_comp = np.vstack([top_row, bottom_plot])
+
+            if frame_comp.shape[:2] != (H_total, W_total):
+                frame_comp = cv2.resize(frame_comp, (W_total, H_total))
+
+            info_txt = f"Step={step_id} Hop={hop:.2f} Prog={prog:.2f}"
+            cv2.putText(frame_comp, info_txt, (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2)
+
+            frames_buffer.append(frame_comp)
+
+        # --- Write Video with MoviePy ---
+        if frames_buffer:
+            print(f"Generating video with {len(frames_buffer)} frames...")
+            # Create clip from list of numpy arrays (RGB)
+            clip = ImageSequenceClip(frames_buffer, fps=4.0)
+
+            # Write file
+            clip.write_videofile(str(out_path), logger=None)
+            print(f"Video saved to: {out_path}")
+        else:
+            print("[WARN] No frames generated.")
+
+    except Exception as e:
+        import traceback
+
+        traceback.print_exc()
+        print(f"[ERROR] {e}")
+
+
+# -----------------------------
+# Main Inference Class
+# -----------------------------
+
+
+class GRMInference:
+    def __init__(self, model_path: str, max_image_num=8, min_pixels=12544, max_pixels=76800):
+        print(f"Loading model from {model_path} ...")
+
+        self.model = LLM(
+            model=model_path,
+            gpu_memory_utilization=0.9,
+            max_model_len=8192,
+            limit_mm_per_prompt={"image": max_image_num},
+            enable_prefix_caching=True,
+            trust_remote_code=True,
+        )
+        self.sampling_params = SamplingParams(temperature=0.1, top_p=0.9, top_k=50, max_tokens=1024)
+
+        self.processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
+
+        if hasattr(self.processor, "image_processor"):
+            self.processor.image_processor.max_pixels = max_pixels
+            self.processor.image_processor.min_pixels = min_pixels
+
+    def inference_batch(self, batch_data: List[Dict]) -> List[Dict]:
+        prompts = []
+        for item in batch_data:
+            images = [Image.open(p).convert("RGB") for p in item["image"]]
+
+            messages = [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[0]},
+                        {"type": "image"},  # Ref Start
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[1]},
+                        {"type": "image"},  # Ref End
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[2]},
+                        {"type": "image"},  # BF High
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[3]},
+                        {"type": "image"},  # BF Left
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[4]},
+                        {"type": "image"},  # BF Right
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[5]},
+                        {"type": "image"},  # AF High
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[6]},
+                        {"type": "image"},  # AF Left
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[7]},
+                        {"type": "image"},  # AF Right
+                        {"type": "text", "text": SYSTEM_PROMPT.format(task=item["task"]).split("<image>")[8]},
+                    ],
+                }
+            ]
+
+            prompt_text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+            prompts.append({"prompt": prompt_text, "multi_modal_data": {"image": images}})
+
+        outputs = self.model.generate(prompts, sampling_params=self.sampling_params, use_tqdm=False)
+
+        results = []
+        for orig_item, out in zip(batch_data, outputs):
+            res_item = orig_item.copy()
+            res_item["pred"] = out.outputs[0].text
+            results.append(res_item)
+        return results
+
+    def run_pipeline(
+        self,
+        cam_high_path: str,
+        cam_left_path: str,
+        cam_right_path: str,
+        out_root: str,
+        task: str,
+        frame_interval: int = 10,
+        batch_size: int = 1,
+        goal_image: Optional[str] = None,
+        eval_mode: str = "incremental",
+        visualize: bool = False,
+    ) -> str:
+        """
+        Main entry point.
+        Args:
+            eval_mode: 'incremental', 'forward', or 'backward'
+        """
+        valid_modes = ["incremental", "forward", "backward"]
+        if eval_mode not in valid_modes:
+            raise ValueError(f"Invalid eval_mode '{eval_mode}'. Must be one of {valid_modes}")
+
+        out_root = Path(out_root)
+        ts = datetime.now().strftime("%y-%m-%d-%H-%M-%S")
+        run_root = out_root / ts
+        cache_root = run_root / ".cache"
+
+        cam_dirs = {
+            "cam_high": cache_root / "cam_high",
+            "cam_left_wrist": cache_root / "cam_left_wrist",
+            "cam_right_wrist": cache_root / "cam_right_wrist",
+        }
+        for d in cam_dirs.values():
+            ensure_dir(d)
+
+        paths = [Path(cam_high_path), Path(cam_left_path), Path(cam_right_path)]
+        types_counts = [get_frame_count(p) for p in paths]
+
+        counts = [tc[1] for tc in types_counts]
+        if len(set(counts)) != 1:
+            raise ValueError(f"Frame count mismatch among cameras: {counts}")
+        total_frames = counts[0]
+
+        indices = make_sample_indices_by_interval(total_frames, frame_interval)
+        print(f"Frames: {total_frames}, Int: {frame_interval}, Mode: {eval_mode}, Indices: {len(indices)}")
+
+        for p, key, (stype, _) in zip(paths, cam_dirs.keys(), types_counts):
+            save_frames(p, cam_dirs[key], indices, stype)
+
+        # Handle Goal Image / Ref End
+        if goal_image is not None and os.path.exists(goal_image):
+            ref_end_path = cache_root / "ref_end.png"
+            shutil.copy(goal_image, ref_end_path)
+            ref_end_path_str = str(ref_end_path)
+            print(f"Using Goal Image: {goal_image}")
+        else:
+            ref_end_path_str = str(cam_dirs["cam_high"] / f"frame_{total_frames - 1:06d}.png")
+            print("No Goal Image provided. Using last frame.")
+
+        # Build Samples based on Mode
+        samples = build_samples_json(run_root, task, indices, ref_end_path_str, mode=eval_mode)
+        json_path = run_root / "sample.json"
+        with open(json_path, "w", encoding="utf-8") as f:
+            json.dump(samples, f, indent=2)
+
+        print(f"Running inference on {len(samples)} samples...")
+        results = []
+
+        for i in tqdm(range(0, len(samples), batch_size)):
+            batch = samples[i : i + batch_size]
+            results.extend(self.inference_batch(batch))
+
+        # --- Post-Processing Logic based on Mode ---
+        prev_prog = 0.0
+
+        for idx, item in enumerate(results):
+            raw = item.get("pred", "")
+            try:
+                val_str = raw.split("<score>")[-1].split("</score>")[0].replace("%", "").strip()
+                raw_score = max(-100.0, min(100.0, float(val_str))) / 100.0
+            except Exception:
+                print(f"[ERR] Failed to parse score for item {idx}: {raw}")
+                raw_score = 0.0
+
+            curr_progress = 0.0
+            hop = 0.0
+
+            if eval_mode == "incremental":
+                # Original Logic: raw_score is incremental change
+                if idx == 0:
+                    curr_progress = raw_score
+                else:
+                    if raw_score >= 0:
+                        curr_progress = prev_prog + (1 - prev_prog) * raw_score
+                    else:
+                        curr_progress = prev_prog + prev_prog * raw_score
+                hop = raw_score  # In incremental, Model Output IS the Hop signal
+
+            elif eval_mode == "forward":
+                # Forward: Model Output IS Progress (Start -> Current)
+                curr_progress = raw_score
+                # Hop is the change in progress from previous step
+                hop = curr_progress - prev_prog
+
+            elif eval_mode == "backward":
+                # Backward: Compare Goal -> Current
+                # raw_score is likely negative (Current is worse than Goal)
+                # Formula: Progress = 1 + ModelOutput
+                curr_progress = 1.0 + raw_score
+                # Hop is the change in progress
+                hop = curr_progress - prev_prog
+
+            # Clamp progress to reasonable bounds [0, 1] for safety?
+            # Not strictly enforcing clamping to allow debugging, but typically progress is 0-1.
+            # curr_progress = max(0.0, min(1.0, curr_progress))
+
+            item["hop"] = hop
+            item["progress"] = curr_progress
+
+            prev_prog = curr_progress
+
+        result_path = run_root / "pred_vllm.json"
+        with open(result_path, "w", encoding="utf-8") as f:
+            json.dump(results, f, indent=2, ensure_ascii=False)
+        print(f"Done. Results saved to {result_path}")
+
+        if visualize:
+            print("Generating visualization video...")
+            plot_video_reward(run_root)
+
+        return str(run_root)
+
+
+# -----------------------------
+# Entry Point
+# -----------------------------
+
+if __name__ == "__main__":
+    MODEL_PATH = "../checkpoints/Robo-Dopamine-GRM-3B"
+    TASK_INSTRUCTION = "organize the table"
+    BASE_DEMO_PATH = "./demo_table"
+    GOAL_IMAGE_PATH = None  # "./demo_table/goal_image.png"
+    OUTPUT_ROOT = "../results"
+
+    agent = GRMInference(model_path=MODEL_PATH, max_image_num=8)
+
+    print("## =====================================")
+    print("## (1) Example for Incremental-Mode")
+    print("## =====================================")
+    mode = "incremental"  # "incremental", "forward", "backward"
+
+    output_dir = agent.run_pipeline(
+        cam_high_path=os.path.join(BASE_DEMO_PATH, "cam_high.mp4"),
+        cam_left_path=os.path.join(BASE_DEMO_PATH, "cam_left_wrist.mp4"),
+        cam_right_path=os.path.join(BASE_DEMO_PATH, "cam_right_wrist.mp4"),
+        out_root=OUTPUT_ROOT,
+        task=TASK_INSTRUCTION,
+        frame_interval=30,
+        batch_size=10,
+        goal_image=GOAL_IMAGE_PATH,
+        eval_mode=mode,
+        visualize=True,
+    )
+
+    print(f"Episode ({BASE_DEMO_PATH}) processed ({mode}). Output at: {output_dir}")
+
+    print("## =====================================")
+    print("## (2) Example for Forward-Mode")
+    print("## =====================================")
+    mode = "forward"  # "incremental", "forward", "backward"
+
+    output_dir = agent.run_pipeline(
+        cam_high_path=os.path.join(BASE_DEMO_PATH, "cam_high.mp4"),
+        cam_left_path=os.path.join(BASE_DEMO_PATH, "cam_left_wrist.mp4"),
+        cam_right_path=os.path.join(BASE_DEMO_PATH, "cam_right_wrist.mp4"),
+        out_root=OUTPUT_ROOT,
+        task=TASK_INSTRUCTION,
+        frame_interval=30,
+        batch_size=10,
+        goal_image=GOAL_IMAGE_PATH,
+        eval_mode=mode,
+        visualize=True,
+    )
+
+    print(f"Episode ({BASE_DEMO_PATH}) processed ({mode}). Output at: {output_dir}")
+
+    print("## =====================================")
+    print("## (3) Example for Backward-Mode")
+    print("## =====================================")
+    mode = "backward"  # "incremental", "forward", "backward"
+
+    output_dir = agent.run_pipeline(
+        cam_high_path=os.path.join(BASE_DEMO_PATH, "cam_high.mp4"),
+        cam_left_path=os.path.join(BASE_DEMO_PATH, "cam_left_wrist.mp4"),
+        cam_right_path=os.path.join(BASE_DEMO_PATH, "cam_right_wrist.mp4"),
+        out_root=OUTPUT_ROOT,
+        task=TASK_INSTRUCTION,
+        frame_interval=30,
+        batch_size=10,
+        goal_image=GOAL_IMAGE_PATH,
+        eval_mode=mode,
+        visualize=True,
+    )
+
+    print(f"Episode ({BASE_DEMO_PATH}) processed ({mode}). Output at: {output_dir}")
diff --git a/robometer/robometer/evals/baselines/rbm_model.py b/robometer/robometer/evals/baselines/rbm_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..4505c191f78720d5ee73f6d989e846372d2f0ab8
--- /dev/null
+++ b/robometer/robometer/evals/baselines/rbm_model.py
@@ -0,0 +1,364 @@
+#!/usr/bin/env python3
+"""
+RBM and ReWiND model class for baseline evaluation.
+
+This class provides a unified interface for loading RBM/ReWiND models from checkpoints
+and computing progress and preference predictions.
+"""
+
+import time
+import torch
+import numpy as np
+from typing import List, Dict, Any, Union
+
+from robometer.utils.setup_utils import setup_batch_collator
+from robometer.utils.save import load_model_from_hf
+from robometer.data.dataset_types import ProgressSample, PreferenceSample
+from robometer.data.datasets.helpers import create_trajectory_from_dict
+from robometer.evals.eval_server import forward_model
+from robometer.utils.logger import get_logger, setup_loguru_logging
+from robometer.models.utils import convert_bins_to_continuous
+
+logger = get_logger()
+
+setup_loguru_logging("TRACE")
+
+
+class RBMModel:
+    """RBM/ReWiND model for baseline evaluation with unified compute methods."""
+
+    def __init__(self, checkpoint_path: str):
+        """Initialize the RBM/ReWiND model wrapper.
+
+        Args:
+            checkpoint_path: Path to model checkpoint (HuggingFace repo ID or local path)
+                           The config.yaml will be loaded from the checkpoint automatically
+        """
+        self.checkpoint_path = checkpoint_path
+
+        # Automatically determine device (cuda:0 if available, else cpu)
+        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        self.device = device
+
+        # Load model, config, processor, and tokenizer using the helper function
+        # This handles loading config.yaml from checkpoint and setting up everything
+        logger.info(f"Loading model from checkpoint: {checkpoint_path}")
+        exp_config, tokenizer, processor, model = load_model_from_hf(
+            model_path=checkpoint_path,
+            device=device,
+        )
+
+        # Store loaded components
+        self.exp_config = exp_config
+        self.model = model
+        self.processor = processor
+        self.tokenizer = tokenizer
+
+        # Optimize model for inference
+        self.model.eval()  # Set to evaluation mode (disables dropout, batch norm updates, etc.)
+
+        # Enable cuDNN benchmarking for faster inference (if using CUDA)
+        if self.device.type == "cuda":
+            torch.backends.cudnn.benchmark = True
+            torch.backends.cudnn.deterministic = False
+
+        # Determine if this is ReWiND or RBM
+        self.is_rewind = "rewind" in exp_config.model.base_model_id.lower()
+        logger.info(f"Model type: {'ReWiND' if self.is_rewind else 'RBM'}")
+
+        # Create batch collator using the loaded config
+        self.batch_collator = setup_batch_collator(
+            processor=processor,
+            tokenizer=tokenizer,
+            cfg=exp_config,
+            is_eval=True,
+        )
+
+        logger.info(f"Model loaded successfully on device: {self.device}")
+
+    def compute_progress(
+        self, frames_array: Union[List, torch.Tensor, np.ndarray], task_description: str = ""
+    ) -> List[float]:
+        """Compute progress prediction for a trajectory.
+
+        Args:
+            frames_array: Array of frames (can be list, tensor, or numpy array)
+            task_description: Task description text
+
+        Returns:
+            List of progress values (0-1) for each frame
+        """
+        # Create a ProgressSample from the inputs
+        traj_dict = {
+            "frames": frames_array,
+            "task": task_description,
+            "num_frames": len(frames_array) if hasattr(frames_array, "__len__") else frames_array.shape[0],
+        }
+
+        trajectory = create_trajectory_from_dict(traj_dict)
+        sample = ProgressSample(trajectory=trajectory)
+
+        # Collate into batch
+        batch_inputs = self.batch_collator([sample])
+
+        # Extract progress_inputs from batch_inputs (batch_collator returns nested structure)
+        progress_inputs = batch_inputs["progress_inputs"]
+
+        # Move to device
+        progress_inputs = {
+            k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k, v in progress_inputs.items()
+        }
+
+        # Forward pass with inference mode for additional optimization
+        with torch.inference_mode():  # Faster than torch.no_grad() for inference-only code
+            model_output, _ = forward_model(self.model, progress_inputs, sample_type="progress")
+
+        # Extract progress logits
+        progress_logits = model_output.progress_logits
+        if progress_logits is None:
+            raise ValueError("No progress logits returned from model")
+
+        # Handle different output formats
+        if isinstance(progress_logits, dict):
+            # RBM format: {"A": tensor, "B": None}
+            progress_tensor = progress_logits.get("A")
+        else:
+            # Direct tensor
+            progress_tensor = progress_logits
+
+        if progress_tensor is None:
+            raise ValueError("No progress logits in 'A' key")
+
+        # Convert to list of floats
+        progress_values = progress_tensor.squeeze().cpu().tolist()
+
+        # Ensure we have the right length
+        if isinstance(progress_values, float):
+            progress_values = [progress_values]
+
+        return progress_values
+
+    def compute_preference(
+        self, chosen_images: List, rejected_images: List, task_description: str = ""
+    ) -> Dict[str, Any]:
+        """Compute preference prediction between two trajectories.
+
+        Args:
+            chosen_images: List of images/frames for the chosen trajectory
+            rejected_images: List of images/frames for the rejected trajectory
+            task_description: Task description text
+
+        Returns:
+            Dictionary containing:
+            - prediction_prob: Probability that chosen is preferred (0.0 to 1.0)
+            - is_correct: True if prediction matches ground truth (always True for chosen)
+            - preference_pred: Binary prediction (1.0 if chosen is preferred, 0.0 otherwise)
+            - Other metadata
+        """
+        start_time = time.time()
+
+        # Create trajectories
+        chosen_traj = create_trajectory_from_dict({
+            "frames": chosen_images,
+            "task": task_description,
+            "num_frames": len(chosen_images),
+        })
+        rejected_traj = create_trajectory_from_dict({
+            "frames": rejected_images,
+            "task": task_description,
+            "num_frames": len(rejected_images),
+        })
+
+        # Create PreferenceSample
+        sample = PreferenceSample(
+            chosen_trajectory=chosen_traj,
+            rejected_trajectory=rejected_traj,
+        )
+
+        # Collate into batch
+        batch_inputs = self.batch_collator([sample])
+
+        # Extract preference_inputs from batch_inputs (batch_collator returns nested structure)
+        preference_inputs = batch_inputs["preference_inputs"]
+
+        # Move to device
+        preference_inputs = {
+            k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k, v in preference_inputs.items()
+        }
+
+        # Forward pass with inference mode for additional optimization
+        with torch.inference_mode():  # Faster than torch.no_grad() for inference-only code
+            model_output, _ = forward_model(self.model, preference_inputs, sample_type="preference")
+
+        # Extract preference logits
+        pref_logits = model_output.pref_logits
+        if pref_logits is None:
+            raise ValueError("No preference logits returned from model")
+
+        # Convert logits to probability
+        pref_probs = torch.sigmoid(pref_logits)
+        prediction_prob = pref_probs.item()
+        preference_pred = 1.0 if prediction_prob > 0.5 else 0.0
+
+        processing_time = time.time() - start_time
+
+        # Build result dict (matching RLVLMF format)
+        result = {
+            "is_correct": True,  # Chosen is always preferred by construction
+            "prediction_prob": float(prediction_prob),
+            "preference_pred": float(preference_pred),
+            "preference_logits": float(pref_logits.item()) if pref_logits is not None else None,
+            "timestamp": time.strftime("%Y-%m-%d %H:%M:%S"),
+            "task": task_description,
+            "num_chosen_frames": len(chosen_images),
+            "num_rejected_frames": len(rejected_images),
+            "processing_time_seconds": processing_time,
+        }
+
+        return result
+
+    def compute_batched_progress(self, samples: List[ProgressSample]) -> List[List[float]]:
+        """Compute progress predictions for a batch of trajectories.
+
+        Args:
+            samples: List of ProgressSample objects
+
+        Returns:
+            List of progress value lists (each inner list contains 0-1 values for each frame)
+        """
+        if not samples:
+            return []
+
+        # Collate into batch
+        batch_inputs = self.batch_collator(samples)
+
+        # Extract progress_inputs from batch_inputs (batch_collator returns nested structure)
+        progress_inputs = batch_inputs["progress_inputs"]
+
+        # Move to device
+        progress_inputs = {
+            k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k, v in progress_inputs.items()
+        }
+
+        # Forward pass with inference mode for additional optimization
+        with torch.inference_mode():  # Faster than torch.no_grad() for inference-only code
+            time_start = time.time()
+            model_output, _ = forward_model(self.model, progress_inputs, sample_type="progress")
+            time_end = time.time()
+            print(f"Time taken for forward pass: {time_end - time_start} seconds")
+
+        # Extract progress logits
+        progress_logits = model_output.progress_logits
+        if progress_logits is None:
+            raise ValueError("No progress logits returned from model")
+
+        # Handle different output formats
+        if isinstance(progress_logits, dict):
+            # RBM format: {"A": tensor, "B": None}
+            progress_tensor = progress_logits.get("A")
+        else:
+            # Direct tensor
+            progress_tensor = progress_logits
+
+        if progress_tensor is None:
+            raise ValueError("No progress logits in 'A' key")
+
+        # progress_tensor shape: [batch_size, num_frames] or [batch_size, 1]
+        batch_size = progress_tensor.shape[0]
+        results = []
+
+        for i in range(batch_size):
+            # Extract progress values for this sample
+            if progress_tensor.ndim == 2:
+                if progress_tensor.shape[1] == 1:
+                    # Single value per sample
+                    progress_values = [float(progress_tensor[i, 0].item())]
+                else:
+                    # Multiple values per sample
+                    progress_values = progress_tensor[i].cpu().tolist()
+            elif progress_tensor.ndim == 3:
+                # Multiple values per sample, discrete multiple bins, convert to continuous
+                progress_values = convert_bins_to_continuous(progress_tensor[i]).cpu().tolist()
+            else:
+                # Unexpected shape
+                raise ValueError(f"Unexpected progress_tensor shape: {progress_tensor.shape}")
+
+            # Ensure we have the right length
+            if isinstance(progress_values, float):
+                progress_values = [progress_values]
+
+            results.append(progress_values)
+
+        return results
+
+    def compute_batched_preference(self, samples: List[PreferenceSample]) -> List[Dict[str, Any]]:
+        """Compute preference predictions for a batch of trajectory pairs.
+
+        Args:
+            samples: List of PreferenceSample objects
+
+        Returns:
+            List of result dictionaries, each containing:
+            - prediction_prob: Probability that chosen is preferred (0.0 to 1.0)
+            - is_correct: True if prediction matches ground truth (always True for chosen)
+            - preference_pred: Binary prediction (1.0 if chosen is preferred, 0.0 otherwise)
+            - Other metadata
+        """
+        if not samples:
+            return []
+
+        start_time = time.time()
+
+        # Collate into batch
+        batch_inputs = self.batch_collator(samples)
+
+        # Extract preference_inputs from batch_inputs (batch_collator returns nested structure)
+        preference_inputs = batch_inputs["preference_inputs"]
+
+        # Move to device
+        preference_inputs = {
+            k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k, v in preference_inputs.items()
+        }
+
+        # Forward pass with inference mode for additional optimization
+        with torch.inference_mode():  # Faster than torch.no_grad() for inference-only code
+            model_output, _ = forward_model(self.model, preference_inputs, sample_type="preference")
+
+        # Extract preference logits
+        pref_logits = model_output.pref_logits
+        if pref_logits is None:
+            raise ValueError("No preference logits returned from model")
+
+        # pref_logits shape: [batch_size]
+        batch_size = pref_logits.shape[0]
+        processing_time = time.time() - start_time
+
+        # Convert logits to probabilities
+        pref_probs = torch.sigmoid(pref_logits)
+        binary_preds = (pref_probs > 0.5).float()
+
+        results = []
+        for i in range(batch_size):
+            sample = samples[i]
+            prediction_prob = float(pref_probs[i].item())
+            preference_pred = float(binary_preds[i].item())
+            pref_logit = float(pref_logits[i].item())
+
+            result = {
+                "is_correct": True,  # Chosen is always preferred by construction
+                "prediction_prob": prediction_prob,
+                "preference_pred": preference_pred,
+                "preference_logits": pref_logit,
+                "timestamp": time.strftime("%Y-%m-%d %H:%M:%S"),
+                "task": sample.chosen_trajectory.task if sample.chosen_trajectory.task else "",
+                "num_chosen_frames": len(sample.chosen_trajectory.frames)
+                if sample.chosen_trajectory.frames is not None
+                else 0,
+                "num_rejected_frames": len(sample.rejected_trajectory.frames)
+                if sample.rejected_trajectory.frames is not None
+                else 0,
+                "processing_time_seconds": processing_time / batch_size,  # Average time per sample
+            }
+            results.append(result)
+
+        return results
diff --git a/robometer/robometer/evals/baselines/rlvlmf.py b/robometer/robometer/evals/baselines/rlvlmf.py
new file mode 100644
index 0000000000000000000000000000000000000000..f747a0279b6d4378f82db578ad2b16e1516ca94a
--- /dev/null
+++ b/robometer/robometer/evals/baselines/rlvlmf.py
@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+"""Direct VLM baseline for preference comparison."""
+
+import os
+import time
+import base64
+import random
+from io import BytesIO
+from typing import List, Dict, Any, Optional, Tuple
+from PIL import Image
+import numpy as np
+
+try:
+    import google.generativeai as genai
+
+    GEMINI_AVAILABLE = True
+except ImportError:
+    GEMINI_AVAILABLE = False
+
+try:
+    import openai
+
+    OPENAI_AVAILABLE = True
+except ImportError:
+    OPENAI_AVAILABLE = False
+
+
+class RLVLMF:
+    """RL-VLM-F baseline for preference comparison between trajectories.
+
+    Uses vision-language models (Gemini or GPT-4V) to directly compare two trajectories
+    and predict which one better achieves a given task goal.
+    """
+
+    def __init__(self, vlm_provider: str = "gemini", temperature: float = 0.0):
+        self.vlm_provider = vlm_provider
+        self.temperature = temperature
+
+        # Setup VLM
+        self._setup_vlm()
+
+    def _setup_vlm(self):
+        """Initialize VLM provider."""
+        if self.vlm_provider == "gemini":
+            api_key = os.environ.get("GEMINI_API_KEY")
+            if not api_key:
+                raise ValueError("Set GEMINI_API_KEY environment variable")
+
+            genai.configure(api_key=api_key)
+            # Original RL-VLM-F used 'gemini-pro-vision' (deprecated); we use a supported model
+            # Using latest flash model for speed and cost-effectiveness
+            self.model = genai.GenerativeModel("gemini-2.5-flash")
+
+        elif self.vlm_provider == "openai":
+            api_key = os.environ.get("OPENAI_API_KEY")
+            if not api_key:
+                raise ValueError("Set OPENAI_API_KEY environment variable")
+
+            openai.api_key = api_key
+            self.client = openai.OpenAI()
+        else:
+            raise ValueError(f"Unknown provider: {self.vlm_provider}")
+
+    def compute_preference(
+        self, chosen_images: List[Image.Image], rejected_images: List[Image.Image], task_description: str = ""
+    ) -> Dict[str, Any]:
+        """Compute preference prediction between trajectories using VLM."""
+        start_time = time.time()
+
+        # DEBUG: Print task that reaches VLM baseline
+        print(f"VLM Baseline received task: '{task_description}'")
+
+        # Select last frame from each trajectory
+        chosen_frame = chosen_images[-1]
+        rejected_frame = rejected_images[-1]
+
+        # Randomize image order to avoid position bias
+        # chosen_is_first = True means chosen goes to Image A, False means chosen goes to Image B
+        chosen_is_first = random.choice([True, False])
+
+        if chosen_is_first:
+            image_a, image_b = chosen_frame, rejected_frame
+            image_a_label, image_b_label = "chosen", "rejected"
+        else:
+            image_a, image_b = rejected_frame, chosen_frame
+            image_a_label, image_b_label = "rejected", "chosen"
+
+        # Query VLM with randomized images
+        prompt = self._build_prompt(task_description)
+
+        if self.vlm_provider == "gemini":
+            preference, raw_response = self._query_gemini(prompt, image_a, image_b)
+        else:
+            preference, raw_response = self._query_openai(prompt, image_a, image_b)
+
+        # Process result - adjust for randomized order
+        # preference is "A"/"B"/"tie" based on VLM's choice of Image A vs Image B
+        # We need to map this back to chosen vs rejected based on randomization
+        if preference == "A":
+            # VLM chose Image A
+            vlm_chose_chosen = chosen_is_first  # If chosen is first, A=chosen, else A=rejected
+        elif preference == "B":
+            # VLM chose Image B
+            vlm_chose_chosen = not chosen_is_first  # If chosen is first, B=rejected, else B=chosen
+        else:  # preference == "tie"
+            vlm_chose_chosen = None  # Tie case
+
+        # Determine correctness
+        if vlm_chose_chosen is True:
+            is_correct = True  # VLM correctly chose the chosen trajectory
+        elif vlm_chose_chosen is None:
+            # For ties, treating as incorrect since we expect clear preferences
+            is_correct = False
+        else:  # vlm_chose_chosen is False
+            is_correct = False  # VLM incorrectly chose the rejected trajectory
+
+        # Convert preference to probability
+        # "A" means chosen was preferred, "B" means rejected was preferred, "tie" is 0.5
+        if preference == "A":
+            prediction_prob = 1.0  # Chosen is preferred
+        elif preference == "B":
+            prediction_prob = 0.0  # Rejected is preferred
+        else:  # tie
+            prediction_prob = 0.5
+
+        # preference_pred: 1 if chosen is selected, 0 otherwise
+        preference_pred = 1.0 if vlm_chose_chosen is True else 0.0
+
+        processing_time = time.time() - start_time
+
+        # Build result with all metadata
+        result = {
+            "is_correct": is_correct,
+            "vlm_preference": preference,
+            "prediction_prob": prediction_prob,
+            "preference_pred": preference_pred,
+            "timestamp": time.strftime("%Y-%m-%d %H:%M:%S"),
+            "task": task_description,
+            "num_chosen_frames": len(chosen_images),
+            "num_rejected_frames": len(rejected_images),
+            "selected_frames": 1,
+            "strategy": "last_frame",
+            "success": True,
+            "error": None,
+            "vlm_response": raw_response[:500] + "...",
+            "vlm_chose_chosen": vlm_chose_chosen,
+            "chosen_is_first": chosen_is_first,
+            "prompt": prompt[:200] + "...",
+            "processing_time_seconds": processing_time,
+        }
+
+        return result
+
+    def _build_prompt(self, task: str) -> str:
+        """Build RL-VLM-F prompt - exact match to original paper."""
+        base = """ Each frame comes from a robot trajectory. (Think causally and use image comparison to verify any confusion between the base of the robot and the end effector)
+1. What is shown in the first image (Image A)?
+2. What is shown in the second image (Image B)?
+3. For this question here is the Goal Text: {goal_text}  
+Is the goal being better achieved in Image A or Image B? 
+Reply a single line of 0 if the goal is better achieved in Image A, or 1 if it is better achieved in Image B.
+Reply -1 if the text is really unsure about either making any progress towards the goal or there is absolutely no difference. (only use this if there is no discernible signs of progress in either image or no discernible difference between the progress in the two images for the given task)
+"""
+
+        if task:
+            goal_text = f"The goal is {task}."
+        else:
+            goal_text = "Which image shows better task execution?"
+
+        return base.format(goal_text=goal_text)
+
+    def _query_gemini(self, prompt: str, chosen: Image.Image, rejected: Image.Image) -> Tuple[str, str]:
+        """Query Gemini for preference."""
+        query = ["Consider the following two images:\nImage 1:", chosen, "Image 2:", rejected, prompt]
+
+        response = self.model.generate_content(
+            query,
+            generation_config=genai.types.GenerationConfig(temperature=self.temperature),
+            safety_settings=[
+                {"category": cat, "threshold": "BLOCK_NONE"}
+                for cat in [
+                    "HARM_CATEGORY_HARASSMENT",
+                    "HARM_CATEGORY_HATE_SPEECH",
+                    "HARM_CATEGORY_SEXUALLY_EXPLICIT",
+                    "HARM_CATEGORY_DANGEROUS",
+                ]
+            ],
+        )
+
+        full_response = response.text
+        # Fix parsing: strip whitespace first, then split and take last non-empty line
+        result = full_response.strip()
+        if "\n" in result:
+            lines = [line.strip() for line in result.split("\n") if line.strip()]
+            result = lines[-1] if lines else ""
+
+        # Parse response
+        if "-1" in result:
+            return "tie", full_response
+        elif "0" in result:
+            return "A", full_response  # Image 1 (chosen)
+        elif "1" in result:
+            return "B", full_response  # Image 2 (rejected)
+        else:
+            return "tie", full_response
+
+    def _query_openai(self, prompt: str, chosen: Image.Image, rejected: Image.Image) -> Tuple[str, str]:
+        """Query GPT-4V for preference."""
+
+        def to_base64(img: Image.Image) -> str:
+            buffer = BytesIO()
+            img.save(buffer, format="JPEG")
+            return base64.b64encode(buffer.getvalue()).decode()
+
+        content = [
+            {"type": "text", "text": "Consider the following two images:\nImage 1:"},
+            {
+                "type": "image_url",
+                "image_url": {"url": f"data:image/jpeg;base64,{to_base64(chosen)}", "detail": "high"},
+            },
+            {"type": "text", "text": "Image 2:"},
+            {
+                "type": "image_url",
+                "image_url": {"url": f"data:image/jpeg;base64,{to_base64(rejected)}", "detail": "high"},
+            },
+            {"type": "text", "text": prompt},
+        ]
+
+        response = self.client.chat.completions.create(
+            model="gpt-4-vision-preview",
+            messages=[{"role": "user", "content": content}],
+            temperature=self.temperature,
+            max_tokens=1000,
+        )
+
+        full_response = response.choices[0].message.content.strip()
+        result = full_response.split("\n")[-1].strip()
+
+        # Parse response
+        if "-1" in result:
+            return "tie", full_response
+        elif "0" in result:
+            return "A", full_response
+        elif "1" in result:
+            return "B", full_response
+        else:
+            return "tie", full_response
diff --git a/robometer/robometer/evals/baselines/robodopamine.py b/robometer/robometer/evals/baselines/robodopamine.py
new file mode 100644
index 0000000000000000000000000000000000000000..9a70d432d018d93f92a627e6d1fc86762ee7953e
--- /dev/null
+++ b/robometer/robometer/evals/baselines/robodopamine.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+"""Robo-Dopamine (GRM) baseline for progress prediction.
+
+Reference: https://github.com/FlagOpen/Robo-Dopamine
+Model: https://huggingface.co/tanhuajie2001/Robo-Dopamine-GRM-3B
+Setup: Clone Robo-Dopamine and set ROBODOPAMINE_PATH=/path/to/Robo-Dopamine
+"""
+
+import os
+import json
+import tempfile
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import cv2
+
+from robometer.utils.logger import get_logger
+from robometer.evals.baselines.rbd_inference import GRMInference
+
+logger = get_logger()
+
+
+class RoboDopamine:
+    """Robo-Dopamine GRM baseline. Uses single-view frames for all three camera inputs."""
+
+    def __init__(
+        self,
+        model_path: str = "tanhuajie2001/Robo-Dopamine-GRM-3B",
+        frame_interval: int = 1,
+        batch_size: int = 1,
+        eval_mode: str = "incremental",
+    ):
+        self.model_path = model_path
+        self.frame_interval = frame_interval
+        self.batch_size = batch_size
+        self.eval_mode = eval_mode
+        self._grm = GRMInference(model_path=model_path, max_image_num=8)
+
+    def compute_progress(
+        self,
+        frames_array: np.ndarray,
+        task_description: str = "",
+        reference_video_path: Optional[str] = None,
+    ) -> np.ndarray:
+        if frames_array is None or frames_array.size == 0:
+            return np.array([], dtype=np.float64)
+
+        num_frames = frames_array.shape[0]
+        with tempfile.TemporaryDirectory() as tmpdir:
+            frames_dir = Path(tmpdir) / "frames"
+            frames_dir.mkdir(parents=True, exist_ok=True)
+            for i in range(num_frames):
+                frame = frames_array[i]
+                if frame.dtype != np.uint8:
+                    frame = np.clip(frame, 0, 255).astype(np.uint8)
+                frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+                cv2.imwrite(
+                    str(frames_dir / f"frame_{i:06d}.png"),
+                    frame_bgr,
+                    [int(cv2.IMWRITE_PNG_COMPRESSION), 3],
+                )
+
+            out_root = Path(tmpdir) / "out"
+            out_root.mkdir(parents=True, exist_ok=True)
+            # run_pipeline accepts either: (1) a dir of .png files (sorted by name), or (2) a video path (.mp4).
+            # We pass a directory of frame_000000.png, frame_000001.png, ...
+            # frame_interval: step between sampled frames (0, interval, 2*interval, ...; last frame always included).
+            # So frame_interval=1 uses every frame; frame_interval=5 uses every 5th frame.
+            run_root = self._grm.run_pipeline(
+                cam_high_path=str(frames_dir),
+                cam_left_path=str(frames_dir),
+                cam_right_path=str(frames_dir),
+                out_root=str(out_root),
+                task=task_description,
+                frame_interval=self.frame_interval,
+                batch_size=self.batch_size,
+                goal_image=None,
+                eval_mode=self.eval_mode,
+                visualize=False,
+            )
+
+            pred_path = Path(run_root) / "pred_vllm.json"
+            with open(pred_path, "r", encoding="utf-8") as f:
+                results = json.load(f)
+
+        progress_list = [0.0]
+        for item in results:
+            p = item.get("progress", 0.0)
+            if isinstance(p, str) and p == "Error":
+                p = progress_list[-1] if progress_list else 0.0
+            progress_list.append(float(p))
+
+        progress_arr = np.clip(np.array(progress_list, dtype=np.float64), 0.0, 1.0)
+        if len(progress_arr) < num_frames:
+            progress_arr = np.pad(
+                progress_arr,
+                (0, num_frames - len(progress_arr)),
+                mode="edge",
+            )
+        elif len(progress_arr) > num_frames:
+            progress_arr = progress_arr[:num_frames]
+
+        return progress_arr
diff --git a/robometer/robometer/evals/baselines/roboreward.py b/robometer/robometer/evals/baselines/roboreward.py
new file mode 100644
index 0000000000000000000000000000000000000000..c4b33afe30fb09e75063c863e7fc9692b7ea44e9
--- /dev/null
+++ b/robometer/robometer/evals/baselines/roboreward.py
@@ -0,0 +1,446 @@
+#!/usr/bin/env python3
+"""
+RoboReward baseline for discrete end-of-episode progress reward prediction.
+
+RoboReward predicts discrete scores (1-5) for task completion:
+- 1: No success
+- 2: Minimal progress
+- 3: Partial completion
+- 4: Near completion
+- 5: Perfect completion
+
+Based on: https://huggingface.co/teetone/RoboReward-8B
+"""
+
+import re
+import tempfile
+import uuid
+from pathlib import Path
+from typing import List, Optional, Union
+import numpy as np
+from PIL import Image
+import torch
+import shutil
+
+from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
+from qwen_vl_utils import process_vision_info
+
+try:
+    from unsloth import FastVisionModel
+
+    HAS_UNSLOTH = True
+except ImportError:
+    HAS_UNSLOTH = False
+
+from robometer.data.collators.utils import convert_frames_to_pil_images, write_mp4
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class RoboReward:
+    """RoboReward baseline for discrete end-of-episode progress reward prediction."""
+
+    def __init__(
+        self,
+        model_path: str = "teetone/RoboReward-8B",
+        max_new_tokens: int = 128,
+        use_unsloth: bool = True,
+    ):
+        """
+        Initialize RoboReward model.
+
+        Args:
+            model_path: HuggingFace model path (e.g., "teetone/RoboReward-8B" or "teetone/RoboReward-4B")
+            max_new_tokens: Maximum number of tokens to generate
+            use_unsloth: Whether to use unsloth for faster inference (default: True)
+        """
+        logger.info(f"Loading RoboReward model: {model_path}")
+
+        # Use unsloth for faster inference if available and requested
+        if use_unsloth and HAS_UNSLOTH:
+            print("Using Unsloth for faster inference")
+            # Load model with unsloth's FastVisionModel
+            self.model, _ = FastVisionModel.from_pretrained(
+                model_path,
+                dtype=torch.bfloat16,
+                device_map="auto",
+                full_finetuning=False,  # Inference only
+            )
+        else:
+            # Standard loading
+            if use_unsloth and not HAS_UNSLOTH:
+                print("Warning: Unsloth requested but not available, using standard loading")
+            self.model = Qwen3VLForConditionalGeneration.from_pretrained(
+                model_path,
+                torch_dtype=torch.bfloat16,
+                device_map="auto",  # Auto device placement is best practice
+            )
+
+        self.processor = AutoProcessor.from_pretrained(
+            model_path, trust_remote_code=True, do_sample_frames=False, fps=1
+        )
+        self.max_new_tokens = max_new_tokens
+        self.model_path = model_path
+
+        print(f"RoboReward model loaded on device: {self.model.device}")
+
+    def _build_prompt(self, task_description: str) -> str:
+        """Build the prompt for RoboReward inference.
+
+        Args:
+            task_description: Task instruction text
+
+        Returns:
+            Formatted prompt string
+        """
+        prompt = """Given the task, assign a discrete progress score reward (1,2,3,4,5) for the robot in the video in the format: ANSWER: <score>
+Rubric for end-of-episode progress (judge only the final state without time limits):
+1 - No Success: Final state shows no goal-relevant change for the command.
+2 - Minimal Progress: Final state shows a small but insufficient change toward the goal.
+3 - Partial Completion: The final state shows good progress toward the goal but violates more than one requirement or a major requirement.
+4 - Near Completion: Final state is correct in region and intent but misses a single minor requirement.
+5 - Perfect Completion: Final state satisfies all requirements.
+
+Task: {task}""".format(task=task_description)
+        return prompt
+
+    def _parse_score(self, output_text: str) -> Optional[int]:
+        """Parse discrete score (1-5) from model output.
+
+        Args:
+            output_text: Model output text
+
+        Returns:
+            Discrete score (1-5) or None if parsing fails
+        """
+        # Look for "ANSWER: <number>" pattern
+        pattern = r"ANSWER:\s*(\d+)"
+        match = re.search(pattern, output_text, re.IGNORECASE)
+        if match:
+            score = int(match.group(1))
+            if 1 <= score <= 5:
+                return score
+
+        # Fallback: look for any single digit 1-5 in the text
+        pattern = r"\b([1-5])\b"
+        matches = re.findall(pattern, output_text)
+        if matches:
+            # Take the last occurrence (most likely the answer)
+            score = int(matches[-1])
+            if 1 <= score <= 5:
+                return score
+
+        return None
+
+    def compute_progress(self, frames_array: np.ndarray, task_description: str = "") -> List[Optional[float]]:
+        """
+        Compute progress prediction for a frame sequence using RoboReward baseline.
+
+        RoboReward predicts a discrete score (1-5) for the end-of-episode state.
+        Since the sampler already uses use_frame_steps to create progressively longer
+        sequences, we just process the single sequence provided here.
+
+        Args:
+            frames_array: (N, H, W, 3) uint8 array from trajectory frames (already a subsequence)
+            task_description: Task description text
+
+        Returns:
+            List of discrete scores (1.0-5.0) for each frame.
+            All frames get the same discrete score (end-of-episode score for this subsequence).
+        """
+        if frames_array is None or frames_array.size == 0:
+            return []
+
+        # Convert frames to PIL Images
+        frames_pil = convert_frames_to_pil_images(frames_array)
+
+        logger.info(f"RoboReward: Converted {len(frames_pil)} frames to PIL Images")
+
+        if not frames_pil:
+            return []
+
+        num_frames = len(frames_pil)
+
+        # Ensure at least 2 frames for video processing (qwen_vl_utils requires minimum 2 frames)
+        if num_frames == 1:
+            # Duplicate the single frame to make it 2 frames
+            frames_pil = [frames_pil[0], frames_pil[0]]
+            num_frames = 2
+
+        # Build prompt
+        prompt = self._build_prompt(task_description)
+
+        # Create temporary directory for frame files
+        # Use individual frame files instead of video to avoid torchcodec memory issues
+        # According to qwen-vl-utils docs, we can pass frames as a list of file paths
+        tmpdir = tempfile.mkdtemp()
+        unique_id = uuid.uuid4().hex
+
+        # Save frames as individual JPEG files (much smaller than video, avoids torchcodec overhead)
+        frame_paths = []
+        for i, frame_pil in enumerate(frames_pil):
+            frame_path = Path(tmpdir) / f"roboreward_{unique_id}_frame_{i:04d}.jpg"
+            # Save as JPEG with reasonable quality to reduce file size
+            frame_pil.save(frame_path, "JPEG", quality=85, optimize=True)
+            frame_paths.append(f"file://{frame_path}")
+
+        logger.info(f"RoboReward: Saved {len(frame_paths)} frames as JPEG files in {tmpdir}")
+
+        # Build message with frames as list of file paths (following Qwen3-VL pattern)
+        message = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "video", "video": frame_paths, "sample_fps": 1.0},
+                    {"type": "text", "text": prompt},
+                ],
+            }
+        ]
+
+        # Apply chat template
+        text = self.processor.apply_chat_template(message, tokenize=False, add_generation_prompt=True)
+
+        # Process vision info (qwen-vl-utils handles resizing)
+        image_inputs, video_inputs, video_kwargs = process_vision_info(
+            [message],
+            image_patch_size=16,
+            return_video_kwargs=True,
+            return_video_metadata=True,
+        )
+
+        # Split videos and metadata (video_inputs is list of (video, video_metadata) tuples)
+        if video_inputs is not None:
+            videos, video_metadatas = zip(*video_inputs)
+            videos, video_metadatas = list(videos), list(video_metadatas)
+        else:
+            videos = None
+            video_metadatas = None
+
+        # Process inputs (do_resize=False since qwen-vl-utils already resized)
+        inputs = self.processor(
+            text=[text],
+            images=image_inputs,
+            videos=videos,
+            video_metadata=video_metadatas,
+            padding=True,
+            return_tensors="pt",
+            do_resize=False,  # qwen-vl-utils already resized
+            **video_kwargs,
+        )
+        inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
+
+        # Generate
+        with torch.no_grad():
+            generated_ids = self.model.generate(
+                **inputs,
+                max_new_tokens=self.max_new_tokens,
+                do_sample=False,  # Deterministic
+            )
+
+        # Decode output
+        generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs["input_ids"], generated_ids)]
+        output_texts = self.processor.batch_decode(
+            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )
+
+        # Parse score
+        output_text = output_texts[0]
+        discrete_score = self._parse_score(output_text)
+        logger.info(f"RoboReward: Discrete score: {discrete_score}")
+
+        if discrete_score is None:
+            print(f"[!] Failed to parse score from output: {output_text}")
+            discrete_score = 1  # Default to minimum score if parsing fails
+
+        # Return same discrete score for all frames in this subsequence
+        # Use original num_frames from frames_array (before duplication)
+        original_num_frames = len(convert_frames_to_pil_images(frames_array))
+
+        # because RoboReward returns a score between 1 and 5, we need to normalize it to 0-1
+        result = [float(discrete_score) / 4.0 - 0.25] * original_num_frames
+
+        # Clean up temporary directory
+        shutil.rmtree(tmpdir, ignore_errors=True)
+
+        return result
+
+    def compute_progress_batched(
+        self, frames_list: List[np.ndarray], task_descriptions: List[str]
+    ) -> List[List[Optional[float]]]:
+        """
+        Compute progress predictions for a batch of frame sequences.
+
+        This processes multiple sequences in a single batch for efficiency.
+        Each sequence can have different lengths but will be processed together.
+
+        Args:
+            frames_list: List of (N, H, W, 3) uint8 arrays, one per sample
+            task_descriptions: List of task description strings, one per sample
+
+        Returns:
+            List of progress predictions, one list per sample.
+            Each inner list contains discrete scores (normalized 0-1) for each frame.
+        """
+        if not frames_list:
+            return []
+
+        batch_size = len(frames_list)
+        if len(task_descriptions) != batch_size:
+            raise ValueError(f"Mismatch: {batch_size} frame arrays but {len(task_descriptions)} task descriptions")
+
+        # Prepare all messages and track original frame counts
+        all_messages = []
+        original_frame_counts = []
+        temp_dirs = []
+
+        for idx, (frames_array, task_desc) in enumerate(zip(frames_list, task_descriptions)):
+            if frames_array is None or frames_array.size == 0:
+                all_messages.append(None)
+                original_frame_counts.append(0)
+                temp_dirs.append(None)
+                continue
+
+            # Convert frames to PIL Images
+            frames_pil = convert_frames_to_pil_images(frames_array)
+            if not frames_pil:
+                all_messages.append(None)
+                original_frame_counts.append(0)
+                temp_dirs.append(None)
+                continue
+
+            original_num_frames = len(frames_pil)
+            original_frame_counts.append(original_num_frames)
+
+            # Ensure at least 2 frames for video processing
+            if len(frames_pil) == 1:
+                frames_pil = [frames_pil[0], frames_pil[0]]
+
+            # Build prompt
+            prompt = self._build_prompt(task_desc)
+
+            # Create temporary directory for frame files
+            tmpdir = tempfile.mkdtemp()
+            temp_dirs.append(tmpdir)
+
+            unique_id = uuid.uuid4().hex
+            frame_paths = []
+            for i, frame_pil in enumerate(frames_pil):
+                frame_path = Path(tmpdir) / f"roboreward_{unique_id}_frame_{i:04d}.jpg"
+                frame_pil.save(frame_path, "JPEG", quality=85, optimize=True)
+                frame_paths.append(f"file://{frame_path}")
+
+            # Build message
+            message = [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "video", "video": frame_paths, "sample_fps": 1.0},
+                        {"type": "text", "text": prompt},
+                    ],
+                }
+            ]
+            all_messages.append(message)
+
+        # Filter out None messages for batched processing
+        valid_indices = [i for i, msg in enumerate(all_messages) if msg is not None]
+        valid_messages = [all_messages[i] for i in valid_indices]
+
+        if not valid_messages:
+            # Clean up temp dirs before returning
+            for tmpdir in temp_dirs:
+                if tmpdir is not None:
+                    shutil.rmtree(tmpdir, ignore_errors=True)
+            return [[] for _ in range(batch_size)]
+
+        # Process all valid messages
+        all_texts = []
+        all_image_inputs = []
+        all_video_inputs = []
+        all_video_kwargs_list = []
+
+        for message in valid_messages:
+            text = self.processor.apply_chat_template(message, tokenize=False, add_generation_prompt=True)
+            all_texts.append(text)
+
+            # Process vision info
+            image_inputs, video_inputs, video_kwargs = process_vision_info(
+                [message],
+                image_patch_size=16,
+                return_video_kwargs=True,
+                return_video_metadata=True,
+            )
+            all_image_inputs.append(image_inputs)
+            all_video_inputs.append(video_inputs)
+            all_video_kwargs_list.append(video_kwargs)
+
+        # Process each sample individually (batching videos with different sizes is complex)
+        # but do it efficiently by reusing the prepared data
+        valid_results = []
+        for i, (text, image_inputs, video_inputs, video_kwargs) in enumerate(
+            zip(all_texts, all_image_inputs, all_video_inputs, all_video_kwargs_list)
+        ):
+            # Split videos and metadata
+            if video_inputs is not None:
+                videos, video_metadatas = zip(*video_inputs)
+                videos, video_metadatas = list(videos), list(video_metadatas)
+            else:
+                videos = None
+                video_metadatas = None
+
+            # Process inputs
+            inputs = self.processor(
+                text=[text],
+                images=image_inputs,
+                videos=videos,
+                video_metadata=video_metadatas,
+                padding=True,
+                return_tensors="pt",
+                do_resize=False,
+                **video_kwargs,
+            )
+            inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
+
+            # Generate
+            with torch.no_grad():
+                generated_ids = self.model.generate(
+                    **inputs,
+                    max_new_tokens=self.max_new_tokens,
+                    do_sample=False,
+                )
+
+            # Decode output
+            generated_ids_trimmed = [
+                out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs["input_ids"], generated_ids)
+            ]
+            output_texts = self.processor.batch_decode(
+                generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+            )
+
+            # Parse score
+            output_text = output_texts[0]
+            discrete_score = self._parse_score(output_text)
+            if discrete_score is None:
+                logger.warning(f"Failed to parse score from output: {output_text}")
+                discrete_score = 1
+
+            valid_results.append(discrete_score)
+
+        # Map results back to original indices
+        results = []
+        valid_result_idx = 0
+        for i in range(batch_size):
+            if i in valid_indices:
+                discrete_score = valid_results[valid_result_idx]
+                valid_result_idx += 1
+                # Normalize to 0-1 and repeat for all frames
+                results.append([float(discrete_score) / 4.0 - 0.25] * original_frame_counts[i])
+            else:
+                results.append([])
+
+        # Clean up all temporary directories
+        for tmpdir in temp_dirs:
+            if tmpdir is not None:
+                shutil.rmtree(tmpdir, ignore_errors=True)
+
+        return results
diff --git a/robometer/robometer/evals/baselines/vlac.py b/robometer/robometer/evals/baselines/vlac.py
new file mode 100644
index 0000000000000000000000000000000000000000..bc458e716af24e50cb4ec938738702abddabbc44
--- /dev/null
+++ b/robometer/robometer/evals/baselines/vlac.py
@@ -0,0 +1,336 @@
+#!/usr/bin/env python3
+"""VLAC baseline for progress prediction.
+
+VLAC (Vision-Language-Action-Critic) is a general-purpose pair-wise critic model
+designed for real-world robot reinforcement learning and data refinement.
+It provides robust evaluation capabilities for task progress prediction and
+task completion verification based on images and task descriptions.
+
+Reference: https://github.com/InternRobotics/VLAC
+Model: https://huggingface.co/InternRobotics/VLAC
+
+Downloading the model:
+    # Option 1: Use Hugging Face CLI
+    pip install huggingface_hub
+    huggingface-cli download InternRobotics/VLAC
+
+    # Option 2: Use Python API (automatic in VLAC.__init__)
+    from robometer.evals.baselines.vlac import VLAC
+    model = VLAC(model_path="InternRobotics/VLAC")  # Auto-downloads if not found
+
+    # Option 3: Use download function directly
+    from robometer.evals.baselines.vlac import download_vlac_model
+    model_path = download_vlac_model(local_dir="./models/vlac")
+"""
+
+import os
+import tempfile
+from typing import List, Dict, Optional
+import numpy as np
+import cv2
+
+from robometer.utils.logger import get_logger
+
+# Disable tqdm globally before importing evo_vlac to prevent progress bars
+try:
+    import tqdm
+
+    # Monkey-patch tqdm to always disable before evo_vlac imports it
+    _original_tqdm = tqdm.tqdm
+
+    def _disabled_tqdm(*args, **kwargs):
+        kwargs["disable"] = True
+        return _original_tqdm(*args, **kwargs)
+
+    tqdm.tqdm = _disabled_tqdm
+except ImportError:
+    pass  # tqdm not available, nothing to patch
+
+try:
+    from evo_vlac import GAC_model
+    from evo_vlac.utils.video_tool import compress_video
+
+    VLAC_AVAILABLE = True
+except ImportError:
+    VLAC_AVAILABLE = False
+
+try:
+    from huggingface_hub import snapshot_download
+
+    HF_AVAILABLE = True
+except ImportError:
+    HF_AVAILABLE = False
+
+logger = get_logger()
+
+
+def download_vlac_model(
+    repo_id: str = "InternRobotics/VLAC", cache_dir: Optional[str] = None, local_dir: Optional[str] = None
+) -> str:
+    """
+    Download VLAC model from Hugging Face.
+
+    Args:
+        repo_id: Hugging Face repository ID (default: "InternRobotics/VLAC")
+        cache_dir: Optional cache directory for Hugging Face downloads
+        local_dir: Optional local directory to download model to (if None, uses cache)
+
+    Returns:
+        Path to the downloaded model directory
+
+    Example:
+        >>> model_path = download_vlac_model()
+        >>> model_path = download_vlac_model(local_dir="./models/vlac")
+    """
+    if not HF_AVAILABLE:
+        raise ImportError("huggingface_hub is required to download models. Install with: pip install huggingface_hub")
+
+    print(f"Downloading VLAC model from {repo_id}...")
+    model_path = snapshot_download(
+        repo_id=repo_id,
+        cache_dir=cache_dir,
+        local_dir=local_dir,
+        local_dir_use_symlinks=False,  # Use actual files, not symlinks
+    )
+    print(f"Model downloaded to: {model_path}")
+    return model_path
+
+
+class VLAC:
+    """VLAC baseline for progress prediction using pair-wise comparison.
+
+    VLAC uses a pair-wise comparison mechanism to predict task progress
+    for each frame in a trajectory. It requires a local model checkpoint
+    to be loaded.
+    """
+
+    def __init__(
+        self,
+        model_path: str,
+        device: str = "cuda:0",
+        model_type: str = "internvl2",
+        temperature: float = 0.5,
+        top_k: int = 1,
+        batch_size: int = 5,
+        skip: int = 5,
+        frame_skip: bool = True,
+        auto_download: bool = True,
+        use_images: bool = False,
+    ):
+        """
+        Initialize VLAC model.
+
+        Args:
+            model_path: Path to local VLAC model checkpoint, or Hugging Face repo ID
+                       (e.g., "InternRobotics/VLAC"). If repo ID and model doesn't exist
+                       locally, will download automatically if auto_download=True.
+            device: Device to run model on (e.g., "cuda:0")
+            model_type: Model type (default: "internvl2")
+            temperature: Temperature for generation
+            top_k: Top-k sampling
+            batch_size: Batch size for processing
+            skip: Pair-wise step size
+            frame_skip: Whether to skip frames for efficiency
+            auto_download: If True and model_path is a Hugging Face repo ID, automatically
+                          download the model if not found locally
+            use_images: If True, use image mode (get_trajectory_critic with image files).
+                       If False, use video mode (web_trajectory_critic with video file).
+        """
+        # Check if model_path exists locally first
+        if not os.path.exists(model_path):
+            # If not found locally and looks like a Hugging Face repo ID (contains "/" but not an absolute path)
+            # and auto_download is enabled, download from Hugging Face
+            is_hf_repo_id = (
+                "/" in model_path
+                and not os.path.isabs(model_path)
+                and not model_path.startswith(".")
+                and not model_path.startswith("~")
+            )
+
+            if is_hf_repo_id and auto_download:
+                if not HF_AVAILABLE:
+                    raise ImportError(
+                        "huggingface_hub is required to download models. Install with: pip install huggingface_hub"
+                    )
+                print(f"Model path '{model_path}' not found locally. Downloading from Hugging Face...")
+                model_path = download_vlac_model(repo_id=model_path)
+            elif not is_hf_repo_id:
+                raise FileNotFoundError(
+                    f"Model path '{model_path}' does not exist. "
+                    f"To download from Hugging Face, use model_path='InternRobotics/VLAC'"
+                )
+
+        self.model_path = model_path
+        self.device = device
+        self.model_type = model_type
+        self.temperature = temperature
+        self.top_k = top_k
+        self.batch_size = batch_size
+        self.skip = skip
+        self.frame_skip = frame_skip
+        self.use_images = use_images
+
+        # Initialize model
+        self.critic = GAC_model(tag="critic")
+        self.critic.init_model(model_path=model_path, model_type=model_type, device_map=device)
+        self.critic.temperature = temperature
+        self.critic.top_k = top_k
+        self.critic.set_config()
+        self.critic.set_system_prompt()
+
+    def compute_progress(
+        self, frames_array: np.ndarray, task_description: str = "", reference_video_path: Optional[str] = None
+    ) -> List[Optional[float]]:
+        """
+        Compute progress predictions for frames using VLAC baseline.
+
+        Args:
+            frames_array: (N, H, W, 3) uint8 array from trajectory frames
+            task_description: Task description text
+            reference_video_path: Optional path to reference video for in-context learning
+
+        Returns:
+            List of task progress values in [0, 1] range for each frame.
+            Values are normalized from VLAC output (which may be [0, 100] or [0, 1]).
+            If no predictions, returns list of zeros.
+        """
+        if frames_array is None or frames_array.size == 0:
+            return []
+
+        # Create temporary directory for intermediate files
+        with tempfile.TemporaryDirectory() as tmpdir:
+            if self.use_images:
+                # Image mode: save frames as image files and use get_trajectory_critic
+                image_list = []
+                for i, frame in enumerate(frames_array):
+                    # Ensure frame is uint8
+                    if frame.dtype != np.uint8:
+                        frame = np.clip(frame, 0, 255).astype(np.uint8)
+                    # Save frame as image
+                    image_path = os.path.join(tmpdir, f"frame_{i:05d}.jpg")
+                    # Convert RGB to BGR for OpenCV
+                    frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+                    cv2.imwrite(image_path, frame_bgr)
+                    image_list.append(image_path)
+
+                # Handle reference images if provided
+                ref_image_list = []
+                if reference_video_path:
+                    # Load reference video frames
+                    ref_cap = cv2.VideoCapture(reference_video_path)
+                    ref_idx = 0
+                    while True:
+                        ret, frame = ref_cap.read()
+                        if not ret:
+                            break
+                        # Save reference frame
+                        ref_image_path = os.path.join(tmpdir, f"ref_frame_{ref_idx:05d}.jpg")
+                        cv2.imwrite(ref_image_path, frame)
+                        ref_image_list.append(ref_image_path)
+                        ref_idx += 1
+                    ref_cap.release()
+
+                # Run VLAC trajectory critic with images
+                # Note: get_trajectory_critic returns (critic_list, value_list) for pair-wise comparisons
+                critic_list, value_list = self.critic.get_trajectory_critic(
+                    task=task_description,
+                    image_list=image_list,
+                    ref_image_list=ref_image_list if ref_image_list else None,
+                    batch_num=self.batch_size,
+                    ref_num=len(ref_image_list) if ref_image_list else 0,
+                    rich=True,  # Output decimal values (True for [0,1] range, False for integer percentage)
+                    reverse_eval=False,
+                )
+            else:
+                # Video mode: use web_trajectory_critic with video file
+                video_path = os.path.join(tmpdir, "trajectory.mp4")
+                self._frames_to_video(frames_array, video_path, fps=1.0)
+
+                # Compress video (VLAC expects compressed video)
+                compressed_video = os.path.join(tmpdir, "compressed.mp4")
+                _, output_fps = compress_video(video_path, compressed_video, fps=1.0)
+
+                # Verify compressed video shape matches original frames
+                cap = cv2.VideoCapture(compressed_video)
+                compressed_frames = []
+                while True:
+                    ret, frame = cap.read()
+                    if not ret:
+                        break
+                    # Convert BGR to RGB for consistency
+                    frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+                    compressed_frames.append(frame_rgb)
+                cap.release()
+                compressed_frames_array = np.array(compressed_frames)  # [T, H, W, C] in RGB
+
+                # Check shape match - frames_array is [T, H, W, C] format (checked in _frames_to_video)
+                assert compressed_frames_array.shape[0] == frames_array.shape[0], (
+                    f"Shape mismatch: original {frames_array.shape} vs compressed {compressed_frames_array.shape}"
+                )
+
+                # Run VLAC trajectory critic with video
+                # Note: web_trajectory_critic returns (result_path, value_list, critic_list, done_list)
+                # where value_list contains progress values in [0, 1] if rich=True, or [0, 100] if rich=False
+                result_path, value_list, critic_list, done_list = self.critic.web_trajectory_critic(
+                    task_description=task_description,
+                    main_video_path=compressed_video,
+                    reference_video_path=None,
+                    batch_num=self.batch_size,
+                    ref_num=0,  # Number of reference images from reference video
+                    think=False,  # Whether to use Chain-of-Thought reasoning
+                    skip=1,  # predict per frame
+                    rich=True,  # Output decimal values (True for [0,1] range, False for integer percentage)
+                    reverse_eval=False,  # Whether to reverse evaluation (for VROC evaluation)
+                    output_path=tmpdir,
+                    fps=float(output_fps),
+                    frame_skip=self.frame_skip,  # Whether to skip frames for efficiency (if False, evaluate each frame)
+                    done_flag=False,  # Whether to output done/task completion flag
+                    in_context_done=False,  # Whether to use reference video for done prediction
+                    done_threshold=0.9,  # Threshold for task completion
+                    video_output=False,  # Whether to output annotated video
+                )
+
+            logger.info(f"value_list: {value_list}")
+
+            # Extract progress values from value_list
+            # value_list contains progress predictions for each frame
+            # VLAC may return values in [0, 100] or [0, 1] range, normalize to [0, 1] if needed
+            if value_list and len(value_list) > 0:
+                # Convert to list of floats
+                progress_list = [float(val) for val in value_list]
+
+                # Check if values need normalization from [0, 100] to [0, 1]
+                max_val = max(progress_list) if progress_list else 0.0
+                if max_val > 1.0:
+                    # Normalize from [0, 100] to [0, 1]
+                    progress_list = [val / 100.0 for val in progress_list]
+
+                # Ensure we have predictions for all frames
+                num_frames = frames_array.shape[0]
+                if len(progress_list) < num_frames:
+                    # Pad with last value
+                    progress_list.extend([progress_list[-1]] * (num_frames - len(progress_list)))
+                elif len(progress_list) > num_frames:
+                    # Truncate to match frame count
+                    progress_list = progress_list[:num_frames]
+
+                return np.array(progress_list)
+            else:
+                # Return zeros for all frames if no predictions
+                return np.array([0.0] * frames_array.shape[0])
+
+    def _frames_to_video(self, frames_array: np.ndarray, output_path: str, fps: float = 5.0):
+        """Convert frames array to video file."""
+        height, width = frames_array.shape[1], frames_array.shape[2]
+        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+        out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
+
+        for frame in frames_array:
+            # Ensure frame is uint8
+            if frame.dtype != np.uint8:
+                frame = np.clip(frame, 0, 255).astype(np.uint8)
+            # Convert RGB to BGR for OpenCV
+            frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+            out.write(frame_bgr)
+
+        out.release()
diff --git a/robometer/robometer/evals/compile_results.py b/robometer/robometer/evals/compile_results.py
new file mode 100644
index 0000000000000000000000000000000000000000..7d9c437c41a921b992e680e9806e38951e8e289d
--- /dev/null
+++ b/robometer/robometer/evals/compile_results.py
@@ -0,0 +1,1283 @@
+#!/usr/bin/env python3
+"""
+Script to compile evaluation results from JSON files.
+"""
+
+import json
+from itertools import combinations, product
+from pathlib import Path
+from typing import Any, Dict, List, Tuple, Optional, Union
+from collections import defaultdict
+
+import numpy as np
+import matplotlib.pyplot as plt
+import seaborn as sns
+import cv2
+import torch
+import torch.nn.functional as F
+from sklearn.metrics import average_precision_score
+from robometer.data.datasets.helpers import load_frames_from_npz
+from robometer.evals.eval_metrics_utils import compute_pearson, compute_spearman, compute_kendall
+from robometer.evals.eval_viz_utils import create_combined_progress_success_plot
+from robometer.models.utils import convert_bins_to_continuous, convert_discrete_target_to_continuous
+
+
+def convert_continuous_to_discrete_bin_roboreward(value: float, num_bins: int) -> int:
+    value = min(max(value, 0.0), 1.0)
+    return round(value * (num_bins - 1))
+
+
+def run_quality_preference_eval(results: List[Dict[str, Any]], data_source: Optional[str] = None) -> Dict[str, Any]:
+    """Run quality_preference evaluation analysis.
+
+    Groups results by task and quality labels (or partial_success for RoboArena),
+    computes preference accuracy per group and aggregate.
+    Returns metrics, task_groups, and task_details similar to policy_ranking.
+    """
+    # Check if data_source contains roboreward or roboarena to determine if we should use partial_success logic
+    use_partial_success = "roboreward" in str(data_source).lower() or "roboarena" in str(data_source).lower()
+
+    # First, gather all predictions and labels, convert to arrays
+    # Note: preference_pred is already binary (0/1) from the trainer
+    all_preds = []
+    all_labels = []
+    all_tasks = []
+    all_quality_combos = []
+    valid_indices = []
+
+    for idx, r in enumerate(results):
+        pred = r.get("preference_pred")
+        label = r.get("preference_labels")
+        if pred is not None and label is not None:
+            pred = float(pred.item()) if pred.size == 1 else float(pred[0])
+            label = float(label.item()) if label.size == 1 else float(label[0])
+
+            # For datasets without partial_success, extract quality combo; for datasets with partial_success, just validate metadata exists
+            chosen_meta = r.get("metadata", {}).get("chosen_metadata", {})
+            rejected_meta = r.get("metadata", {}).get("rejected_metadata", {})
+
+            if use_partial_success:
+                # For datasets with partial_success, just check that partial_success exists (we don't use it)
+                chosen_val = chosen_meta.get("partial_success")
+                rejected_val = rejected_meta.get("partial_success")
+                if chosen_val is None or rejected_val is None:
+                    continue
+            else:
+                # For datasets without partial_success, extract quality combo for later use
+                chosen_val = chosen_meta.get("quality_label")
+                rejected_val = rejected_meta.get("quality_label")
+                if chosen_val is None or rejected_val is None:
+                    continue
+                combo_key = tuple(sorted([chosen_val, rejected_val]))
+                all_quality_combos.append(combo_key)
+
+            all_preds.append(pred)
+            all_labels.append(label)
+            all_tasks.append(r["task"])
+            valid_indices.append(idx)
+
+    if not all_preds:
+        return {"error": "No valid predictions found"}, {}, {}
+
+    # Convert to numpy arrays for vectorized operations
+    # preference_pred is already binary (0/1), so no sigmoid conversion needed
+    all_preds = np.array(all_preds)
+    all_labels = np.array(all_labels)
+    binary_preds = all_preds.astype(float)
+
+    # Group results by task (using valid indices to map back)
+    task_groups = defaultdict(list)
+    task_indices = defaultdict(list)
+    for i, (idx, task) in enumerate(zip(valid_indices, all_tasks)):
+        task_groups[task].append(results[idx])
+        task_indices[task].append(i)
+
+    # Compute preference accuracy per task group using vectorized operations
+    task_details = {}
+    all_correct = 0
+    all_total = 0
+
+    for task, task_results in task_groups.items():
+        task_idx = task_indices[task]
+        task_preds = binary_preds[task_idx]
+        task_labels = all_labels[task_idx]
+
+        task_correct = np.sum(task_preds == task_labels)
+        task_total = len(task_preds)
+        pref_acc = task_correct / task_total if task_total > 0 else 0.0
+
+        task_detail = {
+            "preference_accuracy": pref_acc,
+            "num_correct": int(task_correct),
+            "num_total": task_total,
+        }
+
+        # Only compute quality accuracies for datasets without partial_success
+        if not use_partial_success:
+            # Compute accuracy per quality combination using vectorized operations
+            task_quality_combos = [all_quality_combos[i] for i in task_idx]
+            quality_accs = {}
+            unique_combos = set(task_quality_combos)
+            for combo_key in unique_combos:
+                combo_mask = np.array([qc == combo_key for qc in task_quality_combos])
+                combo_preds = task_preds[combo_mask]
+                combo_labels = task_labels[combo_mask]
+                if len(combo_preds) > 0:
+                    combo_correct = np.sum(combo_preds == combo_labels)
+                    combo_acc = combo_correct / len(combo_preds) if len(combo_preds) > 0 else 0.0
+                    quality_accs[f"{combo_key[0]}_vs_{combo_key[1]}"] = combo_acc
+            task_detail["quality_accuracies"] = quality_accs
+        else:
+            task_detail["quality_accuracies"] = None
+
+        task_details[task] = task_detail
+
+        all_correct += task_correct
+        all_total += task_total
+
+    # Aggregate metrics
+    aggregate_acc = all_correct / all_total if all_total > 0 else 0.0
+
+    metrics = {
+        "preference_accuracy": aggregate_acc,
+    }
+
+    return metrics, task_groups, task_details
+
+
+def run_reward_alignment_eval_per_trajectory(
+    results: List[Dict[str, Any]],
+    progress_pred_type: str,
+    is_discrete_mode: bool,
+    num_bins: int,
+    data_source: Optional[str],
+    use_frame_steps: bool,
+    train_success_head: bool,
+    last_frame_only: bool = False,
+) -> Tuple[Dict[str, Any], List, List, List]:
+    """Run reward_alignment evaluation analysis and create plots for each trajectory.
+
+    For failure datasets, we visualize predictions but skip metric computation.
+
+    Args:
+        use_frame_steps: If True, expects multiple results per trajectory (frame_steps mode).
+                         If False, expects exactly one result per trajectory (whole trajectory mode).
+        train_success_head: Whether the success head is being trained (determines if success predictions exist).
+
+    Returns:
+        Tuple of (metrics, plots, video_frames_list, trajectory_progress_data)
+        where trajectory_progress_data is a list of progress_pred values
+        for each trajectory (one per video in video_frames_list)
+    """
+    # Check if data_source contains roboreward or roboarena to determine if we should use partial_success logic
+    use_partial_success = "roboreward" in str(data_source).lower() or "roboarena" in str(data_source).lower()
+
+    # Check if this is RoboReward (needs MAE metric)
+    is_roboreward = data_source and "roboreward" in str(data_source).lower()
+
+    unique_trajectory_ids = set()
+
+    metrics = {}
+
+    # Collect all success_probs and success_labels for AUPRC computation
+    all_success_probs = []
+    all_success_labels = []
+    for r in results:
+        trajectory_id = r.get("id")
+        if trajectory_id:
+            unique_trajectory_ids.add(trajectory_id)
+
+        # Collect success probabilities and labels for AUPRC
+        if train_success_head:
+            success_probs = r["success_probs"]
+            success_labels = r["success_labels"]
+            all_success_probs.append(success_probs)
+            all_success_labels.append(success_labels)
+
+    # Compute success_auprc across all collected success predictions and labels
+    if all_success_probs and all_success_labels:
+        # Flatten all collected probabilities and labels
+        success_probs_flat = np.concatenate(all_success_probs)
+        success_labels_flat = np.concatenate(all_success_labels)
+
+        # Compute AUPRC if we have valid data
+        if success_probs_flat.size > 0 and len(np.unique(success_labels_flat)) > 1:
+            success_auprc = float(average_precision_score(success_labels_flat, success_probs_flat))
+        else:
+            success_auprc = 0.0
+
+        # Compute positive and negative accuracy
+        if success_probs_flat.size > 0:
+            # Convert probabilities to binary predictions (threshold at 0.5)
+            success_preds_flat = (success_probs_flat > 0.5).astype(float)
+
+            # Compute accuracy for positive samples (where label == 1)
+            positive_mask = success_labels_flat == 1
+            num_positives = positive_mask.sum()
+            if num_positives > 0:
+                positive_correct = ((success_preds_flat == success_labels_flat) & positive_mask).sum()
+                positive_success_acc = float(positive_correct / num_positives)
+
+            # Compute accuracy for negative samples (where label == 0)
+            negative_mask = success_labels_flat == 0
+            num_negatives = negative_mask.sum()
+            if num_negatives > 0:
+                negative_correct = ((success_preds_flat == success_labels_flat) & negative_mask).sum()
+                negative_success_acc = float(negative_correct / num_negatives)
+
+        metrics["success_auprc"] = success_auprc
+        metrics["positive_success_acc"] = positive_success_acc
+        metrics["negative_success_acc"] = negative_success_acc
+
+    loss_per_trajectory = np.zeros(1)
+    loss_trajectories = []
+    pearson_trajectories = []
+    plots = []
+    video_frames_list = []
+    trajectory_progress_data = []
+
+    # Collect absolute deltas between final reward and partial_success for trajectories with partial_success
+    partial_success_deltas = []
+
+    # Collect bins for MAE computation (RoboReward)
+    pred_bins_mae = []
+    gt_bins_mae = []
+
+    for trajectory_id in unique_trajectory_ids:
+        results_for_trajectory = [r for r in results if r.get("id") == trajectory_id]
+
+        # Assert that if use_frame_steps=False, each trajectory should have exactly 1 result
+        if not use_frame_steps:
+            assert len(results_for_trajectory) == 1, (
+                f"Expected exactly 1 result per trajectory when use_frame_steps=False, "
+                f"but found {len(results_for_trajectory)} results for trajectory_id={trajectory_id}"
+            )
+
+        # Sort by frame_step if available (for frame_steps mode)
+        # This orders subsequences from shortest to longest (e.g., [0], [0,1], [0,1,2], ...)
+        # Only sort if there are multiple results (indicating frame_steps mode)
+        if len(results_for_trajectory) > 1:
+            results_for_trajectory.sort(key=lambda r: r.get("metadata", {}).get("frame_step", 0))
+
+        # Get task and quality label from first result
+        task = results_for_trajectory[0]["task"]
+        quality_label = results_for_trajectory[0]["quality_label"]
+        video_path = results_for_trajectory[0]["video_path"]
+        partial_success = results_for_trajectory[0].get("partial_success")
+
+        if is_discrete_mode and partial_success is not None:
+            if isinstance(partial_success, torch.Tensor):
+                # [num_bins] -> [1, 1, num_bins]
+                partial_success_tensor = partial_success[
+                    None, None
+                ]  # to make it 3-dim for convert_discrete_target_to_continuous
+            else:
+                # number -> [1, 1]
+                partial_success_tensor = torch.tensor([partial_success], dtype=torch.float32).unsqueeze(0)
+            partial_success = convert_discrete_target_to_continuous(partial_success_tensor, num_bins=num_bins).item()
+
+        # Step 1: Collect all progress predictions and targets
+        raw_preds = []
+        raw_targets = []
+        traj_pred_logits = []  # For discrete mode: collect full logits
+        traj_target_bins = []  # For discrete mode: collect bin indices
+
+        if not use_frame_steps:
+            # Whole trajectory mode: one result with full progress prediction
+            r = results_for_trajectory[0]
+            raw_preds.append(r["progress_pred"])
+            raw_targets.append(r["target_progress"])
+        else:
+            # Frame steps mode: multiple results, one per subsequence
+            for timestep, r in enumerate(results_for_trajectory):
+                raw_preds.append(r["progress_pred"])
+                raw_targets.append(r["target_progress"])
+
+        # Step 2: Convert all progress predictions and targets to continuous (if discrete mode)
+        # Store logits/bins for loss computation before conversion
+        traj_preds_continuous = []
+        traj_targets_continuous = []
+
+        if not use_frame_steps:
+            # Process single prediction/target
+            pred_array = raw_preds[0]
+            if is_discrete_mode:
+                traj_pred_logits = pred_array
+                # Convert to continuous
+                continuous_preds = convert_bins_to_continuous(torch.tensor(pred_array, dtype=torch.float32)).numpy()
+                traj_preds_continuous.append(continuous_preds)
+            else:
+                traj_preds_continuous.append(pred_array)
+
+            tgt_array = raw_targets[0]
+            if is_discrete_mode:
+                traj_target_bins = tgt_array
+                # Convert to continuous
+                continuous_targets = convert_discrete_target_to_continuous(
+                    torch.tensor(tgt_array[None]), num_bins=num_bins
+                )[0].numpy()
+                traj_targets_continuous.append(continuous_targets)
+            else:
+                traj_targets_continuous.append(tgt_array)
+        else:
+            # Frame steps mode: process each timestep
+            for timestep, (pred_array, tgt_array) in enumerate(zip(raw_preds, raw_targets)):
+                # Process prediction
+                if is_discrete_mode:
+                    # Store logits for loss computation (store all logits from this timestep)
+                    traj_pred_logits.append(pred_array)
+                    # Convert to continuous
+                    continuous_preds = convert_bins_to_continuous(torch.tensor(pred_array, dtype=torch.float32)).numpy()
+                    traj_preds_continuous.append(continuous_preds)
+                else:
+                    traj_preds_continuous.append(pred_array)
+
+                # Process target
+                if is_discrete_mode:
+                    # Store target bins for loss computation (store all bins from this timestep)
+                    traj_target_bins.append(tgt_array)
+                    # Convert to continuous
+                    continuous_targets = convert_discrete_target_to_continuous(
+                        torch.tensor(tgt_array[None]), num_bins=num_bins
+                    )[0].numpy()
+                    traj_targets_continuous.append(continuous_targets)
+                else:
+                    traj_targets_continuous.append(tgt_array)
+
+        # Step 3: Apply last_frame_only logic to continuous predictions/targets
+        traj_preds = []
+        traj_targets = []
+
+        if not use_frame_steps:
+            pred_array = traj_preds_continuous[0]
+            traj_preds = pred_array.flatten()
+            if last_frame_only:
+                traj_preds = pred_array[-1:]
+
+            tgt_array = traj_targets_continuous[0]
+            traj_targets = tgt_array.flatten()
+            if last_frame_only:
+                traj_targets = tgt_array[-1:]
+        else:
+            # Frame steps mode
+            for timestep, (pred_array, tgt_array) in enumerate(zip(traj_preds_continuous, traj_targets_continuous)):
+                if last_frame_only:
+                    pred_val = pred_array[-1]
+                    tgt_val = tgt_array[-1]
+                else:
+                    indx = min(timestep, len(pred_array) - 1)
+                    pred_val = pred_array[indx]
+                    indx = min(timestep, len(tgt_array) - 1)
+                    tgt_val = tgt_array[indx]
+                traj_preds.append(pred_val)
+                traj_targets.append(tgt_val)
+
+        # Step 4: Collect success predictions, labels, and probabilities separately
+        traj_success_preds = []
+        traj_success_labels = []
+        traj_success_probs = []
+
+        if not use_frame_steps:
+            # Whole trajectory mode: process single result
+            r = results_for_trajectory[0]
+            if train_success_head:
+                traj_success_preds = r["success_pred"].flatten()
+                traj_success_labels = r["success_labels"].flatten()
+                traj_success_probs = r["success_probs"].flatten()
+        else:
+            # Frame steps mode: process each result
+            for r in results_for_trajectory:
+                if train_success_head:
+                    traj_success_preds.append(r["success_pred"][-1])
+                    traj_success_labels.append(r["success_labels"][-1])
+                    traj_success_probs.append(r["success_probs"][-1])
+
+        # Convert to numpy arrays
+        traj_preds = np.array(traj_preds)
+        traj_targets = np.array(traj_targets)
+        traj_success = np.array(traj_success_preds)
+        traj_success_labels = np.array(traj_success_labels)
+        traj_success_probs = np.array(traj_success_probs)
+
+        # Load video frames if video path exists
+        frames = None
+        if video_path:
+            frames = load_frames_from_npz(video_path)
+            frames = frames.transpose(0, 3, 1, 2)
+
+            # Resize frames to make them smaller for wandb table display
+            resized_frames = []
+            for frame in frames:
+                frame_resized = cv2.resize(frame.transpose(1, 2, 0), (64, 64))
+                resized_frames.append(frame_resized.transpose(2, 0, 1))
+            frames = np.array(resized_frames)
+
+        video_frames_list.append(frames)
+
+        if progress_pred_type == "relative":
+            traj_preds = np.cumsum(traj_preds)
+            traj_targets = np.cumsum(traj_targets)
+
+        trajectory_progress_data.append(traj_preds.tolist())
+
+        # For trajectories with partial_success, compute absolute delta between final reward and partial_success
+        if use_partial_success and partial_success is not None:
+            final_reward = float(traj_preds[-1])
+            delta = abs(final_reward - partial_success)
+            partial_success_deltas.append(delta)
+
+        # For RoboReward, collect bins for MAE computation
+        if is_roboreward and partial_success is not None:
+            # Get last predicted reward (final reward)
+            final_predicted_reward = float(traj_preds[-1])
+
+            # Convert predicted reward to bin (1-5)
+            pred_bin = convert_continuous_to_discrete_bin_roboreward(final_predicted_reward, num_bins=5)
+
+            # Convert partial_success to bin (0->1, 1->5)
+            gt_bin = convert_continuous_to_discrete_bin_roboreward(partial_success, num_bins=5)
+
+            pred_bins_mae.append(pred_bin)
+            gt_bins_mae.append(gt_bin)
+
+        # Only compute metrics for successful trajectories
+        if quality_label == "successful":
+            # Compute loss based on mode
+            if is_discrete_mode and traj_pred_logits is not None and traj_target_bins is not None:
+                # Discrete mode: compute cross-entropy loss between logits and target bins
+                pred_logits_tensor = torch.tensor(
+                    np.array(traj_pred_logits), dtype=torch.float32
+                )  # [seq_len, num_bins]
+                target_bins_tensor = torch.tensor(traj_target_bins)  # [seq_len, num_bins] or [seq_len]
+                if len(target_bins_tensor.shape) == 1:
+                    target_bins_tensor = target_bins_tensor.long()
+                loss_per_timestep = F.cross_entropy(pred_logits_tensor, target_bins_tensor, reduction="none")
+                traj_loss = float(loss_per_timestep.mean().item())
+            else:
+                # Continuous mode: compute MSE loss
+                traj_loss = float(np.mean((traj_targets - traj_preds) ** 2))
+
+            # Compute Pearson correlation
+            traj_pearson = compute_pearson(traj_targets.tolist(), traj_preds.tolist())
+            # Handle NaN values
+            traj_pearson = float(traj_pearson) if not np.isnan(traj_pearson) else 0.0
+        else:
+            traj_loss = 0.0
+            traj_pearson = 0.0
+
+        # Create a wandb plot for progress predictions and, if available, success predictions
+        # Use the shared helper function from eval_viz_utils
+        # Limit to 10 plots to avoid creating too many
+        if len(plots) < 10:
+            has_success_binary = (
+                train_success_head and traj_success is not None and len(traj_success) == len(traj_preds)
+            )
+
+            title = f"Task: {task} - {quality_label}\nLoss: {traj_loss:.3f}, pearson: {traj_pearson:.2f}"
+            if partial_success is not None:
+                title += f", partial_success: {partial_success:.3f}"
+
+            fig = create_combined_progress_success_plot(
+                progress_pred=traj_preds,
+                num_frames=len(traj_preds),
+                success_binary=traj_success if has_success_binary else None,
+                success_probs=traj_success_probs if train_success_head else None,
+                success_labels=traj_success_labels if train_success_head else None,
+                is_discrete_mode=is_discrete_mode,
+                title=title,
+                loss=traj_loss,
+                pearson=traj_pearson,
+            )
+
+            plots.append(fig)
+
+        # Accumulate metrics only for successful trajectories
+        if quality_label == "successful":
+            loss_trajectories.append(traj_loss)
+            if not np.isnan(traj_pearson):
+                pearson_trajectories.append(traj_pearson)
+
+    if len(unique_trajectory_ids) == 0:
+        loss_per_trajectory = np.nan
+        pearson_per_trajectory = np.nan
+    else:
+        loss_per_trajectory = np.mean(loss_trajectories).item()
+        pearson_per_trajectory = np.mean(pearson_trajectories).item()
+
+    metrics["loss"] = loss_per_trajectory
+    metrics["pearson"] = pearson_per_trajectory
+
+    # Add partial_success delta metric if available
+    if use_partial_success and partial_success_deltas:
+        metrics["partial_success_abs_delta"] = float(np.mean(partial_success_deltas))
+
+    # Add RoboReward MAE metric if available
+    if is_roboreward and pred_bins_mae and gt_bins_mae:
+        mae = _compute_mae_between_bins(pred_bins_mae, gt_bins_mae)
+        metrics["mae"] = mae
+
+    return metrics, plots, video_frames_list, trajectory_progress_data
+
+
+def _compute_mae_between_bins(pred_bins: List[int], gt_bins: List[int]) -> float:
+    """Compute Mean Absolute Error (MAE) between predicted bins and ground truth bins.
+
+    MAE(4, 5) = 1, MAE(3, 5) = 2, etc.
+
+    Args:
+        pred_bins: List of predicted bin values
+        gt_bins: List of ground truth bin values
+
+    Returns:
+        Mean Absolute Error (float)
+    """
+    if len(pred_bins) != len(gt_bins):
+        raise ValueError(
+            f"Length mismatch: pred_bins has {len(pred_bins)} elements, gt_bins has {len(gt_bins)} elements"
+        )
+
+    if len(pred_bins) == 0:
+        return 0.0
+
+    # Compute absolute differences
+    abs_diffs = [abs(pred - gt) for pred, gt in zip(pred_bins, gt_bins)]
+
+    # Return mean
+    return float(np.mean(abs_diffs))
+
+
+def _extract_trajectory_rewards(
+    progress_pred: list | np.ndarray,
+    progress_pred_type: str,
+    is_discrete_mode: bool,
+    aggregation: str = "last",
+) -> float:
+    """Extract trajectory reward using different aggregation methods.
+
+    Args:
+        progress_pred: Progress predictions for a single trajectory.
+                       For both continuous and discrete modes: list of floats in [0, 1]
+                       (discrete mode uses convert_bins_to_continuous to get continuous values)
+        progress_pred_type: "relative" or "absolute"
+        is_discrete_mode: Whether predictions came from discrete mode (now converted to continuous)
+        aggregation: "last", "sum", or "average"
+
+    Returns:
+        Aggregated reward (float)
+    """
+    # Both discrete and continuous modes now use continuous values
+    pred_array = np.array(progress_pred, dtype=np.float32)
+
+    # Apply cumsum if relative (for both discrete and continuous modes)
+    if progress_pred_type == "relative":
+        pred_array = np.cumsum(pred_array)
+
+    # Apply aggregation (same logic for both modes)
+    if aggregation == "last":
+        reward = (
+            pred_array[-1]
+            if pred_array.ndim > 0 and len(pred_array) > 0
+            else (pred_array if pred_array.ndim == 0 else 0)
+        )
+    elif aggregation == "sum":
+        reward = np.sum(pred_array)
+    elif aggregation == "average":
+        reward = np.mean(pred_array)
+    else:
+        raise ValueError(f"Unknown aggregation method: {aggregation}")
+
+    # Always return float (discrete mode now uses continuous values)
+    return float(reward)
+
+
+def _compute_policy_ranking_metrics_partial_success(
+    all_rewards: np.ndarray,
+    all_partial_successes: np.ndarray,
+    all_tasks: List[str],
+    correlation_method: str = "kendall",
+) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+    """Compute policy ranking metrics for datasets with partial_success.
+
+    Args:
+        all_rewards: Array of aggregated rewards
+        all_partial_successes: Array of partial_success values (already converted to discrete bins if needed)
+        all_tasks: List of task names
+
+    Returns:
+        Tuple of (metrics dictionary, task_details dictionary)
+    """
+    all_partial_successes = np.array(all_partial_successes)
+
+    # Group by task
+    task_indices = defaultdict(list)
+    for i, task in enumerate(all_tasks):
+        task_indices[task].append(i)
+
+    if not task_indices:
+        return {}, {}
+
+    task_details = {}
+    all_correct_pairs = []
+    all_total_pairs = []
+    all_spearman_rewind = []
+
+    # Compute ranking accuracy for pairs based on partial_success vs predicted rewards
+    for task, task_idx in task_indices.items():
+        if len(task_idx) < 2:
+            continue
+
+        partial_successes = all_partial_successes[task_idx]
+        predicted_rewards = all_rewards[task_idx]
+
+        n = len(partial_successes)
+        if n < 2:
+            continue
+
+        correct_pairs = 0
+        total_pairs = 0
+
+        for i in range(n):
+            for j in range(i + 1, n):
+                # Skip pairs with None values
+                if partial_successes[i] is None or partial_successes[j] is None:
+                    continue
+                if partial_successes[i] == partial_successes[j]:
+                    continue
+
+                total_pairs += 1
+                # True ranking: partial_success[i] should be ranked higher if it's greater
+                true_label = partial_successes[i] > partial_successes[j]
+                # Predicted ranking: predicted_rewards[i] should be higher if it's greater
+                pred_label = predicted_rewards[i] > predicted_rewards[j]
+
+                if true_label == pred_label:
+                    correct_pairs += 1
+
+        # Compute spearman_rewind (binning between 0 and 1)
+        # Filter out None values for binning
+        valid_mask = np.array([ps is not None for ps in partial_successes])
+        if np.any(valid_mask):
+            valid_partial_successes = np.array([ps for ps in partial_successes if ps is not None])
+            valid_predicted_rewards = predicted_rewards[valid_mask]
+            bin_edges = np.linspace(0, 1, 4)  # Creates bins [0, 1/3), [1/3, 2/3), [2/3, 1]
+            bin_assignments = np.clip(np.digitize(valid_partial_successes, bin_edges[1:], right=False), 0, 2)
+        else:
+            bin_assignments = np.array([])
+            valid_predicted_rewards = np.array([])
+
+        avg_rewards_per_bin = {}
+        bin_ranks = []
+        avg_reward_values = []
+
+        for bin_idx in range(3):
+            mask = bin_assignments == bin_idx
+            if np.any(mask):
+                bin_rewards = valid_predicted_rewards[mask]
+                avg_reward = float(np.mean(bin_rewards))
+                avg_rewards_per_bin[bin_idx] = avg_reward
+                bin_ranks.append(bin_idx)
+                avg_reward_values.append(avg_reward)
+
+        correlation_rewind = None
+        if len(bin_ranks) >= 2:
+            if correlation_method == "kendall":
+                correlation_rewind = compute_kendall(bin_ranks, avg_reward_values)
+            else:  # spearman
+                correlation_rewind = compute_spearman(bin_ranks, avg_reward_values)
+            if not np.isnan(correlation_rewind):
+                all_spearman_rewind.append(correlation_rewind)
+
+        if total_pairs > 0:
+            all_correct_pairs.append(correct_pairs)
+            all_total_pairs.append(total_pairs)
+            task_ranking_acc = correct_pairs / total_pairs
+            task_details[task] = {
+                "ranking_acc": float(task_ranking_acc),
+                f"{correlation_method}_rewind": float(correlation_rewind) if correlation_rewind is not None else None,
+            }
+
+    if not all_total_pairs:
+        return {}, {}
+
+    ranking_acc = None
+    if all_total_pairs:
+        total_correct = sum(all_correct_pairs)
+        total_pairs = sum(all_total_pairs)
+        ranking_acc = total_correct / total_pairs if total_pairs > 0 else 0.0
+
+    metrics = {
+        "ranking_acc_rba": ranking_acc,
+        f"{correlation_method}_rewind_rba": np.mean(all_spearman_rewind).item() if all_spearman_rewind else None,
+    }
+
+    return metrics, task_details
+
+
+def _compute_policy_ranking_metrics_quality_label(
+    all_rewards: np.ndarray,
+    all_quality_labels: list[str],
+    all_tasks: list[str],
+    correlation_method: str = "kendall",
+) -> tuple[dict[str, Any], dict[str, Any]]:
+    """Compute policy ranking metrics for datasets using quality_label.
+
+    Args:
+        all_rewards: Array of aggregated rewards
+        all_quality_labels: List of quality labels
+        all_tasks: List of task names
+
+    Returns:
+        Tuple of (metrics dictionary, task_details dictionary)
+    """
+    # Group by task
+    task_indices = defaultdict(list)
+    for i, task in enumerate(all_tasks):
+        task_indices[task].append(i)
+
+    if not task_indices:
+        return {}, {}
+
+    task_details = {}
+    all_correlations = []
+    all_correlations_rewind = []
+    all_succ_subopt_diffs = []
+    all_subopt_fail_diffs = []
+    all_succ_fail_diffs = []
+    all_correct_pairs = []
+    all_total_pairs = []
+
+    # Track global ranking accuracy for all quality pairs
+    global_pair_correct = {}  # (quality1, quality2) -> correct_count
+    global_pair_total = {}  # (quality1, quality2) -> total_count
+
+    # Non-RoboArena: Use quality_label
+    quality_order = {"failure": 1, "suboptimal": 2, "successful": 3}
+    all_labels = ["failure", "suboptimal", "successful"]
+
+    for task, task_idx in task_indices.items():
+        task_rewards = all_rewards[task_idx]
+        task_quality_labels = [all_quality_labels[i] for i in task_idx]
+
+        quality_to_rewards = {q: [] for q in all_labels}
+        for quality, reward in zip(task_quality_labels, task_rewards):
+            quality_to_rewards[quality].append(float(reward))
+
+        present_labels = [q for q in all_labels if quality_to_rewards[q]]
+
+        if len(present_labels) < 2:
+            continue
+
+        k = len(present_labels)
+        correlation_scores = []
+
+        for labels_combo in combinations(present_labels, k):
+            gold_ranks = [quality_order[q] for q in labels_combo]
+            for rewards_tuple in product(*(quality_to_rewards[q] for q in labels_combo)):
+                if correlation_method == "kendall":
+                    corr = compute_kendall(gold_ranks, list(rewards_tuple))
+                else:  # spearman
+                    corr = compute_spearman(gold_ranks, list(rewards_tuple))
+                if not np.isnan(corr):
+                    correlation_scores.append(corr)
+
+        avg_correlation = float(np.mean(correlation_scores)) if correlation_scores else 0.0
+
+        avg_rewards_per_quality = {}
+        quality_ranks = []
+        avg_reward_values = []
+        for q in present_labels:
+            rewards = np.array(quality_to_rewards[q])
+            if len(rewards) > 0:
+                avg_reward = float(np.mean(rewards))
+                avg_rewards_per_quality[q] = avg_reward
+                quality_ranks.append(quality_order[q])
+                avg_reward_values.append(avg_reward)
+
+        correct_pairs = 0
+        total_pairs = 0
+
+        # Compare every pair of trajectories within this task
+        for i in range(len(task_quality_labels)):
+            for j in range(i + 1, len(task_quality_labels)):
+                quality1 = task_quality_labels[i]
+                quality2 = task_quality_labels[j]
+                reward1 = task_rewards[i]
+                reward2 = task_rewards[j]
+
+                # Skip if same quality label
+                if quality1 == quality2:
+                    continue
+
+                expected_order = quality_order[quality1] > quality_order[quality2]
+                actual_order = reward1 > reward2
+                total_pairs += 1
+                if expected_order == actual_order:
+                    correct_pairs += 1
+
+                # Track global pairs by quality label combination
+                pair_key = tuple(sorted([quality1, quality2]))
+                if pair_key not in global_pair_total:
+                    global_pair_total[pair_key] = 0
+                    global_pair_correct[pair_key] = 0
+                global_pair_total[pair_key] += 1
+                if expected_order == actual_order:
+                    global_pair_correct[pair_key] += 1
+
+        if total_pairs > 0:
+            all_correct_pairs.append(correct_pairs)
+            all_total_pairs.append(total_pairs)
+
+        correlation_rewind = None
+        if len(quality_ranks) >= 2:
+            if correlation_method == "kendall":
+                correlation_rewind = compute_kendall(quality_ranks, avg_reward_values)
+            else:  # spearman
+                correlation_rewind = compute_spearman(quality_ranks, avg_reward_values)
+            if not np.isnan(correlation_rewind):
+                all_correlations_rewind.append(correlation_rewind)
+
+        succ_subopt_diff = None
+        subopt_fail_diff = None
+        succ_fail_diff = None
+
+        if "successful" in avg_rewards_per_quality and "suboptimal" in avg_rewards_per_quality:
+            succ_subopt_diff = avg_rewards_per_quality["successful"] - avg_rewards_per_quality["suboptimal"]
+            all_succ_subopt_diffs.append(succ_subopt_diff)
+
+        if "suboptimal" in avg_rewards_per_quality and "failure" in avg_rewards_per_quality:
+            subopt_fail_diff = avg_rewards_per_quality["suboptimal"] - avg_rewards_per_quality["failure"]
+            all_subopt_fail_diffs.append(subopt_fail_diff)
+
+        if "successful" in avg_rewards_per_quality and "failure" in avg_rewards_per_quality:
+            succ_fail_diff = avg_rewards_per_quality["successful"] - avg_rewards_per_quality["failure"]
+            all_succ_fail_diffs.append(succ_fail_diff)
+
+        task_details[task] = {
+            correlation_method: avg_correlation,
+            f"{correlation_method}_rewind": correlation_rewind,
+            "succ_subopt_diff": succ_subopt_diff,
+            "subopt_fail_diff": subopt_fail_diff,
+            "succ_fail_diff": succ_fail_diff,
+        }
+        all_correlations.append(avg_correlation)
+
+    if len(all_correlations) == 0:
+        return {}, {}
+
+    ranking_acc = None
+    if all_total_pairs:
+        total_correct = sum(all_correct_pairs)
+        total_pairs = sum(all_total_pairs)
+        ranking_acc = total_correct / total_pairs if total_pairs > 0 else 0.0
+
+    # Compute ranking accuracy for all pairs by quality label combination
+    ranking_acc_all_pairs = {}
+    for pair_key in global_pair_total:
+        if global_pair_total[pair_key] > 0:
+            pair_acc = global_pair_correct[pair_key] / global_pair_total[pair_key]
+            pair_name = f"ranking_acc_{pair_key[0]}_vs_{pair_key[1]}"
+            ranking_acc_all_pairs[pair_name] = pair_acc
+
+    # Compute overall ranking accuracy across all pairs
+    overall_ranking_acc_all_pairs = None
+    if global_pair_total:
+        total_correct_all = sum(global_pair_correct.values())
+        total_pairs_all = sum(global_pair_total.values())
+        overall_ranking_acc_all_pairs = total_correct_all / total_pairs_all if total_pairs_all > 0 else 0.0
+
+    metrics = {
+        correlation_method: np.mean(all_correlations).item(),
+        f"{correlation_method}_rewind": np.mean(all_correlations_rewind).item() if all_correlations_rewind else None,
+        "avg_succ_subopt_diff": np.mean(all_succ_subopt_diffs).item() if all_succ_subopt_diffs else None,
+        "min_succ_subopt_diff": np.min(all_succ_subopt_diffs).item() if all_succ_subopt_diffs else None,
+        "max_succ_subopt_diff": np.max(all_succ_subopt_diffs).item() if all_succ_subopt_diffs else None,
+        "avg_subopt_fail_diff": np.mean(all_subopt_fail_diffs).item() if all_subopt_fail_diffs else None,
+        "min_subopt_fail_diff": np.min(all_subopt_fail_diffs).item() if all_subopt_fail_diffs else None,
+        "max_subopt_fail_diff": np.max(all_subopt_fail_diffs).item() if all_subopt_fail_diffs else None,
+        "avg_succ_fail_diff": np.mean(all_succ_fail_diffs).item() if all_succ_fail_diffs else None,
+        "min_succ_fail_diff": np.min(all_succ_fail_diffs).item() if all_succ_fail_diffs else None,
+        "max_succ_fail_diff": np.max(all_succ_fail_diffs).item() if all_succ_fail_diffs else None,
+        "ranking_acc": ranking_acc,
+        "ranking_acc_all_pairs": overall_ranking_acc_all_pairs,
+        **ranking_acc_all_pairs,  # Add individual pair accuracies
+    }
+
+    return metrics, task_details
+
+
+def _compute_policy_ranking_metrics_from_rewards(
+    all_rewards: np.ndarray,
+    use_partial_success: bool,
+    all_partial_successes: Optional[np.ndarray],
+    all_quality_labels: Optional[List[str]],
+    all_tasks: List[str],
+    correlation_method: str = "kendall",
+) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+    """Compute policy ranking metrics from pre-computed trajectory rewards.
+
+    Args:
+        all_rewards: Array of aggregated rewards
+        use_partial_success: Whether this dataset uses partial_success
+        all_partial_successes: Array of partial_success values (already converted to discrete bins if needed)
+        all_quality_labels: List of quality labels (if not use_partial_success)
+        all_tasks: List of task names
+
+    Returns:
+        Tuple of (metrics dictionary, task_details dictionary)
+    """
+    if use_partial_success and all_partial_successes is not None:
+        return _compute_policy_ranking_metrics_partial_success(
+            all_rewards, all_partial_successes, all_tasks, correlation_method
+        )
+    else:
+        return _compute_policy_ranking_metrics_quality_label(
+            all_rewards, all_quality_labels, all_tasks, correlation_method
+        )
+
+
+def run_confusion_matrix_eval(
+    results: List[Dict[str, Any]], progress_pred_type: str, is_discrete_mode: bool, num_bins: int
+) -> Dict[str, Any]:
+    """Run confusion_matrix evaluation analysis."""
+    # First, gather all progress predictions, lang_tasks, and video_tasks
+    all_progress_preds = []
+    all_lang_tasks = []
+    all_video_tasks = []
+    valid_indices = []
+
+    for idx, r in enumerate(results):
+        progress_pred = r.get("progress_pred")
+        if progress_pred is not None and len(progress_pred) > 0:
+            meta = r.get("metadata", {})
+            lang_task = meta.get("lang_task")
+            video_task = meta.get("video_task")
+            if lang_task is not None and video_task is not None:
+                all_progress_preds.append(progress_pred)
+                all_lang_tasks.append(lang_task)
+                all_video_tasks.append(video_task)
+                valid_indices.append(idx)
+
+    if not all_progress_preds:
+        return {"error": "No valid confusion matrix data found"}, np.zeros((1, 1))
+
+    # Group results by confusion matrix task
+    uniq_tasks = set(all_lang_tasks) | set(all_video_tasks)
+    task_to_idx = {task: idx for idx, task in enumerate(uniq_tasks)}
+    num_tasks = len(uniq_tasks)
+
+    # Extract final rewards vectorized
+    all_final_rewards = []
+    for progress_pred in all_progress_preds:
+        pred_array = np.array(progress_pred)
+
+        if is_discrete_mode:
+            # Discrete mode: progress_pred is logits [seq_len, num_bins]
+            # Convert to continuous values using weighted sum of bin centers
+            last_frame_logits = pred_array[-1] if pred_array.ndim > 1 else pred_array
+            continuous_pred = convert_bins_to_continuous(torch.tensor(last_frame_logits, dtype=torch.float32)).item()
+            final_reward = float(continuous_pred)
+        else:
+            # Continuous mode: use last frame value
+            if progress_pred_type == "relative":
+                pred_array = np.cumsum(pred_array)
+            final_reward = float(pred_array[-1] if pred_array.ndim > 0 else pred_array)
+
+        all_final_rewards.append(final_reward)
+
+    all_final_rewards = np.array(all_final_rewards)
+    all_lang_indices = np.array([task_to_idx[task] for task in all_lang_tasks])
+    all_video_indices = np.array([task_to_idx[task] for task in all_video_tasks])
+
+    # Build confusion matrix using vectorized operations
+    confusion_matrix = np.zeros((num_tasks, num_tasks))
+    count_matrix = np.zeros((num_tasks, num_tasks))
+
+    # Use advanced indexing to accumulate rewards
+    np.add.at(confusion_matrix, (all_lang_indices, all_video_indices), all_final_rewards)
+    np.add.at(count_matrix, (all_lang_indices, all_video_indices), 1)
+
+    if np.sum(count_matrix) == 0:
+        return {"error": "No valid confusion matrix data found"}, np.zeros((num_tasks, num_tasks))
+
+    # Calculate average rewards (avoid division by zero)
+    with np.errstate(divide="ignore", invalid="ignore"):
+        confusion_matrix = np.divide(
+            confusion_matrix, count_matrix, out=np.zeros_like(confusion_matrix), where=count_matrix != 0
+        )
+
+    # Create the plot
+    fig = plt.figure(figsize=(8, 8))
+
+    # Create heatmap showing average final rewards
+    sns.heatmap(
+        confusion_matrix,
+        # annot=True,
+        # fmt='.3f',
+        cmap="Blues",  # White to dark blue colormap
+        # xticklabels=list(uniq_tasks),
+        # yticklabels=list(uniq_tasks),
+        # cbar_kws={"label": "Average Final Reward (5 trajs)"},
+        cbar=False,  # Remove the color bar
+    )
+    # plt.xlabel("Language Task", fontsize=12)
+    # plt.ylabel("Video Task", fontsize=12)
+    # plt.xticks(rotation=45, ha="right")
+    # plt.yticks(rotation=0)
+    # Remove xticks and yticks
+    plt.xticks([])
+    plt.yticks([])
+
+    plt.tight_layout()
+
+    # Compute trace - off-diagonal metric
+    n = num_tasks
+    trace = np.trace(confusion_matrix)
+    total_sum = np.sum(confusion_matrix)
+    off_diag_sum = total_sum - trace
+    trace_minus_offdiag = trace - off_diag_sum
+
+    # Normalized version (avg diagonal - avg off-diagonal)
+    avg_diagonal = trace / n if n > 0 else 0.0
+    avg_off_diag = off_diag_sum / (n * n - n) if n > 1 else 0.0
+    normalized_metric = avg_diagonal - avg_off_diag
+
+    metrics = {
+        "trace": float(trace),
+        "off_diagonal_sum": float(off_diag_sum),
+        "trace_minus_offdiag": float(trace_minus_offdiag),
+        "avg_diagonal": float(avg_diagonal),
+        "avg_off_diagonal": float(avg_off_diag),
+        "normalized_trace_minus_offdiag": float(normalized_metric),
+    }
+
+    return fig, confusion_matrix, metrics
+
+
+def run_policy_ranking_eval(
+    results: List[Dict[str, Any]],
+    progress_pred_type: str,
+    is_discrete_mode: bool,
+    num_bins: int,
+    data_source: Optional[str] = None,
+    correlation_method: str = "kendall",
+) -> Dict[str, Any]:
+    """Run policy_ranking evaluation analysis.
+
+    Groups results by trajectory_id (like reward_alignment) and computes policy ranking metrics
+    using "last", "average", and "sum" aggregation methods.
+
+    For datasets without partial_success: Uses quality_label and quality_order for ranking.
+    For datasets with partial_success: Uses partial_success for ranking (no quality_order computation).
+    """
+    # Check if data_source contains roboreward or roboarena to determine if we should use partial_success logic
+    use_partial_success = "roboreward" in str(data_source).lower() or "roboarena" in str(data_source).lower()
+
+    # Group results by trajectory_id
+    unique_trajectory_ids = set()
+    for r in results:
+        trajectory_id = r.get("id")
+        if trajectory_id:
+            unique_trajectory_ids.add(trajectory_id)
+
+    if not unique_trajectory_ids:
+        return {"error": "No valid policy ranking data found"}, {}, {}
+
+    # Collect progress predictions per trajectory
+    trajectory_progress_preds = {}  # trajectory_id -> list of progress_pred arrays
+    trajectory_metadata = {}  # trajectory_id -> {task, quality_label/partial_success, video_path}
+
+    for trajectory_id in unique_trajectory_ids:
+        results_for_trajectory = [r for r in results if r.get("id") == trajectory_id]
+
+        # Sort by frame_step if available (for frame_steps mode)
+        # This orders subsequences from shortest to longest (e.g., [0], [0,1], [0,1,2], ...)
+        # Only sort if there are multiple results (indicating frame_steps mode)
+        if len(results_for_trajectory) > 1:
+            results_for_trajectory.sort(key=lambda r: r.get("metadata", {}).get("frame_step", 0))
+
+        # Collect all progress predictions for this trajectory
+        traj_progress_preds = [
+            r.get("progress_pred") for r in results_for_trajectory if r.get("progress_pred") is not None
+        ]
+        trajectory_progress_preds[trajectory_id] = traj_progress_preds
+
+        metadata = {
+            "task": results_for_trajectory[0].get("task"),
+            "video_path": results_for_trajectory[0].get("video_path"),
+            "partial_success": results_for_trajectory[0].get("partial_success"),
+            "quality_label": results_for_trajectory[0].get("quality_label"),
+        }
+        trajectory_metadata[trajectory_id] = metadata
+
+    if not trajectory_progress_preds:
+        return {"error": "No valid policy ranking data found"}, {}, {}
+
+    # Compute rewards for each trajectory using different aggregation methods
+    all_rewards_last = []
+    all_rewards_avg = []
+    all_rewards_sum = []
+    all_tasks = []
+    all_quality_labels = []
+    all_partial_successes = []
+    all_video_paths = []
+    all_ids = []
+
+    for trajectory_id, progress_preds_list in trajectory_progress_preds.items():
+        metadata = trajectory_metadata[trajectory_id]
+
+        # Process progress predictions: convert logits to continuous values if needed
+        processed_progress_preds = []
+        for progress_pred in progress_preds_list:
+            pred_array = np.array(progress_pred)
+
+            if is_discrete_mode:
+                # Discrete mode: pred_array might be logits [seq_len, num_bins]
+                # Convert to continuous values using weighted sum of bin centers
+                if pred_array.ndim > 1:
+                    # It's logits [seq_len, num_bins], convert to continuous values
+                    continuous_preds = convert_bins_to_continuous(torch.tensor(pred_array, dtype=torch.float32)).numpy()
+                    processed_progress_preds.append(continuous_preds.tolist())
+                elif pred_array.ndim == 1:
+                    # Single frame logits [num_bins], convert to continuous
+                    continuous_pred = convert_bins_to_continuous(torch.tensor(pred_array, dtype=torch.float32)).item()
+                    processed_progress_preds.append([float(continuous_pred)])
+                else:
+                    # Scalar (shouldn't happen, but handle it)
+                    processed_progress_preds.append([float(pred_array)])
+            else:
+                # Continuous mode: pred_array is scalar values
+                if pred_array.ndim > 0:
+                    processed_progress_preds.append(pred_array.tolist())
+                else:
+                    processed_progress_preds.append([float(pred_array)])
+
+        if not processed_progress_preds:
+            continue
+
+        # Take the last prediction from each subsequence (e.g., if max_frames=4, take the 4th prediction)
+        # Then use _extract_trajectory_rewards to compute rewards with different aggregation methods
+        last_predictions = []
+        for pred_list in processed_progress_preds:
+            last_predictions.append(pred_list[-1])
+
+        if not last_predictions:
+            continue
+
+        # Use _extract_trajectory_rewards with the list of last predictions from each subsequence
+        reward_last = _extract_trajectory_rewards(
+            last_predictions,
+            progress_pred_type,
+            is_discrete_mode,
+            aggregation="last",
+        )
+        reward_avg = _extract_trajectory_rewards(
+            last_predictions,
+            progress_pred_type,
+            is_discrete_mode,
+            aggregation="average",
+        )
+        reward_sum = _extract_trajectory_rewards(
+            last_predictions,
+            progress_pred_type,
+            is_discrete_mode,
+            aggregation="sum",
+        )
+
+        # Skip trajectories with None partial_success for datasets with partial_success
+        if use_partial_success:
+            if metadata["partial_success"] is None:
+                continue
+            if is_discrete_mode:
+                if isinstance(metadata["partial_success"], torch.Tensor):
+                    # [num_bins] -> [1, 1, num_bins]
+                    partial_success_tensor = metadata["partial_success"][
+                        None, None
+                    ]  # to make it 3-dim for convert_discrete_target_to_continuous
+                else:
+                    # number -> [1, 1]
+                    partial_success_tensor = torch.tensor([metadata["partial_success"]], dtype=torch.float32).unsqueeze(
+                        0
+                    )
+                metadata["partial_success"] = convert_discrete_target_to_continuous(
+                    partial_success_tensor, num_bins=num_bins
+                ).item()
+            all_partial_successes.append(metadata["partial_success"])
+        else:
+            all_quality_labels.append(metadata["quality_label"])
+
+        all_rewards_last.append(reward_last)
+        all_rewards_avg.append(reward_avg)
+        all_rewards_sum.append(reward_sum)
+
+        all_tasks.append(metadata["task"])
+        all_video_paths.append(metadata.get("video_path"))
+        all_ids.append(trajectory_id)
+
+    all_rewards_last = np.array(all_rewards_last)
+    all_rewards_avg = np.array(all_rewards_avg)
+    all_rewards_sum = np.array(all_rewards_sum)
+
+    # Group by task for building task_groups (for return value)
+    task_groups = {}
+    task_indices = defaultdict(list)
+    for i, task in enumerate(all_tasks):
+        if task not in task_groups:
+            task_groups[task] = []
+
+        task_entry = {
+            "final_predicted_reward_last": all_rewards_last[i],
+            "final_predicted_reward_avg": all_rewards_avg[i],
+            "final_predicted_reward_sum": all_rewards_sum[i],
+            "video_path": all_video_paths[i],
+        }
+
+        # Add id if available
+        if all_ids:
+            task_entry["id"] = all_ids[i]
+
+        # Add specific key based on dataset type
+        if use_partial_success:
+            task_entry["partial_success"] = all_partial_successes[i]
+        else:
+            task_entry["quality_label"] = all_quality_labels[i]
+
+        task_groups[task].append(task_entry)
+        task_indices[task].append(i)
+
+    if not task_groups:
+        return {"error": "No valid policy ranking data found"}, {}, {}
+
+    # Compute policy ranking metrics for each aggregation method
+    all_metrics = {}
+    all_task_details = {}
+
+    for agg_method, rewards in [("last", all_rewards_last), ("avg", all_rewards_avg), ("sum", all_rewards_sum)]:
+        metrics, task_details = _compute_policy_ranking_metrics_from_rewards(
+            rewards,
+            use_partial_success,
+            np.array(all_partial_successes) if use_partial_success and all_partial_successes else None,
+            all_quality_labels if not use_partial_success else None,
+            all_tasks,
+            correlation_method,
+        )
+
+        if metrics:
+            # Prefix metrics with aggregation method
+            prefixed_metrics = {f"{k}_{agg_method}" if k != "error" else k: v for k, v in metrics.items()}
+            all_metrics.update(prefixed_metrics)
+
+            # Merge task details (keep first one, or combine if needed)
+            if not all_task_details:
+                all_task_details = task_details
+            else:
+                # Merge task details by adding aggregation suffix to keys
+                for task, details in task_details.items():
+                    if task not in all_task_details:
+                        all_task_details[task] = {}
+                    for k, v in details.items():
+                        all_task_details[task][f"{k}_{agg_method}"] = v
+
+    if not all_metrics:
+        return {"error": "No valid correlations computed"}, {}, {}
+
+    return all_metrics, task_groups, all_task_details
diff --git a/robometer/robometer/evals/eval_metrics_utils.py b/robometer/robometer/evals/eval_metrics_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..ce97e5e319da09218c052ddf459705c8322b59ee
--- /dev/null
+++ b/robometer/robometer/evals/eval_metrics_utils.py
@@ -0,0 +1,122 @@
+from typing import Any
+
+import numpy as np
+from scipy.stats import spearmanr
+from scipy.stats import pearsonr
+from scipy.stats import kendalltau
+import itertools
+
+
+def compute_pearson(y_true: list[float], y_pred: list[float]) -> float:
+    """Compute Pearson correlation, robust to constant inputs; returns np.nan if undefined."""
+
+    a = np.asarray(y_true, dtype=float)
+    b = np.asarray(y_pred, dtype=float)
+    if a.size == 0 or b.size == 0 or a.size != b.size:
+        return np.nan
+    # If either vector is constant, pearsonr returns nan; keep that behavior
+    try:
+        corr, _ = pearsonr(a, b)
+    except Exception:
+        corr = np.nan
+    return corr
+
+
+def compute_spearman(y_true: list[float], y_pred: list[float]) -> float:
+    """Compute Spearman correlation, robust to constant inputs; returns np.nan if undefined."""
+    a = np.asarray(y_true, dtype=float)
+    b = np.asarray(y_pred, dtype=float)
+    if a.size == 0 or b.size == 0 or a.size != b.size:
+        return np.nan
+    try:
+        corr, _ = spearmanr(a, b)
+    except Exception:
+        corr = np.nan
+    return corr
+
+
+def kendall_tau_a(x, y):
+    C = D = 0
+    n = len(x)
+    for i, j in itertools.combinations(range(n), 2):
+        dx = np.sign(x[i] - x[j])
+        dy = np.sign(y[i] - y[j])
+        if dx == 0 or dy == 0:
+            continue  # τ-a ignores ties
+        if dx == dy:
+            C += 1
+        else:
+            D += 1
+    return (C - D) / (n * (n - 1) / 2)
+
+
+def compute_kendall(y_true: list[float], y_pred: list[float]) -> float:
+    """Compute Kendall tau-b correlation, which handles ties better than Spearman.
+
+    Kendall tau-b is specifically adapted to handle ties in the data.
+    Returns np.nan if undefined or if inputs are invalid.
+
+    See: https://stackoverflow.com/questions/10711395/spearman-correlation-and-ties
+    """
+    a = np.asarray(y_true, dtype=float)
+    b = np.asarray(y_pred, dtype=float)
+    if a.size == 0 or b.size == 0 or a.size != b.size:
+        return np.nan
+    # try:
+    #     # variant='b' specifies Kendall tau-b, which handles ties
+    #     corr, _ = kendalltau(a, b, variant='a')
+    # except Exception:
+    #     import ipdb; ipdb.set_trace()
+    #     corr = np.nan
+    corr = kendall_tau_a(a, b)
+    return corr
+
+
+def compute_preference_accuracy(results: list[dict[str, Any]]) -> dict[str, Any]:
+    """Compute preference accuracy over a list of result dicts.
+    Expects keys 'predicted_preference' and 'preference_label' per sample.
+    Returns dict with accuracy, correct, total, and skipped counts.
+    """
+    correct = 0
+    total = 0
+    skipped = 0
+    for r in results:
+        pred = r.get("predicted_preference")
+        label = r.get("preference_label")
+        if pred is None or label is None:
+            skipped += 1
+            continue
+        if pred == label:
+            correct += 1
+        total += 1
+    acc = (correct / total) if total > 0 else 0.0
+    return {
+        "preference_accuracy": acc,
+        "num_correct": correct,
+        "num_total": total,
+        "num_skipped": skipped,
+    }
+
+
+def compute_preference_accuracy_from_progress(results: list[dict[str, Any]]) -> dict[str, Any]:
+    """Compute preference accuracy by using the final progress predictions."""
+    correct = 0
+    total = 0
+    skipped = 0
+    for r in results:
+        progress_chosen = r.get("progress_pred_chosen")[-1]
+        progress_rejected = r.get("progress_pred_rejected")[-1]
+        label = r.get("preference_label")
+        if progress_chosen is None or progress_rejected is None or label is None:
+            skipped += 1
+            continue
+        if progress_chosen > progress_rejected:
+            correct += 1
+        total += 1
+    acc = (correct / total) if total > 0 else None
+    return {
+        "preference_accuracy": acc,
+        "num_correct": correct,
+        "num_total": total,
+        "num_skipped": skipped,
+    }
diff --git a/robometer/robometer/evals/eval_server.py b/robometer/robometer/evals/eval_server.py
new file mode 100755
index 0000000000000000000000000000000000000000..797e1fd25f0b90f087a41d4f7a1da269b558faa6
--- /dev/null
+++ b/robometer/robometer/evals/eval_server.py
@@ -0,0 +1,797 @@
+#!/usr/bin/env python3
+"""
+FastAPI server to evaluate RBM model batches with a multi-GPU service layer.
+
+Usage example:
+    uv run python robometer/evals/eval_server.py \
+        model_path=aliangdw/Robometer-4B \
+        batch_size=16 \
+        num_gpus=1 \
+        server_port=8001
+
+Endpoints:
+  POST /evaluate_batch        - JSON payload
+  POST /evaluate_batch_npy    - multipart payload with .npy blobs
+
+Response payload per request contains predictions grouped by head:
+  {
+    "outputs_preference": {...},   # Preference logits + optional progress traces
+    "outputs_progress": {...},     # Progress-only trajectories
+  }
+"""
+
+from __future__ import annotations
+
+import asyncio
+import copy
+import io
+import json
+import os
+import queue
+import time
+from concurrent.futures import ThreadPoolExecutor
+from dataclasses import asdict
+from threading import Lock
+from typing import Any, Dict, List, Optional, Tuple
+
+import uvicorn
+
+import numpy as np
+import torch
+from omegaconf import DictConfig
+from hydra import main as hydra_main
+from fastapi import FastAPI, Request
+from fastapi.middleware.cors import CORSMiddleware
+
+from robometer.evals.eval_utils import (
+    parse_npy_form_data,
+    reconstruct_payload_from_npy,
+)
+from robometer.utils.save import load_model_from_hf
+from robometer.configs.eval_configs import EvalServerConfig
+from robometer.configs.experiment_configs import ExperimentConfig
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+from robometer.utils.setup_utils import setup_model_and_processor, setup_batch_collator
+from robometer.models.utils import ModelOutput, convert_bins_to_continuous, convert_bins_to_continuous_hard
+from robometer.utils.config_utils import display_config, convert_hydra_to_dataclass
+from robometer.utils.logger import get_logger, setup_loguru_logging
+
+LOG_LEVEL = "DEBUG"
+setup_loguru_logging(log_level=LOG_LEVEL)
+logger = get_logger()
+logger.info(f"robometer.eval_server logger initialized at level {LOG_LEVEL}")
+
+
+def log_logits(name: str, value: Any) -> None:
+    if isinstance(value, torch.Tensor):
+        logger.debug(f"{name} shape={tuple(value.shape)} values={value.detach().cpu().tolist()}")
+    elif isinstance(value, dict):
+        logger.debug(f"{name} keys={list(value.keys())}")
+        for key, sub_value in value.items():
+            log_logits(f"{name}.{key}", sub_value)
+    elif isinstance(value, list):
+        logger.debug(f"{name}: {value}")
+
+
+def aggregate_frame_step_predictions(
+    outputs: Dict[str, Any],
+    sample_frame_counts: List[int],
+    outputs_success: Optional[Dict[str, Any]] = None,
+) -> Dict[str, Any]:
+    """
+    Aggregate frame-step predictions back into full sequences.
+
+    Args:
+        outputs: Dict containing progress predictions from sub-samples
+        sample_frame_counts: List indicating how many frames each original sample had
+        outputs_success: Optional dict containing success predictions
+
+    Returns:
+        Aggregated outputs with full sequences per original sample
+    """
+    progress_pred = outputs.get("progress_pred", [])
+
+    # Aggregate progress predictions
+    aggregated_progress = []
+    current_idx = 0
+
+    for num_frames in sample_frame_counts:
+        if num_frames == 1:
+            # Non-progress sample or single-frame sample, pass through
+            if current_idx < len(progress_pred):
+                aggregated_progress.append(progress_pred[current_idx])
+                current_idx += 1
+            else:
+                aggregated_progress.append([])
+        else:
+            # Collect predictions from sub-samples and extract last prediction from each
+            sample_predictions = []
+            for i in range(num_frames):
+                if current_idx < len(progress_pred):
+                    sub_pred = progress_pred[current_idx]
+                    # Extract the last (and only meaningful) prediction from this sub-sample
+                    if isinstance(sub_pred, list) and len(sub_pred) > 0:
+                        sample_predictions.append(sub_pred[-1])
+                    current_idx += 1
+            aggregated_progress.append(sample_predictions)
+
+    aggregated_outputs = {"progress_pred": aggregated_progress}
+
+    # Aggregate success predictions if present
+    if outputs_success is not None:
+        success_probs = outputs_success.get("success_probs", [])
+        aggregated_success = []
+        current_idx = 0
+
+        for num_frames in sample_frame_counts:
+            if num_frames == 1:
+                if current_idx < len(success_probs):
+                    aggregated_success.append(success_probs[current_idx])
+                    current_idx += 1
+                else:
+                    aggregated_success.append([])
+            else:
+                # Collect success predictions from sub-samples
+                sample_success = []
+                for i in range(num_frames):
+                    if current_idx < len(success_probs):
+                        sub_success = success_probs[current_idx]
+                        # Extract the last prediction from this sub-sample
+                        if isinstance(sub_success, list) and len(sub_success) > 0:
+                            sample_success.append(sub_success[-1])
+                        current_idx += 1
+                aggregated_success.append(sample_success)
+
+        aggregated_outputs["outputs_success"] = {"success_probs": aggregated_success}
+
+    return aggregated_outputs
+
+
+def forward_model(
+    model: Any, batch_inputs: Dict[str, Any], sample_type: str = "progress"
+) -> Tuple[ModelOutput, Dict[str, Any]]:
+    """Forward pass that mirrors trainer logic (handles ReWiND vs RBM)."""
+    with torch.no_grad():
+        if "rewind" in model.__class__.__name__.lower():
+            model_output, extra = model(
+                video_embeddings=batch_inputs.get("video_embeddings"),
+                text_embeddings=batch_inputs.get("text_embeddings"),
+                sample_type=sample_type,
+                timing_raw=None,
+            )
+        else:
+            model_output, extra = model(
+                input_ids=batch_inputs["input_ids"],
+                attention_mask=batch_inputs["attention_mask"],
+                pixel_values=batch_inputs.get("pixel_values", None),
+                pixel_values_videos=batch_inputs.get("pixel_values_videos", None),
+                image_grid_thw=batch_inputs.get("image_grid_thw", None),
+                video_grid_thw=batch_inputs.get("video_grid_thw", None),
+                second_per_grid_ts=batch_inputs.get("second_per_grid_ts", None),
+                sample_type=sample_type,
+                timing_raw=None,
+            )
+
+    return model_output, extra
+
+
+def process_batch_helper(
+    model_type: str,
+    model: Any,
+    tokenizer: Any,
+    batch_collator: Any,
+    device: torch.device,
+    batch_data: List[Dict[str, Any]],
+    job_id: int = 0,
+    is_discrete_mode: bool = False,
+    num_bins: int = 10,
+    use_frame_steps: bool = False,
+) -> Dict[str, Any]:
+    """Synchronous batch processing on specific GPU."""
+    if not batch_data:
+        raise ValueError("No samples found in batch data")
+
+    logger.debug(f"[job {job_id}] Processing {len(batch_data)} samples on device {device}")
+
+    input_samples: List[Any] = []
+    for sample in batch_data:
+        if isinstance(sample, (PreferenceSample, ProgressSample)):
+            input_samples.append(sample)
+        elif isinstance(sample, dict):
+            sample_type = sample.get("sample_type")
+            if sample_type == "preference":
+                input_samples.append(PreferenceSample(**sample))
+            elif sample_type == "progress":
+                input_samples.append(ProgressSample(**sample))
+            else:
+                raise ValueError(f"Unsupported sample_type: {sample_type}")
+        else:
+            raise ValueError(f"Unsupported sample object type: {type(sample)}")
+
+    # Handle frame steps for progress samples - expand into sub-samples, each subsampled to 4 frames
+    # so they can be batched together (all same size)
+    NUM_SUBSAMPLED_FRAMES = 4
+
+    if use_frame_steps:
+        expanded_samples = []
+        sample_frame_counts = []  # Track how many sub-samples each original sample generates
+
+        for sample in input_samples:
+            if isinstance(sample, ProgressSample):
+                # Get the frames from the trajectory
+                frames = sample.trajectory.frames
+                num_frames = frames.shape[0] if hasattr(frames, "shape") else len(frames)
+
+                # Create sub-samples with increasing frame counts: 0:1, 0:2, 0:3, ..., 0:T
+                # Each sub-sample is subsampled to NUM_SUBSAMPLED_FRAMES frames using linspace
+                for i in range(1, num_frames + 1):
+                    # Use linspace to select frame indices from 0 to i-1
+                    # This ensures all sub-samples have the same number of frames for batching
+                    indices = np.linspace(0, i - 1, NUM_SUBSAMPLED_FRAMES, dtype=int)
+                    sub_frames = frames[indices]
+
+                    sub_trajectory = copy.deepcopy(sample.trajectory)
+                    sub_trajectory.frames = sub_frames
+                    sub_trajectory.frames_shape = (
+                        sub_frames.shape if hasattr(sub_frames, "shape") else (len(sub_frames),)
+                    )
+
+                    # Adjust target_progress and success_label if they exist (also subsample)
+                    if hasattr(sub_trajectory, "target_progress") and sub_trajectory.target_progress is not None:
+                        orig_progress = sub_trajectory.target_progress[:i]
+                        if len(orig_progress) > 0:
+                            sub_trajectory.target_progress = (
+                                np.array(orig_progress)[indices].tolist()
+                                if hasattr(orig_progress, "__len__")
+                                else orig_progress
+                            )
+                    if hasattr(sub_trajectory, "success_label") and sub_trajectory.success_label is not None:
+                        orig_success = sub_trajectory.success_label[:i]
+                        if len(orig_success) > 0:
+                            sub_trajectory.success_label = (
+                                np.array(orig_success)[indices].tolist()
+                                if hasattr(orig_success, "__len__")
+                                else orig_success
+                            )
+
+                    # Create sub-sample
+                    sub_sample = ProgressSample(
+                        trajectory=sub_trajectory,
+                        data_gen_strategy=sample.data_gen_strategy,
+                    )
+                    expanded_samples.append(sub_sample)
+
+                sample_frame_counts.append(num_frames)
+            else:
+                # Non-progress samples are passed through unchanged
+                expanded_samples.append(sample)
+                sample_frame_counts.append(1)
+
+        input_samples = expanded_samples
+        logger.debug(
+            f"[job {job_id}] Expanded {len(sample_frame_counts)} samples into {len(input_samples)} sub-samples with frame steps (each subsampled to {NUM_SUBSAMPLED_FRAMES} frames)"
+        )
+    else:
+        sample_frame_counts = None
+
+    batch_inputs = batch_collator(input_samples)
+
+    # Move inputs to the correct GPU
+    for key, value in batch_inputs["preference_inputs"].items():
+        if isinstance(value, torch.Tensor):
+            batch_inputs["preference_inputs"][key] = value.to(device)
+    for key, value in batch_inputs["progress_inputs"].items():
+        if isinstance(value, torch.Tensor):
+            batch_inputs["progress_inputs"][key] = value.to(device)
+    outputs_preference = None
+    outputs_progress = None
+    outputs_success = None
+
+    num_preferences = batch_inputs.get("num_preferences", 0)
+    num_progress = batch_inputs.get("num_progress", 0)
+    logger.debug(f"[job {job_id}] Batch counts — preference: {num_preferences} progress: {num_progress}")
+
+    if num_preferences > 0:
+        outputs_preference = compute_batch_outputs(
+            model,
+            tokenizer,
+            batch_inputs["preference_inputs"],
+            sample_type="preference",
+            is_discrete_mode=is_discrete_mode,
+            num_bins=num_bins,
+        )
+
+    if num_progress > 0:
+        outputs_progress = compute_batch_outputs(
+            model,
+            tokenizer,
+            batch_inputs["progress_inputs"],
+            sample_type="progress",
+            is_discrete_mode=is_discrete_mode,
+            num_bins=num_bins,
+        )
+
+        if "outputs_success" in outputs_progress:
+            outputs_success = outputs_progress.pop("outputs_success")
+
+        # Aggregate frame-step predictions back into full sequences
+        if use_frame_steps and sample_frame_counts is not None:
+            outputs_progress = aggregate_frame_step_predictions(outputs_progress, sample_frame_counts, outputs_success)
+            if outputs_success is not None:
+                outputs_success = outputs_progress.pop("outputs_success", None)
+
+    return {
+        "outputs_preference": outputs_preference,
+        "outputs_progress": outputs_progress,
+        "outputs_success": outputs_success,
+    }
+
+
+class MultiGPUEvalServer:
+    """Multi-GPU inference server that schedules requests across devices."""
+
+    def __init__(
+        self,
+        model_path: str,
+        num_gpus: int | None = None,
+        max_workers: int | None = None,
+    ):
+        self.model_path = model_path
+        self.num_gpus = num_gpus or torch.cuda.device_count()
+        self.max_workers = max_workers or self.num_gpus
+        self._active_jobs = 0
+        self._job_counter = 0
+        self._completed_jobs = 0
+        self._active_jobs_lock = Lock()
+
+        logger.info(f"Loading experiment config and base model from {self.model_path}")
+        exp_config, tokenizer, processor, reward_model = load_model_from_hf(
+            model_path=self.model_path,
+            device=torch.device("cpu"),
+        )
+
+        self.exp_config: ExperimentConfig = exp_config
+        self.base_tokenizer = tokenizer
+        self.base_processor = processor
+        self.base_model = reward_model
+        self.base_batch_collator = setup_batch_collator(processor, tokenizer, self.exp_config, is_eval=True)
+
+        if self.num_gpus == 0:
+            raise RuntimeError("No CUDA devices available")
+
+        logger.info(
+            f"Initializing multi-GPU eval server: model_path={self.model_path} "
+            f"num_gpus={self.num_gpus} max_workers={self.max_workers}"
+        )
+
+        # Initialize GPU pool
+        self.gpu_pool = queue.Queue(maxsize=self.num_gpus)
+        self.executor = ThreadPoolExecutor(max_workers=self.max_workers)
+        self.gpu_stats = {}
+
+        # Initialize GPUs
+        self._initialize_gpus()
+
+        logger.info("Multi-GPU eval server initialized successfully")
+
+    def _initialize_gpus(self):
+        """Initialize models on all GPUs."""
+        for gpu_id in range(self.num_gpus):
+            device = f"cuda:{gpu_id}"
+            logger.info(f"Loading model replica on GPU {gpu_id} ({device})")
+
+            # Load model on specific GPU
+            tokenizer = copy.deepcopy(self.base_tokenizer)
+            processor = copy.deepcopy(self.base_processor)
+            model = copy.deepcopy(self.base_model)
+            batch_collator = copy.deepcopy(self.base_batch_collator)
+
+            model = model.to(device)
+            model.eval()
+
+            # Initialize GPU stats
+            self.gpu_stats[gpu_id] = {
+                "total_requests": 0,
+                "total_processing_time": 0.0,
+                "last_used": time.time(),
+                "status": "ready",
+            }
+
+            # Add to pool
+            self.gpu_pool.put({
+                "model": model,
+                "processor": processor,
+                "tokenizer": tokenizer,
+                "batch_collator": batch_collator,
+                "device": device,
+                "gpu_id": gpu_id,
+                "created_at": time.time(),
+            })
+
+            logger.info(f"Successfully loaded model on GPU {gpu_id}")
+
+    async def process_batch(self, batch_data: List[Dict[str, Any]]):
+        """Process a batch using whichever GPU is available."""
+        loop = asyncio.get_event_loop()
+
+        # Get GPU from pool (this will block until one is available).
+        # Use the default executor so worker threads remain available for compute.
+        gpu_info = await loop.run_in_executor(None, self.gpu_pool.get)
+        queue_size_after_acquire = self.gpu_pool.qsize()
+        with self._active_jobs_lock:
+            self._job_counter += 1
+            job_id = self._job_counter
+            self._active_jobs += 1
+            active_jobs = self._active_jobs
+        logger.debug(
+            f"[job {job_id}] Acquired GPU {gpu_info['gpu_id']} "
+            f"queue_size={queue_size_after_acquire} active_jobs={active_jobs}"
+        )
+
+        start_time = time.time()
+
+        # Update GPU stats
+        self.gpu_stats[gpu_info["gpu_id"]]["status"] = "processing"
+        self.gpu_stats[gpu_info["gpu_id"]]["last_used"] = start_time
+
+        try:
+            # Determine if discrete mode is enabled
+            progress_loss_type = getattr(self.exp_config.loss, "progress_loss_type", "l2")
+            is_discrete_mode = progress_loss_type.lower() == "discrete"
+            num_bins = getattr(
+                self.exp_config.loss,
+                "progress_discrete_bins",
+                getattr(self.exp_config.model, "progress_discrete_bins", 10),
+            )
+
+            # Extract use_frame_steps flag from batch_data (if present)
+            use_frame_steps = False
+            actual_batch_data = batch_data
+            if isinstance(batch_data, dict) and "samples" in batch_data:
+                use_frame_steps = batch_data.get("use_frame_steps", False)
+                actual_batch_data = batch_data["samples"]
+            elif isinstance(batch_data, dict) and "use_frame_steps" in batch_data:
+                # Legacy format: extract use_frame_steps and assume rest is data
+                use_frame_steps = batch_data.get("use_frame_steps", False)
+                actual_batch_data = [v for k, v in batch_data.items() if k != "use_frame_steps"]
+                if len(actual_batch_data) == 1 and isinstance(actual_batch_data[0], list):
+                    actual_batch_data = actual_batch_data[0]
+
+            # Process batch in thread pool
+            result = await loop.run_in_executor(
+                self.executor,
+                process_batch_helper,
+                self.exp_config.model.model_type,
+                gpu_info["model"],
+                gpu_info["tokenizer"],
+                gpu_info["batch_collator"],
+                gpu_info["device"],
+                actual_batch_data,
+                job_id,
+                is_discrete_mode,
+                num_bins,
+                use_frame_steps,
+            )
+
+            # Update stats
+            processing_time = time.time() - start_time
+            self.gpu_stats[gpu_info["gpu_id"]]["total_requests"] += 1
+            self.gpu_stats[gpu_info["gpu_id"]]["total_processing_time"] += processing_time
+
+            return result
+
+        finally:
+            # Always return GPU to pool and update stats
+            processing_time = time.time() - start_time
+            self.gpu_stats[gpu_info["gpu_id"]]["total_requests"] += 1
+            self.gpu_stats[gpu_info["gpu_id"]]["total_processing_time"] += processing_time
+            self.gpu_stats[gpu_info["gpu_id"]]["status"] = "ready"
+            self.gpu_pool.put(gpu_info)
+            queue_size_after_release = self.gpu_pool.qsize()
+            with self._active_jobs_lock:
+                self._active_jobs -= 1
+                self._completed_jobs += 1
+                active_jobs = self._active_jobs
+                completed_jobs = self._completed_jobs
+            logger.debug(
+                f"[job {job_id}] Completed on GPU {gpu_info['gpu_id']} "
+                f"active_jobs={active_jobs} completed_jobs={completed_jobs} "
+                f"queue_size={queue_size_after_release} "
+                f"processing_time={processing_time:.3f}s"
+            )
+
+    def get_pool_status(self) -> Dict[str, Any]:
+        """Get status of the GPU pool."""
+        return {
+            "total_gpus": self.num_gpus,
+            "available_gpus": self.gpu_pool.qsize(),
+            "max_workers": self.max_workers,
+            "gpu_stats": self.gpu_stats,
+            "pool_size": self.gpu_pool.maxsize,
+        }
+
+    def shutdown(self):
+        """Shutdown the GPU pool and executor."""
+        logger.info("Shutting down GPU pool...")
+        self.executor.shutdown(wait=True)
+        logger.info("GPU pool shutdown complete")
+
+
+def compute_batch_outputs(
+    model: Any,
+    tokenizer: Any,
+    batch_inputs: Dict[str, torch.Tensor],
+    sample_type: str,
+    is_discrete_mode: bool = False,
+    num_bins: int = 10,
+) -> Dict[str, Any]:
+    """
+    Run a forward pass and return the raw head outputs we need for eval logging.
+
+    Args:
+        model: RBM/ReWiND model on the target device.
+        tokenizer: Tokenizer (unused for head-based inference).
+        batch_inputs: Collated inputs for the requested head.
+        sample_type: One of {"preference","progress"}.
+
+    Returns:
+        Dict containing logits/derived predictions keyed by head.
+    """
+    model.eval()
+    logger.debug(f"compute_batch_outputs sample_type={sample_type}")
+    model_output, _ = forward_model(model, batch_inputs, sample_type=sample_type)
+
+    results: Dict[str, Any] = {}
+
+    # Preference logits and metadata
+    if sample_type == "preference" and model_output.pref_logits is not None:
+        logits = model_output.pref_logits.squeeze(-1)
+        probs = torch.sigmoid(logits)
+        preds = (probs > 0.5).long()
+        results.update({
+            "predictions": preds.detach().cpu().tolist(),
+            "prediction_probs": probs.detach().cpu().tolist(),
+            "preference_labels": batch_inputs["preference_labels"].cpu().tolist(),
+        })
+
+        # logger.debug(f"predictions: {results['predictions']}")
+        # logger.debug(f"prediction_probs: {results['prediction_probs']}")
+        # logger.debug(f"preference_labels: {results['preference_labels']}")
+
+    # Progress predictions (only for progress sample type)
+    progress_logits = model_output.progress_logits
+    if progress_logits is not None and isinstance(progress_logits, dict) and sample_type == "progress":
+        progress_pred = []
+        seq_A = progress_logits.get("A")
+
+        # Convert tensor to list
+        seq_A_list = [seq_A[i] for i in range(seq_A.shape[0])] if seq_A is not None else []
+
+        # Process seq_A
+        for seq_A_item in seq_A_list:
+            if seq_A_item is None:
+                progress_pred.append([])
+            elif is_discrete_mode:
+                # seq_A_item is [seq_len, num_bins] logits, convert entire sequence to continuous
+                continuous_pred = convert_bins_to_continuous(seq_A_item.detach().cpu().float())
+                # continuous_pred = convert_bins_to_continuous_hard(seq_A_item.detach().cpu().float())
+                progress_pred.append(continuous_pred.numpy().flatten().tolist())
+            else:
+                progress_pred.append(seq_A_item.detach().cpu().flatten().tolist())
+
+        if not progress_pred:
+            batch_size = len(batch_inputs.get("task", []))
+            progress_pred = [[] for _ in range(batch_size)]
+
+        results["progress_pred"] = progress_pred
+        # logger.debug(f"progress_pred: {progress_pred}")
+        if model_output.success_logits is not None:
+            success_pred = model_output.success_logits["A"]
+            success_probs = torch.sigmoid(success_pred)
+            results["outputs_success"] = {
+                "success_probs": success_probs.detach().cpu().tolist(),
+            }
+            # logger.debug(f"success_probs: {success_probs}")
+
+    return results
+
+
+def create_app(cfg: EvalServerConfig, multi_gpu_server: MultiGPUEvalServer | None = None):
+    app = FastAPI(title="RBM Multi-GPU Evaluation Server")
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+    # Initialize multi-GPU server
+    num_gpus = getattr(cfg, "num_gpus", None)
+    max_workers = getattr(cfg, "max_workers", None)
+
+    multi_gpu_server = multi_gpu_server or MultiGPUEvalServer(cfg.model_path, num_gpus, max_workers)
+    logger.info(f"Multi-GPU eval server initialized with {multi_gpu_server.num_gpus} GPUs")
+
+    @app.post("/evaluate_batch")
+    async def evaluate_batch(batch: Dict[str, Any]) -> Dict[str, Any]:
+        """Evaluate a batch of preference samples using the Multi-GPU server."""
+        logger.debug(f"Received /evaluate_batch request with keys: {list(batch.keys())}")
+        return await multi_gpu_server.process_batch(batch)
+
+    @app.post("/evaluate_batch_npy")
+    async def evaluate_batch_npy(request: Request) -> Dict[str, Any]:
+        """Evaluate a batch with .npy file support for numpy arrays.
+
+        This endpoint handles multipart form data where:
+        - numpy arrays are sent as .npy files
+        - other data is sent as form fields
+        """
+        # Parse form data
+        form_data = await request.form()
+
+        # Extract numpy arrays and other data using shared utility (await async function)
+        numpy_arrays, other_data = await parse_npy_form_data(form_data)
+
+        # Extract use_frame_steps flag from other_data (handle both bool and string)
+        use_frame_steps_value = other_data.pop("use_frame_steps", False)
+        if isinstance(use_frame_steps_value, bool):
+            use_frame_steps = use_frame_steps_value
+        else:
+            use_frame_steps = str(use_frame_steps_value).lower() == "true"
+
+        # Reconstruct the original payload structure (RBM needs torch tensor conversion for embeddings)
+        batch_data = reconstruct_payload_from_npy(
+            numpy_arrays,
+            other_data,
+            trajectory_keys=[
+                "chosen_trajectory",
+                "rejected_trajectory",
+                "reference_trajectory",
+                "traj_sim_trajectory",
+                "traj_diff_trajectory",
+                "trajectory",
+            ],
+            convert_embeddings_to_torch=True,
+        )
+
+        # Add use_frame_steps flag to batch_data
+        batch_payload = {
+            "samples": batch_data,
+            "use_frame_steps": use_frame_steps,
+        }
+
+        # Process the batch
+        logger.debug(
+            f"Received /evaluate_batch_npy request with {len(numpy_arrays)} numpy arrays "
+            f"and {len(other_data)} other fields, use_frame_steps={use_frame_steps}"
+        )
+        return await multi_gpu_server.process_batch(batch_payload)
+
+    @app.get("/gpu_status")
+    def get_gpu_status() -> Dict[str, Any]:
+        """Get status of all GPUs and pool."""
+        return multi_gpu_server.get_pool_status()
+
+    @app.get("/health")
+    def health_check() -> Dict[str, Any]:
+        """Health check endpoint."""
+        status = multi_gpu_server.get_pool_status()
+        return {"status": "healthy", "available_gpus": status["available_gpus"], "total_gpus": status["total_gpus"]}
+
+    @app.get("/model_info")
+    def get_model_info() -> Dict[str, Any]:
+        """Get model information and experiment configuration."""
+
+        def serialize_config(config_obj: Any) -> Any:
+            """Recursively serialize dataclass to dict, handling nested dataclasses."""
+            if hasattr(config_obj, "__dataclass_fields__"):
+                result = {}
+                for field_name, field_value in asdict(config_obj).items():
+                    result[field_name] = serialize_config(field_value)
+                return result
+            elif isinstance(config_obj, dict):
+                return {k: serialize_config(v) for k, v in config_obj.items()}
+            elif isinstance(config_obj, list):
+                return [serialize_config(item) for item in config_obj]
+            elif isinstance(config_obj, (str, int, float, bool, type(None))):
+                return config_obj
+            else:
+                # Fallback: convert to string for non-serializable types
+                return str(config_obj)
+
+        def get_model_architecture_info(model: Any) -> Dict[str, Any]:
+            """Extract model architecture information."""
+            model_info = {
+                "model_class": model.__class__.__name__,
+                "model_module": model.__class__.__module__,
+            }
+
+            # Count parameters
+            try:
+                all_params = sum(p.numel() for p in model.parameters())
+                trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+                frozen_params = all_params - trainable_params
+
+                model_info.update({
+                    "total_parameters": all_params,
+                    "trainable_parameters": trainable_params,
+                    "frozen_parameters": frozen_params,
+                    "trainable_percentage": (100 * trainable_params / all_params) if all_params > 0 else 0.0,
+                })
+            except Exception as e:
+                logger.warning(f"Could not count model parameters: {e}")
+                model_info["parameter_count_error"] = str(e)
+
+            # Get model architecture summary (first few layers)
+            try:
+                architecture_summary = []
+                for name, module in list(model.named_children())[:10]:  # First 10 top-level modules
+                    module_type = module.__class__.__name__
+                    num_params = sum(p.numel() for p in module.parameters())
+                    architecture_summary.append({
+                        "name": name,
+                        "type": module_type,
+                        "parameters": num_params,
+                    })
+                model_info["architecture_summary"] = architecture_summary
+            except Exception as e:
+                logger.warning(f"Could not get architecture summary: {e}")
+                model_info["architecture_summary_error"] = str(e)
+
+            return model_info
+
+        exp_config_dict = serialize_config(multi_gpu_server.exp_config)
+
+        # Get model architecture info from the base model
+        model_arch_info = None
+        try:
+            model_arch_info = get_model_architecture_info(multi_gpu_server.base_model)
+        except Exception as e:
+            logger.warning(f"Could not get model architecture info: {e}")
+            model_arch_info = {"error": str(e)}
+
+        return {
+            "model_path": multi_gpu_server.model_path,
+            "num_gpus": multi_gpu_server.num_gpus,
+            "experiment_config": exp_config_dict,
+            "model_architecture": model_arch_info,
+        }
+
+    @app.on_event("shutdown")
+    async def shutdown_event():
+        """Cleanup on shutdown."""
+        multi_gpu_server.shutdown()
+
+    return app
+
+
+@hydra_main(version_base=None, config_path="../configs", config_name="eval_config_server")
+def main(cfg: DictConfig):
+    """Main entry point for evaluation server using Hydra configuration."""
+    # Convert Hydra config to dataclass
+    eval_cfg = convert_hydra_to_dataclass(cfg, EvalServerConfig)
+
+    # Display the configuration in a nice Rich format
+    display_config(eval_cfg)
+
+    # Ensure pretrained checkpoint is specified
+    if not eval_cfg.model_path:
+        raise ValueError("Eval config must set model_path to a pretrained checkpoint.")
+
+    multi_gpu_server = MultiGPUEvalServer(
+        model_path=eval_cfg.model_path,
+        num_gpus=eval_cfg.num_gpus,
+        max_workers=eval_cfg.max_workers,
+    )
+    display_config(multi_gpu_server.exp_config)
+
+    app = create_app(eval_cfg, multi_gpu_server)
+    print(f"Running multi-GPU eval server on {eval_cfg.server_url}:{eval_cfg.server_port}")
+    print(f"Using {eval_cfg.num_gpus or torch.cuda.device_count()} GPUs")
+    uvicorn.run(app, host=eval_cfg.server_url, port=eval_cfg.server_port)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/robometer/evals/eval_utils.py b/robometer/robometer/evals/eval_utils.py
new file mode 100755
index 0000000000000000000000000000000000000000..8bd5f5398c9f6ea29c1e9ac1012044abf0fa93fc
--- /dev/null
+++ b/robometer/robometer/evals/eval_utils.py
@@ -0,0 +1,455 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import re
+import torch
+import io
+import json
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Union, Optional, Tuple
+from datetime import datetime
+
+import aiohttp
+import numpy as np
+import requests
+import torch
+
+from robometer.data.dataset_types import PreferenceSample, ProgressSample, Trajectory
+from robometer.data.datasets.helpers import linspace_subsample_frames, pad_trajectory_to_max_frames_np
+
+
+def extract_rewards_from_output(outputs: Dict[str, Any]) -> np.ndarray:
+    """
+    Extract rewards from the output dictionary returned by process_batch_helper.
+
+    Args:
+        outputs: Dictionary with 'outputs_preference', 'outputs_progress', 'outputs_similarity'
+                 Should contain 'outputs_progress' with 'progress_pred' key
+
+    Returns:
+        numpy array of rewards (one per sample)
+    """
+    if outputs.get("outputs_progress") is None:
+        raise ValueError("No progress outputs found in batch outputs")
+
+    outputs_progress = outputs["outputs_progress"]
+    progress_pred = outputs_progress.get("progress_pred", [])
+
+    # Extract rewards (last value of each progress prediction)
+    # Each progress_list contains progress values for each frame in the subsequence
+    rewards_list = []
+    for progress_list in progress_pred:
+        try:
+            if isinstance(progress_list, list) and len(progress_list) > 0:
+                # The last value is the reward for this subsequence
+                reward = float(progress_list[-1])
+                # Clamp to [0, 1] range
+                reward = max(0.0, min(1.0, reward))
+            else:
+                # Default to 0.0 if no valid prediction
+                reward = 0.0
+        except (ValueError, TypeError, IndexError) as e:
+            reward = 0.0
+        rewards_list.append(reward)
+
+    return np.array(rewards_list, dtype=np.float32)
+
+
+def extract_success_probs_from_output(outputs: Dict[str, Any]) -> np.ndarray:
+    """
+    Extract success probabilities from the output dictionary returned by process_batch_helper.
+
+    Args:
+        outputs: Dictionary with 'outputs_success'
+                 Should contain 'outputs_success' with 'success_probs' key
+
+    Returns:
+        numpy array of success probabilities (one per sample)
+    """
+    if outputs.get("outputs_success") is None:
+        raise ValueError("No success probabilities outputs found in batch outputs")
+
+    outputs_success = outputs["outputs_success"]
+    success_preds = outputs_success.get("success_probs", [])
+    success_probs = []
+    for success_pred in success_preds:
+        try:
+            if isinstance(success_pred, list) and len(success_pred) > 0:
+                # The last value is the success probability for this subsequence
+                success_prob = float(success_pred[-1])
+                success_probs.append(success_prob)
+            else:
+                success_probs.append(0.0)
+        except (ValueError, TypeError, IndexError) as e:
+            success_probs.append(0.0)
+
+    return np.array(success_probs, dtype=np.float32)
+
+
+def raw_dict_to_sample(
+    raw_data: Union[Tuple[Dict[str, Any], Dict[str, Any]], Dict[str, Any]],
+    max_frames: int = 16,
+    sample_type: str = "progress",
+) -> Union[ProgressSample, PreferenceSample]:
+    """
+    Convert raw data dictionary to a ProgressSample or PreferenceSample.
+
+    Args:
+        raw_data: Dict with 'frames', 'task', 'id', 'metadata', 'video_embeddings', 'text_embedding' or Tuple of (Dict[str, Any], Dict[str, Any])
+        max_frames: Maximum number of frames to use (default: 16)
+        sample_type: Either "progress" or "preference" (default: "progress")
+
+    Returns:
+        ProgressSample or PreferenceSample
+    """
+
+    def _build_trajectory(raw_data: Dict[str, Any], num_frames: int) -> Trajectory:
+        processed_item: Dict[str, Any] = {}
+
+        # Process frames
+        frames_array = raw_data["frames"]
+
+        # Ensure we have the correct shape: (T, H, W, C)
+        if len(frames_array.shape) != 4:
+            raise ValueError(f"Expected 4D array (T, H, W, C), got shape {frames_array.shape}")
+
+        # Convert from CxHxW to HxWxC if needed
+        if frames_array.shape[1] == 3:
+            frames_array = np.transpose(frames_array, (0, 2, 3, 1))
+
+        frames_array, _ = linspace_subsample_frames(frames_array, num_frames)
+        dummy_progress = [0.0] * len(frames_array)
+        frames_array, _ = pad_trajectory_to_max_frames_np(frames_array, dummy_progress, num_frames, pad_from="right")
+
+        if frames_array.size == 0:
+            raise ValueError("No frames processed for example")
+
+        processed_item["frames"] = frames_array
+        processed_item["frames_shape"] = frames_array.shape
+        processed_item["task"] = raw_data["task"]
+        processed_item["lang_vector"] = None
+        processed_item["metadata"] = raw_data.get("metadata", None)
+
+        # Process video embeddings using same helper functions
+        video_embeddings = raw_data.get("video_embeddings")
+        if video_embeddings is not None:
+            video_embeddings, _ = linspace_subsample_frames(video_embeddings, num_frames)
+            dummy_progress_emb = [0.0] * len(video_embeddings)
+            video_embeddings, _ = pad_trajectory_to_max_frames_np(
+                video_embeddings, dummy_progress_emb, num_frames, pad_from="right"
+            )
+
+        text_embedding = raw_data.get("text_embedding")
+
+        # Convert to tensors if they are numpy arrays
+        if video_embeddings is not None and isinstance(video_embeddings, np.ndarray):
+            video_embeddings = torch.tensor(video_embeddings)
+        if text_embedding is not None and isinstance(text_embedding, np.ndarray):
+            text_embedding = torch.tensor(text_embedding)
+
+        processed_item["video_embeddings"] = video_embeddings
+        processed_item["text_embedding"] = text_embedding
+        processed_item["video_shape"] = video_embeddings.shape if video_embeddings is not None else None
+        processed_item["text_shape"] = text_embedding.shape if text_embedding is not None else None
+
+        trajectory = Trajectory(**processed_item)
+        return trajectory
+
+    if sample_type == "progress":
+        assert isinstance(raw_data, dict), "raw_data must be a dictionary"
+        trajectory = _build_trajectory(raw_data=raw_data, num_frames=max_frames)
+        return ProgressSample(trajectory=trajectory)
+    elif sample_type == "preference":
+        assert isinstance(raw_data, tuple), "raw_data must be a tuple"
+        assert len(raw_data) == 2, "raw_data must be a tuple of two dictionaries"
+        trajectories: List[Trajectory] = []
+        for trajectory_data in raw_data:
+            trajectory = _build_trajectory(raw_data=trajectory_data, num_frames=max_frames)
+            trajectories.append(trajectory)
+        return PreferenceSample(chosen_trajectory=trajectories[0], rejected_trajectory=trajectories[1])
+    else:
+        raise ValueError(f"Unsupported sample_type: {sample_type}")
+
+
+def build_payload(
+    samples: list[PreferenceSample | ProgressSample],
+) -> tuple[dict[str, Any], list[dict[str, Any]]]:
+    """Build a payload with numpy array handling.
+
+    Args:
+        samples: List of samples to convert
+
+    Returns:
+        Tuple of (files, sample_data) where:
+        - files: Dict of numpy arrays converted to .npy format
+        - sample_data: List of sample dictionaries with numpy arrays replaced by file references
+    """
+    files = {}
+    sample_data = []
+
+    for sample_idx, sample in enumerate(samples):
+        # Copy the original sample and handle numpy arrays
+        processed_sample = sample.model_dump().copy()
+
+        # Handle trajectory objects with numpy arrays
+        for key in [
+            "chosen_trajectory",
+            "rejected_trajectory",
+            "trajectory",
+        ]:
+            if key in processed_sample and isinstance(processed_sample[key], dict):
+                trajectory = processed_sample[key]
+
+                # Convert numpy arrays to .npy files
+                numpy_fields = ["frames", "lang_vector", "video_embeddings", "text_embedding"]
+                for field_name in numpy_fields:
+                    # if it is a tensor, first convert it to a numpy array
+                    if field_name in trajectory and isinstance(trajectory[field_name], torch.Tensor):
+                        trajectory[field_name] = trajectory[field_name].numpy()
+
+                    if field_name in trajectory and isinstance(trajectory[field_name], np.ndarray):
+                        # Convert numpy array to .npy file
+                        buf = io.BytesIO()
+                        np.save(buf, trajectory[field_name])
+                        buf.seek(0)
+                        file_key = f"sample_{sample_idx}_{key}_{field_name}"
+                        files[file_key] = (
+                            f"sample_{sample_idx}_{key}_{field_name}.npy",
+                            buf,
+                            "application/octet-stream",
+                        )
+                        trajectory[field_name] = {"__numpy_file__": file_key}
+
+        sample_data.append(processed_sample)
+
+    return files, sample_data
+
+
+def post_batch(url: str, payload: dict[str, Any], timeout_s: float = 120.0) -> dict[str, Any]:
+    """POST a batch payload to the evaluation server and return parsed JSON."""
+    resp = requests.post(url.rstrip("/") + "/evaluate_batch", json=payload, timeout=timeout_s)
+    resp.raise_for_status()
+    return resp.json()
+
+
+def post_batch_npy(
+    url: str, files: dict[str, Any], sample_data: list[dict[str, Any]], timeout_s: float = 120.0
+) -> dict[str, Any]:
+    """POST batch using .npy format for numpy arrays."""
+    # Convert sample_data to form data
+    data = {f"sample_{i}": json.dumps(sample) for i, sample in enumerate(sample_data)}
+
+    # Send as multipart form data
+    resp = requests.post(url.rstrip("/") + "/evaluate_batch_npy", files=files, data=data, timeout=timeout_s)
+    resp.raise_for_status()
+    return resp.json()
+
+
+async def post_batch_npy_async(
+    session: aiohttp.ClientSession,
+    url: str,
+    files: dict[str, Any],
+    sample_data: list[dict[str, Any]],
+    timeout_s: float = 120.0,
+) -> dict[str, Any]:
+    """Async version of post_batch_npy using aiohttp."""
+    # Create FormData for aiohttp
+    form_data = aiohttp.FormData()
+
+    # Add files
+    for key, (filename, file_obj, content_type) in files.items():
+        form_data.add_field(key, file_obj, filename=filename, content_type=content_type)
+
+    # Add sample data
+    for i, sample in enumerate(sample_data):
+        form_data.add_field(f"sample_{i}", json.dumps(sample))
+
+    headers = {"Connection": "close"}
+    # Send as multipart form data using aiohttp
+    timeout = aiohttp.ClientTimeout(total=timeout_s)
+    async with session.post(
+        url.rstrip("/") + "/evaluate_batch_npy", data=form_data, timeout=timeout, headers=headers
+    ) as resp:
+        resp.raise_for_status()
+        return await resp.json()
+
+
+async def parse_npy_form_data(form_data: Any) -> Tuple[Dict[str, np.ndarray], Dict[str, Any]]:
+    """Parse multipart form data to extract numpy arrays and other data.
+
+    Args:
+        form_data: FastAPI form data from request.form()
+
+    Returns:
+        Tuple of (numpy_arrays dict, other_data dict)
+    """
+    numpy_arrays = {}
+    other_data = {}
+
+    for key, value in form_data.items():
+        # Check if this is a file upload (UploadFile object)
+        if hasattr(value, "filename") and value.filename:
+            # This is a file upload
+            if value.filename.endswith(".npy"):
+                # Load .npy file (await async read)
+                content = await value.read()
+                buf = io.BytesIO(content)
+                array = np.load(buf)
+                numpy_arrays[key] = array
+            else:
+                # Non-.npy file, skip for now
+                continue
+        else:
+            # This is a string value (form field)
+            try:
+                # Try to parse as JSON
+                other_data[key] = json.loads(value)
+            except (json.JSONDecodeError, TypeError):
+                # Keep as string if not JSON
+                other_data[key] = value
+
+    return numpy_arrays, other_data
+
+
+def reconstruct_payload_from_npy(
+    numpy_arrays: Dict[str, np.ndarray],
+    other_data: Dict[str, Any],
+    trajectory_keys: Optional[List[str]] = None,
+    convert_embeddings_to_torch: bool = False,
+) -> List[Dict[str, Any]]:
+    """Reconstruct the original payload structure from .npy files and form data.
+
+    The client sends data in this format:
+    - Files: sample_0_chosen_trajectory_frames.npy, sample_0_trajectory_frames.npy, etc.
+    - Data: sample_0, sample_1, etc. (each containing the full sample JSON with numpy file references)
+
+    Args:
+        numpy_arrays: Dictionary of numpy arrays loaded from .npy files
+        other_data: Dictionary of other form data
+        trajectory_keys: List of trajectory keys to process (default: common keys)
+        convert_embeddings_to_torch: Whether to convert embeddings to torch tensors
+
+    Returns:
+        List of reconstructed sample dictionaries
+    """
+    if trajectory_keys is None:
+        trajectory_keys = [
+            "chosen_trajectory",
+            "rejected_trajectory",
+            "trajectory",
+        ]
+
+    samples = []
+
+    # Process each sample
+    for i in range(len(other_data)):
+        sample_key = f"sample_{i}"
+        if sample_key in other_data:
+            # Get the sample data - might already be parsed or might be a string
+            sample_data = other_data[sample_key]
+            if isinstance(sample_data, str):
+                # Parse the sample JSON if it's a string
+                sample_data = json.loads(sample_data)
+
+            # Replace numpy file references with actual arrays
+            for key, value in sample_data.items():
+                if key in trajectory_keys:
+                    if isinstance(value, dict):
+                        for traj_key, traj_value in value.items():
+                            if isinstance(traj_value, dict) and traj_value.get("__numpy_file__"):
+                                # Replace with actual numpy array
+                                file_key = traj_value["__numpy_file__"]
+                                if file_key in numpy_arrays:
+                                    value[traj_key] = numpy_arrays[file_key]
+
+                            # Convert embeddings to torch if requested
+                            if convert_embeddings_to_torch and traj_key in ["video_embeddings", "text_embedding"]:
+                                if traj_key in value and value[traj_key] is not None:
+                                    if isinstance(value[traj_key], np.ndarray):
+                                        value[traj_key] = torch.tensor(value[traj_key])
+                                    elif isinstance(value[traj_key], list):
+                                        value[traj_key] = torch.tensor(value[traj_key])
+
+            samples.append(sample_data)
+
+    return samples
+
+
+def find_video_files(directory: str) -> list[str]:
+    """Find all video files in a directory.
+
+    Args:
+        directory: Path to directory containing video files
+
+    Returns:
+        List of paths to video files
+    """
+    video_extensions = {".mp4", ".avi", ".mov", ".mkv", ".webm", ".flv", ".wmv", ".m4v"}
+    video_files = []
+
+    directory_path = Path(directory)
+    if not directory_path.is_dir():
+        return []
+
+    for file_path in directory_path.iterdir():
+        if file_path.is_file() and file_path.suffix.lower() in video_extensions:
+            video_files.append(str(file_path))
+
+    video_files.sort()
+    return video_files
+
+
+def infer_task_from_video_name(video_path: str) -> str:
+    """Infer task name from video filename.
+
+    Task is everything before the comma (if comma exists), or everything before success/fail/failure.
+
+    Args:
+        video_path: Path to video file
+
+    Returns:
+        Inferred task name
+    """
+    video_name = Path(video_path).stem  # Get filename without extension
+
+    # If there's a comma, task is everything before the comma
+    if "," in video_name:
+        task_part = video_name.split(",")[0]
+    else:
+        # Otherwise, split by underscore and remove success/fail/failure suffixes
+        parts = video_name.split("_")
+        filtered_parts = []
+        for part in parts:
+            part_lower = part.lower()
+            if part_lower not in ["success", "fail", "failure"]:
+                filtered_parts.append(part)
+
+        if not filtered_parts:
+            return "Complete the task"
+
+        task_part = "_".join(filtered_parts)
+
+    # Split by underscore and join with spaces
+    task_words = task_part.split("_")
+    task = " ".join(task_words)
+
+    if task:
+        # Capitalize first letter of first word, keep rest as is
+        task = task[0].upper() + task[1:] if len(task) > 1 else task.upper()
+    else:
+        task = "Complete the task"
+
+    return task
+
+
+def setup_output_directory(output_dir: Optional[str], video_path: Optional[str] = None) -> str:
+    """Create output directory and return path."""
+    if output_dir:
+        save_dir = output_dir
+    else:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        save_dir = os.path.join(".", f"eval_outputs/{timestamp}")
+
+    os.makedirs(save_dir, exist_ok=True)
+    return save_dir
diff --git a/robometer/robometer/evals/eval_viz_utils.py b/robometer/robometer/evals/eval_viz_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..801a106054e8cdf723054bb4575655056fd92aa0
--- /dev/null
+++ b/robometer/robometer/evals/eval_viz_utils.py
@@ -0,0 +1,205 @@
+#!/usr/bin/env python3
+"""
+Utility functions for visualization in RBM evaluations.
+"""
+
+from typing import Optional
+import os
+import logging
+import tempfile
+import numpy as np
+import matplotlib.pyplot as plt
+import decord
+
+logger = logging.getLogger(__name__)
+
+
+def create_combined_progress_success_plot(
+    progress_pred: np.ndarray,
+    num_frames: int,
+    success_binary: Optional[np.ndarray] = None,
+    success_probs: Optional[np.ndarray] = None,
+    success_labels: Optional[np.ndarray] = None,
+    is_discrete_mode: bool = False,
+    title: Optional[str] = None,
+    loss: Optional[float] = None,
+    pearson: Optional[float] = None,
+) -> plt.Figure:
+    """Create a combined plot with progress, success binary, and success probabilities.
+
+    This function creates a unified plot with 1 subplot (progress only) or 3 subplots
+    (progress, success binary, success probs), similar to the one used in compile_results.py.
+
+    Args:
+        progress_pred: Progress predictions array
+        num_frames: Number of frames
+        success_binary: Optional binary success predictions
+        success_probs: Optional success probability predictions
+        success_labels: Optional ground truth success labels
+        is_discrete_mode: Optional; unused, kept for API compatibility
+        title: Optional title for the plot (if None, auto-generated from loss/pearson)
+        loss: Optional loss value to display in title
+        pearson: Optional pearson correlation to display in title
+
+    Returns:
+        matplotlib Figure object
+    """
+    # Determine if we should show success plots
+    has_success_binary = success_binary is not None and len(success_binary) == len(progress_pred)
+
+    if has_success_binary:
+        # Three subplots: progress, success (binary), success_probs
+        fig, axs = plt.subplots(1, 3, figsize=(15, 3.5))
+        ax = axs[0]  # Progress subplot
+        ax2 = axs[1]  # Success subplot (binary)
+        ax3 = axs[2]  # Success probs subplot
+    else:
+        # Single subplot: progress only
+        fig, ax = plt.subplots(figsize=(6, 3.5))
+        ax2 = None
+        ax3 = None
+
+    # Plot progress
+    ax.plot(progress_pred, linewidth=2)
+    ax.set_ylabel("Progress")
+
+    # Build title
+    if title is None:
+        title_parts = ["Progress"]
+        if loss is not None:
+            title_parts.append(f"Loss: {loss:.3f}")
+        if pearson is not None:
+            title_parts.append(f"Pearson: {pearson:.2f}")
+        title = ", ".join(title_parts)
+    fig.suptitle(title)
+
+    # Set y-limits and ticks (always continuous since discrete is converted before this function)
+    ax.set_ylim(0, 1)
+    ax.spines["right"].set_visible(False)
+    ax.spines["top"].set_visible(False)
+    y_ticks = [0, 0.2, 0.4, 0.6, 0.8, 1.0]
+    ax.set_yticks(y_ticks)
+
+    # Setup success binary subplot
+    if ax2 is not None:
+        ax2.step(range(len(success_binary)), success_binary, where="post", linewidth=2, label="Predicted", color="blue")
+        # Add ground truth success labels as green line if available
+        if success_labels is not None and len(success_labels) == len(success_binary):
+            ax2.step(
+                range(len(success_labels)),
+                success_labels,
+                where="post",
+                linewidth=2,
+                label="Ground Truth",
+                color="green",
+            )
+        ax2.set_ylabel("Success (Binary)")
+        ax2.set_ylim(-0.05, 1.05)
+        ax2.spines["right"].set_visible(False)
+        ax2.spines["top"].set_visible(False)
+        ax2.set_yticks([0, 1])
+        ax2.legend()
+
+    # Setup success probs subplot if available
+    if ax3 is not None and success_probs is not None:
+        ax3.plot(range(len(success_probs)), success_probs, linewidth=2, label="Success Prob", color="purple")
+        # Add ground truth success labels as green line if available
+        if success_labels is not None and len(success_labels) == len(success_probs):
+            ax3.step(
+                range(len(success_labels)),
+                success_labels,
+                where="post",
+                linewidth=2,
+                label="Ground Truth",
+                color="green",
+                linestyle="--",
+            )
+        ax3.set_ylabel("Success Probability")
+        ax3.set_ylim(-0.05, 1.05)
+        ax3.spines["right"].set_visible(False)
+        ax3.spines["top"].set_visible(False)
+        ax3.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0])
+        ax3.legend()
+
+    plt.tight_layout()
+    return fig
+
+
+def extract_frames(video_path: str, fps: float = 1.0, max_frames: int = 64) -> np.ndarray:
+    """Extract frames from video file as numpy array (T, H, W, C).
+
+    Supports both local file paths and URLs (e.g., HuggingFace Hub URLs).
+    Uses the provided ``fps`` to control how densely frames are sampled from
+    the underlying video, but caps the total number of frames at ``max_frames``
+    to prevent memory issues.
+
+    Args:
+        video_path: Path to video file or URL
+        fps: Frames per second to extract (default: 1.0)
+        max_frames: Maximum number of frames to extract (default: 64). This prevents
+            memory issues with long videos or high FPS settings.
+
+    Returns:
+        numpy array of shape (T, H, W, C) containing extracted frames, or None if error
+    """
+    if video_path is None:
+        return None
+
+    if isinstance(video_path, tuple):
+        video_path = video_path[0]
+
+    # Check if it's a URL or local file
+    is_url = video_path.startswith(("http://", "https://"))
+    is_local_file = os.path.exists(video_path) if not is_url else False
+
+    if not is_url and not is_local_file:
+        logger.warning(f"Video path does not exist: {video_path}")
+        return None
+
+    try:
+        # decord.VideoReader can handle both local files and URLs
+        vr = decord.VideoReader(video_path, num_threads=1)
+        total_frames = len(vr)
+
+        # Determine native FPS; fall back to a reasonable default if unavailable
+        try:
+            native_fps = float(vr.get_avg_fps())
+        except Exception:
+            native_fps = 1.0
+
+        # If user-specified fps is invalid or None, default to native fps
+        if fps is None or fps <= 0:
+            fps = native_fps
+
+        # Compute how many frames we want based on desired fps
+        # num_frames ≈ total_duration * fps = total_frames * (fps / native_fps)
+        if native_fps > 0:
+            desired_frames = int(round(total_frames * (fps / native_fps)))
+        else:
+            desired_frames = total_frames
+
+        # Clamp to [1, total_frames]
+        desired_frames = max(1, min(desired_frames, total_frames))
+
+        # IMPORTANT: Cap at max_frames to prevent memory issues
+        # This is critical when fps is high or videos are long
+        if desired_frames > max_frames:
+            logger.warning(
+                f"Requested {desired_frames} frames but capping at {max_frames} "
+                f"to prevent memory issues (video has {total_frames} frames at {native_fps:.2f} fps, "
+                f"requested extraction at {fps:.2f} fps)"
+            )
+            desired_frames = max_frames
+
+        # Evenly sample indices to match the desired number of frames
+        if desired_frames == total_frames:
+            frame_indices = list(range(total_frames))
+        else:
+            frame_indices = np.linspace(0, total_frames - 1, desired_frames, dtype=int).tolist()
+
+        frames_array = vr.get_batch(frame_indices).asnumpy()  # Shape: (T, H, W, C)
+        del vr
+        return frames_array
+    except Exception as e:
+        logger.error(f"Error extracting frames from {video_path}: {e}")
+        return None
diff --git a/robometer/robometer/evals/run_baseline_eval.py b/robometer/robometer/evals/run_baseline_eval.py
new file mode 100644
index 0000000000000000000000000000000000000000..748f74c26df9bd04e748889653646ce083365278
--- /dev/null
+++ b/robometer/robometer/evals/run_baseline_eval.py
@@ -0,0 +1,871 @@
+#!/usr/bin/env python3
+"""
+Script to run baseline evaluations (GVL, RL-VLM-F, VLAC) on datasets.
+
+Usage:
+    # Run VLAC progress evaluation (requires separate dependency set due to trl conflict)
+    uv run --extra vlac --python .venv-vlac/bin/python robometer/evals/run_baseline_eval.py \
+        reward_model=vlac \
+        custom_eval.eval_types=[reward_alignment] \
+        custom_eval.reward_alignment=[jesbu1_roboreward_rfm_roboreward_test] \
+        custom_eval.use_frame_steps=false \
+        max_frames=8 \
+        custom_eval.reward_alignment_max_trajectories=null
+    
+    # Run RoboReward-8B progress evaluation
+    # reward_model=roboreward automatically loads configs/reward_model/roboreward.yaml
+    uv run python robometer/evals/run_baseline_eval.py \
+        reward_model=roboreward \
+        custom_eval.eval_types=[reward_alignment] \
+        custom_eval.reward_alignment=[franka] \
+        custom_eval.use_frame_steps=true \
+        max_frames=8
+    
+    # Run RBM model progress evaluation (reward alignment)
+    # reward_model=rbm loads configs/reward_model/rbm.yaml
+    uv run python robometer/evals/run_baseline_eval.py \
+        reward_model=rbm \
+        model_path=rewardfm/rbm_qwen_pref_prog_4frames_all_strategy \
+        custom_eval.eval_types=[reward_alignment] \
+        custom_eval.reward_alignment=[jesbu1_roboreward_rfm_roboreward_test] \
+        custom_eval.reward_alignment_max_trajectories=null \
+        max_frames=4 \
+        model_config.batch_size=32
+"""
+
+import copy
+import hashlib
+import json
+import os
+import re
+from dataclasses import asdict
+from typing import Optional, Dict, Any, List, Union
+from datetime import datetime
+
+import numpy as np
+from PIL import Image
+import matplotlib.pyplot as plt
+import imageio
+from io import BytesIO
+from hydra import main as hydra_main
+from omegaconf import DictConfig
+from tqdm import tqdm
+
+from robometer.configs.eval_configs import BaselineEvalConfig
+from robometer.configs.experiment_configs import DataConfig
+from robometer.utils.setup_utils import setup_custom_eval_dataset
+from robometer.data.datasets.base import resolve_dataset_keys
+from robometer.utils.distributed import is_rank_0
+from robometer.utils.logger import get_logger
+from robometer.utils.config_utils import display_config, convert_hydra_to_dataclass
+from robometer.data.dataset_types import PreferenceSample, ProgressSample
+from robometer.data.collators.utils import convert_frames_to_pil_images, frames_to_numpy_array
+from robometer.evals.baselines.rlvlmf import RLVLMF
+from robometer.evals.baselines.gvl import GVL
+from robometer.evals.baselines.vlac import VLAC
+
+from robometer.evals.baselines.robodopamine import RoboDopamine
+from robometer.evals.baselines.roboreward import RoboReward
+from robometer.evals.baselines.rbm_model import RBMModel
+from robometer.evals.compile_results import (
+    run_quality_preference_eval,
+    run_reward_alignment_eval_per_trajectory,
+    run_policy_ranking_eval,
+    run_confusion_matrix_eval,
+)
+
+logger = get_logger()
+
+
+def _create_plot_with_video_gif(
+    fig: plt.Figure,
+    video_frames: Optional[np.ndarray],
+    output_path: str,
+    plot_width: int = 800,
+    video_height: int = 224,
+    fps: int = 2,
+) -> None:
+    """Create a GIF combining a static plot with animated video frames side by side.
+
+    Args:
+        fig: Matplotlib figure to include as static plot
+        video_frames: Video frames array of shape [T, C, H, W] or [T, H, W, C]
+        output_path: Path to save the GIF
+        plot_width: Width of the plot in pixels
+        video_height: Height of the video frames in pixels
+        fps: Frames per second for the GIF
+    """
+    if video_frames is None or video_frames.size == 0:
+        # If no video, just save the plot as PNG
+        fig.savefig(output_path.replace(".gif", ".png"), dpi=150, bbox_inches="tight")
+        return
+
+    # Convert matplotlib figure to PIL Image
+    buf = BytesIO()
+    fig.savefig(buf, format="png", dpi=150, bbox_inches="tight")
+    buf.seek(0)
+    plot_img = Image.open(buf)
+    plot_img = plot_img.convert("RGB")
+
+    # Resize plot to desired width while maintaining aspect ratio
+    plot_aspect = plot_img.height / plot_img.width
+    plot_height = int(plot_width * plot_aspect)
+    plot_img = plot_img.resize((plot_width, plot_height), Image.Resampling.LANCZOS)
+
+    # Process video frames
+    # video_frames is [T, C, H, W] - need to convert to [T, H, W, C] for PIL
+    if video_frames.ndim == 4:
+        if video_frames.shape[1] == 3 or video_frames.shape[1] == 1:  # [T, C, H, W]
+            video_frames = video_frames.transpose(0, 2, 3, 1)  # [T, H, W, C]
+        # Now it's [T, H, W, C]
+
+    # Resize video frames to match video_height
+    num_frames = video_frames.shape[0]
+    frame_height, frame_width = video_frames.shape[1], video_frames.shape[2]
+    video_aspect = frame_height / frame_width
+    video_width = int(video_height / video_aspect)
+
+    # Create combined frames
+    combined_frames = []
+    for t in range(num_frames):
+        frame = video_frames[t]
+
+        # Ensure uint8
+        if frame.dtype != np.uint8:
+            if frame.max() <= 1.0:
+                frame = (frame * 255).astype(np.uint8)
+            else:
+                frame = np.clip(frame, 0, 255).astype(np.uint8)
+
+        # Convert to PIL Image
+        if frame.shape[2] == 1:  # Grayscale
+            frame_pil = Image.fromarray(frame[:, :, 0], mode="L").convert("RGB")
+        else:
+            frame_pil = Image.fromarray(frame, mode="RGB")
+
+        # Resize video frame
+        frame_pil = frame_pil.resize((video_width, video_height), Image.Resampling.LANCZOS)
+
+        # Combine plot and video side by side
+        # Use the maximum height and pad if needed
+        max_height = max(plot_height, video_height)
+        combined = Image.new("RGB", (plot_width + video_width, max_height), color="white")
+
+        # Paste plot on the left
+        plot_y = (max_height - plot_height) // 2
+        combined.paste(plot_img, (0, plot_y))
+
+        # Paste video frame on the right
+        video_y = (max_height - video_height) // 2
+        combined.paste(frame_pil, (plot_width, video_y))
+
+        # Convert to numpy array for imageio
+        combined_frames.append(np.array(combined))
+
+    # Save as GIF
+    imageio.mimwrite(output_path, combined_frames, fps=fps, loop=0)
+    plt.close(fig)  # Close figure to free memory
+
+
+def _make_json_serializable(obj: Any) -> Any:
+    """Convert numpy types and other non-serializable objects to JSON-serializable types."""
+    if isinstance(obj, (np.integer, np.int64, np.int32)):
+        return int(obj)
+    elif isinstance(obj, (np.floating, np.float64, np.float32)):
+        return float(obj)
+    elif isinstance(obj, np.ndarray):
+        return obj.tolist()
+    elif isinstance(obj, dict):
+        return {k: _make_json_serializable(v) for k, v in obj.items()}
+    elif isinstance(obj, (list, tuple)):
+        return [_make_json_serializable(item) for item in obj]
+    elif obj is None:
+        return None
+    else:
+        return obj
+
+
+def _shorten_single_name(name: str, mapping: dict) -> str:
+    """Shorten a single dataset name using the mapping."""
+    # Check if there's a direct mapping
+    if name in mapping:
+        return mapping[name]
+
+    # Also check after normalizing (in case the input has special chars)
+    normalized = name.replace("-", "_").replace("/", "_")
+    if normalized in mapping:
+        return mapping[normalized]
+
+    # No mapping found, clean up the name
+    for char in ["/", "\\", ":", "*", "?", '"', "<", ">", "|", " ", ","]:
+        name = name.replace(char, "_")
+    name = re.sub(r"_+", "_", name)
+    return name.strip("_")
+
+
+def _shorten_dataset_name(dataset_name: Union[str, List[str]], max_length: int = 60) -> str:
+    """Shorten dataset name for use in filenames.
+
+    Uses DS_SHORT_NAME_MAPPING for known datasets, otherwise truncates with hash suffix.
+    For lists, each element is shortened individually then combined.
+
+    Args:
+        dataset_name: Dataset name (string or list of strings)
+        max_length: Maximum length for the output string (default 60)
+
+    Returns:
+        Shortened, filesystem-safe string for use in filenames
+    """
+    from robometer.data.datasets.name_mapping import DS_SHORT_NAME_MAPPING
+
+    # Handle list by shortening each element individually
+    if isinstance(dataset_name, list):
+        shortened_parts = [_shorten_single_name(str(x), DS_SHORT_NAME_MAPPING) for x in dataset_name]
+        name_str = "_".join(shortened_parts)
+    else:
+        name_str = _shorten_single_name(str(dataset_name), DS_SHORT_NAME_MAPPING)
+
+    # Collapse multiple underscores
+    name_str = re.sub(r"_+", "_", name_str)
+    name_str = name_str.strip("_")
+
+    # If short enough, return as-is
+    if len(name_str) <= max_length:
+        return name_str
+
+    # Generate short hash of original name for uniqueness
+    original_str = "_".join(str(x) for x in dataset_name) if isinstance(dataset_name, list) else str(dataset_name)
+    hash_suffix = hashlib.md5(original_str.encode()).hexdigest()[:8]
+
+    # Truncate and add hash
+    # Reserve space for hash suffix: "_" + 8 chars = 9 chars
+    truncate_length = max_length - 9
+    truncated = name_str[:truncate_length].rstrip("_")
+
+    return f"{truncated}_{hash_suffix}"
+
+
+def process_preference_sample(sample: PreferenceSample, model: RLVLMF) -> Dict[str, Any]:
+    """Process a single preference sample with baseline."""
+    chosen_traj = sample.chosen_trajectory
+    rejected_traj = sample.rejected_trajectory
+
+    # Convert frames to PIL Images
+    chosen_images = convert_frames_to_pil_images(chosen_traj.frames)
+    rejected_images = convert_frames_to_pil_images(rejected_traj.frames)
+
+    assert chosen_traj.task == rejected_traj.task, "Chosen and rejected trajectories must have the same task"
+
+    # Compute preference
+    result = model.compute_preference(
+        chosen_images=chosen_images,
+        rejected_images=rejected_images,
+        task_description=chosen_traj.task,
+    )
+
+    chosen_metadata = {
+        "quality_label": chosen_traj.quality_label,
+        "data_source": chosen_traj.data_source,
+        "task": chosen_traj.task,
+        "id": chosen_traj.id,
+        "video_path": chosen_traj.frames if isinstance(chosen_traj.frames, str) else None,
+    }
+    if chosen_traj.partial_success is not None:
+        chosen_metadata["partial_success"] = chosen_traj.partial_success
+
+    rejected_metadata = {
+        "quality_label": rejected_traj.quality_label,
+        "data_source": rejected_traj.data_source,
+        "task": rejected_traj.task,
+        "id": rejected_traj.id,
+        "video_path": rejected_traj.frames if isinstance(rejected_traj.frames, str) else None,
+    }
+    if rejected_traj.partial_success is not None:
+        rejected_metadata["partial_success"] = rejected_traj.partial_success
+
+    prediction_prob = result.get("prediction_prob")
+    is_correct = result.get("is_correct")
+    preference_pred = result.get("preference_pred")
+
+    return {
+        "preference_pred": float(preference_pred)
+        if preference_pred is not None
+        else (float(prediction_prob) if prediction_prob is not None else None),
+        "preference_labels": 1.0,  # Always 1.0 because chosen trajectory is always preferred by construction
+        "is_correct": bool(is_correct) if is_correct is not None else None,
+        "task": chosen_traj.task,
+        "data_source": chosen_traj.data_source or rejected_traj.data_source,
+        "chosen_data_gen_strategy": chosen_traj.data_gen_strategy,
+        "rejected_data_gen_strategy": rejected_traj.data_gen_strategy,
+        "metadata": {
+            "chosen_metadata": chosen_metadata,
+            "rejected_metadata": rejected_metadata,
+        },
+    }
+
+
+def process_progress_sample(
+    sample: ProgressSample,
+    model: Union[GVL, VLAC, RoboReward, RoboDopamine],
+) -> Dict[str, Any]:
+    """Process a single progress sample with baseline."""
+    traj = sample.trajectory
+
+    # Get frames array
+    frames_array = frames_to_numpy_array(traj.frames)
+
+    logger.info(f"Processing progress sample with {len(frames_array)} frames")
+
+    if frames_array is None or frames_array.size == 0:
+        logger.warning("No frames found in trajectory")
+        return None
+
+    progress_pred = model.compute_progress(frames_array, task_description=traj.task)
+
+    # Convert to numpy array and normalize to [0, 1] if needed
+    progress_array = np.array([p if p is not None else 0.0 for p in progress_pred])
+    # Note: GVL/VLAC already return normalized [0, 1] values, so no division by 100 needed
+
+    # Build metadata dict - get video_path and frame_step from trajectory metadata
+    metadata = {}
+    if traj.id is not None:
+        metadata["id"] = traj.id
+    if traj.metadata is not None:
+        metadata.update(traj.metadata)
+
+    # Build result dict
+    result = {
+        "progress_pred": progress_array,
+        "task": traj.task,
+        "data_source": traj.data_source,
+        "data_gen_strategy": traj.data_gen_strategy,
+        "metadata": metadata,
+        "id": traj.id,
+        "video_path": metadata.get("video_path"),
+        "partial_success": traj.partial_success,
+        "target_progress": np.array(traj.target_progress),
+        "quality_label": traj.quality_label,
+    }
+
+    return result
+
+
+def process_batched_rbm_samples(
+    dataset,
+    model: RBMModel,
+    batch_size: int = 32,
+) -> List[Dict[str, Any]]:
+    """Process RBM/ReWiND samples using batched computation with minibatching.
+
+    Args:
+        dataset: Dataset object that supports indexing (e.g., CustomEvalDataset)
+        model: RBMModel instance
+        batch_size: Batch size for processing samples
+
+    Returns:
+        List of result dictionaries in the same format as process_progress_sample/process_preference_sample
+    """
+    dataset_len = len(dataset)
+
+    # Group indices by sample type (iterate once, only storing indices)
+    progress_indices = []
+    preference_indices = []
+
+    for i in range(dataset_len):
+        sample = dataset[i]
+        if isinstance(sample, ProgressSample):
+            progress_indices.append(i)
+        elif isinstance(sample, PreferenceSample):
+            preference_indices.append(i)
+        else:
+            logger.warning(f"Unknown sample type: {type(sample)}")
+
+    results = []
+
+    # Process progress samples in minibatches
+    if progress_indices:
+        for batch_start in tqdm(range(0, len(progress_indices), batch_size), desc="Processing progress batches"):
+            batch_indices = progress_indices[batch_start : batch_start + batch_size]
+            batch = [dataset[i] for i in batch_indices]
+            progress_preds = model.compute_batched_progress(batch)
+            for sample, progress_pred in zip(batch, progress_preds):
+                traj = sample.trajectory
+
+                # Build metadata dict - get video_path and frame_step from trajectory metadata
+                metadata = {}
+                if traj.id is not None:
+                    metadata["id"] = traj.id
+                if traj.metadata is not None:
+                    metadata.update(traj.metadata)
+
+                # Build result dict
+                result = {
+                    "progress_pred": np.array(progress_pred),
+                    "task": traj.task,
+                    "data_source": traj.data_source,
+                    "data_gen_strategy": traj.data_gen_strategy,
+                    "metadata": metadata,
+                    "id": traj.id,
+                    "video_path": metadata.get("video_path"),
+                    "partial_success": traj.partial_success,
+                    "target_progress": np.array(traj.target_progress),
+                    "quality_label": traj.quality_label,
+                }
+                results.append(result)
+
+    # Process preference samples in minibatches
+    if preference_indices:
+        for batch_start in tqdm(range(0, len(preference_indices), batch_size), desc="Processing preference batches"):
+            batch_indices = preference_indices[batch_start : batch_start + batch_size]
+            batch = [dataset[i] for i in batch_indices]
+            preference_results = model.compute_batched_preference(batch)
+            for sample, result in zip(batch, preference_results):
+                chosen_traj = sample.chosen_trajectory
+                rejected_traj = sample.rejected_trajectory
+
+                chosen_metadata = {
+                    "quality_label": chosen_traj.quality_label,
+                    "data_source": chosen_traj.data_source,
+                    "task": chosen_traj.task,
+                    "id": chosen_traj.id,
+                    "video_path": chosen_traj.frames if isinstance(chosen_traj.frames, str) else None,
+                }
+                if chosen_traj.partial_success is not None:
+                    chosen_metadata["partial_success"] = chosen_traj.partial_success
+
+                rejected_metadata = {
+                    "quality_label": rejected_traj.quality_label,
+                    "data_source": rejected_traj.data_source,
+                    "task": rejected_traj.task,
+                    "id": rejected_traj.id,
+                    "video_path": rejected_traj.frames if isinstance(rejected_traj.frames, str) else None,
+                }
+                if rejected_traj.partial_success is not None:
+                    rejected_metadata["partial_success"] = rejected_traj.partial_success
+
+                prediction_prob = result.get("prediction_prob")
+                is_correct = result.get("is_correct")
+                preference_pred = result.get("preference_pred")
+
+                formatted_result = {
+                    "preference_pred": float(preference_pred)
+                    if preference_pred is not None
+                    else (float(prediction_prob) if prediction_prob is not None else None),
+                    "preference_labels": 1.0,  # Always 1.0 because chosen trajectory is always preferred by construction
+                    "is_correct": bool(is_correct) if is_correct is not None else None,
+                    "task": chosen_traj.task,
+                    "data_source": chosen_traj.data_source or rejected_traj.data_source,
+                    "chosen_data_gen_strategy": chosen_traj.data_gen_strategy,
+                    "rejected_data_gen_strategy": rejected_traj.data_gen_strategy,
+                    "metadata": {
+                        "chosen_metadata": chosen_metadata,
+                        "rejected_metadata": rejected_metadata,
+                    },
+                }
+                results.append(formatted_result)
+
+    return results
+
+
+def run_baseline_evaluation(cfg: BaselineEvalConfig, base_data_cfg: DataConfig) -> Dict[str, Any]:
+    """Run baseline evaluation on datasets."""
+
+    # Initialize model based on reward_model type
+    model_config_dict = (
+        asdict(cfg.model_config) if hasattr(cfg.model_config, "__dataclass_fields__") else cfg.model_config.__dict__
+    )
+
+    if cfg.reward_model == "rlvlmf":
+        model = RLVLMF(**model_config_dict)
+    elif cfg.reward_model == "gvl":
+        # API key is read from GEMINI_API_KEY environment variable if not provided
+        model = GVL(max_frames=cfg.max_frames, **model_config_dict)
+    elif cfg.reward_model == "vlac":
+        if not cfg.model_path:
+            raise ValueError("model_path is required for VLAC baseline")
+        model = VLAC(model_path=cfg.model_path, **model_config_dict)
+    elif cfg.reward_model == "robodopamine":
+        if not cfg.model_path:
+            raise ValueError("model_path is required for Robo-Dopamine baseline")
+        model = RoboDopamine(model_path=cfg.model_path, **model_config_dict)
+    elif cfg.reward_model == "roboreward":
+        model = RoboReward(model_path=cfg.model_path or "teetone/RoboReward-4B", **model_config_dict)
+    elif cfg.reward_model in ["rewind", "rbm"]:
+        if not cfg.model_path:
+            raise ValueError("model_path is required for RBM/ReWiND reward model")
+        model = RBMModel(checkpoint_path=cfg.model_path)
+    else:
+        raise ValueError(
+            f"Unknown reward_model: {cfg.reward_model}. Must be 'rlvlmf', 'gvl', 'vlac', 'robodopamine', 'roboreward', 'rbm', or 'rewind'"
+        )
+
+    all_metrics = {}
+
+    # Process each evaluation type
+    for eval_type in cfg.custom_eval.eval_types:
+        logger.info(f"=" * 80)
+        logger.info(f"Running {eval_type} evaluation with {cfg.reward_model} reward model")
+        logger.info(f"=" * 80)
+
+        # Get datasets for this eval type
+        eval_datasets = getattr(cfg.custom_eval, eval_type, [])
+        if not eval_datasets:
+            logger.warning(f"No datasets specified for {eval_type}, skipping")
+            continue
+
+        # Create eval_type subdirectory: {model_type}/{eval_type}/...
+        eval_type_dir = os.path.join(cfg.output_dir, eval_type)
+        os.makedirs(eval_type_dir, exist_ok=True)
+        logger.info(f"Saving results to: {eval_type_dir}")
+
+        eval_type_metrics = {}
+
+        logger.info(f"Eval datasets: {eval_datasets}")
+
+        resolved_datasets = []
+
+        for eval_dataset in eval_datasets:
+            if isinstance(eval_dataset, list):
+                resolved_datasets.append(eval_dataset)
+            else:
+                resolved_datasets.extend(resolve_dataset_keys([eval_dataset], split="eval"))
+
+            logger.info(f"Resolved datasets for {eval_type}: {resolved_datasets}")
+
+        for dataset_name in resolved_datasets:
+            # Create short name for filenames (dataset names can be very long)
+            short_dataset_name = _shorten_dataset_name(dataset_name)
+
+            # Resolve dataset keys
+            if isinstance(dataset_name, list):
+                resolved_dataset_name = dataset_name
+            else:
+                resolved_dataset_name = resolve_dataset_keys([dataset_name], split="eval")
+
+            logger.info(f"Resolved datasets for {eval_type}: {resolved_datasets}")
+
+            # Create data config for this dataset (similar to trainer)
+            eval_data_cfg = copy.deepcopy(base_data_cfg)
+            eval_data_cfg.dataset_type = "rbm"
+            eval_data_cfg.eval_datasets = resolved_dataset_name
+
+            # Setup dataset
+            sampler_kwargs = {
+                "random_seed": cfg.custom_eval.custom_eval_random_seed,
+            }
+
+            if cfg.reward_model == "roboreward":
+                sampler_kwargs["pad_frames"] = False
+
+            if eval_type == "reward_alignment":
+                sampler_kwargs["max_trajectories"] = cfg.custom_eval.reward_alignment_max_trajectories
+                sampler_kwargs["use_frame_steps"] = cfg.custom_eval.use_frame_steps
+                sampler_kwargs["subsample_n_frames"] = cfg.custom_eval.subsample_n_frames
+                sampler_kwargs["pad_frames"] = cfg.custom_eval.pad_frames
+            elif eval_type == "policy_ranking":
+                sampler_kwargs["num_examples_per_quality_pr"] = cfg.custom_eval.num_examples_per_quality_pr
+                sampler_kwargs["num_partial_successes"] = cfg.custom_eval.num_partial_successes
+                sampler_kwargs["max_tasks"] = cfg.custom_eval.policy_ranking_max_tasks
+                sampler_kwargs["use_frame_steps"] = cfg.custom_eval.use_frame_steps
+                sampler_kwargs["pad_frames"] = cfg.custom_eval.pad_frames
+            elif eval_type == "confusion_matrix":
+                sampler_kwargs["pad_frames"] = cfg.custom_eval.pad_frames
+            elif "quality_preference" in eval_type:
+                sampler_kwargs["comparisons_per_task"] = cfg.custom_eval.comparisons_per_task
+                sampler_kwargs["max_comparisons"] = cfg.custom_eval.max_comparisons
+
+            dataset = setup_custom_eval_dataset(
+                cfg=eval_data_cfg, sampler_type=eval_type, verbose=True, sampler_kwargs=sampler_kwargs
+            )
+
+            # Process samples
+            eval_results = []
+
+            if cfg.reward_model in ["rewind", "rbm"]:
+                # For RBM/ReWiND, process dataset using indices to avoid materializing entire dataset
+                logger.info(f"Processing {len(dataset)} samples in batches for RBM/ReWiND")
+
+                model_config_dict = (
+                    asdict(cfg.model_config)
+                    if hasattr(cfg.model_config, "__dataclass_fields__")
+                    else cfg.model_config.__dict__
+                )
+                batch_results = process_batched_rbm_samples(dataset, model, batch_size=model_config_dict["batch_size"])
+                eval_results.extend(batch_results)
+            else:
+                # For other models, process samples one at a time
+                for i, sample in enumerate(tqdm(dataset, desc=f"Processing {dataset_name}")):
+                    if cfg.reward_model == "rlvlmf" and isinstance(sample, PreferenceSample):
+                        result = process_preference_sample(sample, model)
+                        if result:
+                            eval_results.append(result)
+                    elif cfg.reward_model in ["gvl", "vlac", "roboreward", "robodopamine"] and isinstance(
+                        sample, ProgressSample
+                    ):
+                        # Handle ProgressSamples for gvl/vlac/roboreward (including confusion_matrix)
+                        result = process_progress_sample(sample, model)
+                        if result:
+                            eval_results.append(result)
+                    else:
+                        logger.warning(f"Sample type mismatch: reward_model={cfg.reward_model}, sample={type(sample)}")
+
+            logger.info(f"Processed {len(eval_results)} samples from {dataset_name}")
+
+            # Save results to JSON
+            if eval_type_dir:
+                results_file = os.path.join(eval_type_dir, f"{short_dataset_name}_results.json")
+                with open(results_file, "w") as f:
+                    json.dump(_make_json_serializable(eval_results), f, indent=2)
+                logger.info(f"Saved results to {results_file}")
+
+            # Compute metrics using the same functions as the trainer
+            if eval_results:
+                # Determine data_source from first result
+                data_source = eval_results[0].get("data_source") if eval_results else None
+
+                if eval_type == "quality_preference":
+                    # Quality preference evaluation for rlvlmf, rbm, rewind
+                    if cfg.reward_model not in ["rlvlmf", "rbm", "rewind"]:
+                        raise ValueError(
+                            f"quality_preference evaluation only supported for rlvlmf, rbm, rewind, got {cfg.reward_model}"
+                        )
+
+                    metrics_dict, task_groups, task_details = run_quality_preference_eval(
+                        results=eval_results,
+                        data_source=data_source,
+                    )
+                    # Save task_groups and task_details if available
+                    if eval_type_dir:
+                        task_groups_file = os.path.join(eval_type_dir, f"{short_dataset_name}_task_groups.json")
+                        task_details_file = os.path.join(eval_type_dir, f"{short_dataset_name}_task_details.json")
+                        with open(task_groups_file, "w") as f:
+                            json.dump(_make_json_serializable(task_groups), f, indent=2)
+                        with open(task_details_file, "w") as f:
+                            json.dump(_make_json_serializable(task_details), f, indent=2)
+                        logger.info(f"Saved task_groups to {task_groups_file}")
+                        logger.info(f"Saved task_details to {task_details_file}")
+
+                    # Extract metrics from the returned dict
+                    for key, value in metrics_dict.items():
+                        if isinstance(value, (int, float)):
+                            eval_type_metrics[f"{short_dataset_name}/{key}"] = float(value)
+
+                    # Write metrics incrementally after each dataset
+                    if eval_type_dir:
+                        _write_metrics_incremental(eval_type_dir, eval_type_metrics)
+
+                elif eval_type == "confusion_matrix":
+                    # Confusion matrix evaluation for gvl, vlac, roboreward, rbm, rewind
+                    if cfg.reward_model not in ["gvl", "vlac", "roboreward", "robodopamine", "rbm", "rewind"]:
+                        raise ValueError(
+                            f"confusion_matrix evaluation only supported for gvl, vlac, roboreward, robodopamine, rbm, rewind, got {cfg.reward_model}"
+                        )
+
+                    # run_confusion_matrix_eval returns (fig, confusion_matrix, metrics)
+                    fig, confusion_matrix, metrics_dict = run_confusion_matrix_eval(
+                        results=eval_results,
+                        progress_pred_type="absolute_wrt_total_frames",  # Baselines use absolute progress
+                        is_discrete_mode=False,  # Baselines output continuous values
+                        num_bins=None,
+                    )
+
+                    # Save confusion matrix plot
+                    if fig and eval_type_dir:
+                        plot_path = os.path.join(eval_type_dir, f"{short_dataset_name}_confusion_matrix.png")
+                        fig.savefig(plot_path, dpi=150, bbox_inches="tight")
+                        plt.close(fig)
+                        logger.info(f"Saved confusion matrix plot to {plot_path}")
+
+                        # Save confusion matrix as numpy array
+                        matrix_path = os.path.join(eval_type_dir, f"{short_dataset_name}_confusion_matrix.npy")
+                        np.save(matrix_path, confusion_matrix)
+                        logger.info(f"Saved confusion matrix array to {matrix_path}")
+
+                    # Extract metrics from the returned dict (if any)
+                    for key, value in metrics_dict.items():
+                        if isinstance(value, (int, float)):
+                            eval_type_metrics[f"{short_dataset_name}/{key}"] = float(value)
+
+                    # Write metrics incrementally after each dataset
+                    if eval_type_dir:
+                        _write_metrics_incremental(eval_type_dir, eval_type_metrics)
+
+                else:
+                    # Progress evaluation (reward_alignment, policy_ranking) for gvl, vlac, roboreward, rbm, rewind
+                    if cfg.reward_model not in ["gvl", "vlac", "roboreward", "robodopamine", "rbm", "rewind"]:
+                        raise ValueError(
+                            f"Progress evaluation only supported for gvl, vlac, roboreward, robodopamine, rbm, rewind, got {cfg.reward_model}"
+                        )
+
+                    if eval_type == "reward_alignment":
+                        metrics_dict, plots, video_frames_list, _ = run_reward_alignment_eval_per_trajectory(
+                            results=eval_results,
+                            progress_pred_type="absolute_wrt_total_frames",  # Baselines use absolute progress
+                            is_discrete_mode=False,  # Baselines output continuous values
+                            num_bins=None,
+                            data_source=data_source,
+                            use_frame_steps=cfg.custom_eval.use_frame_steps,
+                            train_success_head=False,  # Baselines don't have success head
+                            last_frame_only=False,
+                        )
+                        # Save plots with videos as GIFs if available
+                        if plots and eval_type_dir:
+                            plots_dir = os.path.join(eval_type_dir, f"{short_dataset_name}_plots")
+                            os.makedirs(plots_dir, exist_ok=True)
+                            for i, fig in enumerate(plots[:10]):
+                                video_frames = video_frames_list[i] if i < len(video_frames_list) else None
+                                gif_path = os.path.join(plots_dir, f"trajectory_{i:04d}.gif")
+                                _create_plot_with_video_gif(fig, video_frames, gif_path)
+                            logger.info(f"Saved {len(plots)} plot+video GIFs to {plots_dir}")
+                    elif eval_type == "policy_ranking":
+                        metrics_dict, task_groups, task_details = run_policy_ranking_eval(
+                            results=eval_results,
+                            progress_pred_type="absolute_wrt_total_frames",  # Baselines use absolute progress
+                            is_discrete_mode=False,  # Baselines output continuous values
+                            num_bins=None,
+                            data_source=data_source,
+                            correlation_method="kendall",
+                        )
+                        # Save task_groups and task_details if available
+                        if eval_type_dir:
+                            task_groups_file = os.path.join(eval_type_dir, f"{short_dataset_name}_task_groups.json")
+                            task_details_file = os.path.join(eval_type_dir, f"{short_dataset_name}_task_details.json")
+                            with open(task_groups_file, "w") as f:
+                                json.dump(_make_json_serializable(task_groups), f, indent=2)
+                            with open(task_details_file, "w") as f:
+                                json.dump(_make_json_serializable(task_details), f, indent=2)
+                            logger.info(f"Saved task_groups to {task_groups_file}")
+                            logger.info(f"Saved task_details to {task_details_file}")
+                    else:
+                        raise ValueError(f"Unknown eval_type: {eval_type}")
+
+                    # Extract metrics from the returned dict
+                    for key, value in metrics_dict.items():
+                        if isinstance(value, (int, float)):
+                            eval_type_metrics[f"{short_dataset_name}/{key}"] = float(value)
+
+                    # Write metrics incrementally after each dataset
+                    if eval_type_dir:
+                        _write_metrics_incremental(eval_type_dir, eval_type_metrics)
+
+        all_metrics[eval_type] = eval_type_metrics
+
+    return all_metrics
+
+
+def _write_metrics_incremental(eval_type_dir: str, eval_type_metrics: Dict[str, float]):
+    """Write metrics to file incrementally after each dataset.
+
+    Args:
+        eval_type_dir: Directory for this eval_type (e.g., ./baseline_eval_output/robometer/policy_ranking/)
+        eval_type_metrics: Dictionary of metrics (keyed by dataset_name/metric_name)
+    """
+    metrics_file = os.path.join(eval_type_dir, "metrics.json")
+
+    # Write metrics to file (overwrite since we're accumulating in eval_type_metrics)
+    try:
+        with open(metrics_file, "w") as f:
+            json.dump(_make_json_serializable(eval_type_metrics), f, indent=2)
+        logger.info(f"Updated metrics file: {metrics_file}")
+    except IOError as e:
+        logger.error(f"Could not write metrics file: {e}")
+
+
+def _normalize_model_path(model_path: Optional[str]) -> str:
+    """Normalize model path for use in directory names.
+
+    Handles HuggingFace paths like 'rewardfm/rbm-base' or local paths.
+    Replaces slashes, special characters with underscores.
+
+    Args:
+        model_path: Model path string (e.g., 'rewardfm/rbm-base', '/path/to/model')
+
+    Returns:
+        Normalized string safe for directory names
+    """
+    if not model_path:
+        return ""
+
+    # Get the basename if it's an absolute path
+    if model_path.startswith("/"):
+        # For absolute paths, use the last two components (parent/name)
+        parts = model_path.rstrip("/").split("/")
+        model_path = "_".join(parts[-2:]) if len(parts) >= 2 else parts[-1]
+
+    # Replace slashes with underscores
+    normalized = model_path.replace("/", "_")
+
+    # Replace other special characters that might cause issues
+    for char in [":", "\\", " ", ".", ","]:
+        normalized = normalized.replace(char, "_")
+
+    # Remove leading/trailing underscores and collapse multiple underscores
+    normalized = re.sub(r"_+", "_", normalized)
+    normalized = normalized.strip("_")
+
+    return normalized
+
+
+@hydra_main(version_base=None, config_path="../configs", config_name="baseline_eval_config")
+def main(cfg: DictConfig):
+    """Main entry point for baseline evaluation."""
+    # Convert Hydra config to dataclass
+    baseline_cfg = convert_hydra_to_dataclass(cfg, BaselineEvalConfig)
+
+    # Display config
+    display_config(baseline_cfg)
+
+    # Validate reward model
+    if baseline_cfg.reward_model not in ["gvl", "vlac", "rlvlmf", "roboreward", "robodopamine", "rbm", "rewind"]:
+        raise ValueError(
+            f"reward_model must be 'gvl', 'vlac', 'rlvlmf', 'roboreward', 'robodopamine', 'rbm', or 'rewind', got {baseline_cfg.reward_model}"
+        )
+
+    # Setup output directory: {model_type}_{model_path}/{eval_type}/...
+    if baseline_cfg.output_dir is None:
+        normalized_path = _normalize_model_path(baseline_cfg.model_path)
+
+        if normalized_path:
+            dir_name = f"{baseline_cfg.reward_model}_{normalized_path}"
+        else:
+            dir_name = baseline_cfg.reward_model
+
+        baseline_cfg.output_dir = os.path.join("./baseline_eval_output", dir_name)
+
+    os.makedirs(baseline_cfg.output_dir, exist_ok=True)
+    logger.info(f"Output directory: {baseline_cfg.output_dir}")
+
+    # Create data config with default settings
+    # Datasets will be set per eval type during processing
+    data_cfg = DataConfig(
+        max_frames=baseline_cfg.max_frames,
+        load_embeddings=True if "rewind" in baseline_cfg.reward_model else False,
+    )
+
+    display_config(data_cfg)
+
+    # Run evaluation
+    metrics = run_baseline_evaluation(baseline_cfg, data_cfg)
+
+    # Save all metrics summary to model root directory
+    if metrics and is_rank_0():
+        metrics_file = os.path.join(baseline_cfg.output_dir, "all_metrics.json")
+        metrics_serializable = {}
+        for k, v in metrics.items():
+            if isinstance(v, dict):
+                metrics_serializable[k] = {
+                    k2: float(v2) if isinstance(v2, (int, float, np.number)) else v2 for k2, v2 in v.items()
+                }
+            else:
+                metrics_serializable[k] = v
+
+        with open(metrics_file, "w") as f:
+            json.dump(metrics_serializable, f, indent=2)
+        logger.info(f"Saved all metrics summary to: {metrics_file}")
+
+    logger.info("\nBaseline evaluation complete!")
+    return metrics
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/robometer/models/__init__.py b/robometer/robometer/models/__init__.py
new file mode 100755
index 0000000000000000000000000000000000000000..66bd49f0c62fc9c3bd7489d5f6627775c107284c
--- /dev/null
+++ b/robometer/robometer/models/__init__.py
@@ -0,0 +1,8 @@
+from .rewind_transformer import ReWiNDTransformer, ReWINDTransformerConfig
+from .rbm import RBM
+
+__all__ = [
+    "RBM",
+    "ReWiNDTransformer",
+    "ReWINDTransformerConfig",
+]
diff --git a/robometer/robometer/models/heads.py b/robometer/robometer/models/heads.py
new file mode 100644
index 0000000000000000000000000000000000000000..68f7c7ab511907d60656e62df4d2927adb504a1c
--- /dev/null
+++ b/robometer/robometer/models/heads.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+"""
+Shared prediction heads plugin for RBM and ReWiND models.
+Provides a mixin class that models can inherit from to get prediction heads.
+"""
+
+import torch
+import torch.nn as nn
+from typing import Optional
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class PredictionHeadsMixin(nn.Module):
+    """
+    Mixin class that provides prediction heads for reward models.
+
+    Models should inherit from this mixin and pass hidden_dim and optionally model_config
+    via **kwargs to super().__init__().
+    """
+
+    def __init__(
+        self,
+        *args,
+        hidden_dim: Optional[int] = None,
+        model_config: Optional[object] = None,
+        dropout: float = 0.1,
+        **kwargs,
+    ):
+        """
+        Initialize prediction heads if parameters are provided.
+
+        Args:
+            hidden_dim: Input hidden dimension for all heads (if None, heads won't be initialized)
+            model_config: Optional model config object with loss settings (if provided, extracts progress settings from it)
+            dropout: Dropout rate for all heads
+        """
+        super().__init__(*args, **kwargs)
+
+        if hidden_dim is not None:
+            # Extract progress settings from model_config if provided
+            if model_config is not None:
+                progress_output_size = 1  # Default: continuous output
+                progress_use_sigmoid = True
+                use_discrete_progress = False
+                # Check for progress_loss_type in model_config (direct attribute or under loss)
+                progress_loss_type = None
+                if hasattr(model_config, "progress_loss_type"):
+                    progress_loss_type = model_config.progress_loss_type
+                elif hasattr(model_config, "loss") and hasattr(model_config.loss, "progress_loss_type"):
+                    progress_loss_type = model_config.loss.progress_loss_type
+
+                if progress_loss_type and progress_loss_type.lower() == "discrete":
+                    use_discrete_progress = True
+                    if hasattr(model_config, "progress_discrete_bins"):
+                        progress_output_size = model_config.progress_discrete_bins
+                    elif hasattr(model_config, "loss") and hasattr(model_config.loss, "progress_discrete_bins"):
+                        progress_output_size = model_config.loss.progress_discrete_bins
+                    else:
+                        progress_output_size = 10  # Default bins
+                    progress_use_sigmoid = False
+                # Set use_discrete_progress attribute for models to use
+                self.use_discrete_progress = use_discrete_progress
+            else:
+                # Default values if no model_config
+                progress_output_size = 1
+                progress_use_sigmoid = True
+                self.use_discrete_progress = False
+
+            logger.info(f"PredictionHeadsMixin. __init__: use_discrete_progress: {self.use_discrete_progress}")
+
+            # Progress head
+            progress_layers = [
+                nn.Linear(hidden_dim, hidden_dim // 2),
+                nn.LayerNorm(hidden_dim // 2),
+                nn.GELU(),
+                nn.Dropout(dropout),
+                nn.Linear(hidden_dim // 2, progress_output_size),
+            ]
+            if progress_use_sigmoid:
+                progress_layers.append(nn.Sigmoid())
+            self.progress_head = nn.Sequential(*progress_layers)
+
+            # Preference head
+            self.preference_head = nn.Sequential(
+                nn.Linear(hidden_dim, hidden_dim // 2),
+                nn.LayerNorm(hidden_dim // 2),
+                nn.GELU(),
+                nn.Dropout(dropout),
+                nn.Linear(hidden_dim // 2, 1),
+            )
+
+            # Success head
+            self.success_head = nn.Sequential(
+                nn.Linear(hidden_dim, hidden_dim // 2),
+                nn.LayerNorm(hidden_dim // 2),
+                nn.GELU(),
+                nn.Dropout(dropout),
+                nn.Linear(hidden_dim // 2, 1),
+            )
diff --git a/robometer/robometer/models/rbm.py b/robometer/robometer/models/rbm.py
new file mode 100755
index 0000000000000000000000000000000000000000..112c0daa0479cae64089eec5543fe60df72337f1
--- /dev/null
+++ b/robometer/robometer/models/rbm.py
@@ -0,0 +1,993 @@
+#!/usr/bin/env python3
+"""
+Robometer implementation.
+Contains the RBM class with three prediction heads for different objectives.
+
+Note: make sure that the forward pass uses all of the
+heads or there will be some problems with FSDP sharding.
+"""
+
+import torch
+import torch.nn as nn
+from transformers import PreTrainedModel, Qwen2_5_VLModel
+
+try:
+    from transformers import Qwen3VLModel
+except ImportError:
+    Qwen3VLModel = None
+
+# from transformers import AutoModelForImageTextToText as Molmo2VLModel  # Molmo2 uses AutoModelForImageTextToText
+from transformers import SmolVLMModel
+import torch.distributed as dist
+
+from robometer.models.utils import ModelOutput
+from robometer.models.heads import PredictionHeadsMixin
+from robometer.utils.timer import _timer
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+
+
+def squeeze_last_safe(x: torch.Tensor) -> torch.Tensor:
+    if x.ndim > 1 and x.shape[-1] == 1:
+        return x.squeeze(-1)
+    return x
+
+
+class RBM(PredictionHeadsMixin, PreTrainedModel):
+    """Robometer (RBM) with three prediction heads for different objectives.
+
+    Supports multiple base model architectures:
+    - Qwen2.5-VL (Qwen2_5_VLModel)
+    - SmolVLM (AutoModelForImageTextToText)
+    """
+
+    # unused param i think
+    config_class = Qwen2_5_VLModel.config_class
+
+    # Declare support for SDPA and Flash Attention (will delegate to underlying model), needed for Qwen3
+    _supports_sdpa = True
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config, processor, tokenizer, base_model=None, base_model_id=None, model_config=None):
+        if "SmolVLM" in base_model_id:
+            hidden_size = config.text_config.hidden_size
+            self.model_cls = SmolVLMModel
+        elif "Qwen2.5" in base_model_id:
+            hidden_size = config.hidden_size
+            self.model_cls = Qwen2_5_VLModel
+        elif "Qwen3" in base_model_id:
+            hidden_size = config.text_config.hidden_size
+            self.model_cls = Qwen3VLModel
+        elif "Molmo" in base_model_id:
+            # Molmo2 is based on Qwen3 architecture
+            hidden_size = config.text_config.hidden_size
+            self.model_cls = Qwen3VLModel
+            # self.model_cls = Molmo2VLModel
+        else:
+            raise ValueError(f"Unsupported base model: {base_model_id}")
+
+        super().__init__(
+            config,
+            hidden_dim=hidden_size,
+            model_config=model_config,
+            dropout=0.1,
+        )
+
+        if base_model is not None:
+            self.model = base_model
+        else:
+            self.model = self.model_cls(config)
+
+        self.config_class = self.model_cls.config_class
+        self.base_model_id = base_model_id
+
+        self.model_dtype = self.model.dtype
+        self.progress_head = self.progress_head.to(dtype=self.model_dtype)
+        self.preference_head = self.preference_head.to(dtype=self.model_dtype)
+        self.success_head = self.success_head.to(dtype=self.model_dtype)
+
+        self.processor = processor
+        self.tokenizer = tokenizer
+        self.model_config = model_config
+
+        self.average_temporal_patches = self.model_config.average_temporal_patches
+        self.use_per_frame_progress_token = self.model_config.use_per_frame_progress_token
+        self.use_multi_image = self.model_config.use_multi_image
+
+        # Frame pooling strategy for multi-image mode (used when NOT using per-frame progress tokens).
+        # - mean: average pool patch tokens in the frame span
+        # - boundary: use the last patch token in the frame span
+        # - attention: learned attention pooling over patch tokens in the frame span
+        self.frame_pooling = getattr(self.model_config, "frame_pooling", "mean")
+        self.frame_pooling_attn_temperature = float(getattr(self.model_config, "frame_pooling_attn_temperature", 1.0))
+        if self.frame_pooling_attn_temperature <= 0:
+            raise ValueError("frame_pooling_attn_temperature must be > 0")
+        # Always create the attention pooling projection so checkpoints can be loaded across pooling modes.
+        self.frame_pool_attn = nn.Linear(hidden_size, 1, bias=False).to(dtype=self.model_dtype)
+
+        # Validate that use_per_frame_progress_token requires use_multi_image
+        if self.use_per_frame_progress_token and not self.use_multi_image:
+            raise ValueError(
+                "use_per_frame_progress_token=True requires use_multi_image=True. "
+                "Per-frame progress tokens can only be used in multi-image mode."
+            )
+
+        # Molmo2 only supports multi-image mode, not video
+        if "Molmo" in self.base_model_id and not self.use_multi_image:
+            raise ValueError(
+                "Molmo2 does not support video mode (use_multi_image=False). "
+                "Please set data.use_multi_image=True to use Molmo2 with multi-image input."
+            )
+
+    def gradient_checkpointing_enable(self, **kwargs):
+        """Delegates gradient checkpointing enabling to the base model."""
+        self.model.gradient_checkpointing_enable(**kwargs)
+
+    def gradient_checkpointing_disable(self, **kwargs):
+        """Delegates gradient checkpointing disabling to the base model."""
+        self.model.gradient_checkpointing_disable(**kwargs)
+
+    def generate(self, *args, **kwargs):
+        """Delegates generation to the base model."""
+        return self.model.generate(*args, **kwargs)
+
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        """Delegates input preparation for generation to the base model."""
+        return self.model.prepare_inputs_for_generation(*args, **kwargs)
+
+    def _extract_hidden_states_from_token_pairs(
+        self,
+        hidden_state: torch.Tensor,
+        input_ids: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Extract image/video frame embeddings from hidden states by finding token pairs and mean pooling.
+
+        This is a general function that works for both SmolVLM and Qwen multi-image mode.
+        It automatically detects which model is being used based on the base_model_id:
+        - SmolVLM: Uses the same token for start and end: <fake_token_around_image>
+        - Qwen: Uses different tokens: <|vision_start|> and <|vision_end|>
+
+        Args:
+            hidden_state: Hidden states tensor [seq_len, hidden_dim]
+            input_ids: Input token IDs [seq_len]
+
+        Returns:
+            frame_embeddings: Tensor [num_frames, hidden_dim] containing mean-pooled
+                            embeddings for each frame/image between token pairs
+        """
+        # Detect model type and get appropriate tokenizer and tokens
+        is_molmo = "Molmo" in self.base_model_id
+        if "SmolVLM" in self.base_model_id:
+            # SmolVLM mode: same token appears in pairs
+            tokenizer = self.tokenizer
+            start_token = "<fake_token_around_image>"
+            end_token = None  # Same token for both start and end
+            use_same_token = True
+            use_molmo_mode = False
+        elif is_molmo:
+            # Molmo2 mode: <low_res_im_start> followed by <im_patch> tokens
+            tokenizer = self.processor.tokenizer
+            start_token = "<low_res_im_start>"
+            end_token = None  # No explicit end token
+            patch_token = "<im_patch>"
+            use_same_token = False
+            use_molmo_mode = True
+        else:
+            # Qwen mode: different start and end tokens
+            tokenizer = self.processor.tokenizer
+            start_token = "<|vision_start|>"
+            end_token = "<|vision_end|>"
+            use_same_token = False
+            use_molmo_mode = False
+
+        # Get token IDs
+        start_token_id = tokenizer.convert_tokens_to_ids(start_token)
+
+        # Find all positions where start tokens appear
+        start_positions = (input_ids == start_token_id).nonzero(as_tuple=True)[0]
+
+        if len(start_positions) == 0:
+            raise ValueError(
+                f"No {start_token} tokens found in input_ids. Token ID {start_token_id} not found in sequence."
+            )
+
+        # Handle different pairing modes
+        if use_same_token:
+            # SmolVLM mode: same token appears in pairs
+            if len(start_positions) % 2 != 0:
+                raise ValueError(
+                    f"Expected even number of {start_token} tokens (pairs), but found {len(start_positions)} tokens."
+                )
+
+            # Group tokens into pairs (every two consecutive tokens form a pair)
+            token_pairs = []
+            for i in range(0, len(start_positions), 2):
+                token_pairs.append((start_positions[i].item(), start_positions[i + 1].item()))
+        elif use_molmo_mode:
+            # Molmo2 mode: <low_res_im_start> followed by <im_patch> tokens
+            patch_token_id = tokenizer.convert_tokens_to_ids(patch_token)
+            im_patch_positions = (input_ids == patch_token_id).nonzero(as_tuple=True)[0]
+
+            token_pairs = []
+            for start_idx, start_pos in enumerate(start_positions):
+                start_pos_val = start_pos.item()
+                # Find the last consecutive im_patch token after this start
+                patches_after_start = im_patch_positions[im_patch_positions > start_pos]
+                if len(patches_after_start) > 0:
+                    # Find where patches stop (at next image start or end of sequence)
+                    if start_idx + 1 < len(start_positions):
+                        next_start = start_positions[start_idx + 1].item()
+                        patches_for_this_image = patches_after_start[patches_after_start < next_start]
+                    else:
+                        patches_for_this_image = patches_after_start
+                    if len(patches_for_this_image) > 0:
+                        end_pos = patches_for_this_image[-1].item()
+                        token_pairs.append((start_pos_val, end_pos))
+        else:
+            # Qwen mode: different start and end tokens
+            end_token_id = tokenizer.convert_tokens_to_ids(end_token)
+
+            # Find all positions where end tokens appear
+            end_positions = (input_ids == end_token_id).nonzero(as_tuple=True)[0]
+
+            if len(end_positions) == 0:
+                raise ValueError(
+                    f"No {end_token} tokens found in input_ids. Token ID {end_token_id} not found in sequence."
+                )
+
+            if len(start_positions) != len(end_positions):
+                raise ValueError(
+                    f"Mismatched number of tokens: "
+                    f"found {len(start_positions)} {start_token} tokens "
+                    f"and {len(end_positions)} {end_token} tokens."
+                )
+
+            # Pair up start and end tokens (they should appear in order: start, end, start, end, ...)
+            token_pairs = []
+            for i in range(len(start_positions)):
+                start_pos = start_positions[i].item()
+                end_pos = end_positions[i].item()
+
+                if start_pos >= end_pos:
+                    raise ValueError(
+                        f"Invalid token pair at index {i}: "
+                        f"{start_token} at {start_pos}, {end_token} at {end_pos}. "
+                        f"Start must come before end."
+                    )
+                token_pairs.append((start_pos, end_pos))
+
+        # Extract hidden states between token pairs
+        frame_embeddings = []
+        for start_pos, end_pos in token_pairs:
+            # Extract hidden states between the token pair (exclusive of the tokens themselves)
+            # Add 1 to start_pos to exclude the start token, end_pos is exclusive
+            frame_tokens = hidden_state[start_pos + 1 : end_pos]
+
+            if frame_tokens.shape[0] == 0:
+                # If no tokens between the pair, use the token positions themselves
+                # This shouldn't happen normally, but handle it gracefully
+                frame_embedding = (hidden_state[start_pos] + hidden_state[end_pos]) / 2.0
+            else:
+                if self.frame_pooling == "mean":
+                    frame_embedding = frame_tokens.mean(dim=0)  # [hidden_dim]
+                elif self.frame_pooling == "boundary":
+                    # Use the last patch token as a boundary summary of the frame span.
+                    frame_embedding = frame_tokens[-1]
+                elif self.frame_pooling == "attention":
+                    # Learned attention pooling over patch tokens in this frame span.
+                    # scores: [num_tokens]
+                    scores = self.frame_pool_attn(frame_tokens).squeeze(-1) / self.frame_pooling_attn_temperature
+                    weights = torch.softmax(scores, dim=0).unsqueeze(-1)  # [num_tokens, 1]
+                    frame_embedding = (weights * frame_tokens).sum(dim=0)
+                else:
+                    raise ValueError(f"Unsupported frame_pooling: {self.frame_pooling}")
+
+            frame_embeddings.append(frame_embedding)
+
+        if len(frame_embeddings) == 0:
+            return torch.empty(0, hidden_state.shape[-1], device=hidden_state.device, dtype=hidden_state.dtype)
+
+        return torch.stack(frame_embeddings)  # [num_frames, hidden_dim]
+
+    def _extract_progress_from_trajectory(
+        self,
+        hidden_state: torch.Tensor,
+        start_position: int,
+        video_grid_thw: list[int],  # [T, H, W]
+        merge_size: int,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Extract progress and success predictions from a trajectory's hidden states.
+
+        Args:
+            hidden_state: Hidden states tensor [seq_len, hidden_dim]
+            start_position: Starting position in the sequence for this trajectory
+            video_grid_thw: Video grid dimensions [T, H, W] where T is number of temporal patch groups,
+                           H and W are spatial grid dimensions
+            merge_size: Merge size for patch grouping
+
+        Returns:
+            tuple: (progress_logits [T], success_logits [T])
+        """
+        T, H, W = video_grid_thw
+
+        if T == 0:
+            return torch.empty(0, device=hidden_state.device), torch.empty(0, device=hidden_state.device)
+
+        # Calculate tokens per frame: (H * W) // merge_size^2
+        tokens_per_frame = (H * W) // (merge_size**2)
+
+        if self.average_temporal_patches:
+            # Average all tokens within each temporal patch group
+            temporal_patch_tokens = []
+            current_pos = start_position
+            for t_idx in range(T):
+                start_idx = current_pos
+                end_idx = current_pos + tokens_per_frame
+                patch_tokens = hidden_state[start_idx:end_idx]  # [tokens_per_frame, hidden_dim]
+                patch_embedding = patch_tokens.mean(dim=0)  # [hidden_dim] - averaged
+                temporal_patch_tokens.append(patch_embedding)
+                current_pos = end_idx
+            boundary_hidden_states = torch.stack(temporal_patch_tokens)  # [T, hidden_dim]
+        else:
+            # Use last token (boundary) of each temporal patch group
+            frame_boundary_positions = []
+            current_pos = start_position
+            for _frame_idx in range(T):
+                frame_end = current_pos + tokens_per_frame
+                frame_boundary_positions.append(frame_end)
+                current_pos = frame_end
+
+            trajectory_boundaries = torch.tensor(frame_boundary_positions, device=hidden_state.device)
+            boundary_hidden_states = hidden_state[trajectory_boundaries]  # [T, hidden_dim]
+
+        assert boundary_hidden_states.shape[0] == T, f"Expected {T} frames, got {boundary_hidden_states.shape[0]}"
+        progress_output = self.progress_head(boundary_hidden_states)  # [T, 1] or [T, num_bins] for discrete
+        if self.use_discrete_progress:
+            progress = progress_output  # [T, num_bins] - keep logits
+        else:
+            progress = squeeze_last_safe(progress_output)  # [T]
+        success = squeeze_last_safe(self.success_head(boundary_hidden_states))  # [T]
+
+        return progress, success
+
+    def _extract_hidden_state_from_token(
+        self,
+        hidden_state: torch.Tensor,
+        input_ids: torch.Tensor,
+        token_name: str,
+    ) -> torch.Tensor | list[torch.Tensor]:
+        """
+        Extract hidden states at specific token positions.
+
+        Args:
+            hidden_state: Hidden states tensor [B, seq_len, hidden_dim] or [seq_len, hidden_dim]
+            input_ids: Input token IDs [B, seq_len] or [seq_len]
+            token_name: Name of the token to find (e.g., "<|prog_token|>", "<|pref_token|>")
+
+        Returns:
+            If exactly one token per sequence: tensor [B, hidden_dim]
+            If multiple tokens per sequence (variable counts): list of tensors, one per batch item [num_tokens_i, hidden_dim]
+        """
+        # Handle both batched and unbatched inputs
+        is_batched = hidden_state.dim() == 3
+        if not is_batched:
+            hidden_state = hidden_state.unsqueeze(0)  # [1, seq_len, hidden_dim]
+            input_ids = input_ids.unsqueeze(0)  # [1, seq_len]
+
+        B = input_ids.shape[0]
+
+        # Get tokenizer (works for both SmolVLM and Qwen)
+        if "SmolVLM" in self.base_model_id:
+            tokenizer = self.tokenizer
+        else:
+            tokenizer = self.processor.tokenizer
+
+        # Get token ID
+        token_id = tokenizer.convert_tokens_to_ids(token_name)
+
+        # Find all token positions across batch - vectorized
+        token_mask = input_ids == token_id  # [B, seq_len]
+        batch_indices, positions = token_mask.nonzero(as_tuple=True)  # both [total_tokens]
+
+        if len(positions) == 0:
+            raise ValueError(f"{token_name} not found in any sequence")
+
+        # Extract all hidden states at token positions at once - vectorized
+        all_hidden = hidden_state[batch_indices, positions]  # [total_tokens, hidden_dim]
+
+        # Count tokens per batch sample
+        token_counts = torch.bincount(batch_indices, minlength=B)  # [B]
+
+        # Check if all sequences have exactly one token
+        if (token_counts == 1).all():
+            # Return as [B, hidden_dim] tensor
+            return all_hidden.reshape(B, -1)
+        else:
+            # Split by batch index into list of tensors (variable token counts)
+            return list(torch.split(all_hidden, token_counts.tolist()))
+
+    def _forward_smolvlm(
+        self,
+        input_ids,
+        attention_mask,
+        pixel_values,
+        sample_type,
+        timing_raw,
+        **kwargs,
+    ):
+        """Forward pass for SmolVLM model. Returns (ModelOutput, timing_raw)."""
+        model_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            **kwargs,
+        }
+        with _timer("time/rbm_forward", timing_raw=timing_raw):
+            outputs = self.model(**model_kwargs, output_hidden_states=True, return_dict=True)
+
+        hidden_state = outputs.hidden_states[-1]  # [B, seq_len, hidden_dim]
+        progress_logits = {"A": None, "B": None}
+        success_logits = {"A": None, "B": None}
+
+        # Create output
+        output = ModelOutput()
+
+        # Handle token-based extraction first if using per-frame progress tokens
+        if self.use_per_frame_progress_token:
+            progress_logits, success_logits, pref_or_sim_logits = self._process_token_extraction(
+                hidden_state, input_ids, sample_type
+            )
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+            if pref_or_sim_logits is not None:
+                output.pref_logits = pref_or_sim_logits
+        else:
+            # Process frames normally
+            with _timer("time/progress_logits", timing_raw=timing_raw):
+                progress_logits, success_logits = self._process_smolvlm_frames(hidden_state, input_ids, sample_type)
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+
+            # Handle preference token if needed
+            if sample_type == "preference":
+                token_hidden = self._extract_hidden_state_from_token(hidden_state, input_ids, "<|pref_token|>")
+                output.pref_logits = self.preference_head(token_hidden)
+
+        return output, timing_raw
+
+    def _process_smolvlm_frames(self, hidden_state, input_ids, sample_type):
+        """Process SmolVLM frames and return progress/success logits."""
+        B = input_ids.shape[0]
+        progress_logits_A, progress_logits_B = [], []
+        success_logits_A, success_logits_B = [], []
+
+        for i in range(B):
+            frame_embeddings = self._extract_hidden_states_from_token_pairs(hidden_state[i], input_ids[i])
+            if frame_embeddings.shape[0] == 0:
+                raise ValueError(f"No frame embeddings extracted for sample {i}")
+
+            if sample_type == "progress":
+                traj_A, traj_B = frame_embeddings, None
+            else:
+                mid = frame_embeddings.shape[0] // 2
+                traj_A, traj_B = frame_embeddings[:mid], frame_embeddings[mid:]
+
+            # Trajectory A
+            prog_A = self.progress_head(traj_A)
+            progress_logits_A.append(prog_A if self.use_discrete_progress else squeeze_last_safe(prog_A))
+            success_logits_A.append(squeeze_last_safe(self.success_head(traj_A)))
+
+            # Trajectory B
+            if traj_B is not None:
+                prog_B = self.progress_head(traj_B)
+                progress_logits_B.append(prog_B if self.use_discrete_progress else squeeze_last_safe(prog_B))
+                success_logits_B.append(squeeze_last_safe(self.success_head(traj_B)))
+            else:
+                progress_logits_B.append(None)
+                success_logits_B.append(None)
+
+        return (
+            {
+                "A": torch.stack(progress_logits_A) if progress_logits_A else None,
+                "B": torch.stack(progress_logits_B) if progress_logits_B and progress_logits_B[0] is not None else None,
+            },
+            {
+                "A": torch.stack(success_logits_A) if success_logits_A else None,
+                "B": torch.stack(success_logits_B) if success_logits_B and success_logits_B[0] is not None else None,
+            },
+        )
+
+    def _forward_qwen(
+        self,
+        input_ids,
+        attention_mask,
+        pixel_values,
+        pixel_values_videos,
+        image_grid_thw,
+        video_grid_thw,
+        sample_type,
+        timing_raw,
+        second_per_grid_ts=None,
+        **kwargs,
+    ):
+        """Forward pass for Qwen2.5/Qwen3 models. Returns (ModelOutput, timing_raw)."""
+        model_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            "pixel_values_videos": pixel_values_videos,
+            "image_grid_thw": image_grid_thw,
+            "video_grid_thw": video_grid_thw,
+            "second_per_grid_ts": second_per_grid_ts,
+            **kwargs,
+        }
+        with _timer("time/rbm_forward", timing_raw=timing_raw):
+            # Qwen3 models may need output_hidden_states=True and use hidden_states instead of last_hidden_state
+            is_qwen3 = "Qwen3" in self.base_model_id or (
+                hasattr(self.model, "config") and "Qwen3" in str(type(self.model))
+            )
+            if is_qwen3:
+                outputs = self.model(**model_kwargs, output_hidden_states=True, return_dict=True)
+                # Qwen3 uses hidden_states tuple, take the last layer
+                hidden_state = (
+                    outputs.hidden_states[-1] if hasattr(outputs, "hidden_states") else outputs.last_hidden_state
+                )
+            else:
+                outputs = self.model(**model_kwargs)
+                hidden_state = outputs.last_hidden_state  # [B, seq_len, hidden_dim]
+
+        progress_logits = {"A": None, "B": None}
+        success_logits = {"A": None, "B": None}
+
+        # Create output
+        output = ModelOutput()
+
+        if self.use_per_frame_progress_token:
+            progress_logits, success_logits, pref_or_sim_logits = self._process_token_extraction(
+                hidden_state, input_ids, sample_type
+            )
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+            if pref_or_sim_logits is not None:
+                output.pref_logits = pref_or_sim_logits
+        else:
+            # Process frames normally
+            vision_start_token_id = self.processor.tokenizer.convert_tokens_to_ids("<|vision_start|>")
+            vision_end_token_id = self.processor.tokenizer.convert_tokens_to_ids("<|vision_end|>")
+            split_token_id = self.processor.tokenizer.convert_tokens_to_ids("<|split_token|>")
+
+            tps = getattr(getattr(self.processor, "video_processor", None), "temporal_patch_size", 2)
+            merge_size = getattr(getattr(self.processor, "video_processor", None), "merge_size", 14)
+
+            if self.use_multi_image:
+                progress_logits, success_logits = self._process_multi_image_frames(
+                    hidden_state,
+                    input_ids,
+                    sample_type,
+                    vision_start_token_id,
+                    vision_end_token_id,
+                    split_token_id,
+                    timing_raw,
+                )
+            else:
+                progress_logits, success_logits = self._process_video_frames(
+                    hidden_state,
+                    input_ids,
+                    video_grid_thw,
+                    sample_type,
+                    vision_start_token_id,
+                    split_token_id,
+                    tps,
+                    merge_size,
+                    timing_raw,
+                )
+
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+
+            # Handle preference token if needed
+            if sample_type == "preference":
+                token_hidden = self._extract_hidden_state_from_token(hidden_state, input_ids, "<|pref_token|>")
+                output.pref_logits = self.preference_head(token_hidden)
+
+        return output, timing_raw
+
+    def _forward_molmo(
+        self,
+        input_ids,
+        attention_mask,
+        pixel_values,
+        image_grid_thw,
+        sample_type,
+        timing_raw,
+        image_grids=None,
+        image_token_pooling=None,
+        image_num_crops=None,
+        **kwargs,
+    ):
+        """Forward pass for Molmo2 models (multi-image only). Returns (ModelOutput, timing_raw)."""
+        model_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            "image_grid_thw": image_grid_thw,
+            "image_grids": image_grids,
+            "image_token_pooling": image_token_pooling,
+            "image_num_crops": image_num_crops,
+            **kwargs,
+        }
+        with _timer("time/rbm_forward", timing_raw=timing_raw):
+            # Qwen3 models may need output_hidden_states=True and use hidden_states instead of last_hidden_state
+            is_qwen3 = "Qwen3" in self.base_model_id or (
+                hasattr(self.model, "config") and "Qwen3" in str(type(self.model))
+            )
+            if is_qwen3:
+                outputs = self.model(**model_kwargs, output_hidden_states=True, return_dict=True)
+                # Qwen3 uses hidden_states tuple, take the last layer
+                hidden_state = (
+                    outputs.hidden_states[-1] if hasattr(outputs, "hidden_states") else outputs.last_hidden_state
+                )
+            else:
+                outputs = self.model(**model_kwargs)
+                hidden_state = outputs.last_hidden_state  # [B, seq_len, hidden_dim]
+
+        progress_logits = {"A": None, "B": None}
+        success_logits = {"A": None, "B": None}
+
+        # Create output
+        output = ModelOutput()
+
+        # Handle token-based extraction first if using per-frame progress tokens
+        if self.use_per_frame_progress_token:
+            progress_logits, success_logits, pref_or_sim_logits = self._process_token_extraction(
+                hidden_state, input_ids, sample_type
+            )
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+            if pref_or_sim_logits is not None:
+                output.pref_logits = pref_or_sim_logits
+        else:
+            # Process frames normally
+            vision_start_token_id = self.processor.tokenizer.convert_tokens_to_ids("<low_res_im_start>")
+            split_token_id = self.processor.tokenizer.convert_tokens_to_ids("<|split_token|>")
+
+            progress_logits, success_logits = self._process_multi_image_frames(
+                hidden_state, input_ids, sample_type, vision_start_token_id, None, split_token_id, timing_raw
+            )
+
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+
+            # Handle preference token if needed
+            if sample_type == "preference":
+                token_hidden = self._extract_hidden_state_from_token(hidden_state, input_ids, "<|pref_token|>")
+                output.pref_logits = self.preference_head(token_hidden)
+
+        return output, timing_raw
+
+    def _process_multi_image_frames(
+        self,
+        hidden_state,
+        input_ids,
+        sample_type,
+        vision_start_token_id,
+        vision_end_token_id,
+        split_token_id,
+        timing_raw,
+    ):
+        """Process frames in multi-image mode (shared by Qwen and Molmo)."""
+        progress_logits_A = []
+        progress_logits_B = []
+        success_logits_A = []
+        success_logits_B = []
+
+        all_trajectory_A_frames = []
+        all_trajectory_B_frames = []
+        trajectory_A_lengths = []
+        trajectory_B_lengths = []
+        has_trajectory_B = sample_type != "progress"
+
+        with _timer("time/progress_logits", timing_raw=timing_raw):
+            # First pass: extract all frame embeddings
+            for i, seq_ids in enumerate(input_ids):
+                vision_start_positions = (seq_ids == vision_start_token_id).nonzero(as_tuple=True)[0]
+                if len(vision_start_positions) == 0:
+                    raise ValueError(f"vision_start_token (id={vision_start_token_id}) not found in sequence {i}")
+
+                # Extract embeddings using _extract_hidden_states_from_token_pairs
+                # (handles both Qwen and Molmo token patterns)
+                frame_embeddings = self._extract_hidden_states_from_token_pairs(hidden_state[i], seq_ids)
+
+                if frame_embeddings.shape[0] == 0:
+                    raise ValueError(f"No frame embeddings extracted for sample {i}")
+
+                # Split into trajectories
+                if sample_type == "progress":
+                    trajectory_A_frames = frame_embeddings
+                    trajectory_B_frames = None
+                else:
+                    split_positions = (seq_ids == split_token_id).nonzero(as_tuple=True)[0]
+                    if len(split_positions) == 0:
+                        raise ValueError(f"split_token not found in sequence {i}")
+                    split_pos = split_positions[0].item()
+                    traj_a_count = sum(1 for pos in vision_start_positions if pos.item() < split_pos)
+                    trajectory_A_frames = frame_embeddings[:traj_a_count]
+                    trajectory_B_frames = frame_embeddings[traj_a_count:]
+
+                all_trajectory_A_frames.append(trajectory_A_frames)
+                trajectory_A_lengths.append(trajectory_A_frames.shape[0])
+
+                if trajectory_B_frames is not None:
+                    all_trajectory_B_frames.append(trajectory_B_frames)
+                    trajectory_B_lengths.append(trajectory_B_frames.shape[0])
+                else:
+                    all_trajectory_B_frames.append(None)
+                    trajectory_B_lengths.append(0)
+
+            # Batch process trajectory A
+            if all_trajectory_A_frames:
+                batched_A = torch.cat(all_trajectory_A_frames, dim=0)
+                progress_A_out = self.progress_head(batched_A)
+                success_A_out = squeeze_last_safe(self.success_head(batched_A))
+
+                if self.use_discrete_progress:
+                    progress_A_split = torch.split(progress_A_out, trajectory_A_lengths, dim=0)
+                else:
+                    progress_A_split = torch.split(squeeze_last_safe(progress_A_out), trajectory_A_lengths, dim=0)
+                success_A_split = torch.split(success_A_out, trajectory_A_lengths, dim=0)
+
+                for prog, succ in zip(progress_A_split, success_A_split):
+                    progress_logits_A.append(prog)
+                    success_logits_A.append(succ)
+
+            # Batch process trajectory B
+            if has_trajectory_B:
+                valid_B = [(f, l) for f, l in zip(all_trajectory_B_frames, trajectory_B_lengths) if f is not None]
+                if valid_B:
+                    valid_B_frames, valid_B_lengths = zip(*valid_B)
+                    batched_B = torch.cat(valid_B_frames, dim=0)
+                    progress_B_out = self.progress_head(batched_B)
+                    success_B_out = squeeze_last_safe(self.success_head(batched_B))
+
+                    if self.use_discrete_progress:
+                        progress_B_split = torch.split(progress_B_out, list(valid_B_lengths), dim=0)
+                    else:
+                        progress_B_split = torch.split(squeeze_last_safe(progress_B_out), list(valid_B_lengths), dim=0)
+                    success_B_split = torch.split(success_B_out, list(valid_B_lengths), dim=0)
+
+                    valid_idx = 0
+                    for frame in all_trajectory_B_frames:
+                        if frame is not None:
+                            progress_logits_B.append(progress_B_split[valid_idx])
+                            success_logits_B.append(success_B_split[valid_idx])
+                            valid_idx += 1
+                        else:
+                            progress_logits_B.append(None)
+                            success_logits_B.append(None)
+                else:
+                    progress_logits_B = [None] * len(all_trajectory_A_frames)
+                    success_logits_B = [None] * len(all_trajectory_A_frames)
+            else:
+                progress_logits_B = [None] * len(all_trajectory_A_frames)
+                success_logits_B = [None] * len(all_trajectory_A_frames)
+
+        progress_logits = {
+            "A": torch.stack(progress_logits_A) if progress_logits_A else None,
+            "B": torch.stack(progress_logits_B) if progress_logits_B and progress_logits_B[0] is not None else None,
+        }
+        success_logits = {
+            "A": torch.stack(success_logits_A) if success_logits_A else None,
+            "B": torch.stack(success_logits_B) if success_logits_B and success_logits_B[0] is not None else None,
+        }
+        return progress_logits, success_logits
+
+    def _process_video_frames(
+        self,
+        hidden_state,
+        input_ids,
+        video_grid_thw,
+        sample_type,
+        vision_start_token_id,
+        split_token_id,
+        tps,
+        merge_size,
+        timing_raw,
+    ):
+        """Process frames in video mode (Qwen only). Returns (progress_logits, success_logits)."""
+        progress_logits_A = []
+        progress_logits_B = []
+        success_logits_A = []
+        success_logits_B = []
+
+        with _timer("time/progress_logits", timing_raw=timing_raw):
+            for i, seq_ids in enumerate(input_ids):
+                vision_start_positions = (seq_ids == vision_start_token_id).nonzero(as_tuple=True)[0]
+                if len(vision_start_positions) == 0:
+                    raise ValueError(f"vision_start_token not found in sequence {i}")
+
+                if video_grid_thw is None or i >= len(video_grid_thw):
+                    raise ValueError(f"video_grid_thw required for video mode")
+
+                # Trajectory A
+                grid_idx_A = i if sample_type == "progress" else i * tps
+                progress_A, success_A = self._extract_progress_from_trajectory(
+                    hidden_state[i], vision_start_positions[0].item(), video_grid_thw[grid_idx_A], merge_size
+                )
+                progress_logits_A.append(progress_A)
+                success_logits_A.append(success_A)
+
+                # Trajectory B (if not progress sample)
+                if sample_type != "progress":
+                    grid_idx_B = i * tps + 1
+                    if grid_idx_B >= len(video_grid_thw):
+                        raise ValueError(f"video_grid_thw index {grid_idx_B} out of bounds")
+                    progress_B, success_B = self._extract_progress_from_trajectory(
+                        hidden_state[i], vision_start_positions[1].item(), video_grid_thw[grid_idx_B], merge_size
+                    )
+                    progress_logits_B.append(progress_B)
+                    success_logits_B.append(success_B)
+                else:
+                    progress_logits_B.append(None)
+                    success_logits_B.append(None)
+
+        progress_logits = {
+            "A": torch.stack(progress_logits_A) if progress_logits_A else None,
+            "B": torch.stack(progress_logits_B) if progress_logits_B[0] is not None else None,
+        }
+        success_logits = {
+            "A": torch.stack(success_logits_A) if success_logits_A else None,
+            "B": torch.stack(success_logits_B) if success_logits_B[0] is not None else None,
+        }
+        return progress_logits, success_logits
+
+    def _apply_heads_to_hidden_states(
+        self, hidden_states_list: list[torch.Tensor]
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply progress and success heads to a list of hidden states.
+
+        Args:
+            hidden_states_list: List of tensors, one per batch item [num_tokens_i, hidden_dim]
+
+        Returns:
+            tuple: (progress_logits, success_logits)
+                - progress_logits: [B, max_tokens] or [B, max_tokens, num_bins] for discrete (padded)
+                - success_logits: [B, max_tokens] (padded)
+        """
+        progress_list = []
+        success_list = []
+        for hidden in hidden_states_list:
+            if hidden.shape[0] > 0:
+                progress_output = self.progress_head(hidden)
+                if self.use_discrete_progress:
+                    progress_list.append(progress_output)
+                else:
+                    progress_list.append(squeeze_last_safe(progress_output))
+                success_list.append(squeeze_last_safe(self.success_head(hidden)))
+            else:
+                progress_list.append(torch.empty(0, device=hidden.device))
+                success_list.append(torch.empty(0, device=hidden.device))
+
+        progress = torch.stack(progress_list) if progress_list else None
+        success = torch.stack(success_list) if success_list else None
+        return progress, success
+
+    def _process_token_extraction(
+        self,
+        hidden_state: torch.Tensor,
+        input_ids: torch.Tensor,
+        sample_type: str,
+    ) -> tuple[dict, dict, torch.Tensor | None]:
+        """
+        Process token-based extraction for progress and preference predictions.
+
+        Returns:
+            tuple: (progress_logits, success_logits, pref_logits)
+                - progress_logits: dict with "A" and/or "B" keys
+                - success_logits: dict with "A" and/or "B" keys
+                - pref_logits: preference logits, or None
+        """
+        progress_logits = {"A": None, "B": None}
+        success_logits = {"A": None, "B": None}
+        pref_or_sim_logits = None
+
+        # Extract all <|prog_token|> hidden states (returns list of tensors when multiple tokens)
+        all_prog_token_hidden = self._extract_hidden_state_from_token(hidden_state, input_ids, "<|prog_token|>")
+
+        if sample_type == "progress":
+            # For progress samples, all tokens belong to trajectory A
+            hidden_states_A = all_prog_token_hidden
+            progress_logits["A"], success_logits["A"] = self._apply_heads_to_hidden_states(hidden_states_A)
+        elif sample_type == "preference":
+            # For preference, assume equal number of tokens for A and B
+            hidden_states_A = []
+            hidden_states_B = []
+            for tokens in all_prog_token_hidden:
+                num_tokens = tokens.shape[0]
+                if num_tokens % 2 != 0:
+                    raise ValueError(f"Expected even number of <|prog_token|> tokens, got {num_tokens}")
+                mid = num_tokens // 2
+                hidden_states_A.append(tokens[:mid])
+                hidden_states_B.append(tokens[mid:])
+
+            progress_logits["A"], success_logits["A"] = self._apply_heads_to_hidden_states(hidden_states_A)
+            progress_logits["B"], success_logits["B"] = self._apply_heads_to_hidden_states(hidden_states_B)
+
+            token_hidden = self._extract_hidden_state_from_token(hidden_state, input_ids, "<|pref_token|>")
+            pref_or_sim_logits = self.preference_head(token_hidden)
+
+        return progress_logits, success_logits, pref_or_sim_logits
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        pixel_values=None,
+        pixel_values_videos=None,
+        image_grid_thw=None,
+        video_grid_thw=None,
+        sample_type=None,  # "preference", "progress"
+        second_per_grid_ts=None,
+        timing_raw=None,
+        # Molmo2-specific parameters
+        image_grids=None,
+        image_token_pooling=None,
+        image_num_crops=None,
+        video_grids=None,
+        video_token_pooling=None,
+        **kwargs,
+    ):
+        """
+        Forward pass for the RBM (Robometer).
+
+        Dispatches to model-specific forward methods:
+        - SmolVLM: _forward_smolvlm
+        - Qwen2.5/Qwen3: _forward_qwen
+        - Molmo2: _forward_molmo
+
+        Returns:
+            tuple: (ModelOutput, timing_raw dict)
+        """
+        if timing_raw is None:
+            timing_raw = {}
+
+        # Dispatch to model-specific forward
+        if "SmolVLM" in self.base_model_id:
+            return self._forward_smolvlm(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                pixel_values=pixel_values,
+                sample_type=sample_type,
+                timing_raw=timing_raw,
+                **kwargs,
+            )
+        elif "Molmo" in self.base_model_id:
+            return self._forward_molmo(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                pixel_values=pixel_values,
+                image_grid_thw=image_grid_thw,
+                sample_type=sample_type,
+                timing_raw=timing_raw,
+                image_grids=image_grids,
+                image_token_pooling=image_token_pooling,
+                image_num_crops=image_num_crops,
+                **kwargs,
+            )
+        else:
+            # Qwen2.5 / Qwen3
+            return self._forward_qwen(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                pixel_values=pixel_values,
+                pixel_values_videos=pixel_values_videos,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                sample_type=sample_type,
+                timing_raw=timing_raw,
+                second_per_grid_ts=second_per_grid_ts,
+                **kwargs,
+            )
diff --git a/robometer/robometer/models/rewind_transformer.py b/robometer/robometer/models/rewind_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..60f4fc1a7c5916290f2d7bacea4d942dd545265f
--- /dev/null
+++ b/robometer/robometer/models/rewind_transformer.py
@@ -0,0 +1,396 @@
+#!/usr/bin/env python3
+"""
+ReWiND Transformer implementation.
+Contains the ReWINDTransformer class with three prediction heads for different objectives.
+
+Note: make sure that the forward pass uses all of the
+heads or there will be some problems with FSDP sharding.
+"""
+
+import einops
+import torch
+import torch.nn as nn
+from transformers import PreTrainedModel, AutoConfig, AutoModel
+from transformers import PretrainedConfig
+from robometer.models.utils import ModelOutput
+from robometer.models.heads import PredictionHeadsMixin
+
+
+def mean_pooling(model_output, attention_mask):
+    token_embeddings = model_output[0]  # First element of model_output contains all token embeddings
+    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
+
+
+class ReWINDTransformerConfig(PretrainedConfig):
+    model_type = "rewind_transformer"
+
+    def __init__(
+        self,
+        video_feature_dim: int = 768,
+        text_feature_dim: int = 384,
+        hidden_dim: int = 512,
+        num_layers: int = 4,
+        num_attention_heads: int = 8,
+        dropout: float = 0.1,
+        max_len: int = 16,
+        causal_mask: bool = False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.video_feature_dim = video_feature_dim
+        self.text_feature_dim = text_feature_dim
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.dropout = dropout
+        self.max_len = max_len
+        self.causal_mask = causal_mask
+
+
+class ReWiNDTransformer(PredictionHeadsMixin, PreTrainedModel):
+    """ReWiND Transformer with three prediction heads for different objectives."""
+
+    config_class = ReWINDTransformerConfig
+
+    def __init__(self, config, processor=None, tokenizer=None, image_encoder=None, text_encoder=None):
+        rewind_config = config.rewind
+
+        video_feature_dim = rewind_config.video_feature_dim
+        text_feature_dim = rewind_config.text_feature_dim
+
+        if image_encoder is not None:
+            video_feature_dim = image_encoder.config.hidden_size
+        if text_encoder is not None:
+            text_feature_dim = text_encoder.config.hidden_size
+
+        super().__init__(
+            config=config.rewind,
+            model_config=config,
+            hidden_dim=rewind_config.hidden_dim,
+            dropout=rewind_config.dropout,
+        )
+
+        self.image_encoder = image_encoder
+        self.text_encoder = text_encoder
+        self.tokenizer = tokenizer
+        self.processor = processor
+
+        self.video_proj = nn.Linear(video_feature_dim, rewind_config.hidden_dim)
+        self.text_proj = nn.Linear(text_feature_dim, rewind_config.hidden_dim)
+
+        self.first_embedding_A = nn.Parameter(torch.randn(1, 1, rewind_config.hidden_dim))
+        self.first_embedding_B = nn.Parameter(torch.randn(1, 1, rewind_config.hidden_dim))
+
+        encoder_layer = nn.TransformerEncoderLayer(
+            d_model=rewind_config.hidden_dim,
+            nhead=rewind_config.num_attention_heads,
+            dim_feedforward=rewind_config.hidden_dim * 4,
+            dropout=rewind_config.dropout,
+            batch_first=True,
+        )
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=rewind_config.num_layers)
+
+        if rewind_config.use_per_frame_progress_token:
+            self.text_position_embedding = nn.Parameter(torch.randn(1, 1, rewind_config.hidden_dim))
+            self.prog_token_A = nn.Parameter(torch.randn(1, rewind_config.max_len, rewind_config.hidden_dim))
+            self.prog_token_B = nn.Parameter(torch.randn(1, rewind_config.max_len, rewind_config.hidden_dim))
+
+        # Prediction token for preference
+        self.preference_token = nn.Parameter(torch.randn(1, 1, rewind_config.hidden_dim))
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        pixel_values_videos=None,
+        video_embeddings=None,
+        text_embeddings=None,
+        sample_type=None,  # "preference", "progress"
+        timing_raw=None,
+        **kwargs,
+    ):
+        """Forward pass for ReWiND Transformer."""
+        if timing_raw is None:
+            timing_raw = {}
+
+        use_precomputed = video_embeddings is not None and text_embeddings is not None
+
+        if use_precomputed:
+            B, T, D_video = video_embeddings.shape
+            D_text = text_embeddings.shape[1]
+
+            # Project embeddings to hidden dimension
+            video_embeddings = self.video_proj(video_embeddings.view(-1, D_video)).view(B, T, -1)  # [B, T, hidden_dim]
+            text_embeddings = self.text_proj(text_embeddings)  # [B, hidden_dim]
+        else:
+            # Use raw inputs with encoders
+            if pixel_values_videos is None:
+                raise ValueError("pixel_values_videos is required when not using precomputed embeddings")
+
+            B, T, C, H, W = pixel_values_videos.shape
+
+            # processing text inputs
+            with torch.no_grad():
+                text_embeddings = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask)
+                text_embeddings = mean_pooling(text_embeddings, attention_mask)  # [B, text_hidden_dim]
+                text_embeddings = self.text_proj(text_embeddings)  # [B, D]
+
+            # processing video inputs
+            # T should contain both chosen and rejected trajectories concatenated together
+            pixel_values_videos = pixel_values_videos.view(B * T, C, H, W)
+            video_embeddings = self.image_encoder(
+                pixel_values=pixel_values_videos
+            ).pooler_output  # [B, vision_hidden_dim]
+            video_embeddings = self.video_proj(video_embeddings)  # [B * T, D]
+            video_embeddings = video_embeddings.view(B, T, -1)  # [B, T, D]
+
+        output = ModelOutput()
+
+        if sample_type == "preference":
+            half = T // 2
+            video_embeddings_A = video_embeddings[:, :half].clone()
+            video_embeddings_B = video_embeddings[:, half:].clone()
+
+            # Add the first embedding to the beginning of each sequence
+            first_frame_emb_A = einops.repeat(self.first_embedding_A, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+            first_frame_emb_B = einops.repeat(self.first_embedding_B, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+            video_embeddings_A[:, 0:1] += first_frame_emb_A
+            video_embeddings_B[:, 0:1] += first_frame_emb_B
+
+            if self.config.use_per_frame_progress_token:
+                # Add position embedding to text
+                text_pos = einops.repeat(self.text_position_embedding, "1 1 d -> b 1 d", b=B)
+                text_emb = text_embeddings.unsqueeze(1) + text_pos
+
+                # Get progress tokens - slice to match sequence length
+                prog_token_A = einops.repeat(self.prog_token_A[:, :half, :], "1 t d -> b t d", b=B)  # [B, half, D]
+                prog_token_B = einops.repeat(self.prog_token_B[:, :half, :], "1 t d -> b t d", b=B)  # [B, half, D]
+
+                # Interleave frames with progress tokens: [frame, prog_token, frame, prog_token, ...]
+                sequence_A = []
+                for i in range(half):
+                    sequence_A.append(video_embeddings_A[:, i : i + 1, :])  # [B, 1, D]
+                    sequence_A.append(prog_token_A[:, i : i + 1, :])  # [B, 1, D]
+
+                sequence_B = []
+                for i in range(half):
+                    sequence_B.append(video_embeddings_B[:, i : i + 1, :])  # [B, 1, D]
+                    sequence_B.append(prog_token_B[:, i : i + 1, :])  # [B, 1, D]
+
+                # Prediction token
+                pred_token = einops.repeat(self.preference_token, "1 1 d -> b 1 d", b=B)
+
+                # Concatenate: [text, frame1_A, prog_A, frame2_A, prog_A, ..., frame1_B, prog_B, frame2_B, prog_B, ..., pred]
+                full_sequence = torch.cat([text_emb] + sequence_A + sequence_B + [pred_token], dim=1)  # [B, L, D]
+
+                # Build causal mask if enabled
+                mask = None
+                if self.config.causal_mask:
+                    L = full_sequence.shape[1]
+                    mask = torch.triu(torch.ones((L, L), dtype=torch.bool, device=full_sequence.device), diagonal=1)
+
+                full_embeddings = self.transformer(full_sequence, src_key_padding_mask=None, mask=mask)
+
+                # Extract progress token embeddings
+                D = full_embeddings.shape[-1]
+                prog_embeddings_A = []
+                prog_embeddings_B = []
+
+                idx_A_start = 1
+                idx_B_start = 1 + 2 * half
+                idx_pred = 1 + 4 * half
+
+                for i in range(half):
+                    prog_idx_A = idx_A_start + 2 * i + 1
+                    prog_embeddings_A.append(full_embeddings[:, prog_idx_A, :])  # [B, D]
+
+                    prog_idx_B = idx_B_start + 2 * i + 1
+                    prog_embeddings_B.append(full_embeddings[:, prog_idx_B, :])  # [B, D]
+
+                prog_embeddings_A = torch.stack(prog_embeddings_A, dim=1)  # [B, half, D]
+                prog_embeddings_B = torch.stack(prog_embeddings_B, dim=1)  # [B, half, D]
+
+                # Apply heads
+                progress_A_logits = self.progress_head(prog_embeddings_A.reshape(-1, D))
+                if self.use_discrete_progress:
+                    # Discrete: [b*t, num_bins] -> [b, t, num_bins]
+                    progress_A_logits = einops.rearrange(progress_A_logits, "(b t) ... -> b t ...", b=B, t=half)
+                else:
+                    # Continuous: [b*t, 1] -> [b, t]
+                    progress_A_logits = einops.rearrange(progress_A_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                progress_B_logits = self.progress_head(prog_embeddings_B.reshape(-1, D))
+                if self.use_discrete_progress:
+                    # Discrete: [b*t, num_bins] -> [b, t, num_bins]
+                    progress_B_logits = einops.rearrange(progress_B_logits, "(b t) ... -> b t ...", b=B, t=half)
+                else:
+                    # Continuous: [b*t, 1] -> [b, t]
+                    progress_B_logits = einops.rearrange(progress_B_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                progress_logits = {"A": progress_A_logits, "B": progress_B_logits}
+                output.progress_logits = progress_logits
+
+                success_A_logits = self.success_head(prog_embeddings_A.reshape(-1, D))
+                success_A_logits = einops.rearrange(success_A_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                success_B_logits = self.success_head(prog_embeddings_B.reshape(-1, D))
+                success_B_logits = einops.rearrange(success_B_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                success_logits = {"A": success_A_logits, "B": success_B_logits}
+                output.success_logits = success_logits
+
+                pred_class_token = full_embeddings[:, idx_pred, :]  # [B, D]
+            else:
+                # Regular mode: use frame embeddings directly
+                pred_token = einops.repeat(self.preference_token, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+
+                token_sequence = torch.cat(
+                    [text_embeddings.unsqueeze(1), video_embeddings_A, video_embeddings_B, pred_token], dim=1
+                )  # shape: [B, 2*T + 1, D]
+
+                # Build causal mask if enabled
+                mask = None
+                if self.config.causal_mask:
+                    L = token_sequence.shape[1]
+                    mask = torch.triu(torch.ones((L, L), dtype=torch.bool, device=token_sequence.device), diagonal=1)
+
+                token_embeddings = self.transformer(token_sequence, src_key_padding_mask=None, mask=mask)
+                D = token_embeddings.shape[-1]
+
+                final_embeddings_A = token_embeddings[:, 1 : 1 + half, :]  # avoid the text embedding
+                final_embeddings_B = token_embeddings[:, 1 + half : -1, :]  # avoid the text embedding
+
+                progress_A_logits = self.progress_head(final_embeddings_A.reshape(-1, D))
+                if self.use_discrete_progress:
+                    # Discrete: [b*t, num_bins] -> [b, t, num_bins]
+                    progress_A_logits = einops.rearrange(progress_A_logits, "(b t) ... -> b t ...", b=B, t=half)
+                else:
+                    # Continuous: [b*t, 1] -> [b, t]
+                    progress_A_logits = einops.rearrange(progress_A_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                progress_B_logits = self.progress_head(final_embeddings_B.reshape(-1, D))
+                if self.use_discrete_progress:
+                    # Discrete: [b*t, num_bins] -> [b, t, num_bins]
+                    progress_B_logits = einops.rearrange(progress_B_logits, "(b t) ... -> b t ...", b=B, t=half)
+                else:
+                    # Continuous: [b*t, 1] -> [b, t]
+                    progress_B_logits = einops.rearrange(progress_B_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                progress_logits = {"A": progress_A_logits, "B": progress_B_logits}
+                output.progress_logits = progress_logits
+
+                # Predict success for all frames
+                success_A_logits = self.success_head(final_embeddings_A.reshape(-1, D))
+                success_A_logits = einops.rearrange(success_A_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                success_B_logits = self.success_head(final_embeddings_B.reshape(-1, D))
+                success_B_logits = einops.rearrange(success_B_logits, "(b t) 1 -> b t", b=B, t=half)
+
+                success_logits = {"A": success_A_logits, "B": success_B_logits}
+                output.success_logits = success_logits
+
+                pred_class_token = token_embeddings[:, -1, :]  # [B, D]
+
+            if sample_type == "preference":
+                output.pref_logits = self.preference_head(pred_class_token)
+
+        elif sample_type == "progress":
+            if self.config.use_per_frame_progress_token:
+                # Add position embedding to text
+                text_pos_emb = einops.repeat(self.text_position_embedding, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+                text_emb = text_embeddings.unsqueeze(1).clone()
+                text_emb += text_pos_emb
+
+                # Add the first embedding to the beginning of the sequence
+                first_frame_emb = einops.repeat(self.first_embedding_A, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+                video_embeddings = video_embeddings.clone()
+                video_embeddings[:, 0:1] += first_frame_emb
+
+                # Get progress tokens - slice to match sequence length
+                prog_token_A = einops.repeat(self.prog_token_A[:, :T, :], "1 t d -> b t d", b=B)  # [B, T, D]
+
+                # Interleave frames with progress tokens: [frame, prog_token, frame, prog_token, ...]
+                sequence = []
+                for i in range(T):
+                    sequence.append(video_embeddings[:, i : i + 1, :])  # [B, 1, D]
+                    sequence.append(prog_token_A[:, i : i + 1, :])  # [B, 1, D]
+
+                # Concatenate: [text, frame1, prog_A, frame2, prog_A, ...]
+                token_sequence = torch.cat([text_emb] + sequence, dim=1)  # [B, 2*T + 1, D]
+
+                # Build causal mask if enabled
+                mask = None
+                if self.config.causal_mask:
+                    L = token_sequence.shape[1]
+                    mask = torch.triu(torch.ones((L, L), dtype=torch.bool, device=token_sequence.device), diagonal=1)
+
+                token_embeddings = self.transformer(token_sequence, src_key_padding_mask=None, mask=mask)
+                D = token_embeddings.shape[-1]
+
+                # Extract progress token embeddings
+                prog_embeddings = []
+                for i in range(T):
+                    prog_idx = 1 + 2 * i + 1
+                    prog_embeddings.append(token_embeddings[:, prog_idx, :])  # [B, D]
+
+                prog_embeddings = torch.stack(prog_embeddings, dim=1)  # [B, T, D]
+
+                # Progress prediction for all frames
+                progress_logits = self.progress_head(prog_embeddings.reshape(-1, D))
+                if self.use_discrete_progress:
+                    # Discrete: [b*t, num_bins] -> [b, t, num_bins]
+                    progress_logits = einops.rearrange(progress_logits, "(b t) ... -> b t ...", b=B, t=T)
+                else:
+                    # Continuous: [b*t, 1] -> [b, t]
+                    progress_logits = einops.rearrange(progress_logits, "(b t) 1 -> b t", b=B, t=T)
+
+                # Predict success for all frames
+                success_logits = self.success_head(prog_embeddings.reshape(-1, D))
+                success_logits = einops.rearrange(success_logits, "(b t) 1 -> b t", b=B, t=T)
+            else:
+                # Regular mode: use frame embeddings directly
+                first_frame_emb = einops.repeat(self.first_embedding_A, "1 1 d -> b 1 d", b=B)  # [B, 1, D]
+
+                # [B, T, D]
+                video_embeddings = video_embeddings.clone()
+                video_embeddings[:, 0:1] += first_frame_emb
+
+                token_sequence = torch.cat(
+                    [text_embeddings.unsqueeze(1), video_embeddings], dim=1
+                )  # shape: [B, T + 1, D]
+
+                # Build causal mask if enabled
+                mask = None
+                if self.config.causal_mask:
+                    L = token_sequence.shape[1]
+                    mask = torch.triu(torch.ones((L, L), dtype=torch.bool, device=token_sequence.device), diagonal=1)
+
+                token_embeddings = self.transformer(token_sequence, src_key_padding_mask=None, mask=mask)
+                D = token_embeddings.shape[-1]
+                final_embeddings = token_embeddings[:, 1:, :]  # avoid the text embedding
+
+                # Progress prediction for all frames
+                progress_logits = self.progress_head(final_embeddings)
+                if self.use_discrete_progress:
+                    # Discrete: [b, t, num_bins] - keep as is
+                    pass
+                else:
+                    # Continuous: [b, t, 1] -> [b, t]
+                    progress_logits = progress_logits.squeeze(-1)
+
+                # Predict success for all frames
+                success_logits = self.success_head(final_embeddings)
+                success_logits = success_logits.squeeze(-1)
+
+            progress_logits = {"A": progress_logits, "B": None}
+            success_logits = {"A": success_logits, "B": None}
+            output.progress_logits = progress_logits
+            output.success_logits = success_logits
+
+        return output, timing_raw
+
+
+# Register the model and configs with transformers
+AutoConfig.register("rewind_transformer", ReWINDTransformerConfig)
+AutoModel.register(ReWINDTransformerConfig, ReWiNDTransformer)
diff --git a/robometer/robometer/models/utils.py b/robometer/robometer/models/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..9e38c6f8118d164946208f1c64ed4aef4ba5f3e3
--- /dev/null
+++ b/robometer/robometer/models/utils.py
@@ -0,0 +1,77 @@
+import torch
+import numpy as np
+from dataclasses import dataclass
+
+
+@dataclass
+class ModelOutput:
+    pref_logits: torch.Tensor | None = None
+    success_logits: torch.Tensor | None = None
+    progress_logits: torch.Tensor | None = None
+
+    hidden_states: torch.Tensor | None = None
+
+
+def convert_bins_to_continuous(bin_logits: torch.Tensor | np.ndarray) -> torch.Tensor | np.ndarray:
+    """Convert discrete bins to continuous progress values in [0, 1] via weighted sum of bin centers."""
+    num_bins = bin_logits.shape[-1]
+    if (bin_logits.sum(dim=-1) == 1).all():
+        bin_probs = bin_logits
+    else:
+        bin_probs = (
+            torch.softmax(bin_logits, dim=-1)
+            if isinstance(bin_logits, torch.Tensor)
+            else np.softmax(bin_logits, axis=-1)
+        )
+    bin_centers = (
+        torch.linspace(0.0, 1.0, num_bins, device=bin_logits.device, dtype=bin_logits.dtype)
+        if isinstance(bin_logits, torch.Tensor)
+        else np.linspace(0.0, 1.0, num_bins)
+    )
+    return (
+        (bin_probs * bin_centers).sum(dim=-1)
+        if isinstance(bin_logits, torch.Tensor)
+        else (bin_probs * bin_centers).sum(axis=-1)
+    )
+
+
+def convert_bins_to_continuous_hard(
+    bin_logits: torch.Tensor | np.ndarray,
+) -> torch.Tensor | np.ndarray:
+    """
+    Convert discrete bins to a continuous value in [0, 1]
+    by selecting the argmax bin and returning its center.
+    """
+    num_bins = bin_logits.shape[-1]
+
+    if isinstance(bin_logits, torch.Tensor):
+        idx = torch.argmax(bin_logits, dim=-1)
+        bin_centers = torch.linspace(
+            0.0,
+            1.0,
+            num_bins,
+            device=bin_logits.device,
+            dtype=bin_logits.dtype,
+        )
+        return bin_centers[idx]
+
+    else:
+        idx = np.argmax(bin_logits, axis=-1)
+        bin_centers = np.linspace(0.0, 1.0, num_bins)
+        return bin_centers[idx]
+
+
+def convert_bin_index_to_continuous(bin_index: torch.Tensor | np.ndarray, num_bins: int) -> torch.Tensor | np.ndarray:
+    """Convert discrete bin index to continuous progress value in [0, 1] by selecting the argmax bin and returning its center."""
+    bin_centers = torch.linspace(0.0, 1.0, num_bins, device=bin_index.device, dtype=bin_index.dtype)
+    return bin_centers[bin_index.long()]
+
+
+def convert_discrete_target_to_continuous(
+    target: torch.Tensor | np.ndarray, num_bins: int
+) -> torch.Tensor | np.ndarray:
+    """Convert discrete target to continuous progress value in [0, 1] by selecting the argmax bin and returning its center."""
+    if len(target.shape) == 2:
+        return convert_bin_index_to_continuous(target, num_bins)
+    else:
+        return convert_bins_to_continuous(target)
diff --git a/robometer/robometer/trainers/__init__.py b/robometer/robometer/trainers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..d92c4ea2d8a775a3e40ef0fe7d56b985a5a58bf0
--- /dev/null
+++ b/robometer/robometer/trainers/__init__.py
@@ -0,0 +1,4 @@
+from .rewind_trainer import ReWiNDTrainer
+from .rbm_heads_trainer import RBMHeadsTrainer
+
+__all__ = ["RBMHeadsTrainer", "ReWiNDTrainer"]
diff --git a/robometer/robometer/trainers/rbm_heads_trainer.py b/robometer/robometer/trainers/rbm_heads_trainer.py
new file mode 100755
index 0000000000000000000000000000000000000000..b1684efee31bc5666db624b8a754ade3a745e477
--- /dev/null
+++ b/robometer/robometer/trainers/rbm_heads_trainer.py
@@ -0,0 +1,2607 @@
+import collections
+import copy
+import io
+import json
+import os
+import random
+from typing import Dict, List, Tuple, Optional, Any, final
+
+import matplotlib.animation as animation
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from sklearn.metrics import average_precision_score
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from transformers import Trainer
+
+from robometer.data.datasets.name_mapping import DS_SHORT_NAME_MAPPING
+from robometer.evals.compile_results import (
+    run_quality_preference_eval,
+    run_reward_alignment_eval_per_trajectory,
+    run_confusion_matrix_eval,
+    run_policy_ranking_eval,
+)
+from robometer.models.utils import ModelOutput, convert_bins_to_continuous, convert_discrete_target_to_continuous
+from robometer.utils.distributed import banner, get_rank, is_rank_0, log_fsdp_diagnostics
+from robometer.utils.logger import Logger, get_logger, log_memory_usage
+from robometer.utils.metrics import compute_spearman_correlation
+from robometer.utils.setup_utils import setup_batch_collator, setup_custom_eval_dataset
+from robometer.utils.tensor_utils import t2n
+from robometer.utils.timer import _timer
+from robometer.utils.video_utils import create_policy_ranking_grid
+
+logger = get_logger()
+
+
+def seed_worker(worker_id):
+    """Set random seed for dataloader workers."""
+    import random
+
+    worker_seed = torch.initial_seed() % 2**32
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+
+
+def reduce_metrics_with_accelerate(metrics: Dict[str, Any], accelerator, aggregate_method="sum"):
+    """
+    Reduce multiple scalar metrics using Accelerate's built-in methods.
+    Handles cases where different processes have different metric keys.
+    metrics: dict of {name: float or tensor}
+    Returns dict with averaged metrics across all ranks.
+    """
+    if not metrics:
+        return metrics
+
+    # Step 1: Gather all metric keys from all processes
+    local_keys = list(metrics.keys())
+    all_keys_gathered = accelerator.gather_for_metrics(local_keys)
+
+    # Step 2: Create union of all keys across all processes
+    all_unique_keys = set()
+    for keys_from_process in all_keys_gathered:
+        if isinstance(keys_from_process, list):
+            all_unique_keys.update(keys_from_process)
+        else:
+            # Handle single key case
+            all_unique_keys.add(keys_from_process)
+
+    all_unique_keys = sorted(all_unique_keys)
+
+    # Step 3: Create synchronized metrics dict with 0.0 for missing keys
+    synchronized_metrics = {}
+    for key in all_unique_keys:
+        if key in metrics:
+            synchronized_metrics[key] = metrics[key]
+        else:
+            # This process doesn't have this metric, use 0.0
+            synchronized_metrics[key] = 0.0
+
+    # Step 4: Now reduce all metrics (all processes have same keys)
+    result_metrics = {}
+
+    for key, value in synchronized_metrics.items():
+        try:
+            # Convert to tensor on accelerator device
+            if torch.is_tensor(value):
+                tensor_val = value.to(accelerator.device, dtype=torch.float32)
+            else:
+                tensor_val = torch.tensor(float(value), dtype=torch.float32, device=accelerator.device)
+
+            # Check for NaN values before reduction
+            if torch.isnan(tensor_val).any():
+                logger.warning(f"NaN detected in metric '{key}', using 0.0")
+                tensor_val = torch.tensor(0.0, dtype=torch.float32, device=accelerator.device)
+
+            # Check for infinity values
+            if torch.isinf(tensor_val).any():
+                logger.warning(f"Infinity detected in metric '{key}', using 0.0")
+                tensor_val = torch.tensor(0.0, dtype=torch.float32, device=accelerator.device)
+
+            # Use accelerator's reduce method - all processes participate
+            reduced_val = accelerator.reduce(tensor_val, reduction=aggregate_method)
+
+            # Final check for NaN in reduced result
+            if torch.isnan(reduced_val).any():
+                logger.warning(f"NaN in reduced result for metric '{key}', using fallback")
+                result_metrics[key] = 0.0
+            else:
+                result_metrics[key] = reduced_val.item()
+
+        except Exception as metric_error:
+            # If individual metric fails, keep original value (or 0.0 if missing)
+            logger.warning(f"Failed to reduce metric '{key}': {metric_error}")
+            if key in metrics:
+                original_val = float(metrics[key]) if not torch.is_tensor(metrics[key]) else metrics[key].item()
+                result_metrics[key] = 0.0 if np.isnan(original_val) else original_val
+            else:
+                result_metrics[key] = 0.0
+
+    # Step 5: Return all reduced metrics (all processes should have the same keys after reduction)
+    # Return all keys from result_metrics to ensure we get all metrics across all processes
+    return result_metrics
+
+
+class RBMHeadsTrainer(Trainer):
+    def __init__(self, config, *args, logger=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.config = config
+        self.log_metadata = collections.defaultdict(float)
+        self.global_metadata = collections.defaultdict(float)
+        self.timing_raw = collections.defaultdict(float)
+        self._ddp_static_graph_set = False  # Flag to track if DDP static graph has been set
+        self._fsdp_diagnostics_logged = False  # Flag to track if FSDP diagnostics have been logged
+
+        if logger is not None:
+            self.logger = logger
+        else:
+            log_level = self.config.logging.log_level
+            self.logger = Logger(
+                log_to=self.config.logging.log_to,
+                output_dir=getattr(self.args, "output_dir", "./logs"),
+                is_main_process=is_rank_0(),
+                log_level=log_level,
+            )
+
+        # Use loguru logger after it's been initialized
+        loguru_logger = get_logger()
+        loguru_logger.info(f"DDP find_unused_parameters: {getattr(self.args, 'ddp_find_unused_parameters', 'N/A')}")
+
+    def create_optimizer(self):
+        """
+        Override to create optimizer with separate parameter groups for vision encoder layers.
+        If vision_encoder_lr is set, the last N vision encoder layers will use that LR,
+        while all other parameters use the default learning rate.
+        """
+        # Check if we need to create parameter groups for vision encoder
+        vision_encoder_lr = self.config.training.vision_encoder_lr
+        vision_encoder_num_layers = self.config.training.vision_encoder_num_layers
+
+        if vision_encoder_lr is None or vision_encoder_lr <= 0:
+            # No special vision encoder LR, use default optimizer
+            return super().create_optimizer()
+
+        # Get the model
+        model = self.model
+        if not hasattr(model, "model") or not hasattr(model.model, "visual"):
+            logger.warning(
+                "vision_encoder_lr is set but model doesn't have visual encoder. "
+                "Using default optimizer without parameter groups."
+            )
+            return super().create_optimizer()
+
+        # Get vision encoder blocks
+        visual_encoder = model.model.visual
+        if not hasattr(visual_encoder, "blocks"):
+            logger.warning(
+                "vision_encoder_lr is set but visual encoder doesn't have blocks. "
+                "Using default optimizer without parameter groups."
+            )
+            return super().create_optimizer()
+
+        blocks = visual_encoder.blocks
+        total_blocks = len(blocks)
+
+        if vision_encoder_num_layers > total_blocks:
+            logger.warning(
+                f"vision_encoder_num_layers ({vision_encoder_num_layers}) is greater than "
+                f"total blocks ({total_blocks}). Using all blocks for vision encoder LR."
+            )
+            vision_encoder_num_layers = total_blocks
+
+        # Identify parameters for last N layers
+        vision_encoder_params = []
+        other_params = []
+
+        # Get all parameters and their names
+        for name, param in model.named_parameters():
+            if not param.requires_grad:
+                continue
+
+            # Check if this parameter belongs to the last N vision encoder blocks
+            is_vision_encoder_param = False
+            if "visual.blocks" in name:
+                # Extract block index from parameter name
+                # Format: model.visual.blocks.{idx}.{rest}
+                try:
+                    parts = name.split("visual.blocks.")
+                    if len(parts) > 1:
+                        block_part = parts[1].split(".")[0]
+                        block_idx = int(block_part)
+                        # Check if this is one of the last N blocks
+                        if block_idx >= (total_blocks - vision_encoder_num_layers):
+                            is_vision_encoder_param = True
+                except (ValueError, IndexError):
+                    # If we can't parse the block index, skip this parameter
+                    pass
+
+            if is_vision_encoder_param:
+                vision_encoder_params.append(param)
+            else:
+                other_params.append(param)
+
+        if not vision_encoder_params:
+            logger.warning(
+                "No vision encoder parameters found for parameter groups. "
+                "Using default optimizer without parameter groups."
+            )
+            return super().create_optimizer()
+
+        # Use AdamW as default (same as HuggingFace Trainer)
+        optimizer_kwargs = {
+            "betas": (
+                self.args.adam_beta1 if hasattr(self.args, "adam_beta1") else 0.9,
+                self.args.adam_beta2 if hasattr(self.args, "adam_beta2") else 0.999,
+            ),
+            "eps": self.args.adam_epsilon if hasattr(self.args, "adam_epsilon") else 1e-8,
+            "weight_decay": self.args.weight_decay,
+        }
+
+        # Create parameter groups with different learning rates
+        param_groups = [
+            {
+                "params": other_params,
+                "lr": self.args.learning_rate,
+                **optimizer_kwargs,
+            },
+            {
+                "params": vision_encoder_params,
+                "lr": vision_encoder_lr,
+                **optimizer_kwargs,
+            },
+        ]
+
+        optimizer = torch.optim.AdamW(param_groups)
+
+        logger.info(
+            f"Created optimizer with parameter groups: "
+            f"{len(other_params)} params at LR={self.args.learning_rate}, "
+            f"{len(vision_encoder_params)} vision encoder params (last {vision_encoder_num_layers} blocks) at LR={vision_encoder_lr}"
+        )
+        self.optimizer = optimizer
+
+        return optimizer
+
+    def _post_checkpoint_load_reset(self):
+        """
+        Reset model and optimizer state after loading from checkpoint.
+        This addresses issues where checkpoint loading can leave stale gradients
+        or computational graph state that causes crashes during training.
+        """
+        logger.info("Performing post-checkpoint load reset...")
+
+        # Ensure model is in training mode
+        self.model.train()
+
+        # Clear any cached gradients or computational graph state
+        # NOTE: We don't clear optimizer.state or param_groups as that breaks the lr_scheduler
+        try:
+            # Zero out any existing gradients
+            if hasattr(self, "optimizer") and self.optimizer is not None:
+                self.optimizer.zero_grad(set_to_none=True)
+        except Exception as e:
+            logger.warning(f"Could not clear gradients: {e}")
+
+        # Clear CUDA cache if available
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+
+        logger.info("Post-checkpoint load reset complete")
+
+    def _normalize_list_like(self, value):
+        """Convert None/scalars/tuples to a list so we can safely gather across ranks."""
+        if value is None:
+            return []
+        if isinstance(value, list):
+            return list(value)
+        if isinstance(value, tuple):
+            return list(value)
+        return [value]
+
+    def _gather_list_across_processes(self, value):
+        """Gather Python lists (or list-like) across all ranks."""
+        normalized = self._normalize_list_like(value)
+        if not dist.is_initialized():
+            return normalized
+
+        world_size = dist.get_world_size()
+        gathered = [None] * world_size
+        dist.all_gather_object(gathered, normalized)
+
+        flattened = []
+        for proc_list in gathered:
+            if not proc_list:
+                continue
+            flattened.extend(proc_list)
+        return flattened
+
+    def _gather_metadata_fields(self, sample_inputs: Dict[str, Any], fields: List[str]) -> Dict[str, Any]:
+        """Gather heterogeneous metadata fields (lists, tensors) across processes."""
+        gathered = {}
+        for field in fields:
+            field_value = sample_inputs.get(field)
+            if torch.is_tensor(field_value):
+                gathered[field] = self.accelerator.gather_for_metrics(field_value)
+            else:
+                gathered[field] = self._gather_list_across_processes(field_value)
+        return gathered
+
+    def _truncate_metadata_lists(self, metadata: dict, max_len: int) -> dict:
+        """Ensure metadata lists align with tensor batch sizes without altering tensors."""
+        if max_len < 0:
+            return metadata
+        for key, value in metadata.items():
+            if isinstance(value, list):
+                metadata[key] = value[:max_len]
+        return metadata
+
+    def _get_learning_rate(self):
+        """
+        Override to safely get learning rate, handling cases where scheduler hasn't been stepped yet.
+        """
+        try:
+            if hasattr(self, "lr_scheduler") and self.lr_scheduler is not None:
+                last_lrs = self.lr_scheduler.get_last_lr()
+                if last_lrs:
+                    return last_lrs[0]
+            # Fallback to optimizer's learning rate
+            if hasattr(self, "optimizer") and self.optimizer is not None:
+                if self.optimizer.param_groups:
+                    return self.optimizer.param_groups[0]["lr"]
+            # Last resort: return configured learning rate
+            return self.args.learning_rate
+        except Exception as e:
+            logger.warning(f"Could not get learning rate: {e}")
+            return self.args.learning_rate
+
+    def train(self, resume_from_checkpoint=None, **kwargs):
+        """
+        Override train method to perform post-checkpoint reset.
+        """
+        # If resuming from checkpoint, set flag for reset in first training step
+        if resume_from_checkpoint is not None:
+            logger.info(f"Resuming from checkpoint: {resume_from_checkpoint}")
+            self._just_resumed_from_checkpoint = True
+
+        # Call parent train method
+        result = super().train(resume_from_checkpoint=resume_from_checkpoint, **kwargs)
+
+        return result
+
+    def training_step(self, model, inputs, num_items_in_batch=None):
+        """
+        Perform a training step and log custom losses.
+        """
+        logger.debug("training_step: Starting")
+
+        if not self._fsdp_diagnostics_logged:
+            log_fsdp_diagnostics(model, accelerator=self.accelerator, logger=logger)
+            self._fsdp_diagnostics_logged = True
+
+        # Check if we just resumed from checkpoint (first step after resume)
+        if hasattr(self, "_just_resumed_from_checkpoint") and self._just_resumed_from_checkpoint:
+            self._post_checkpoint_load_reset()
+            self._just_resumed_from_checkpoint = False
+
+        self.timing_raw = {}
+
+        # Initialize log_metadata
+        self.log_metadata = {}
+
+        # Safety check: ensure model is in training mode and gradients are properly set up
+        if not model.training:
+            logger.warning("Model not in training mode, setting to train mode")
+            model.train()
+
+        # Clear any stale gradients before starting
+        if hasattr(self, "optimizer") and self.optimizer is not None:
+            self.optimizer.zero_grad(set_to_none=True)
+
+        with _timer("time/training_step", timing_raw=self.timing_raw):
+            loss = super().training_step(model, inputs, num_items_in_batch)
+
+        # Extract the separate batches
+        preference_inputs = inputs.get("preference_inputs", {})
+        progress_inputs = inputs.get("progress_inputs", {})
+        num_preferences = inputs.get("num_preferences", 0)
+        num_progress = inputs.get("num_progress", 0)
+
+        logger.trace(f"num_preferences: {num_preferences}, num_progress: {num_progress}")
+
+        if num_preferences > 0 and preference_inputs:
+            rejected_data_gen_strategy = preference_inputs["rejected_data_gen_strategy"]
+            if isinstance(rejected_data_gen_strategy, list) and len(rejected_data_gen_strategy) > 0:
+                for s in rejected_data_gen_strategy:
+                    self.global_metadata[f"pref_{s}"] += 1
+
+            data_sources = preference_inputs.get("data_source", None)
+            if data_sources is not None:
+                for ds in data_sources:
+                    self.global_metadata[f"total_{ds}"] += 1.0
+
+        if num_progress > 0 and progress_inputs:
+            data_gen_strategy = progress_inputs["data_gen_strategy"]
+            if isinstance(data_gen_strategy, list) and len(data_gen_strategy) > 0:
+                for s in data_gen_strategy:
+                    self.global_metadata[f"prog_{s}"] += 1
+
+            data_sources = progress_inputs.get("data_source", None)
+            if data_sources is not None:
+                for ds in data_sources:
+                    self.global_metadata[f"total_{ds}"] += 1.0
+
+        # Update global metadata for training
+        # add to total batch size and sum across all processes
+        self.global_metadata["total_samples"] += num_preferences + num_progress
+        self.global_metadata["total_preferences"] += num_preferences
+        self.global_metadata["total_progress"] += num_progress
+
+        logger.trace("finished updating global metadata")
+
+        # self._update_resample_attempt_metrics(inputs)
+
+        # logger.trace("update resample attempt metrics")
+
+        # Log custom losses at specified intervals (using our custom logger only)
+        if self.state.global_step % self.args.logging_steps == 0:
+            try:
+                self._log_metadata()
+            except Exception as e:
+                logger.warning(f"Error logging metadata: {e}")
+
+        # Log GPU memory usage at every training step for diagnostics
+        log_memory_usage(f"Step {self.state.global_step}")
+
+        return loss
+
+    def _get_optimizer_stats(self):
+        """Get optimizer and gradient statistics for logging."""
+        optim_stats = {}
+
+        if not hasattr(self, "optimizer") or self.optimizer is None:
+            return optim_stats
+
+        # Get learning rates for each parameter group
+        for i, param_group in enumerate(self.optimizer.param_groups):
+            lr = param_group.get("lr", 0.0)
+            optim_stats[f"optim/lr_group_{i}"] = lr
+
+        # If only one param group, also log as optim/lr for convenience
+        if len(self.optimizer.param_groups) == 1:
+            optim_stats["optim/lr"] = self.optimizer.param_groups[0].get("lr", 0.0)
+
+        # Compute gradient norms across all model parameters
+        total_norm = 0.0
+        num_params_with_grad = 0
+        max_grad_norm = 0.0
+        min_grad_norm = float("inf")
+
+        for p in self.model.parameters():
+            if p.grad is not None:
+                param_norm = p.grad.data.norm(2).item()
+                total_norm += param_norm**2
+                num_params_with_grad += 1
+                max_grad_norm = max(max_grad_norm, param_norm)
+                min_grad_norm = min(min_grad_norm, param_norm)
+
+        if num_params_with_grad > 0:
+            total_norm = total_norm**0.5
+            optim_stats["optim/preclip_grad_norm"] = total_norm
+            optim_stats["optim/preclip_grad_norm_max"] = max_grad_norm
+            optim_stats["optim/preclip_grad_norm_min"] = min_grad_norm if min_grad_norm != float("inf") else 0.0
+            optim_stats["optim/num_params_with_grad"] = num_params_with_grad
+
+        # Compute parameter norms across all model parameters
+        total_param_norm = 0.0
+        max_param_norm = 0.0
+        min_param_norm = float("inf")
+        param_norms = []
+
+        for name, p in self.model.named_parameters():
+            if not p.requires_grad or p.data is None:
+                continue
+            param_norm = p.data.norm(2).item()
+            total_param_norm += param_norm**2
+            max_param_norm = max(max_param_norm, param_norm)
+            min_param_norm = min(min_param_norm, param_norm)
+            param_norms.append((name, param_norm))
+
+        if param_norms:
+            total_param_norm = total_param_norm**0.5
+            optim_stats["optim/param_norm"] = total_param_norm
+            optim_stats["optim/param_norm_max"] = max_param_norm
+            optim_stats["optim/param_norm_min"] = min_param_norm if min_param_norm != float("inf") else 0.0
+
+        # Get optimizer state statistics (e.g., momentum, variance for Adam)
+        if hasattr(self.optimizer, "state") and len(self.optimizer.state) > 0:
+            # For Adam-like optimizers, log average momentum and variance
+            exp_avg_norms = []
+            exp_avg_sq_norms = []
+
+            for state in self.optimizer.state.values():
+                if "exp_avg" in state:
+                    exp_avg_norms.append(state["exp_avg"].norm(2).item())
+                if "exp_avg_sq" in state:
+                    exp_avg_sq_norms.append(state["exp_avg_sq"].norm(2).item())
+
+            if exp_avg_norms:
+                optim_stats["optim/exp_avg_norm_mean"] = np.mean(exp_avg_norms)
+                optim_stats["optim/exp_avg_norm_max"] = np.max(exp_avg_norms)
+            if exp_avg_sq_norms:
+                optim_stats["optim/exp_avg_sq_norm_mean"] = np.mean(exp_avg_sq_norms)
+                optim_stats["optim/exp_avg_sq_norm_max"] = np.max(exp_avg_sq_norms)
+
+        # Log top 10 parameters with largest gradient norms
+        param_grad_norms = []
+        for name, p in self.model.named_parameters():
+            if not p.requires_grad or p.grad is None:
+                continue
+            grad_norm = p.grad.data.norm(2).item()
+            param_grad_norms.append((name, grad_norm))
+
+        if param_grad_norms:
+            # Sort by gradient norm (descending) and take top 10
+            param_grad_norms.sort(key=lambda x: x[1], reverse=True)
+            for i, (name, grad_norm) in enumerate(param_grad_norms[:5]):
+                # Shorten parameter name for cleaner logging
+                short_name = name.replace("model.", "").replace("module.", "")
+                optim_stats[f"optim/top_preclip_grad_norm_{i + 1}_{short_name}"] = grad_norm
+
+        if param_norms:
+            # Sort by parameter norm (descending) and take top 10
+            param_norms.sort(key=lambda x: x[1], reverse=True)
+            for i, (name, param_norm) in enumerate(param_norms[:5]):
+                short_name = name.replace("model.", "").replace("module.", "")
+                optim_stats[f"optim/top_param_norm_{i + 1}_{short_name}"] = param_norm
+
+        return optim_stats
+
+    def _update_resample_attempt_metrics(self, inputs: Dict[str, Any]) -> None:
+        """Aggregate resample attempt statistics across processes."""
+        if not hasattr(self, "accelerator"):
+            return
+
+        local_pairs: List[Tuple[str, float]] = []
+
+        for key in ("preference_inputs", "progress_inputs"):
+            sample_inputs = inputs.get(key) or {}
+            resample_attempts = sample_inputs.get("resample_attempts")
+            if resample_attempts is None:
+                continue
+
+            if torch.is_tensor(resample_attempts):
+                attempts_tensor = resample_attempts.to(self.accelerator.device, dtype=torch.float32).view(-1)
+            else:
+                attempts_tensor = torch.tensor(
+                    resample_attempts, dtype=torch.float32, device=self.accelerator.device
+                ).view(-1)
+
+            if attempts_tensor.numel() == 0:
+                continue
+
+            sample_category = key.replace("_inputs", "")
+            strategies = sample_inputs.get("data_gen_strategy")
+            if strategies is None:
+                raise ValueError(
+                    f"Expected data_gen_strategy for {sample_category} samples when logging resample attempts."
+                )
+
+            if len(strategies) != attempts_tensor.numel():
+                raise ValueError(
+                    f"Mismatch between resample attempts ({attempts_tensor.numel()}) and strategies "
+                    f"({len(strategies)}) for {sample_category} samples."
+                )
+
+            strategy_labels = [f"{sample_category}/{str(strategy)}" for strategy in strategies]
+
+            for attempt_value, strategy_label in zip(attempts_tensor.tolist(), strategy_labels):
+                local_pairs.append((strategy_label, float(attempt_value)))
+
+        if dist.is_initialized():
+            world_size = dist.get_world_size()
+            gathered_lists: List[List[Tuple[str, float]]] = [None] * world_size
+            dist.all_gather_object(gathered_lists, local_pairs)
+            flat_pairs = [pair for proc_pairs in gathered_lists for pair in proc_pairs]
+        else:
+            flat_pairs = local_pairs
+
+        if not flat_pairs:
+            return
+
+        all_attempts = [attempt for _, attempt in flat_pairs]
+        self.log_metadata["data/resample_min"] = float(min(all_attempts))
+        self.log_metadata["data/resample_max"] = float(max(all_attempts))
+        self.log_metadata["data/resample_mean"] = float(sum(all_attempts) / len(all_attempts))
+
+        strategy_values: Dict[str, List[float]] = collections.defaultdict(list)
+        for label, attempt in flat_pairs:
+            strategy_values[label].append(attempt)
+
+        for label, values in strategy_values.items():
+            if not values:
+                continue
+            safe_label = label.replace("/", "_").replace(" ", "_")
+            strategy_min = float(min(values))
+            strategy_max = float(max(values))
+            strategy_mean = float(sum(values) / len(values))
+            self.log_metadata[f"data/resample_min_{safe_label}"] = strategy_min
+            self.log_metadata[f"data/resample_max_{safe_label}"] = strategy_max
+            self.log_metadata[f"data/resample_mean_{safe_label}"] = strategy_mean
+
+    def _log_metadata(self):
+        """Log custom RBM losses to wandb and console."""
+        if not self.log_metadata:
+            return
+
+        logger.trace("logging metadata, starting to aggregate metrics")
+
+        # Use local metrics (no aggregation needed for individual GPU metrics)
+        log_metadata = reduce_metrics_with_accelerate(self.log_metadata, self.accelerator, aggregate_method="mean")
+
+        logger.trace("finished aggregating metrics")
+
+        training_step_time = self.timing_raw.get("time/training_step", 0.0)
+        it_per_sec = 1.0 / training_step_time if training_step_time > 0 else 0.0
+
+        # Prepare logging data using aggregated losses
+        log_data = {
+            "step": self.state.global_step,
+            "epoch": self.state.epoch,
+            "train/it_per_sec": it_per_sec,
+            **self.timing_raw,
+            **log_metadata,
+        }
+
+        # Log global metadata
+        logger.trace("logging global metadata")
+        global_metadata = reduce_metrics_with_accelerate(self.global_metadata, self.accelerator, aggregate_method="sum")
+        logger.trace("finished aggregating global metadata")
+
+        # Convert counts to fractions of total samples
+        total_samples = global_metadata["total_samples"]
+        log_global = {
+            f"counts/{key}": value / total_samples for key, value in global_metadata.items() if key != "total_samples"
+        }
+
+        log_data.update(log_global)
+
+        # Log optimizer and gradient statistics
+        optim_stats = self._get_optimizer_stats()
+        log_data.update(optim_stats)
+
+        # make sure values are floats so they are loggable into wandb reports
+        log_data = {k: float(v) for k, v in log_data.items()}
+
+        self.logger.log_scalars(log_data, step=self.state.global_step + 1)
+
+        if is_rank_0():
+            logger.info(f"Step {self.state.global_step}, Epoch {self.state.epoch:.2f}:")
+            logger.info("-" * 50)
+            logger.info(f"  train/it_per_sec: {it_per_sec:.4f}")
+            for key in log_global:
+                logger.info(f"  {key}: {log_global[key]}")
+
+            rounded_times = {k: round(v, 2) for k, v in self.timing_raw.items()}
+            logger.info(f"Timing raw: {rounded_times}")
+
+            # Log optimizer stats to console
+            # if optim_stats:
+            #     logger.info(f"Optimizer stats: {optim_stats}")
+
+    def _make_eval_dataloader(self, dataset):
+        """Create a distributed evaluation dataloader with proper sampling."""
+        collator = setup_batch_collator(self.model.processor, self.model.tokenizer, self.config, is_eval=True)
+
+        dl = DataLoader(
+            dataset,
+            batch_size=self.config.training.per_device_eval_batch_size,
+            collate_fn=collator,
+            num_workers=self.args.dataloader_num_workers,
+            pin_memory=self.args.dataloader_pin_memory,
+            drop_last=False,
+            # Force persistent_workers=False for eval to prevent memory leaks across datasets
+            persistent_workers=False,
+            worker_init_fn=seed_worker,
+        )
+        prepared_dl = self.accelerator.prepare(dl)
+        return prepared_dl
+
+    def _setup_eval_dataset(self, eval_type, eval_dataset):
+        """Setup dataset and dataloader for evaluation."""
+        eval_cfg = copy.deepcopy(self.config.data)
+
+        # explicitly set dataset type to rbm for custom eval datasets
+        eval_cfg.dataset_type = "rbm"
+
+        if isinstance(eval_dataset, list):
+            eval_cfg.eval_datasets = eval_dataset
+        else:
+            eval_cfg.eval_datasets = [eval_dataset]
+
+        # Create custom eval dataset with the appropriate sampler
+        sampler_kwargs = {}
+        sampler_kwargs["random_seed"] = self.config.custom_eval.custom_eval_random_seed
+
+        if eval_type == "reward_alignment":
+            sampler_kwargs["max_trajectories"] = self.config.custom_eval.reward_alignment_max_trajectories
+            sampler_kwargs["frame_step"] = (
+                2 if (self.config.trainer_cls == "rbm_heads" and not self.config.data.use_multi_image) else 1
+            )
+            sampler_kwargs["use_frame_steps"] = self.config.custom_eval.use_frame_steps
+        elif eval_type == "policy_ranking":
+            sampler_kwargs["num_examples_per_quality_pr"] = self.config.custom_eval.num_examples_per_quality_pr
+            sampler_kwargs["num_partial_successes"] = self.config.custom_eval.num_partial_successes
+            sampler_kwargs["max_tasks"] = self.config.custom_eval.policy_ranking_max_tasks
+            sampler_kwargs["frame_step"] = (
+                2 if (self.config.trainer_cls == "rbm_heads" and not self.config.data.use_multi_image) else 1
+            )
+            # sampler_kwargs["use_frame_steps"] = self.config.custom_eval.use_frame_steps
+            # we only care about the final predicted progress for ranking
+            sampler_kwargs["use_frame_steps"] = False
+        elif eval_type == "quality_preference":
+            sampler_kwargs["comparisons_per_task"] = self.config.custom_eval.comparisons_per_task
+            sampler_kwargs["max_comparisons"] = self.config.custom_eval.max_comparisons
+
+        dataset = setup_custom_eval_dataset(
+            eval_cfg, sampler_type=eval_type, verbose=False, sampler_kwargs=sampler_kwargs
+        )
+        # Explicitly delete eval_cfg after dataset creation to free memory
+        del eval_cfg
+
+        logger.info(f"  Dataset size: {len(dataset)}")
+        # log_memory_usage(f"After creating dataset")
+
+        dataloader = self._make_eval_dataloader(dataset)
+        logger.info(f"  Dataloader created with {len(dataloader)} batches")
+        # log_memory_usage(f"After creating dataloader")
+
+        # Ensure model is in eval mode and clear any gradient buffers
+        self.model.eval()
+        # Explicitly zero any gradients that might exist (shouldn't, but safety measure)
+        if hasattr(self, "optimizer") and self.optimizer is not None:
+            self.optimizer.zero_grad(set_to_none=True)
+
+        # Clear cache before starting evaluation
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+        # log_memory_usage(f"After clearing cache, before eval loop")
+
+        return dataset, dataloader
+
+    def _process_batch_progress_eval(self, batch, eval_type):
+        """Process a batch for progress-based evaluations (reward_alignment, policy_ranking, confusion_matrix)."""
+        logger.trace(f"    Processing {eval_type} batch")
+        progress_samples = batch["progress_inputs"]
+        logger.trace(f"    Calling forward_model for progress")
+        with torch.no_grad():
+            outputs, _ = self.forward_model(self.model, progress_samples, sample_type="progress")
+        logger.trace(f"    Forward pass complete")
+
+        progress_logits = outputs.progress_logits
+        progress_pred = progress_logits["A"]
+
+        # Gather everything
+        progress_pred = self.accelerator.gather_for_metrics(progress_pred)
+        target_progress = self.accelerator.gather_for_metrics(progress_samples["target_progress"])
+
+        # Gather metadata fields
+        metadata_fields = [
+            "task",
+            "data_source",
+            "data_gen_strategy",
+            "quality_labels",
+            "metadata",
+            "partial_success",
+        ]
+        gathered_metadata_dict = self._gather_metadata_fields(progress_samples, metadata_fields)
+        num_progress_samples = progress_pred.shape[0] if progress_pred is not None else 0
+        gathered_metadata_dict = self._truncate_metadata_lists(gathered_metadata_dict, num_progress_samples)
+
+        # Handle success predictions if needed
+        success_pred_gathered = None
+        success_probs_gathered = None
+        success_labels_gathered = None
+        if self.config.model.train_success_head:
+            success_pred = outputs.success_logits["A"]
+            success_probs = torch.sigmoid(success_pred)
+            success_binary = (success_probs > 0.5).float()
+            success_labels = progress_samples.get("success_labels")
+            success_pred_gathered = self.accelerator.gather_for_metrics(success_binary)
+            success_probs_gathered = self.accelerator.gather_for_metrics(success_probs)
+            success_labels_gathered = self.accelerator.gather_for_metrics(success_labels)
+
+            # Clean up intermediate tensors (but keep gathered tensors for eval_results)
+            del (
+                success_pred,
+                success_binary,
+                success_probs,
+                success_labels,
+            )
+
+        # Build eval_results on all processes for compute_eval_metrics
+        batch_results = []
+        for i in range(len(progress_pred)):
+            metadata = gathered_metadata_dict["metadata"][i]
+            sample_result = {
+                "task": gathered_metadata_dict["task"][i],
+                "target_progress": t2n(target_progress[i]),
+                "progress_pred": t2n(progress_pred[i]),
+                "data_source": gathered_metadata_dict["data_source"][i],
+                "data_gen_strategy": gathered_metadata_dict["data_gen_strategy"][i],
+                "quality_label": gathered_metadata_dict["quality_labels"][i],
+                "metadata": metadata,
+                "id": metadata["id"],
+                "video_path": metadata["video_path"],
+                "partial_success": gathered_metadata_dict["partial_success"][i],
+            }
+            if success_pred_gathered is not None:
+                sample_result["success_pred"] = t2n(success_pred_gathered[i])
+            if success_probs_gathered is not None:
+                sample_result["success_probs"] = t2n(success_probs_gathered[i])
+            if success_labels_gathered is not None:
+                sample_result["success_labels"] = t2n(success_labels_gathered[i])
+            batch_results.append(sample_result)
+
+        # Clean up gathered tensors and metadata after building results
+        del progress_pred, target_progress, gathered_metadata_dict
+        if success_pred_gathered is not None:
+            del success_pred_gathered
+        if success_probs_gathered is not None:
+            del success_probs_gathered
+        if success_labels_gathered is not None:
+            del success_labels_gathered
+
+        return batch_results, outputs
+
+    def _process_batch_preference_eval(self, batch):
+        """Process a batch for preference-based evaluations (quality_preference)."""
+        logger.trace(f"    Processing quality_preference batch")
+        preference_samples = batch["preference_inputs"]
+        logger.trace(f"    Calling forward_model for preference")
+        with torch.no_grad():
+            outputs, _ = self.forward_model(self.model, preference_samples, sample_type="preference")
+        logger.trace(f"    Forward pass complete")
+        pref_logits = outputs.pref_logits
+
+        # Gather predictions and labels across all ranks
+        pref_logits = self.accelerator.gather_for_metrics(pref_logits)
+        preference_labels = self.accelerator.gather_for_metrics(preference_samples["preference_labels"])
+
+        # Convert logits to binary predictions (0/1): apply sigmoid, then threshold at 0.5
+        pref_probs = torch.sigmoid(pref_logits)
+        binary_preds = (pref_probs > 0.5).float()
+
+        # Gather non-tensor metadata using helper (handles single and multi GPU)
+        gathered_pref_metadata = self._gather_metadata_fields(
+            preference_samples,
+            [
+                "task",
+                "data_source",
+                "chosen_data_gen_strategy",
+                "rejected_data_gen_strategy",
+                "metadata",
+            ],
+        )
+        num_pref_samples = binary_preds.shape[0] if binary_preds is not None else 0
+        gathered_pref_metadata = self._truncate_metadata_lists(gathered_pref_metadata, num_pref_samples)
+
+        # Build eval_results on all processes for compute_eval_metrics
+        batch_results = []
+        for i in range(len(binary_preds)):
+            if binary_preds[i] is None:
+                continue
+            sample_result = {
+                "task": gathered_pref_metadata["task"][i],
+                "preference_pred": t2n(binary_preds[i]),
+                "preference_logits": t2n(pref_logits[i]),
+                "preference_labels": t2n(preference_labels[i]),
+                "data_source": gathered_pref_metadata["data_source"][i],
+                "chosen_data_gen_strategy": gathered_pref_metadata["chosen_data_gen_strategy"][i],
+                "rejected_data_gen_strategy": gathered_pref_metadata["rejected_data_gen_strategy"][i],
+                "metadata": gathered_pref_metadata["metadata"][i],
+            }
+            batch_results.append(sample_result)
+
+        # Clean up gathered tensors and metadata after building results
+        del pref_logits, pref_probs, binary_preds, preference_labels, gathered_pref_metadata
+
+        return batch_results, outputs
+
+    def _compute_and_log_eval_metrics(self, eval_type, eval_results, ds_name, eval_step):
+        """Compute metrics and create visualizations for evaluation results."""
+        # Initialize variables to None to ensure they exist for cleanup
+        plots = None
+        video_frames_list = None
+        trajectory_progress_data = None
+        task_groups = None
+        task_details = None
+        confusion_plot = None
+        confusion_matrix = None
+
+        # if dataset name is too long, truncate it
+        if len(ds_name) > 90:
+            ds_name = ds_name[:90] + "..."
+
+        is_discrete_mode = self.config.loss.progress_loss_type.lower() == "discrete"
+        num_bins = self.config.loss.progress_discrete_bins if is_discrete_mode else None
+
+        data_source = None
+        if eval_results and len(eval_results) > 0:
+            data_source = eval_results[0]["data_source"]
+
+        if eval_type == "reward_alignment":
+            eval_metrics, plots, video_frames_list, trajectory_progress_data = run_reward_alignment_eval_per_trajectory(
+                eval_results,
+                self.config.data.progress_pred_type,
+                is_discrete_mode,
+                num_bins,
+                data_source,
+                use_frame_steps=self.config.custom_eval.use_frame_steps,
+                train_success_head=self.config.model.train_success_head,
+                last_frame_only=False,
+            )
+            # log_memory_usage(f"After compute_eval_metrics (reward_alignment)")
+
+            banner(
+                f"{eval_type} evaluation: {len(eval_results)} samples",
+                f"Metrics: {eval_metrics}",
+                inner_padding=1,
+            )
+
+            # Build rows of (video, figure)
+            rows = []
+            for plot, frames in zip(plots, video_frames_list):
+                if frames is not None:
+                    rows.append((frames, plot))
+
+            if video_frames_list and plots:
+                # Log individual images to wandb: first frame + plot side-by-side
+                if self.logger.enabled("wandb"):
+                    # Filter valid pairs and limit to 10
+                    valid_pairs = [
+                        (frames, plot)
+                        for frames, plot in zip(video_frames_list, plots)
+                        if frames is not None and plot is not None
+                    ]
+                    valid_pairs = valid_pairs[:10]
+
+                    combined_images = []
+                    for idx, (frames, plot) in enumerate(valid_pairs):
+                        # Convert frames from (T, C, H, W) to (T, H, W, C) for display
+                        if len(frames.shape) == 4 and frames.shape[1] == 3:
+                            frames_rgb = frames.transpose(0, 2, 3, 1)
+                        else:
+                            frames_rgb = frames
+
+                        # Ensure frames are uint8 in [0, 255]
+                        if frames_rgb.max() <= 1.0:
+                            frames_rgb = (frames_rgb * 255).astype(np.uint8)
+                        else:
+                            frames_rgb = np.clip(frames_rgb, 0, 255).astype(np.uint8)
+
+                        # Convert plot to image at original resolution
+                        plot_fig = plot
+                        buf = io.BytesIO()
+                        plot_fig.savefig(buf, format="png", dpi=plot_fig.dpi, bbox_inches="tight")
+                        buf.seek(0)
+                        from PIL import Image
+
+                        plot_img_pil = Image.open(buf)
+                        # Convert RGBA to RGB if necessary
+                        if plot_img_pil.mode == "RGBA":
+                            plot_img_pil = plot_img_pil.convert("RGB")
+                        plot_img = np.array(plot_img_pil)
+                        buf.close()
+
+                        # Get plot dimensions
+                        plot_h, plot_w = plot_img.shape[:2]
+
+                        # Get first and last frames and resize to match plot height
+                        first_frame = frames_rgb[0]
+                        last_frame = frames_rgb[-1]
+
+                        # Calculate width to maintain aspect ratio, or use plot height
+                        frame_h, frame_w = first_frame.shape[:2]
+                        aspect_ratio = frame_w / frame_h
+                        video_h = plot_h
+                        video_w = int(plot_h * aspect_ratio)
+
+                        # Resize first frame
+                        first_frame_pil = Image.fromarray(first_frame)
+                        first_frame_resized = first_frame_pil.resize((video_w, video_h), Image.Resampling.LANCZOS)
+                        first_frame_resized = np.array(first_frame_resized)
+
+                        # Resize last frame
+                        last_frame_pil = Image.fromarray(last_frame)
+                        last_frame_resized = last_frame_pil.resize((video_w, video_h), Image.Resampling.LANCZOS)
+                        last_frame_resized = np.array(last_frame_resized)
+
+                        # Combine first frame, last frame, and plot side-by-side
+                        combined_image = np.hstack([first_frame_resized, last_frame_resized, plot_img])
+                        combined_images.append(combined_image)
+
+                        # Log individual image to wandb
+                        tag = f"reward_alignment_samples/{ds_name}/reward_sample_{idx}"
+                        self.logger.log_image(tag, combined_image, step=eval_step)
+
+                    # Create combined figure with all samples stacked vertically
+                    if combined_images:
+                        # Find the maximum width to ensure all images have the same width
+                        max_width = max(img.shape[1] for img in combined_images)
+                        # Resize all images to have the same width
+                        resized_combined = []
+                        for img in combined_images:
+                            if img.shape[1] != max_width:
+                                img_pil = Image.fromarray(img)
+                                # Maintain aspect ratio by calculating new height
+                                aspect = img.shape[0] / img.shape[1]
+                                new_height = int(max_width * aspect)
+                                img_resized = img_pil.resize((max_width, new_height), Image.Resampling.LANCZOS)
+                                resized_combined.append(np.array(img_resized))
+                            else:
+                                resized_combined.append(img)
+
+                        # Stack all images vertically
+                        combined_figure = np.vstack(resized_combined)
+
+                        # Log combined figure to wandb
+                        combined_tag = f"reward_alignment_samples/{ds_name}/all_samples_combined"
+                        self.logger.log_image(combined_tag, combined_figure, step=eval_step)
+
+            # if rows and self.logger.enabled("wandb"):
+            #     self.logger.log_video_table(
+            #         f"reward_alignment_samples/{ds_name}",
+            #         videos_and_figures=rows,
+            #         columns=["video", "progress_plot"],
+            #         step=eval_step,
+            #     )
+
+            # # Create and log 3x3 grid of videos with progress overlays
+            # if video_frames_list and self.logger.enabled("wandb"):
+            #     grid_video = create_video_grid_with_progress(
+            #         video_frames_list,
+            #         trajectory_progress_data,
+            #         grid_size=(3, 3),
+            #         max_videos=9,
+            #         is_discrete_mode=is_discrete_mode,
+            #     )
+            #     if grid_video is not None:
+            #         self.logger.log_video(
+            #             f"reward_alignment_grid/{ds_name}",
+            #             grid_video,
+            #             fps=2,
+            #             step=eval_step,
+            #         )
+            #         del grid_video
+
+            # For tensorboard (no table support), log each video and its figure separately
+            # if self.logger.enabled("tensorboard"):
+            #     for idx, frames in enumerate(video_frames_list):
+            #         if frames is not None:
+            #             self.logger.log_video(
+            #                 f"reward_alignment_video/{ds_name}/{idx}",
+            #                 frames,
+            #                 fps=2,
+            #                 step=eval_step,
+            #             )
+            #     for idx, plot in enumerate(plots):
+            #         self.logger.log_figure(f"{ds_name}/reward_alignment_plot/{idx}", plot, step=eval_step)
+
+            # Close all plots to avoid accumulating open figures
+            for plot in plots:
+                plt.close(plot)
+
+            # Explicitly delete to free memory and set to None for outer cleanup
+            # log_memory_usage(f"Before deleting plots/videos")
+            del plots, video_frames_list, trajectory_progress_data, rows
+            plots = None
+            video_frames_list = None
+            trajectory_progress_data = None
+            # log_memory_usage(f"After deleting plots/videos")
+        elif eval_type == "policy_ranking":
+            eval_metrics, task_groups, task_details = run_policy_ranking_eval(
+                eval_results,
+                self.config.data.progress_pred_type,
+                is_discrete_mode,
+                num_bins,
+                data_source,
+                correlation_method="kendall",
+            )
+            # log_memory_usage(f"After compute_eval_metrics (policy_ranking)")
+
+            banner(
+                f"{eval_type} evaluation: {len(eval_results)} samples",
+                f"Metrics: {eval_metrics}",
+                inner_padding=1,
+            )
+
+            # Check if any trajectory has partial_success to determine visualization type
+            use_partial_success = False
+            if task_groups:
+                # Check first task group to see if it has partial_success
+                first_group = next(iter(task_groups.values()))
+                if first_group and len(first_group) > 0:
+                    use_partial_success = first_group[0].get("partial_success") is not None
+
+            data = []
+            if use_partial_success:
+                # Visualization for datasets with partial_success: show partial vs predicted rewards for each aggregation type
+                for task, group in task_groups.items():
+                    partial_successes = np.array([t["partial_success"] for t in group]).round(2)
+                    predicted_rewards_last = np.array([t["final_predicted_reward_last"] for t in group]).round(2)
+                    predicted_rewards_avg = np.array([t["final_predicted_reward_avg"] for t in group]).round(2)
+                    predicted_rewards_sum = np.array([t["final_predicted_reward_sum"] for t in group]).round(2)
+                    partial_successes = partial_successes.tolist()
+                    predicted_rewards_last = predicted_rewards_last.tolist()
+                    predicted_rewards_avg = predicted_rewards_avg.tolist()
+                    predicted_rewards_sum = predicted_rewards_sum.tolist()
+                    data.append([
+                        task,
+                        f"partial:{partial_successes}",
+                        f"predicted_last:{predicted_rewards_last}",
+                        f"predicted_avg:{predicted_rewards_avg}",
+                        f"predicted_sum:{predicted_rewards_sum}",
+                    ])
+                columns = [
+                    "task",
+                    "partial_successes",
+                    "predicted_rewards_last",
+                    "predicted_rewards_avg",
+                    "predicted_rewards_sum",
+                ]
+            else:
+                # Standard policy ranking visualization: show quality labels and rewards for each aggregation type
+                for task, group in task_groups.items():
+                    quality_to_rews_last = collections.defaultdict(list)
+                    quality_to_rews_avg = collections.defaultdict(list)
+                    quality_to_rews_sum = collections.defaultdict(list)
+                    for t in group:
+                        rew_last = t["final_predicted_reward_last"]
+                        rew_avg = t["final_predicted_reward_avg"]
+                        rew_sum = t["final_predicted_reward_sum"]
+                        quality_label = t["quality_label"]
+                        quality_to_rews_last[quality_label].append(rew_last)
+                        quality_to_rews_avg[quality_label].append(rew_avg)
+                        quality_to_rews_sum[quality_label].append(rew_sum)
+
+                    for q, r in quality_to_rews_last.items():
+                        quality_to_rews_last[q] = np.array(r).round(2).tolist()
+                    for q, r in quality_to_rews_avg.items():
+                        quality_to_rews_avg[q] = np.array(r).round(2).tolist()
+                    for q, r in quality_to_rews_sum.items():
+                        quality_to_rews_sum[q] = np.array(r).round(2).tolist()
+
+                    quality_to_rews_last_str = ",".join([f"{q}:{r}" for q, r in quality_to_rews_last.items()])
+                    quality_to_rews_avg_str = ",".join([f"{q}:{r}" for q, r in quality_to_rews_avg.items()])
+                    quality_to_rews_sum_str = ",".join([f"{q}:{r}" for q, r in quality_to_rews_sum.items()])
+
+                    # Get task details for differences (using last aggregation for differences)
+                    task_detail = task_details.get(task, {})
+                    succ_subopt = task_detail.get("succ_subopt_diff")
+                    subopt_fail = task_detail.get("subopt_fail_diff")
+                    succ_fail = task_detail.get("succ_fail_diff")
+
+                    # Format differences
+                    diff_str = []
+                    if succ_subopt is not None:
+                        diff_str.append(f"succ-subopt:{succ_subopt:.2f}")
+                    if subopt_fail is not None:
+                        diff_str.append(f"subopt-fail:{subopt_fail:.2f}")
+                    if succ_fail is not None:
+                        diff_str.append(f"succ-fail:{succ_fail:.2f}")
+                    diff_str = ",".join(diff_str) if diff_str else "N/A"
+
+                    data.append([
+                        task,
+                        quality_to_rews_last_str,
+                        quality_to_rews_avg_str,
+                        quality_to_rews_sum_str,
+                        diff_str,
+                    ])
+
+                columns = [
+                    "task",
+                    "quality_and_rews_last",
+                    "quality_and_rews_avg",
+                    "quality_and_rews_sum",
+                    "avg_differences",
+                ]
+
+            table_name = f"policy_ranking_samples/{ds_name}"
+
+            self.logger.log_table(
+                table_name,
+                data=data,
+                columns=columns,
+                step=eval_step,
+            )
+
+            # Save policy ranking samples as JSON metadata
+            # Convert table data (list of lists) to list of dictionaries
+            samples_metadata = []
+            for row in data:
+                sample_dict = {col: val for col, val in zip(columns, row)}
+                samples_metadata.append(sample_dict)
+
+            # Save to policy_ranking_samples folder
+            output_dir = self.args.output_dir
+            samples_dir = os.path.join(output_dir, "policy_ranking_samples", f"step_{eval_step}")
+            os.makedirs(samples_dir, exist_ok=True)
+
+            filename = f"{ds_name}.json"
+            filepath = os.path.join(samples_dir, filename)
+
+            with open(filepath, "w") as f:
+                json.dump(samples_metadata, f, indent=2)
+            logger.info(f"Saved {len(samples_metadata)} policy ranking samples to {filepath}")
+
+            # # Create and log grid of frame pairs with progress annotations
+            # if self.logger.enabled("wandb"):
+            #     grid_image = create_policy_ranking_grid(
+            #         eval_results, grid_size=(2, 2), max_samples=4, is_discrete_mode=is_discrete_mode
+            #     )
+            #     if grid_image is not None:
+            #         self.logger.log_image(
+            #             f"policy_ranking_grid/{ds_name}",
+            #             grid_image,
+            #             step=eval_step,
+            #         )
+            #         del grid_image
+
+            # log_memory_usage(f"Before deleting policy_ranking data")
+            del data, task_groups, task_details
+            task_groups = None
+            task_details = None
+            # log_memory_usage(f"After deleting policy_ranking data")
+        elif eval_type == "confusion_matrix":
+            confusion_plot, confusion_matrix, eval_metrics = run_confusion_matrix_eval(
+                eval_results, self.config.data.progress_pred_type, is_discrete_mode, num_bins
+            )
+            # log_memory_usage(f"After compute_eval_metrics (confusion_matrix)")
+
+            banner(
+                f"{eval_type} evaluation: {len(eval_results)} samples",
+                f"Metrics: {eval_metrics}",
+                inner_padding=1,
+            )
+
+            self.logger.log_figure(f"eval_cm/{ds_name}", confusion_plot, step=eval_step)
+            plt.close(confusion_plot)
+            # log_memory_usage(f"Before deleting confusion_matrix data")
+            del confusion_plot, confusion_matrix
+            confusion_plot = None
+            confusion_matrix = None
+            # log_memory_usage(f"After deleting confusion_matrix data")
+        elif "quality_preference" in eval_type:
+            eval_metrics, task_groups, task_details = run_quality_preference_eval(eval_results, data_source=data_source)
+            # log_memory_usage(f"After compute_eval_metrics (quality_preference)")
+
+            banner(
+                "Completed evaluation",
+                f"{eval_type} evaluation: {len(eval_results)} samples",
+                "Metrics",
+                f"{eval_metrics}",
+                inner_padding=1,
+            )
+
+            data = []
+            for task, details in task_details.items():
+                task_acc = details["preference_accuracy"]
+                quality_accs = details["quality_accuracies"]
+                quality_accs_str = ",".join([f"{k}:{round(v, 3)}" for k, v in quality_accs.items()])
+                num_correct = details["num_correct"]
+                num_total = details["num_total"]
+                data.append([
+                    task,
+                    round(task_acc, 3),
+                    quality_accs_str if quality_accs_str else "N/A",
+                    f"{num_correct}/{num_total}",
+                ])
+            columns = ["task", "preference_accuracy", "quality_accuracies", "num_correct/total"]
+
+            table_name = f"quality_preference_samples/{ds_name}"
+
+            self.logger.log_table(
+                table_name,
+                data=data,
+                columns=columns,
+                step=eval_step,
+            )
+
+            # Save quality preference samples as JSON metadata
+            # Convert table data (list of lists) to list of dictionaries
+            samples_metadata = []
+            for row in data:
+                sample_dict = {col: val for col, val in zip(columns, row)}
+                samples_metadata.append(sample_dict)
+
+            # Save to quality_preference_samples folder
+            output_dir = self.args.output_dir
+            samples_dir = os.path.join(output_dir, "quality_preference_samples", f"step_{eval_step}")
+            os.makedirs(samples_dir, exist_ok=True)
+
+            filename = f"{ds_name}.json"
+            filepath = os.path.join(samples_dir, filename)
+
+            with open(filepath, "w") as f:
+                json.dump(samples_metadata, f, indent=2)
+            logger.info(f"Saved {len(samples_metadata)} quality preference samples to {filepath}")
+
+            # log_memory_usage(f"Before deleting quality_preference data")
+            del data, task_groups, task_details
+            task_groups = None
+            task_details = None
+            # log_memory_usage(f"After deleting quality_preference data")
+        else:
+            raise ValueError(f"Unsupported eval type: {eval_type}")
+
+        # Clean up eval-specific outputs
+        if plots is not None:
+            del plots
+        if video_frames_list is not None:
+            del video_frames_list
+        if trajectory_progress_data is not None:
+            del trajectory_progress_data
+        if task_groups is not None:
+            del task_groups
+        if task_details is not None:
+            del task_details
+        if confusion_plot is not None:
+            del confusion_plot
+        if confusion_matrix is not None:
+            del confusion_matrix
+
+        return eval_metrics
+
+    def _save_eval_results_json(self, eval_results, eval_type, ds_name):
+        """Save eval_results as JSON file.
+
+        Args:
+            eval_results: List of evaluation result dictionaries
+            eval_type: Type of evaluation (e.g., "reward_alignment", "policy_ranking")
+            ds_name: Dataset name
+        """
+
+        def serialize_value(value):
+            """Recursively serialize a value to JSON-compatible format."""
+            if isinstance(value, np.ndarray):
+                return value.tolist()
+            elif isinstance(value, (np.integer, np.floating)):
+                return float(value)
+            elif isinstance(value, (np.bool_, bool)):
+                return bool(value)
+            elif isinstance(value, dict):
+                return {k: serialize_value(v) for k, v in value.items()}
+            elif isinstance(value, list):
+                return [serialize_value(v) for v in value]
+            elif isinstance(value, (int, float, str, type(None))):
+                return value
+            else:
+                # Fallback: try to convert to string
+                return str(value)
+
+        # Serialize eval_results
+        serialized_results = [serialize_value(result) for result in eval_results]
+
+        # Create output directory if it doesn't exist
+        output_dir = self.args.output_dir
+        eval_results_dir = os.path.join(output_dir, "eval_results")
+        os.makedirs(eval_results_dir, exist_ok=True)
+
+        # Create filename: {eval_type}_{ds_name}.json
+        filename = f"{eval_type}_{ds_name}.json"
+        filepath = os.path.join(eval_results_dir, filename)
+
+        # Save to JSON file
+        with open(filepath, "w") as f:
+            json.dump(serialized_results, f, indent=2)
+        logger.info(f"Saved {len(eval_results)} eval results to: {filepath}")
+
+    def _cleanup_eval_dataset(self, dataset, dataloader, eval_results):
+        """Clean up dataset, dataloader, and eval_results after evaluation."""
+        logger.info(f"  [Rank {get_rank()}] Cleaning up dataset and eval_results")
+        # log_memory_usage(f"Before cleanup")
+
+        # Aggressive cleanup to prevent memory leaks
+        # First, delete eval_results which can be large
+        logger.debug(f"  [Rank {get_rank()}] Deleting eval_results")
+        del eval_results
+
+        # For the dataloader, we need to ensure worker processes are shut down
+        # The accelerator.prepare() wraps the dataloader, so we need to clean both
+        # Access the underlying dataloader if it exists and has workers
+        logger.debug(f"  [Rank {get_rank()}] Shutting down dataloader workers")
+        try:
+            if hasattr(dataloader, "_loader"):
+                # Accelerator-wrapped dataloader
+                underlying_dl = dataloader._loader
+            else:
+                underlying_dl = dataloader
+
+            # Shutdown workers if they exist
+            if hasattr(underlying_dl, "_iterator") and underlying_dl._iterator is not None:
+                logger.debug(f"  [Rank {get_rank()}] Calling _shutdown_workers()")
+                underlying_dl._iterator._shutdown_workers()
+                underlying_dl._iterator = None
+                logger.debug(f"  [Rank {get_rank()}] Workers shut down successfully")
+        except (AttributeError, RuntimeError) as e:
+            logger.debug(f"  [Rank {get_rank()}] Could not explicitly shutdown workers: {e}")
+
+        # Delete dataloader and dataset
+        logger.debug(f"  [Rank {get_rank()}] Deleting dataloader and dataset")
+        del dataloader, dataset
+        # log_memory_usage(f"After deleting dataloader and dataset")
+
+        # Force garbage collection
+        import gc
+
+        logger.debug(f"  [Rank {get_rank()}] Running garbage collection")
+        gc.collect()
+        # log_memory_usage(f"After first gc.collect()")
+        gc.collect()  # Call twice for cyclic references
+        # log_memory_usage(f"After second gc.collect()")
+
+        # Clear GPU cache
+        if torch.cuda.is_available():
+            logger.debug(f"  [Rank {get_rank()}] Clearing CUDA cache")
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+        logger.debug(f"  [Rank {get_rank()}] Cleanup complete")
+
+    def _run_single_eval_dataset(self, eval_type, eval_dataset, eval_step):
+        """Run evaluation for a single dataset."""
+        logger.info(f"  Processing dataset: {eval_dataset}")
+        # log_memory_usage(f"Before dataset {eval_dataset}")
+
+        dataset_for_mapping = eval_dataset[0] if isinstance(eval_dataset, list) else eval_dataset
+        ds_name = DS_SHORT_NAME_MAPPING.get(dataset_for_mapping, dataset_for_mapping)
+        timing_key = f"time/eval_dataset/{eval_type}/{ds_name}"
+
+        with _timer(timing_key, timing_raw=self.timing_raw):
+            # Setup dataset and dataloader
+            dataset, dataloader = self._setup_eval_dataset(eval_type, eval_dataset)
+
+            eval_results = []
+            batch_idx = 0
+            # Create tqdm iterator explicitly so we can close it properly
+            dataloader_iter = tqdm(
+                dataloader,
+                desc=f"Running {eval_type}, ds: {eval_dataset}, batch size: {self.config.training.per_device_eval_batch_size}",
+                disable=not is_rank_0(),
+            )
+
+            for batch in dataloader_iter:
+                logger.trace(f"  Processing batch {batch_idx}")
+                # if batch_idx % 10 == 0:  # Log memory every 10 batches
+                #     log_memory_usage(f"Batch {batch_idx}/{len(dataloader)}")
+
+                batch = self._prepare_inputs(batch)
+
+                # Log batch composition
+                num_pref = batch.get("num_preferences", 0)
+                num_prog = batch.get("num_progress", 0)
+                num_sim = batch.get("num_similarities", 0)
+                logger.trace(f"  Batch {batch_idx}: pref={num_pref}, prog={num_prog}, sim={num_sim}")
+                batch_idx += 1
+
+                # Process batch based on eval type
+                if eval_type in ["reward_alignment", "policy_ranking", "confusion_matrix"]:
+                    batch_results, outputs = self._process_batch_progress_eval(batch, eval_type)
+                    eval_results.extend(batch_results)
+                elif "quality_preference" in eval_type:
+                    batch_results, outputs = self._process_batch_preference_eval(batch)
+                    eval_results.extend(batch_results)
+                # Clean up batch tensors and free memory after each batch
+                # Free memory after each batch to prevent OOM
+                del batch, outputs
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
+                # # Log memory after cleanup every 10 batches
+                # if (batch_idx - 1) % 10 == 0:
+                #     log_memory_usage(f"After batch {batch_idx - 1} cleanup")
+
+            # Close tqdm iterator to release any held references
+            dataloader_iter.close()
+            del dataloader_iter
+
+            logger.info(f"  [Rank {get_rank()}] Finished processing {len(eval_results)} eval results")
+            # log_memory_usage(f"After eval loop, before compute_eval_metrics")
+
+            # Compute metrics and create visualizations (only on main process)
+            eval_metrics = {}
+            if self.accelerator.is_main_process:
+                logger.info(f"  [Rank {get_rank()}] Starting metric computation")
+                eval_metrics = self._compute_and_log_eval_metrics(eval_type, eval_results, ds_name, eval_step)
+
+                # Save eval_results as JSON
+                self._save_eval_results_json(eval_results, eval_type, ds_name)
+                logger.info(f"  [Rank {get_rank()}] Finished metric computation")
+            else:
+                logger.info(f"  [Rank {get_rank()}] Skipping metric computation (not main process)")
+
+            # Cleanup
+            logger.info(f"  [Rank {get_rank()}] Starting dataset cleanup")
+            self._cleanup_eval_dataset(dataset, dataloader, eval_results)
+            logger.info(f"  [Rank {get_rank()}] Finished dataset cleanup")
+
+            # log_memory_usage(f"After cleanup for {eval_dataset}")
+
+            # Store timing for this eval_dataset
+            eval_dataset_time = self.timing_raw.get(timing_key, 0.0)
+            logger.info(
+                f"  [Rank {get_rank()}] Finished {eval_type} for {eval_dataset} (took {eval_dataset_time:.2f} seconds)"
+            )
+            logger.info("-" * 80)
+
+            return eval_metrics, ds_name
+
+    def _run_custom_evaluations(self, eval_step=None):
+        """
+        Run custom evaluations.
+
+        Args:
+            eval_step: Step number to use for logging. If None, uses self.state.global_step.
+                      This ensures consistent step logging to prevent wandb warnings.
+        """
+        if eval_step is None:
+            eval_step = self.state.global_step
+
+        logger.info("=" * 100)
+        logger.info("STARTING CUSTOM EVALUATIONS")
+        # log_memory_usage("Before custom evaluations")
+        logger.info("=" * 100)
+
+        metrics = collections.defaultdict(dict)
+        eval_types = self.config.custom_eval.eval_types
+
+        EVAL_TYPE_SHORT = {
+            "reward_alignment": "rew_align",
+            "confusion_matrix": "cm",
+            "policy_ranking": "p_rank",
+            "quality_preference": "pref",
+            "quality_preference_roboarena": "pref_robo",
+        }
+
+        banner("Running custom evaluations", f"Custom evaluations: {eval_types}")
+
+        eval_type_timings = {}
+        eval_dataset_timings = {}
+
+        for eval_type in eval_types:
+            logger.info("=" * 80)
+            logger.info(f"Running evaluation for: {eval_type}")
+            # log_memory_usage(f"Before {eval_type}")
+            logger.info("=" * 80)
+
+            eval_datasets_name = getattr(self.config.custom_eval, eval_type)
+
+            with _timer(f"time/eval_type/{eval_type}", timing_raw=self.timing_raw):
+                for eval_dataset in eval_datasets_name:
+                    eval_metrics, ds_name = self._run_single_eval_dataset(eval_type, eval_dataset, eval_step)
+                    metrics[ds_name][eval_type] = eval_metrics
+
+                    # Store timing for this eval_dataset
+                    dataset_for_mapping = eval_dataset[0] if isinstance(eval_dataset, list) else eval_dataset
+                    ds_name_mapped = DS_SHORT_NAME_MAPPING.get(dataset_for_mapping, dataset_for_mapping)
+                    timing_key = f"time/eval_dataset/{eval_type}/{ds_name_mapped}"
+                    eval_dataset_time = self.timing_raw.get(timing_key, 0.0)
+                    eval_dataset_timings[timing_key] = eval_dataset_time
+
+                # log_memory_usage(f"After completing all datasets for {eval_type}")
+
+            eval_type_time = self.timing_raw.get(f"time/eval_type/{eval_type}", 0.0)
+            eval_type_timings[f"time/eval_type/{eval_type}"] = eval_type_time
+            logger.info(f"Finished eval_type: {eval_type} (took {eval_type_time:.2f} seconds)")
+            logger.info("=" * 80)
+
+        flat_metrics = {}
+        for ds_name, eval_type_metric in metrics.items():
+            for eval_type, metric in eval_type_metric.items():
+                eval_type_short = EVAL_TYPE_SHORT[eval_type]
+                # Add to flat metrics dict with full names
+                for k, v in metric.items():
+                    if isinstance(v, (int, float)):
+                        metric_name = f"eval_{eval_type_short}/{k}_{ds_name}"
+                        flat_metrics[metric_name] = v
+
+        # Prepare metrics for callbacks (all processes should have the same metrics)
+        callback_metrics = flat_metrics
+
+        # Prepare wandb metrics and log (only on main process)
+        if self.accelerator.is_main_process:
+            # Convert callback_metrics to float for wandb logging
+            to_log = {k: float(v) for k, v in callback_metrics.items()}
+            to_log["epoch"] = self.state.epoch
+
+            # Add timing metrics
+            for timing_key, timing_value in eval_type_timings.items():
+                to_log[timing_key] = float(timing_value)
+            for timing_key, timing_value in eval_dataset_timings.items():
+                to_log[timing_key] = float(timing_value)
+
+            self.logger.log_scalars(to_log, step=eval_step)
+
+            # Log timing summary to console
+            if is_rank_0():
+                logger.info("=" * 80)
+                logger.info("Custom Evaluation Timing Summary")
+                logger.info("=" * 80)
+                logger.info("Per eval_type:")
+                for timing_key, timing_value in sorted(eval_type_timings.items()):
+                    logger.info(f"  {timing_key}: {timing_value:.2f} seconds")
+                logger.info("Per eval_dataset:")
+                for timing_key, timing_value in sorted(eval_dataset_timings.items()):
+                    logger.info(f"  {timing_key}: {timing_value:.2f} seconds")
+                logger.info("=" * 80)
+
+        banner("Finished running custom evaluations!")
+        # log_memory_usage("After all evaluations, before cleanup")
+
+        # Reset model to training mode and clear any cached states to prevent leakage
+        self.model.train()
+        # Ensure gradients are cleared before returning to training
+        if hasattr(self, "optimizer") and self.optimizer is not None:
+            self.optimizer.zero_grad(set_to_none=True)
+
+        # Aggressive cleanup to prevent OOM after evaluation
+        import gc
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()  # Call twice
+        gc.collect()
+        gc.collect()
+
+        # Clean up large objects
+        del metrics
+
+        # log_memory_usage("After final cleanup")
+        logger.info("=" * 100)
+        logger.info("FINISHED CUSTOM EVALUATIONS")
+        logger.info("=" * 100)
+
+        # Final synchronization barrier to ensure all processes finish together
+        if dist.is_initialized():
+            logger.info(f"  [Rank {get_rank()}] Waiting at barrier in _run_custom_evaluations")
+            dist.barrier()
+            logger.info(f"  [Rank {get_rank()}] Passed barrier in _run_custom_evaluations")
+
+        return callback_metrics
+
+    def evaluate(self, eval_dataset=None, ignore_keys=None) -> Dict[str, float]:
+        """
+        Override evaluate method to implement custom RBM evaluation metrics.
+        """
+        eval_step = self.state.global_step + 1
+
+        # Save current training mode and set to eval mode
+        was_training = self.model.training
+        self.model.eval()
+        metrics = {}
+
+        # Run evaluation
+        if self.config.training.run_default_eval:
+            # Get the evaluation dataset
+            eval_dataloader = self.get_eval_dataloader(eval_dataset)
+
+            outputs = []
+            with _timer("time/evaluate", timing_raw=self.timing_raw):
+                with torch.no_grad():
+                    for _step, inputs in tqdm(
+                        enumerate(eval_dataloader),
+                        total=len(eval_dataloader),
+                        desc="Evaluating",
+                        disable=not is_rank_0(),
+                    ):
+                        # Move inputs to device
+                        inputs = self._prepare_inputs(inputs)
+
+                        _, loss_dicts = self.compute_loss(self.model, inputs, return_outputs=True, training=False)
+                        outputs.append(loss_dicts)
+
+            # assume that we already called .item() on the outputs
+            keys = list(outputs[0].keys())
+            for key in keys:
+                metrics[key] = [output[key] for output in outputs if key in output]
+                metrics[key] = np.array(metrics[key]).mean()
+
+            # Aggregate metrics across all processes using accelerator
+            metrics = reduce_metrics_with_accelerate(metrics, self.accelerator, aggregate_method="mean")
+            metrics["time/evaluate"] = self.timing_raw["time/evaluate"]
+
+        # Run the custom evaluations
+        custom_eval_should_run = (
+            self.config.training.custom_eval_steps
+            and self.state.global_step % self.config.training.custom_eval_steps == 0
+        )
+        if custom_eval_should_run:
+            with _timer("time/custom_evaluations", timing_raw=self.timing_raw):
+                custom_metrics = self._run_custom_evaluations(eval_step=eval_step)
+
+            metrics.update(custom_metrics)
+            # Add custom evaluation time
+            metrics["time/custom_evaluations"] = self.timing_raw["time/custom_evaluations"]
+
+            if is_rank_0():
+                logger.info(f"Custom evaluations took {self.timing_raw['time/custom_evaluations']:.2f} seconds")
+
+        if metrics:
+            if is_rank_0():
+                banner("Evaluation Results (Aggregated)", inner_padding=1)
+                for key, value in metrics.items():
+                    logger.info(f"{key}: {value:.6f}")
+                logger.info("=" * 50)
+
+            if is_rank_0():
+                self.logger.log_scalars(metrics, step=eval_step)
+
+            # Trigger the callback handler with all metrics
+            self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, metrics)
+
+        # CRITICAL: Final barrier OUTSIDE the if metrics block to ensure ALL ranks wait
+        # This is the absolute final barrier before returning from evaluate(), ensuring no training can start
+        # until all evaluation is completely done, regardless of whether metrics were computed
+        if dist.is_initialized():
+            logger.debug(f"[Rank {get_rank()}] Waiting at final barrier before returning from evaluate()")
+            dist.barrier()
+            logger.debug(f"[Rank {get_rank()}] Passed final barrier, about to return from evaluate()")
+
+        # Restore original training mode to prevent state leakage
+        self.model.train(was_training)
+        # Ensure gradients are cleared before returning to training
+        if hasattr(self, "optimizer") and self.optimizer is not None:
+            self.optimizer.zero_grad(set_to_none=True)
+        # Clear any cached states that might persist from evaluation
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            # Synchronize to ensure all operations are complete
+            torch.cuda.synchronize()
+
+        return metrics
+
+    def compute_loss(self, model, inputs, return_outputs=False, training=True, **kwargs):
+        """Compute loss for separate preference and progress batches."""
+        logger.trace("compute_loss: Starting")
+
+        # Set static graph for DDP on first training step to handle multiple forward passes
+        # Preference and progress losses are computed in separate forward passes.
+        if (
+            training
+            and not self._ddp_static_graph_set
+            and getattr(self.accelerator.gradient_state, "sync_gradients", True)
+            and hasattr(model, "module")
+        ):
+            if hasattr(model.module, "_set_static_graph"):
+                logger.info("Setting DDP static graph mode for multiple forward passes")
+                model.module._set_static_graph()
+                self._ddp_static_graph_set = True
+            elif hasattr(model, "_set_static_graph"):
+                logger.info("Setting DDP static graph mode for multiple forward passes")
+                model._set_static_graph()
+                self._ddp_static_graph_set = True
+
+        # Extract the separate batches
+        preference_inputs = inputs.get("preference_inputs", {})
+        progress_inputs = inputs.get("progress_inputs", {})
+
+        num_preferences = inputs.get("num_preferences", 0)
+        num_progress = inputs.get("num_progress", 0)
+
+        total_loss = 0
+        log_metadata = {}
+
+        logger.trace(f"Num preferences: {num_preferences}, Num progress: {num_progress}")
+
+        # Compute preference loss if we have preference samples
+        if num_preferences > 0 and preference_inputs and self.config.model.train_preference_head:
+            with _timer("time/compute_preference_loss", timing_raw=self.timing_raw):
+                preference_loss, loss_dict = self._compute_preference_loss(
+                    model, preference_inputs, return_outputs=True, training=training
+                )
+                if not torch.isnan(preference_loss).any():
+                    total_loss += preference_loss
+                else:
+                    logger.warning(f"NaN detected in preference loss, replacing with 0.0")
+                log_metadata.update(loss_dict)
+
+        # Compute progress loss if we have progress samples
+        if num_progress > 0 and progress_inputs and self.config.model.train_progress_head:
+            with _timer("time/compute_progress_loss", timing_raw=self.timing_raw):
+                progress_loss, loss_dict = self._compute_progress_loss(
+                    model, progress_inputs, return_outputs=True, training=training
+                )
+                if not torch.isnan(progress_loss).any():
+                    total_loss += progress_loss
+                else:
+                    logger.warning(f"NaN detected in progress loss, replacing with 0.0")
+                log_metadata.update(loss_dict)
+
+        for key, value in log_metadata.items():
+            logger.trace(f"{key}: {value}, type: {type(value)}")
+            if isinstance(value, torch.Tensor):
+                logger.trace(f"\t{key}: shape={value.shape}")
+        # Check for NaN in total loss before returning
+        if torch.isnan(total_loss).any():
+            logger.warning(f"NaN detected in total_loss, replacing with 0.0")
+            total_loss = torch.tensor(0.0, device=total_loss.device, dtype=total_loss.dtype)
+
+        # Always store custom losses for logging (even when return_outputs=False)
+        self.log_metadata = log_metadata
+
+        if return_outputs:
+            # Combine outputs from all loss functions
+            extra_info = {**log_metadata, "total_loss": total_loss.item()}
+            return total_loss, extra_info
+
+        return total_loss
+
+    def _compute_success_loss_helper(
+        self, success_logits, target_progress, success_labels, progress_loss_mask=None, quality_labels=None
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, Dict[str, torch.Tensor]]:
+        """
+        Helper function to compute success prediction loss.
+
+        Computes binary cross-entropy loss for frames with:
+        - progress < min_success (label=0, failure)
+        - progress > max_success (label=1, success)
+        - ignores frames in between
+
+        The loss is automatically balanced by applying a weight to the minority class
+        (whichever has fewer samples - positives or negatives) so both classes contribute equally.
+
+        Args:
+            success_logits: Success prediction logits (can be tensor or list of tensors)
+            target_progress: Target progress tensors (can be tensor or list of tensors)
+            success_labels: Success labels from batch (computed in collator) (can be tensor or list of tensors)
+            progress_loss_mask: Per-sample mask tensor of shape (batch_size,) with 1.0 for samples
+                where we should compute progress/success loss (e.g., successful, rewound, different_task)
+            quality_labels: Optional list of quality labels (e.g., "successful", "suboptimal", "failure") for each sample.
+                If a trajectory has quality_label in ["suboptimal", "failure", "failed"], we always predict success=0
+                and verify that success_labels are all 0s.
+
+        Returns:
+            tuple: (success_loss, success_accuracy, success_auprc, metrics)
+                   The loss is already balanced via per-sample weighting of the minority class
+        """
+        # Get base thresholds from config
+        min_success = self.config.data.min_success
+
+        # Handle Qwen/Molmo downsampling: take every 2nd frame if using Qwen/Molmo and NOT using multi_image
+        # In multi_image mode, we already get one embedding per frame, so no downsampling needed
+        # Ensure success_logits matches target_progress length after downsampling
+        if (
+            "Qwen" in self.config.model.base_model_id or "Molmo" in self.config.model.base_model_id
+        ) and not self.config.data.use_multi_image:
+            success_logits = success_logits[:, ::2]
+            target_progress = target_progress[:, ::2]
+            success_labels = success_labels[:, ::2]
+
+        # Handle suboptimal/failure trajectories: always predict success=0 and verify labels are all 0s
+        # Create a mask for suboptimal/failure trajectories: always include all frames for these trajectories
+        quality_mask = None
+        if quality_labels is not None:
+            batch_size = success_logits.shape[0]
+            seq_len = success_logits.shape[1]
+            quality_mask = torch.zeros(batch_size, seq_len, device=success_logits.device, dtype=torch.float32)
+
+            for i in range(batch_size):
+                quality_label = quality_labels[i]
+                if quality_label is not None and quality_label.lower() in ("suboptimal", "failure", "failed"):
+                    # Verify that success_labels are all 0s for this trajectory
+                    sample_success_labels = success_labels[i]
+                    if not (sample_success_labels == 0.0).all():
+                        logger.debug(
+                            f"Trajectory {i} has quality_label='{quality_label}' but success_labels are not all 0s. "
+                            f"Found non-zero labels: {(sample_success_labels != 0.0).sum().item()} out of {len(sample_success_labels)}"
+                        )
+                        import ipdb
+
+                        ipdb.set_trace()
+
+                    # Include all frames for this trajectory in the mask
+                    quality_mask[i, :] = 1.0
+
+        if self.config.loss.progress_loss_type.lower() == "discrete":
+            target_progress = convert_discrete_target_to_continuous(
+                target_progress, num_bins=self.config.loss.progress_discrete_bins
+            )
+
+        # We predict success for frames where progress < min_success or the frame is a success
+        combined_mask = ((target_progress < min_success) | (success_labels > 0.5)).float()
+
+        # Incorporate quality mask: always include all frames for suboptimal/failure trajectories
+        if quality_mask is not None:
+            combined_mask = torch.maximum(combined_mask, quality_mask)
+
+        # if progress_loss_mask is not None:
+        #     combined_mask = combined_mask * progress_loss_mask
+
+        # Clamp logits to prevent extreme values and gradient issues
+        success_logits = torch.clamp(success_logits, min=-50.0, max=50.0)
+
+        # Compute class counts for balancing
+        num_positives = (success_labels * combined_mask).sum()
+        num_negatives = ((1 - success_labels) * combined_mask).sum()
+
+        # Compute per-sample weights to balance classes
+        # Weight the minority class so both classes contribute equally to the loss
+        # success_loss_weight = max(num_pos, num_neg) / min(num_pos, num_neg)
+        # Applied to whichever class has fewer samples
+        if num_positives > 0 and num_negatives > 0:
+            if num_positives < num_negatives:
+                # Positives are minority - weight them up
+                success_loss_weight = (num_negatives / num_positives).detach()
+                sample_weights = torch.where(
+                    success_labels > 0.5,
+                    success_loss_weight * combined_mask,
+                    combined_mask,
+                )
+            else:
+                # Negatives are minority (or equal) - weight them up
+                success_loss_weight = (num_positives / num_negatives).detach()
+                sample_weights = torch.where(
+                    success_labels > 0.5,
+                    combined_mask,
+                    success_loss_weight * combined_mask,
+                )
+        else:
+            success_loss_weight = torch.tensor(1.0, device=success_logits.device, dtype=success_logits.dtype)
+            sample_weights = combined_mask
+
+        # Compute BCE loss with per-sample weights (includes combined_mask)
+        loss = F.binary_cross_entropy_with_logits(
+            success_logits,
+            success_labels,
+            weight=sample_weights,
+            reduction="none",
+        )
+        combined_mask_index = combined_mask.bool()
+        loss = (loss * combined_mask) / (sample_weights + 1e-8)
+        success_loss = loss[combined_mask_index].mean()
+
+        # Compute accuracy per sample
+        success_preds = (torch.sigmoid(success_logits) > 0.5).float()
+        correct = (success_preds == success_labels).float()
+        masked_correct = correct * combined_mask
+
+        # Compute per-sample positive and negative accuracy tensors (like masked_correct)
+        # Only include values for samples of the corresponding class, use NaN for other class
+        # positive_correct: correct (0 or 1) for positive samples (label==1), NaN for negative samples
+        positive_mask = (success_labels == 1) & (combined_mask > 0)
+        positive_correct_tensor = torch.where(
+            positive_mask, correct, torch.tensor(float("nan"), device=correct.device, dtype=torch.float32)
+        )
+
+        # negative_correct: correct (0 or 1) for negative samples (label==0), NaN for positive samples
+        negative_mask = (success_labels == 0) & (combined_mask > 0)
+        negative_correct_tensor = torch.where(
+            negative_mask, correct, torch.tensor(float("nan"), device=correct.device, dtype=torch.float32)
+        )
+
+        # Compute weighted accuracy (balanced accuracy) - scalar values for logging
+        # Weight each class's accuracy by inverse of its frequency
+        positive_correct_sum = (correct[combined_mask_index] * success_labels[combined_mask_index]).sum()
+        negative_correct_sum = (correct[combined_mask_index] * (1 - success_labels[combined_mask_index])).sum()
+
+        if num_positives > 0 and num_negatives > 0:
+            # Balanced accuracy: average of recall for each class
+            positive_acc_scalar = positive_correct_sum / (num_positives + 1e-8)
+            negative_acc_scalar = negative_correct_sum / (num_negatives + 1e-8)
+            weighted_acc = (positive_acc_scalar + negative_acc_scalar) / 2.0
+        else:
+            weighted_acc = masked_correct.sum() / (combined_mask.sum() + 1e-8)
+            # Set accuracies to 0.0 when we can't compute them properly
+            positive_acc_scalar = torch.tensor(0.0, device=success_loss.device, dtype=torch.float32)
+            negative_acc_scalar = torch.tensor(0.0, device=success_loss.device, dtype=torch.float32)
+
+        success_acc = masked_correct.sum() / (combined_mask.sum() + 1e-8)
+
+        # Compute AUPRC (Area Under Precision-Recall Curve)
+        success_probs = torch.sigmoid(success_logits)
+        success_probs_flat = success_probs[combined_mask > 0]
+        success_labels_flat = success_labels[combined_mask > 0]
+
+        # Compute AUPRC across all valid frames
+        if success_probs_flat.numel() > 0 and len(torch.unique(success_labels_flat)) > 1:
+            auprc = average_precision_score(
+                t2n(success_labels_flat),
+                t2n(success_probs_flat),
+            )
+            batch_auprc = torch.tensor(auprc, device=success_loss.device, dtype=torch.float32)
+        else:
+            batch_auprc = torch.tensor(0.0, device=success_loss.device, dtype=torch.float32)
+
+        metrics = {
+            "masked_correct": masked_correct,
+            "masked_loss": loss,
+            "weighted_accuracy": weighted_acc,
+            "positive_accuracy": positive_correct_tensor,  # Per-sample tensor like masked_correct
+            "negative_accuracy": negative_correct_tensor,  # Per-sample tensor like masked_correct
+            "positive_accuracy_scalar": positive_acc_scalar,  # Scalar for logging
+            "negative_accuracy_scalar": negative_acc_scalar,  # Scalar for logging
+            "success_loss_weight": success_loss_weight,
+            "num_positives": num_positives,
+            "num_negatives": num_negatives,
+        }
+
+        return success_loss, success_acc, batch_auprc, metrics
+
+    def _compute_progress_loss_helper(
+        self,
+        progress_pred: torch.Tensor,
+        target_progress: torch.Tensor,
+        mask: torch.Tensor,
+        predict_last_frame_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Dict[str, torch.Tensor]]:
+        """
+        Helper function to compute progress loss.
+
+        Args:
+            progress_pred: Progress prediction tensors of shape (batch_size, seq_len) for L1/L2 loss,
+                          or (batch_size, seq_len, num_bins) for discrete loss (logits)
+            target_progress: Target progress tensors of shape (batch_size, seq_len) with values in [0, 1]
+            mask: Per-sample mask tensor of shape (batch_size, seq_len) with 1.0 for samples where we should compute loss
+            predict_last_frame_mask: Optional mask tensor of shape (batch_size, seq_len) with 1.0 for frames with partial_success.
+                                    If provided, this takes precedence over config.
+
+        Returns:
+            tuple: (masked_loss, spearman_correlations, metrics)
+        """
+        # Handle Qwen downsampling: take every 2nd frame if using Qwen and NOT using multi_image
+        # In multi_image mode, we already get one embedding per frame, so no downsampling needed
+        if (
+            "Qwen" in self.config.model.base_model_id or "Molmo" in self.config.model.base_model_id
+        ) and not self.config.data.use_multi_image:
+            target_progress = target_progress[:, ::2]
+            mask = mask[:, ::2]
+            if predict_last_frame_mask is not None:
+                predict_last_frame_mask = predict_last_frame_mask[:, ::2]
+
+        # Apply predict_last_frame_mask if provided
+        # The mask defaults to all 1s (include all frames) unless explicitly set to 0
+        if predict_last_frame_mask is not None:
+            # Use the mask from batch inputs (already handles partial_success)
+            # Ensure shapes match
+            if predict_last_frame_mask.shape != mask.shape:
+                # If predict_last_frame_mask is [batch_size, seq_len] and mask is [batch_size, 1], expand mask first
+                if mask.shape[1] == 1 and predict_last_frame_mask.shape[1] > 1:
+                    mask = mask.expand_as(predict_last_frame_mask)
+                elif predict_last_frame_mask.shape[1] == 1 and mask.shape[1] > 1:
+                    predict_last_frame_mask = predict_last_frame_mask.expand_as(mask)
+
+            mask = mask * predict_last_frame_mask
+        elif self.config.loss.predict_last_frame_progress:
+            # Fallback to config-based logic: create a mask that only selects the last frame for each sequence
+            last_frame_mask = torch.zeros_like(target_progress, dtype=torch.float32)
+            last_frame_mask[:, -1] = 1.0  # Set last frame to 1.0 for all sequences
+            mask = mask * last_frame_mask
+
+        # Determine loss type from config
+        loss_type = self.config.loss.progress_loss_type.lower()
+
+        masked_correct = None
+
+        # Set loss function based on loss type
+        if loss_type == "discrete":
+            # Discrete loss: target progress is already binned in data sampling
+            num_bins = self.config.loss.progress_discrete_bins
+
+            # Target progress is already discrete bins [0, num_bins-1] from data sampling
+            # Convert to long tensor
+            if len(target_progress.shape) == 2:
+                target_bins = target_progress.long()  # [batch_size, seq_len]
+                # Ensure bins are in valid range [0, num_bins-1]
+                target_bins = torch.clamp(target_bins, 0, num_bins - 1)
+
+                # progress_pred should be [batch_size, seq_len, num_bins] logits
+                # Reshape for cross-entropy: [batch_size * seq_len, num_bins] and [batch_size * seq_len]
+                batch_size, seq_len = target_bins.shape
+                target_bins_flat = target_bins.view(batch_size * seq_len)  # [B*T]
+            else:
+                target_bins = target_progress
+                # if we're using C51-style soft bins
+                batch_size, seq_len, num_bins = target_bins.shape
+                target_bins_flat = target_bins.view(batch_size * seq_len, num_bins)
+
+            # Check if progress_pred has the correct shape for discrete mode
+            if len(progress_pred.shape) == 2:
+                # Model is outputting [batch_size, seq_len] instead of [batch_size, seq_len, num_bins]
+                # This means the model wasn't configured for discrete mode
+                raise ValueError(
+                    f"Discrete loss requires progress_pred shape [batch_size, seq_len, num_bins], "
+                    f"but got shape {progress_pred.shape}. "
+                    f"The model's progress head may not be configured for discrete mode. "
+                    f"Check that loss.progress_loss_type='discrete' is set before model initialization."
+                )
+
+            if progress_pred.shape[:2] != (batch_size, seq_len) or progress_pred.shape[2] != num_bins:
+                raise ValueError(
+                    f"Shape mismatch: progress_pred has shape {progress_pred.shape}, "
+                    f"but expected [batch_size={batch_size}, seq_len={seq_len}, num_bins={num_bins}]"
+                )
+
+            progress_pred_flat = progress_pred.view(batch_size * seq_len, num_bins)  # [B*T, num_bins]
+            # Mask shape may be [B, 1] or [B, seq_len] depending on downsampling/last_frame_mask
+            # Ensure it matches target_bins shape [B, seq_len] before flattening
+            if mask.shape[1] != seq_len:
+                # Mask is [B, 1], expand to [B, seq_len]
+                mask_expanded = mask.expand(batch_size, seq_len)  # [B, seq_len]
+            else:
+                # Mask is already [B, seq_len]
+                mask_expanded = mask
+            mask_flat = mask_expanded.flatten()  # [B*T]
+
+            # Compute cross-entropy loss per sample
+            loss_per_sample_flat = F.cross_entropy(progress_pred_flat, target_bins_flat, reduction="none")  # [B*T]
+
+            # Compute accuracy: compare predicted bins (argmax) with target bins
+            pred_bins_flat = torch.argmax(progress_pred_flat, dim=-1)  # [B*T]
+            if len(target_bins_flat.shape) == 2:
+                correct_flat = (pred_bins_flat == torch.argmax(target_bins_flat, dim=-1)).float()  # [B*T]
+            else:
+                correct_flat = (pred_bins_flat == target_bins_flat).float()  # [B*T]
+
+            # Apply mask and reshape back
+            masked_loss_flat = loss_per_sample_flat * mask_flat  # [B*T]
+            masked_correct_flat = correct_flat * mask_flat  # [B*T]
+            loss_per_sample = loss_per_sample_flat.view(batch_size, seq_len)  # [B, T]
+            masked_loss = masked_loss_flat.view(batch_size, seq_len)  # [B, T]
+            masked_correct = masked_correct_flat.view(batch_size, seq_len)  # [B, T]
+        elif loss_type == "l1":
+            loss_fn = F.l1_loss
+        else:
+            loss_fn = F.mse_loss
+
+        # Compute loss_per_sample and masked_loss for L1/L2
+        if loss_type != "discrete":
+            loss_per_sample = loss_fn(progress_pred.float(), target_progress.float(), reduction="none")
+            masked_loss = loss_per_sample * mask
+
+        # For discrete mode, convert predictions back to continuous for spearman correlation
+        # For L1/L2, use predictions as-is
+        if loss_type == "discrete":
+            progress_pred_for_corr = convert_bins_to_continuous(progress_pred)
+            target_progress_for_corr = convert_discrete_target_to_continuous(target_progress, num_bins=num_bins)
+        else:
+            progress_pred_for_corr = progress_pred
+            target_progress_for_corr = target_progress
+
+        if mask.shape[1] != target_progress_for_corr.shape[1]:
+            repeated_mask = mask.repeat(1, target_progress_for_corr.shape[1])
+        else:
+            repeated_mask = mask
+        masked_spearman_corr = compute_spearman_correlation(
+            progress_pred_for_corr, target_progress_for_corr, aggregate=False, mask=repeated_mask
+        )
+        masked_spearman_corr = masked_spearman_corr.detach()
+
+        # Average per sample, then take mean across batch
+        # TODO: might need to change this if the mask is per timestep too
+        progress_loss = masked_loss.mean(dim=1).sum(dim=0) / (mask.sum() + 1e-8)
+        spearman_corr = masked_spearman_corr.mean()
+
+        # Keep track of the per-sample metrics
+        metrics = {"masked_loss": masked_loss, "masked_spearman_corr": masked_spearman_corr}
+
+        # Add progress accuracy for discrete mode
+        if loss_type == "discrete" and masked_correct is not None:
+            metrics["masked_progress_accuracy"] = masked_correct
+
+        return progress_loss, spearman_corr, metrics
+
+    def _add_stratified_metrics(
+        self,
+        outputs_dict: Dict[str, Any],
+        prefix: str,
+        strategy_values: Optional[List[str]],
+        data_source_values: List[str],
+        metrics: Dict[str, torch.Tensor],
+        loss_mask: Optional[torch.Tensor] = None,
+    ) -> None:
+        """
+        Add stratified metrics (by strategy and data source) to outputs_dict.
+
+        Args:
+            outputs_dict: Dictionary to update with metrics
+            prefix: Prefix for metric keys (e.g., "train" or "eval")
+            strategy_values: List of strategy values to split by (e.g., data_gen_strategy), or None to skip
+            data_source_values: List of data source values to split by
+            metrics: Dictionary of metric tensors, e.g., {"acc": tensor, "loss": tensor, "margin": tensor}
+            loss_mask: Optional[torch.Tensor] = None,
+        """
+        device = self.accelerator.device
+
+        # Flatten loss_mask to 1D if it has extra dimensions (e.g., [batch_size, 1] -> [batch_size])
+        if loss_mask is not None:
+            loss_mask = loss_mask.squeeze()
+
+        # Split by strategy
+        if strategy_values is not None:
+            strats = set(strategy_values)
+            for strat in strats:
+                mask = torch.tensor(
+                    [1 if s == strat else 0 for s in strategy_values], device=device, requires_grad=False
+                )
+                # Apply mask to each metric and compute mean
+                for metric_name, metric_tensor in metrics.items():
+                    if len(metric_tensor.shape) == 0:
+                        continue
+                    if loss_mask is not None:
+                        masked_metric = metric_tensor[(mask == 1) & (loss_mask == 1)].detach()
+                    else:
+                        masked_metric = metric_tensor[mask == 1].detach()
+
+                    # get mean over non-nan values
+                    non_nan_masked_metric = masked_metric[~torch.isnan(masked_metric)]
+                    if non_nan_masked_metric.numel() > 0:
+                        mean_value = non_nan_masked_metric.mean().item()
+                        outputs_dict[f"{prefix}_strat_{metric_name}/{strat}"] = mean_value
+
+        # Split by data source
+        data_sources = set(data_source_values)
+        for data_source in data_sources:
+            mask = torch.tensor(
+                [1 if s == data_source else 0 for s in data_source_values], device=device, requires_grad=False
+            )
+            # Apply mask to each metric and compute mean
+            for metric_name, metric_tensor in metrics.items():
+                if len(metric_tensor.shape) == 0:
+                    continue
+                if loss_mask is not None:
+                    masked_metric = metric_tensor[(mask == 1) & (loss_mask == 1)].detach()
+                else:
+                    masked_metric = metric_tensor[mask == 1].detach()
+                non_nan_masked_metric = masked_metric[~torch.isnan(masked_metric)]
+                if non_nan_masked_metric.numel() > 0:
+                    mean_value = non_nan_masked_metric.mean().item()
+                    outputs_dict[f"{prefix}_ds_{metric_name}/{data_source}"] = mean_value
+
+    def forward_model(self, model, inputs, sample_type="progress"):
+        """Forward pass for the model."""
+        logger.trace(f"forward_model: Starting forward pass for sample_type={sample_type}")
+
+        with _timer("time/forward", timing_raw=self.timing_raw):
+            if "rewind" in self.config.model.base_model_id:
+                logger.trace("forward_model: Using ReWiND model path")
+                model_output, model_timing_raw = model(
+                    input_ids=inputs.get("input_ids"),
+                    attention_mask=inputs.get("attention_mask"),
+                    pixel_values=inputs.get("pixel_values", None),
+                    pixel_values_videos=inputs.get("pixel_values_videos", None),
+                    video_embeddings=inputs.get("video_embeddings", None),
+                    text_embeddings=inputs.get("text_embeddings", None),
+                    sample_type=sample_type,
+                    timing_raw=self.timing_raw,
+                )
+            else:
+                logger.trace("forward_model: Using Qwen/Molmo/RBM model path, calling model forward")
+                logger.trace(
+                    f"forward_model: input_ids shape: {inputs['input_ids'].shape if 'input_ids' in inputs else 'N/A'}"
+                )
+
+                # Build model kwargs - include both Qwen and Molmo2 specific parameters
+                model_kwargs = {
+                    "input_ids": inputs["input_ids"],
+                    "attention_mask": inputs["attention_mask"],
+                    "pixel_values": inputs.get("pixel_values", None),
+                    "pixel_values_videos": inputs.get("pixel_values_videos", None),
+                    # Qwen-specific parameters
+                    "image_grid_thw": inputs.get("image_grid_thw", None),
+                    "video_grid_thw": inputs.get("video_grid_thw", None),
+                    "second_per_grid_ts": inputs.get("second_per_grid_ts", None),
+                    # Molmo2-specific parameters
+                    "image_grids": inputs.get("image_grids", None),
+                    "image_token_pooling": inputs.get("image_token_pooling", None),
+                    "image_num_crops": inputs.get("image_num_crops", None),
+                    "video_grids": inputs.get("video_grids", None),
+                    "video_token_pooling": inputs.get("video_token_pooling", None),
+                    # Common parameters
+                    "sample_type": sample_type,
+                    "timing_raw": self.timing_raw,
+                }
+                model_output, model_timing_raw = model(**model_kwargs)
+                logger.trace("forward_model: Model forward pass completed")
+
+            logger.trace("forward_model: Updating timing and returning")
+            self.timing_raw.update(model_timing_raw)
+            return model_output, model_timing_raw
+
+    def _compute_progress_loss(self, model, inputs, return_outputs=False, training=True, stratify_by_strategy=True):
+        """
+        Compute progress prediction loss.
+
+        Args:
+            model: The model to use for forward pass
+            inputs: Input dictionary containing progress data
+            return_outputs: Whether to return detailed outputs dict
+            training: Whether in training mode
+            stratify_by_strategy: Whether to stratify metrics by data_gen_strategy (default: True)
+                                 Set to False for single-frame training where strategies aren't used
+        """
+        model_output, _ = self.forward_model(model, inputs, sample_type="progress")
+        progress_logits = model_output.progress_logits
+        progress_pred = progress_logits["A"]
+        progress_target = inputs["target_progress"]
+        progress_target_mask = inputs["target_progress_mask"].unsqueeze(-1)
+        predict_last_frame_mask = inputs["predict_last_frame_mask"]
+
+        progress_loss, spearman_corr, progress_metrics = self._compute_progress_loss_helper(
+            progress_pred, progress_target, progress_target_mask, predict_last_frame_mask=predict_last_frame_mask
+        )
+        final_loss = 0
+
+        final_loss += progress_loss
+        if self.config.model.train_success_head:
+            success_logits = model_output.success_logits
+            success_pred = success_logits["A"]
+            success_labels = inputs["success_labels"]
+
+            quality_labels = inputs.get("quality_labels", None)
+            success_loss, success_accuracy, success_auprc, success_metrics = self._compute_success_loss_helper(
+                success_pred,
+                progress_target,
+                success_labels,
+                progress_loss_mask=progress_target_mask,
+                quality_labels=quality_labels,
+            )
+            # success_loss is already balanced via per-sample weighting of minority class
+            if not torch.isnan(success_loss).any():
+                final_loss += success_loss
+            else:
+                logger.warning(f"NaN detected in success loss")
+
+        # Check for NaN in final loss
+        if torch.isnan(final_loss).any():
+            if training:
+                import ipdb
+
+                ipdb.set_trace()
+            logger.warning(f"NaN detected in progress loss, replacing with 0.0")
+            final_loss = torch.tensor(0.0, device=final_loss.device, dtype=final_loss.dtype)
+
+        if return_outputs:
+            outputs_dict = {}
+
+            prefix = "train" if training else "eval"
+            stratified_metrics = {
+                "spearman_corr": progress_metrics["masked_spearman_corr"],
+                "prog_loss": progress_metrics["masked_loss"],
+            }
+
+            strategy_values = inputs.get("data_gen_strategy") if stratify_by_strategy else None
+            self._add_stratified_metrics(
+                outputs_dict,
+                prefix,
+                strategy_values,
+                inputs["data_source"],
+                stratified_metrics,
+                progress_target_mask,
+            )
+
+            outputs_dict.update({
+                f"{prefix}/prog_loss": progress_loss.item(),
+                f"{prefix}/spearman_corr": spearman_corr.item(),
+            })
+
+            # Add progress accuracy for discrete mode
+            if "masked_progress_accuracy" in progress_metrics:
+                # Expand mask to match masked_progress_accuracy shape [batch_size, seq_len]
+                masked_progress_accuracy = progress_metrics["masked_progress_accuracy"]
+                batch_size, seq_len = masked_progress_accuracy.shape
+                if progress_target_mask.shape[1] != seq_len:
+                    mask_expanded = progress_target_mask.expand(batch_size, seq_len)
+                else:
+                    mask_expanded = progress_target_mask
+                progress_accuracy = masked_progress_accuracy.sum() / (mask_expanded.sum() + 1e-8)
+                outputs_dict[f"{prefix}/prog_accuracy"] = progress_accuracy.item()
+
+            if self.config.model.train_success_head:
+                weighted_accuracy = success_metrics["weighted_accuracy"]
+                positive_accuracy_scalar = success_metrics["positive_accuracy_scalar"]
+                negative_accuracy_scalar = success_metrics["negative_accuracy_scalar"]
+                success_loss_weight = success_metrics["success_loss_weight"]
+                outputs_dict.update({
+                    f"{prefix}/success_loss": success_loss.item(),
+                    f"{prefix}/success_accuracy": success_accuracy.item(),
+                    f"{prefix}/success_auprc": success_auprc.item(),
+                    f"{prefix}/weighted_success_accuracy": weighted_accuracy.item()
+                    if torch.is_tensor(weighted_accuracy)
+                    else weighted_accuracy,
+                    f"{prefix}/positive_success_accuracy": positive_accuracy_scalar.item()
+                    if torch.is_tensor(positive_accuracy_scalar)
+                    else positive_accuracy_scalar,
+                    f"{prefix}/negative_success_accuracy": negative_accuracy_scalar.item()
+                    if torch.is_tensor(negative_accuracy_scalar)
+                    else negative_accuracy_scalar,
+                    f"{prefix}/success_loss_weight": success_loss_weight.item()
+                    if torch.is_tensor(success_loss_weight)
+                    else success_loss_weight,
+                    f"{prefix}/success_num_positives": success_metrics["num_positives"].item(),
+                    f"{prefix}/success_num_negatives": success_metrics["num_negatives"].item(),
+                })
+
+        if not return_outputs:
+            return final_loss
+
+        return final_loss, outputs_dict
+
+    def _compute_preference_loss(self, model, inputs, return_outputs=False, training=True):
+        """Compute preference prediction loss using Bradley-Terry model."""
+        model_outputs, model_timing_raw = self.forward_model(model, inputs, sample_type="preference")
+        progress_logits = model_outputs.progress_logits
+
+        # Get preference labels (1 if first trajectory is preferred, 0 if second trajectory is preferred)
+        preference_labels = inputs["preference_labels"]
+
+        # Get preference scores from the preference head
+        preference_scores = model_outputs.pref_logits.squeeze(-1)  # [batch_size]
+
+        # Clamp logits to prevent extreme values and gradient issues
+        preference_scores = torch.clamp(preference_scores, min=-50.0, max=50.0)
+
+        # Binary cross entropy loss for preference prediction
+        preference_loss_all = F.binary_cross_entropy_with_logits(
+            preference_scores, preference_labels.float(), reduction="none"
+        )
+        preference_loss = preference_loss_all.mean()
+
+        final_loss = 0
+
+        if not torch.isnan(preference_loss).any():
+            final_loss += preference_loss
+        else:
+            logger.warning(f"NaN detected in preference loss")
+
+        # =========================================================================================
+        # Compute progress and success loss for the first trajectory in the paired samples
+        # =========================================================================================
+        target_progress_A = inputs["target_progress_A"]
+        target_progress_A_mask = inputs["target_progress_A_mask"].unsqueeze(-1)
+        data_gen_strat = inputs["trajectory_A_data_gen_strategy"]
+        logger.warning(f"DATA GEN STRAT FOR TRAJ A: {data_gen_strat}")
+        logger.warning(f"DATA SOURCE FOR TRAJ A: {inputs['trajectory_A_data_source']}")
+        # logger.warning(f"PREFERENCE LABELS: {inputs['preference_labels']}")
+
+        if self.config.model.train_progress_head and self.config.training.predict_pref_progress:
+            progress_pred_A = progress_logits["A"]
+            predict_last_frame_mask_A = inputs["predict_last_frame_mask_A"]
+            progress_loss_A, spearman_corr_A, progress_metrics_A = self._compute_progress_loss_helper(
+                progress_pred_A,
+                target_progress_A,
+                target_progress_A_mask,
+                predict_last_frame_mask=predict_last_frame_mask_A,
+            )
+            final_loss += progress_loss_A
+
+        if self.config.model.train_success_head:
+            success_logits = model_outputs.success_logits
+            success_logits = success_logits["A"]
+            success_labels_A = inputs["success_labels_A"]
+
+            quality_labels_A = inputs.get("trajectory_A_quality_label", None)
+            success_loss, success_accuracy, success_auprc, success_metrics_A = self._compute_success_loss_helper(
+                success_logits,
+                target_progress_A,
+                success_labels_A,
+                progress_loss_mask=target_progress_A_mask,
+                quality_labels=quality_labels_A,
+            )
+            # success_loss is already balanced via per-sample weighting of minority class
+            if not torch.isnan(success_loss).any():
+                final_loss += success_loss
+            else:
+                logger.warning(f"NaN detected in success loss")
+
+        # Check for NaN in final loss
+        if torch.isnan(final_loss).any():
+            logger.warning(f"NaN detected in preference loss")
+
+        if return_outputs:
+            outputs_dict = {}
+
+            prefix = "train" if training else "eval"
+            rejected_data_gen_strategy = inputs["rejected_data_gen_strategy"]
+
+            if self.config.model.train_progress_head and self.config.training.predict_pref_progress:
+                outputs_dict.update({
+                    f"{prefix}/pref_prog_loss": progress_loss_A.item(),
+                    f"{prefix}/pref_prog_spearman_corr": spearman_corr_A.item(),
+                })
+
+                # Add progress accuracy for discrete mode
+                if "masked_progress_accuracy" in progress_metrics_A:
+                    # Expand mask to match masked_progress_accuracy shape [batch_size, seq_len]
+                    masked_progress_accuracy = progress_metrics_A["masked_progress_accuracy"]
+                    batch_size, seq_len = masked_progress_accuracy.shape
+                    if target_progress_A_mask.shape[1] != seq_len:
+                        mask_expanded = target_progress_A_mask.expand(batch_size, seq_len)
+                    else:
+                        mask_expanded = target_progress_A_mask
+                    progress_accuracy_A = masked_progress_accuracy.sum() / (mask_expanded.sum() + 1e-8)
+                    outputs_dict[f"{prefix}/pref_prog_accuracy"] = progress_accuracy_A.item()
+
+                stratified_progress_metrics = {
+                    "spearman_corr": progress_metrics_A["masked_spearman_corr"],
+                    "prog_loss": progress_metrics_A["masked_loss"],
+                }
+
+                self._add_stratified_metrics(
+                    outputs_dict,
+                    prefix,
+                    inputs["trajectory_A_data_gen_strategy"],
+                    inputs["trajectory_A_data_source"],
+                    stratified_progress_metrics,
+                    target_progress_A_mask,
+                )
+
+            if self.config.model.train_success_head:
+                weighted_accuracy = success_metrics_A["weighted_accuracy"]
+                positive_accuracy_scalar = success_metrics_A["positive_accuracy_scalar"]
+                negative_accuracy_scalar = success_metrics_A["negative_accuracy_scalar"]
+                success_loss_weight = success_metrics_A["success_loss_weight"]
+                outputs_dict.update({
+                    f"{prefix}/pref_success_loss": success_loss.item(),
+                    f"{prefix}/pref_success_accuracy": success_accuracy.item(),
+                    f"{prefix}/pref_success_auprc": success_auprc.item(),
+                    f"{prefix}/pref_weighted_success_accuracy": weighted_accuracy.item()
+                    if torch.is_tensor(weighted_accuracy)
+                    else weighted_accuracy,
+                    f"{prefix}/pref_positive_success_accuracy": positive_accuracy_scalar.item()
+                    if torch.is_tensor(positive_accuracy_scalar)
+                    else positive_accuracy_scalar,
+                    f"{prefix}/pref_negative_success_accuracy": negative_accuracy_scalar.item()
+                    if torch.is_tensor(negative_accuracy_scalar)
+                    else negative_accuracy_scalar,
+                    f"{prefix}/pref_success_loss_weight": success_loss_weight.item()
+                    if torch.is_tensor(success_loss_weight)
+                    else success_loss_weight,
+                })
+
+                stratified_success_metrics = {
+                    "success_loss": success_metrics_A["masked_loss"],
+                    "success_acc": success_metrics_A["masked_correct"],
+                    "success_pos_acc": success_metrics_A["positive_accuracy"],
+                    "success_neg_acc": success_metrics_A["negative_accuracy"],
+                }
+                self._add_stratified_metrics(
+                    outputs_dict,
+                    prefix,
+                    inputs["trajectory_A_data_gen_strategy"],
+                    inputs["trajectory_A_data_source"],
+                    stratified_success_metrics,
+                    target_progress_A_mask,
+                )
+
+            if preference_loss is not None:
+                # Compute preference accuracy for training monitoring
+                preference_probs = torch.sigmoid(preference_scores)
+                preference_predictions = (preference_probs > 0.5).float()
+                preference_accuracy = (preference_predictions == preference_labels).float()
+
+                # Prepare metrics for stratification
+                stratified_metrics = {
+                    "pref_acc": preference_accuracy,
+                    "pref_loss": preference_loss_all,
+                }
+
+                self._add_stratified_metrics(
+                    outputs_dict,
+                    prefix,
+                    rejected_data_gen_strategy,
+                    inputs["data_source"],
+                    stratified_metrics,
+                )
+
+                outputs_dict.update({
+                    f"{prefix}/preference_loss": preference_loss.item(),
+                    f"{prefix}/preference_accuracy": preference_accuracy.mean().item(),
+                })
+                return final_loss, outputs_dict
+
+        return final_loss
diff --git a/robometer/robometer/trainers/rewind_trainer.py b/robometer/robometer/trainers/rewind_trainer.py
new file mode 100755
index 0000000000000000000000000000000000000000..53a5f3c5db9328d5dc74b40a49c11e78367d4501
--- /dev/null
+++ b/robometer/robometer/trainers/rewind_trainer.py
@@ -0,0 +1,19 @@
+from .rbm_heads_trainer import RBMHeadsTrainer
+from robometer.utils.timer import _timer
+
+
+class ReWiNDTrainer(RBMHeadsTrainer):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward_model(self, model, inputs, sample_type="progress"):
+        """Forward pass for the model."""
+        with _timer("time/forward", timing_raw=self.timing_raw):
+            model_outputs, model_timing_raw = model(
+                video_embeddings=inputs["video_embeddings"],
+                text_embeddings=inputs["text_embeddings"],
+                sample_type=sample_type,
+                timing_raw=self.timing_raw,
+            )
+            self.timing_raw.update(model_timing_raw)
+            return model_outputs, model_timing_raw
diff --git a/robometer/robometer/utils/__init__.py b/robometer/robometer/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/robometer/robometer/utils/config_utils.py b/robometer/robometer/utils/config_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..d866d031f9a8d5b6726b6778278d7df884f020f5
--- /dev/null
+++ b/robometer/robometer/utils/config_utils.py
@@ -0,0 +1,41 @@
+#!/usr/bin/env python3
+"""
+Utility functions for working with Hydra configs and dataclasses.
+"""
+
+from typing import TypeVar, Type
+from omegaconf import OmegaConf, DictConfig
+from rich.console import Console
+
+from robometer.utils.distributed import is_rank_0
+
+T = TypeVar("T")
+
+
+def display_config(cfg):
+    """Display the configuration in a nice Rich format."""
+    if not is_rank_0():
+        return  # Only display config on rank 0
+
+    console = Console()
+    console.print(cfg)
+
+
+def convert_hydra_to_dataclass(cfg: DictConfig, dataclass_type: Type[T]) -> T:
+    """
+    Convert Hydra DictConfig to a dataclass instance.
+
+    Args:
+        cfg: Hydra DictConfig to convert
+        dataclass_type: The dataclass type to convert to (e.g., ExperimentConfig, OfflineEvalConfig)
+
+    Returns:
+        Instance of the specified dataclass type
+    """
+    # Convert to dict and then to dataclass
+    # Use structured config if available, otherwise convert manually
+    if OmegaConf.is_struct(cfg):
+        cfg_dict = OmegaConf.to_container(cfg, resolve=True, structured_config_mode="convert")
+    else:
+        cfg_dict = OmegaConf.to_container(cfg, resolve=True)
+    return dataclass_type(**cfg_dict)
diff --git a/robometer/robometer/utils/distributed.py b/robometer/robometer/utils/distributed.py
new file mode 100755
index 0000000000000000000000000000000000000000..8fbd5cea5787bd1a927a8cd79fda968ef8e0357a
--- /dev/null
+++ b/robometer/robometer/utils/distributed.py
@@ -0,0 +1,273 @@
+import os
+from rich import print as rprint
+from typing import Optional
+
+
+def is_rank_0():
+    """Check if current process is rank 0 (main process)."""
+    # First check environment variables (most reliable for accelerate)
+    if "LOCAL_RANK" in os.environ:
+        return int(os.environ.get("LOCAL_RANK", 0)) == 0
+
+    # Fallback to torch.distributed
+    try:
+        import torch.distributed as dist
+
+        if dist.is_initialized():
+            return dist.get_rank() == 0
+        else:
+            return True  # If not distributed, consider it rank 0
+    except:
+        return True  # If distributed module not available, consider it rank 0
+
+
+def rank_0_print(*args, verbose=True, **kwargs):
+    """Print only if on rank 0."""
+    if is_rank_0() and verbose:
+        rprint(*args, **kwargs)
+
+
+def get_rank():
+    """Get the current rank (process index) in distributed training.
+
+    Returns:
+        int: Rank number, or 0 if not in distributed training or rank cannot be determined.
+    """
+    # First check environment variables (most reliable for accelerate)
+    if "LOCAL_RANK" in os.environ:
+        return int(os.environ.get("LOCAL_RANK", 0))
+
+    # Also check RANK for global rank
+    if "RANK" in os.environ:
+        return int(os.environ.get("RANK", 0))
+
+    # Fallback to torch.distributed
+    try:
+        import torch.distributed as dist
+
+        if dist.is_initialized():
+            return dist.get_rank()
+    except:
+        pass
+
+    # Default to 0 if not distributed
+    return 0
+
+
+def banner(*lines, inner_padding=3):
+    rank_0_print("\n" + "#" * 60)
+
+    # top inner padding
+    for _ in range(inner_padding):
+        rank_0_print("#")
+
+    # content lines
+    for line in lines:
+        rank_0_print("# " + line)
+
+    # bottom inner padding
+    for _ in range(inner_padding):
+        rank_0_print("#")
+
+    rank_0_print("#" * 60 + "\n")
+
+
+def log_fsdp_diagnostics(model, accelerator=None, logger=None):
+    """
+    Log comprehensive FSDP diagnostics including parameter counts, types, and FSDP wrapping status.
+    This helps verify if FSDP is properly configured and working.
+
+    Args:
+        model: The model to check (may be wrapped in FSDP/DDP)
+        accelerator: Optional Accelerate accelerator object to check FSDP plugin configuration
+        logger: Optional logger instance (defaults to loguru logger if not provided)
+    """
+    if not is_rank_0():
+        return
+
+    # Import logger if not provided
+    if logger is None:
+        from robometer.utils.logger import get_logger
+
+        logger = get_logger()
+
+    logger.info("=" * 80)
+    logger.info("FSDP DIAGNOSTICS (After Accelerate/Trainer Wrapping)")
+    logger.info("=" * 80)
+
+    # Check if FSDP is available
+    try:
+        from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+
+        has_fsdp_v1 = True
+    except ImportError:
+        has_fsdp_v1 = False
+        FSDP = None
+
+    try:
+        from torch.distributed._composable.fsdp import FSDPModule
+
+        has_fsdp_v2 = True
+    except ImportError:
+        has_fsdp_v2 = False
+        FSDPModule = None
+
+    logger.info(f"FSDP v1 (FullyShardedDataParallel) available: {has_fsdp_v1}")
+    logger.info(f"FSDP v2 (FSDPModule) available: {has_fsdp_v2}")
+
+    # Unwrap model if it's wrapped in DDP/FSDP for checking
+    # The model passed to training_step might already be unwrapped, but let's check
+    unwrapped_model = model
+    if hasattr(model, "module"):
+        unwrapped_model = model.module
+
+    # Check if model is wrapped in FSDP
+    is_fsdp_wrapped = False
+    fsdp_version = 0
+    if has_fsdp_v1 and isinstance(model, FSDP):
+        is_fsdp_wrapped = True
+        fsdp_version = 1
+        logger.info("Model is wrapped in FSDP v1 (FullyShardedDataParallel)")
+    elif has_fsdp_v2 and isinstance(model, FSDPModule):
+        is_fsdp_wrapped = True
+        fsdp_version = 2
+        logger.info("Model is wrapped in FSDP v2 (FSDPModule)")
+    else:
+        logger.info("Model root is NOT wrapped in FSDP (checking submodules...)")
+
+    # Check for FSDP-wrapped submodules
+    fsdp_wrapped_modules = []
+    if has_fsdp_v1:
+        for name, module in model.named_modules():
+            if isinstance(module, FSDP):
+                fsdp_wrapped_modules.append((name, "FSDP_v1"))
+    if has_fsdp_v2:
+        for name, module in model.named_modules():
+            if isinstance(module, FSDPModule):
+                fsdp_wrapped_modules.append((name, "FSDP_v2"))
+
+    if fsdp_wrapped_modules:
+        logger.info(f"Found {len(fsdp_wrapped_modules)} FSDP-wrapped submodules:")
+        for name, fsdp_type in fsdp_wrapped_modules:
+            logger.info(f"  - {name} ({fsdp_type})")
+    else:
+        logger.info("⚠ No FSDP-wrapped submodules found - FSDP may not be active")
+
+    # Parameter statistics
+    all_params = list(model.parameters())
+    trainable_params = [p for p in all_params if p.requires_grad]
+    frozen_params = [p for p in all_params if not p.requires_grad]
+
+    # Count parameters by device
+    device_counts = {}
+    for param in all_params:
+        device = str(param.device)
+        device_counts[device] = device_counts.get(device, 0) + param.numel()
+
+    logger.info("\nParameter Statistics:")
+    logger.info(f"  Total parameters: {sum(p.numel() for p in all_params):,}")
+    logger.info(f"  Trainable parameters: {sum(p.numel() for p in trainable_params):,}")
+    logger.info(f"  Frozen parameters: {sum(p.numel() for p in frozen_params):,}")
+    if sum(p.numel() for p in all_params) > 0:
+        logger.info(
+            f"  Trainable %: {100 * sum(p.numel() for p in trainable_params) / sum(p.numel() for p in all_params):.4f}%"
+        )
+
+    logger.info("\nParameters by device:")
+    for device, count in sorted(device_counts.items()):
+        total = sum(device_counts.values())
+        if total > 0:
+            logger.info(f"  {device}: {count:,} parameters ({100 * count / total:.2f}%)")
+
+    # Parameter types and dtypes
+    dtype_counts = {}
+    for param in all_params:
+        dtype = str(param.dtype)
+        dtype_counts[dtype] = dtype_counts.get(dtype, 0) + param.numel()
+
+    logger.info("\nParameters by dtype:")
+    for dtype, count in sorted(dtype_counts.items()):
+        total = sum(dtype_counts.values())
+        if total > 0:
+            logger.info(f"  {dtype}: {count:,} parameters ({100 * count / total:.2f}%)")
+
+    # Check for FSDP-specific attributes if wrapped
+    if is_fsdp_wrapped:
+        logger.info("\nFSDP Configuration (Root Model):")
+        if fsdp_version == 1:
+            # FSDP v1 attributes
+            if hasattr(model, "sharding_strategy"):
+                logger.info(f"  Sharding strategy: {model.sharding_strategy}")
+            if hasattr(model, "mixed_precision"):
+                logger.info(f"  Mixed precision: {model.mixed_precision}")
+            if hasattr(model, "cpu_offload"):
+                logger.info(f"  CPU offload: {model.cpu_offload}")
+            if hasattr(model, "use_orig_params"):
+                logger.info(f"  Use orig params: {model.use_orig_params}")
+        elif fsdp_version == 2:
+            # FSDP v2 attributes
+            logger.info("  FSDP v2 detected (composable FSDP)")
+
+    # Check Accelerate FSDP plugin configuration
+    if accelerator is not None:
+        logger.info("\nAccelerate Configuration:")
+        if hasattr(accelerator.state, "fsdp_plugin"):
+            fsdp_plugin = accelerator.state.fsdp_plugin
+            logger.info(f"  FSDP Plugin found: {type(fsdp_plugin).__name__}")
+            if hasattr(fsdp_plugin, "sharding_strategy"):
+                logger.info(f"  Sharding strategy: {fsdp_plugin.sharding_strategy}")
+            if hasattr(fsdp_plugin, "mixed_precision_policy"):
+                logger.info(f"  Mixed precision policy: {fsdp_plugin.mixed_precision_policy}")
+            if hasattr(fsdp_plugin, "cpu_offload_policy"):
+                logger.info(f"  CPU offload policy: {fsdp_plugin.cpu_offload_policy}")
+        else:
+            logger.info("  FSDP Plugin: Not found (FSDP may not be configured)")
+
+    # Check gradient checkpointing status
+    logger.info("\nGradient Checkpointing:")
+    gradient_checkpointing_enabled = False
+    gradient_checkpointing_info = []
+
+    # Check unwrapped model first
+    if hasattr(unwrapped_model, "is_gradient_checkpointing"):
+        gradient_checkpointing_enabled = unwrapped_model.is_gradient_checkpointing
+        gradient_checkpointing_info.append(f"  Root model: {gradient_checkpointing_enabled}")
+    elif hasattr(unwrapped_model, "gradient_checkpointing"):
+        gradient_checkpointing_enabled = unwrapped_model.gradient_checkpointing
+        gradient_checkpointing_info.append(f"  Root model: {gradient_checkpointing_enabled}")
+    else:
+        gradient_checkpointing_info.append("  Root model: is_gradient_checkpointing attribute not found")
+
+    # Check if the underlying model (e.g., base_model) has gradient checkpointing
+    if hasattr(unwrapped_model, "model"):
+        base_model = unwrapped_model.model
+        if hasattr(base_model, "is_gradient_checkpointing"):
+            base_gc = base_model.is_gradient_checkpointing
+            gradient_checkpointing_info.append(f"  Base model: {base_gc}")
+        elif hasattr(base_model, "gradient_checkpointing"):
+            base_gc = base_model.gradient_checkpointing
+            gradient_checkpointing_info.append(f"  Base model: {base_gc}")
+
+    # Check wrapped model as well
+    if hasattr(model, "is_gradient_checkpointing") and model != unwrapped_model:
+        wrapped_gc = model.is_gradient_checkpointing
+        gradient_checkpointing_info.append(f"  Wrapped model: {wrapped_gc}")
+
+    for info in gradient_checkpointing_info:
+        logger.info(info)
+
+    if not gradient_checkpointing_info or all("not found" in info.lower() for info in gradient_checkpointing_info):
+        logger.info("  Note: Could not determine gradient checkpointing status from model attributes")
+
+    # Check distributed environment
+    import torch.distributed as dist
+
+    if dist.is_available() and dist.is_initialized():
+        logger.info(f"\nDistributed Environment:")
+        logger.info(f"  World size: {dist.get_world_size()}")
+        logger.info(f"  Rank: {dist.get_rank()}")
+        logger.info(f"  Backend: {dist.get_backend()}")
+    else:
+        logger.info("\nDistributed Environment: Not initialized (single process)")
+
+    logger.info("=" * 80)
diff --git a/robometer/robometer/utils/embedding_utils.py b/robometer/robometer/utils/embedding_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..b420eba656b2c541418f1c45b2296263df02fdde
--- /dev/null
+++ b/robometer/robometer/utils/embedding_utils.py
@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+
+
+def compute_video_embeddings(
+    frames_array: np.ndarray,
+    dinov2_model,
+    dinov2_processor,
+    batch_size: int = 32,
+    use_autocast: bool = True,
+    use_tqdm: bool = False,
+) -> torch.Tensor:
+    """
+    Compute DINOv2 embeddings for video frames.
+
+    Args:
+        frames_array: Video frames as numpy array (T, H, W, C)
+        dinov2_model: DINOv2 model instance
+        dinov2_processor: DINOv2 processor instance
+        batch_size: Batch size for processing frames
+        use_autocast: Whether to use mixed precision (autocast) for computation
+
+    Returns:
+        Video embeddings as torch tensor (T, D) where D is embedding dimension
+
+    Raises:
+        ValueError: If DINOv2 model or processor is not initialized
+    """
+    if dinov2_model is None or dinov2_processor is None:
+        raise ValueError("DINOv2 model not initialized. Set precompute_embeddings=True in config.")
+
+    device = next(dinov2_model.parameters()).device
+    embeddings = []
+
+    # Process frames in batches
+    num_frames = frames_array.shape[0]
+
+    autocast_context = torch.amp.autocast(device_type="cuda", enabled=use_autocast and torch.cuda.is_available())
+
+    with torch.no_grad(), autocast_context:
+        for i in tqdm(range(0, num_frames, batch_size), disable=not use_tqdm):
+            end_idx = min(i + batch_size, num_frames)
+            batch_frames = frames_array[i:end_idx]
+
+            # Convert numpy frames to PIL images for processing
+            if batch_frames.dtype != np.uint8:
+                batch_frames = (batch_frames * 255).astype(np.uint8)
+
+            pil_images = [Image.fromarray(frame) for frame in batch_frames]
+
+            # Process with DINOv2
+            inputs = dinov2_processor(images=pil_images, return_tensors="pt")
+            inputs = {k: v.to(device, non_blocking=True) for k, v in inputs.items()}
+
+            outputs = dinov2_model(**inputs)
+            # Use pooler_output for global representation
+            batch_embeddings = outputs.pooler_output  # (batch_size, embedding_dim)
+            embeddings.append(batch_embeddings.cpu())
+
+    # Concatenate all embeddings
+    video_embeddings = torch.cat(embeddings, dim=0)  # (T, embedding_dim)
+    return video_embeddings
+
+
+def compute_text_embeddings(
+    text: str,
+    sentence_model,
+    use_autocast: bool = True,
+    show_progress_bar: bool = False,
+) -> torch.Tensor:
+    """
+    Compute sentence transformer embeddings for text.
+
+    Args:
+        text: Text description
+        sentence_model: Sentence transformer model instance
+        use_autocast: Whether to use mixed precision (autocast) for computation
+        show_progress_bar: Whether to show progress bar during encoding
+
+    Returns:
+        Text embedding as torch tensor (D,) where D is embedding dimension
+
+    Raises:
+        ValueError: If sentence model is not initialized
+    """
+    if sentence_model is None:
+        raise ValueError("Sentence model not initialized. Set precompute_embeddings=True in config.")
+
+    autocast_context = torch.amp.autocast(device_type="cuda", enabled=use_autocast and torch.cuda.is_available())
+
+    with torch.no_grad(), autocast_context:
+        embedding = sentence_model.encode(
+            text,
+            convert_to_tensor=True,
+            show_progress_bar=show_progress_bar,
+            batch_size=1,
+        )
+        return embedding.cpu()
diff --git a/robometer/robometer/utils/fsdp/base.py b/robometer/robometer/utils/fsdp/base.py
new file mode 100755
index 0000000000000000000000000000000000000000..d7415892a5cd7e7a5164a487cec48530b20dae9f
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/base.py
@@ -0,0 +1,32 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Sharding manager to implement HybridEngine
+"""
+
+from verl import DataProto
+
+
+class BaseShardingManager:
+    def __enter__(self):
+        pass
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        pass
+
+    def preprocess_data(self, data: DataProto) -> DataProto:
+        return data
+
+    def postprocess_data(self, data: DataProto) -> DataProto:
+        return data
diff --git a/robometer/robometer/utils/fsdp/device.py b/robometer/robometer/utils/fsdp/device.py
new file mode 100755
index 0000000000000000000000000000000000000000..ee9e279d212a94d54c8c31a375022bda1416e06b
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/device.py
@@ -0,0 +1,57 @@
+# Copyright 2025 Bytedance Ltd. and/or its affiliates
+#
+# This code is inspired by the torchtune.
+# https://github.com/pytorch/torchtune/blob/main/torchtune/utils/_device.py
+#
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license in https://github.com/pytorch/torchtune/blob/main/LICENSE
+
+import logging
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+
+def is_torch_npu_available() -> bool:
+    """Check the availability of NPU"""
+    try:
+        import torch_npu  # noqa: F401
+
+        return torch.npu.is_available()
+    except ImportError:
+        return False
+
+
+is_cuda_available = torch.cuda.is_available()
+is_npu_available = is_torch_npu_available()
+
+
+def get_device_name() -> str:
+    """Function that gets the torch.device based on the current machine.
+    This currently only supports CPU, CUDA, NPU.
+    Returns:
+        device
+    """
+    if is_cuda_available:
+        device = "cuda"
+    elif is_npu_available:
+        device = "npu"
+    else:
+        device = "cpu"
+    return device
+
+
+def get_torch_device() -> any:
+    """Return the corresponding torch attribute based on the device type string.
+    Returns:
+        module: The corresponding torch device namespace, or torch.cuda if not found.
+    """
+    device_name = get_device_name()
+    try:
+        return getattr(torch, device_name)
+    except AttributeError:
+        logger.warning(f"Device namespace '{device_name}' not found in torch, try to load torch.cuda.")
+        return torch.cuda
diff --git a/robometer/robometer/utils/fsdp/distributed.py b/robometer/robometer/utils/fsdp/distributed.py
new file mode 100755
index 0000000000000000000000000000000000000000..a7e97f9d3ceed143cdf8a63b374ccf980f75f90d
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/distributed.py
@@ -0,0 +1,48 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Utilities for distributed training."""
+
+import os
+
+import torch.distributed
+from robometer.utils.device import get_torch_device, is_cuda_available
+
+
+def initialize_global_process_group(timeout_second=36000):
+    from datetime import timedelta
+
+    # Check if process group is already initialized
+    if torch.distributed.is_initialized():
+        local_rank = int(os.environ["LOCAL_RANK"])
+        rank = int(os.environ["RANK"])
+        world_size = int(os.environ["WORLD_SIZE"])
+        get_torch_device().set_device(local_rank)
+        return local_rank, rank, world_size
+
+    # Initialize process group if not already initialized
+    torch.distributed.init_process_group(
+        "nccl" if is_cuda_available else "hccl", timeout=timedelta(seconds=timeout_second)
+    )
+    local_rank = int(os.environ["LOCAL_RANK"])
+    rank = int(os.environ["RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+
+    if torch.distributed.is_initialized():
+        get_torch_device().set_device(local_rank)
+    return local_rank, rank, world_size
+
+
+def destroy_global_process_group():
+    if torch.distributed.is_initialized():
+        torch.distributed.destroy_process_group()
diff --git a/robometer/robometer/utils/fsdp/fsdp_ulysses.py b/robometer/robometer/utils/fsdp/fsdp_ulysses.py
new file mode 100755
index 0000000000000000000000000000000000000000..6e5d1b7ac3c49c915f9687ef77eb27b30002ff11
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/fsdp_ulysses.py
@@ -0,0 +1,69 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains a resharding manager that binds weights from FSDP zero3 to XPerfGPT
+"""
+
+from robometer.utils.base import BaseShardingManager
+from robometer.utils.protocol import DataProto, all_gather_data_proto
+from robometer.utils.ulysses import get_ulysses_sequence_parallel_group, set_ulysses_sequence_parallel_group
+from torch.distributed.device_mesh import DeviceMesh
+
+
+class FSDPUlyssesShardingManager(BaseShardingManager):
+    """
+    Sharding manager to support data resharding when using FSDP + Ulysses
+    """
+
+    def __init__(self, device_mesh: DeviceMesh):
+        super().__init__()
+        self.device_mesh = device_mesh
+        self.seed_offset = 12345
+
+    def __enter__(self):
+        if self.device_mesh is not None:
+            # We have a global SP group
+            # so we have to change to use model-specific sp group
+            self.prev_sp_group = get_ulysses_sequence_parallel_group()
+            set_ulysses_sequence_parallel_group(self.device_mesh["sp"].get_group())
+            # TODO: check how to set seed for each model
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        # restore random states
+        if self.device_mesh is not None:
+            # revert to previous sp group
+            set_ulysses_sequence_parallel_group(self.prev_sp_group)
+            # TODO: check how to set seed for each model
+
+    def preprocess_data(self, data: DataProto) -> DataProto:
+        """
+        AllGather data from sp region
+        This is because the data is first sharded along the FSDP dimension as we utilize the DP_COMPUTE
+        In Ulysses, we need to make sure the same data is used across a SP group
+        """
+        if self.device_mesh is not None:
+            group = self.device_mesh["sp"].get_group()
+
+            all_gather_data_proto(data=data, process_group=group)
+        return data
+
+    def postprocess_data(self, data: DataProto) -> DataProto:
+        """
+        Split the data to follow FSDP partition
+        """
+        if self.device_mesh is not None:
+            sp_size = self.device_mesh["sp"].size()
+            sp_rank = self.device_mesh["sp"].get_local_rank()
+            data = data.chunk(chunks=sp_size)[sp_rank]
+        return data
diff --git a/robometer/robometer/utils/fsdp/fsdp_utils.py b/robometer/robometer/utils/fsdp/fsdp_utils.py
new file mode 100755
index 0000000000000000000000000000000000000000..e32a7726cbc5d274b407adad914ca11f0b891adf
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/fsdp_utils.py
@@ -0,0 +1,520 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import itertools
+import json
+import math
+import os
+from collections import OrderedDict
+from contextlib import contextmanager, nullcontext
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from packaging import version
+from torch.distributed import DeviceMesh
+from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp._runtime_utils import _lazy_init
+from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
+from transformers.trainer_pt_utils import get_module_class_from_name
+from verl.utils.device import get_device_name, get_torch_device
+
+if version.parse(torch.__version__) >= version.parse("2.6"):
+    from torch.distributed.fsdp import CPUOffloadPolicy, FSDPModule, MixedPrecisionPolicy, fully_shard
+elif version.parse(torch.__version__) >= version.parse("2.4"):
+    from torch.distributed._composable.fsdp import CPUOffloadPolicy, FSDPModule, MixedPrecisionPolicy, fully_shard
+else:
+    fully_shard, MixedPrecisionPolicy, FSDPModule, CPUOffloadPolicy = None, None, None, None
+
+
+def init_fn(x: torch.nn.Module):
+    if torch.distributed.get_rank() != 0:
+        x = x.to_empty(device=get_torch_device().current_device(), recurse=False)
+        get_torch_device().empty_cache()
+    return x
+
+
+def get_init_weight_context_manager(use_meta_tensor=True, mesh: DeviceMesh = None):
+    from accelerate import init_empty_weights
+
+    def cpu_init_weights():
+        return torch.device("cpu")
+
+    if use_meta_tensor:
+        if mesh is None:
+            init_context = init_empty_weights if torch.distributed.get_rank() != 0 else cpu_init_weights
+        else:
+            init_context = init_empty_weights if mesh.get_coordinate()[-1] != 0 else cpu_init_weights
+    else:
+        init_context = cpu_init_weights
+    return init_context
+
+
+# Copyright 2020-present the HuggingFace Inc. team.
+# Adapted from https://github.com/huggingface/transformers/src/transformers/trainer.py
+def get_fsdp_wrap_policy(module, config=None, is_lora=False):
+    """Get FSDP wrap policy for the module.
+
+    Args:
+        module: The module to get wrap policy for
+        config: Configuration for wrap policy
+        is_lora: Whether to enable lambda policy for LoRA modules
+    """
+    if config is None:
+        config = {}
+
+    # NOTE: This is a temporary workaround to be compatible with the OmegaConf & dataclass. We will remove this once we have make all config in verl from OmegaConf to data class.
+    def _get_attr(attr_name, default_value=None):
+        if hasattr(config, "get"):
+            return config.get(attr_name, default_value)
+        else:
+            return config.__getattribute__(attr_name)
+
+    if _get_attr("disable", False):
+        return None
+
+    default_transformer_cls_names_to_wrap = getattr(module, "_no_split_modules", None)
+    fsdp_transformer_layer_cls_to_wrap = _get_attr(
+        "transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap
+    )
+    min_num_params = _get_attr("min_num_params", 0)
+    auto_wrap_policy = None
+
+    policies = []
+
+    from torch.distributed.fsdp.wrap import _or_policy, lambda_auto_wrap_policy
+
+    # Add lambda policy for LoRA modules if is_lora is True
+    if is_lora:
+
+        def lambda_policy_fn(module):
+            return bool(
+                len(list(module.named_children())) == 0
+                and getattr(module, "weight", None) is not None
+                and module.weight.requires_grad
+            )
+
+        lambda_policy = functools.partial(lambda_auto_wrap_policy, lambda_fn=lambda_policy_fn)
+        policies.append(lambda_policy)
+
+    if min_num_params > 0:
+        size_policy = functools.partial(size_based_auto_wrap_policy, min_num_params=min_num_params)
+        policies.append(size_policy)
+    elif fsdp_transformer_layer_cls_to_wrap is not None:
+        transformer_cls_to_wrap = set()
+        for layer_class in fsdp_transformer_layer_cls_to_wrap:
+            transformer_cls = get_module_class_from_name(module, layer_class)
+            if transformer_cls is None:
+                raise Exception("Could not find the transformer layer class to wrap in the model.")
+            else:
+                transformer_cls_to_wrap.add(transformer_cls)
+
+        print(f"transformer_cls_to_wrap: {transformer_cls_to_wrap}")
+        transformer_policy = functools.partial(
+            transformer_auto_wrap_policy,
+            transformer_layer_cls=transformer_cls_to_wrap,
+        )
+        policies.append(transformer_policy)
+
+    if len(policies) > 0:
+        auto_wrap_policy = functools.partial(_or_policy, policies=policies)
+
+    return auto_wrap_policy
+
+
+@torch.no_grad()
+def offload_fsdp_model_to_cpu(model: FSDP, empty_cache: bool = True):
+    if fsdp_version(model) == 2:
+        offload_fsdp2_model_to_cpu(model, empty_cache)
+        return
+
+    assert isinstance(model, FSDP)
+    # lazy init FSDP model
+    _lazy_init(model, model)
+    assert model._is_root, "Only support root model offloading to CPU"
+    for handle in model._all_handles:
+        if handle._offload_params:
+            continue
+        flat_param = handle.flat_param
+        assert (
+            flat_param.data.data_ptr() == flat_param._local_shard.data_ptr()
+            and id(flat_param.data) != id(flat_param._local_shard)
+            and flat_param.data.size() == flat_param._local_shard.size()
+        )
+        handle.flat_param_to(torch.device("cpu"), non_blocking=True)
+        # the following still keeps id(._local_shard) != id(.data)
+        flat_param._local_shard = flat_param.data
+        assert id(flat_param._local_shard) != id(flat_param.data)
+    if empty_cache:
+        get_torch_device().empty_cache()
+
+
+@torch.no_grad()
+def offload_fsdp2_model_to_cpu(model, empty_cache: bool = True):
+    for param in model.parameters():
+        param.data = param.data.to(torch.device("cpu"), non_blocking=True)
+    if empty_cache:
+        get_torch_device().empty_cache()
+
+
+@torch.no_grad()
+def load_fsdp_model_to_gpu(model: FSDP):
+    if fsdp_version(model) == 2:
+        load_fsdp2_model_to_gpu(model)
+        return
+
+    assert isinstance(model, FSDP)
+    # lazy init FSDP model
+    _lazy_init(model, model)
+    assert model._is_root, "Only support root model loading to GPU"
+    device_id = get_torch_device().current_device()
+    for handle in model._all_handles:
+        if handle._offload_params:
+            continue
+        flat_param = handle.flat_param
+        handle.flat_param_to(torch.device(f"{get_device_name()}:{device_id}"), non_blocking=True)
+        # the following still keeps id(._local_shard) != id(.data)
+        flat_param._local_shard = flat_param.data
+
+
+@torch.no_grad()
+def load_fsdp2_model_to_gpu(model):
+    device = torch.cuda.current_device()
+    for param in model.parameters():
+        param.data = param.data.to(device, non_blocking=True)
+
+
+@torch.no_grad()
+def offload_fsdp_optimizer(optimizer):
+    if not optimizer.state:
+        return
+    for param_group in optimizer.param_groups:
+        for param in param_group["params"]:
+            state = optimizer.state[param]
+            for key, value in state.items():
+                if isinstance(value, torch.Tensor):
+                    state[key] = value.to("cpu", non_blocking=True)
+
+
+@torch.no_grad()
+def load_fsdp_optimizer(optimizer, device_id):
+    if not optimizer.state:
+        return
+    for param_group in optimizer.param_groups:
+        for param in param_group["params"]:
+            state = optimizer.state[param]
+            for key, value in state.items():
+                if isinstance(value, torch.Tensor):
+                    state[key] = value.to(device_id, non_blocking=True)
+
+
+@contextmanager
+def meta_device_init():
+    """
+    Create model parameters with meta device.
+
+    Note buffers in model will still be initialized in default device (e.g., CPU),
+    since the buffers can be non-persistent and filled with expected values that can
+    NOT be captured in meta device.
+    """
+    device = torch.device("meta")
+    old_register_parameter = nn.Module.register_parameter
+    registered = set()
+
+    def register_empty_parameter(module, name, param):
+        old_register_parameter(module, name, param)
+        # we will skip register shared parameters as it
+        # is already registered previously
+        if param is not None and param not in registered:
+            param_cls = type(module._parameters[name])
+            kwargs = module._parameters[name].__dict__
+            kwargs["requires_grad"] = param.requires_grad
+            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
+            registered.add(module._parameters[name])
+
+    try:
+        nn.Module.register_parameter = register_empty_parameter
+        yield
+    finally:
+        registered.clear()
+        nn.Module.register_parameter = old_register_parameter
+
+
+def parallel_load_safetensors(filepath):
+    """
+    Parallel load safetensors from huggingface checkpoint
+
+    Huggingface checkpoint contains:
+
+    - config.json: a json file for model configuration
+    - model.safetensor.index.json: a json file for safetensors (parameters & buffers) index
+    - model-000x-of-ooxx.safetensors: a binary file for safetensors (parameters & buffers) chunks
+
+    Or (when model is small),
+
+    - model.safetensors: a binary file for all parameters and buffers
+
+    Each rank will own a part of model chunks and load them directly into GPU memory.
+    """
+    from safetensors.torch import load_file
+
+    safetensors2param = {}
+
+    index_file = os.path.join(filepath, "model.safetensors.index.json")
+    if os.path.exists(index_file):
+        index = json.load(open(index_file, "rb"))
+        for param_name, filename in index["weight_map"].items():
+            safetensors2param.setdefault(filename, []).append(param_name)
+    else:
+        # in this case, the model is small and we can load it all at once
+        param_file = os.path.join(filepath, "model.safetensors")
+        assert os.path.exists(param_file), f"Cannot find {param_file}"
+        states = load_file(param_file)
+        for param_name in states:
+            safetensors2param.setdefault("model.safetensors", []).append(param_name)
+        del states
+
+    total_files = len(safetensors2param)
+    ckpt_chunks = sorted(safetensors2param.keys())
+    world_size = dist.get_world_size()
+    size = math.ceil(total_files / world_size)
+    ckpt_chunks = [ckpt_chunks[rank * size : rank * size + size] for rank in range(world_size)]
+
+    shard_states = {}
+    device = get_torch_device().current_device()
+    for rank, files in enumerate(ckpt_chunks):
+        if rank == dist.get_rank():
+            for file in files:
+                file = os.path.join(filepath, file)
+                states = load_file(file, device=device)
+                # print(f"rank {rank} loading {file}...")
+                shard_states.update(states)
+        else:
+            for file in files:
+                for param_name in safetensors2param[file]:
+                    shard_states[param_name] = rank
+    return shard_states
+
+
+def parallel_init_module_fn(module: torch.nn.Module, shard_states: dict[str, torch.nn.Parameter]):
+    """
+    Generate a function to initialize sub-modules in the `module` with `shard_states`
+    from huggingface checkpoint.
+
+    Args:
+        module (torch.nn.Module): the global module to be initialized
+        shard_states (Dict[str, torch.nn.Parameter]): the shard states from huggingface checkpoint
+
+    Returns:
+        init_fn (Callable): a function to initialize sub-modules in the `module` with `shard_states`
+    """
+
+    state2fqn = {}
+    for name, state in itertools.chain(
+        module.named_parameters(remove_duplicate=False), module.named_buffers(remove_duplicate=False)
+    ):
+        state2fqn.setdefault(state, []).append(name)
+    # remove standalone parameters and buffers
+    shared = {s for s, names in state2fqn.items() if len(names) > 1}
+    materialized_states = {}
+
+    @torch.no_grad()
+    def create_and_sync_state(param_name, state, is_param):
+        assert param_name in shard_states, f"{param_name} not loaded"
+        device = get_torch_device().current_device()
+        if is_param:
+            param = torch.nn.Parameter(torch.empty_like(state.data, device=device), requires_grad=state.requires_grad)
+        else:  # buffer
+            param = torch.empty_like(state.data, device=device)
+        loaded = shard_states[param_name]
+        if isinstance(loaded, (torch.nn.Parameter, torch.Tensor)):
+            # NOTE: loaded.dtype can be different with param.dtype
+            param.data.copy_(loaded.data)
+            dist.broadcast(param.data, src=dist.get_rank())
+        else:
+            assert isinstance(loaded, int)  # the rank that holds the state
+            dist.broadcast(param.data, src=loaded)
+        shard_states.pop(param_name)
+        del loaded
+        return param
+
+    def init_fn(sub_mod: torch.nn.Module, recurse: bool = True):
+        param_and_buffers = tuple(sub_mod.named_parameters(recurse=False)) + tuple(sub_mod.named_buffers(recurse=False))
+        # param_and_buffers = sorted(sub_mod.named_parameters(recurse=False), key=lambda x: x[0])
+        for name, state in param_and_buffers:
+            if not state.is_meta:
+                continue
+            is_param = name in sub_mod._parameters
+            fqn = state2fqn[state].pop(0)
+            # non-persistent buffers will not be saved in state dict, we can safely skip it
+            if (not is_param) and fqn not in shard_states:
+                if state.is_meta:
+                    raise RuntimeError(
+                        f"find a non-persistent buffer ({fqn}) initiated with device meta. Such buffer is not saved in checkpoint and user should guarantee to init in CPU / GPU device."
+                    )
+                continue
+            # for shared parameter, we get it from the first time it is created
+            if state in shared:
+                if state not in materialized_states:
+                    materialized_states[state] = create_and_sync_state(fqn, state, is_param)
+                else:
+                    if fqn in shard_states:
+                        shard_states.pop(fqn)
+                materialize_state = materialized_states[state]
+            # for not shared parameter, we create it directly
+            else:
+                materialize_state = create_and_sync_state(fqn, state, is_param)
+            if is_param:
+                sub_mod._parameters[name] = materialize_state
+            else:
+                sub_mod._buffers[name] = materialize_state
+        if recurse:
+            for module in sub_mod.children():
+                init_fn(module, recurse=True)
+
+        # for debug
+        # if len(shard_states) == 0: print("clear")
+        return sub_mod
+
+    return init_fn
+
+
+def fsdp_version(model):
+    if isinstance(model, FSDP):
+        return 1
+    elif isinstance(model, FSDPModule):
+        return 2
+    else:
+        return 0
+
+
+def get_fsdp_state_ctx(model, state_type, state_cfg, optim_cfg):
+    if fsdp_version(model) == 1:
+        return FSDP.state_dict_type(model, state_type, state_cfg, optim_cfg)
+    else:
+        return nullcontext()
+
+
+def fsdp2_load_full_state_dict(model: torch.nn.Module, full_state: dict, device_mesh=None, cpu_offload=None):
+    """
+    Loads the full state dict (could be only on rank 0) into the sharded model. This is done by broadcasting the
+    parameters from rank 0 to all other ranks. This function modifies the model in-place.
+
+    Args:
+        model (`torch.nn.Module`): The model to load the state dict into
+        full_state (`dict`): The full state dict to load, can only be on rank 0
+    """
+    from torch.distributed.checkpoint.state_dict import StateDictOptions, set_model_state_dict
+
+    # To broadcast, it needs to be instantiated in the GPU.
+    if dist.get_rank() == 0:
+        model = model.to(device=torch.cuda.current_device(), non_blocking=True)
+    else:
+        model = model.to_empty(device=torch.cuda.current_device())
+
+    cpu_offload = cpu_offload is not None
+    options = StateDictOptions(full_state_dict=True, cpu_offload=cpu_offload, broadcast_from_rank0=True)
+    set_model_state_dict(model, full_state, options=options)
+
+    # rotary_emb is not in state_dict, so we need to broadcast it manually
+    for _name, buf in model.named_buffers():
+        dist.broadcast(buf, src=0)
+
+    if cpu_offload:
+        model.to("cpu", non_blocking=True)
+        for buf in model.buffers():
+            buf.data = buf.data.to(torch.cuda.current_device())
+
+
+def apply_fsdp2(model, fsdp_kwargs, config):
+    """model: AutoModelForCausalLM"""
+    assert CPUOffloadPolicy is not None, "PyTorch version >= 2.4 is required for using fully_shard API (FSDP2)"
+
+    default_transformer_cls_names_to_wrap = getattr(model, "_no_split_modules", None)
+    fsdp_transformer_layer_cls_to_wrap = config.get("wrap_policy", {}).get(
+        "transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap
+    )
+
+    if isinstance(fsdp_transformer_layer_cls_to_wrap, str):
+        fsdp_transformer_layer_cls_to_wrap = [fsdp_transformer_layer_cls_to_wrap]
+
+    assert len(fsdp_transformer_layer_cls_to_wrap) > 0 and fsdp_transformer_layer_cls_to_wrap[0] is not None
+
+    modules = []
+    for _name, module in model.named_modules():
+        if module.__class__.__name__ in fsdp_transformer_layer_cls_to_wrap or (
+            isinstance(module, nn.Embedding) and not model.config.tie_word_embeddings
+        ):
+            modules.append(module)
+
+    for _idx, module in enumerate(modules):
+        fully_shard(module, **fsdp_kwargs)
+    fully_shard(model, **fsdp_kwargs)  # fsdp2 will not reshard_after_forward for root module
+
+
+def fsdp2_clip_grad_norm_(parameters, max_norm, norm_type=2.0, error_if_nonfinite=False, foreach=None):
+    """torch.nn.utils.clip_grad_norm_ cann't run on cpu parameter DTensor"""
+    from torch.nn.utils.clip_grad import _clip_grads_with_norm_, _get_total_norm
+
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    else:
+        # prevent generators from being exhausted
+        parameters = list(parameters)
+    grads = [p.grad for p in parameters if p.grad is not None]
+    total_norm = _get_total_norm(grads, norm_type, error_if_nonfinite, foreach)
+    total_norm = total_norm.to(torch.cuda.current_device(), non_blocking=True)
+    _clip_grads_with_norm_(parameters, max_norm, total_norm, foreach)
+    return total_norm
+
+
+def layered_summon_lora_params(fsdp_module) -> OrderedDict:
+    from peft.utils.save_and_load import get_peft_model_state_dict
+
+    def __prefix_submodules(module, prefix):
+        for name, submodule in module.named_modules():
+            if name.startswith(prefix) and "." not in name[len(prefix) :]:
+                yield name, submodule
+
+    lora_params = OrderedDict()
+    prefix_list = [
+        # fsdp
+        "_fsdp_wrapped_module.base_model.model.",
+        "_fsdp_wrapped_module.base_model.model.model.",
+        "_fsdp_wrapped_module.base_model.model.model.layers.",
+        # fsdp2
+        "base_model.model.",
+        "base_model.model.model.",
+        "base_model.model.model.layers.",
+    ]
+    peft_model = getattr(fsdp_module, "_fsdp_wrapped_module", fsdp_module)
+    for prefix in prefix_list:
+        for name, submodule in __prefix_submodules(fsdp_module, prefix):
+            prefix = name.replace("_fsdp_wrapped_module.base_model.model.", "base_model.model.")
+            if name.endswith(".model") or name.endswith(".layers"):
+                continue
+            if fsdp_version(submodule) > 0:
+                with FSDP.summon_full_params(submodule, writeback=False):
+                    sub_lora_params = get_peft_model_state_dict(peft_model, state_dict=submodule.state_dict())
+                    sub_lora_params = {
+                        f"{prefix}.{name}": param.full_tensor().detach().cpu()
+                        if hasattr(param, "full_tensor")
+                        else param.detach().cpu()
+                        for name, param in sub_lora_params.items()
+                    }
+                    lora_params.update(sub_lora_params)
+                    submodule._is_root = False
+                torch.cuda.empty_cache()
+    return lora_params
diff --git a/robometer/robometer/utils/fsdp/protocol.py b/robometer/robometer/utils/fsdp/protocol.py
new file mode 100755
index 0000000000000000000000000000000000000000..5cc7b42e106dfe5a2f92574a80f73306b19e2305
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/protocol.py
@@ -0,0 +1,945 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Implement base data transfer protocol between any two functions, modules.
+We can subclass Protocol to define more detailed batch info with specific keys
+"""
+
+import contextlib
+import copy
+import logging
+import os
+import pickle
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import numpy as np
+import pandas as pd
+import ray
+import tensordict
+import torch
+import torch.distributed
+from packaging import version
+from tensordict import TensorDict
+from torch.utils.data import DataLoader
+from verl.utils.device import get_torch_device
+from verl.utils.py_functional import union_two_dict
+from verl.utils.torch_functional import allgather_dict_tensors
+
+__all__ = ["DataProto", "union_tensor_dict"]
+
+with contextlib.suppress(Exception):
+    tensordict.set_lazy_legacy(False).set()
+
+
+class _DataProtoConfigMeta(type):
+    _config = {}
+
+    auto_padding_key = "_verl_auto_padding"
+
+    @property
+    def auto_padding(cls):
+        enabled_by_env = os.getenv("VERL_AUTO_PADDING", "FALSE").upper() in ["TRUE", "1"]
+        return enabled_by_env or cls._config.get(cls.auto_padding_key, False)
+
+    @auto_padding.setter
+    def auto_padding(cls, enabled: bool):
+        assert isinstance(enabled, bool), f"enabled must be a boolean, got {enabled} as {type(enabled)}"
+        cls._config[cls.auto_padding_key] = enabled
+
+
+class DataProtoConfig(metaclass=_DataProtoConfigMeta):
+    pass
+
+
+_padding_size_key = "_padding_size_key_x123d"
+
+
+def pad_dataproto_to_divisor(data: "DataProto", size_divisor: int):
+    """Pad a DataProto to size divisible by size_divisor
+
+    Args:
+        size_divisor (int): size divisor
+
+    Returns:
+        data: (DataProto): the padded DataProto
+        pad_size (int)
+    """
+    assert isinstance(data, DataProto), "data must be a DataProto"
+    if len(data) % size_divisor != 0:
+        pad_size = size_divisor - len(data) % size_divisor
+        padding_protos = []
+        remaining_pad = pad_size
+        while remaining_pad > 0:
+            take_size = min(remaining_pad, len(data))
+            padding_protos.append(data[:take_size])
+            remaining_pad -= take_size
+        data_padded = DataProto.concat([data, *padding_protos])
+    else:
+        if len(data) == 0:
+            logging.warning("padding a DataProto with no item, no changed made")
+        pad_size = 0
+        data_padded = data
+    return data_padded, pad_size
+
+
+def unpad_dataproto(data: "DataProto", pad_size):
+    if pad_size != 0:
+        data = data[:-pad_size]
+    return data
+
+
+def union_tensor_dict(tensor_dict1: TensorDict, tensor_dict2: TensorDict) -> TensorDict:
+    """Union two tensordicts."""
+    assert tensor_dict1.batch_size == tensor_dict2.batch_size, (
+        f"Two tensor dict must have identical batch size. Got {tensor_dict1.batch_size} and {tensor_dict2.batch_size}"
+    )
+    for key in tensor_dict2.keys():
+        if key not in tensor_dict1.keys():
+            tensor_dict1[key] = tensor_dict2[key]
+        else:
+            assert tensor_dict1[key].equal(tensor_dict2[key]), (
+                f"{key} in tensor_dict1 and tensor_dict2 are not the same object"
+            )
+
+    return tensor_dict1
+
+
+def union_numpy_dict(tensor_dict1: dict[str, np.ndarray], tensor_dict2: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    for key, val in tensor_dict2.items():
+        if key in tensor_dict1:
+            assert isinstance(tensor_dict2[key], np.ndarray)
+            assert isinstance(tensor_dict1[key], np.ndarray)
+            # to properly deal with nan and object type
+            assert pd.DataFrame(tensor_dict2[key]).equals(pd.DataFrame(tensor_dict1[key])), (
+                f"{key} in tensor_dict1 and tensor_dict2 are not the same object"
+            )
+        tensor_dict1[key] = val
+
+    return tensor_dict1
+
+
+def list_of_dict_to_dict_of_list(list_of_dict: list[dict]):
+    if len(list_of_dict) == 0:
+        return {}
+    keys = list_of_dict[0].keys()
+    output = {key: [] for key in keys}
+    for data in list_of_dict:
+        for key, item in data.items():
+            assert key in output
+            output[key].append(item)
+    return output
+
+
+def fold_batch_dim(data: "DataProto", new_batch_size):
+    """
+    Fold a batch dim from [bsz, xxx] into [new_bsz, bsz // new_bsz, xxx]
+    """
+    batch_size = data.batch.batch_size[0]
+
+    assert batch_size % new_batch_size == 0
+
+    tensor: TensorDict = data.batch
+    non_tensor = data.non_tensor_batch
+
+    tensor = tensor.view(new_batch_size, -1)
+    tensor.auto_batch_size_(batch_dims=1)
+
+    for key, val in non_tensor.items():
+        non_tensor[key] = np.reshape(val, newshape=(new_batch_size, -1, *val.shape[1:]))
+
+    return type(data)(batch=tensor, non_tensor_batch=non_tensor, meta_info=data.meta_info)
+
+
+def unfold_batch_dim(data: "DataProto", batch_dims=2):
+    """
+    Unfold the first n dims as new batch dim
+    """
+    tensor: TensorDict = data.batch
+    non_tensor = data.non_tensor_batch
+    tensor.auto_batch_size_(batch_dims=batch_dims)
+    tensor = tensor.view(-1)
+
+    batch_size = tensor.batch_size[0]
+
+    non_tensor_new = {}
+
+    for key, val in non_tensor.items():
+        non_tensor_new[key] = np.reshape(val, newshape=(batch_size, *val.shape[batch_dims:]))
+
+    return type(data)(batch=tensor, non_tensor_batch=non_tensor_new, meta_info=data.meta_info)
+
+
+def collate_fn(x: list["DataProtoItem"]):
+    batch = []
+    non_tensor_batch = []
+    for data in x:
+        batch.append(data.batch)
+        non_tensor_batch.append(data.non_tensor_batch)
+    batch = torch.stack(batch).contiguous()
+    non_tensor_batch = list_of_dict_to_dict_of_list(non_tensor_batch)
+    for key, val in non_tensor_batch.items():
+        non_tensor_batch[key] = np.array(val, dtype=object)
+    return DataProto(batch=batch, non_tensor_batch=non_tensor_batch)
+
+
+@dataclass
+class DataProtoItem:
+    # TODO(zhangchi.usc1992) add consistency check
+    batch: TensorDict = None
+    non_tensor_batch: dict = field(default_factory=dict)
+    meta_info: dict = field(default_factory=dict)
+
+
+@dataclass
+class DataProto:
+    """
+    A DataProto is a data structure that aims to provide a standard protocol for data exchange between functions.
+    It contains a batch (TensorDict) and a meta_info (Dict). The batch is a TensorDict https://pytorch.org/tensordict/.
+    TensorDict allows you to manipulate a dictionary of Tensors like a single Tensor. Ideally, the tensors with the
+    same batch size should be put inside batch.
+    """
+
+    batch: TensorDict = None
+    non_tensor_batch: dict = field(default_factory=dict)
+    meta_info: dict = field(default_factory=dict)
+
+    def __post_init__(self):
+        # perform necessary checking
+        self.check_consistency()
+
+    def __len__(self):
+        if self.batch is not None:
+            return self.batch.batch_size[0]
+        elif self.non_tensor_batch is not None and len(self.non_tensor_batch) > 0:
+            random_key = next(iter(self.non_tensor_batch.keys()))
+            return self.non_tensor_batch[random_key].shape[0]
+        else:
+            return 0
+
+    def __getitem__(self, item):
+        """
+        Enhanced indexing for DataProto objects.
+
+        Args:
+            item: Can be one of:
+                - int: A single index
+                - slice: A slice object (start:stop:step)
+                - list: A list of indices
+                - numpy.ndarray: An array of indices
+                - torch.Tensor: A tensor of indices
+
+        Returns:
+            DataProto: For all indexing types except single integers
+            DataProtoItem: Only for single integer indices
+        """
+        # Case 1: Slice object - use the slice method
+        if isinstance(item, slice):
+            return self.slice(item.start, item.stop, item.step)
+
+        # Case 2: List, numpy array, or torch tensor - use sel_idxs
+        elif isinstance(item, (list, np.ndarray, torch.Tensor)):
+            return self.select_idxs(item)
+
+        # Case 3: Single integer - return DataProtoItem for backward compatibility
+        elif isinstance(item, (int, np.integer)):
+            tensor_data = self.batch[item] if self.batch is not None else None
+            non_tensor_data = {key: val[item] for key, val in self.non_tensor_batch.items()}
+            return DataProtoItem(batch=tensor_data, non_tensor_batch=non_tensor_data, meta_info=self.meta_info)
+
+        # # Case 4: Unsupported type
+        else:
+            raise TypeError(f"Indexing with {type(item)} is not supported")
+
+    def __getstate__(self):
+        import io
+
+        buffer = io.BytesIO()
+        if version.parse(tensordict.__version__) >= version.parse("0.5.0") and self.batch is not None:
+            self.batch = self.batch.contiguous()
+            self.batch = self.batch.consolidate()
+        torch.save(self.batch, buffer)
+        buffer_bytes = buffer.getvalue()
+        return buffer_bytes, self.non_tensor_batch, self.meta_info
+
+    def __setstate__(self, data):
+        import io
+
+        batch_deserialized_bytes, non_tensor_batch, meta_info = data
+        batch_deserialized = io.BytesIO(initial_bytes=batch_deserialized_bytes)
+        batch = torch.load(
+            batch_deserialized,
+            weights_only=False,
+            map_location="cpu" if not get_torch_device().is_available() else None,
+        )
+        self.batch = batch
+        self.non_tensor_batch = non_tensor_batch
+        self.meta_info = meta_info
+
+    def save_to_disk(self, filepath):
+        with open(filepath, "wb") as f:
+            pickle.dump(self, f)
+
+    @staticmethod
+    def load_from_disk(filepath) -> "DataProto":
+        with open(filepath, "rb") as f:
+            data = pickle.load(f)
+            return data
+
+    def print_size(self, prefix=""):
+        size_of_tensordict = 0
+        if self.batch is None:
+            for _key, tensor in self.batch.items():
+                size_of_tensordict += tensor.element_size() * tensor.numel()
+        size_of_numpy_array = 0
+        for _key, numpy_array in self.non_tensor_batch.items():
+            size_of_numpy_array += numpy_array.nbytes
+
+        size_of_numpy_array /= 1024**3
+        size_of_tensordict /= 1024**3
+
+        message = f"Size of tensordict: {size_of_tensordict} GB, size of non_tensor_batch: {size_of_numpy_array} GB"
+
+        if prefix:
+            message = f"{prefix}, " + message
+        print(message)
+
+    def check_consistency(self):
+        """Check the consistency of the DataProto. Mainly for batch and non_tensor_batch
+        We expose this function as a public one so that user can call themselves directly
+        """
+        if self.batch is not None:
+            assert len(self.batch.batch_size) == 1, "only support num_batch_dims=1"
+
+        if self.non_tensor_batch is not None:
+            for key, val in self.non_tensor_batch.items():
+                assert isinstance(val, np.ndarray)
+
+        if self.batch is not None and self.non_tensor_batch is not None and len(self.non_tensor_batch) != 0:
+            # TODO: we can actually lift this restriction if needed
+            assert len(self.batch.batch_size) == 1, "only support num_batch_dims=1 when non_tensor_batch is not empty."
+
+            batch_size = self.batch.batch_size[0]
+            for key, val in self.non_tensor_batch.items():
+                assert isinstance(val, np.ndarray), (
+                    f"data in the non_tensor_batch must be a numpy.array with dtype=object, but for {key=}, got {type(val)=}"
+                )
+                assert val.shape[0] == batch_size, (
+                    f"key {key} length {len(val)} is not equal to batch size {batch_size}"
+                )
+
+    @classmethod
+    def from_single_dict(cls, data: dict[str, torch.Tensor | np.ndarray], meta_info=None, auto_padding=False):
+        """Create a DataProto from a dict of tensors and non_tensors"""
+        tensors = {}
+        non_tensors = {}
+
+        for key, val in data.items():
+            if isinstance(val, torch.Tensor):
+                tensors[key] = val
+            elif isinstance(val, np.ndarray):
+                non_tensors[key] = val
+            else:
+                raise ValueError(f"Unsupported type in data {type(val)}")
+
+        return cls.from_dict(tensors=tensors, non_tensors=non_tensors, meta_info=meta_info, auto_padding=auto_padding)
+
+    @classmethod
+    def from_dict(
+        cls,
+        tensors: dict[str, torch.Tensor] | None = None,
+        non_tensors=None,
+        meta_info=None,
+        num_batch_dims=1,
+        auto_padding=False,
+    ):
+        """Create a DataProto from a dict of tensors. This assumes that
+        1. All the tensor in tensors have the same dim0
+        2. Only dim0 is the batch dim
+        """
+
+        assert num_batch_dims > 0, "num_batch_dims must be greater than zero"
+        if non_tensors is not None:
+            assert num_batch_dims == 1, "only support num_batch_dims=1 when non_tensors is not None."
+
+        if tensors is None:
+            tensors = {}
+        if meta_info is None:
+            meta_info = {}
+        if non_tensors is None:
+            non_tensors = {}
+
+        assert isinstance(non_tensors, dict)
+
+        # get and check batch size
+        batch_size = None
+        pivot_key = None
+        for key, tensor in tensors.items():
+            if batch_size is None:
+                batch_size = tensor.shape[:num_batch_dims]
+                pivot_key = key
+            else:
+                current_batch = tensor.shape[:num_batch_dims]
+                assert batch_size == current_batch, (
+                    f"Not all the tensor in tensors have the same batch size with batch_dims={num_batch_dims}. Got {pivot_key} has {batch_size}, {key} has {current_batch}"
+                )
+
+        for key, val in non_tensors.items():
+            if not isinstance(val, np.ndarray):
+                non_tensors[key] = np.array(val, dtype=object)
+
+        tensor_dict = TensorDict(source=tensors, batch_size=batch_size) if tensors else None
+        if auto_padding:
+            meta_info[DataProtoConfig.auto_padding_key] = True
+        return cls(batch=tensor_dict, non_tensor_batch=non_tensors, meta_info=meta_info)
+
+    def to(self, device) -> "DataProto":
+        """move the batch to device
+
+        Args:
+            device (torch.device, str): torch device
+
+        Returns:
+            DataProto: the current DataProto
+
+        """
+        if self.batch is not None:
+            self.batch = self.batch.to(device)
+        return self
+
+    def select(self, batch_keys=None, non_tensor_batch_keys=None, meta_info_keys=None, deepcopy=False) -> "DataProto":
+        """Select a subset of the DataProto via batch_keys and meta_info_keys
+
+        Args:
+            batch_keys (list, optional): a list of strings indicating the keys in batch to select
+            meta_info_keys (list, optional): a list of keys indicating the meta info to select
+
+        Returns:
+            DataProto: the DataProto with the selected batch_keys and meta_info_keys
+        """
+        # TODO (zhangchi.usc1992) whether to copy
+        if batch_keys is not None:
+            batch_keys = tuple(batch_keys)
+            sub_batch = self.batch.select(*batch_keys)
+        else:
+            sub_batch = self.batch
+
+        if non_tensor_batch_keys is not None:
+            non_tensor_batch = {key: val for key, val in self.non_tensor_batch.items() if key in non_tensor_batch_keys}
+        else:
+            non_tensor_batch = self.non_tensor_batch
+
+        if deepcopy:
+            non_tensor_batch = copy.deepcopy(non_tensor_batch)
+
+        if meta_info_keys is not None:
+            sub_meta_info = {key: val for key, val in self.meta_info.items() if key in meta_info_keys}
+        else:
+            sub_meta_info = self.meta_info
+
+        if deepcopy:
+            sub_meta_info = copy.deepcopy(sub_meta_info)
+
+        return type(self)(batch=sub_batch, non_tensor_batch=non_tensor_batch, meta_info=sub_meta_info)
+
+    def select_idxs(self, idxs):
+        """
+        Select specific indices from the DataProto.
+
+        Args:
+            idxs (torch.Tensor or numpy.ndarray or list): Indices to select
+
+        Returns:
+            DataProto: A new DataProto containing only the selected indices
+        """
+        if isinstance(idxs, list):
+            idxs = torch.tensor(idxs)
+            if idxs.dtype != torch.bool:
+                idxs = idxs.type(torch.int32)
+
+        if isinstance(idxs, np.ndarray):
+            idxs_np = idxs
+            idxs_torch = torch.from_numpy(idxs)
+        else:  # torch.Tensor
+            idxs_torch = idxs
+            idxs_np = idxs.detach().cpu().numpy()
+
+        batch_size = int(idxs_np.sum()) if idxs_np.dtype == bool else idxs_np.shape[0]
+
+        if self.batch is not None:
+            # Use TensorDict's built-in indexing capabilities
+            selected_batch = TensorDict(
+                source={key: tensor[idxs_torch] for key, tensor in self.batch.items()},
+                batch_size=(batch_size,),
+                device=self.batch.device,
+            )
+        else:
+            selected_batch = None
+
+        selected_non_tensor = {}
+        for key, val in self.non_tensor_batch.items():
+            selected_non_tensor[key] = val[idxs_np]
+
+        return type(self)(batch=selected_batch, non_tensor_batch=selected_non_tensor, meta_info=self.meta_info)
+
+    def slice(self, start=None, end=None, step=None):
+        """
+        Slice the DataProto and return a new DataProto object.
+        This is an improved version of direct slicing which returns a DataProtoItem.
+
+        Args:
+            start (int, optional): Start index. Defaults to None (start from beginning).
+            end (int, optional): End index (exclusive). Defaults to None (go to end).
+            step (int, optional): Step size. Defaults to None (step=1).
+
+        Returns:
+            DataProto: A new DataProto containing the sliced data
+
+        Examples:
+            # Using the slice method directly
+            sliced_data = data_proto.slice(10, 20)
+
+            # Using enhanced indexing (returns DataProto)
+            sliced_data = data_proto[10:20]
+            sliced_data = data_proto[::2]  # Every other element
+
+            # Using list indexing (returns DataProto)
+            indices = [1, 5, 10]
+            selected_data = data_proto[indices]
+
+            # Single index still returns DataProtoItem
+            single_item = data_proto[5]
+        """
+        # Create a slice object
+        slice_obj = slice(start, end, step)
+
+        # Handle the batch data
+        if self.batch is not None:
+            # Use TensorDict's built-in slicing capabilities
+            sliced_batch = self.batch[slice_obj]
+        else:
+            sliced_batch = None
+
+        # Handle the non-tensor batch data
+        sliced_non_tensor = {}
+        for key, val in self.non_tensor_batch.items():
+            sliced_non_tensor[key] = val[slice_obj]
+
+        # Return a new DataProto object
+        return type(self)(batch=sliced_batch, non_tensor_batch=sliced_non_tensor, meta_info=self.meta_info)
+
+    def pop(self, batch_keys=None, non_tensor_batch_keys=None, meta_info_keys=None) -> "DataProto":
+        """Pop a subset of the DataProto via `batch_keys` and `meta_info_keys`
+
+        Args:
+            batch_keys (list, optional): a list of strings indicating the keys in batch to pop
+            meta_info_keys (list, optional): a list of keys indicating the meta info to pop
+
+        Returns:
+            DataProto: the DataProto with the poped batch_keys and meta_info_keys
+        """
+        if batch_keys is None:
+            batch_keys = []
+        if meta_info_keys is None:
+            meta_info_keys = []
+        if non_tensor_batch_keys is None:
+            non_tensor_batch_keys = []
+
+        tensors = {}
+        # tensor batch
+        for key in batch_keys:
+            assert key in self.batch.keys()
+            tensors[key] = self.batch.pop(key)
+        non_tensors = {}
+        # non tensor batch
+        for key in non_tensor_batch_keys:
+            assert key in self.non_tensor_batch.keys()
+            non_tensors[key] = self.non_tensor_batch.pop(key)
+        meta_info = {}
+        for key in meta_info_keys:
+            assert key in self.meta_info.keys()
+            meta_info[key] = self.meta_info.pop(key)
+        return DataProto.from_dict(tensors=tensors, non_tensors=non_tensors, meta_info=meta_info)
+
+    def rename(self, old_keys=None, new_keys=None) -> "DataProto":
+        """
+        Note that this function only rename the key in the batch
+        """
+
+        def validate_input(keys):
+            if keys is not None:
+                if isinstance(keys, str):
+                    keys = [keys]
+                elif isinstance(keys, list):
+                    pass
+                else:
+                    raise TypeError(f"keys must be a list or a string, but got {type(keys)}")
+            return keys
+
+        old_keys = validate_input(old_keys)
+        new_keys = validate_input(new_keys)
+
+        if len(new_keys) != len(old_keys):
+            raise ValueError(
+                f"new_keys and old_keys must have the same length, but got {len(new_keys)} and {len(old_keys)}"
+            )
+
+        self.batch.rename_key_(tuple(old_keys), tuple(new_keys))
+
+        return self
+
+    def union(self, other: "DataProto") -> "DataProto":
+        """Union with another DataProto. Union batch and meta_info separately.
+        Throw an error if
+
+        - there are conflict keys in batch and they are not equal
+        - the batch size of two data batch is not the same
+        - there are conflict keys in meta_info and they are not the same.
+
+        Args:
+            other (DataProto): another DataProto to union
+
+        Returns:
+            DataProto: the DataProto after union
+        """
+        self.batch = union_tensor_dict(self.batch, other.batch)
+        self.non_tensor_batch = union_numpy_dict(self.non_tensor_batch, other.non_tensor_batch)
+        self.meta_info = union_two_dict(self.meta_info, other.meta_info)
+        return self
+
+    def make_iterator(self, mini_batch_size, epochs, seed=None, dataloader_kwargs=None):
+        r"""Make an iterator from the DataProto. This is built upon that TensorDict can be used as a normal Pytorch
+        dataset. See https://pytorch.org/tensordict/tutorials/data_fashion for more details.
+
+
+        Args:
+            mini_batch_size (int): mini-batch size when iterating the dataset. We require that ``batch.batch_size[0] % mini_batch_size == 0``.
+            epochs (int): number of epochs when iterating the dataset.
+            dataloader_kwargs (Any): internally, it returns a DataLoader over the batch. The dataloader_kwargs is the kwargs passed to the DataLoader.
+
+        Returns:
+            Iterator: an iterator that yields a mini-batch data at a time. The total number of iteration steps is ``self.batch.batch_size * epochs // mini_batch_size``
+        """
+        assert self.batch.batch_size[0] % mini_batch_size == 0, f"{self.batch.batch_size[0]} % {mini_batch_size} != 0"
+        # we can directly create a dataloader from TensorDict
+        if dataloader_kwargs is None:
+            dataloader_kwargs = {}
+
+        if seed is not None:
+            generator = torch.Generator()
+            generator.manual_seed(seed)
+        else:
+            generator = None
+
+        assert isinstance(dataloader_kwargs, dict)
+        train_dataloader = DataLoader(
+            dataset=self, batch_size=mini_batch_size, collate_fn=collate_fn, generator=generator, **dataloader_kwargs
+        )
+
+        def get_data():
+            for _ in range(epochs):
+                for d in train_dataloader:
+                    d.meta_info = self.meta_info
+                    yield d
+
+        return iter(get_data())
+
+    def is_padding_enabled(self):
+        """
+        Check if padding is enabled for the DataProto.
+        Returns:
+            bool: True if padding is enabled, False otherwise.
+        """
+        dataproto_specific_padding = self.meta_info.get(DataProtoConfig.auto_padding_key, False)
+        return dataproto_specific_padding or DataProtoConfig.auto_padding
+
+    def padding(self, padding_size, padding_candidate=""):
+        """Pad the DataProto by concating with padding_candidate.repeat(padding_size)
+
+        Args:
+            padding_size (int): the number of repeated padding_candidate
+            padding_candidate: the item to be repeated and appended to the DataProto, only supporting ["first", "last"]
+        """
+        if padding_size == 0:
+            return
+        padding_candidate = self.select_idxs([0 if padding_candidate == "first" else len(self) - 1])
+        padding_part = padding_candidate.repeat(padding_size)
+        padded_dp = DataProto.concat([self, padding_part])
+        self.batch = padded_dp.batch
+        self.non_tensor_batch = padded_dp.non_tensor_batch
+
+    def chunk(self, chunks: int) -> list["DataProto"]:
+        """Split the batch among dim=0 into chunks. The meta_info is passed to each DataProto after split.
+
+        Args:
+            chunks (int): the number of chunks to split on dim=0
+
+        Returns:
+            List[DataProto]: a list of DataProto after splitting
+        """
+        if not self.is_padding_enabled():
+            assert len(self) % chunks == 0, (
+                f"only support equal chunk. Got size of DataProto {len(self)} and chunk {chunks}."
+            )
+
+        bsz_in_batch = None
+        if self.batch is not None:
+            batch_lst = self.batch.chunk(chunks=chunks, dim=0)
+            bsz_in_batch = np.array([batch.batch_size[0] for batch in batch_lst])
+            chunk_indices = np.cumsum(bsz_in_batch)[:-1]
+        else:
+            batch_lst = [None for _ in range(chunks)]
+
+        non_tensor_batch_lst = [{} for _ in range(chunks)]
+        for key, val in self.non_tensor_batch.items():
+            assert isinstance(val, np.ndarray)
+            if bsz_in_batch is not None:
+                non_tensor_lst = np.array_split(val, chunk_indices.tolist())
+            else:
+                non_tensor_lst = np.array_split(val, chunks)
+            assert len(non_tensor_lst) == chunks
+            for i in range(chunks):
+                non_tensor_batch_lst[i][key] = non_tensor_lst[i]
+
+        output = []
+        for i in range(chunks):
+            output.append(
+                type(self)(batch=batch_lst[i], non_tensor_batch=non_tensor_batch_lst[i], meta_info=self.meta_info)
+            )
+
+        return output
+
+    @staticmethod
+    def concat(data: list["DataProto"]) -> "DataProto":
+        """Concat a list of DataProto. The batch is concatenated among dim=0.
+        The meta_info is assumed to be identical and will use the first one.
+
+        Args:
+            data (List[DataProto]): list of DataProto
+
+        Returns:
+            DataProto: concatenated DataProto
+        """
+        batch_lst = []
+        for batch in data:
+            batch_lst.append(batch.batch)
+        new_batch = torch.cat(batch_lst, dim=0) if batch_lst[0] is not None else None
+
+        non_tensor_batch = list_of_dict_to_dict_of_list(list_of_dict=[d.non_tensor_batch for d in data])
+        for key, val in non_tensor_batch.items():
+            non_tensor_batch[key] = np.concatenate(val, axis=0)
+
+        cls = type(data[0]) if len(data) > 0 else DataProto
+        return cls(batch=new_batch, non_tensor_batch=non_tensor_batch, meta_info=data[0].meta_info)
+
+    def reorder(self, indices):
+        """
+        Note that this operation is in-place
+        """
+        indices_np = indices.detach().numpy()
+        self.batch = self.batch[indices]
+        self.non_tensor_batch = {key: val[indices_np] for key, val in self.non_tensor_batch.items()}
+
+    def repeat(self, repeat_times=2, interleave=True):
+        """
+        Repeat the batch data a specified number of times.
+
+        Args:
+            repeat_times (int): Number of times to repeat the data.
+            interleave (bool): Whether to interleave the repeated data.
+
+        Returns:
+            DataProto: A new DataProto with repeated data.
+        """
+        if self.batch is not None:
+            if interleave:
+                # Interleave the data
+                repeated_tensors = {
+                    key: tensor.repeat_interleave(repeat_times, dim=0) for key, tensor in self.batch.items()
+                }
+            else:
+                # Stack the data
+                repeated_tensors = {
+                    key: tensor.unsqueeze(0).expand(repeat_times, *tensor.shape).reshape(-1, *tensor.shape[1:])
+                    for key, tensor in self.batch.items()
+                }
+
+            repeated_batch = TensorDict(
+                source=repeated_tensors,
+                batch_size=(self.batch.batch_size[0] * repeat_times,),
+            )
+        else:
+            repeated_batch = None
+
+        repeated_non_tensor_batch = {}
+        for key, val in self.non_tensor_batch.items():
+            if interleave:
+                repeated_non_tensor_batch[key] = np.repeat(val, repeat_times, axis=0)
+            else:
+                repeated_non_tensor_batch[key] = np.tile(val, (repeat_times,) + (1,) * (val.ndim - 1))
+
+        return type(self)(
+            batch=repeated_batch,
+            non_tensor_batch=repeated_non_tensor_batch,
+            meta_info=self.meta_info,
+        )
+
+    def unfold_column_chunks(self, n_split: int, split_keys: list[str] | None = None):
+        """Split along the second dim into `n_split`, unfold it to the first dim (batch dim)
+        Useful in passing grouped tensors that doesn't want to be shuffled in dataset.
+        keys not in split_keys are repeated to match the shape
+        Note that if the `split_keys` is not provided, it will repeat all the keys in the second dim.
+        """
+        if self.batch is not None:
+            unfolded_batch = {}
+            for key in self.batch.keys():
+                if key in split_keys if split_keys is not None else False:
+                    shape = list(self.batch[key].shape)
+                    shape[0] = self.batch[key].shape[0] * n_split
+                    shape[1] = self.batch[key].shape[1] // n_split
+                    unfolded_batch[key] = self.batch[key].reshape(*shape)
+                else:
+                    unfolded_batch[key] = torch.repeat_interleave(self.batch[key], n_split, dim=0)
+            # locate the `unfolded_batch` as a TensorDict on the same device as the original batch
+            unfolded_batch = TensorDict(
+                source=unfolded_batch, batch_size=(self.batch.batch_size[0] * n_split,), device=self.batch.device
+            )
+        else:
+            unfolded_batch = None
+
+        repeated_non_tensor_batch = {}
+        for key, val in self.non_tensor_batch.items():
+            if key in split_keys:
+                shape = list(val.shape)
+                shape[0] = val.shape[0] * n_split
+                shape[1] = val.shape[1] // n_split
+                repeated_non_tensor_batch[key] = val.reshape(*shape)
+            else:
+                repeated_non_tensor_batch[key] = np.repeat(val, n_split, axis=0)
+
+        return type(self)(
+            batch=unfolded_batch,
+            non_tensor_batch=repeated_non_tensor_batch,
+            meta_info=self.meta_info,
+        )
+
+    def sample_level_repeat(self, repeat_times):
+        """
+        Repeat each row of the batch data a specified number of times.
+
+        Args:
+            repeat_times (torch.tensor, list, tuple, ndarray):  Number of times to repeat the data.
+
+        Returns:
+            DataProto: A new DataProto with repeated data.
+        """
+        if isinstance(repeat_times, tuple):
+            repeat_times = list(repeat_times)
+        elif isinstance(repeat_times, torch.Tensor):
+            assert len(repeat_times.shape) == 1
+            repeat_times = repeat_times.tolist()
+        elif isinstance(repeat_times, np.ndarray):
+            assert len(repeat_times.shape) == 1
+            repeat_times = repeat_times.tolist()
+        else:
+            assert isinstance(repeat_times, list), (
+                f"repeat_times type must be in [list, torch.Tensor, np.ndarray, tuple], got {type(repeat_times)}"
+            )
+        repeat_times = torch.tensor(repeat_times)
+
+        if self.batch is not None:
+            # Interleave the data
+            repeated_tensors = {
+                key: tensor.repeat_interleave(repeat_times, dim=0) for key, tensor in self.batch.items()
+            }
+
+            repeated_batch = TensorDict(
+                source=repeated_tensors,
+                batch_size=(repeat_times.sum().item(),),
+                device=self.batch.device,
+            )
+        else:
+            repeated_batch = None
+
+        repeated_non_tensor_batch = {}
+        for key, val in self.non_tensor_batch.items():
+            repeated_non_tensor_batch[key] = np.repeat(val, repeat_times, axis=0)
+
+        return type(self)(
+            batch=repeated_batch,
+            non_tensor_batch=repeated_non_tensor_batch,
+            meta_info=self.meta_info,
+        )
+
+
+@dataclass
+class DataProtoFuture:
+    """
+    DataProtoFuture aims to eliminate actual data fetching on driver. By doing so, the driver doesn't have to wait
+    for data so that asynchronous execution becomes possible.
+    DataProtoFuture contains a list of futures from another WorkerGroup of size world_size.
+    - collect_fn is a Callable that reduces the list of futures to a DataProto
+    - dispatch_fn is a Callable that partitions the DataProto into a list of DataProto of size world_size and then select
+
+    Potential issue: we can optimize dispatch_fn(collect_fn) such that only needed data is fetched on destination
+    - DataProtoFuture only supports directly passing from the output of a method to another input. You can't perform any
+    operation on the DataProtoFuture in driver.
+    """
+
+    collect_fn: Callable
+    futures: list[ray.ObjectRef]
+    dispatch_fn: Callable = None
+
+    @staticmethod
+    def concat(data: list[ray.ObjectRef]) -> "DataProtoFuture":
+        output = DataProtoFuture(collect_fn=DataProto.concat, futures=data)
+        return output
+
+    def chunk(self, chunks: int) -> list["DataProtoFuture"]:
+        from functools import partial
+
+        arg_future_lst = []
+        for i in range(chunks):
+            # note that we can't directly pass i and chunks
+            def dispatch_fn(x, i, chunks):
+                return x.chunk(chunks=chunks)[i]
+
+            arg_future = DataProtoFuture(
+                collect_fn=self.collect_fn, dispatch_fn=partial(dispatch_fn, i=i, chunks=chunks), futures=self.futures
+            )
+            arg_future_lst.append(arg_future)
+        return arg_future_lst
+
+    def get(self):
+        output = ray.get(self.futures)  # dp_size.
+        for o in output:
+            assert isinstance(o, DataProto)
+        output = self.collect_fn(output)  # select dp, concat
+        if self.dispatch_fn is not None:
+            output = self.dispatch_fn(output)  # split in batch dim, select using dp
+        return output
+
+
+def all_gather_data_proto(data: DataProto, process_group):
+    # Note that this is an inplace operator just like torch.distributed.all_gather
+    group_size = torch.distributed.get_world_size(group=process_group)
+    assert isinstance(data, DataProto)
+    prev_device = data.batch.device
+    data.batch = data.batch.to(get_torch_device().current_device())
+    data.batch = allgather_dict_tensors(data.batch.contiguous(), size=group_size, group=process_group, dim=0)
+    data.batch = data.batch.to(prev_device)
+    # all gather non_tensor_batch
+    all_non_tensor_batch = [None for _ in range(group_size)]
+    torch.distributed.all_gather_object(all_non_tensor_batch, data.non_tensor_batch, group=process_group)
+    data.non_tensor_batch = {k: np.concatenate([d[k] for d in all_non_tensor_batch]) for k in data.non_tensor_batch}
diff --git a/robometer/robometer/utils/fsdp/ulysses.py b/robometer/robometer/utils/fsdp/ulysses.py
new file mode 100755
index 0000000000000000000000000000000000000000..906dc35ab1312cad4fd23703a6f228749891276c
--- /dev/null
+++ b/robometer/robometer/utils/fsdp/ulysses.py
@@ -0,0 +1,320 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Utilities for DeepSpeed Ulysses Sequence Parallelism.
+DeepSpeed Ulysses Paper: https://arxiv.org/abs/2309.14509
+Inspired from: https://github.com/deepspeedai/DeepSpeed/blob/master/deepspeed/sequence/layer.py
+"""
+
+from typing import Any
+
+import torch
+import torch.distributed as dist
+from torch import Tensor
+from torch.distributed import ProcessGroup
+
+_ULYSSES_SEQUENCE_PARALLEL_GROUP = None
+
+
+def set_ulysses_sequence_parallel_group(group: dist.ProcessGroup):
+    """
+    Set ulysses sequence parallel process group.
+    """
+    global _ULYSSES_SEQUENCE_PARALLEL_GROUP
+    _ULYSSES_SEQUENCE_PARALLEL_GROUP = group
+
+
+def get_ulysses_sequence_parallel_group() -> dist.ProcessGroup | None:
+    """
+    Get ulysses sequence parallel process group.
+    """
+    global _ULYSSES_SEQUENCE_PARALLEL_GROUP
+    return _ULYSSES_SEQUENCE_PARALLEL_GROUP
+
+
+def get_ulysses_sequence_parallel_world_size(group: ProcessGroup = None) -> int:
+    """
+    Get ulysses sequence parallel world size.
+    """
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    return dist.get_world_size(group) if group else 1
+
+
+def get_ulysses_sequence_parallel_rank(group: ProcessGroup = None) -> int:
+    """
+    Get ulysses sequence parallel rank.
+    """
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    return dist.get_rank(group) if group else 0
+
+
+def gather_seq_scatter_heads(
+    x: Tensor,
+    seq_dim: int,
+    head_dim: int,
+    unpadded_dim_size: int = 0,
+    group: ProcessGroup = None,
+) -> Tensor:
+    """
+    A func to sync embedding input with alltoall in sequence parallel
+    gather sequence dimension and scatter head dim:
+    e.g. seq_dim: 1, head_dim: 2
+    [bsz, seq/n, h, ...] -> [bsz, seq, h/n, ...]
+    """
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    if not group:
+        return x
+    sp_world = get_ulysses_sequence_parallel_world_size(group)
+    x = SeqAllToAll.apply(group, x, head_dim, seq_dim)
+    if unpadded_dim_size and unpadded_dim_size % sp_world != 0:
+        padding_size = x.size(seq_dim) - unpadded_dim_size
+        x = _unpad_tensor(x, seq_dim, padding_size)
+    return x
+
+
+def gather_heads_scatter_seq(x: Tensor, head_dim: int, seq_dim: int, group: ProcessGroup = None) -> Tensor:
+    """
+    A func to sync attention result with alltoall in sequence parallel
+    gather head dimension and scatter seq dim:
+    e.g. seq_dim: 1, head_dim: 2
+    [bsz, seq, h/n, ...] -> [bsz, seq/n, h, ...]
+    """
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    if not group:
+        return x
+    dim_size = x.size(seq_dim)
+    sp_world = get_ulysses_sequence_parallel_world_size(group)
+    if dim_size % sp_world != 0:
+        padding_size = sp_world - (dim_size % sp_world)
+        x = _pad_tensor(x, seq_dim, padding_size)
+    return SeqAllToAll.apply(group, x, seq_dim, head_dim, False)
+
+
+def _pad_tensor(x: Tensor, dim: int, padding_size: int) -> Tensor:
+    shape = list(x.shape)
+    shape[dim] = padding_size
+    pad = torch.zeros(shape, dtype=x.dtype, device=x.device)
+    return torch.cat([x, pad], dim=dim)
+
+
+def _unpad_tensor(x: Tensor, dim: int, padding_size: int) -> Tensor:
+    slc = [slice(None)] * len(x.shape)
+    slc[dim] = slice(0, -padding_size)
+    return x[slc]
+
+
+def slice_input_tensor(x: Tensor, dim: int, padding: bool = True, group: ProcessGroup = None) -> Tensor:
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    sp_world_size = dist.get_world_size(group)
+    sp_rank = get_ulysses_sequence_parallel_rank()
+    dim_size = x.size(dim)
+    # pad before slice
+    if padding and dim_size % sp_world_size:
+        padding_size = sp_world_size - (dim_size % sp_world_size)
+        x = _pad_tensor(x, dim, padding_size)
+    # slice the input tensor
+    parts = x.size(dim) // sp_world_size
+    slc = [slice(None)] * len(x.shape)
+    slc[dim] = slice(sp_rank * parts, (sp_rank + 1) * parts)
+    return x[slc].contiguous()
+
+
+def all_to_all_tensor(
+    local_input: Tensor,
+    scatter_dim: int,
+    gather_dim: int,
+    group: dist.ProcessGroup | None = None,
+    async_op: bool = False,
+):
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    seq_world_size = dist.get_world_size(group)
+    input_list = [t.contiguous() for t in torch.tensor_split(local_input, seq_world_size, scatter_dim)]
+    output_list = [torch.empty_like(input_list[0]) for _ in range(seq_world_size)]
+    comm = dist.all_to_all(output_list, input_list, group=group, async_op=async_op)
+    if async_op:
+
+        def wait():
+            comm.wait()
+            return torch.cat(output_list, dim=gather_dim).contiguous()
+
+        return wait
+    return torch.cat(output_list, dim=gather_dim).contiguous()
+
+
+def all_gather_tensor(local_tensor: Tensor, group: dist.ProcessGroup | None = None, async_op: bool = False):
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    sp_world_size = dist.get_world_size(group=group)
+    output_shape = list(local_tensor.shape)
+    output_shape[0] = output_shape[0] * sp_world_size
+    output = torch.empty(output_shape, dtype=local_tensor.dtype, device=local_tensor.device)
+    dist.all_gather_into_tensor(output, local_tensor, group=group, async_op=async_op)
+    return output
+
+
+class SeqAllToAll(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx: Any,
+        group: dist.ProcessGroup,
+        local_input: Tensor,
+        scatter_dim: int,
+        gather_dim: int,
+        async_op: bool = False,
+    ) -> Tensor:
+        ctx.group = group
+        ctx.scatter_dim = scatter_dim
+        ctx.gather_dim = gather_dim
+        ctx.async_op = async_op
+        return all_to_all_tensor(local_input, scatter_dim, gather_dim, group, async_op)
+
+    @staticmethod
+    def backward(ctx: Any, *grad_output: Tensor) -> tuple[None, Tensor, None, None]:
+        input_t = torch.cat(grad_output[1:], dim=ctx.gather_dim).contiguous() if ctx.async_op else grad_output[0]
+        return (
+            None,
+            all_to_all_tensor(input_t, ctx.gather_dim, ctx.scatter_dim, ctx.group, False),
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+class Gather(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx: Any,
+        group: dist.ProcessGroup,
+        local_tensor: Tensor,
+        gather_dim: int,
+        grad_scaler: bool = True,
+        async_op=False,
+    ) -> Tensor:
+        ctx.group = group
+        ctx.gather_dim = gather_dim
+        ctx.grad_scaler = grad_scaler
+        ctx.async_op = async_op
+
+        sp_world_size = dist.get_world_size(group=group)
+        ctx.sp_world_size = sp_world_size
+
+        sp_rank = dist.get_rank(group=group)
+        ctx.sp_rank = sp_rank
+
+        local_shape = list(local_tensor.size())
+        split_size = local_shape[0]
+        part_size = local_shape[gather_dim]  # store original size
+        ctx.part_size = part_size
+
+        output = all_gather_tensor(local_tensor, group, async_op)
+        return torch.cat(output.split(split_size, dim=0), dim=gather_dim)
+
+    @staticmethod
+    def backward(ctx: Any, grad_output: Tensor) -> Any:
+        if ctx.grad_scaler:
+            grad_output = grad_output * ctx.sp_world_size
+        return (
+            None,
+            grad_output.split(ctx.part_size, dim=ctx.gather_dim)[ctx.sp_rank].contiguous(),
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+def gather_outpus_and_unpad(
+    x: Tensor,
+    gather_dim: int,
+    unpad_dim: int | None = None,
+    padding_size: int = 0,
+    grad_scaler: bool = True,
+    group: dist.ProcessGroup | None = None,
+):
+    """
+    Gather a tensor across a process group and optionally unpad its padded elements.
+
+    Args:
+        x (Tensor): Input tensor to gather.
+        gather_dim (int): Dimension along which to gather across ranks.
+        unpad_dim (int, optional): Dimension from which to remove padding. If None, no unpadding.
+        padding_size (int): Number of padding elements to remove on `unpad_dim`. Defaults to 0.
+        grad_scaler (bool): Whether to apply gradient scaling during gather. Defaults to True.
+        group (ProcessGroup, optional): Process group for gathering. If None, uses
+            `get_ulysses_sequence_parallel_group()`. If still None, returns `x` unchanged.
+
+    Returns:
+        Tensor: The gathered tensor, with padding removed if requested.
+    """
+    group = get_ulysses_sequence_parallel_group() if group is None else group
+    if group is None:
+        return x
+    x = Gather.apply(group, x, gather_dim, grad_scaler)
+    if unpad_dim is not None:
+        assert isinstance(padding_size, int), "padding size is not given or is not an integer"
+        if padding_size == 0:
+            return x
+        x = _unpad_tensor(x, unpad_dim, padding_size)
+    return x
+
+
+def ulysses_pad(input_ids_rmpad: torch.Tensor, position_ids_rmpad: torch.Tensor | None = None, sp_size: int = 1):
+    if position_ids_rmpad is not None:
+        assert position_ids_rmpad.size(0) == 1
+        assert input_ids_rmpad.size(1) == position_ids_rmpad.size(1)
+    if sp_size <= 1:
+        return input_ids_rmpad, position_ids_rmpad, 0
+    _, total_seq_len = input_ids_rmpad.shape
+    pad_size = (sp_size - total_seq_len % sp_size) % sp_size
+    if pad_size > 0:
+        input_ids_rmpad = torch.nn.functional.pad(input_ids_rmpad, (0, pad_size), value=0)
+        if position_ids_rmpad is not None:
+            pad_pos_ids = torch.arange(pad_size, device=position_ids_rmpad.device).unsqueeze(0)
+            position_ids_rmpad = torch.cat((position_ids_rmpad, pad_pos_ids), dim=-1)
+    return input_ids_rmpad, position_ids_rmpad, pad_size
+
+
+def ulysses_pad_and_slice_inputs(
+    input_ids_rmpad: torch.Tensor, position_ids_rmpad: torch.Tensor | None = None, sp_size: int = 1
+):
+    """
+    Pad and slice input_ids to be divisible by sp_size
+    Pad position_ids to be divisible by sp_size.
+
+    Note both input_ids_rmpad and position_ids_rmpad will be padded and sliced.
+
+    The is the utility of pre-forward for ulysses sequence parallelism
+
+    Args:
+        input_ids_rmpad: shape of [bsz, seqlen]
+        position_ids_rmpad: shape of [bsz, seqlen], where bsz must be 1
+        sp_size (int): ulysses sequence parallelism size
+
+    Returns:
+        torch.Tensor: padded and sliced input_ids
+        torch.Tensor: padded and sliced position_ids
+        int: pad size
+    """
+    input_ids_rmpad, position_ids_rmpad, pad_size = ulysses_pad(input_ids_rmpad, position_ids_rmpad, sp_size)
+    input_ids_rmpad = slice_input_tensor(input_ids_rmpad, dim=1, padding=False)
+    if position_ids_rmpad is not None:
+        position_ids_rmpad = slice_input_tensor(position_ids_rmpad, dim=1, padding=False)
+    return input_ids_rmpad, position_ids_rmpad, pad_size
+
+
+def validate_ulysses_config(num_heads, ulysses_sequence_size):
+    if ulysses_sequence_size > 1:
+        assert num_heads % ulysses_sequence_size == 0, (
+            f"num_heads ({num_heads}) must be divisible by ulysses sequence size({ulysses_sequence_size})"
+        )
diff --git a/robometer/robometer/utils/logger.py b/robometer/robometer/utils/logger.py
new file mode 100644
index 0000000000000000000000000000000000000000..bf3a4186e25c85161620c2c7d6aa1a7f830730f0
--- /dev/null
+++ b/robometer/robometer/utils/logger.py
@@ -0,0 +1,441 @@
+import os
+import sys
+import json
+from typing import Dict, Any, Iterable, Optional, List
+import numpy as np
+import torch
+import wandb
+from torch.utils.tensorboard import SummaryWriter
+import matplotlib.pyplot as plt
+from loguru import logger as loguru_logger
+
+from robometer.utils.distributed import get_rank, is_rank_0
+
+# Define custom log levels for more granular debugging
+# Standard levels: CRITICAL=50, ERROR=40, WARNING=30, INFO=20, DEBUG=10
+# TRACE (5) already exists in loguru by default
+DEBUG2_LEVEL = 8  # Intermediate debug level between TRACE (5) and DEBUG (10)
+
+
+def _add_custom_log_levels():
+    """Add custom log levels to loguru for more granular debugging control."""
+    # TRACE level (5) already exists in loguru by default, no need to add it
+    # Only add DEBUG2 level (intermediate verbose)
+    # Check if level already exists to avoid errors on multiple calls
+    try:
+        loguru_logger.level("DEBUG2", no=DEBUG2_LEVEL, color="<dim><cyan>")
+    except ValueError:
+        # Level already exists, which is fine
+        pass
+
+
+def setup_loguru_logging(log_level: str = "INFO", output_dir: Optional[str] = None):
+    """
+    Initialize loguru logger with rank-aware formatting and log level.
+    Uses loguru's default format to preserve automatic rich formatting for booleans, numbers, etc.
+
+    Supported log levels (from most to least verbose):
+    - TRACE: Most verbose debugging (level 5)
+    - DEBUG2: Intermediate debug level (level 8)
+    - DEBUG: Standard debug level (level 10)
+    - INFO: Informational messages (level 20)
+    - WARNING: Warning messages (level 30)
+    - ERROR: Error messages (level 40)
+    - CRITICAL: Critical errors (level 50)
+
+    Args:
+        log_level: Logging level (e.g., "TRACE", "DEBUG2", "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL")
+        output_dir: Optional directory to write log files to
+    """
+    # Add custom log levels first
+    _add_custom_log_levels()
+
+    loguru_logger.remove()
+    rank = get_rank()
+    rank_prefix = f"[Rank {rank}] "
+    format_string = (
+        "<green>{time:YYYY-MM-DD HH:mm:ss}</green> | "
+        "<level>{level: <8}</level> | "
+        f"{rank_prefix}"
+        "<cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> | "
+        "<level>{message}</level>"
+    )
+
+    # Console handler using sys.stderr to preserve loguru's rich formatting
+    loguru_logger.add(
+        sys.stderr,
+        format=format_string,
+        level=log_level.upper(),
+        colorize=True,
+    )
+
+    # Optional file handler if output_dir is provided
+    if output_dir and is_rank_0():
+        log_file = os.path.join(output_dir, "training.log")
+        loguru_logger.add(
+            log_file,
+            format="{time:YYYY-MM-DD HH:mm:ss} | {level: <8} | {name}:{function}:{line} | {message}\n",
+            level=log_level.upper(),
+            rotation="10 MB",
+            retention="7 days",
+            encoding="utf-8",
+        )
+
+
+class Logger:
+    """Logger for metrics (wandb/tensorboard). For console logging, use loguru logger directly."""
+
+    def __init__(
+        self,
+        log_to: Iterable[str] | None,
+        output_dir: str,
+        is_main_process: bool = True,
+        wandb_run: Optional[Any] = None,
+        log_level: str = "INFO",
+    ):
+        backends = [b.lower() for b in (list(log_to) if log_to is not None else [])]
+        self._use_wandb = "wandb" in backends
+        self._use_tb = "tensorboard" in backends
+        self._is_main = bool(is_main_process)
+
+        # Setup loguru for console logging
+        setup_loguru_logging(log_level=log_level, output_dir=output_dir if is_main_process else None)
+
+        self._wandb_run = wandb.run if (self._use_wandb and self._is_main) else None
+
+        self._tb_writer = None
+        if self._use_tb and self._is_main:
+            logging_dir = os.path.join(output_dir, "tb")
+            os.makedirs(logging_dir, exist_ok=True)
+            self._tb_writer = SummaryWriter(log_dir=logging_dir)
+
+    def init_wandb(
+        self,
+        project: Optional[str],
+        entity: Optional[str],
+        name: Optional[str],
+        config: Optional[dict],
+        notes: Optional[str] = None,
+        mode: Optional[str] = None,
+        resume_id: Optional[str] = None,
+    ):
+        if not (self._use_wandb and self._is_main):
+            return None
+        if self._wandb_run is not None:
+            return self._wandb_run
+        init_kwargs = {"project": project, "entity": entity, "name": name, "config": config}
+        if notes:
+            init_kwargs["notes"] = notes
+        # Use offline mode if specified, or check environment variable
+        if mode is None:
+            mode = os.environ.get("WANDB_MODE", "online")
+        init_kwargs["mode"] = mode
+        # Resume existing run if resume_id is provided
+        if resume_id:
+            init_kwargs["id"] = resume_id
+            init_kwargs["resume"] = "must"
+        self._wandb_run = wandb.init(**init_kwargs)
+        return self._wandb_run
+
+    def enabled(self, backend: str) -> bool:
+        backend = backend.lower()
+        if backend == "wandb":
+            return self._use_wandb and self._is_main and (self._wandb_run is not None)
+        if backend == "tensorboard":
+            return self._use_tb and self._is_main and self._tb_writer is not None
+        return False
+
+    def log_scalars(self, scalars: Dict[str, float], step: Optional[int] = None):
+        if not self._is_main:
+            return
+        if self.enabled("wandb"):
+            self._wandb_run.log(scalars, step=step)
+        if self.enabled("tensorboard"):
+            for k, v in scalars.items():
+                if isinstance(v, (int, float)):
+                    self._tb_writer.add_scalar(k, float(v), global_step=step)
+
+    def log_figure(self, tag: str, figure, step: Optional[int] = None):
+        if not self._is_main:
+            return
+        if self.enabled("wandb"):
+            self._wandb_run.log({tag: wandb.Image(figure)}, step=step)
+        if self.enabled("tensorboard"):
+            self._tb_writer.add_figure(tag, figure, global_step=step)
+
+    def log_image(self, tag: str, image: np.ndarray, step: Optional[int] = None):
+        """
+        Log a static image.
+        - image: numpy array in (H, W, C) format, uint8, RGB
+        """
+        if not self._is_main:
+            return
+        if self.enabled("wandb"):
+            # Ensure image is in (H, W, C) format and uint8
+            if image.ndim == 3 and image.shape[2] == 3:
+                if image.dtype != np.uint8:
+                    image = np.clip(image, 0, 255).astype(np.uint8)
+                self._wandb_run.log({tag: wandb.Image(image)}, step=step)
+        if self.enabled("tensorboard"):
+            # Convert to (C, H, W) for tensorboard
+            if image.ndim == 3 and image.shape[2] == 3:
+                image_chw = image.transpose(2, 0, 1)
+                if image_chw.dtype != np.uint8:
+                    image_chw = np.clip(image_chw, 0, 255).astype(np.uint8)
+                self._tb_writer.add_image(tag, image_chw, global_step=step)
+
+    def log_video_table(
+        self,
+        tag: str,
+        videos_and_figures: List[tuple],
+        columns: List[str],
+        step: Optional[int] = None,
+        fps: int = 6,
+    ):
+        """
+        Log a table where first column can be video (wandb), second a figure, etc.
+        videos_and_figures: list of tuples e.g. [(video_array_or_path, figure), ...]
+        Only supported for wandb; TensorBoard has no native table/video support.
+        """
+        if not self._is_main:
+            return
+
+        if self.enabled("wandb"):
+            rows = []
+            for item in videos_and_figures:
+                row = []
+                for x in item:
+                    if x is None:
+                        row.append(None)
+                    else:
+                        # Wrap images/figures
+                        if hasattr(x, "savefig") or getattr(x, "__class__", type("x", (), {})).__name__ == "Figure":
+                            row.append(wandb.Image(x))
+                            # Free matplotlib figure memory after wrapping
+                            plt.close(x)
+                        # Wrap file path as video
+                        elif isinstance(x, str):
+                            row.append(wandb.Video(x, fps=fps, format="gif"))
+                        else:
+                            # Try to interpret as array/tensor video
+                            arr = None
+                            if isinstance(x, np.ndarray):
+                                arr = x
+                            elif isinstance(x, torch.Tensor):
+                                arr = x.detach().cpu().numpy()
+                            if arr is not None and arr.ndim == 4:
+                                # wandb.Video expects T x C x H x W format
+                                # If last dimension is 3 (or 1), it's THWC format, convert to TCHW
+                                if arr.shape[-1] in (1, 3) and arr.shape[1] not in (1, 3):
+                                    arr = np.transpose(arr, (0, 3, 1, 2))  # T x H x W x C -> T x C x H x W
+                                row.append(wandb.Video(arr, fps=fps, format="gif"))
+                            else:
+                                # Fallback: store raw value
+                                row.append(x)
+                rows.append(row)
+            # if step is not None:
+            #     tag_with_step = f"{tag}/step_{step}"
+            # else:
+            #     tag_with_step = tag
+            self._wandb_run.log({tag: wandb.Table(data=rows, columns=columns)}, step=step)
+
+    def add_text(self, tag: str, text: str, step: Optional[int] = None):
+        if not self._is_main:
+            return
+        if self.enabled("tensorboard"):
+            self._tb_writer.add_text(tag, text, global_step=step)
+        # For wandb, text can be added via log_scalars or a media panel; skip for simplicity
+        if self.enabled("wandb"):
+            # Store as a simple text panel by wrapping in a dict
+            self._wandb_run.log({tag: wandb.Html(f"<pre>{text}</pre>")}, step=step)
+
+    def log_table(self, tag: str, data: List[List[Any]], columns: List[str], step: Optional[int] = None):
+        """
+        Log a generic table (wandb only). TensorBoard has no native table support.
+        """
+        if not self._is_main:
+            return
+        if self.enabled("wandb"):
+            # if step is not None:
+            #     tag_with_step = f"{tag}/step_{step}"
+            # else:
+            #     tag_with_step = tag
+            self._wandb_run.log({tag: wandb.Table(data=data, columns=columns)}, step=step)
+
+    def log_video(self, tag: str, video: Any, fps: int = 10, step: Optional[int] = None):
+        """
+        Log a single video clip.
+        - For wandb: accepts file path or numpy/torch array; arrays are expected as T x H x W x C.
+        - For TensorBoard: accepts numpy/torch array; converted to 1 x C x T x H x W with values in [0,1].
+        """
+        if not self._is_main:
+            return
+        # wandb
+        if self.enabled("wandb"):
+            if isinstance(video, str):
+                self._wandb_run.log({tag: wandb.Video(video, fps=fps)}, step=step)
+            else:
+                arr = None
+                if isinstance(video, np.ndarray):
+                    arr = video
+                elif isinstance(video, torch.Tensor):
+                    arr = video.detach().cpu().numpy()
+                if arr is not None and arr.ndim == 4:
+                    # wandb.Video expects T x C x H x W format
+                    # If last dimension is 3 (or 1), it's THWC format, convert to TCHW
+                    if arr.shape[-1] in (1, 3) and arr.shape[1] not in (1, 3):
+                        arr = np.transpose(arr, (0, 3, 1, 2))  # T x H x W x C -> T x C x H x W
+                    self._wandb_run.log({tag: wandb.Video(arr, fps=fps, format="mp4")}, step=step)
+        # tensorboard
+        if self.enabled("tensorboard"):
+            tens = None
+            if isinstance(video, torch.Tensor):
+                tens = video.detach().cpu()
+            elif isinstance(video, np.ndarray):
+                tens = torch.from_numpy(video)
+            if tens is not None and tens.dim() == 4:
+                # Convert to C x T x H x W
+                if tens.shape[-1] in (1, 3):  # T x H x W x C
+                    tens = tens.permute(0, 3, 1, 2)  # T x C x H x W
+                if tens.dtype != torch.float32:
+                    tens = tens.float()
+                if tens.numel() > 0 and tens.max().item() > 1.0:
+                    tens = tens / 255.0
+                tens = tens.permute(1, 0, 2, 3).unsqueeze(0)  # 1 x C x T x H x W
+                self._tb_writer.add_video(tag, tens, global_step=step, fps=fps)
+
+    def close(self):
+        if self._tb_writer is not None:
+            self._tb_writer.flush()
+            self._tb_writer.close()
+
+    def write_wandb_info(self, output_dir: str, run_name: str):
+        """
+        Persist basic wandb run information alongside outputs, if wandb is active.
+        Safe to call even if wandb isn't enabled or initialized.
+        """
+        if not self.enabled("wandb"):
+            return
+        run = self._wandb_run
+        if run is None:
+            return
+        info = {
+            "wandb_id": run.id,
+            "wandb_name": run.name or run_name,
+            "wandb_project": run.project,
+            "wandb_entity": run.entity,
+            "wandb_url": run.url,
+        }
+        # Include notes if available
+        if hasattr(run, "notes") and run.notes:
+            info["wandb_notes"] = run.notes
+        os.makedirs(output_dir, exist_ok=True)
+        path = os.path.join(output_dir, "wandb_info.json")
+        with open(path, "w") as f:
+            json.dump(info, f, indent=2)
+
+
+def get_logger():
+    """Get the loguru logger instance for structured logging."""
+    return loguru_logger
+
+
+def rank_0_info(*args, **kwargs):
+    """Log info message only on rank 0 (main process)."""
+    if is_rank_0():
+        loguru_logger.info(*args, **kwargs)
+
+
+def rank_0_warning(*args, **kwargs):
+    """Log warning message only on rank 0 (main process)."""
+    if is_rank_0():
+        loguru_logger.warning(*args, **kwargs)
+
+
+def rank_0_debug(*args, **kwargs):
+    """Log debug message only on rank 0 (main process)."""
+    if is_rank_0():
+        loguru_logger.debug(*args, **kwargs)
+
+
+def trace(*args, **kwargs):
+    """Log trace message (most verbose debug level)."""
+    loguru_logger.trace(*args, **kwargs)
+
+
+def debug2(*args, **kwargs):
+    """Log debug2 message (intermediate debug level)."""
+    loguru_logger.debug2(*args, **kwargs)
+
+
+def rank_0_trace(*args, **kwargs):
+    """Log trace message only on rank 0 (main process)."""
+    if is_rank_0():
+        loguru_logger.trace(*args, **kwargs)
+
+
+def rank_0_debug2(*args, **kwargs):
+    """Log debug2 message only on rank 0 (main process)."""
+    if is_rank_0():
+        loguru_logger.debug2(*args, **kwargs)
+
+
+def log_memory_usage(prefix="", rank=None, output_dir=None):
+    """Log GPU and CPU memory usage in a readable format.
+
+    Args:
+        prefix: Prefix string to identify the logging point
+        rank: Process rank (auto-detected if None)
+        output_dir: Directory to write memory log file (uses OUTPUT_DIR env var if None)
+
+    Returns:
+        str: Memory usage string
+    """
+    import psutil
+    import datetime
+
+    if rank is None:
+        rank = get_rank()
+
+    memory_info = []
+
+    # GPU memory
+    if torch.cuda.is_available():
+        for i in range(torch.cuda.device_count()):
+            allocated = torch.cuda.memory_allocated(i) / 1024**3  # GB
+            reserved = torch.cuda.memory_reserved(i) / 1024**3  # GB
+            max_allocated = torch.cuda.max_memory_allocated(i) / 1024**3  # GB
+            memory_info.append(
+                f"GPU{i}: {allocated:.2f}GB alloc, {reserved:.2f}GB reserved, {max_allocated:.2f}GB peak"
+            )
+
+    # CPU memory
+    process = psutil.Process()
+    mem_info = process.memory_info()
+    cpu_mem_rss = mem_info.rss / 1024**3  # GB - Resident Set Size (actual physical memory)
+    cpu_mem_vms = mem_info.vms / 1024**3  # GB - Virtual Memory Size
+
+    # Get system-wide memory stats
+    system_mem = psutil.virtual_memory()
+    system_total = system_mem.total / 1024**3  # GB
+    system_available = system_mem.available / 1024**3  # GB
+    system_percent = system_mem.percent
+
+    memory_info.append(
+        f"CPU: {cpu_mem_rss:.2f}GB RSS, {cpu_mem_vms:.2f}GB VMS | "
+        f"System: {system_available:.2f}GB/{system_total:.2f}GB avail ({system_percent:.1f}% used)"
+    )
+
+    memory_str = " | ".join(memory_info) if memory_info else "No memory info available"
+    loguru_logger.info(f"[Rank {rank}] {prefix} Memory: {memory_str}")
+
+    # Also write to a memory log file for post-mortem analysis
+    if output_dir is None:
+        output_dir = os.environ.get("OUTPUT_DIR", ".")
+
+    memory_log_file = os.path.join(output_dir, f"memory_log_rank{rank}.txt")
+    with open(memory_log_file, "a") as f:
+        timestamp = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        f.write(f"{timestamp} | {prefix} | {memory_str}\n")
+
+    return memory_str
diff --git a/robometer/robometer/utils/metrics.py b/robometer/robometer/utils/metrics.py
new file mode 100755
index 0000000000000000000000000000000000000000..3b911b1d684d1534f0e002d61bfbf0344a591a86
--- /dev/null
+++ b/robometer/robometer/utils/metrics.py
@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+"""
+Utility functions for computing evaluation metrics.
+"""
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+def compute_spearman_correlation(
+    pred: torch.Tensor, target: torch.Tensor, aggregate: bool = True, mask: torch.Tensor = None
+) -> torch.Tensor:
+    """
+    Compute Spearman correlation between prediction and target tensors.
+
+    Args:
+        pred: Prediction tensor
+        target: Target tensor
+        aggregate: If True and tensors are 2D, returns mean correlation. If False and 2D, returns per-sample correlations.
+        mask: Optional mask tensor of shape (N, T) with 1.0 for valid frames, 0.0 for masked frames. Only used for 2D tensors.
+
+    Returns:
+        Spearman correlation coefficient (scalar for 1D or aggregate=True, tensor for 2D with aggregate=False)
+    """
+    from scipy.stats import spearmanr
+
+    # NumPy doesn't support bf16/half; cast to float32 before moving to CPU
+    pred_f32 = pred.detach().to(dtype=torch.float32)
+    target_f32 = target.detach().to(dtype=torch.float32)
+
+    pred_np = pred_f32.cpu().numpy()
+    target_np = target_f32.cpu().numpy()
+
+    # Handle 1D arrays
+    if pred_np.ndim == 1 and target_np.ndim == 1:
+        correlation, _ = spearmanr(pred_np, target_np)
+        if np.isnan(correlation):
+            correlation = 0.0
+        return torch.tensor(correlation, device=pred.device, dtype=torch.float32)
+
+    # Handle 2D arrays (batch, sequence)
+    elif pred_np.ndim == 2 and target_np.ndim == 2:
+        if mask is not None:
+            # Convert mask to numpy
+            mask_np = mask.detach().to(dtype=torch.float32).cpu().numpy()
+            # Ensure mask matches shape
+            if mask_np.shape != pred_np.shape:
+                raise ValueError(f"Mask shape {mask_np.shape} must match pred shape {pred_np.shape}")
+
+        correlations = []
+        for i, (p, t) in enumerate(zip(pred_np, target_np, strict=False)):
+            # Apply mask if provided
+            if mask is not None:
+                valid_mask = mask_np[i] > 0.0
+                p = p[valid_mask]
+                t = t[valid_mask]
+
+            if len(p) > 1 and len(t) > 1:  # Need at least 2 points for correlation
+                corr, _ = spearmanr(p, t)
+                if not np.isnan(corr):
+                    correlations.append(corr)
+                else:
+                    correlations.append(0.0)
+            else:
+                correlations.append(0.0)
+
+        if correlations:
+            correlations_tensor = torch.tensor(correlations, device=pred.device, dtype=torch.float32)
+            if aggregate:
+                return correlations_tensor.mean()
+            else:
+                return correlations_tensor
+        else:
+            result = torch.tensor(0.0, device=pred.device, dtype=torch.float32)
+            return result if aggregate else result.unsqueeze(0).expand(pred.shape[0])
+
+    else:
+        raise ValueError(f"Unsupported tensor dimensions: pred={pred_np.ndim}, target={target_np.ndim}")
+
+
+def compute_auc(scores: torch.Tensor, labels: torch.Tensor) -> float:
+    """
+    Compute Area Under the Curve (AUC) for binary classification.
+
+    Args:
+        scores: Model prediction scores/logits
+        labels: Binary labels (0 or 1)
+
+    Returns:
+        AUC score
+    """
+    try:
+        from sklearn.metrics import roc_auc_score
+
+        scores_np = scores.detach().cpu().numpy()
+        labels_np = labels.detach().cpu().numpy()
+
+        # Handle edge cases
+        if len(np.unique(labels_np)) < 2:
+            return 0.5  # Default AUC for single class
+
+        auc = roc_auc_score(labels_np, scores_np)
+        return auc
+
+    except ImportError:
+        # Fallback implementation if sklearn is not available
+        return manual_auc(scores, labels)
+
+
+def manual_auc(scores: torch.Tensor, labels: torch.Tensor) -> float:
+    """
+    Manual implementation of AUC as fallback.
+
+    Args:
+        scores: Model prediction scores/logits
+        labels: Binary labels (0 or 1)
+
+    Returns:
+        AUC score
+    """
+    # Sort by scores in descending order
+    sorted_indices = torch.argsort(scores, descending=True)
+    sorted_labels = labels[sorted_indices]
+
+    # Count positive and negative samples
+    n_pos = torch.sum(labels == 1).item()
+    n_neg = torch.sum(labels == 0).item()
+
+    if n_pos == 0 or n_neg == 0:
+        return 0.5  # Default AUC for single class
+
+    # Calculate true positive rate and false positive rate
+    tp = 0
+    fp = 0
+    prev_score = float("inf")
+
+    area = 0.0
+
+    for i, label in enumerate(sorted_labels):
+        if scores[sorted_indices[i]] != prev_score:
+            # Calculate area under the curve
+            area += trapezoid_area(fp / n_neg, tp / n_pos, fp / n_neg, tp / n_pos)
+            prev_score = scores[sorted_indices[i]]
+
+        if label == 1:
+            tp += 1
+        else:
+            fp += 1
+
+    # Add final area
+    area += trapezoid_area(fp / n_neg, tp / n_pos, 1.0, 1.0)
+
+    return area
+
+
+def trapezoid_area(x1: float, y1: float, x2: float, y2: float) -> float:
+    """
+    Calculate area of trapezoid using trapezoid rule.
+
+    Args:
+        x1, y1: First point coordinates
+        x2, y2: Second point coordinates
+
+    Returns:
+        Area of trapezoid
+    """
+    return (x2 - x1) * (y1 + y2) / 2
+
+
+def compute_accuracy(predictions: torch.Tensor, targets: torch.Tensor) -> float:
+    """
+    Compute accuracy for binary classification.
+
+    Args:
+        predictions: Binary predictions (0 or 1)
+        targets: Binary targets (0 or 1)
+
+    Returns:
+        Accuracy score
+    """
+    return (predictions == targets).float().mean().item()
+
+
+def compute_mse(predictions: torch.Tensor, targets: torch.Tensor) -> float:
+    """
+    Compute Mean Squared Error.
+
+    Args:
+        predictions: Prediction tensor
+        targets: Target tensor
+
+    Returns:
+        MSE value
+    """
+    return F.mse_loss(predictions, targets).item()
+
+
+def compute_mae(predictions: torch.Tensor, targets: torch.Tensor) -> float:
+    """
+    Compute Mean Absolute Error.
+
+    Args:
+        predictions: Prediction tensor
+        targets: Target tensor
+
+    Returns:
+        MAE value
+    """
+    return F.l1_loss(predictions, targets).item()
diff --git a/robometer/robometer/utils/save.py b/robometer/robometer/utils/save.py
new file mode 100644
index 0000000000000000000000000000000000000000..a8baea15a715d2bf3e1e72e78f0f524f1a7f031a
--- /dev/null
+++ b/robometer/robometer/utils/save.py
@@ -0,0 +1,973 @@
+import os
+import re
+import shutil
+import json
+from dataclasses import dataclass, field
+from typing import Optional, List, Tuple
+from datetime import datetime
+import numpy as np
+import gc
+import torch
+from transformers import TrainerCallback, TrainerState, TrainerControl, TrainingArguments, Trainer
+from transformers.trainer_utils import get_last_checkpoint
+from accelerate.state import AcceleratorState
+from huggingface_hub import HfApi, snapshot_download
+from huggingface_hub.utils import HfHubHTTPError
+from peft import PeftModel
+from .upload_to_hub import upload_model_to_hub
+from robometer.utils.distributed import is_rank_0
+from robometer.utils.logger import loguru_logger as logger
+from robometer.configs.experiment_configs import (
+    ExperimentConfig,
+    ModelConfig,
+    LossConfig,
+    DataConfig,
+)
+from pathlib import Path
+from dataclasses import fields
+from typing import Any, Optional, Tuple
+import yaml
+
+
+def _apply_loaded_section_to_dataclass(instance: Any, loaded: dict, valid_names: set) -> None:
+    """Set attributes on instance from loaded dict only for valid field names."""
+    for key, value in loaded.items():
+        if key in valid_names and value is not None:
+            setattr(instance, key, value)
+
+
+def update_cfg_with_pretrained_ckpt(
+    cfg: ExperimentConfig,
+    resume_from_checkpoint: Optional[str],
+) -> None:
+    """
+    When resuming from a HuggingFace (or local) checkpoint, load its config.yaml
+    and update: cfg.model (full replace), and only progress_loss_type and
+    progress_discrete_bins on cfg.loss and cfg.data.
+    """
+    if not resume_from_checkpoint:
+        return
+
+    hub_token = os.environ.get("HF_TOKEN")
+    is_hub = "/" in resume_from_checkpoint and not resume_from_checkpoint.startswith(("/", "./", "../"))
+    config_path: Optional[str] = None
+
+    if is_hub:
+        repo_id, revision = parse_hf_model_id_and_revision(resume_from_checkpoint, model_name="checkpoint")
+        try:
+            from huggingface_hub import hf_hub_download
+
+            config_path = hf_hub_download(
+                repo_id=repo_id, filename="config.yaml", revision=revision, token=hub_token
+            )
+            logger.info(f"Loaded checkpoint config from Hub: {repo_id}@{revision or 'latest'}")
+        except Exception as e:
+            logger.warning(f"Could not load config from checkpoint repo: {e}")
+            return
+    else:
+        resolved = resolve_checkpoint_path(resume_from_checkpoint, hub_token=hub_token)
+        if not resolved:
+            return
+        for candidate in [Path(resolved) / "config.yaml", Path(resolved).parent / "config.yaml"]:
+            if candidate.is_file():
+                config_path = str(candidate)
+                logger.info(f"Loaded checkpoint config from local: {config_path}")
+                break
+        if not config_path:
+            return
+
+    with open(config_path) as f:
+        loaded = yaml.safe_load(f)
+    if not isinstance(loaded, dict):
+        return
+
+    # Replace entire model config except use_peft (keep current run's choice); only sync progress-loss + use_multi_image for loss and data
+    model_names = {f.name for f in fields(ModelConfig)} - {"use_peft"}
+    progress_loss_fields = {"progress_loss_type", "progress_discrete_bins"}
+    loss_names = progress_loss_fields & {f.name for f in fields(LossConfig)}
+    data_sync_fields = progress_loss_fields | {"use_multi_image", "use_per_frame_progress_token"} 
+    data_names = data_sync_fields & {f.name for f in fields(DataConfig)}
+
+    model_loaded = loaded.get("model")
+    loss_loaded = loaded.get("loss")
+    data_loaded = loaded.get("data")
+
+    if model_loaded and isinstance(model_loaded, dict):
+        _apply_loaded_section_to_dataclass(cfg.model, model_loaded, model_names)
+        logger.info("Updated model config from checkpoint (full replace)")
+    if loss_loaded and isinstance(loss_loaded, dict) and loss_names:
+        _apply_loaded_section_to_dataclass(cfg.loss, loss_loaded, loss_names)
+    if data_loaded and isinstance(data_loaded, dict) and data_names:
+        _apply_loaded_section_to_dataclass(cfg.data, data_loaded, data_names)
+    if loss_names or data_names:
+        logger.info(
+            "Updated from checkpoint: progress_loss_type, progress_discrete_bins (loss); progress_loss_type, progress_discrete_bins, use_multi_image (data)"
+        )
+
+def resolve_checkpoint_path(checkpoint_path: Optional[str], hub_token: Optional[str] = None) -> Optional[str]:
+    """
+    Resolve checkpoint path, supporting local paths and HuggingFace Hub with @ notation.
+
+    Args:
+        checkpoint_path: Path to checkpoint. Can be:
+            - None: No checkpoint to load
+            - Local path: /path/to/checkpoint
+            - HF repo: username/model-name (loads best tag automatically)
+            - HF repo with tag: username/model-name@tag-name
+        hub_token: Optional HuggingFace token for private repos
+
+    Returns:
+        Resolved local path to checkpoint, or None if no checkpoint
+    """
+    if not checkpoint_path:
+        return None
+
+    # If it's a local path, return as-is
+    if checkpoint_path.startswith("/") or checkpoint_path.startswith("./") or checkpoint_path.startswith("../"):
+        logger.info(f"Using local checkpoint: {checkpoint_path}")
+        return checkpoint_path
+
+    # Check if it looks like a HuggingFace repo (contains /)
+    if "/" in checkpoint_path:
+        repo_id, revision = parse_hf_model_id_and_revision(checkpoint_path, model_name="checkpoint")
+
+        # Download from HuggingFace Hub
+        logger.info(f"Downloading checkpoint from HuggingFace Hub: {repo_id}@{revision or 'latest'}")
+        local_path = snapshot_download(
+            repo_id=repo_id,
+            revision=revision,
+            token=hub_token,
+            allow_patterns=["*.safetensors", "*.bin", "*.json", "*.txt", "*.model", "*.yaml"],
+        )
+        logger.info(f"Downloaded checkpoint to: {local_path}")
+        return local_path
+
+    # Otherwise, treat as local path
+    logger.info(f"Using checkpoint: {checkpoint_path}")
+    return checkpoint_path
+
+
+def parse_hf_model_id_and_revision(hf_model_id: str, model_name: str = "model") -> Tuple[str, Optional[str]]:
+    """
+    Parse HuggingFace model ID and determine which revision (tag) to load.
+
+    Supports explicit revisions via repo@revision format, or automatically
+    finds the best tag if no explicit revision is provided.
+
+    Args:
+        hf_model_id: HuggingFace model repository ID or local path, optionally with @revision
+        model_name: Name of the model type for logging (e.g., "ReWiND model", "Qwen model")
+
+    Returns:
+        Tuple of (repo_id, revision_to_load) where:
+        - repo_id: The repository ID without the @revision suffix
+        - revision_to_load: The revision/tag to load, or None for latest
+    """
+    # Allow users to specify explicit revisions via repo@revision
+    if "@" in hf_model_id:
+        repo_id, explicit_revision = hf_model_id.split("@", 1)
+    else:
+        repo_id, explicit_revision = hf_model_id, None
+
+    revision_to_load = explicit_revision
+
+    # Check if this is a HuggingFace repo (not a local path) and find best tag
+    if "/" in repo_id and not repo_id.startswith("/"):
+        if revision_to_load:
+            logger.info(f"Loading {model_name} {repo_id} at explicit revision '{revision_to_load}'")
+        else:
+            best_tag, best_score = find_best_model_tag(repo_id)
+            if best_tag:
+                revision_to_load = best_tag
+                logger.info(f"Loading {model_name} from best tag: {repo_id}@{revision_to_load} (score: {best_score})")
+            else:
+                logger.info(f"No best tag found, loading latest revision of {repo_id}")
+    else:
+        logger.info(f"Loading local/explicit {model_name} from {repo_id}")
+
+    return repo_id, revision_to_load
+
+
+def find_best_model_tag(hf_model_id: str, hub_token: Optional[str] = None) -> Tuple[Optional[str], Optional[float]]:
+    """
+    Find the best model tag from HuggingFace Hub by parsing tag names and extracting scores.
+
+    Expected tag format: "best-{metric_short}-{score:.4f}-step-{step}"
+    Example: "best-p-rank-spearman-mw-0.8500-step-123" or "best-avg-3metrics-0.7234-step-456"
+
+    Args:
+        hf_model_id: HuggingFace model ID (e.g., "aliangdw/rewind-debug")
+        hub_token: Optional HuggingFace token for private repos
+
+    Returns:
+        tuple: (best_tag_name, best_score) or (None, None) if no valid tags found
+    """
+    try:
+        api = HfApi(token=hub_token)
+
+        # Check if repository exists
+        if not api.repo_exists(repo_id=hf_model_id, repo_type="model"):
+            logger.info(f"Repository {hf_model_id} does not exist")
+            return None, None
+
+        # Get all tags for the repository
+        tags = api.list_repo_refs(repo_id=hf_model_id, repo_type="model").tags
+
+        if not tags:
+            logger.info(f"No tags found in repository {hf_model_id}")
+            return None, None
+
+        logger.info(f"Found {len(tags)} tags in {hf_model_id}: {[tag.name for tag in tags]}")
+
+        best_tag = None
+        best_score = float("-inf")
+
+        # Parse each tag to extract score
+        for tag in tags:
+            tag_name = tag.name
+
+            # Match our tag pattern: "best-{metric_short}-{score}-step-{step}"
+            # Examples: "best-p-rank-spearman-mw-0.8500-step-123" or "best-avg-3metrics-0.7234-step-456"
+            # Score can be positive or negative (e.g., 0.8500 or -1.2300)
+            pattern = r"best-.*?-(-?\d+\.\d+)-step-\d+"
+            match = re.search(pattern, tag_name)
+
+            if match:
+                try:
+                    score = float(match.group(1))
+                    logger.info(f"Parsed tag '{tag_name}': score = {score}")
+
+                    if score > best_score:
+                        best_score = score
+                        best_tag = tag_name
+
+                except ValueError:
+                    logger.info(f"Could not parse score from tag '{tag_name}'")
+                    continue
+            else:
+                logger.info(f"Tag '{tag_name}' does not match expected pattern")
+
+        if best_tag:
+            logger.info(f"Best tag found: '{best_tag}' with score {best_score}")
+        else:
+            logger.info("No valid tags found matching the expected pattern")
+
+        return best_tag, best_score
+
+    except Exception as e:
+        logger.info(f"Error finding best tag for {hf_model_id}: {e}")
+        return None, None
+
+
+class SaveBestCallback(TrainerCallback):
+    """
+    Save a checkpoint whenever `metric_name` improves.
+    Works in DDP/accelerate: only rank 0 writes checkpoints.
+    Optionally keeps top-k best checkpoints and uploads to Hub.
+    Also saves a 'latest' checkpoint at regular intervals.
+    """
+
+    def __init__(
+        self,
+        metric_names: List[str] = None,
+        greater_is_better: List[bool] = None,
+        keep_top_k: int = 1,
+        save_every: Optional[int] = None,
+        upload_to_hub: bool = False,
+        hub_save_every: Optional[int] = None,
+        hub_token: Optional[str] = None,
+        hub_private: bool = False,
+        base_model: str = "Qwen/Qwen2.5-VL-3B-Instruct",
+    ):
+        super().__init__()
+        self.metric_names = metric_names or ["custom_eval/p_rank_spearman_mw"]
+        self.greater_is_better = greater_is_better or [True]
+
+        # Validate inputs
+        if len(self.metric_names) != len(self.greater_is_better):
+            raise ValueError(
+                f"metric_names ({len(self.metric_names)}) and greater_is_better ({len(self.greater_is_better)}) must have the same length"
+            )
+        self.keep_top_k = keep_top_k
+        self.save_every = save_every
+        self.upload_to_hub = upload_to_hub
+        self.hub_save_every = hub_save_every  # Frequency for Hub uploads (None = upload every checkpoint)
+        self.hub_token = hub_token
+        self.hub_private = hub_private
+        self.base_model = base_model
+        self._best_val = None
+        self._saved: List[Tuple[float, str]] = []  # list of (score, path), sorted from best -> worst
+        self._uploaded: List[
+            Tuple[float, str, str]
+        ] = []  # list of (score, tag_name, commit_id), sorted from best -> worst
+        self._trainer = None  # Will be set when callback is registered
+        self._last_save_step = -1  # Track last step where we saved 'latest'
+        self._last_best_save_step = -1  # Track last step where we saved 'best'
+        self._last_hub_upload_step = -1  # Track last step where we uploaded 'best' to Hub
+        self._last_latest_hub_upload_step = -1  # Track last step where we uploaded 'latest' to Hub
+        self._previous_latest_ckpt_dir = None  # Track previous 'latest' checkpoint directory
+        self._previous_latest_hub_tag = None  # Track previous 'latest' Hub tag
+        self._run_timestamp = datetime.now().strftime("%Y%m%d-%H%M%S")  # Static timestamp for this run
+
+    def setup_trainer_reference(self, trainer: Trainer):
+        """Set the trainer reference for later use in callbacks"""
+        self._trainer = trainer
+
+    def _compute_averaged_score(self, metrics: dict) -> Tuple[float, List[str]]:
+        """
+        Compute averaged score from multiple metrics.
+
+        Returns:
+            tuple: (averaged_score, missing_metrics)
+        """
+        available_scores = []
+        missing_metrics = []
+
+        for metric_name, is_better in zip(self.metric_names, self.greater_is_better):
+            if metric_name in metrics:
+                score = float(metrics[metric_name])
+                # Normalize score: if lower is better, negate it so higher normalized score is better
+                normalized_score = score if is_better else -score
+                available_scores.append(normalized_score)
+            else:
+                missing_metrics.append(metric_name)
+
+        if not available_scores:
+            return float("-inf"), missing_metrics
+
+        # Return average of normalized scores
+        return np.mean(available_scores), missing_metrics
+
+    def _is_main_process(self, trainer: Trainer) -> bool:
+        try:
+            return trainer.is_world_process_zero() and is_rank_0()
+        except:
+            return (not AcceleratorState().distributed_type) or AcceleratorState().is_main_process
+
+    def _build_metric_short_name(self) -> str:
+        """Build a short metric name for checkpoint naming."""
+        if len(self.metric_names) == 1:
+            return self.metric_names[0].split("/")[-1]
+        else:
+            return f"avg-{len(self.metric_names)}metrics"
+
+    def _build_metrics_detail_string(self, metrics: dict) -> str:
+        """Build a detailed metrics string for logging."""
+        metrics_detail = []
+        for name in self.metric_names:
+            if name in metrics:
+                metrics_detail.append(f"{name}:{metrics[name]:.4f}")
+        return " | ".join(metrics_detail) if metrics_detail else "no metrics"
+
+    def _build_individual_scores_string(self, metrics: dict) -> str:
+        """Build individual scores string for commit messages."""
+        individual_scores = []
+        for name in self.metric_names:
+            if name in metrics:
+                individual_scores.append(f"{name.split('/')[-1]}={metrics[name]:.4f}")
+        return ", ".join(individual_scores) if individual_scores else "no metrics"
+
+    def _get_hub_model_id(self, args: TrainingArguments) -> str:
+        """Get the Hub model ID from output directory with timestamp."""
+        base_name = args.output_dir.split("/")[-1].replace("_", "-")
+        base_name = re.sub(r"-+", "-", base_name)
+        base_name = base_name.strip("-")
+        return f"rewardfm/{base_name}-{self._run_timestamp}"
+
+    def _clean_tag_name(self, tag_name: str) -> str:
+        """Clean tag name for HuggingFace repo naming requirements."""
+        tag_name = tag_name.replace("_", "-").replace(",", "")
+        tag_name = re.sub(r"-+", "-", tag_name)
+        tag_name = tag_name.strip("-")
+        return tag_name
+
+    def _save_checkpoint_files(self, args: TrainingArguments, ckpt_dir: str, metrics: dict = None, step: int = None):
+        """Save model, trainer state files, and metrics.
+
+        For PEFT models, uses standard PeftModel.save_pretrained() to save adapter weights.
+        Also saves custom heads (progress_head, etc.) which are part of the RBM wrapper.
+
+        Note: This should only be called from rank 0 in the current implementation.
+        """
+        model = self._trainer.model
+
+        # Check if we're using PEFT
+        is_peft = False
+        base_model = None
+        if hasattr(model, "model") and isinstance(model.model, PeftModel):
+            is_peft = True
+            base_model = model.model
+            logger.info("Detected PEFT model - using standard PeftModel.save_pretrained() for adapter weights")
+
+        if is_peft:
+            # For PEFT models, we need to save:
+            # 1. Adapter weights using PeftModel.save_pretrained() (saves adapter_model.safetensors + adapter_config.json)
+            # 2. Custom heads (progress_head, etc.) from RBM wrapper
+            # 3. Other RBM-specific parameters (frame_pool_attn, video_proj, text_proj, etc.)
+
+            # Save adapter weights using standard PEFT method
+            logger.info("Saving PEFT adapter weights using PeftModel.save_pretrained()")
+            base_model.save_pretrained(ckpt_dir)
+
+            # Save custom heads and other RBM-specific parameters
+            # These are not part of the PeftModel, so we save them separately
+            rbm_state_dict = {}
+            for name, param in model.named_parameters():
+                # Include custom heads and other RBM-specific parameters
+                # Exclude base model parameters (handled by PEFT) and adapter parameters (already saved)
+                if (
+                    "progress_head" in name
+                    or "preference_head" in name
+                    or "similarity_head" in name
+                    or "success_head" in name
+                ):
+                    rbm_state_dict[name] = param.data.cpu()
+                elif "frame_pool_attn" in name or "video_proj" in name or "text_proj" in name:
+                    rbm_state_dict[name] = param.data.cpu()
+
+            # Also save non-parameter buffers if any
+            for name, buffer in model.named_buffers():
+                if (
+                    "progress_head" in name
+                    or "preference_head" in name
+                    or "similarity_head" in name
+                    or "success_head" in name
+                ):
+                    rbm_state_dict[name] = buffer.cpu()
+
+            if rbm_state_dict:
+                # Save custom heads to a separate file
+                from safetensors.torch import save_file
+
+                custom_heads_path = os.path.join(ckpt_dir, "custom_heads.safetensors")
+                save_file(rbm_state_dict, custom_heads_path)
+                logger.info(f"Saved {len(rbm_state_dict)} custom head parameters to {custom_heads_path}")
+
+            # Verify that adapter weights were saved
+            adapter_config_path = os.path.join(ckpt_dir, "adapter_config.json")
+            adapter_model_paths = [
+                os.path.join(ckpt_dir, "adapter_model.safetensors"),
+                os.path.join(ckpt_dir, "adapter_model.bin"),  # Fallback to .bin if safetensors not available
+            ]
+
+            if os.path.exists(adapter_config_path):
+                logger.info(f"PEFT adapter config saved to {adapter_config_path}")
+            else:
+                logger.warning(f"PEFT adapter config not found at {adapter_config_path}")
+
+            # Check for adapter model files
+            adapter_model_found = any(os.path.exists(p) for p in adapter_model_paths)
+            if adapter_model_found:
+                adapter_path = next(p for p in adapter_model_paths if os.path.exists(p))
+                logger.info(f"PEFT adapter weights saved to {adapter_path}")
+            else:
+                logger.warning(
+                    "PEFT adapter weights file not found - adapter weights may not have been saved correctly"
+                )
+        else:
+            # For non-PEFT models, use standard save_model()
+            self._trainer.save_model(ckpt_dir)
+
+        if args.should_save:
+            self._trainer.save_state()  # trainer_state.json etc. in output_dir
+            # save the trainer_state.json to the actual checkpoint directory
+            shutil.copy(os.path.join(args.output_dir, "trainer_state.json"), ckpt_dir)
+
+        # Save metrics to JSON file
+        if metrics is not None:
+            metrics_file = os.path.join(ckpt_dir, "metrics.json")
+            metrics_to_save = {
+                "step": step,
+                "metrics": {
+                    k: float(v) if isinstance(v, (int, float, np.number)) else str(v) for k, v in metrics.items()
+                },
+            }
+            with open(metrics_file, "w") as f:
+                json.dump(metrics_to_save, f, indent=2)
+            logger.info(f"📊 Saved metrics to {metrics_file}")
+
+        # Save random state from training dataset
+        if self._trainer and hasattr(self._trainer, "train_dataset"):
+            try:
+                train_dataset = self._trainer.train_dataset
+                # Handle RepeatedDataset wrapper if present
+                if hasattr(train_dataset, "dataset"):
+                    train_dataset = train_dataset.dataset
+
+                if hasattr(train_dataset, "get_random_state"):
+                    random_state = train_dataset.get_random_state()
+                    random_state_file = os.path.join(ckpt_dir, "dataset_random_state.json")
+                    with open(random_state_file, "w") as f:
+                        json.dump(random_state, f, indent=2)
+                    logger.info(f"Saved dataset random state to {random_state_file}")
+            except Exception as e:
+                logger.warning(f"Could not save random state: {e}")
+
+    def _cleanup_memory(self):
+        """Perform memory cleanup."""
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        gc.collect()
+
+    def _upload_checkpoint_to_hub(
+        self, ckpt_dir: str, hub_model_id: str, tag_name: str, commit_message: str
+    ) -> Tuple[str, str]:
+        """Upload checkpoint to Hub and return URL and commit ID."""
+        hub_url, commit_id = upload_model_to_hub(
+            model_dir=ckpt_dir,
+            hub_model_id=hub_model_id,
+            private=self.hub_private,
+            token=self.hub_token,
+            commit_message=commit_message,
+            base_model=self.base_model,
+            tag_name=tag_name,
+        )
+        return hub_url, commit_id
+
+    def on_evaluate(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, metrics, **kwargs):
+        """
+        Callback triggered after evaluation.
+        Metrics are already gathered across all processes by the trainer before being passed here.
+        """
+        step = state.global_step
+
+        # Only rank 0 needs to process metrics and save checkpoints
+        if not self._is_main_process(self._trainer):
+            logger.debug("Skipping checkpoint save (not main process)")
+            return control
+
+        logger.info(f"SaveBestCallback.on_evaluate called at step {step} with {len(metrics)} metrics")
+
+        score, missing_metrics = self._compute_averaged_score(metrics)
+
+        if missing_metrics:
+            logger.warning(f"⚠️ Metrics {missing_metrics} not found in evaluation metrics")
+            logger.warning(f"Available metrics: {metrics.keys()}")
+            # If all metrics are missing, use a dummy score for filename but still save
+            if score == float("-inf"):  # All metrics missing
+                score_for_filename = 0.0  # Dummy value for filename
+                logger.warning("⚠️ All metrics missing, using dummy score 0.0 in checkpoint filename")
+            else:
+                score_for_filename = score
+        else:
+            score_for_filename = score
+
+        improved = (self._best_val is None) or (score > self._best_val)
+
+        # Check if this score is worth saving (top-k logic)
+        should_save = False
+        if len(self._saved) < self.keep_top_k:
+            # We haven't reached top-k yet, always save
+            should_save = True
+        else:
+            # Check if this score beats the worst in our top-k
+            worst_score = self._saved[-1][0]  # Last item is worst (sorted best -> worst)
+            should_save = score > worst_score  # Always use > since we normalized scores
+
+        if should_save and self._trainer:
+            # Update overall best for reference (only if we have a valid score)
+            if improved and score != float("-inf"):
+                self._best_val = score
+
+            # Make a descriptive dir name
+            step = state.global_step
+            metric_short = self._build_metric_short_name()
+            tag = f"{metric_short}={score_for_filename:.4f}_step={step}"
+            ckpt_dir = os.path.join(args.output_dir, f"ckpt-{tag}")
+
+            metrics_str = self._build_metrics_detail_string(metrics)
+            logger.info(
+                f"💾 Saving ckpt: {ckpt_dir} | avg_score: {score_for_filename:.6f} | {metrics_str} (rank {len(self._saved) + 1}/{self.keep_top_k})"
+            )
+
+            # Create checkpoint directory
+            os.makedirs(ckpt_dir, exist_ok=True)
+
+            # Save model, trainer state, and metrics
+            self._save_checkpoint_files(args, ckpt_dir, metrics, step)
+            self._cleanup_memory()
+
+            # Track that we saved a best checkpoint at this step
+            self._last_best_save_step = step
+
+            # Add to saved list and sort (always best -> worst since we normalized scores)
+            self._saved.append((score, ckpt_dir))
+            self._saved.sort(key=lambda x: x[0], reverse=True)
+
+            # Remove old checkpoint if we exceed keep_top_k
+            if len(self._saved) > self.keep_top_k:
+                _, path_to_rm = self._saved.pop(-1)
+                logger.info(f"🗑️ Removing old checkpoint: {path_to_rm}")
+                if os.path.isdir(path_to_rm):
+                    shutil.rmtree(path_to_rm, ignore_errors=True)
+
+            # Upload to Hub if enabled and frequency check passes
+            should_upload_to_hub = False
+            if self.upload_to_hub:
+                if self.hub_save_every is None:
+                    # Upload every checkpoint if no frequency is set
+                    should_upload_to_hub = True
+                else:
+                    # Check if it's time to upload based on frequency
+                    if self._last_hub_upload_step == -1:
+                        # First upload
+                        should_upload_to_hub = True
+                    elif (step - self._last_hub_upload_step) >= self.hub_save_every:
+                        # Enough steps have passed
+                        should_upload_to_hub = True
+
+            if should_upload_to_hub:
+                hub_model_id = self._get_hub_model_id(args)
+                tag_name = self._clean_tag_name(f"best-{metric_short}-{score_for_filename:.4f}-step-{step}")
+                individual_scores_str = self._build_individual_scores_string(metrics)
+                commit_message = (
+                    f"Checkpoint: avg_score={score_for_filename:.4f} at step {step} | {individual_scores_str}"
+                )
+
+                logger.info(f"🚀 Uploading to Hub: {hub_model_id}")
+
+                hub_url, commit_id = self._upload_checkpoint_to_hub(
+                    ckpt_dir=ckpt_dir,
+                    hub_model_id=hub_model_id,
+                    tag_name=tag_name,
+                    commit_message=commit_message,
+                )
+                logger.info(f"✅ Successfully uploaded to: {hub_url}")
+                logger.info(f"🏷️ Tagged as: {tag_name}")
+
+                # Track that we uploaded to Hub at this step
+                self._last_hub_upload_step = step
+
+                # Add to uploaded list and sort (always best -> worst since we normalized scores)
+                self._uploaded.append((score, tag_name, commit_id))
+                self._uploaded.sort(key=lambda x: x[0], reverse=True)
+
+                # Remove old tags if we exceed keep_top_k
+                api = HfApi(token=self.hub_token)
+                if len(self._uploaded) > self.keep_top_k:
+                    _, old_tag, _ = self._uploaded.pop(-1)
+                    logger.info(f"🗑️ Removing old Hub tag: {old_tag}")
+                    api.delete_tag(repo_id=hub_model_id, repo_type="model", tag=old_tag)
+                    logger.info(f"✅ Deleted tag: {old_tag}")
+
+                # Aggressive memory cleanup after upload to prevent OOM
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                    torch.cuda.synchronize()
+                gc.collect()
+                logger.info("🧹 Cleaned up memory after Hub upload")
+            elif self.upload_to_hub and self.hub_save_every is not None:
+                # Hub upload is enabled but not time yet
+                steps_until_upload = self.hub_save_every - (step - self._last_hub_upload_step)
+                logger.info(f"⏭️ Skipping Hub upload (saving locally only). Next upload in {steps_until_upload} steps")
+
+        # Save 'latest' checkpoint if save_every is configured and it's time to save
+        # Do this after processing best checkpoints so we have the gathered metrics
+        # Skip if we just saved a best checkpoint at this step
+        if self.save_every is not None and state.global_step > 0 and state.global_step % self.save_every == 0:
+            if state.global_step != self._last_save_step and state.global_step != self._last_best_save_step:
+                self._save_latest_checkpoint(args, state, metrics)
+                self._last_save_step = state.global_step
+            elif state.global_step == self._last_best_save_step:
+                logger.info(
+                    f"⏭️ Skipping 'latest' checkpoint save at step {state.global_step} (already saved as 'best')"
+                )
+
+        # Additional cleanup on all ranks after the entire on_evaluate callback
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+        gc.collect()
+
+        return control
+
+    def _save_latest_checkpoint(self, args: TrainingArguments, state: TrainerState, metrics: dict):
+        """Save a 'latest' checkpoint with metrics and step in the tag.
+        Tracks and deletes the previous 'latest' checkpoint.
+
+        Args:
+            args: Training arguments
+            state: Trainer state
+            metrics: Evaluation metrics dictionary
+        """
+        if not self._trainer:
+            return
+
+        # Compute score and build tag similar to best checkpoints
+        score, missing_metrics = self._compute_averaged_score(metrics)
+        step = state.global_step
+        metric_short = self._build_metric_short_name()
+
+        # Build tag with metrics and step
+        tag = f"latest-{metric_short}={score:.4f}_step={step}"
+        ckpt_dir = os.path.join(args.output_dir, f"ckpt-{tag}")
+
+        metrics_str = self._build_metrics_detail_string(metrics)
+        logger.info(f"💾 Saving 'latest' checkpoint at step {step} to {ckpt_dir} | {metrics_str}")
+
+        # Remove old 'latest' checkpoint if it exists
+        if self._previous_latest_ckpt_dir and os.path.isdir(self._previous_latest_ckpt_dir):
+            logger.info(f"🗑️ Removing previous 'latest' checkpoint: {self._previous_latest_ckpt_dir}")
+            shutil.rmtree(self._previous_latest_ckpt_dir, ignore_errors=True)
+
+        # Save model, trainer state, and metrics
+        self._save_checkpoint_files(args, ckpt_dir, metrics, step)
+        logger.info(f"✅ Saved 'latest' checkpoint at step {step}")
+
+        # Upload to Hub if enabled and frequency check passes
+        should_upload_latest_to_hub = False
+        if self.upload_to_hub:
+            if self.hub_save_every is None:
+                # Upload every checkpoint if no frequency is set
+                should_upload_latest_to_hub = True
+            else:
+                # Check if it's time to upload based on frequency
+                if self._last_latest_hub_upload_step == -1:
+                    # First upload
+                    should_upload_latest_to_hub = True
+                elif (step - self._last_latest_hub_upload_step) >= self.hub_save_every:
+                    # Enough steps have passed
+                    should_upload_latest_to_hub = True
+
+        if should_upload_latest_to_hub:
+            hub_model_id = self._get_hub_model_id(args)
+            tag_name = self._clean_tag_name(f"latest-{metric_short}-{score:.4f}-step-{step}")
+            individual_scores_str = self._build_individual_scores_string(metrics)
+            commit_message = f"Latest checkpoint: avg_score={score:.4f} at step {step} | {individual_scores_str}"
+
+            # Delete previous 'latest' Hub tag if it exists
+            api = HfApi(token=self.hub_token)
+            if self._previous_latest_hub_tag:
+                try:
+                    logger.info(f"🗑️ Removing previous 'latest' Hub tag: {self._previous_latest_hub_tag}")
+                    api.delete_tag(repo_id=hub_model_id, repo_type="model", tag=self._previous_latest_hub_tag)
+                    logger.info(f"✅ Deleted previous tag: {self._previous_latest_hub_tag}")
+                except Exception as e:
+                    logger.warning(f"⚠️ Could not delete previous Hub tag {self._previous_latest_hub_tag}: {e}")
+
+            logger.info(f"🚀 Uploading 'latest' checkpoint to Hub: {hub_model_id}")
+
+            hub_url, commit_id = self._upload_checkpoint_to_hub(
+                ckpt_dir=ckpt_dir,
+                hub_model_id=hub_model_id,
+                tag_name=tag_name,
+                commit_message=commit_message,
+            )
+            logger.info(f"✅ Successfully uploaded 'latest' to: {hub_url}")
+            logger.info(f"🏷️ Tagged as: {tag_name}")
+
+            # Track this as the new previous latest
+            self._previous_latest_hub_tag = tag_name
+            self._last_latest_hub_upload_step = step
+        elif self.upload_to_hub and self.hub_save_every is not None:
+            # Hub upload is enabled but not time yet
+            steps_until_upload = self.hub_save_every - (step - self._last_latest_hub_upload_step)
+            logger.info(
+                f"⏭️ Skipping Hub upload for 'latest' (saving locally only). Next upload in {steps_until_upload} steps"
+            )
+
+        # Track this as the new previous latest checkpoint directory
+        self._previous_latest_ckpt_dir = ckpt_dir
+
+        # Memory cleanup after saving model
+        self._cleanup_memory()
+
+
+def load_model_from_hf(
+    model_path: str,
+    device: torch.device,
+    hub_token: Optional[str] = None,
+) -> Tuple[Optional[ExperimentConfig], Optional[Any], Optional[Any], Optional[Any]]:
+    """
+    Load reward model config and model from HuggingFace or local checkpoint.
+
+    This mirrors the logic used by the training/eval scripts:
+    - Resolve checkpoint path (supports HF Hub with @ notation)
+    - Locate config.yaml locally (if model_path is a directory) or download from HF
+    - Use custom YAML loader for ReWiND configs
+    - Filter config keys to ExperimentConfig
+    - Clear training/logging sections
+    - Load model artifacts via setup_model_and_processor
+
+    Args:
+        model_path: HuggingFace model repository ID or local checkpoint path.
+                   Supports @ notation for tags: username/model@tag-name
+        device: Device to load model on
+        hub_token: Optional HuggingFace token for private repos
+
+    Returns:
+        Tuple of (exp_config, tokenizer, processor, reward_model)
+    """
+    # Resolve checkpoint path (handles HF Hub downloads with @ notation)
+    resolved_path = resolve_checkpoint_path(model_path, hub_token=hub_token)
+    if resolved_path is None:
+        raise ValueError(f"Could not resolve checkpoint path: {model_path}")
+
+    config_path: Optional[str] = None
+
+    # Parse repo_id and revision (tag) from model_path if using @tag format
+    # This is used for downloading config.yaml if needed
+    if "@" in model_path:
+        repo_id, revision = model_path.split("@", 1)
+    else:
+        repo_id, revision = model_path, None
+
+    resolved_path = Path(resolved_path)
+
+    if resolved_path.exists():
+        # Local checkpoint: look for config.yaml
+        candidate_paths = [
+            resolved_path / "config.yaml",
+            resolved_path.parent / "config.yaml",
+        ]
+        config_path = None
+        for candidate in candidate_paths:
+            if candidate.is_file():
+                config_path = candidate
+                break
+
+        # If config.yaml not found locally, try to download it from Hub
+        if config_path is None:
+            try:
+                from huggingface_hub import hf_hub_download
+            except ImportError:
+                raise ImportError("huggingface_hub not available. Install with: pip install huggingface_hub")
+
+            # Check if this is a HuggingFace repo (not a local path)
+            if "/" in repo_id and not repo_id.startswith("/"):
+                logger.info(
+                    f"config.yaml not found locally, downloading from HuggingFace Hub: {repo_id}@{revision or 'latest'}"
+                )
+                try:
+                    config_path = hf_hub_download(
+                        repo_id=repo_id, filename="config.yaml", revision=revision, token=hub_token
+                    )
+                    logger.info(f"Downloaded config.yaml to: {config_path}")
+                except Exception as e:
+                    logger.warning(f"Could not download config.yaml from Hub: {e}")
+                    raise ValueError(f"config.yaml not found locally and could not be downloaded from Hub: {e}")
+            else:
+                raise ValueError(f"config.yaml not found in checkpoint directory or parent directory: {resolved_path}")
+    else:
+        try:
+            from huggingface_hub import hf_hub_download
+        except ImportError:
+            raise ImportError("huggingface_hub not available. Install with: pip install huggingface_hub")
+        # Download config with revision if specified
+        logger.info(f"Downloading config.yaml from HuggingFace Hub: {repo_id}@{revision or 'latest'}")
+        config_path = hf_hub_download(repo_id=repo_id, filename="config.yaml", revision=revision, token=hub_token)
+        logger.info(f"Downloaded config.yaml to: {config_path}")
+
+    with open(config_path) as f:
+        yaml_text = f.read()
+
+    class _ConfigSafeLoader(yaml.SafeLoader):
+        pass
+
+    _ConfigSafeLoader.add_constructor(
+        "tag:yaml.org,2002:python/object:robometer.models.rewind_transformer.ReWINDTransformerConfig",
+        lambda loader, node: loader.construct_mapping(node),
+    )
+
+    model_config_dict = yaml.load(yaml_text, Loader=_ConfigSafeLoader)
+
+    valid_keys = {f.name for f in fields(ExperimentConfig)}
+    filtered_config = {k: v for k, v in model_config_dict.items() if k in valid_keys}
+
+    exp_config = ExperimentConfig(**filtered_config)
+    # Use resolved_path for loading the actual model
+    # Import here to avoid circular dependency with setup_utils
+    from robometer.utils.setup_utils import setup_model_and_processor
+
+    # Extract PEFT config from the loaded experiment config
+    peft_config = exp_config.peft if hasattr(exp_config, "peft") and exp_config.model.use_peft else None
+
+    tokenizer, processor, reward_model = setup_model_and_processor(
+        exp_config.model, str(resolved_path), peft_config=peft_config
+    )
+    reward_model = reward_model.to(device)
+    reward_model.eval()
+
+    return exp_config, tokenizer, processor, reward_model
+
+
+def load_wandb_run_info(model_path: str, hub_token: Optional[str] = None) -> Optional[dict[str, Any]]:
+    """
+    Retrieve saved wandb metadata for a checkpoint.
+
+    Checks for a local `wandb_info.json` (written during training) and, if the
+    checkpoint lives on HuggingFace, falls back to parsing the README that
+    `upload_to_hub.py` generates (which embeds wandb fields).
+
+    Args:
+        model_path: HuggingFace model repository ID or local checkpoint path.
+                   Supports @ notation for tags: username/model@tag-name
+        hub_token: Optional HuggingFace token for private repos
+    """
+
+    def _load_json(path: Path) -> Optional[dict[str, Any]]:
+        try:
+            with open(path) as f:
+                return json.load(f)
+        except (FileNotFoundError, json.JSONDecodeError):
+            return None
+
+    # Resolve checkpoint path first
+    resolved_path = resolve_checkpoint_path(model_path, hub_token=hub_token)
+    if resolved_path:
+        path = Path(resolved_path)
+        if path.exists():
+            candidates = []
+            if path.is_file():
+                candidates.append(path.parent / "wandb_info.json")
+            else:
+                candidates.append(path / "wandb_info.json")
+                candidates.append(path.parent / "wandb_info.json")
+            for candidate in candidates:
+                info = _load_json(candidate)
+                if info:
+                    return info
+
+    try:
+        from huggingface_hub import hf_hub_download
+    except ImportError:
+        return None
+
+    # Parse repo_id and revision (tag) from model_path if using @tag format
+    if "@" in model_path:
+        repo_id, revision = model_path.split("@", 1)
+    else:
+        repo_id, revision = model_path, None
+
+    try:
+        readme_path = hf_hub_download(
+            repo_id=repo_id, filename="README.md", revision=revision, token=hub_token, local_files_only=False
+        )
+    except Exception:
+        return None
+
+    try:
+        readme_text = Path(readme_path).read_text()
+    except OSError:
+        return None
+
+    wandb_info: dict[str, Any] = {}
+
+    run_match = re.search(r"\*\*Wandb Run\*\*:\s*\[(?P<name>.+?)\]\((?P<url>.+?)\)", readme_text)
+    if run_match:
+        wandb_info["wandb_name"] = run_match.group("name")
+        wandb_info["wandb_url"] = run_match.group("url")
+
+    id_match = re.search(r"\*\*Wandb ID\*\*:\s*`(?P<id>[^`]+)`", readme_text)
+    if id_match:
+        wandb_info["wandb_id"] = id_match.group("id")
+
+    project_match = re.search(r"\*\*Project\*\*:\s*(?P<project>[^\n]+)", readme_text)
+    if project_match:
+        wandb_info["wandb_project"] = project_match.group("project").strip()
+
+    entity_match = re.search(r"\*\*Entity\*\*:\s*(?P<entity>[^\n]+)", readme_text)
+    if entity_match:
+        wandb_info["wandb_entity"] = entity_match.group("entity").strip()
+
+    return wandb_info or None
diff --git a/robometer/robometer/utils/setup_utils.py b/robometer/robometer/utils/setup_utils.py
new file mode 100755
index 0000000000000000000000000000000000000000..c2e042036c4651a485155beabd948ad714385ee1
--- /dev/null
+++ b/robometer/robometer/utils/setup_utils.py
@@ -0,0 +1,1203 @@
+#!/usr/bin/env python3
+"""
+Shared setup utilities for RBM training.
+This file contains setup functions that can be reused across different training scripts.
+"""
+
+from unsloth import FastVisionModel
+
+import re
+import os
+from pathlib import Path
+from typing import Tuple, Optional, Any
+import torch
+from safetensors.torch import load_file
+from peft import LoraConfig, get_peft_model, PeftModel
+import bitsandbytes as bnb
+from huggingface_hub import HfApi
+from transformers import (
+    AutoImageProcessor,
+    AutoModel,
+    AutoModelForImageTextToText,
+    AutoProcessor,
+    AutoTokenizer,
+    Qwen2_5_VLForConditionalGeneration,
+    Qwen2_5_VLModel,
+    TrainingArguments,
+    BitsAndBytesConfig,
+)
+
+# Try to import Qwen3 models if available
+try:
+    from transformers import Qwen3VLForConditionalGeneration, Qwen3VLModel
+
+    HAS_QWEN3 = True
+except ImportError:
+    HAS_QWEN3 = False
+    Qwen3VLForConditionalGeneration = None
+    Qwen3VLModel = None
+
+from robometer.configs.experiment_configs import (
+    DataConfig,
+    ExperimentConfig,
+    ModelConfig,
+    PEFTConfig,
+    TrainingConfig,
+)
+from robometer.data.collators import BaseCollator, ReWiNDBatchCollator, RBMBatchCollator
+from robometer.data.datasets import (
+    RBMDataset,
+    StrategyFirstDataset,
+    BaseDataset,
+    RepeatedDataset,
+)
+from robometer.data.datasets.custom_eval import CustomEvalDataset
+from robometer.models import RBM, ReWiNDTransformer, ReWINDTransformerConfig
+from robometer.utils.logger import get_logger
+
+logger = get_logger()
+from robometer.utils.save import parse_hf_model_id_and_revision, resolve_checkpoint_path
+
+
+def _load_checkpoint_weights_from_safetensors(
+    model, checkpoint_path: str, cfg: ModelConfig, load_adapters: bool = True
+) -> None:
+    """
+    Load checkpoint weights from safetensors files in a checkpoint directory.
+    Includes verification for PEFT adapters and progress_head.
+
+    This is needed when using Unsloth, as we can't use from_pretrained on checkpoints.
+    Instead, we load the base model with Unsloth first, then manually load the checkpoint weights.
+
+    Args:
+        model: The model to load weights into
+        checkpoint_path: Path to checkpoint directory containing safetensors files
+        cfg: Model configuration for verification
+        load_adapters: If False, skip loading adapter weights (assumes already loaded via PeftModel.from_pretrained)
+    """
+
+    checkpoint_path = Path(checkpoint_path)
+    if not checkpoint_path.exists():
+        raise ValueError(f"Checkpoint path does not exist: {checkpoint_path}")
+
+    if not checkpoint_path.is_dir():
+        raise ValueError(f"Checkpoint path is not a directory: {checkpoint_path}")
+
+    # Collect all safetensors files
+    safetensors_files = list(checkpoint_path.glob("*.safetensors"))
+    if not safetensors_files:
+        raise ValueError(f"No safetensors files found in checkpoint directory: {checkpoint_path}")
+
+    logger.info(f"Loading checkpoint weights from {len(safetensors_files)} safetensors file(s) in {checkpoint_path}")
+
+    # Capture before weights for verification (adapter and progress_head)
+    before_weights = {}
+    before_progress_head = model.progress_head[0].weight.clone()
+    before_weights["progress_head"] = before_progress_head
+
+    # Capture adapter weights before loading (if PEFT is enabled)
+    adapter_keys_before = []
+    sample_adapter_keys = []
+    if cfg.use_peft:
+        model_state_dict = model.state_dict()
+        adapter_keys_before = [k for k in model_state_dict.keys() if "lora_A" in k or "lora_B" in k]
+        if adapter_keys_before:
+            # Sample a few adapter weights to verify they change
+            sample_adapter_keys = adapter_keys_before[:3]  # Check first 3 adapters
+            for key in sample_adapter_keys:
+                before_weights[f"adapter_{key}"] = model_state_dict[key].clone()
+            logger.info(f"Found {len(adapter_keys_before)} adapter parameters in model before loading checkpoint")
+        else:
+            logger.warning("No adapter parameters found in model - PEFT may not be applied correctly")
+
+    # Load all safetensors files and merge into a single state dict
+    checkpoint_state_dict = {}
+    for safetensors_file in safetensors_files:
+        logger.debug(f"Loading weights from {safetensors_file.name}")
+        file_state_dict = load_file(str(safetensors_file))
+        checkpoint_state_dict.update(file_state_dict)
+
+    # Check what adapter keys are in checkpoint
+    checkpoint_adapter_keys = [k for k in checkpoint_state_dict.keys() if "lora_A" in k or "lora_B" in k]
+    logger.info(f"Found {len(checkpoint_adapter_keys)} adapter parameters in checkpoint")
+    if checkpoint_adapter_keys:
+        logger.debug(f"Sample checkpoint adapter keys: {checkpoint_adapter_keys[:5]}")
+
+    # If load_adapters=False, filter out adapter keys (they were already loaded via PeftModel.from_pretrained)
+    if not load_adapters:
+        logger.info("Skipping adapter weights (already loaded via PeftModel.from_pretrained)")
+        checkpoint_state_dict = {
+            k: v for k, v in checkpoint_state_dict.items() if "lora_A" not in k and "lora_B" not in k
+        }
+
+    # Get model's expected state dict keys
+    model_state_dict = model.state_dict()
+    model_keys = set(model_state_dict.keys())
+
+    # Log sample model keys to understand structure
+    sample_model_keys = [k for k in list(model_keys)[:10]]
+    logger.info(f"Sample model keys (first 10): {sample_model_keys}")
+
+    # Log sample checkpoint keys to understand structure
+    sample_ckpt_keys = [k for k in list(checkpoint_state_dict.keys())[:10]]
+    logger.info(f"Sample checkpoint keys (first 10): {sample_ckpt_keys}")
+
+    # Check if model has adapter keys and what structure they use
+    model_adapter_keys = [k for k in model_keys if "lora_A" in k or "lora_B" in k]
+    if model_adapter_keys:
+        logger.info(f"Found {len(model_adapter_keys)} adapter keys in model")
+        logger.debug(f"Sample model adapter keys: {model_adapter_keys[:3]}")
+
+    # Remap checkpoint keys to match model structure
+    # For PEFT models wrapped in RBM: checkpoint has "model.model." but model expects "model.base_model.model.model."
+    # Try multiple strategies: direct match, map "model.model." -> "model.base_model.model.model.", etc.
+    remapped_state_dict = {}
+    remapped_count = 0
+    direct_match_count = 0
+    remap_strategies = {}
+
+    for ckpt_key, ckpt_value in checkpoint_state_dict.items():
+        if ckpt_key in model_keys:
+            # Direct match - use as is
+            remapped_state_dict[ckpt_key] = ckpt_value
+            direct_match_count += 1
+        else:
+            # Try different remapping strategies
+            potential_keys = []
+
+            # Strategy 1 (PEFT): Map "model.model." -> "model.base_model.model.model." (for PEFT wrapped in RBM)
+            # This handles the case where Unsloth saved the full model from model.model
+            if ckpt_key.startswith("model.model."):
+                # For PEFT: model.model.* -> model.base_model.model.model.*
+                peft_key = ckpt_key.replace("model.model.", "model.base_model.model.model.", 1)
+                potential_keys.append(peft_key)
+                # Also try: model.model.* -> model.* (fallback)
+                potential_keys.append(ckpt_key.replace("model.model.", "model.", 1))
+
+            # Strategy 2: Remove "model." prefix entirely
+            if ckpt_key.startswith("model."):
+                potential_keys.append(ckpt_key.replace("model.", "", 1))
+
+            # Strategy 3: Add "model." prefix if missing
+            if not ckpt_key.startswith("model."):
+                potential_keys.append(f"model.{ckpt_key}")
+
+            # Strategy 4: Try adding "base_model." between "model." and the rest (for non-PEFT wrapped models)
+            if ckpt_key.startswith("model.") and not ckpt_key.startswith("model.base_model."):
+                parts = ckpt_key.split(".", 1)
+                if len(parts) == 2:
+                    potential_keys.append(f"model.base_model.{parts[1]}")
+
+            # Try each potential key
+            matched = False
+            for potential_key in potential_keys:
+                if potential_key in model_keys:
+                    remapped_state_dict[potential_key] = ckpt_value
+                    remapped_count += 1
+                    strategy = f"{ckpt_key} -> {potential_key}"
+                    remap_strategies[strategy] = remap_strategies.get(strategy, 0) + 1
+                    if remapped_count <= 5:  # Log first 5 remappings
+                        logger.debug(f"Remapped: {strategy}")
+                    matched = True
+                    break
+
+            if not matched:
+                # Key still doesn't match, will be in unexpected_keys
+                remapped_state_dict[ckpt_key] = ckpt_value
+
+    if remapped_count > 0:
+        logger.info(
+            f"Remapped {remapped_count} checkpoint keys to match model structure (direct matches: {direct_match_count})"
+        )
+        if remap_strategies:
+            logger.debug(f"Remapping strategies used: {dict(list(remap_strategies.items())[:5])}")
+
+    # Load remapped state dict into model with strict=False to handle missing keys
+    missing_keys, unexpected_keys = model.load_state_dict(remapped_state_dict, strict=False)
+
+    # Filter missing keys - base model keys are expected for PEFT checkpoints
+    base_model_missing = [
+        k
+        for k in missing_keys
+        if any(
+            pattern in k for pattern in ["visual.", "language_model.", "text_encoder.", "text_model.", "embed_tokens"]
+        )
+    ]
+    other_missing = [k for k in missing_keys if k not in base_model_missing]
+
+    if base_model_missing:
+        logger.info(f"Missing base model keys (expected for PEFT checkpoints): {len(base_model_missing)} keys")
+    if other_missing:
+        logger.warning(f"Missing non-base model keys: {len(other_missing)} keys")
+        logger.debug(
+            f"Missing keys: {other_missing[:10]}..." if len(other_missing) > 10 else f"Missing keys: {other_missing}"
+        )
+
+    # Filter unexpected keys - adapter keys might be expected if structure differs slightly
+    adapter_unexpected = [k for k in unexpected_keys if "lora_A" in k or "lora_B" in k]
+    other_unexpected = [k for k in unexpected_keys if k not in adapter_unexpected]
+
+    if adapter_unexpected:
+        logger.warning(f"Unexpected adapter keys in checkpoint (not in model): {len(adapter_unexpected)} keys")
+        logger.debug(
+            f"Unexpected adapter keys: {adapter_unexpected[:10]}..."
+            if len(adapter_unexpected) > 10
+            else f"Unexpected adapter keys: {adapter_unexpected}"
+        )
+    if other_unexpected:
+        logger.warning(f"Unexpected non-adapter keys: {len(other_unexpected)} keys")
+        logger.debug(
+            f"Unexpected keys: {other_unexpected[:10]}..."
+            if len(other_unexpected) > 10
+            else f"Unexpected keys: {other_unexpected}"
+        )
+
+    # Verify progress_head loaded correctly
+    after_progress_head = model.progress_head[0].weight
+    progress_head_loaded = not torch.allclose(before_progress_head, after_progress_head, atol=1e-6)
+
+    logger.info(
+        f"Progress head - Before: shape={before_progress_head.shape}, sum={before_progress_head.sum():.6f} | "
+        f"After: shape={after_progress_head.shape}, sum={after_progress_head.sum():.6f} | "
+        f"Loaded: {progress_head_loaded}"
+    )
+
+    if not progress_head_loaded:
+        logger.error("Progress head weights did not change after loading checkpoint!")
+        logger.error("This indicates the checkpoint weights were not loaded correctly.")
+        import ipdb
+
+        ipdb.set_trace()  # Breakpoint if progress_head didn't load
+
+    # Verify adapter weights loaded correctly (if PEFT is enabled)
+    adapter_loaded_correctly = True
+    if cfg.use_peft and adapter_keys_before:
+        model_state_dict_after = model.state_dict()
+        adapter_keys_after = [k for k in model_state_dict_after.keys() if "lora_A" in k or "lora_B" in k]
+
+        logger.info(f"Adapter keys - Before: {len(adapter_keys_before)} | After: {len(adapter_keys_after)}")
+
+        # Check if sample adapter weights changed
+        for key in sample_adapter_keys:
+            if key in before_weights:
+                before_adapter = before_weights[f"adapter_{key}"]
+                if key in model_state_dict_after:
+                    after_adapter = model_state_dict_after[key]
+                    adapter_changed = not torch.allclose(before_adapter, after_adapter, atol=1e-6)
+                    logger.info(
+                        f"Adapter {key} - Before: shape={before_adapter.shape}, sum={before_adapter.sum():.6f} | "
+                        f"After: shape={after_adapter.shape}, sum={after_adapter.sum():.6f} | "
+                        f"Loaded: {adapter_changed}"
+                    )
+                    if not adapter_changed:
+                        logger.warning(f"Adapter {key} weights did not change after loading checkpoint!")
+                        adapter_loaded_correctly = False
+                else:
+                    logger.warning(f"Adapter key {key} not found in model after loading!")
+                    adapter_loaded_correctly = False
+
+        # Check how many adapter keys from checkpoint were actually loaded
+        # Need to check both original and remapped keys (matching the remapping strategies above)
+        loaded_adapter_keys = []
+        for ckpt_key in checkpoint_adapter_keys:
+            # Check if original key exists in model
+            if ckpt_key in model_state_dict_after:
+                loaded_adapter_keys.append(ckpt_key)
+            else:
+                # Try remapping strategies to find the actual key used in model
+                potential_keys = []
+                if ckpt_key.startswith("model.model."):
+                    # Strategy 1: PEFT wrapped in RBM
+                    potential_keys.append(ckpt_key.replace("model.model.", "model.base_model.model.model.", 1))
+                    # Strategy 2: Fallback
+                    potential_keys.append(ckpt_key.replace("model.model.", "model.", 1))
+
+                # Check if any remapped key exists in model
+                for remapped_key in potential_keys:
+                    if remapped_key in model_state_dict_after:
+                        loaded_adapter_keys.append(ckpt_key)  # Count original key as loaded
+                        break
+        logger.info(f"Loaded {len(loaded_adapter_keys)}/{len(checkpoint_adapter_keys)} adapter keys from checkpoint")
+
+        if len(loaded_adapter_keys) == 0:
+            logger.error("No adapter weights were loaded from checkpoint!")
+            adapter_loaded_correctly = False
+        elif len(loaded_adapter_keys) < len(checkpoint_adapter_keys) * 0.5:  # Less than 50% loaded
+            logger.warning(
+                f"Only {len(loaded_adapter_keys)}/{len(checkpoint_adapter_keys)} adapter keys loaded - may indicate structure mismatch"
+            )
+            adapter_loaded_correctly = False
+
+    if not adapter_loaded_correctly:
+        logger.error("Adapter weights did not load correctly!")
+        import ipdb
+
+        ipdb.set_trace()  # Breakpoint if adapters didn't load correctly
+
+    logger.info(f"Successfully loaded checkpoint weights from {checkpoint_path}")
+
+
+def _load_base_model_with_unsloth(
+    cfg: ModelConfig,
+    torch_dtype: torch.dtype,
+    extra_kwargs: dict,
+    peft_config: Optional[PEFTConfig] = None,
+    loading_from_checkpoint: bool = False,
+    apply_peft: bool = True,
+) -> Tuple[Any, Any]:
+    """
+    Load base model using Unsloth's FastVisionModel.
+
+    Args:
+        cfg: Model configuration
+        torch_dtype: Torch dtype to use
+        extra_kwargs: Extra kwargs for model loading (e.g., attn_implementation)
+        peft_config: Optional PEFT configuration
+        loading_from_checkpoint: If True, skip PEFT application (checkpoint already has weights)
+        apply_peft: If False, do not apply PEFT (e.g. when loading from checkpoint that has no adapter files; PEFT added later in train.py)
+
+    Returns:
+        Tuple of (base_model, tokenizer)
+    """
+    logger.info("Using Unsloth for faster training with Qwen model")
+
+    # Load model with unsloth
+    base_model, tokenizer = FastVisionModel.from_pretrained(
+        cfg.base_model_id,
+        load_in_4bit=cfg.quantization,  # Use 4bit if quantization is enabled
+        use_gradient_checkpointing="unsloth",  # Use unsloth's optimized checkpointing
+        dtype=torch_dtype,  # Set the dtype from config,
+        full_finetuning=True if not cfg.use_peft else False,
+        device_map=None,
+        attn_implementation=extra_kwargs["attn_implementation"],
+        trust_remote_code=True,
+    )
+
+    # Apply PEFT if enabled (skip when apply_peft=False, e.g. checkpoint has no adapter files; train.py will add PEFT later)
+    if apply_peft and cfg.use_peft and peft_config:
+        if loading_from_checkpoint:
+            logger.info("Applying PEFT configuration to base model (needed to load adapter weights from checkpoint)")
+        else:
+            logger.info("Applying PEFT configuration to base model")
+        base_model = FastVisionModel.get_peft_model(
+            base_model,
+            finetune_vision_layers=cfg.train_vision_encoder,
+            finetune_language_layers=cfg.train_language_model,
+            finetune_attention_modules=True,
+            finetune_mlp_modules=True,
+            r=peft_config.r,
+            lora_alpha=peft_config.lora_alpha,
+            lora_dropout=peft_config.lora_dropout,
+            bias=peft_config.bias,
+        )
+    elif cfg.use_peft and not apply_peft:
+        logger.info("Skipping PEFT here; checkpoint has no adapter files, PEFT will be added in train.py")
+
+    # Extract inner model after PEFT is applied (if needed for RBM wrapper)
+    # IMPORTANT: After FastVisionModel.get_peft_model(), base_model.model should be a PeftModel
+    # We keep it as PeftModel so we can use PeftModel.from_pretrained() later
+    if cfg.model_type == "default":
+        # Extract the inner model, which should be a PeftModel if PEFT was applied
+        inner_model = base_model.model
+        # Check if it's a PeftModel - if so, keep it as PeftModel for proper adapter loading
+        if cfg.use_peft and isinstance(inner_model, PeftModel):
+            logger.info(
+                "Base model inner model is a PeftModel - will use PeftModel.from_pretrained() for adapter loading"
+            )
+            base_model = inner_model
+        else:
+            base_model = inner_model
+
+    return base_model, tokenizer
+
+
+def _load_base_model_standard(
+    cfg: ModelConfig,
+    torch_dtype: torch.dtype,
+    extra_kwargs: dict,
+    bnb: Optional[BitsAndBytesConfig],
+) -> Any:
+    """
+    Load base model using standard transformers loading.
+
+    Args:
+        cfg: Model configuration
+        torch_dtype: Torch dtype to use
+        extra_kwargs: Extra kwargs for model loading (e.g., attn_implementation)
+        bnb: Optional BitsAndBytesConfig for quantization
+
+    Returns:
+        Base model
+    """
+    # Check if it's Molmo, Qwen3 or Qwen2/2.5
+    is_molmo = "Molmo" in cfg.base_model_id
+    is_qwen3 = ("Qwen3" in cfg.base_model_id or "qwen3" in cfg.base_model_id.lower()) and HAS_QWEN3
+
+    # Select appropriate model classes based on version and model type
+    if is_molmo:
+        # Molmo2 uses AutoModelForImageTextToText with trust_remote_code
+        base_model = AutoModelForImageTextToText.from_pretrained(
+            cfg.base_model_id,
+            torch_dtype=torch_dtype,
+            trust_remote_code=cfg.trust_remote_code,
+            **extra_kwargs,
+            quantization_config=bnb,
+        )
+        # Extract the base model for RBM
+        base_model = base_model.model
+        logger.info("Using Molmo2 models")
+    elif is_qwen3:
+        qwen_model_cls = Qwen3VLModel
+        base_model = qwen_model_cls.from_pretrained(
+            cfg.base_model_id,
+            torch_dtype=torch_dtype,
+            **extra_kwargs,
+            quantization_config=bnb,
+        )
+        logger.info("Using Qwen3 models")
+    else:
+        qwen_model_cls = Qwen2_5_VLModel if cfg.model_type == "default" else Qwen2_5_VLForConditionalGeneration
+        base_model = qwen_model_cls.from_pretrained(
+            cfg.base_model_id,
+            torch_dtype=torch_dtype,
+            **extra_kwargs,
+            quantization_config=bnb,
+        )
+        logger.info("Using Qwen2/2.5 models")
+
+    return base_model
+
+
+def _setup_processor_and_tokenizer(cfg: ModelConfig) -> AutoProcessor:
+    """
+    Setup processor and tokenizer for the model.
+
+    Args:
+        cfg: Model configuration
+
+    Returns:
+        Processor
+    """
+    if "SmolVLM" in cfg.base_model_id:
+        processor = AutoProcessor.from_pretrained(
+            cfg.base_model_id,
+            trust_remote_code=cfg.trust_remote_code,
+            padding_side="left",
+            size={"longest_edge": 512},
+            max_image_size={"longest_edge": 512},
+            use_fast=True,
+        )
+        logger.info(f"SmolVLM Processor: {processor}")
+    elif "Qwen" in cfg.base_model_id or "Molmo" in cfg.base_model_id:
+        processor = AutoProcessor.from_pretrained(
+            cfg.base_model_id,
+            trust_remote_code=cfg.trust_remote_code,
+            do_sample_frames=False,  # disable frame sampling here since we do this in the data generator
+            # padding_side="left",
+            padding_side="right",
+        )
+        logger.info(f"Qwen Processor: {processor}")
+    else:
+        raise ValueError(f"Invalid base model id: {cfg.base_model_id}")
+
+    if processor.tokenizer.pad_token is None:
+        processor.tokenizer.pad_token = processor.tokenizer.eos_token
+
+    return processor
+
+
+def _add_special_tokens_and_resize(cfg: ModelConfig, processor: AutoProcessor, base_model: Any) -> None:
+    """
+    Add RBM special tokens and resize token embeddings if needed.
+
+    Args:
+        cfg: Model configuration
+        processor: Processor with tokenizer
+        base_model: Base model to resize embeddings for
+    """
+    # Add RBM special tokens if they don't exist
+    special_tokens = [
+        "<|split_token|>",
+        "<|reward_token|>",
+        "<|pref_token|>",
+        "<|sim_token|>",
+        "<|prog_token|>",  # Per-frame progress token
+    ]
+    logger.info(f"Before adding special tokens: {len(processor.tokenizer.get_vocab())}")
+    num_added = 0
+    for token in special_tokens:
+        if token not in processor.tokenizer.get_vocab():
+            added = processor.tokenizer.add_special_tokens({"additional_special_tokens": [token]})
+            num_added += added
+
+    logger.info(f"Added {num_added} special tokens")
+
+    base_model.resize_token_embeddings(len(processor.tokenizer))
+    logger.info(f"Resized token embeddings to {len(processor.tokenizer)}")
+
+    # import ipdb; ipdb.set_trace()
+    # # Resize token embeddings if new tokens were added
+    # vocab_size = (
+    #     base_model.config.text_config.vocab_size
+    #     if ("Qwen3" in cfg.base_model_id or "Molmo" in cfg.base_model_id)
+    #     else base_model.config.vocab_size
+    # )
+
+    # if len(processor.tokenizer) != vocab_size:
+    #     logger.info(f"Resizing token embeddings from {vocab_size} to {len(processor.tokenizer)}")
+
+    #     is_molmo = "Molmo" in cfg.base_model_id
+    #     if is_molmo:
+    #         # Custom resize for Molmo2 - its Molmo2Embedding stores embedding as a Parameter directly
+    #         new_vocab_size = len(processor.tokenizer)
+    #         _embed_layer = base_model.get_input_embeddings()
+
+    #         # Check if embedding is a Parameter (tensor) directly, or an nn.Embedding
+    #         if hasattr(_embed_layer, "embedding"):
+    #             old_embed_attr = _embed_layer.embedding
+
+    #             # Case 1: embedding is a Parameter (raw tensor)
+    #             if isinstance(old_embed_attr, torch.nn.Parameter):
+    #                 old_num_tokens, embedding_dim = old_embed_attr.shape
+
+    #                 # Create new parameter with expanded vocab
+    #                 new_embed_data = torch.zeros(
+    #                     new_vocab_size, embedding_dim, device=old_embed_attr.device, dtype=old_embed_attr.dtype
+    #                 )
+
+    #                 # Copy existing weights
+    #                 new_embed_data[:old_num_tokens] = old_embed_attr.data
+
+    #                 # Initialize new token embeddings using mean of existing embeddings
+    #                 mean_embedding = old_embed_attr.data.mean(dim=0)
+    #                 new_embed_data[old_num_tokens:] = mean_embedding.unsqueeze(0).expand(
+    #                     new_vocab_size - old_num_tokens, -1
+    #                 )
+
+    #                 # Replace the embedding Parameter
+    #                 _embed_layer.embedding = torch.nn.Parameter(new_embed_data)
+
+    #                 # Also update config to reflect new vocab size
+    #                 base_model.config.text_config.vocab_size = new_vocab_size
+
+    #                 logger.info(
+    #                     f"Custom resized Molmo2 embeddings (Parameter) from {old_num_tokens} to {new_vocab_size}"
+    #                 )
+
+    #             # Case 2: embedding is an nn.Embedding with .weight
+    #             elif hasattr(old_embed_attr, "weight"):
+    #                 old_num_tokens, embedding_dim = old_embed_attr.weight.shape
+
+    #                 # Create new embedding layer with expanded vocab
+    #                 new_embedding = torch.nn.Embedding(
+    #                     new_vocab_size,
+    #                     embedding_dim,
+    #                     device=old_embed_attr.weight.device,
+    #                     dtype=old_embed_attr.weight.dtype,
+    #                 )
+
+    #                 # Copy existing weights
+    #                 new_embedding.weight.data[:old_num_tokens] = old_embed_attr.weight.data
+
+    #                 # Initialize new token embeddings using mean of existing embeddings
+    #                 mean_embedding = old_embed_attr.weight.data.mean(dim=0)
+    #                 new_embedding.weight.data[old_num_tokens:] = mean_embedding.unsqueeze(0).expand(
+    #                     new_vocab_size - old_num_tokens, -1
+    #                 )
+
+    #                 # Replace the nested embedding
+    #                 _embed_layer.embedding = new_embedding
+
+    #                 # Also update config to reflect new vocab size
+    #                 base_model.config.text_config.vocab_size = new_vocab_size
+
+    #                 logger.info(
+    #                     f"Custom resized Molmo2 embeddings (Embedding) from {old_num_tokens} to {new_vocab_size}"
+    #                 )
+    #             else:
+    #                 logger.warning(f"Cannot resize Molmo2 embeddings - unknown embedding type: {type(old_embed_attr)}")
+    #         else:
+    #             logger.warning(f"Cannot resize Molmo2 embeddings - no embedding attribute found")
+    #     else:
+    #         base_model.resize_token_embeddings(len(processor.tokenizer))
+    #         logger.info(f"Resized token embeddings to {len(processor.tokenizer)}")
+
+
+def _verify_checkpoint_loading(cfg: ModelConfig, model: Any, before_weights: dict) -> None:
+    """
+    Verify that checkpoint weights were loaded correctly by comparing before/after weights.
+
+    Args:
+        cfg: Model configuration
+        model: The model after loading checkpoint
+        before_weights: Dictionary of weights before loading (keys: visual, progress_head, lm_embed_tokens, lm_layer)
+    """
+
+    if "Qwen2.5" in cfg.base_model_id:
+        after_visual = model.model.visual.blocks[0].mlp.down_proj.weight
+        after_progress_head = model.progress_head[0].weight
+        after_lm_embed_tokens = model.model.language_model.embed_tokens.weight
+        after_lm_layer = model.model.language_model.layers[0].mlp.up_proj.weight
+    elif "Qwen3" in cfg.base_model_id or "Molmo" in cfg.base_model_id:
+        after_visual = model.model.visual.blocks[0].mlp.linear_fc1.weight
+        after_progress_head = model.progress_head[0].weight
+        after_lm_embed_tokens = model.model.language_model.embed_tokens.weight
+        after_lm_layer = model.model.language_model.layers[0].mlp.up_proj.weight
+    else:
+        return
+
+    before_visual = before_weights["visual"]
+    before_progress_head = before_weights["progress_head"]
+    before_lm_embed_tokens = before_weights["lm_embed_tokens"]
+    before_lm_layer = before_weights["lm_layer"]
+
+    logger.info(
+        f"Before visual: {before_visual.shape}, {before_visual.sum()} | After visual: {after_visual.shape}, {after_visual.sum()}"
+    )
+    logger.info(
+        f"Before progress head: {before_progress_head.shape}, {before_progress_head.sum()} | After progress head: {after_progress_head.shape}, {after_progress_head.sum()}"
+    )
+    logger.info(
+        f"Before LM embed tokens: {before_lm_embed_tokens.shape}, {before_lm_embed_tokens.sum()} | After LM embed tokens: {after_lm_embed_tokens.shape}, {after_lm_embed_tokens.sum()}"
+    )
+    logger.info(
+        f"Before LM layer: {before_lm_layer.shape}, {before_lm_layer.sum()} | After LM layer: {after_lm_layer.shape}, {after_lm_layer.sum()}"
+    )
+
+    # check that before and after are different
+    if torch.allclose(before_visual, after_visual):
+        logger.warning("Before and after visual are the same! Check if you loaded the pretrained model correctly")
+    if torch.allclose(before_progress_head, after_progress_head):
+        logger.warning(
+            "Before and after progress head are the same! Check if you loaded the pretrained model correctly"
+        )
+    if torch.allclose(before_lm_embed_tokens, after_lm_embed_tokens):
+        logger.warning(
+            "Before and after LM embed tokens are the same! Check if you loaded the pretrained model correctly"
+        )
+
+
+def setup_model_and_processor(
+    cfg: ModelConfig, hf_model_id: str = "", peft_config: PEFTConfig = None
+) -> tuple[AutoProcessor, RBM]:
+    """
+    Shared function to set up model, processor, and tokenizer for both training and evaluation.
+
+    Args:
+        cfg: Model configuration
+        hf_model_id: Optional HuggingFace model ID to load from
+
+    Note:
+        When use_unsloth is enabled for Qwen models:
+        - The model will be loaded using unsloth's FastVisionModel
+        - Automatically uses optimized gradient checkpointing
+        - If use_peft is enabled, applies unsloth's optimized PEFT configuration
+        - Use unsloth/Qwen models for best performance (e.g., unsloth/Qwen2.5-VL-3B-Instruct-bnb-4bit)
+    """
+
+    # Convert string dtype to torch dtype (used across all model loading paths)
+    torch_dtype = getattr(torch, cfg.torch_dtype, torch.bfloat16)
+    logger.info(f"Using torch dtype: {torch_dtype}")
+
+    # Check if unsloth should be used
+    use_unsloth = cfg.use_unsloth and "Qwen" in cfg.base_model_id
+
+    if use_unsloth:
+        logger.info("Unsloth mode enabled for faster training")
+
+    # If quantization is enabled, use bitsandbytes (unless using unsloth)
+    if cfg.quantization and not use_unsloth:
+        bnb = BitsAndBytesConfig(
+            load_in_8bit=True,
+            llm_int8_enable_fp32_cpu_offload=True,
+        )
+    else:
+        bnb = None
+
+    try:
+        import flash_attn
+
+        logger.info("Flash Attention 2 CUDA is available")
+        has_flash_attn = True
+    except:
+        logger.info("Flash Attention 2 CUDA is not available")
+        has_flash_attn = False
+
+    if has_flash_attn:
+        extra_kwargs = {"attn_implementation": "flash_attention_2"}
+    else:
+        extra_kwargs = {"attn_implementation": "sdpa"}
+
+    # Determine if we're loading from a checkpoint and whether it has PEFT adapter files
+    loading_from_checkpoint = bool(hf_model_id)
+    checkpoint_path_for_load: Optional[str] = None
+    has_adapter_files = True  # default when not loading
+    if hf_model_id:
+        hub_token = os.environ.get("HF_TOKEN")
+        checkpoint_path_for_load = resolve_checkpoint_path(hf_model_id, hub_token=hub_token)
+        if checkpoint_path_for_load:
+            adapter_config_path = os.path.join(checkpoint_path_for_load, "adapter_config.json")
+            adapter_model_paths = [
+                os.path.join(checkpoint_path_for_load, "adapter_model.safetensors"),
+                os.path.join(checkpoint_path_for_load, "adapter_model.bin"),
+            ]
+            has_adapter_files = os.path.exists(adapter_config_path) and any(
+                os.path.exists(p) for p in adapter_model_paths
+            )
+            if cfg.use_peft and not has_adapter_files:
+                logger.info(
+                    "Checkpoint has no PEFT adapter files; will load base+heads then add PEFT in train.py"
+                )
+        else:
+            has_adapter_files = True  # treat as True so we don't skip PEFT
+    # When loading from checkpoint without adapters but use_peft: build base without PEFT, load weights, then train.py adds PEFT
+    apply_peft_before_wrap = cfg.use_peft and (not hf_model_id or has_adapter_files)
+
+    # Load processor and tokenizer
+    if "SmolVLM" in cfg.base_model_id or "Qwen" in cfg.base_model_id or "Molmo" in cfg.base_model_id:
+        if "SmolVLM" in cfg.base_model_id:
+            processor = AutoProcessor.from_pretrained(
+                cfg.base_model_id,
+                trust_remote_code=cfg.trust_remote_code,
+                padding_side="left",
+                size={"longest_edge": 512},
+                max_image_size={"longest_edge": 512},
+                use_fast=True,
+            )
+            logger.info(f"SmolVLM Processor: {processor}")
+
+            base_model = AutoModelForImageTextToText.from_pretrained(
+                cfg.base_model_id,
+                torch_dtype=torch_dtype,
+                **extra_kwargs,
+                quantization_config=bnb,
+            )
+            model_cls = RBM
+
+        elif "Qwen" in cfg.base_model_id or "Molmo" in cfg.base_model_id:
+            # Load base model (with or without Unsloth)
+            if use_unsloth:
+                base_model, tokenizer = _load_base_model_with_unsloth(
+                    cfg,
+                    torch_dtype,
+                    extra_kwargs,
+                    peft_config,
+                    loading_from_checkpoint=loading_from_checkpoint,
+                    apply_peft=apply_peft_before_wrap,
+                )
+            else:
+                base_model = _load_base_model_standard(cfg, torch_dtype, extra_kwargs, bnb)
+                tokenizer = None  # Will be loaded with processor
+
+            model_cls = RBM
+
+            # Setup processor and tokenizer
+            processor = _setup_processor_and_tokenizer(cfg)
+            if tokenizer is None:
+                tokenizer = processor.tokenizer
+
+        else:
+            raise ValueError(f"Invalid base model id: {cfg.base_model_id}")
+
+        # CRITICAL: Ensure PEFT is applied to base_model BEFORE wrapping in RBM (when checkpoint has adapters or we're not loading)
+        if apply_peft_before_wrap and cfg.use_peft and not isinstance(base_model, PeftModel):
+            logger.warning("PEFT is enabled but base_model is not a PeftModel. Applying PEFT now...")
+            if peft_config is None:
+                raise ValueError("PEFT is enabled but peft_config is None. Cannot apply PEFT without configuration.")
+
+            # Apply PEFT to base_model
+            from peft import LoraConfig, get_peft_model
+
+            lora_config = LoraConfig(
+                r=peft_config.r,
+                lora_alpha=peft_config.lora_alpha,
+                target_modules=peft_config.target_modules,
+                lora_dropout=peft_config.lora_dropout,
+                bias=peft_config.bias,
+            )
+            base_model = get_peft_model(base_model, lora_config)
+            logger.info("Applied PEFT to base_model before wrapping in RBM")
+
+        # Verify PEFT was applied when we expect it
+        if apply_peft_before_wrap and cfg.use_peft:
+            if isinstance(base_model, PeftModel):
+                logger.info("Confirmed: base_model is a PeftModel - ready to load adapter weights from checkpoint")
+            else:
+                logger.error("CRITICAL: PEFT is enabled but base_model is not a PeftModel after applying PEFT!")
+                raise ValueError(
+                    "Failed to apply PEFT to base_model. Cannot load adapter weights without PeftModel structure."
+                )
+
+        # Add special tokens and resize embeddings
+        _add_special_tokens_and_resize(cfg, processor, base_model)
+
+        # Initialize RBM model wrapper with the pre-loaded base model
+        logger.info("Initializing RBM model...")
+        tokenizer = processor.tokenizer
+
+        model = model_cls(
+            config=base_model.config,
+            processor=processor,
+            tokenizer=tokenizer,
+            base_model=base_model,
+            base_model_id=cfg.base_model_id,
+            model_config=cfg,  # Pass ModelConfig for RBM-specific settings
+        )
+
+        # Load checkpoint if provided
+        if hf_model_id:
+            repo_id, revision_to_load = parse_hf_model_id_and_revision(hf_model_id, model_name="model")
+            checkpoint_path = checkpoint_path_for_load
+            if checkpoint_path is None:
+                hub_token = os.environ.get("HF_TOKEN")
+                checkpoint_path = resolve_checkpoint_path(hf_model_id, hub_token=hub_token)
+            if checkpoint_path is None:
+                raise ValueError(f"Could not resolve checkpoint path: {hf_model_id}")
+
+            # When use_peft and checkpoint has adapter files: load via PeftModel.from_pretrained + custom heads
+            # When use_peft and checkpoint has NO adapter files: we built base without PEFT; load weights only; train.py will add PEFT
+            if cfg.use_peft:
+                if has_adapter_files:
+                    logger.info("Loading PEFT adapters using standard PeftModel.from_pretrained() method")
+                    if not isinstance(model.model, PeftModel):
+                        logger.error("CRITICAL: model.model is not a PeftModel! Cannot load adapter weights.")
+                        raise ValueError(
+                            "model.model is not a PeftModel. "
+                            "This should not happen if PEFT was applied correctly before wrapping in RBM."
+                        )
+                    logger.info("Checkpoint contains PEFT adapter files - loading using PeftModel.from_pretrained()")
+                    try:
+                        model.model = PeftModel.from_pretrained(model.model, checkpoint_path)
+                        logger.info("Successfully loaded PEFT adapters using PeftModel.from_pretrained()")
+                        logger.info("Loading custom heads only (adapters already loaded)")
+                        _load_checkpoint_weights_from_safetensors(model, checkpoint_path, cfg, load_adapters=False)
+                    except Exception as e:
+                        logger.warning(f"PeftModel.from_pretrained() failed: {e}")
+                        logger.info("Falling back to manual loading for all weights")
+                        _load_checkpoint_weights_from_safetensors(model, checkpoint_path, cfg, load_adapters=True)
+                else:
+                    # Checkpoint has no adapter files; we built base without PEFT, load base+heads only; train.py will add PEFT
+                    logger.info("Checkpoint has no PEFT adapter files - loading base + custom heads (PEFT will be added in train.py)")
+                    _load_checkpoint_weights_from_safetensors(model, checkpoint_path, cfg, load_adapters=True)
+            else:
+                # For non-PEFT models, we can use from_pretrained as before
+                # Capture before weights for verification
+                before_weights = {}
+                if "Qwen2.5" in cfg.base_model_id:
+                    before_weights = {
+                        "visual": model.model.visual.blocks[0].mlp.down_proj.weight,
+                        "progress_head": model.progress_head[0].weight,
+                        "lm_embed_tokens": model.model.language_model.embed_tokens.weight,
+                        "lm_layer": model.model.language_model.layers[0].mlp.up_proj.weight,
+                    }
+                elif "Qwen3" in cfg.base_model_id or "Molmo" in cfg.base_model_id:
+                    before_weights = {
+                        "visual": model.model.visual.blocks[0].mlp.linear_fc1.weight,
+                        "progress_head": model.progress_head[0].weight,
+                        "lm_embed_tokens": model.model.language_model.embed_tokens.weight,
+                        "lm_layer": model.model.language_model.layers[0].mlp.up_proj.weight,
+                    }
+
+                # Load the model from the evaluation path
+                model = model_cls.from_pretrained(
+                    repo_id,
+                    processor=processor,
+                    tokenizer=tokenizer,
+                    base_model=base_model,
+                    base_model_id=cfg.base_model_id,
+                    model_config=cfg,
+                    revision=revision_to_load,
+                )
+
+                # Verify weights were loaded
+                if before_weights:
+                    _verify_checkpoint_loading(cfg, model, before_weights)
+
+    # elif "rewind_transformer" in cfg.base_model_id or "rewind_scale_transformer" in cfg.base_model_id:
+    elif "rewind" in cfg.base_model_id:
+        # Initialize new model with encoders
+        # Pretrained image and text encoders
+        image_encoder = AutoModel.from_pretrained("facebook/dinov2-base")
+        text_encoder = AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L12-v2")
+        processor = AutoImageProcessor.from_pretrained("facebook/dinov2-base", use_fast=True)
+        tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L12-v2")
+
+        if hf_model_id:
+            repo_id, revision_to_load = parse_hf_model_id_and_revision(hf_model_id, model_name="ReWiND model")
+
+            model = ReWiNDTransformer.from_pretrained(
+                repo_id,
+                processor=processor,
+                image_encoder=image_encoder,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                revision=revision_to_load,
+            )
+        else:
+            train_img = cfg.train_vision_encoder
+            train_text = cfg.train_language_model
+
+            for p in image_encoder.parameters():
+                p.requires_grad = train_img
+
+            for p in text_encoder.parameters():
+                p.requires_grad = train_text
+
+            logger.info("Initializing ReWiND model...")
+            model = ReWiNDTransformer(
+                config=cfg,
+                processor=processor,
+                tokenizer=tokenizer,
+                image_encoder=image_encoder,
+                text_encoder=text_encoder,
+            )
+
+    logger.info("Model architecture initialized")
+    logger.info(f"Model architecture: {model}")
+
+    # Configure which parts of the model to train based on config
+    # IMPORTANT: When using PEFT (via Unsloth or standard), PEFT already handles freezing
+    # base model parameters. We should NOT override requires_grad on base model params.
+    peft_applied = cfg.use_peft and (cfg.use_unsloth or cfg.peft_vision_encoder)
+
+    # Helper function to check if a parameter is part of the base model (vision/language)
+    def is_base_model_param(name: str) -> bool:
+        """Check if parameter belongs to base model (vision/language) that PEFT handles."""
+        base_model_patterns = ["visual", "vision", "language_model", "text_encoder", "text_model", "image_encoder"]
+        return any(pattern in name for pattern in base_model_patterns)
+
+    # Helper function to check if a parameter is a prediction head
+    def is_prediction_head(name: str) -> bool:
+        """Check if parameter belongs to a prediction head."""
+        head_patterns = ["progress_head", "success_head", "preference_head", "similarity_head"]
+        return any(pattern in name for pattern in head_patterns)
+
+    for name, param in model.named_parameters():
+        # 1. Handle prediction heads - always controlled by their individual flags
+        if is_prediction_head(name):
+            if "progress_head" in name:
+                param.requires_grad = cfg.train_progress_head
+            elif "success_head" in name:
+                param.requires_grad = cfg.train_success_head
+            elif "preference_head" in name:
+                param.requires_grad = cfg.train_preference_head
+            elif "similarity_head" in name:
+                param.requires_grad = False
+
+        # 2. Handle base model parameters (vision/language) - skip if PEFT is applied
+        elif is_base_model_param(name):
+            if peft_applied:
+                # PEFT handles freezing/unfreezing - don't override
+                continue
+
+            # Set requires_grad based on config flags
+            if "visual" in name or "vision" in name or "vision_model" in name:
+                param.requires_grad = cfg.train_vision_encoder
+            elif "language_model" in name or "text_encoder" in name or "text_model" in name:
+                param.requires_grad = cfg.train_language_model
+            elif "image_encoder" in name:
+                param.requires_grad = cfg.train_vision_encoder
+
+        # 3. Handle special cases
+        elif "lm_head" in name:
+            # Language modeling head should not be trainable for RBM
+            param.requires_grad = False
+
+        # 4. All other parameters (custom RBM parameters like frame_pool_attn, video_proj, text_proj)
+        # should always be trainable
+        else:
+            param.requires_grad = True
+
+    logger.info("Training configuration:")
+    logger.info(f"  - Vision encoder: {cfg.train_vision_encoder}")
+    logger.info(f"  - Language model: {cfg.train_language_model}")
+    logger.info(f"  - Progress head: {cfg.train_progress_head}")
+    logger.info(f"  - Success head: {getattr(cfg, 'train_success_head', False)}")
+    logger.info(f"  - Preference head: {cfg.train_preference_head}")
+
+    # When use_peft, skip verbose param list here; it will be printed after PEFT in setup_peft_model
+    if not cfg.use_peft:
+        for name, param in model.named_parameters():
+            if param.requires_grad:
+                logger.info(f"{name:60} | {param.shape} | RG: {param.requires_grad}")
+
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    all_params = sum(p.numel() for p in model.parameters())
+    logger.info(
+        f"trainable params: {trainable_params:,} || all params: {all_params:,} || trainable%: {100 * trainable_params / all_params:.4f}"
+    )
+    return tokenizer, processor, model
+
+
+def _get_vl_inner_model(rbm_model: RBM):
+    """Return the inner VL model (with .visual and .language_model) from RBM, unwrapping PeftModel if present."""
+    m = rbm_model.model
+    if isinstance(m, PeftModel):
+        return m.get_base_model()
+    return m
+
+
+def setup_peft_model(rbm_model: RBM, cfg: PEFTConfig) -> RBM:
+    """Shared function to apply PEFT configuration to the model."""
+
+    logger.info("Using PEFT/LoRA training...")
+    lora_config = LoraConfig(
+        r=cfg.r,
+        lora_alpha=cfg.lora_alpha,
+        target_modules=cfg.target_modules,
+        lora_dropout=cfg.lora_dropout,
+        bias=cfg.bias,
+    )
+    inner = _get_vl_inner_model(rbm_model)
+    if cfg.peft_vision_encoder:
+        logger.info("Attaching LoRA to the vision encoder...")
+        inner.visual = get_peft_model(inner.visual, lora_config)
+    else:
+        # Default: attach LoRA to the language model so adapter weights are actually added
+        logger.info("Attaching LoRA to the language model...")
+        inner.language_model = get_peft_model(inner.language_model, lora_config)
+
+    # Print all trainable parameters after PEFT so adapter weights (lora_A, lora_B, etc.) are visible
+    logger.info("Trainable parameters (after PEFT):")
+    for name, param in rbm_model.named_parameters():
+        if param.requires_grad:
+            logger.info(f"  {name:70} | {param.shape} | RG: {param.requires_grad}")
+    trainable_params = sum(p.numel() for p in rbm_model.parameters() if p.requires_grad)
+    all_params = sum(p.numel() for p in rbm_model.parameters())
+    logger.info(
+        f"AFTER PEFT: trainable params: {trainable_params:,} || all params: {all_params:,} || trainable%: {100 * trainable_params / all_params:.4f}"
+    )
+    return rbm_model
+
+
+def create_training_arguments(cfg: TrainingConfig, output_dir: str, is_eval: bool = False) -> TrainingArguments:
+    """Shared function to create TrainingArguments for both training and evaluation"""
+
+    # Base arguments that are the same for both training and evaluation
+    base_args = {
+        "output_dir": output_dir,
+        "per_device_train_batch_size": cfg.per_device_train_batch_size,
+        "gradient_accumulation_steps": cfg.gradient_accumulation_steps,
+        "ddp_find_unused_parameters": cfg.ddp_find_unused_parameters,
+        "learning_rate": cfg.learning_rate,
+        "save_strategy": cfg.save_strategy,
+        "logging_steps": cfg.logging_steps,
+        "save_steps": cfg.save_steps,
+        "bf16": cfg.bf16,
+        "fp16": cfg.fp16,
+        "remove_unused_columns": cfg.remove_unused_columns,
+        "gradient_checkpointing": cfg.gradient_checkpointing,
+        "dataloader_pin_memory": cfg.dataloader_pin_memory,
+        "dataloader_num_workers": cfg.dataloader_num_workers,
+        "dataloader_persistent_workers": cfg.dataloader_persistent_workers,
+        "save_safetensors": True,
+        "save_total_limit": 2,
+        # Evaluation settings
+        "eval_strategy": cfg.evaluation_strategy,
+        "per_device_eval_batch_size": cfg.per_device_eval_batch_size,
+        "do_eval": cfg.do_eval,
+        "prediction_loss_only": cfg.prediction_loss_only,
+        "lr_scheduler_type": cfg.lr_scheduler_type,
+        "warmup_steps": cfg.warmup_steps,
+        "warmup_ratio": cfg.warmup_ratio,
+        "max_grad_norm": cfg.max_grad_norm,
+        "weight_decay": cfg.weight_decay,
+        "disable_tqdm": False,
+        # # Compile settings
+        # "torch_compile": True,
+        # "torch_compile_mode": "max-autotune",
+        # "torch_compile_backend": "inductor",
+    }
+
+    # Add eval_steps if evaluation_strategy is "steps"
+    if cfg.evaluation_strategy == "steps" and cfg.eval_steps is not None:
+        base_args["eval_steps"] = cfg.eval_steps
+
+    if is_eval:
+        # Evaluation-specific arguments
+        base_args.update({
+            "per_device_eval_batch_size": 2,
+            "num_train_epochs": -1,
+            "max_steps": 1,
+            "report_to": [],
+        })
+    else:
+        # Training-specific arguments
+        base_args.update({
+            "num_train_epochs": cfg.num_train_epochs if cfg.num_train_epochs is not None else 1,
+            "max_steps": cfg.max_steps if cfg.max_steps is not None else -1,
+            # Disable HuggingFace's automatic logging - we use custom Logger class instead
+            "report_to": [],
+        })
+
+    return TrainingArguments(**base_args)
+
+
+def setup_dataset(cfg: DataConfig, is_eval: bool = False, sampler_kwargs=None, **kwargs) -> BaseDataset:
+    """Shared function to create Dataset for training or evaluation
+
+    Args:
+        cfg: Data configuration
+        is_eval: Whether this is for evaluation
+        sampler_kwargs: Optional dictionary of keyword arguments to pass to samplers
+        **kwargs: Additional keyword arguments to pass to dataset
+    """
+    dataset_cls = {
+        "rbm": RBMDataset,
+        "strategy_first": StrategyFirstDataset,
+    }
+
+    if cfg.dataset_type not in dataset_cls:
+        raise ValueError(f"Unknown dataset_type: {cfg.dataset_type}. Must be one of: {list(dataset_cls.keys())}")
+
+    if sampler_kwargs is None:
+        sampler_kwargs = {}
+
+    sampler_kwargs["random_seed"] = cfg.seed
+    kwargs["sampler_kwargs"] = sampler_kwargs
+
+    # Create the base dataset
+    dataset = dataset_cls[cfg.dataset_type](config=cfg, is_evaluation=is_eval, **kwargs)
+
+    if not is_eval:
+        dataset = RepeatedDataset(dataset)
+    return dataset
+
+
+def setup_custom_eval_dataset(cfg: DataConfig, sampler_type: str, verbose=True, sampler_kwargs=None):
+    """Setup a custom evaluation dataset with a specific sampler."""
+    custom_eval_dataset = CustomEvalDataset(sampler_type, cfg, verbose=verbose, sampler_kwargs=sampler_kwargs)
+
+    return custom_eval_dataset
+
+
+def setup_batch_collator(
+    processor: AutoProcessor, tokenizer: AutoTokenizer, cfg: ExperimentConfig, is_eval: bool = False
+) -> BaseCollator:
+    """Shared function to create BatchCollator"""
+    collator_kwargs = {
+        "processor": processor,
+        "resized_height": cfg.data.resized_height,
+        "resized_width": cfg.data.resized_width,
+        "base_model_id": cfg.model.base_model_id,
+        "use_multi_image": cfg.data.use_multi_image,
+        "prog_pref": cfg.training.predict_pref_progress,
+        "use_per_frame_progress_token": getattr(cfg.data, "use_per_frame_progress_token", False),
+        "shuffle_progress_frames": cfg.data.shuffle_progress_frames,
+        "inference": is_eval,
+    }
+    # Check for unsupported Molmo2 video mode
+    if "Molmo" in cfg.model.base_model_id and not cfg.data.use_multi_image:
+        raise ValueError(
+            "Molmo2 implementation does not yet support video mode as it requires extra imports (use_multi_image=False). "
+            "Please set data.use_multi_image=True to use Molmo2 with multi-image input."
+        )
+
+    if "Qwen" in cfg.model.base_model_id or "SmolVLM" in cfg.model.base_model_id or "Molmo" in cfg.model.base_model_id:
+        batch_collator = RBMBatchCollator(**collator_kwargs)
+    # elif "rewind_transformer" in cfg.model.base_model_id:
+    elif "rewind" in cfg.model.base_model_id:
+        batch_collator = ReWiNDBatchCollator(
+            **collator_kwargs, tokenizer=tokenizer, load_embeddings=cfg.data.load_embeddings
+        )
+    return batch_collator
diff --git a/robometer/robometer/utils/tensor_utils.py b/robometer/robometer/utils/tensor_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..fb25cdfc161a39e4be0afe4527320fd678bd722e
--- /dev/null
+++ b/robometer/robometer/utils/tensor_utils.py
@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""
+Tensor utility functions for common tensor operations.
+"""
+
+import torch
+import numpy as np
+
+
+def t2n(tensor, dtype=torch.float32):
+    """
+    Convert tensor to numpy array.
+
+    Args:
+        tensor: PyTorch tensor to convert
+        dtype: Target dtype for conversion (default: torch.float32)
+
+    Returns:
+        numpy.ndarray: Converted numpy array
+    """
+    if tensor is None:
+        return None
+    if isinstance(tensor, np.ndarray):
+        return tensor
+    if not torch.is_tensor(tensor):
+        return np.array(tensor)
+    return tensor.detach().to(dtype=dtype).cpu().numpy()
diff --git a/robometer/robometer/utils/timer.py b/robometer/robometer/utils/timer.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b9df0b76ad36b586340b37b9533011513f3cf44
--- /dev/null
+++ b/robometer/robometer/utils/timer.py
@@ -0,0 +1,40 @@
+import time
+from contextlib import contextmanager
+
+from codetiming import Timer
+
+
+@contextmanager
+def timer(name: str, verbose: bool = False):
+    """Context manager for timing operations."""
+    start_time = time.time()
+    if verbose:
+        print(f"    Starting {name}...")
+    try:
+        yield
+    finally:
+        end_time = time.time()
+        duration = end_time - start_time
+        if verbose:
+            print(f"    {name} completed in {duration:.2f}s")
+
+
+@contextmanager
+def _timer(name: str, timing_raw: dict[str, float]):
+    """Context manager for timing code execution.
+
+    This utility function measures the execution time of code within its context
+    and accumulates the timing information in the provided dictionary.
+
+    Args:
+        name (str): The name/identifier for this timing measurement.
+        timing_raw (Dict[str, float]): Dictionary to store timing information.
+
+    Yields:
+        None: This is a context manager that yields control back to the code block.
+    """
+    with Timer(name=name, logger=None) as timer:
+        yield
+    if name not in timing_raw:
+        timing_raw[name] = 0
+    timing_raw[name] += timer.last
diff --git a/robometer/robometer/utils/upload_to_hub.py b/robometer/robometer/utils/upload_to_hub.py
new file mode 100755
index 0000000000000000000000000000000000000000..905939112139ee19afd93d7742a6371cc84e6827
--- /dev/null
+++ b/robometer/robometer/utils/upload_to_hub.py
@@ -0,0 +1,241 @@
+#!/usr/bin/env python3
+"""
+Simple script to upload a trained model to HuggingFace Hub.
+
+uv run python robometer/utils/upload_to_hub.py \
+    --model_dir ./checkpoints/my_model \
+    --hub_model_id rewardfm/my-rbm-model \
+    --private \
+    --commit_message "Upload trained RBM model checkpoint" \
+    --base_model "Qwen/Qwen3-VL-4B-Instruct"
+"""
+
+import argparse
+import json
+import os
+from pathlib import Path
+
+from huggingface_hub import HfApi, login
+
+
+def validate_model_directory(model_dir: Path) -> bool:
+    """Validate that the directory contains a valid model."""
+    required_files = ["config.json"]
+
+    # Check for required files
+    for file in required_files:
+        if not (model_dir / file).exists():
+            print(f"❌ ERROR: Required file {file} not found in {model_dir}")
+            return False
+
+    # Check for model files (either single or sharded)
+    has_model_files = False
+
+    # Check for safetensors files
+    safetensors_files = list(model_dir.glob("*.safetensors"))
+    if safetensors_files:
+        has_model_files = True
+        print(f"Found {len(safetensors_files)} safetensors files")
+
+    # Check for pytorch files
+    pytorch_files = list(model_dir.glob("pytorch_model*.bin"))
+    if pytorch_files:
+        has_model_files = True
+        print(f"Found {len(pytorch_files)} pytorch files")
+
+    if not has_model_files:
+        print("❌ ERROR: No model files found (*.safetensors or pytorch_model*.bin)")
+        return False
+
+    print("✅ Model directory validation passed")
+    return True
+
+
+def create_model_card(model_dir: Path, base_model: str, model_name: str):
+    """Create or update the model card."""
+    readme_path = model_dir / "README.md"
+
+    # Try to read existing config to get more info
+    config_path = model_dir / "config.json"
+    try:
+        with open(config_path) as f:
+            config = json.load(f)
+        model_type = config.get("model_type", "unknown")
+        config.get("architectures", ["unknown"])
+    except:
+        model_type = "unknown"
+
+    # Try to read wandb info if available
+    wandb_info_path = model_dir.parent / "wandb_info.json"
+    wandb_section = ""
+    if wandb_info_path.exists():
+        try:
+            with open(wandb_info_path) as f:
+                wandb_info = json.load(f)
+            wandb_notes = wandb_info.get("wandb_notes", "")
+            notes_section = f"\n- **Notes**: {wandb_notes}" if wandb_notes else ""
+            wandb_section = f"""
+## Training Run
+
+- **Wandb Run**: [{wandb_info.get("wandb_name", "N/A")}]({wandb_info.get("wandb_url", "#")})
+- **Wandb ID**: `{wandb_info.get("wandb_id", "N/A")}`
+- **Project**: {wandb_info.get("wandb_project", "N/A")}{notes_section}
+"""
+        except Exception as e:
+            print(f"⚠️ WARNING: Could not read wandb info: {e}")
+            wandb_section = ""
+
+    model_card_content = f"""---
+license: apache-2.0
+base_model: {base_model}
+tags:
+- reward_model
+- rbm
+- preference_comparisons
+library_name: transformers
+---
+
+# {model_name}
+
+## Model Details
+
+- **Base Model**: {base_model}
+- **Model Type**: {model_type}
+{wandb_section}
+## Citation
+
+If you use this model, please cite:
+"""
+
+    with open(readme_path, "w") as f:
+        f.write(model_card_content)
+
+    print("Created/updated model card (README.md)")
+
+
+def upload_model_to_hub(
+    model_dir: str,
+    hub_model_id: str,
+    private: bool = False,
+    token: str | None = None,
+    commit_message: str = "Upload model",
+    base_model: str = "Qwen/Qwen2.5-VL-3B-Instruct",
+    tag_name: str | None = None,
+):
+    """
+    Upload model directory to HuggingFace Hub.
+
+    Args:
+        model_dir: Path to the model directory
+        hub_model_id: HuggingFace model ID (username/model-name)
+        private: Whether to make the model private
+        token: HuggingFace token
+        commit_message: Commit message for the upload
+        base_model: Base model name for the model card
+        tag_name: Optional tag name to create after upload
+
+    Returns:
+        tuple: (hub_url, commit_id) - URL of the uploaded model and the commit ID
+    """
+
+    model_path = Path(model_dir)
+
+    # Validate model directory
+    if not model_path.exists():
+        raise ValueError(f"Model directory does not exist: {model_path}")
+
+    if not validate_model_directory(model_path):
+        raise ValueError("Model directory validation failed")
+
+    # Create/update model card
+    create_model_card(model_path, base_model, hub_model_id)
+
+    # Login to HuggingFace
+    if token:
+        login(token=token)
+        print("Logged in to HuggingFace Hub")
+    elif os.getenv("HF_TOKEN"):
+        login(token=os.getenv("HF_TOKEN"))
+        print("Logged in using HF_TOKEN environment variable")
+    else:
+        print("⚠️ WARNING: No HuggingFace token provided. You may need to login manually.")
+
+    # Upload to Hub
+    print(f"Uploading model to: {hub_model_id}")
+    print(f"Private: {private}")
+    print(f"Model directory: {model_path}")
+
+    api = HfApi()
+
+    # Create the repository if it doesn't exist
+    try:
+        api.create_repo(repo_id=hub_model_id, repo_type="model", private=private, exist_ok=True)
+        print(f"Repository {hub_model_id} created/verified")
+    except Exception as e:
+        print(f"⚠️ WARNING: Could not create repository (may already exist): {e}")
+
+    # Upload the entire directory
+    commit_info = api.upload_folder(
+        folder_path=str(model_path), repo_id=hub_model_id, commit_message=commit_message, repo_type="model"
+    )
+
+    commit_id = commit_info.oid
+    print(f"✅ Successfully uploaded model to: https://huggingface.co/{hub_model_id}")
+    print(f"📋 Commit ID: {commit_id}")
+
+    # Also upload the config.yaml which is in the directory above
+    config_path = model_path.parent / "config.yaml"
+    if config_path.exists():
+        print(f"Uploading config.yaml to: {hub_model_id}")
+        api.upload_file(
+            path_or_fileobj=str(config_path),
+            path_in_repo="config.yaml",
+            repo_id=hub_model_id,
+            repo_type="model",
+        )
+
+    # Create tag if requested
+    if tag_name:
+        api.create_tag(
+            repo_id=hub_model_id, repo_type="model", tag=tag_name, revision=commit_id, tag_message=commit_message
+        )
+        print(f"🏷️ Created tag: {tag_name}")
+
+    hub_url = f"https://huggingface.co/{hub_model_id}"
+    return hub_url, commit_id
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Upload model to HuggingFace Hub")
+    parser.add_argument("--model_dir", type=str, required=True, help="Path to the model directory")
+    parser.add_argument("--hub_model_id", type=str, required=True, help="HuggingFace model ID (username/model-name)")
+    parser.add_argument("--private", action="store_true", help="Make the model private")
+    parser.add_argument("--token", type=str, help="HuggingFace token (or set HF_TOKEN env var)")
+    parser.add_argument("--commit_message", type=str, default="Upload RBM model", help="Commit message for the upload")
+    parser.add_argument(
+        "--base_model", type=str, default="Qwen/Qwen2.5-VL-3B-Instruct", help="Base model name for the model card"
+    )
+
+    args = parser.parse_args()
+
+    try:
+        url, commit_id = upload_model_to_hub(
+            model_dir=args.model_dir,
+            hub_model_id=args.hub_model_id,
+            private=args.private,
+            token=args.token,
+            commit_message=args.commit_message,
+            base_model=args.base_model,
+        )
+
+        print("\n🎉 Upload completed successfully!")
+        print(f"Model URL: {url}")
+        print(f"Commit ID: {commit_id}")
+
+    except Exception as e:
+        print(f"❌ ERROR: Upload failed: {e}")
+        raise
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/robometer/utils/video_utils.py b/robometer/robometer/utils/video_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4d0c43e586c0fc9895a9a195d1eb31d5967ea785
--- /dev/null
+++ b/robometer/robometer/utils/video_utils.py
@@ -0,0 +1,603 @@
+import base64
+import io
+import os
+import random
+from typing import Any, Optional, Union
+
+import cv2
+import numpy as np
+from PIL import Image
+
+from robometer.data.datasets.helpers import load_frames_from_npz
+
+
+def extract_frames_from_video(video_path: str, fps: int = 1) -> np.ndarray:
+    """
+    Extract frames from video file at specified FPS.
+
+    Args:
+        video_path: Path to the .mp4 file
+        fps: Frames per second to extract (default: 1)
+
+    Returns:
+        numpy array of shape (num_frames, H, W, C) with extracted frames
+    """
+    if not os.path.exists(video_path):
+        raise FileNotFoundError(f"Video file not found: {video_path}")
+
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise ValueError(f"Could not open video file: {video_path}")
+
+    # Get video properties
+    int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    video_fps = cap.get(cv2.CAP_PROP_FPS)
+
+    # Calculate frame interval for target FPS
+    frame_interval = max(1, int(video_fps / fps))
+
+    frames = []
+    frame_count = 0
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        if frame_count % frame_interval == 0:
+            # Convert BGR to RGB
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame_rgb)
+
+        frame_count += 1
+
+    cap.release()
+
+    if not frames:
+        raise ValueError(f"No frames extracted from video: {video_path}")
+
+    return np.array(frames)
+
+
+def _ensure_numpy_frames(frames: Any, frames_shape: tuple[int, int, int, int] | None = None) -> np.ndarray:
+    """Ensure frames are a numpy array of shape (T, H, W, C).
+
+    Accepts bytes (with shape), numpy array, list of numpy frames, or single frame.
+    """
+    if frames is None:
+        return np.empty((0,))
+
+    # Bytes -> numpy using provided shape
+    if isinstance(frames, (bytes, bytearray)):
+        if frames_shape is None:
+            # Fallback: interpret as uint8 flat array (cannot reshape reliably)
+            arr = np.frombuffer(frames, dtype=np.uint8)
+            return arr
+        if isinstance(frames_shape, list):
+            frames_shape = tuple(frames_shape)
+        try:
+            return np.frombuffer(frames, dtype=np.uint8).reshape(frames_shape)
+        except Exception:
+            return np.frombuffer(frames, dtype=np.uint8)
+
+    # Already a numpy array
+    if isinstance(frames, np.ndarray):
+        return frames
+
+    # List of numpy arrays
+    if isinstance(frames, list) and all(isinstance(f, np.ndarray) for f in frames):
+        return np.stack(frames, axis=0)
+
+    # Unsupported (e.g., file paths) – return as empty; upstream should handle
+    return np.empty((0,))
+
+
+def decode_frames_b64(frames_b64: list[str]) -> list[Image.Image]:
+    images: list[Image.Image] = []
+    for s in frames_b64:
+        buf = io.BytesIO(base64.b64decode(s))
+        img = Image.open(buf).convert("RGB")
+        images.append(img)
+    return images
+
+
+def frames_to_base64_images(frames: Any, frames_shape: tuple[int, int, int, int] | None = None) -> list[str]:
+    """Convert frames to a list of base64-encoded JPEG strings.
+
+    Frames can be ndarray (T,H,W,C), bytes + shape, list of ndarray, or a single frame.
+    """
+    arr = _ensure_numpy_frames(frames, frames_shape)
+    if arr.size == 0:
+        return []
+
+    # Normalize to (T, H, W, C)
+    if arr.ndim == 3:  # single frame (H,W,C)
+        arr = arr[None, ...]
+    elif arr.ndim != 4:
+        # Unknown shape: cannot encode reliably
+        return []
+
+    encoded: list[str] = []
+    for i in range(arr.shape[0]):
+        frame = arr[i]
+        if frame.dtype != np.uint8:
+            frame = frame.astype(np.uint8)
+        img = Image.fromarray(frame)
+        buf = io.BytesIO()
+        img.save(buf, format="PNG")
+        encoded.append(base64.b64encode(buf.getvalue()).decode("utf-8"))
+    return encoded
+
+
+def add_text_overlay(
+    frame: np.ndarray,
+    text: str,
+    position: tuple[int, int] = (10, 10),
+    font_scale: float = 0.5,
+    color: tuple[int, int, int] = (255, 255, 255),
+    thickness: int = 1,
+    bg_color: Optional[tuple[int, int, int]] = None,
+) -> np.ndarray:
+    """
+    Add text overlay to a frame.
+
+    Args:
+        frame: Frame in (H, W, C) format, uint8, RGB
+        text: Text to add
+        position: (x, y) position of text (bottom-left corner of text)
+        font_scale: Font scale
+        color: Text color (RGB format, will be converted to BGR for cv2)
+        thickness: Text thickness
+        bg_color: Optional background color (RGB format, will be converted to BGR for cv2)
+
+    Returns:
+        Frame with text overlay (H, W, C), RGB
+    """
+    frame_with_text = frame.copy()
+
+    # Ensure frame has 3 channels
+    if frame_with_text.ndim != 3 or frame_with_text.shape[2] != 3:
+        raise ValueError(f"Expected frame with shape (H, W, 3), got {frame_with_text.shape}")
+
+    # Convert RGB to BGR for cv2 (cv2 uses BGR)
+    frame_bgr = cv2.cvtColor(frame_with_text, cv2.COLOR_RGB2BGR)
+
+    # Convert colors from RGB to BGR for cv2
+    color_bgr = (color[2], color[1], color[0])
+    bg_color_bgr = (bg_color[2], bg_color[1], bg_color[0]) if bg_color is not None else None
+
+    # Get text size for background box
+    (text_width, text_height), baseline = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, font_scale, thickness)
+
+    # Draw background box if specified
+    if bg_color_bgr is not None:
+        cv2.rectangle(
+            frame_bgr,
+            (position[0] - 5, position[1] - text_height - baseline - 5),
+            (position[0] + text_width + 5, position[1] + 5),
+            bg_color_bgr,
+            -1,
+        )
+
+    # Draw text (position is bottom-left corner)
+    cv2.putText(frame_bgr, text, position, cv2.FONT_HERSHEY_SIMPLEX, font_scale, color_bgr, thickness, cv2.LINE_AA)
+
+    # Convert back to RGB
+    frame_with_text = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+
+    return frame_with_text
+
+
+def create_video_grid_with_progress(
+    video_frames_list: list[Optional[np.ndarray]],
+    trajectory_progress_data: list[Union[list[float], list[int], np.ndarray, None]],
+    grid_size: tuple[int, int] = (3, 3),
+    max_videos: int = 9,
+    is_discrete_mode: bool = False,
+) -> Optional[np.ndarray]:
+    """
+    Create a grid of videos with progress information overlaid on each video.
+
+    Args:
+        video_frames_list: List of videos, each in (T, C, H, W) format or None
+        trajectory_progress_data: List of progress_pred values, one per video.
+                                  Each element is a list/array of progress predictions.
+        grid_size: Tuple of (rows, cols) for the grid
+        max_videos: Maximum number of videos to sample
+        is_discrete_mode: Whether predictions are in discrete mode
+
+    Returns:
+        Grid video in (T, C, H, W) format, or None if insufficient valid videos
+    """
+    # Filter out None videos
+    valid_videos = [(idx, v) for idx, v in enumerate(video_frames_list) if v is not None]
+
+    grid_cells = grid_size[0] * grid_size[1]
+    if len(valid_videos) == 0:
+        return None
+
+    # Sample available videos (up to grid_cells)
+    num_to_sample = min(grid_cells, len(valid_videos))
+    sampled_videos = random.sample(valid_videos, num_to_sample)
+
+    # Get corresponding progress data (assume alignment)
+    valid_items = [(v_idx, v, trajectory_progress_data[v_idx]) for v_idx, v in sampled_videos]
+
+    # Find maximum time dimension across valid videos
+    max_time = max(v.shape[0] for _, v, _ in valid_items) if valid_items else 1
+
+    # Resize and normalize videos to same size for grid
+    processed_videos = []
+    target_h, target_w = 128, 128  # Target size for each cell in grid
+
+    for _, video, progress_pred in valid_items:
+        # Pad video to max_time by repeating the last frame
+        if video.shape[0] < max_time:
+            padding = np.repeat(video[-1:], max_time - video.shape[0], axis=0)
+            video = np.concatenate([video, padding], axis=0)
+
+        # Pad progress_pred to match video length if needed
+        if progress_pred is not None and isinstance(progress_pred, (list, np.ndarray)):
+            pred_arr = np.array(progress_pred) if isinstance(progress_pred, list) else progress_pred
+            if len(pred_arr) < max_time:
+                padding = np.repeat([pred_arr[-1]], max_time - len(pred_arr), axis=0)
+                progress_pred = np.concatenate([pred_arr, padding], axis=0)
+            else:
+                progress_pred = pred_arr
+
+        # Convert (T, C, H, W) to (T, H, W, C) for processing
+        video = video.transpose(0, 2, 3, 1)
+
+        # Resize each frame and add progress overlay
+        resized_frames = []
+        for t, frame in enumerate(video):
+            # Ensure uint8 format
+            if frame.dtype != np.uint8:
+                frame = np.clip(frame * 255, 0, 255).astype(np.uint8)
+            frame_resized = cv2.resize(frame, (target_w, target_h))
+
+            # Add progress text overlay in bottom right corner
+            if progress_pred is not None:
+                # Get progress value for this frame (index by t)
+                if isinstance(progress_pred, (list, np.ndarray)):
+                    pred_elem = progress_pred[t]
+                    if is_discrete_mode:
+                        # Discrete mode: apply argmax to get predicted bin index
+                        pred_val = float(np.argmax(pred_elem))
+                    else:
+                        # Continuous mode: convert to float directly
+                        pred_val = float(pred_elem)
+                else:
+                    pred_val = float(progress_pred)
+
+                # Format text (only show predicted value)
+                progress_text = f"P:{pred_val:.2f}"
+
+                # Calculate position (bottom right, with padding)
+                text_x = target_w - 60
+                text_y = target_h - 10
+
+                # Add text with background
+                frame_resized = add_text_overlay(
+                    frame_resized,
+                    progress_text,
+                    position=(text_x, text_y),
+                    font_scale=0.4,
+                    color=(255, 255, 255),  # White text
+                    thickness=1,
+                    bg_color=(0, 0, 0),  # Black background
+                )
+
+            resized_frames.append(frame_resized)
+
+        # Convert back to (T, C, H, W)
+        video_resized = np.array(resized_frames).transpose(0, 3, 1, 2)
+        processed_videos.append(video_resized)
+
+    # Fill remaining grid cells with black videos
+    num_black_videos = grid_cells - len(processed_videos)
+    for _ in range(num_black_videos):
+        # Create black video: (T, C, H, W) with all zeros (black)
+        black_video = np.zeros((max_time, 3, target_h, target_w), dtype=np.uint8)
+        processed_videos.append(black_video)
+
+    # Arrange videos in grid
+    grid_frames = []
+    for t in range(max_time):
+        grid_rows = []
+        for row in range(grid_size[0]):
+            row_frames = []
+            for col in range(grid_size[1]):
+                vid_idx = row * grid_size[1] + col
+                frame = processed_videos[vid_idx][t]  # (C, H, W)
+                # Convert to (H, W, C) for concatenation
+                frame = frame.transpose(1, 2, 0)
+                row_frames.append(frame)
+            # Concatenate horizontally
+            row_concat = np.concatenate(row_frames, axis=1)  # (H, total_W, C)
+            grid_rows.append(row_concat)
+        # Concatenate vertically
+        grid_frame = np.concatenate(grid_rows, axis=0)  # (total_H, total_W, C)
+        # Convert back to (C, H, W)
+        grid_frame = grid_frame.transpose(2, 0, 1)
+        grid_frames.append(grid_frame)
+
+    # Stack frames: (T, C, H, W)
+    grid_video = np.stack(grid_frames, axis=0)
+    return grid_video
+
+
+def create_frame_pair_with_progress(
+    eval_result: dict, target_h: int = 224, target_w: int = 224, is_discrete_mode: bool = False
+) -> Optional[np.ndarray]:
+    """
+    Create a horizontal row of frames from a trajectory with progress annotations.
+    The number of frames is determined by the length of progress_pred or target_progress.
+    Each row represents one trajectory, with frames displayed horizontally.
+
+    Args:
+        eval_result: Evaluation result dict with video_path, progress_pred, target_progress
+        target_h: Target height for frames
+        target_w: Target width for frames
+
+    Returns:
+        Single row image with frames horizontally concatenated in (C, H, W) format, or None if unavailable
+    """
+    video_path = eval_result.get("video_path")
+    if video_path is None:
+        return None
+
+    # Get progress values to determine number of frames
+    progress_pred = eval_result.get("progress_pred")
+    target_progress = eval_result.get("target_progress")
+
+    # Determine number of frames from progress_pred or target_progress length
+    if isinstance(progress_pred, (list, np.ndarray)) and len(progress_pred) > 0:
+        num_frames = len(progress_pred)
+    elif isinstance(target_progress, (list, np.ndarray)) and len(target_progress) > 0:
+        num_frames = len(target_progress)
+    else:
+        # Fallback: if both are scalars or None, default to 1 frame
+        num_frames = 1
+
+    # Load frames from video
+    frames = load_frames_from_npz(video_path)
+    frames = frames.transpose(0, 3, 1, 2)  # (T, C, H, W)
+
+    if frames.shape[0] < 1:
+        return None
+
+    # Process frames - use evenly spaced indices to sample frames from video
+    # This handles cases where video has more frames than progress_pred indicates
+    processed_frames = []
+    for i in range(num_frames):
+        # Calculate frame index - evenly space across available frames
+        if num_frames == 1:
+            frame_idx = 0
+        else:
+            frame_idx = int((i / (num_frames - 1)) * (frames.shape[0] - 1)) if frames.shape[0] > 1 else 0
+        frame_idx = min(frame_idx, frames.shape[0] - 1)
+
+        frame = frames[frame_idx]  # (C, H, W)
+        frame_hwc = frame.transpose(1, 2, 0)  # (H, W, C)
+        if frame_hwc.dtype != np.uint8:
+            frame_hwc = np.clip(frame_hwc * 255, 0, 255).astype(np.uint8)
+        frame_resized = cv2.resize(frame_hwc, (target_w, target_h))
+
+        # Get progress values for this frame index
+        if isinstance(progress_pred, (list, np.ndarray)):
+            pred_elem = progress_pred[i]
+            if is_discrete_mode:
+                # Discrete mode: apply argmax to get predicted bin index
+                pred_val = float(np.argmax(pred_elem))
+            else:
+                # Continuous mode: convert to float directly
+                pred_val = float(pred_elem)
+        else:
+            pred_val = float(progress_pred) if progress_pred is not None else 0.0
+
+        if isinstance(target_progress, (list, np.ndarray)):
+            # Target is already a discrete bin index (integer) in discrete mode, or continuous in continuous mode
+            target_val = float(target_progress[i])
+        else:
+            target_val = float(target_progress) if target_progress is not None else 0.0
+
+        # Add progress annotation to frame
+        progress_text = f"P:{pred_val:.2f} T:{target_val:.2f}"
+        text_x = target_w - 110
+        text_y = target_h - 10
+        frame_resized = add_text_overlay(
+            frame_resized,
+            progress_text,
+            position=(text_x, text_y),
+            font_scale=0.4,
+            color=(255, 255, 255),
+            thickness=1,
+            bg_color=(0, 0, 0),
+        )
+
+        processed_frames.append(frame_resized)
+
+    # Concatenate all frames horizontally
+    if len(processed_frames) == 0:
+        return None
+    combined_frame = np.concatenate(processed_frames, axis=1)  # (H, num_frames*W, C)
+
+    # Add text label above the frame row (task and quality_label/partial_success)
+    task = eval_result.get("task")
+    quality_label = eval_result.get("quality_label")
+    partial_success = eval_result.get("partial_success")
+
+    # Build label text parts (before wrapping)
+    label_parts = []
+    if task is not None:
+        label_parts.append(f"Task: {task}")
+    if partial_success is not None:
+        # RoboArena: use partial_success
+        label_parts.append(f"Partial: {partial_success:.2f}")
+    elif quality_label is not None:
+        # Non-RoboArena: use quality_label
+        label_parts.append(f"Quality: {quality_label}")
+
+    if label_parts:
+        # Get available width for text (frame width - padding on both sides)
+        available_width = combined_frame.shape[1] - 20  # 10px padding on each side
+        font_scale = 0.5
+        thickness = 1
+        font = cv2.FONT_HERSHEY_SIMPLEX
+
+        # Helper function to wrap text to fit available width
+        def wrap_text(text, max_width):
+            """Wrap text into multiple lines that fit within max_width."""
+            words = text.split()
+            lines = []
+            current_line = words[0] if words else ""
+
+            for word in words[1:]:
+                # Check if adding the next word exceeds width
+                test_line = current_line + " " + word
+                (text_width, _), _ = cv2.getTextSize(test_line, font, font_scale, thickness)
+
+                if text_width <= max_width:
+                    current_line = test_line
+                else:
+                    # Current line is full, start a new line
+                    lines.append(current_line)
+                    current_line = word
+
+            if current_line:
+                lines.append(current_line)
+
+            return lines
+
+        # Wrap each label part and collect all lines
+        label_lines = []
+        for part in label_parts:
+            wrapped = wrap_text(part, available_width)
+            label_lines.extend(wrapped)
+
+        # Calculate label height based on number of lines
+        line_height = 20  # Height per line
+        line_spacing = 5  # Spacing between lines
+        label_height = len(label_lines) * line_height + (len(label_lines) - 1) * line_spacing + 10  # Add padding
+
+        # Ensure minimum label height
+        min_label_height = 30
+        if label_height < min_label_height:
+            label_height = min_label_height
+
+        label_frame = np.ones((label_height, combined_frame.shape[1], 3), dtype=np.uint8) * 255  # White background
+
+        # Add each line of text
+        for i, label_text in enumerate(label_lines):
+            y_position = label_height - 5 - (len(label_lines) - 1 - i) * (line_height + line_spacing)
+            label_frame = add_text_overlay(
+                label_frame,
+                label_text,
+                position=(10, y_position),  # Left-aligned
+                font_scale=font_scale,
+                color=(0, 0, 0),  # Black text
+                thickness=thickness,
+                bg_color=None,  # No background needed (already white)
+            )
+
+        # Concatenate label above the frame row
+        combined_frame = np.concatenate([label_frame, combined_frame], axis=0)  # (H + label_height, num_frames*W, C)
+
+    # Convert back to (C, H, W) format (no time dimension)
+    combined_frame_chw = combined_frame.transpose(2, 0, 1)  # (C, H + label_height, num_frames*W)
+
+    return combined_frame_chw
+
+
+def create_policy_ranking_grid(
+    eval_results: list[dict],
+    grid_size: tuple[int, int] = (2, 2),
+    max_samples: int = 4,
+    border_width: int = 4,
+    is_discrete_mode: bool = False,
+) -> Optional[np.ndarray]:
+    """
+    Create a vertical stack of trajectory rows, each showing all frames horizontally.
+    Each row represents one trajectory with all its frames displayed horizontally.
+
+    Args:
+        eval_results: List of evaluation results with video_path, progress_pred, target_progress
+        grid_size: Not used anymore (kept for compatibility), but ignored
+        max_samples: Maximum number of samples (trajectories) to display (default: 4)
+        border_width: Width of border between rows in pixels
+
+    Returns:
+        Stacked image in (H, W, C) format, uint8, RGB, or None if unavailable
+    """
+    # Filter results with valid video_paths
+    valid_results = [r for r in eval_results if r.get("video_path") is not None]
+
+    if len(valid_results) == 0:
+        return None
+
+    # Limit to max_samples
+    num_to_sample = min(max_samples, len(valid_results))
+
+    # Sample random results
+    sampled_results = random.sample(valid_results, num_to_sample)
+
+    # Create frame rows for each sampled result (each row is one trajectory)
+    frame_rows = []
+    target_h, target_w = 224, 224
+
+    for result in sampled_results:
+        frame_row = create_frame_pair_with_progress(result, target_h, target_w, is_discrete_mode=is_discrete_mode)
+        if frame_row is not None:
+            frame_rows.append(frame_row)
+
+    if len(frame_rows) == 0:
+        return None
+
+    # Get the maximum width across all rows to align them
+    max_width = max(row.shape[2] for row in frame_rows)  # (C, H, W) -> W dimension
+    max_height = max(row.shape[1] for row in frame_rows)  # (C, H, W) -> H dimension
+
+    # Pad all rows to the same width and height
+    aligned_rows = []
+    border_color = np.array([128, 128, 128], dtype=np.uint8)  # Gray border
+
+    for row in frame_rows:
+        # Convert to (H, W, C) for processing
+        row_hwc = row.transpose(1, 2, 0)  # (H, W, C)
+        current_height, current_width = row_hwc.shape[0], row_hwc.shape[1]
+
+        # Pad width if necessary (pad on the right)
+        if current_width < max_width:
+            width_padding = max_width - current_width
+            right_padding = np.ones((current_height, width_padding, 3), dtype=np.uint8) * 255  # White padding
+            row_hwc = np.concatenate([row_hwc, right_padding], axis=1)
+
+        # Pad height if necessary (pad on the bottom)
+        if current_height < max_height:
+            height_padding = max_height - current_height
+            bottom_padding = np.ones((height_padding, row_hwc.shape[1], 3), dtype=np.uint8) * 255  # White padding
+            row_hwc = np.concatenate([row_hwc, bottom_padding], axis=0)
+
+        aligned_rows.append(row_hwc)
+
+    # Stack rows vertically with borders between them
+    stacked_rows = []
+    for i, row in enumerate(aligned_rows):
+        stacked_rows.append(row)
+        # Add horizontal border below (except for last row)
+        if i < len(aligned_rows) - 1:
+            h_border = np.tile(border_color, (border_width, row.shape[1], 1))
+            stacked_rows.append(h_border)
+
+    # Concatenate vertically
+    if len(stacked_rows) == 0:
+        return None
+
+    grid_frame = np.concatenate(stacked_rows, axis=0)  # (total_H, max_W, C)
+
+    # Return as static image in (H, W, C) format, uint8
+    if grid_frame.dtype != np.uint8:
+        grid_frame = np.clip(grid_frame, 0, 255).astype(np.uint8)
+
+    return grid_frame
diff --git a/robometer/scripts/bench_video_robometer.py b/robometer/scripts/bench_video_robometer.py
new file mode 100644
index 0000000000000000000000000000000000000000..8fe35d4ced31985836e49942ec8cd7a94368d9da
--- /dev/null
+++ b/robometer/scripts/bench_video_robometer.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+"""Run Robometer progress scoring on the BenchVideo clips (no marks).
+
+For each episode in a category's tasks.json:
+  - load mp4 -> sample to <= max_frames at fps
+  - frame_steps inference -> per-frame progress curve
+  - save results_robometer.json + overview_robometer.png
+
+Output: BenchVideo/eval_results/<Category>/<id>/
+"""
+import argparse
+import json
+import os
+from pathlib import Path
+
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+from benchmark_progress_mark_local import (
+    RobometerLocalRunner,
+    load_video_frames_with_indices,
+)
+
+BENCH_ROOT = Path("/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/Benchmark_Videos/BenchVideo")
+EVAL_ROOT = BENCH_ROOT / "eval_results"
+MODEL_PATH = "/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/Robometer/models/Robometer-4B"
+
+
+def write_overview(out_path, frames, progress, title):
+    fig, ax = plt.subplots(figsize=(11, 4.6))
+    ax.plot(frames[: len(progress)], progress, marker="o", linewidth=1.5, markersize=2.8, color="#b9482f")
+    ax.set_ylim(-0.05, 1.05)
+    ax.set_xlim(0, max(5, int(max(frames) * 1.02)) if frames else 5)
+    ax.set_xlabel("Original Frame"); ax.set_ylabel("Progress (normalized)")
+    ax.set_title(title); ax.grid(True, linestyle="--", alpha=0.35)
+    fig.tight_layout(); fig.savefig(out_path, dpi=180); plt.close(fig)
+
+
+def run_category(runner, category, fps, max_frames, fallback):
+    cat_dir = BENCH_ROOT / category
+    data = json.loads((cat_dir / "tasks.json").read_text(encoding="utf-8"))
+    out_root = EVAL_ROOT / category
+    out_root.mkdir(parents=True, exist_ok=True)
+
+    for ep in data["episodes"]:
+        name = ep["episode"]; stem = name[:-4]
+        task = (ep.get("tasks") or [fallback])[0] or fallback
+        src = cat_dir / name
+        ep_out = out_root / stem
+        if (ep_out / "results_robometer.json").exists():
+            print(f"[skip] {category}/{name} (results_robometer.json exists)")
+            continue
+        ep_out.mkdir(parents=True, exist_ok=True)
+
+        print(f"\n[RUN] {category}/{name}  task={task!r}")
+        frames, sampled_idx, total, native_fps = load_video_frames_with_indices(
+            src, fps=fps, max_frames=max_frames, required_frames=[])
+        progress, success = runner.compute_rewards_per_frame(
+            frames, task, inference_mode="frame_steps",
+            prefix_sample_frames=8, prefix_batch_size=1)
+        progress = [float(x) for x in progress]
+
+        results = {
+            "category": category, "episode": name,
+            "original_file": ep.get("original_file"), "task": task,
+            "num_total_frames": total, "native_fps": native_fps, "sample_fps": fps,
+            "num_sampled_frames": len(sampled_idx),
+            "sampled_original_frame_indices": sampled_idx,
+            "progress_curve": progress,
+        }
+        (ep_out / "results_robometer.json").write_text(
+            json.dumps(results, ensure_ascii=False, indent=2), encoding="utf-8")
+        write_overview(ep_out / "overview_robometer.png", sampled_idx, progress,
+                       f"Robometer {category}/{stem}: {task[:50]}")
+        rng = f"{min(progress):.2f}~{max(progress):.2f}" if progress else "empty"
+        print(f"[OK] {ep_out}  (progress {rng})")
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--category", default="all")
+    ap.add_argument("--fps", type=float, default=2.0)
+    ap.add_argument("--max-frames", type=int, default=128)
+    ap.add_argument("--instruction-fallback", default="Complete the task shown in the video.")
+    args = ap.parse_args()
+
+    runner = RobometerLocalRunner(model_path=MODEL_PATH)
+    cats = ["Count", "State", "Sequence"] if args.category == "all" else [args.category]
+    for c in cats:
+        run_category(runner, c, args.fps, args.max_frames, args.instruction_fallback)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/benchmark_progress_mark_local.py b/robometer/scripts/benchmark_progress_mark_local.py
new file mode 100644
index 0000000000000000000000000000000000000000..f7d532202ca0b9102e36ab7ca6749223c1feef01
--- /dev/null
+++ b/robometer/scripts/benchmark_progress_mark_local.py
@@ -0,0 +1,1018 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import gc
+import json
+import os
+import re
+from collections import defaultdict
+from pathlib import Path
+from typing import Optional
+
+import av
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+import torch
+
+from robometer.data.dataset_types import ProgressSample, Trajectory
+from robometer.evals.eval_server import compute_batch_outputs
+from robometer.utils.save import load_model_from_hf
+from robometer.utils.setup_utils import setup_batch_collator
+
+matplotlib.use("Agg")
+av.logging.set_level(av.logging.ERROR)
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PROJECT_ROOT = Path(__file__).resolve().parents[2]
+WORKSPACE_ROOT = Path(__file__).resolve().parents[3]
+DEFAULT_MODEL_PATH = PROJECT_ROOT / "models" / "Robometer-4B"
+DEFAULT_VIDEOS_ROOT = WORKSPACE_ROOT / "Videos"
+DEFAULT_OUTPUT_ROOT = REPO_ROOT / "Benchmark_Eval"
+DEFAULT_CHUNK = "chunk-000"
+DEFAULT_CAMERA_DIR = "observation.images.wrist_image_left"
+DEFAULT_MAX_FRAMES = 128
+DEFAULT_MIN_FRAMES = 32
+FRAME_BACKOFF_CANDIDATES = [512, 384, 320, 256, 224, 192, 160, 144, 128, 112, 96, 80, 64, 48, 32, 24, 16, 8]
+DEFAULT_INFERENCE_MODE = "frame_steps"
+DEFAULT_PREFIX_SAMPLE_FRAMES = 4
+DEFAULT_PREFIX_BATCH_SIZE = 1
+DEFAULT_DESCRIPTION_EPISODES_PATH = WORKSPACE_ROOT / "Robo-Dopamine" / "Evaluation" / "Benchmark_Eval" / "Description" / "episodes.jsonl"
+
+
+def normalize_episode_name(name: str) -> str:
+    name = name.strip()
+    if name.endswith(".mp4"):
+        return name
+    if name.startswith("episode_"):
+        return f"{name}.mp4"
+    if name.isdigit():
+        return f"episode_{int(name):06d}.mp4"
+    return f"episode_{name}.mp4"
+
+
+def chunk_to_dir_name(chunk: str) -> str:
+    m = re.fullmatch(r"chunk-(\d+)(?:_(.+))?", chunk)
+    if not m:
+        return f"Chunk{chunk.split('-')[-1]}"
+    suffix = m.group(2)
+    return f"Chunk{m.group(1)}_{suffix}" if suffix else f"Chunk{m.group(1)}"
+
+
+def parse_episode_input(token: str) -> str:
+    return normalize_episode_name(token)
+
+
+def load_json(path: Path) -> dict:
+    with path.open("r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def load_episode_catalog(episode_tasks_path: Path, annotations_path: Path) -> list[str]:
+    if episode_tasks_path.exists():
+        data = load_json(episode_tasks_path)
+        catalog = []
+        for ep in data.get("episodes", []):
+            episode = ep.get("episode")
+            if isinstance(episode, str) and episode.strip():
+                catalog.append(episode.strip())
+        if catalog:
+            return catalog
+    data = load_json(annotations_path)
+    return [ep["episode"] for ep in data.get("episodes", []) if ep.get("episode")]
+
+
+def load_episode_record(annotations_path: Path, episode_mp4: str) -> dict:
+    data = load_json(annotations_path)
+    for ep in data.get("episodes", []):
+        if ep.get("episode") == episode_mp4:
+            return ep
+    raise ValueError(f"Episode not found in annotations: {episode_mp4}")
+
+
+def load_instruction(episode_tasks_path: Path, episode_mp4: str) -> str:
+    if episode_tasks_path.exists():
+        data = load_json(episode_tasks_path)
+        for ep in data.get("episodes", []):
+            if ep.get("episode") != episode_mp4:
+                continue
+            for task in ep.get("tasks") or []:
+                if isinstance(task, str) and task.strip():
+                    return task.strip()
+
+    m = re.fullmatch(r"episode_(\d+)\.mp4", episode_mp4)
+    if m and DEFAULT_DESCRIPTION_EPISODES_PATH.exists():
+        target_idx = int(m.group(1))
+        with DEFAULT_DESCRIPTION_EPISODES_PATH.open("r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                obj = json.loads(line)
+                if obj.get("episode_index") != target_idx:
+                    continue
+                for task in obj.get("tasks") or []:
+                    if isinstance(task, str) and task.strip():
+                        return task.strip()
+                break
+    return "robot manipulation task"
+
+
+def extract_progress_pairs(episode_record: dict) -> list[tuple[str, int, int]]:
+    marks = [a for a in episode_record.get("annotations", []) if a.get("type") == "progress_mark"]
+    camera_ref = episode_record.get("camera_reference")
+    if camera_ref:
+        ref_marks = [m for m in marks if m.get("camera") == camera_ref]
+        if ref_marks:
+            marks = ref_marks
+
+    grouped: dict[str, list[int]] = defaultdict(list)
+    for mark in marks:
+        frame = mark.get("frame")
+        if isinstance(frame, int):
+            grouped[str(mark.get("group_label", "UNK"))].append(frame)
+
+    pairs: list[tuple[str, int, int]] = []
+    for label in sorted(grouped):
+        frames = sorted(set(grouped[label]))
+        if len(frames) >= 2:
+            pairs.append((label, frames[0], frames[1]))
+    return pairs
+
+
+def load_progress_mark_catalog(annotations_path: Path) -> list[str]:
+    data = load_json(annotations_path)
+    catalog = []
+    for ep in data.get("episodes", []):
+        episode = ep.get("episode")
+        if isinstance(episode, str) and episode.strip() and extract_progress_pairs(ep):
+            catalog.append(episode.strip())
+    return catalog
+
+
+def parse_episode_selection(token: str, catalog: list[str], progress_catalog: list[str], require_progress_marks: bool = True) -> list[str]:
+    aliases = {
+        "fisrt10": "first10",
+        "frist10": "first10",
+        "first 10": "first10",
+    }
+    token = token.strip()
+    token_lower = aliases.get(token.lower(), token.lower())
+
+    if token_lower == "first10":
+        return (progress_catalog if require_progress_marks else catalog)[:10]
+    if token_lower == "all":
+        return progress_catalog if require_progress_marks else catalog
+
+    plus_match = re.fullmatch(r"(.+)\+(\d+)", token)
+    if plus_match:
+        start_episode = parse_episode_input(plus_match.group(1).strip())
+        count = int(plus_match.group(2))
+        if count <= 0:
+            raise ValueError("Count after '+' must be >= 1.")
+        if start_episode not in catalog:
+            raise ValueError(f"Invalid episode selection: {plus_match.group(1).strip()}.")
+        if require_progress_marks and start_episode not in progress_catalog:
+            raise ValueError(f"{start_episode[:-4]} does not have usable progress_mark pairs.")
+        base_catalog = progress_catalog if require_progress_marks else catalog
+        start_idx = base_catalog.index(start_episode)
+        return base_catalog[start_idx:start_idx + count]
+
+    selected = []
+    for part in token.split(","):
+        part = part.strip()
+        if not part:
+            continue
+        episode_mp4 = parse_episode_input(part)
+        if episode_mp4 not in catalog:
+            raise ValueError(f"Invalid episode selection: {part}.")
+        if require_progress_marks and episode_mp4 not in progress_catalog:
+            raise ValueError(f"{episode_mp4[:-4]} does not have usable progress_mark pairs.")
+        selected.append(episode_mp4)
+
+    return selected
+
+
+def select_progress_window(
+    start_episode_raw: str,
+    count: int,
+    catalog: list[str],
+    progress_catalog: list[str],
+    require_progress_marks: bool = True,
+) -> list[str]:
+    start_episode = parse_episode_input(start_episode_raw)
+    if count <= 0:
+        raise ValueError("Count must be >= 1.")
+    if start_episode not in catalog:
+        raise ValueError(f"Invalid episode selection: {start_episode_raw}.")
+    if require_progress_marks and start_episode not in progress_catalog:
+        raise ValueError(f"{start_episode[:-4]} does not have usable progress_mark pairs.")
+    base_catalog = progress_catalog if require_progress_marks else catalog
+    start_idx = base_catalog.index(start_episode)
+    return base_catalog[start_idx:start_idx + count]
+
+
+def result_path_for(output_root: Path, episode_mp4: str) -> Path:
+    return output_root / episode_mp4[:-4] / "results.json"
+
+
+def chunk_output_dir_name(chunk: str, tag: str | None) -> str:
+    base = chunk_to_dir_name(chunk)
+    return f"{base}_{tag}" if tag else base
+
+
+def parse_chunk_selection(chunk_arg: str) -> list[str]:
+    chunks = [part.strip() for part in chunk_arg.split(",") if part.strip()]
+    if not chunks:
+        raise ValueError("At least one chunk must be provided.")
+    return chunks
+
+
+def cleanup_cuda_memory() -> None:
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+
+def is_cuda_oom_error(exc: BaseException) -> bool:
+    if isinstance(exc, torch.OutOfMemoryError):
+        return True
+    return "out of memory" in str(exc).lower()
+
+
+def build_frame_retry_schedule(max_frames: int, min_frames: int, adaptive: bool) -> list[int]:
+    max_frames = max(1, int(max_frames))
+    min_frames = max(1, min(int(min_frames), max_frames))
+    if not adaptive or max_frames == min_frames:
+        return [max_frames]
+
+    schedule = [max_frames]
+    for candidate in FRAME_BACKOFF_CANDIDATES:
+        if min_frames <= candidate < max_frames:
+            schedule.append(candidate)
+    if schedule[-1] != min_frames:
+        schedule.append(min_frames)
+    return schedule
+
+
+def resolve_episode_list(
+    episode_arg: Optional[str],
+    count: int,
+    catalog: list[str],
+    progress_catalog: list[str],
+    require_progress_marks: bool,
+) -> list[str]:
+    if episode_arg:
+        token = episode_arg.strip()
+        token_lower = token.lower()
+        if count > 1:
+            return select_progress_window(token, count, catalog, progress_catalog, require_progress_marks=require_progress_marks)
+        if token_lower in {"first10", "all"} or "," in token or "+" in token:
+            return parse_episode_selection(token, catalog, progress_catalog, require_progress_marks=require_progress_marks)
+        episode_mp4 = parse_episode_input(token)
+        if episode_mp4 not in catalog:
+            raise ValueError(f"Invalid episode selection: {token}.")
+        if require_progress_marks and episode_mp4 not in progress_catalog:
+            raise ValueError(f"{episode_mp4[:-4]} does not have usable progress_mark pairs.")
+        return [episode_mp4]
+
+    base_catalog = progress_catalog if require_progress_marks else catalog
+    print(f"[EPISODES] {len(base_catalog)} episodes available")
+    user_selection = input("Episodes (000287 / 000287+5 / 000287,000330 / first10 / all): ").strip()
+    aliases = {"fisrt10": "first10", "frist10": "first10", "first 10": "first10"}
+    token_lower = aliases.get(user_selection.lower(), user_selection.lower())
+    if token_lower not in {"first10", "all"} and "," not in user_selection and "+" not in user_selection:
+        count_raw = input("How many from this start? (default 5): ").strip()
+        window_count = int(count_raw) if count_raw else 5
+        return select_progress_window(user_selection, window_count, catalog, progress_catalog, require_progress_marks=require_progress_marks)
+    return parse_episode_selection(user_selection, catalog, progress_catalog, require_progress_marks=require_progress_marks)
+
+
+def load_video_frames_with_indices(
+    video_path: Path,
+    *,
+    fps: float,
+    max_frames: int,
+    required_frames: list[int],
+) -> tuple[np.ndarray, list[int], int, float]:
+    all_frames, native_fps = load_all_video_frames(video_path)
+    sampled_frames, sampled_indices = sample_video_frames_with_indices(
+        all_frames,
+        native_fps=native_fps,
+        fps=fps,
+        max_frames=max_frames,
+        required_frames=required_frames,
+    )
+    return sampled_frames, sampled_indices, len(all_frames), native_fps
+
+
+def load_all_video_frames(video_path: Path) -> tuple[list[np.ndarray], float]:
+    all_frames: list[np.ndarray] = []
+    with av.open(str(video_path)) as container:
+        stream = container.streams.video[0]
+        rate = stream.average_rate or stream.guessed_rate
+        native_fps = float(rate) if rate is not None else 1.0
+        for frame in container.decode(stream):
+            all_frames.append(frame.to_ndarray(format="rgb24"))
+
+    if not all_frames:
+        raise RuntimeError(f"Could not extract frames from video: {video_path}")
+    return all_frames, native_fps
+
+
+def sample_video_frames_with_indices(
+    all_frames: list[np.ndarray],
+    *,
+    native_fps: float,
+    fps: float,
+    max_frames: int,
+    required_frames: list[int],
+) -> tuple[np.ndarray, list[int]]:
+    total_frames = len(all_frames)
+    if fps <= 0:
+        fps = native_fps
+
+    if native_fps > 0:
+        desired_frames = int(round(total_frames * (fps / native_fps)))
+    else:
+        desired_frames = total_frames
+    desired_frames = max(1, min(desired_frames, total_frames, max_frames))
+
+    if desired_frames == total_frames:
+        sampled_indices = list(range(total_frames))
+    else:
+        base_indices = np.linspace(0, total_frames - 1, desired_frames, dtype=int).tolist()
+        snapped_indices = list(base_indices)
+        used_slots: set[int] = set()
+
+        # Snap the nearest sampling slots onto the required progress_mark frames.
+        for frame_idx in sorted(set(required_frames)):
+            clamped = max(0, min(total_frames - 1, int(frame_idx)))
+            slot = min(
+                (idx for idx in range(desired_frames) if idx not in used_slots),
+                key=lambda idx: (abs(base_indices[idx] - clamped), idx),
+            )
+            snapped_indices[slot] = clamped
+            used_slots.add(slot)
+
+        sampled_indices = sorted(set(snapped_indices))
+        if len(sampled_indices) < desired_frames:
+            for idx in base_indices:
+                if idx not in sampled_indices:
+                    sampled_indices.append(idx)
+                    if len(sampled_indices) == desired_frames:
+                        break
+        if len(sampled_indices) < desired_frames:
+            for idx in range(total_frames):
+                if idx not in sampled_indices:
+                    sampled_indices.append(idx)
+                    if len(sampled_indices) == desired_frames:
+                        break
+        sampled_indices = sorted(sampled_indices[:desired_frames])
+
+    sampled_frames = np.stack([all_frames[idx] for idx in sampled_indices], axis=0)
+    return sampled_frames, sampled_indices
+
+
+def extract_frame(video_path: Path, frame_idx: int) -> Optional[np.ndarray]:
+    with av.open(str(video_path)) as container:
+        stream = container.streams.video[0]
+        for idx, frame in enumerate(container.decode(stream)):
+            if idx == frame_idx:
+                return frame.to_ndarray(format="rgb24")
+    return None
+
+
+class RobometerLocalRunner:
+    def __init__(self, model_path: str, device: Optional[torch.device] = None):
+        if device is None:
+            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+        self.device = device
+        self.model_path = model_path
+        self.exp_config = None
+        self.tokenizer = None
+        self.processor = None
+        self.reward_model = None
+        self.batch_collator = None
+        self.is_discrete = False
+        self.num_bins = 10
+        self._load_model_components()
+
+    def _load_model_components(self) -> None:
+        print(f"Loading model from {self.model_path} ...")
+
+        self.exp_config, self.tokenizer, self.processor, self.reward_model = load_model_from_hf(
+            model_path=self.model_path,
+            device=self.device,
+        )
+        self.reward_model.eval()
+        self.batch_collator = setup_batch_collator(
+            self.processor,
+            self.tokenizer,
+            self.exp_config,
+            is_eval=True,
+        )
+
+        loss_config = getattr(self.exp_config, "loss", None)
+        self.is_discrete = (
+            getattr(loss_config, "progress_loss_type", "l2").lower() == "discrete"
+            if loss_config
+            else False
+        )
+        self.num_bins = (
+            getattr(loss_config, "progress_discrete_bins", None)
+            or getattr(self.exp_config.model, "progress_discrete_bins", 10)
+        )
+
+    def reload_model(self) -> None:
+        if self.reward_model is not None:
+            del self.reward_model
+        cleanup_cuda_memory()
+        self._load_model_components()
+
+    @torch.inference_mode()
+    def _run_progress_samples(self, progress_samples: list[ProgressSample]) -> tuple[list[list[float]], list[list[float]]]:
+        batch = self.batch_collator(progress_samples)
+
+        progress_inputs = batch["progress_inputs"]
+        for key, value in progress_inputs.items():
+            if hasattr(value, "to"):
+                progress_inputs[key] = value.to(self.device)
+
+        results = compute_batch_outputs(
+            self.reward_model,
+            self.tokenizer,
+            progress_inputs,
+            sample_type="progress",
+            is_discrete_mode=self.is_discrete,
+            num_bins=self.num_bins,
+        )
+
+        progress_pred = results.get("progress_pred", []) or []
+        outputs_success = results.get("outputs_success", {})
+        success_probs = outputs_success.get("success_probs", []) if outputs_success else []
+        return progress_pred, success_probs
+
+    @torch.inference_mode()
+    def _compute_rewards_whole_trajectory(self, video_frames: np.ndarray, task: str) -> tuple[np.ndarray, np.ndarray]:
+        traj = Trajectory(
+            frames=video_frames,
+            frames_shape=tuple(video_frames.shape),
+            task=task,
+            id="0",
+            metadata={"subsequence_length": int(video_frames.shape[0])},
+            video_embeddings=None,
+        )
+        progress_sample = ProgressSample(trajectory=traj, sample_type="progress")
+        progress_pred, success_probs = self._run_progress_samples([progress_sample])
+
+        progress_array = (
+            np.array(progress_pred[0], dtype=np.float32)
+            if progress_pred and len(progress_pred) > 0
+            else np.array([], dtype=np.float32)
+        )
+
+        success_array = (
+            np.array(success_probs[0], dtype=np.float32)
+            if success_probs and len(success_probs) > 0
+            else np.array([], dtype=np.float32)
+        )
+
+        return progress_array, success_array
+
+    @torch.inference_mode()
+    def _compute_rewards_frame_steps(
+        self,
+        video_frames: np.ndarray,
+        task: str,
+        *,
+        prefix_sample_frames: int,
+        prefix_batch_size: int,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        total_frames = int(video_frames.shape[0])
+        if total_frames <= 0:
+            return np.array([], dtype=np.float32), np.array([], dtype=np.float32)
+
+        prefix_sample_frames = max(1, int(prefix_sample_frames))
+        batch_size = max(1, int(prefix_batch_size))
+
+        prefix_samples: list[ProgressSample] = []
+        metadata = {"subsequence_length": total_frames}
+        for prefix_len in range(1, total_frames + 1):
+            indices = np.linspace(0, prefix_len - 1, prefix_sample_frames, dtype=int)
+            sub_frames = video_frames[indices]
+            traj = Trajectory(
+                frames=sub_frames,
+                frames_shape=tuple(sub_frames.shape),
+                task=task,
+                id="0",
+                metadata=metadata,
+                video_embeddings=None,
+            )
+            prefix_samples.append(ProgressSample(trajectory=traj, sample_type="progress"))
+
+        progress_values: list[float] = []
+        success_values: list[float] = []
+        cursor = 0
+
+        while cursor < len(prefix_samples):
+            current_samples = prefix_samples[cursor:cursor + batch_size]
+            try:
+                progress_pred, success_probs = self._run_progress_samples(current_samples)
+            except RuntimeError as exc:
+                if not is_cuda_oom_error(exc):
+                    raise
+                if batch_size == 1:
+                    raise
+                next_batch_size = max(1, batch_size // 2)
+                print(
+                    f"[OOM] frame_steps batch_size={batch_size} failed; retrying with batch_size={next_batch_size}"
+                )
+                cleanup_cuda_memory()
+                batch_size = next_batch_size
+                continue
+
+            for sub_pred in progress_pred:
+                if not sub_pred:
+                    raise RuntimeError("Robometer returned empty progress predictions for a frame-step prefix")
+                progress_values.append(float(sub_pred[-1]))
+
+            for idx in range(len(current_samples)):
+                sub_success = success_probs[idx] if idx < len(success_probs) else []
+                success_values.append(float(sub_success[-1]) if sub_success else np.nan)
+
+            cursor += len(current_samples)
+
+        success_array = np.array(success_values, dtype=np.float32)
+        if success_array.size > 0 and np.isnan(success_array).all():
+            success_array = np.array([], dtype=np.float32)
+
+        return np.array(progress_values, dtype=np.float32), success_array
+
+    @torch.inference_mode()
+    def compute_rewards_per_frame(
+        self,
+        video_frames: np.ndarray,
+        task: str,
+        *,
+        inference_mode: str = DEFAULT_INFERENCE_MODE,
+        prefix_sample_frames: int = DEFAULT_PREFIX_SAMPLE_FRAMES,
+        prefix_batch_size: int = DEFAULT_PREFIX_BATCH_SIZE,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        if inference_mode == "frame_steps":
+            return self._compute_rewards_frame_steps(
+                video_frames=video_frames,
+                task=task,
+                prefix_sample_frames=prefix_sample_frames,
+                prefix_batch_size=prefix_batch_size,
+            )
+        if inference_mode == "whole":
+            return self._compute_rewards_whole_trajectory(video_frames=video_frames, task=task)
+        raise ValueError(f"Unsupported inference_mode: {inference_mode}")
+
+
+def compute_rewards_per_frame_local(
+    model_path: str,
+    video_frames: np.ndarray,
+    task: str,
+    device: Optional[torch.device] = None,
+) -> tuple[np.ndarray, np.ndarray]:
+    runner = RobometerLocalRunner(model_path=model_path, device=device)
+    return runner.compute_rewards_per_frame(video_frames=video_frames, task=task)
+
+
+def map_frame_to_sample(frame_idx: int, sampled_indices: list[int]) -> tuple[int, int]:
+    nearest_pos = min(range(len(sampled_indices)), key=lambda idx: abs(sampled_indices[idx] - frame_idx))
+    return nearest_pos, sampled_indices[nearest_pos]
+
+
+def build_mark_points(
+    frame_pairs: list[tuple[str, int, int]],
+    sampled_indices: list[int],
+    progress_pred: np.ndarray,
+) -> list[dict]:
+    points = []
+    for label, f0, f1 in frame_pairs:
+        for suffix, frame_idx in (("1", f0), ("2", f1)):
+            sample_pos, sampled_frame = map_frame_to_sample(frame_idx, sampled_indices)
+            progress = float(progress_pred[sample_pos])
+            points.append(
+                {
+                    "group": label,
+                    "tag": f"{label}{suffix}",
+                    "original_frame": frame_idx,
+                    "sample_index": sample_pos,
+                    "sampled_frame": sampled_frame,
+                    "frame_error": sampled_frame - frame_idx,
+                    "progress": progress,
+                }
+            )
+    return points
+
+
+def write_overview_plot(
+    output_dir: Path,
+    sampled_indices: list[int],
+    progress_pred: np.ndarray,
+    mark_points: list[dict],
+) -> None:
+    fig, ax = plt.subplots(figsize=(10.5, 4.8))
+    ax.plot(sampled_indices, progress_pred, linewidth=2.0, color="#444444")
+
+    palette = plt.get_cmap("tab10")
+    grouped: dict[str, list[dict]] = defaultdict(list)
+    for point in mark_points:
+        grouped[point["group"]].append(point)
+
+    for idx, label in enumerate(sorted(grouped)):
+        pts = sorted(grouped[label], key=lambda item: item["original_frame"])
+        color = palette(idx % 10)
+        ax.plot(
+            [pt["sampled_frame"] for pt in pts],
+            [pt["progress"] for pt in pts],
+            marker="o",
+            linewidth=1.8,
+            markersize=7,
+            color=color,
+        )
+        for pt in pts:
+            ax.text(
+                pt["sampled_frame"],
+                pt["progress"] + 0.03,
+                f"{pt['tag']} {pt['progress']:.3f}",
+                color=color,
+                fontsize=9,
+                ha="center",
+            )
+
+    ax.set_xlabel("Original Frame")
+    ax.set_ylabel("Progress")
+    ax.set_ylim(-0.05, 1.05)
+    ax.set_title("Robometer Progress Curve")
+    ax.grid(True, linestyle="--", alpha=0.35)
+    fig.tight_layout()
+    fig.savefig(output_dir / "overview.png", dpi=180)
+    plt.close(fig)
+
+
+def write_group_cards(
+    output_dir: Path,
+    video_path: Path,
+    sampled_indices: list[int],
+    progress_pred: np.ndarray,
+    mark_points: list[dict],
+) -> None:
+    groups_dir = output_dir / "groups"
+    groups_dir.mkdir(parents=True, exist_ok=True)
+    palette = plt.get_cmap("tab10")
+
+    grouped: dict[str, list[dict]] = defaultdict(list)
+    for point in mark_points:
+        grouped[point["group"]].append(point)
+
+    for idx, label in enumerate(sorted(grouped)):
+        pts = sorted(grouped[label], key=lambda item: item["original_frame"])
+        color = palette(idx % 10)
+        first = pts[0]
+        second = pts[-1]
+        img0 = extract_frame(video_path, first["original_frame"])
+        img1 = extract_frame(video_path, second["original_frame"])
+
+        frame_gap = abs(second["original_frame"] - first["original_frame"])
+        x_min = max(0, min(first["original_frame"], second["original_frame"]) - max(25, frame_gap))
+        x_max = max(first["original_frame"], second["original_frame"]) + max(25, frame_gap)
+
+        local_x = []
+        local_y = []
+        for x, y in zip(sampled_indices, progress_pred, strict=False):
+            if x_min <= x <= x_max:
+                local_x.append(x)
+                local_y.append(float(y))
+        if not local_x:
+            local_x = sampled_indices
+            local_y = progress_pred.tolist()
+
+        fig = plt.figure(figsize=(9.5, 6.5))
+        gs = fig.add_gridspec(2, 2, height_ratios=[1.45, 1.0])
+        ax_img0 = fig.add_subplot(gs[0, 0])
+        ax_img1 = fig.add_subplot(gs[0, 1])
+        ax_curve = fig.add_subplot(gs[1, :])
+
+        for ax, img, title in (
+            (ax_img0, img0, f"{first['tag']} frame {first['original_frame']}"),
+            (ax_img1, img1, f"{second['tag']} frame {second['original_frame']}"),
+        ):
+            ax.axis("off")
+            ax.set_title(title, fontsize=10)
+            if img is None:
+                ax.text(0.5, 0.5, "frame read failed", ha="center", va="center", fontsize=9)
+            else:
+                ax.imshow(img)
+
+        ax_curve.plot(local_x, local_y, marker="o", linewidth=1.6, markersize=4, color="#666666")
+        ax_curve.plot(
+            [pt["sampled_frame"] for pt in pts],
+            [pt["progress"] for pt in pts],
+            marker="o",
+            linewidth=2.0,
+            markersize=7,
+            color=color,
+        )
+        for pt in pts:
+            ax_curve.text(
+                pt["sampled_frame"],
+                pt["progress"] + 0.03,
+                f"{pt['tag']} {pt['progress']:.3f}",
+                color=color,
+                fontsize=9,
+                ha="center",
+            )
+
+        ax_curve.set_xlim(x_min, x_max)
+        ax_curve.set_ylim(-0.05, 1.05)
+        ax_curve.set_xlabel("Original Frame")
+        ax_curve.set_ylabel("Progress")
+        ax_curve.set_title(f"{label} Group")
+        ax_curve.grid(True, linestyle="--", alpha=0.35)
+        ax_curve.text(
+            0.015,
+            0.97,
+            "\n".join(
+                f"{pt['tag']}: frame={pt['original_frame']}, sampled={pt['sampled_frame']}, progress={pt['progress']:.3f}"
+                for pt in pts
+            ),
+            transform=ax_curve.transAxes,
+            va="top",
+            ha="left",
+            fontsize=9,
+            bbox={"facecolor": "white", "edgecolor": "#cccccc", "alpha": 0.92, "boxstyle": "round,pad=0.35"},
+        )
+        fig.tight_layout()
+        fig.savefig(groups_dir / f"{label}.png", dpi=180)
+        plt.close(fig)
+
+
+def save_results(output_dir: Path, payload: dict) -> None:
+    with (output_dir / "results.json").open("w", encoding="utf-8") as f:
+        json.dump(payload, f, indent=2, ensure_ascii=False)
+        f.write("\n")
+
+
+def run_episode(
+    *,
+    episode_mp4: str,
+    chunk: str,
+    fps: float,
+    max_frames: int,
+    min_frames: int,
+    adaptive_max_frames: bool,
+    inference_mode: str,
+    prefix_sample_frames: int,
+    prefix_batch_size: int,
+    runner: RobometerLocalRunner,
+    overwrite: bool,
+    videos_root_base: Path,
+    allow_no_marks: bool,
+    tag: str | None = None,
+) -> None:
+    direct_chunk_dir = videos_root_base / chunk
+    if direct_chunk_dir.exists():
+        videos_root = direct_chunk_dir
+    else:
+        chunk_filtered = f"{chunk}_filtered"
+        videos_root = videos_root_base / chunk_filtered
+    annotations_path = videos_root / "annotations.json"
+    episode_tasks_path = videos_root / "episode_tasks.json"
+    video_path = videos_root / DEFAULT_CAMERA_DIR / episode_mp4
+    output_dir = DEFAULT_OUTPUT_ROOT / chunk_output_dir_name(chunk, tag) / episode_mp4[:-4]
+
+    if not overwrite and (output_dir / "results.json").exists():
+        print(f"[SKIP] {episode_mp4[:-4]} already done")
+        return
+
+    episode_record = load_episode_record(annotations_path, episode_mp4)
+    instruction = load_instruction(episode_tasks_path, episode_mp4)
+    frame_pairs = extract_progress_pairs(episode_record)
+    if not frame_pairs and not allow_no_marks:
+        raise ValueError(f"{episode_mp4} has no usable progress_mark pairs")
+    if not video_path.exists():
+        raise FileNotFoundError(f"Video not found: {video_path}")
+
+    required_frames = [frame for _, f0, f1 in frame_pairs for frame in (f0, f1)]
+
+    print(f"[RUN] {episode_mp4[:-4]}")
+    print(f"Loading frames from {video_path} ...")
+    all_frames, native_fps = load_all_video_frames(video_path)
+    total_frames = len(all_frames)
+    retry_schedule = (
+        build_frame_retry_schedule(max_frames, min_frames, adaptive_max_frames)
+        if inference_mode == "whole"
+        else [max_frames]
+    )
+    print(
+        f"[MODE] {inference_mode}"
+        + (
+            f" (prefix_sample_frames={prefix_sample_frames}, prefix_batch_size={prefix_batch_size})"
+            if inference_mode == "frame_steps"
+            else ""
+        )
+    )
+
+    progress_pred: np.ndarray | None = None
+    success_probs: np.ndarray | None = None
+    sampled_indices: list[int] | None = None
+    used_max_frames = retry_schedule[0]
+
+    for attempt_idx, frame_budget in enumerate(retry_schedule, start=1):
+        frames, sampled_indices = sample_video_frames_with_indices(
+            all_frames,
+            native_fps=native_fps,
+            fps=fps,
+            max_frames=frame_budget,
+            required_frames=required_frames,
+        )
+        print(
+            f"Loaded {total_frames} total frames; sampled {len(frames)} frames at fps={fps:g} "
+            f"(max_frames={frame_budget}, try {attempt_idx}/{len(retry_schedule)})"
+        )
+        try:
+            progress_pred, success_probs = runner.compute_rewards_per_frame(
+                video_frames=frames,
+                task=instruction,
+                inference_mode=inference_mode,
+                prefix_sample_frames=prefix_sample_frames,
+                prefix_batch_size=prefix_batch_size,
+            )
+            used_max_frames = frame_budget
+            break
+        except RuntimeError as exc:
+            if inference_mode != "whole" or not is_cuda_oom_error(exc):
+                raise
+            del frames
+            cleanup_cuda_memory()
+            if attempt_idx == len(retry_schedule):
+                raise
+            next_budget = retry_schedule[attempt_idx]
+            print(
+                f"[OOM] {episode_mp4[:-4]} hit CUDA OOM at max_frames={frame_budget}; "
+                f"retrying with max_frames={next_budget}"
+            )
+            runner.reload_model()
+
+    if progress_pred is None or success_probs is None or sampled_indices is None:
+        raise RuntimeError(f"Failed to compute rewards for {episode_mp4[:-4]}")
+    if progress_pred.size == 0:
+        raise RuntimeError("Robometer returned empty progress predictions")
+    if progress_pred.size != len(sampled_indices):
+        raise RuntimeError(
+            f"Progress length mismatch: got {progress_pred.size} predictions for {len(sampled_indices)} sampled frames"
+        )
+
+    mark_points = build_mark_points(frame_pairs, sampled_indices, progress_pred)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    np.save(str(output_dir / "progress.npy"), progress_pred)
+    np.save(str(output_dir / "success_probs.npy"), success_probs)
+
+    payload = {
+        "episode": episode_mp4,
+        "chunk": chunk,
+        "instruction": instruction,
+        "video_path": str(video_path),
+        "num_total_frames": total_frames,
+        "native_fps": native_fps,
+        "sample_fps": float(fps),
+        "num_sampled_frames": len(sampled_indices),
+        "max_frames_used": used_max_frames,
+        "inference_mode": inference_mode,
+        "prefix_sample_frames": int(prefix_sample_frames),
+        "sampled_original_frame_indices": sampled_indices,
+        "progress_marks": [{"group": label, "frame_pair": [f0, f1]} for label, f0, f1 in frame_pairs],
+        "mark_points": mark_points,
+        "progress_min": float(np.min(progress_pred)),
+        "progress_max": float(np.max(progress_pred)),
+        "progress_mean": float(np.mean(progress_pred)),
+    }
+    save_results(output_dir, payload)
+    write_overview_plot(output_dir, sampled_indices, progress_pred, mark_points)
+    write_group_cards(output_dir, video_path, sampled_indices, progress_pred, mark_points)
+
+    print(f"[OK] Saved results to {output_dir}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Run Robometer local inference aligned to progress_mark frames.")
+    parser.add_argument("--episode", default=None, help="000287 / episode_000287 / episode_000287.mp4")
+    parser.add_argument(
+        "--count",
+        type=int,
+        default=1,
+        help="when --episode is set, run this many progress_mark episodes from that start",
+    )
+    parser.add_argument("--chunk", default=DEFAULT_CHUNK)
+    parser.add_argument("--videos-root", type=Path, default=DEFAULT_VIDEOS_ROOT)
+    parser.add_argument("--model-path", default=str(DEFAULT_MODEL_PATH))
+    parser.add_argument("--fps", type=float, default=3.0)
+    parser.add_argument("--max-frames", type=int, default=DEFAULT_MAX_FRAMES)
+    parser.add_argument("--min-frames", type=int, default=DEFAULT_MIN_FRAMES)
+    parser.add_argument(
+        "--inference-mode",
+        choices=["frame_steps", "whole"],
+        default=DEFAULT_INFERENCE_MODE,
+        help="frame_steps: stable dense curve via prefix inference; whole: one forward pass on the full sampled trajectory",
+    )
+    parser.add_argument("--prefix-sample-frames", type=int, default=DEFAULT_PREFIX_SAMPLE_FRAMES)
+    parser.add_argument("--tag", type=str, default=None,
+                        help="Optional suffix appended to chunk output dir (e.g. 'original' → Chunk000_original)")
+    parser.add_argument("--prefix-batch-size", type=int, default=DEFAULT_PREFIX_BATCH_SIZE)
+    parser.add_argument(
+        "--adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_true",
+        help="on CUDA OOM, retry an episode with smaller frame budgets",
+    )
+    parser.add_argument(
+        "--no-adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_false",
+        help="disable dynamic frame-budget retry",
+    )
+    parser.add_argument("--overwrite", action="store_true", help="rerun episodes even if results.json already exists")
+    parser.add_argument("--allow-no-marks", action="store_true", help="allow episodes without progress_mark pairs")
+    parser.set_defaults(adaptive_max_frames=True)
+    args = parser.parse_args()
+
+    if "CUDA_VISIBLE_DEVICES" not in os.environ:
+        user_gpu = input("GPU id (default 5): ").strip()
+        os.environ["CUDA_VISIBLE_DEVICES"] = user_gpu if user_gpu else "5"
+
+    chunks = parse_chunk_selection(args.chunk)
+    runner = RobometerLocalRunner(model_path=args.model_path)
+    ran_any = False
+
+    for chunk in chunks:
+        direct_chunk_dir = args.videos_root / chunk
+        if direct_chunk_dir.exists():
+            videos_root = direct_chunk_dir
+        else:
+            chunk_filtered = f"{chunk}_filtered"
+            videos_root = args.videos_root / chunk_filtered
+        annotations_path = videos_root / "annotations.json"
+        episode_tasks_path = videos_root / "episode_tasks.json"
+        catalog = load_episode_catalog(episode_tasks_path, annotations_path)
+        progress_catalog = load_progress_mark_catalog(annotations_path)
+        require_progress_marks = not bool(args.allow_no_marks)
+
+        episode_list = resolve_episode_list(
+            episode_arg=args.episode,
+            count=args.count,
+            catalog=catalog,
+            progress_catalog=progress_catalog,
+            require_progress_marks=require_progress_marks,
+        )
+
+        output_root = DEFAULT_OUTPUT_ROOT / chunk_output_dir_name(chunk, args.tag)
+        if not args.overwrite:
+            filtered = []
+            for episode_mp4 in episode_list:
+                if result_path_for(output_root, episode_mp4).exists():
+                    print(f"[SKIP] {episode_mp4[:-4]} already done")
+                    continue
+                filtered.append(episode_mp4)
+            episode_list = filtered
+
+        if not episode_list:
+            print(f"[OK] {chunk}: nothing to run.")
+            continue
+
+        ran_any = True
+        print(f"[CHUNK] {chunk}")
+        print(f"[QUEUE] {', '.join(ep[:-4] for ep in episode_list)}")
+        for episode_mp4 in episode_list:
+            run_episode(
+                episode_mp4=episode_mp4,
+                chunk=chunk,
+                fps=float(args.fps),
+                max_frames=int(args.max_frames),
+                min_frames=int(args.min_frames),
+                adaptive_max_frames=bool(args.adaptive_max_frames),
+                inference_mode=str(args.inference_mode),
+                prefix_sample_frames=int(args.prefix_sample_frames),
+                prefix_batch_size=int(args.prefix_batch_size),
+                runner=runner,
+                overwrite=args.overwrite,
+                videos_root_base=args.videos_root,
+                allow_no_marks=bool(args.allow_no_marks),
+                tag=args.tag,
+            )
+
+    if not ran_any:
+        print("[OK] Nothing to run.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/count_trajectories.py b/robometer/scripts/count_trajectories.py
new file mode 100755
index 0000000000000000000000000000000000000000..6554a8003aafd10a9547698a04d21886ffaae13d
--- /dev/null
+++ b/robometer/scripts/count_trajectories.py
@@ -0,0 +1,370 @@
+#!/usr/bin/env python3
+"""
+Script to count trajectories in each processed dataset.
+
+This script reads the preprocess.yaml configuration and counts the number of
+trajectories in each dataset/subset combination by loading the processed datasets.
+"""
+
+import os
+import json
+import yaml
+from pathlib import Path
+from datasets import Dataset
+from rich.console import Console
+from rich.table import Table
+from rich import print as rprint
+
+
+def load_preprocess_config(config_path: str = "robometer/configs/preprocess.yaml") -> dict:
+    """Load the preprocess configuration file."""
+    with open(config_path, "r") as f:
+        config = yaml.safe_load(f)
+    return config
+
+
+def get_cache_dir_for_dataset(base_cache_dir: str, dataset_path: str, subset: str) -> str:
+    """
+    Construct the cache directory path for a dataset/subset pair.
+
+    Following the logic from robometer/data/scripts/preprocess_datasets.py:
+    cache_key = f"{dataset_path}/{subset}"
+    individual_cache_dir = os.path.join(cache_dir, cache_key.replace("/", "_").replace(":", "_"))
+    """
+    cache_key = f"{dataset_path}/{subset}"
+    individual_cache_dir = os.path.join(base_cache_dir, cache_key.replace("/", "_").replace(":", "_"))
+    return individual_cache_dir
+
+
+def count_trajectories_in_dataset(cache_dir: str) -> dict:
+    """
+    Count trajectories in a processed dataset directory.
+
+    Returns a dictionary with:
+    - 'count': number of trajectories
+    - 'exists': whether the dataset exists
+    - 'error': any error message
+    """
+    result = {"count": 0, "exists": False, "error": None}
+
+    # Check if cache directory exists
+    if not os.path.exists(cache_dir):
+        result["error"] = "Cache directory not found"
+        return result
+
+    # Check for dataset_info.json
+    info_file = os.path.join(cache_dir, "dataset_info.json")
+    if not os.path.exists(info_file):
+        result["error"] = "dataset_info.json not found"
+        return result
+
+    # Check for processed_dataset directory
+    dataset_cache_dir = os.path.join(cache_dir, "processed_dataset")
+    if not os.path.exists(dataset_cache_dir):
+        result["error"] = "processed_dataset directory not found"
+        return result
+
+    result["exists"] = True
+
+    try:
+        # Load the dataset to count trajectories
+        dataset = Dataset.load_from_disk(dataset_cache_dir, keep_in_memory=False)
+        result["count"] = len(dataset)
+
+        # Also read dataset_info.json for additional metadata
+        with open(info_file, "r") as f:
+            info = json.load(f)
+            result["info"] = info
+
+    except Exception as e:
+        result["error"] = f"Error loading dataset: {str(e)}"
+
+    return result
+
+
+def main():
+    """Main function to count trajectories in all processed datasets."""
+    console = Console()
+
+    # Get the processed datasets path from environment variable
+    cache_dir = os.environ.get("ROBOMETER_PROCESSED_DATASETS_PATH", "")
+
+    if not cache_dir:
+        console.print(
+            "[bold red]Error:[/bold red] ROBOMETER_PROCESSED_DATASETS_PATH environment variable is not set!"
+        )
+        console.print("Please set it to the directory containing your processed datasets.")
+        console.print("Example: export ROBOMETER_PROCESSED_DATASETS_PATH=/path/to/robometer/processed_datasets")
+        return
+
+    console.print(f"[bold]Using cache directory:[/bold] {cache_dir}")
+
+    # Load preprocess configuration
+    config_path = "robometer/configs/preprocess.yaml"
+    if not os.path.exists(config_path):
+        console.print(f"[bold red]Error:[/bold red] Config file not found: {config_path}")
+        return
+
+    config = load_preprocess_config(config_path)
+    console.print(f"[bold]Loaded config from:[/bold] {config_path}\n")
+
+    # Process training datasets
+    train_datasets = config.get("train_datasets", [])
+    train_subsets = config.get("train_subsets", [])
+
+    # Process evaluation datasets
+    eval_datasets = config.get("eval_datasets", [])
+    eval_subsets = config.get("eval_subsets", [])
+
+    # Create tables for display
+    train_table = Table(title="Training Datasets - Trajectory Counts", show_header=True, header_style="bold magenta")
+    train_table.add_column("Dataset", style="cyan", width=40)
+    train_table.add_column("Subset", style="green", width=35)
+    train_table.add_column("Trajectories", justify="right", style="yellow")
+    train_table.add_column("Status", style="white")
+
+    eval_table = Table(title="Evaluation Datasets - Trajectory Counts", show_header=True, header_style="bold magenta")
+    eval_table.add_column("Dataset", style="cyan", width=40)
+    eval_table.add_column("Subset", style="green", width=35)
+    eval_table.add_column("Trajectories", justify="right", style="yellow")
+    eval_table.add_column("Status", style="white")
+
+    # Count training datasets
+    console.print("[bold blue]Processing Training Datasets...[/bold blue]")
+    train_total = 0
+    train_found = 0
+    train_dataset_aggregates = {}  # To store aggregated counts per dataset
+    train_rows = []  # Collect rows for sorting
+
+    for dataset_path, dataset_subsets in zip(train_datasets, train_subsets):
+        if dataset_path not in train_dataset_aggregates:
+            train_dataset_aggregates[dataset_path] = {"count": 0, "found": 0, "total_subsets": 0}
+
+        for subset in dataset_subsets:
+            train_dataset_aggregates[dataset_path]["total_subsets"] += 1
+            individual_cache_dir = get_cache_dir_for_dataset(cache_dir, dataset_path, subset)
+            result = count_trajectories_in_dataset(individual_cache_dir)
+
+            if result["exists"] and result["error"] is None:
+                count_str = f"{result['count']:,}"
+                status = "✓ Found"
+                status_style = "green"
+                train_total += result["count"]
+                train_found += 1
+                train_dataset_aggregates[dataset_path]["count"] += result["count"]
+                train_dataset_aggregates[dataset_path]["found"] += 1
+                count_value = result["count"]
+            else:
+                count_str = "N/A"
+                status = f"✗ {result['error']}"
+                status_style = "red"
+                count_value = -1  # Use -1 for sorting (will appear at end)
+
+            train_rows.append((
+                count_value,
+                dataset_path,
+                subset,
+                count_str,
+                f"[{status_style}]{status}[/{status_style}]",
+            ))
+
+    # Sort by count (descending), then add to table
+    train_rows.sort(key=lambda x: x[0], reverse=True)
+    for _, dataset_path, subset, count_str, status_display in train_rows:
+        train_table.add_row(dataset_path, subset, count_str, status_display)
+
+    # Count evaluation datasets
+    console.print("[bold blue]Processing Evaluation Datasets...[/bold blue]")
+    eval_total = 0
+    eval_found = 0
+    eval_dataset_aggregates = {}  # To store aggregated counts per dataset
+    eval_rows = []  # Collect rows for sorting
+
+    for dataset_path, dataset_subsets in zip(eval_datasets, eval_subsets):
+        if dataset_path not in eval_dataset_aggregates:
+            eval_dataset_aggregates[dataset_path] = {"count": 0, "found": 0, "total_subsets": 0}
+
+        for subset in dataset_subsets:
+            eval_dataset_aggregates[dataset_path]["total_subsets"] += 1
+            individual_cache_dir = get_cache_dir_for_dataset(cache_dir, dataset_path, subset)
+            result = count_trajectories_in_dataset(individual_cache_dir)
+
+            if result["exists"] and result["error"] is None:
+                count_str = f"{result['count']:,}"
+                status = "✓ Found"
+                status_style = "green"
+                eval_total += result["count"]
+                eval_found += 1
+                eval_dataset_aggregates[dataset_path]["count"] += result["count"]
+                eval_dataset_aggregates[dataset_path]["found"] += 1
+                count_value = result["count"]
+            else:
+                count_str = "N/A"
+                status = f"✗ {result['error']}"
+                status_style = "red"
+                count_value = -1  # Use -1 for sorting (will appear at end)
+
+            eval_rows.append((
+                count_value,
+                dataset_path,
+                subset,
+                count_str,
+                f"[{status_style}]{status}[/{status_style}]",
+            ))
+
+    # Sort by count (descending), then add to table
+    eval_rows.sort(key=lambda x: x[0], reverse=True)
+    for _, dataset_path, subset, count_str, status_display in eval_rows:
+        eval_table.add_row(dataset_path, subset, count_str, status_display)
+
+    # Display results
+    console.print()
+    console.print(train_table)
+    console.print()
+    console.print(eval_table)
+    console.print()
+
+    # Create aggregated tables
+    train_agg_table = Table(
+        title="Training Datasets - Aggregated by Dataset", show_header=True, header_style="bold magenta"
+    )
+    train_agg_table.add_column("Dataset", style="cyan", width=50)
+    train_agg_table.add_column("Total Trajectories", justify="right", style="yellow")
+    train_agg_table.add_column("Subsets Found/Total", justify="center", style="green")
+
+    # Sort by count (descending)
+    train_agg_items = sorted(train_dataset_aggregates.items(), key=lambda x: x[1]["count"], reverse=True)
+
+    for dataset_path, agg in train_agg_items:
+        count_str = f"{agg['count']:,}" if agg["count"] > 0 else "0"
+        subsets_str = f"{agg['found']}/{agg['total_subsets']}"
+
+        # Color code based on whether all subsets were found
+        if agg["found"] == agg["total_subsets"] and agg["total_subsets"] > 0:
+            subsets_display = f"[green]{subsets_str}[/green]"
+        elif agg["found"] > 0:
+            subsets_display = f"[yellow]{subsets_str}[/yellow]"
+        else:
+            subsets_display = f"[red]{subsets_str}[/red]"
+
+        train_agg_table.add_row(dataset_path, count_str, subsets_display)
+
+    eval_agg_table = Table(
+        title="Evaluation Datasets - Aggregated by Dataset", show_header=True, header_style="bold magenta"
+    )
+    eval_agg_table.add_column("Dataset", style="cyan", width=50)
+    eval_agg_table.add_column("Total Trajectories", justify="right", style="yellow")
+    eval_agg_table.add_column("Subsets Found/Total", justify="center", style="green")
+
+    # Sort by count (descending)
+    eval_agg_items = sorted(eval_dataset_aggregates.items(), key=lambda x: x[1]["count"], reverse=True)
+
+    for dataset_path, agg in eval_agg_items:
+        count_str = f"{agg['count']:,}" if agg["count"] > 0 else "0"
+        subsets_str = f"{agg['found']}/{agg['total_subsets']}"
+
+        # Color code based on whether all subsets were found
+        if agg["found"] == agg["total_subsets"] and agg["total_subsets"] > 0:
+            subsets_display = f"[green]{subsets_str}[/green]"
+        elif agg["found"] > 0:
+            subsets_display = f"[yellow]{subsets_str}[/yellow]"
+        else:
+            subsets_display = f"[red]{subsets_str}[/red]"
+
+        eval_agg_table.add_row(dataset_path, count_str, subsets_display)
+
+    # Display aggregated tables
+    console.print(train_agg_table)
+    console.print()
+    console.print(eval_agg_table)
+    console.print()
+
+    # Summary
+    summary_table = Table(title="Summary", show_header=True, header_style="bold magenta")
+    summary_table.add_column("Category", style="cyan")
+    summary_table.add_column("Total Trajectories", justify="right", style="yellow")
+    summary_table.add_column("Datasets Found", justify="right", style="green")
+
+    summary_table.add_row("Training", f"{train_total:,}", str(train_found))
+    summary_table.add_row("Evaluation", f"{eval_total:,}", str(eval_found))
+    summary_table.add_row(
+        "[bold]Grand Total[/bold]",
+        f"[bold]{train_total + eval_total:,}[/bold]",
+        f"[bold]{train_found + eval_found}[/bold]",
+    )
+
+    console.print(summary_table)
+    console.print()
+
+    # Save results to JSON
+    output_file = "trajectory_counts.json"
+
+    # Prepare aggregated data for JSON
+    train_aggregates_json = {}
+    for dataset_path, agg in train_dataset_aggregates.items():
+        train_aggregates_json[dataset_path] = {
+            "total_trajectories": agg["count"],
+            "subsets_found": agg["found"],
+            "total_subsets": agg["total_subsets"],
+        }
+
+    eval_aggregates_json = {}
+    for dataset_path, agg in eval_dataset_aggregates.items():
+        eval_aggregates_json[dataset_path] = {
+            "total_trajectories": agg["count"],
+            "subsets_found": agg["found"],
+            "total_subsets": agg["total_subsets"],
+        }
+
+    results = {
+        "cache_dir": cache_dir,
+        "training_datasets": [],
+        "training_datasets_aggregated": train_aggregates_json,
+        "evaluation_datasets": [],
+        "evaluation_datasets_aggregated": eval_aggregates_json,
+        "summary": {
+            "train_total_trajectories": train_total,
+            "train_datasets_found": train_found,
+            "eval_total_trajectories": eval_total,
+            "eval_datasets_found": eval_found,
+            "grand_total_trajectories": train_total + eval_total,
+            "total_datasets_found": train_found + eval_found,
+        },
+    }
+
+    # Add training dataset details
+    for dataset_path, dataset_subsets in zip(train_datasets, train_subsets):
+        for subset in dataset_subsets:
+            individual_cache_dir = get_cache_dir_for_dataset(cache_dir, dataset_path, subset)
+            result = count_trajectories_in_dataset(individual_cache_dir)
+            results["training_datasets"].append({
+                "dataset": dataset_path,
+                "subset": subset,
+                "cache_dir": individual_cache_dir,
+                "trajectory_count": result["count"] if result["exists"] else None,
+                "exists": result["exists"],
+                "error": result["error"],
+            })
+
+    # Add evaluation dataset details
+    for dataset_path, dataset_subsets in zip(eval_datasets, eval_subsets):
+        for subset in dataset_subsets:
+            individual_cache_dir = get_cache_dir_for_dataset(cache_dir, dataset_path, subset)
+            result = count_trajectories_in_dataset(individual_cache_dir)
+            results["evaluation_datasets"].append({
+                "dataset": dataset_path,
+                "subset": subset,
+                "cache_dir": individual_cache_dir,
+                "trajectory_count": result["count"] if result["exists"] else None,
+                "exists": result["exists"],
+                "error": result["error"],
+            })
+
+    with open(output_file, "w") as f:
+        json.dump(results, f, indent=2)
+
+    console.print(f"[bold green]Results saved to:[/bold green] {output_file}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/download_data.sh b/robometer/scripts/download_data.sh
new file mode 100755
index 0000000000000000000000000000000000000000..4f736eb1e7dbb08153da0b775d14067cc23e6a7e
--- /dev/null
+++ b/robometer/scripts/download_data.sh
@@ -0,0 +1,158 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# ------------------------------------------------------------------
+# Fast dataset setup via either Hugging Face CLI downloads (default)
+# or git-lfs clones from the Hub.
+#
+#  * HF CLI path requires: `pip install huggingface_hub` (provides `hf` CLI)
+#  * Git path requires: `git` and `git-lfs`
+# ------------------------------------------------------------------
+
+# ------------------------------
+# Parse command‑line arguments
+# ------------------------------
+METHOD=${RFM_DOWNLOAD_METHOD:-hf}   # hf | git
+BASE_DIR_DEFAULT=${ROBOMETER_DATASET_PATH:-${RFM_DATASET_PATH:-./robometer_dataset}}
+BASE_DIR="$BASE_DIR_DEFAULT"
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --git)           METHOD="git"; shift ;;
+    --hf)            METHOD="hf"; shift ;;
+    --method=*)      METHOD="${1#*=}"; shift ;;
+    --method)        METHOD="$2"; shift 2 ;;
+    --dir|--base-dir|-d)
+                      BASE_DIR="$2"; shift 2 ;;
+    *)               BASE_DIR="$1"; shift ;;
+  esac
+done
+
+# ------------------------------
+# Sanity checks for the chosen method
+# ------------------------------
+case "$METHOD" in
+  hf)
+    if ! command -v hf >/dev/null 2>&1; then
+      echo "Error: 'hf' CLI not found. Install with:" >&2
+      echo "   uv pip install huggingface_hub (or ensure your venv is activated)" >&2
+      exit 1
+    fi
+    ;;
+  git)
+    if ! command -v git >/dev/null 2>&1; then
+      echo "Error: git not found. Please install git." >&2
+      exit 1
+    fi
+    if ! git lfs version >/dev/null 2>&1; then
+      echo "Warning: git-lfs not found. You may end up with pointer files." >&2
+      echo "   Install git-lfs for full downloads." >&2
+    fi
+    ;;
+  *)
+    echo "Error: Unknown METHOD='${METHOD}'. Use 'hf' or 'git'." >&2
+    exit 1
+    ;;
+esac
+
+mkdir -p "${BASE_DIR}"
+
+# ------------------------------
+# Helper to download a dataset repo
+# ------------------------------
+download_dataset() {
+  local repo_id="$1"                # e.g., abraranwar/libero_rfm
+  local name="${repo_id##*/}"       # last path segment as folder name
+  local target_dir="${BASE_DIR}/${name}"
+
+  echo "Downloading ${repo_id} -> ${target_dir} via ${METHOD}"
+  if [[ "$METHOD" == "hf" ]]; then
+    hf download "${repo_id}" \
+      --repo-type dataset \
+      --local-dir "${target_dir}"
+  else
+    local url="https://huggingface.co/datasets/${repo_id}.git"
+    if [[ -d "${target_dir}/.git" ]]; then
+      echo "Updating existing clone at ${target_dir}"
+      git -C "${target_dir}" remote set-url origin "${url}" || true
+      git -C "${target_dir}" fetch --all --tags
+      git -C "${target_dir}" pull --ff-only
+    else
+      git clone "${url}" "${target_dir}"
+    fi
+    if git lfs version >/dev/null 2>&1; then
+      git -C "${target_dir}" lfs install --local >/dev/null 2>&1 || true
+      git -C "${target_dir}" lfs pull || true
+    fi
+  fi
+}
+
+# ------------------------------
+# Retry helper – keeps trying until success
+# ------------------------------
+retry_until_success() {
+  local cmd="$1"
+  local desc="${2:-$cmd}"
+
+  echo "=== ${desc} ==="
+  until eval "$cmd"; do
+    echo "❌ $desc failed – retrying in 30s …"
+    sleep 30
+  done
+  echo "✅ $desc succeeded."
+}
+
+# ------------------------------
+# List of all repos to download (including the ones that were previously commented out)
+# ------------------------------
+repos=(
+  abraranwar/libero_rfm
+  abraranwar/agibotworld_alpha_rfm   
+  abraranwar/agibotworld_alpha_headcam_rfm   
+  abraranwar/usc_koch_rewind_rfm
+  ykorkmaz/libero_failure_rfm
+  aliangdw/metaworld
+  jesbu1/oxe_rfm
+  jesbu1/galaxea_rfm
+  jesbu1/molmoact_rfm
+  jesbu1/ph2d_rfm
+  jesbu1/epic_rfm
+  jesbu1/failsafe_rfm
+  jesbu1/h2r_rfm
+  jesbu1/roboarena_0825_rfm
+  jesbu1/oxe_rfm_eval
+  anqil/rh20t_subset_rfm # can be replaced with anqil/rh20t_rfm full dataset
+  jesbu1/humanoid_everyday_rfm
+  jesbu1/motif_rfm
+  jesbu1/auto_eval_rfm
+  jesbu1/soar_rfm 
+  jesbu1/racer_rfm
+  jesbu1/egodex_rfm
+  aliangdw/usc_xarm_policy_ranking
+  aliangdw/usc_franka_policy_ranking
+  aliangdw/utd_so101_policy_ranking
+  aliangdw/utd_so101_human
+  jesbu1/mit_franka_p-rank_rfm
+  jesbu1/utd_so101_clean_policy_ranking_top
+  jesbu1/utd_so101_clean_policy_ranking_wrist
+  jesbu1/usc_koch_human_robot_paired
+  jesbu1/usc_koch_p_ranking_rfm
+  #jesbu1/roboreward_rfm
+  jesbu1/roboreward_rfm_high_res
+  jesbu1/rfm_new_mit_franka_rfm_nowrist
+  ykorkmaz/usc_trossen_rfm
+  aliangdw/robofac_rbm
+)
+
+# ------------------------------
+# Download each repo with retry logic
+# ------------------------------
+for repo in "${repos[@]}"; do
+  # Skip lines that are still commented out (start with '#')
+  [[ "$repo" == \#* ]] && continue
+
+  retry_until_success "download_dataset $repo" "$repo"
+done
+
+echo ""
+echo "Done. Set ROBOMETER_DATASET_PATH=${BASE_DIR} for training/eval."
diff --git a/robometer/scripts/download_processed_datasets.sh b/robometer/scripts/download_processed_datasets.sh
new file mode 100755
index 0000000000000000000000000000000000000000..ccf7c4f001d759140de034cbdc4c707d858bc712
--- /dev/null
+++ b/robometer/scripts/download_processed_datasets.sh
@@ -0,0 +1,10 @@
+#!/bin/bash
+# check ROBOMETER_PROCESSED_DATASETS_PATH (or RBM_PROCESSED_DATASETS_PATH) is set
+if [ -z "${ROBOMETER_PROCESSED_DATASETS_PATH:-$RBM_PROCESSED_DATASETS_PATH}" ]; then
+    echo "ROBOMETER_PROCESSED_DATASETS_PATH (or RBM_PROCESSED_DATASETS_PATH) is not set"
+    exit 1
+fi
+# increase the number of open files limit
+ulimit -n 65535
+# download processed datasets
+hf download jesbu1/processed_datasets --repo-type dataset --local-dir=${ROBOMETER_PROCESSED_DATASETS_PATH:-$RBM_PROCESSED_DATASETS_PATH}
\ No newline at end of file
diff --git a/robometer/scripts/example_inference.py b/robometer/scripts/example_inference.py
new file mode 100644
index 0000000000000000000000000000000000000000..4bc94ec2e6442b3aef59a602c849f29cda336ea8
--- /dev/null
+++ b/robometer/scripts/example_inference.py
@@ -0,0 +1,482 @@
+#!/usr/bin/env python3
+"""
+Client script for the RBM eval server. No robometer dependency.
+
+Sends a video (or .npy/.npz frames) and task instruction to a running eval server,
+then saves per-frame progress and success predictions plus an optional plot.
+
+Example:
+  # Start the server first (in another terminal):
+  #   uv run python robometer/evals/eval_server.py --config_path=robometer/configs/config.yaml --host=0.0.0.0 --port=8000
+
+  python scripts/example_inference.py --eval-server-url http://localhost:8000 --video /path/to/video.mp4 --task "Pick up the red block"
+"""
+
+from __future__ import annotations
+
+import argparse
+import io
+import json
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import requests
+
+import matplotlib.pyplot as plt
+
+
+def create_combined_progress_success_plot(
+    progress_pred: np.ndarray,
+    num_frames: int,
+    success_binary: Optional[np.ndarray] = None,
+    success_probs: Optional[np.ndarray] = None,
+    success_labels: Optional[np.ndarray] = None,
+    is_discrete_mode: bool = False,
+    title: Optional[str] = None,
+    loss: Optional[float] = None,
+    pearson: Optional[float] = None,
+) -> Any:
+    """Create a combined plot with progress, success binary, and success probabilities.
+
+    This function creates a unified plot with 1 subplot (progress only) or 3 subplots
+    (progress, success binary, success probs), similar to the one used in compile_results.py.
+
+    Args:
+        progress_pred: Progress predictions array
+        num_frames: Number of frames
+        success_binary: Optional binary success predictions
+        success_probs: Optional success probability predictions
+        success_labels: Optional ground truth success labels
+        is_discrete_mode: Whether progress is in discrete mode (deprecated, kept for compatibility)
+        title: Optional title for the plot (if None, auto-generated from loss/pearson)
+        loss: Optional loss value to display in title
+        pearson: Optional pearson correlation to display in title
+
+    Returns:
+        matplotlib Figure object
+    """
+    # Determine if we should show success plots
+    has_success_binary = success_binary is not None and len(success_binary) == len(progress_pred)
+
+    if has_success_binary:
+        # Three subplots: progress, success (binary), success_probs
+        fig, axs = plt.subplots(1, 3, figsize=(15, 3.5))
+        ax = axs[0]  # Progress subplot
+        ax2 = axs[1]  # Success subplot (binary)
+        ax3 = axs[2]  # Success probs subplot
+    else:
+        # Single subplot: progress only
+        fig, ax = plt.subplots(figsize=(6, 3.5))
+        ax2 = None
+        ax3 = None
+
+    # Plot progress
+    ax.plot(progress_pred, linewidth=2)
+    ax.set_ylabel("Progress")
+
+    # Build title
+    if title is None:
+        title_parts = ["Progress"]
+        if loss is not None:
+            title_parts.append(f"Loss: {loss:.3f}")
+        if pearson is not None:
+            title_parts.append(f"Pearson: {pearson:.2f}")
+        title = ", ".join(title_parts)
+    fig.suptitle(title)
+
+    # Set y-limits and ticks (always continuous since discrete is converted before this function)
+    ax.set_ylim(0, 1)
+    ax.spines["right"].set_visible(False)
+    ax.spines["top"].set_visible(False)
+    y_ticks = [0, 0.2, 0.4, 0.6, 0.8, 1.0]
+    ax.set_yticks(y_ticks)
+
+    # Setup success binary subplot
+    if ax2 is not None:
+        ax2.step(range(len(success_binary)), success_binary, where="post", linewidth=2, label="Predicted", color="blue")
+        # Add ground truth success labels as green line if available
+        if success_labels is not None and len(success_labels) == len(success_binary):
+            ax2.step(
+                range(len(success_labels)),
+                success_labels,
+                where="post",
+                linewidth=2,
+                label="Ground Truth",
+                color="green",
+            )
+        ax2.set_ylabel("Success (Binary)")
+        ax2.set_ylim(-0.05, 1.05)
+        ax2.spines["right"].set_visible(False)
+        ax2.spines["top"].set_visible(False)
+        ax2.set_yticks([0, 1])
+        ax2.legend()
+
+    # Setup success probs subplot if available
+    if ax3 is not None and success_probs is not None:
+        ax3.plot(range(len(success_probs)), success_probs, linewidth=2, label="Success Prob", color="purple")
+        # Add ground truth success labels as green line if available
+        if success_labels is not None and len(success_labels) == len(success_probs):
+            ax3.step(
+                range(len(success_labels)),
+                success_labels,
+                where="post",
+                linewidth=2,
+                label="Ground Truth",
+                color="green",
+                linestyle="--",
+            )
+        ax3.set_ylabel("Success Probability")
+        ax3.set_ylim(-0.05, 1.05)
+        ax3.spines["right"].set_visible(False)
+        ax3.spines["top"].set_visible(False)
+        ax3.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0])
+        ax3.legend()
+
+    plt.tight_layout()
+    return fig
+
+
+def extract_frames(video_path: str, fps: float = 1.0) -> np.ndarray:
+    """Extract frames from video file as numpy array (T, H, W, C).
+
+    Supports both local file paths and URLs (e.g., HuggingFace Hub URLs).
+    Uses the provided ``fps`` to control how densely frames are sampled from
+    the underlying video; there is no additional hard cap on the number of frames.
+
+    Args:
+        video_path: Path to video file or URL
+        fps: Frames per second to extract (default: 1.0)
+
+    Returns:
+        numpy array of shape (T, H, W, C) containing extracted frames, or None if error
+    """
+    if video_path is None:
+        raise ValueError("video_path is None")
+
+    if isinstance(video_path, tuple):
+        video_path = video_path[0]
+
+    # Check if it's a URL or local file
+    is_url = video_path.startswith(("http://", "https://"))
+    is_local_file = os.path.exists(video_path) if not is_url else False
+
+    if not is_url and not is_local_file:
+        raise FileNotFoundError(video_path)
+
+    try:
+        import decord  # type: ignore
+
+        # decord.VideoReader can handle both local files and URLs
+        vr = decord.VideoReader(video_path, num_threads=1)
+        total_frames = len(vr)
+
+        # Determine native FPS; fall back to a reasonable default if unavailable
+        try:
+            native_fps = float(vr.get_avg_fps())
+        except Exception:
+            native_fps = 1.0
+
+        # If user-specified fps is invalid or None, default to native fps
+        if fps is None or fps <= 0:
+            fps = native_fps
+
+        # Compute how many frames we want based on desired fps
+        # num_frames ≈ total_duration * fps = total_frames * (fps / native_fps)
+        if native_fps > 0:
+            desired_frames = int(round(total_frames * (fps / native_fps)))
+        else:
+            desired_frames = total_frames
+
+        # Clamp to [1, total_frames]
+        desired_frames = max(1, min(desired_frames, total_frames))
+
+        # Evenly sample indices to match the desired number of frames
+        if desired_frames == total_frames:
+            frame_indices = list(range(total_frames))
+        else:
+            frame_indices = np.linspace(0, total_frames - 1, desired_frames, dtype=int).tolist()
+
+        frames_array = vr.get_batch(frame_indices).asnumpy()  # Shape: (T, H, W, C)
+        del vr
+        return frames_array
+    except Exception as e:
+        raise RuntimeError(f"Error extracting frames from {video_path}: {e}")
+
+
+def load_frames_input(
+    video_or_array_path: str,
+    *,
+    fps: float = 1.0,
+) -> np.ndarray:
+    """
+    Load frames from a video file (path or URL) or from a .npy/.npz file.
+
+    Video: uses decord; fps controls sampling density.
+    .npy/.npz: expects array shape (T, H, W, C) or (T, C, H, W); for .npz uses
+    key 'arr_0' or the first array.
+
+    Returns:
+        Frames as uint8 array (T, H, W, C). Raises on failure.
+    """
+    if video_or_array_path.endswith(".npy"):
+        frames_array = np.load(video_or_array_path)
+    elif video_or_array_path.endswith(".npz"):
+        with np.load(video_or_array_path, allow_pickle=False) as npz:
+            if "arr_0" in npz:
+                frames_array = npz["arr_0"].copy()
+            else:
+                frames_array = next(iter(npz.values())).copy()
+    else:
+        frames_array = extract_frames(video_or_array_path, fps=fps)
+
+    if frames_array is None or frames_array.size == 0:
+        raise RuntimeError("Could not extract frames from input.")
+
+    if frames_array.dtype != np.uint8:
+        frames_array = np.clip(frames_array, 0, 255).astype(np.uint8)
+    if frames_array.ndim == 4:
+        if frames_array.shape[1] in (1, 3) and frames_array.shape[-1] not in (1, 3):
+            frames_array = frames_array.transpose(0, 2, 3, 1)
+
+    return frames_array
+
+
+def _numpy_to_npy_file_tuple(arr: np.ndarray, filename: str) -> Tuple[str, io.BytesIO, str]:
+    buf = io.BytesIO()
+    np.save(buf, arr)
+    buf.seek(0)
+    return (filename, buf, "application/octet-stream")
+
+
+def build_multipart_payload(samples: List[Dict[str, Any]]) -> Tuple[Dict[str, Any], Dict[str, str]]:
+    """
+    Convert a list of sample dicts into:
+      - files: mapping for requests.post(files=...)
+      - data: mapping for requests.post(data=...) with sample_{i} JSON strings
+
+    Numpy arrays inside trajectory fields are moved to .npy blobs and replaced by
+    {"__numpy_file__": <file_key>} references.
+    """
+    files: Dict[str, Any] = {}
+    data: Dict[str, str] = {}
+
+    numpy_fields = ["frames", "lang_vector", "video_embeddings"]
+
+    for i, sample in enumerate(samples):
+        sample_copy = json.loads(json.dumps(sample, default=str))  # make JSON-serializable shell
+        traj = sample.get("trajectory", {})
+        traj_copy = sample_copy.get("trajectory", {})
+
+        for field in numpy_fields:
+            val = traj.get(field, None)
+            if val is None:
+                continue
+
+            # torch.Tensor -> numpy (if torch is available)
+            if hasattr(val, "detach") and hasattr(val, "cpu"):
+                val = val.detach().cpu().numpy()
+
+            if isinstance(val, np.ndarray):
+                file_key = f"sample_{i}_trajectory_{field}"
+                files[file_key] = _numpy_to_npy_file_tuple(val, f"{file_key}.npy")
+                traj_copy[field] = {"__numpy_file__": file_key}
+            else:
+                traj_copy[field] = val
+
+        # Keep a helpful frames_shape (as list of ints) if present
+        if "frames_shape" in traj_copy and isinstance(traj_copy["frames_shape"], (tuple, list)):
+            traj_copy["frames_shape"] = [int(x) for x in traj_copy["frames_shape"]]
+
+        sample_copy["trajectory"] = traj_copy
+        data[f"sample_{i}"] = json.dumps(sample_copy)
+
+    return files, data
+
+
+def post_evaluate_batch_npy(
+    eval_server_url: str,
+    samples: List[Dict[str, Any]],
+    timeout_s: float = 120.0,
+    use_frame_steps: bool = False,
+) -> Dict[str, Any]:
+    files, data = build_multipart_payload(samples)
+    data["use_frame_steps"] = "true" if use_frame_steps else "false"
+    url = eval_server_url.rstrip("/") + "/evaluate_batch_npy"
+    resp = requests.post(url, files=files, data=data, timeout=timeout_s)
+    resp.raise_for_status()
+    return resp.json()
+
+
+def extract_rewards_from_server_output(outputs: Dict[str, Any]) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Parse server JSON into per-frame progress and success probability arrays.
+
+    Returns:
+        progress_array: Per-frame progress (reward) predictions for the first sample.
+        success_array: Per-frame success probabilities, or empty array if not in response.
+    """
+    outputs_progress = outputs.get("outputs_progress")
+    if outputs_progress is None:
+        raise ValueError("No `outputs_progress` in server response")
+    progress_pred = outputs_progress.get("progress_pred", [])
+
+    if progress_pred and len(progress_pred) > 0:
+        progress_array = np.array(progress_pred[0], dtype=np.float32)
+    else:
+        progress_array = np.array([], dtype=np.float32)
+
+    outputs_success = outputs.get("outputs_success", {})
+    success_probs = outputs_success.get("success_probs", []) if outputs_success else []
+    if success_probs and len(success_probs) > 0:
+        success_array = np.array(success_probs[0], dtype=np.float32)
+    else:
+        success_array = np.array([], dtype=np.float32)
+
+    return progress_array, success_array
+
+
+def make_progress_sample(
+    frames: np.ndarray,
+    task: str,
+    sample_id: str,
+    subsequence_length: int,
+) -> Dict[str, Any]:
+    return {
+        "sample_type": "progress",
+        "trajectory": {
+            "frames": frames,
+            "frames_shape": tuple(frames.shape),
+            "task": task,
+            "id": sample_id,
+            "metadata": {"subsequence_length": int(subsequence_length)},
+            "video_embeddings": None,
+        },
+    }
+
+
+def compute_rewards_per_frame(
+    eval_server_url: str,
+    video_frames: np.ndarray,
+    task: str,
+    timeout_s: float = 120.0,
+    use_frame_steps: bool = False,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Send the full trajectory to the eval server and get per-frame progress and success.
+
+    Args:
+        use_frame_steps: If True, server expands into frame-step sub-samples (0:1, 0:2, ...)
+            and aggregates; can improve alignment with training. If False, one forward pass on
+            the full trajectory (subsampled to fixed frames on the server).
+
+    Returns:
+        progress: Per-frame progress (reward) predictions.
+        success_probs: Per-frame success probabilities (empty if model has no success head).
+    """
+    T = int(video_frames.shape[0])
+    sample = make_progress_sample(
+        frames=video_frames,
+        task=task,
+        sample_id="0",
+        subsequence_length=T,
+    )
+    outputs = post_evaluate_batch_npy(
+        eval_server_url, [sample], timeout_s=timeout_s, use_frame_steps=use_frame_steps
+    )
+    return extract_rewards_from_server_output(outputs)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Get per-frame progress and success predictions from an RBM eval server.",
+        epilog="Start the server with: uv run python robometer/evals/eval_server.py --config_path=robometer/configs/config.yaml --host=0.0.0.0 --port=8000",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--eval-server-url",
+        type=str,
+        default="http://localhost:8000",
+        help="Eval server base URL (default: http://localhost:8000)",
+    )
+    parser.add_argument(
+        "--video",
+        type=str,
+        required=True,
+        help="Path or URL to a video, or a .npy/.npz file with frames (T,H,W,C) or (T,C,H,W)",
+    )
+    parser.add_argument("--task", type=str, required=True, help="Task instruction describing the trajectory")
+    parser.add_argument(
+        "--fps",
+        type=float,
+        default=1.0,
+        help="Frames per second when sampling from video (default: 1.0)",
+    )
+    parser.add_argument("--timeout-s", type=float, default=120.0, help="HTTP request timeout in seconds (default: 120)")
+    parser.add_argument(
+        "--use-frame-steps",
+        action="store_true",
+        help="If set, server uses frame-step expansion (0:1, 0:2, ...) and aggregates; can improve reward alignment",
+    )
+    parser.add_argument(
+        "--success-threshold",
+        type=float,
+        default=0.5,
+        help="Threshold for binary success curve in the plot (default: 0.5)",
+    )
+    parser.add_argument(
+        "--out",
+        type=str,
+        default=None,
+        help="Output path for rewards .npy (default: <video_stem>_rewards.npy)",
+    )
+    args = parser.parse_args()
+
+    video_path = Path(args.video)
+    out_path = Path(args.out) if args.out is not None else video_path.with_name(video_path.stem + "_rewards.npy")
+
+    frames = load_frames_input(str(args.video), fps=float(args.fps))
+
+    rewards, success_probs = compute_rewards_per_frame(
+        eval_server_url=args.eval_server_url,
+        video_frames=frames,
+        task=args.task,
+        timeout_s=float(args.timeout_s),
+        use_frame_steps=args.use_frame_steps,
+    )
+
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    np.save(str(out_path), rewards)
+    success_path = out_path.with_name(out_path.stem + "_success_probs.npy")
+    np.save(str(success_path), success_probs)
+
+    show_success = success_probs.size > 0 and success_probs.size == rewards.size
+    success_binary = (success_probs > float(args.success_threshold)).astype(np.int32) if show_success else None
+    fig = create_combined_progress_success_plot(
+        progress_pred=rewards,
+        num_frames=int(frames.shape[0]),
+        success_binary=success_binary,
+        success_probs=success_probs if show_success else None,
+        success_labels=None,
+        title=f"Progress/Success — {video_path.name}",
+    )
+    plot_path = out_path.with_name(out_path.stem + "_progress_success.png")
+    fig.savefig(str(plot_path), dpi=200)
+    plt.close(fig)
+
+    summary = {
+        "video": str(video_path),
+        "num_frames": int(frames.shape[0]),
+        "out_npy": str(out_path),
+        "out_success_probs_npy": str(success_path),
+        "out_plot_png": str(plot_path),
+        "reward_min": float(np.min(rewards)) if rewards.size else None,
+        "reward_max": float(np.max(rewards)) if rewards.size else None,
+        "reward_mean": float(np.mean(rewards)) if rewards.size else None,
+    }
+    print(json.dumps(summary, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/example_inference_local.py b/robometer/scripts/example_inference_local.py
new file mode 100644
index 0000000000000000000000000000000000000000..5b184551b684a91578189fce920153fdbf54b262
--- /dev/null
+++ b/robometer/scripts/example_inference_local.py
@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+"""
+Run RBM inference locally: load a checkpoint from HuggingFace and compute per-frame progress
+and success for a video (or .npy/.npz frames) and task instruction. Writes rewards .npy,
+success-probs .npy, and a progress/success plot. Requires the robometer package.
+
+Example:
+  python scripts/example_inference_local.py \\
+    --model-path aliangdw/qwen4b_pref_prog_succ_8_frames_all_part2 \\
+    --video /path/to/video.mp4 \\
+    --task "Pick up the red block and place it in the bin"
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+from benchmark_progress_mark_local import (
+    DEFAULT_INFERENCE_MODE,
+    DEFAULT_MAX_FRAMES,
+    DEFAULT_MIN_FRAMES,
+    DEFAULT_PREFIX_BATCH_SIZE,
+    DEFAULT_PREFIX_SAMPLE_FRAMES,
+    RobometerLocalRunner,
+    build_frame_retry_schedule,
+    is_cuda_oom_error,
+    load_all_video_frames,
+    sample_video_frames_with_indices,
+)
+from robometer.evals.eval_viz_utils import create_combined_progress_success_plot
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Run RBM inference locally: load model from HuggingFace and compute per-frame progress and success.",
+        epilog="Outputs: <out>.npy (rewards), <out>_success_probs.npy, <out>_progress_success.png",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument("--model-path", default="../../models/Robometer-4B", help="HuggingFace model id or local checkpoint path")
+    parser.add_argument("--video", default="example_videos/soar_put_green_stick_in_brown_bowl.mp4", help="Video path/URL or .npy/.npz with frames (T,H,W,C)")
+    parser.add_argument("--task", default="Put green stick in brown bowl", help="Task instruction for the trajectory")
+    parser.add_argument("--fps", type=float, default=1.0, help="FPS when sampling from video (default: 1.0)")
+    parser.add_argument("--max-frames", type=int, default=DEFAULT_MAX_FRAMES, help="Max frames to extract from video (default: 128)")
+    parser.add_argument("--min-frames", type=int, default=DEFAULT_MIN_FRAMES, help="Minimum retry frame budget after OOM (default: 32)")
+    parser.add_argument(
+        "--inference-mode",
+        choices=["frame_steps", "whole"],
+        default=DEFAULT_INFERENCE_MODE,
+        help="frame_steps matches benchmark behavior; whole does a single forward pass on the full sampled trajectory",
+    )
+    parser.add_argument("--prefix-sample-frames", type=int, default=DEFAULT_PREFIX_SAMPLE_FRAMES, help="Frames per prefix in frame_steps mode (default: 4)")
+    parser.add_argument("--prefix-batch-size", type=int, default=DEFAULT_PREFIX_BATCH_SIZE, help="Batch size for prefix inference in frame_steps mode (default: 1)")
+    parser.add_argument(
+        "--adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_true",
+        help="On CUDA OOM in whole mode, retry with a smaller frame budget",
+    )
+    parser.add_argument(
+        "--no-adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_false",
+        help="Disable frame-budget retry on CUDA OOM",
+    )
+    parser.add_argument(
+        "--success-threshold",
+        type=float,
+        default=0.5,
+        help="Threshold for binary success in plot (default: 0.5)",
+    )
+    parser.add_argument("--out", default=None, help="Output path for rewards .npy (default: <video_stem>_rewards.npy)")
+    parser.set_defaults(adaptive_max_frames=True)
+    args = parser.parse_args()
+
+    video_path = Path(args.video)
+    
+    # Create output directory: scripts/outputs/{video_name}/
+    if args.out is not None:
+        out_path = Path(args.out)
+    else:
+        output_dir = Path(__file__).parent / "outputs" / video_path.stem
+        output_dir.mkdir(parents=True, exist_ok=True)
+        out_path = output_dir / f"{video_path.stem}_rewards.npy"
+
+    runner = RobometerLocalRunner(model_path=args.model_path)
+
+    all_frames, native_fps = load_all_video_frames(video_path)
+    retry_schedule = (
+        build_frame_retry_schedule(args.max_frames, args.min_frames, bool(args.adaptive_max_frames))
+        if args.inference_mode == "whole"
+        else [int(args.max_frames)]
+    )
+
+    frames = None
+    rewards = None
+    success_probs = None
+    used_max_frames = retry_schedule[0]
+
+    for attempt_idx, frame_budget in enumerate(retry_schedule, start=1):
+        frames, _ = sample_video_frames_with_indices(
+            all_frames,
+            native_fps=native_fps,
+            fps=float(args.fps),
+            max_frames=int(frame_budget),
+            required_frames=[],
+        )
+        print(
+            f"Loaded {len(all_frames)} total frames; sampled {len(frames)} frames at fps={float(args.fps):g} "
+            f"(max_frames={int(frame_budget)}, try {attempt_idx}/{len(retry_schedule)})"
+        )
+        try:
+            rewards, success_probs = runner.compute_rewards_per_frame(
+                video_frames=frames,
+                task=args.task,
+                inference_mode=args.inference_mode,
+                prefix_sample_frames=int(args.prefix_sample_frames),
+                prefix_batch_size=int(args.prefix_batch_size),
+            )
+            used_max_frames = int(frame_budget)
+            break
+        except RuntimeError as exc:
+            if args.inference_mode != "whole" or not is_cuda_oom_error(exc) or attempt_idx == len(retry_schedule):
+                raise
+            next_budget = retry_schedule[attempt_idx]
+            print(
+                f"[OOM] whole inference hit CUDA OOM at max_frames={int(frame_budget)}; "
+                f"retrying with max_frames={int(next_budget)}"
+            )
+            runner.reload_model()
+
+    if rewards is None or success_probs is None or frames is None:
+        raise RuntimeError("Robometer inference did not produce outputs.")
+
+    # Save results (directory already created)
+    np.save(str(out_path), rewards)
+    success_path = out_path.with_name(out_path.stem + "_success_probs.npy")
+    np.save(str(success_path), success_probs)
+
+    show_success = success_probs.size > 0 and success_probs.size == rewards.size
+    success_binary = (success_probs > float(args.success_threshold)).astype(np.int32) if show_success else None
+    fig = create_combined_progress_success_plot(
+        progress_pred=rewards,
+        num_frames=int(frames.shape[0]),
+        success_binary=success_binary,
+        success_probs=success_probs if show_success else None,
+        success_labels=None,
+        title=f"Progress/Success — {video_path.name}",
+    )
+    plot_path = out_path.with_name(out_path.stem + "_progress_success.png")
+    fig.savefig(str(plot_path), dpi=200)
+    plt.close(fig)
+
+    summary = {
+        "video": str(video_path),
+        "num_frames": int(frames.shape[0]),
+        "inference_mode": args.inference_mode,
+        "max_frames_used": int(used_max_frames),
+        "model_path": args.model_path,
+        "out_rewards": str(out_path),
+        "out_success_probs": str(success_path),
+        "out_plot": str(plot_path),
+        "reward_min": float(np.min(rewards)) if rewards.size else None,
+        "reward_max": float(np.max(rewards)) if rewards.size else None,
+        "reward_mean": float(np.mean(rewards)) if rewards.size else None,
+    }
+    print(json.dumps(summary, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/example_libero_robometer_wrapper.py b/robometer/scripts/example_libero_robometer_wrapper.py
new file mode 100644
index 0000000000000000000000000000000000000000..fc248981c5ccbef903eacd9495b8b6d0106c5884
--- /dev/null
+++ b/robometer/scripts/example_libero_robometer_wrapper.py
@@ -0,0 +1,736 @@
+"""
+Example reward-model wrapper for LIBERO environments.
+"""
+
+from __future__ import annotations
+
+from collections import deque
+from dataclasses import dataclass
+from typing import Any, Deque, Dict, List, Optional, Sequence, Tuple, Union
+import os
+import argparse
+import sys
+sys.path.append(os.path.join(os.path.dirname(__file__), "..", "LIBERO"))
+
+import numpy as np
+import gymnasium as gym
+import gymnasium.vector as gym_vector
+
+from robometer.evals.eval_utils import raw_dict_to_sample, extract_rewards_from_output, extract_success_probs_from_output
+from robometer.evals.eval_server import process_batch_helper
+from robometer.utils.setup_utils import setup_batch_collator
+from robometer.utils.tensor_utils import t2n
+from robometer.utils.save import load_model_from_hf
+
+class GymToGymnasiumWrapper(gym.Env):
+    """
+    A wrapper to convert a classic Gym environment to a Gymnasium-like interface.
+    It adapts `reset()` and `step()` signatures, handles info dict changes, and supports compatibility.
+    """
+
+    def __init__(self, env, time_limit: int = None):
+        super().__init__()  # make sure Env is initialized
+        self.env = env
+        # Action space remains the same
+        if hasattr(self.env, "action_space"):
+            self.action_space = self.env.action_space
+        if hasattr(self.env, "observation_space"):
+            self.observation_space = self.env.observation_space
+        self.reward_range = getattr(env, "reward_range", None)
+        self.metadata = getattr(env, "metadata", {})
+        self.time_limit = time_limit
+        self.current_step = 0
+
+    def reset(self, *, seed=None, options=None):
+        # Reset step counter
+        self.current_step = 0
+        # Gym reset sometimes does not support 'seed' or 'options'
+        if seed is not None:
+            try:
+                obs = self.env.reset(seed=seed)
+            except TypeError:
+                self.env.seed(seed)
+                obs = self.env.reset()
+        else:
+            obs = self.env.reset()
+        info = {}
+        if isinstance(obs, tuple) and len(obs) == 2:
+            obs, info = obs
+        return obs, info
+
+    def step(self, action):
+        result = self.env.step(action)
+        self.current_step += 1
+        if len(result) == 4:
+            obs, reward, done, info = result
+            terminated = done
+            # Gymnasium expects terminated, truncated
+            if self.time_limit is not None and self.current_step >= self.time_limit:
+                truncated = True
+            else:
+                truncated = info.get("TimeLimit.truncated", False)
+            return obs, reward, terminated, truncated, info
+        elif len(result) == 5:
+            # Already modern API
+            return result
+        else:
+            raise ValueError("Unexpected number of outputs from env.step")
+
+    def render(self, *args, **kwargs):
+        return self.env.render(*args, **kwargs)
+
+    def close(self):
+        return self.env.close()
+
+    def __getattr__(self, name):
+        # Forward other attributes/methods to original env
+        return getattr(self.env, name)
+
+
+
+@dataclass
+class RewardModelStepOutput:
+    # Kept for potential downstream debugging/typing; not required by wrappers.
+    reward: float
+    success_prob: float
+    per_key_rewards: Dict[str, float]
+    per_key_success_probs: Dict[str, float]
+
+
+class _RewardModelInferenceMixin:
+    """
+    Shared reward-model inference logic with local model.
+    """
+
+    def __init__(
+        self,
+        model_path: str,
+        device: str,
+        max_frames: Optional[int] = None,
+    ):
+        if model_path is not None:
+            reward_model_config, tokenizer, processor, reward_model = load_model_from_hf(
+                model_path=model_path,
+                device=device,
+            )
+            reward_model.eval()
+
+            self.reward_model = reward_model
+            self.reward_model_config = reward_model_config
+
+        if self.reward_model is None:
+            raise ValueError("reward_model must be provided")
+
+        # Determine max_frames
+        if max_frames is not None:
+            self.max_frames = int(max_frames)
+        elif self.reward_model_config is not None:
+            self.max_frames = int(getattr(getattr(self.reward_model_config, "data", None), "max_frames", 16))
+        else:
+            self.max_frames = 16
+
+        # Local model path: set up collator once
+        self.processor = None
+        self.tokenizer = None
+        self.batch_collator = None
+        self._model_device = None
+        self._model_type = None
+        if self.reward_model is not None:
+            self.processor = getattr(self.reward_model, "processor", None)
+            self.tokenizer = getattr(self.reward_model, "tokenizer", None)
+            if self.processor is None or self.tokenizer is None:
+                raise ValueError(
+                    "processor and tokenizer must be available on reward_model "
+                    "(reward_model.processor / reward_model.tokenizer)"
+                )
+
+            # Ensure multi-image behavior is enabled (matches SPUR buffer)
+            if self.reward_model_config is not None:
+                data_cfg = getattr(self.reward_model_config, "data", None)
+                if data_cfg is not None and hasattr(data_cfg, "use_multi_image") and not data_cfg.use_multi_image:
+                    data_cfg.use_multi_image = True
+
+            # Resolve model type/device once
+            self._model_type = getattr(getattr(self.reward_model_config, "model", None), "model_type", None)
+            if self._model_type is None:
+                raise ValueError("reward_model_config.model.model_type is required for local reward inference")
+            self._model_device = getattr(self.reward_model, "device", None)
+            if self._model_device is None:
+                try:
+                    import torch
+
+                    self._model_device = next(self.reward_model.parameters()).device
+                    if isinstance(self._model_device, torch.device):
+                        self._model_device = str(self._model_device)
+                except Exception:
+                    self._model_device = None
+
+            self.batch_collator = setup_batch_collator(
+                self.processor, self.tokenizer, self.reward_model_config, is_eval=True
+            )
+
+    def _compute_rewards_batch(
+        self, batch_raw: List[Dict[str, Any]]
+    ) -> Tuple[List[float], List[float]]:
+        """
+        Returns lists: (progress_rewards, success_probs).
+        """
+        if len(batch_raw) == 0:
+            return [], []
+
+        samples = [
+            raw_dict_to_sample(raw_data=raw, max_frames=self.max_frames, sample_type="progress")
+            for raw in batch_raw
+        ]
+
+        is_discrete_mode = (
+            self.reward_model_config is not None
+            and getattr(getattr(self.reward_model_config, "loss", None), "progress_loss_type", None) == "discrete"
+        )
+        num_bins = (
+            getattr(getattr(self.reward_model_config, "loss", None), "progress_discrete_bins", None)
+            if self.reward_model_config is not None
+            else None
+        )
+        outputs = process_batch_helper(
+            model_type=self._model_type,
+            model=self.reward_model,
+            tokenizer=self.tokenizer,
+            batch_collator=self.batch_collator,
+            device=self._model_device,
+            batch_data=[s.model_dump() for s in samples],
+            job_id=0,
+            is_discrete_mode=bool(is_discrete_mode),
+            num_bins=num_bins,
+        )
+        rewards = extract_rewards_from_output(outputs)
+        success_probs = extract_success_probs_from_output(outputs)
+        return rewards.tolist(), success_probs.tolist()
+
+
+class LiberoRobometerRewardWrapper(gym.Wrapper, _RewardModelInferenceMixin):
+    """
+    Non-vector LIBERO wrapper that replaces rewards with reward-model predictions.
+    """
+
+    def __init__(
+        self,
+        env,
+        model_path: str,
+        device: str,
+        reward_relabeling_keys: Sequence[str],
+        *,
+        use_relative_rewards: bool = False,
+        add_estimated_reward: bool = False,
+        use_success_detection: bool = False,
+        success_detection_duration: int = 2,
+        success_detection_threshold: float = 0.65,
+        max_frames: Optional[int] = None,
+    ):
+        self.env = GymToGymnasiumWrapper(env, time_limit=400)
+        gym.Wrapper.__init__(self, self.env)
+        _RewardModelInferenceMixin.__init__(
+            self,
+            model_path=model_path,
+            device=device,
+            max_frames=max_frames,
+        )
+
+        self.reward_relabeling_keys = list(reward_relabeling_keys)
+        if len(self.reward_relabeling_keys) == 0:
+            raise ValueError("reward_relabeling_keys must be non-empty")
+        
+        # Action space remains the same
+        if not hasattr(self.env, "action_space"):
+            self.action_space = gym.spaces.Box(low=-1.0, high=1.0, shape=(7,), dtype=np.float32)    
+        else:
+            self.action_space = self.env.action_space
+
+        self.use_relative_rewards = bool(use_relative_rewards)
+        self.add_estimated_reward = bool(add_estimated_reward)
+        self.use_success_detection = bool(use_success_detection)
+        self.success_detection_duration = int(success_detection_duration)
+        self.success_detection_threshold = float(success_detection_threshold)
+
+        self._frames: Dict[str, Deque[np.ndarray]] = {}
+        self.language_instruction = self.env.language_instruction
+        self.episode_id = 0
+        self._step_in_episode: int = 0
+        self._prev_reward: float = 0.0
+        self._success_window: Deque[float] = deque(maxlen=self.success_detection_duration)
+
+    def _get_language_instruction(self, obs: Dict[str, Any], info: Dict[str, Any]) -> Optional[str]:
+        if isinstance(info, dict) and "language_instruction" in info:
+            return info.get("language_instruction")
+        if isinstance(obs, dict) and isinstance(obs.get("prompt"), str):
+            return obs.get("prompt")
+        return self.language_instruction
+
+    def reset(self, **kwargs):
+        obs, info = self.env.reset(**kwargs)
+        self.language_instruction = self.env.language_instruction
+        self.episode_id += 1
+
+        self._frames = {k: [] for k in self.reward_relabeling_keys}
+        self._step_in_episode = 0
+        self._prev_reward = 0.0
+        self._success_window = deque(maxlen=self.success_detection_duration)
+
+        if isinstance(obs, dict):
+            for k in self.reward_relabeling_keys:
+                if k in obs:
+                    self._frames[k].append(t2n(obs[k]))
+        return obs, info
+
+    def step(self, action):
+        obs, env_reward, terminated, truncated, info = self.env.step(action)
+
+        if not isinstance(info, dict):
+            info = {} if info is None else dict(info)
+        # In LIBERO, done is only True when task succeeds, so success = done
+        # But don't overwrite if already present in info
+        if "success" not in info:
+            info["success"] = terminated
+        if terminated:
+            assert env_reward == 1.0, "Reward should be 1.0 when task succeeds"
+        
+        env_reward -= 1  # reward is -1, 0
+
+        if isinstance(obs, dict):
+            for k in self.reward_relabeling_keys:
+                if k in obs:
+                    self._frames[k].append(t2n(obs[k]))
+
+        # Prepare per-key inputs for this timestep
+        per_key_rewards: Dict[str, float] = {}
+        per_key_success: Dict[str, float] = {}
+
+        for key_idx, key in enumerate(self.reward_relabeling_keys):
+            frames = np.stack(list(self._frames[key]), axis=0) if len(self._frames[key]) > 0 else np.array([])
+            raw = dict(
+                frames=frames,
+                task=self.language_instruction,
+                id=self.episode_id,
+                metadata=dict(
+                        subsequence_length=len(self._frames[key]) if self._frames[key] is not None else 0,
+                ),
+                video_embeddings=None,
+                text_embedding=None,
+            )
+            rewards, success_probs = self._compute_rewards_batch([raw])
+            per_key_rewards[key] = rewards[0]
+            per_key_success[key] = success_probs[0]
+
+        pred_reward = np.mean(list(per_key_rewards.values())) if per_key_rewards else 0.0
+        success_prob = np.mean(list(per_key_success.values())) if per_key_success else 0.0
+
+        # Relative reward option
+        if self.use_relative_rewards:
+            current = pred_reward
+            pred_reward = pred_reward - self._prev_reward
+            self._prev_reward = current
+            if terminated or truncated:
+                self._prev_reward = 0.0
+
+        # Success detection option
+        if self.use_success_detection:
+            self._success_window.append(success_prob)
+            if len(self._success_window) == self.success_detection_duration:
+                votes = sum(1 for p in self._success_window if p >= self.success_detection_threshold)
+                if votes > (self.success_detection_duration / 2):
+                    terminated = True
+                    info["success"] = True
+                    info["success_from_reward_model"] = True
+
+        # Decide what reward to return
+        if self.add_estimated_reward:
+            out_reward = env_reward + pred_reward
+        else:
+            out_reward = pred_reward
+
+        info["env_reward"] = env_reward
+        info["predicted_reward"] = pred_reward
+        info["success_prob"] = success_prob
+        info["predicted_rewards_by_key"] = per_key_rewards
+        info["success_probs_by_key"] = per_key_success
+        info["step_in_episode"] = int(self._step_in_episode)
+
+        self._step_in_episode += 1
+
+        # If the env auto-resets under the hood, start a fresh history when done/truncated.
+        if terminated or truncated:
+            self._frames = {k: [] for k in self.reward_relabeling_keys}
+            self.language_instruction = self.env.language_instruction
+            self._step_in_episode = 0
+            self._success_window = deque(maxlen=self.success_detection_duration)
+
+        return obs, out_reward, terminated, truncated, info
+
+
+class VectorLiberoRobometerRewardWrapper(gym_vector.VectorWrapper, _RewardModelInferenceMixin):
+    """
+    Vectorized LIBERO wrapper that replaces rewards with reward-model predictions per env.
+    """
+
+    def __init__(
+        self,
+        env: gym_vector.VectorEnv,
+        model_path: str,
+        device: str,
+        reward_relabeling_keys: Sequence[str],
+        *,
+        use_relative_rewards: bool = False,
+        add_estimated_reward: bool = False,
+        replace_reward: bool = True,
+        use_success_detection: bool = False,
+        success_detection_duration: int = 2,
+        success_detection_threshold: float = 0.65,
+        max_frames: Optional[int] = None,
+    ):
+        gym_vector.VectorWrapper.__init__(self, env)
+        _RewardModelInferenceMixin.__init__(
+            self,
+            model_path=model_path,
+            device=device,
+            max_frames=max_frames,
+        )
+
+        self.reward_relabeling_keys = list(reward_relabeling_keys)
+        if len(self.reward_relabeling_keys) == 0:
+            raise ValueError("reward_relabeling_keys must be non-empty")
+
+        self.use_relative_rewards = bool(use_relative_rewards)
+        self.add_estimated_reward = bool(add_estimated_reward)
+        self.replace_reward = bool(replace_reward)
+        self.use_success_detection = bool(use_success_detection)
+        self.success_detection_duration = int(success_detection_duration)
+        self.success_detection_threshold = float(success_detection_threshold)
+
+        self._n = int(getattr(self.env, "num_envs", 1))
+        self._frames: List[Dict[str, Deque[np.ndarray]]] = []
+        self._language_instructions: List[Optional[str]] = []
+        self._episode_ids: List[int] = []
+        self._step_in_episode: List[int] = []
+        self._prev_rewards: List[float] = []
+        self._success_windows: List[Deque[float]] = []
+
+        self._init_state()
+
+    def _init_state(self):
+        self._n = int(getattr(self.env, "num_envs", self._n))
+        self._frames = [
+            {k: deque(maxlen=self.max_frames) for k in self.reward_relabeling_keys} for _ in range(self._n)
+        ]
+        self._language_instructions = [None for _ in range(self._n)]
+        self._episode_ids = [0 for _ in range(self._n)]
+        self._step_in_episode = [0 for _ in range(self._n)]
+        self._prev_rewards = [0.0 for _ in range(self._n)]
+        self._success_windows = [deque(maxlen=self.success_detection_duration) for _ in range(self._n)]
+
+    def _get_language_instruction_vec(self, obs: Dict[str, Any], info: Any) -> List[Optional[str]]:
+        getter = getattr(self.env, "get_language_instruction", None)
+        if callable(getter):
+            try:
+                instr = getter()
+                if isinstance(instr, str):
+                    return [instr] * self._n
+            except Exception:
+                pass
+
+        # Try prompt in obs
+        if isinstance(obs, dict) and "prompt" in obs:
+            p = obs["prompt"]
+            if isinstance(p, list) and len(p) == self._n:
+                return [str(x) for x in p]
+            if isinstance(p, np.ndarray) and p.shape[0] == self._n:
+                return [str(x) for x in p.tolist()]
+        # Fallback: single instruction attribute (shared across envs)
+        shared = getattr(self.env, "language_instruction", None)
+        return [shared] * self._n
+
+    def reset(self, **kwargs):
+        obs, info = self.env.reset(**kwargs)
+        self._init_state()
+
+        if isinstance(obs, dict):
+            instrs = self._get_language_instruction_vec(obs, info)
+            for i in range(self._n):
+                self._language_instructions[i] = instrs[i]
+                self._episode_ids[i] += 1
+
+            for k in self.reward_relabeling_keys:
+                if k in obs:
+                    arr = t2n(obs[k])
+                    if arr is not None and arr.shape[0] == self._n:
+                        for i in range(self._n):
+                            self._frames[i][k].append(arr[i])
+
+        return obs, info
+
+    def step(self, actions):
+        obs, env_rewards, terminateds, truncateds, info = self.env.step(actions)
+
+        # Normalize arrays
+        env_rewards_np = t2n(env_rewards)
+        terminateds_np = t2n(terminateds).astype(bool)
+        truncateds_np = t2n(truncateds).astype(bool)
+
+        if env_rewards_np is None:
+            env_rewards_np = np.zeros((self._n,), dtype=np.float64)
+
+        # In LIBERO, done is only True when task succeeds, so success = terminated.
+        # Mirror the non-vector wrapper's reward shift (0/1 -> -1/0).
+        for i in range(self._n):
+            if bool(terminateds_np[i]):
+                assert float(env_rewards_np[i]) == 1.0, "Reward should be 1.0 when task succeeds"
+        env_rewards_shifted = env_rewards_np.astype(np.float64) - 1.0
+
+        # Gymnasium VectorEnv may auto-reset in the same step; if so, terminal obs is in info["final_observation"]
+        final_obs = None
+        if isinstance(info, dict) and "final_observation" in info:
+            final_obs = info.get("final_observation")
+
+        reset_instrs = self._get_language_instruction_vec(obs, info) if isinstance(obs, dict) else [None] * self._n
+        task_for_model: List[Optional[str]] = [
+            (self._language_instructions[i] if self._language_instructions[i] is not None else reset_instrs[i])
+            for i in range(self._n)
+        ]
+
+        # Update frame histories using the correct observation for this transition.
+        # If SAME_STEP autoreset is enabled, use terminal obs from final_observation when available.
+        if isinstance(obs, dict):
+            for k in self.reward_relabeling_keys:
+                if k not in obs:
+                    continue
+                arr_reset = t2n(obs[k])
+                if arr_reset is None or arr_reset.shape[0] != self._n:
+                    continue
+                for i in range(self._n):
+                    frame_i = arr_reset[i]
+                    if final_obs is not None and i < len(final_obs) and final_obs[i] is not None:
+                        fo_i = final_obs[i]
+                        if isinstance(fo_i, dict) and k in fo_i:
+                            frame_i = t2n(fo_i[k])
+                    self._frames[i][k].append(frame_i)
+
+        # Batch reward computation per key across envs
+        per_env_per_key_reward: Dict[str, List[float]] = {k: [0.0] * self._n for k in self.reward_relabeling_keys}
+        per_env_per_key_success: Dict[str, List[float]] = {k: [0.0] * self._n for k in self.reward_relabeling_keys}
+
+        for key_idx, key in enumerate(self.reward_relabeling_keys):
+            batch_raw: List[Dict[str, Any]] = []
+            for i in range(self._n):
+                frames = np.stack(list(self._frames[i][key]), axis=0) if len(self._frames[i][key]) > 0 else np.array([])
+                batch_raw.append(
+                    dict(
+                        frames=frames,
+                        task=task_for_model[i],
+                        id=int(self._episode_ids[i]),
+                        metadata=dict(subsequence_length=len(self._frames[i][key])),
+                        video_embeddings=None,
+                        text_embedding=None,
+                    )
+                )
+
+            rewards_k, success_k = self._compute_rewards_batch(batch_raw)
+            for i in range(self._n):
+                per_env_per_key_reward[key][i] = rewards_k[i] if i < len(rewards_k) else 0.0
+                per_env_per_key_success[key][i] = success_k[i] if i < len(success_k) else 0.0
+
+        # Aggregate across keys
+        pred_rewards_abs = np.zeros((self._n,), dtype=np.float64)
+        success_probs = np.zeros((self._n,), dtype=np.float64)
+        for i in range(self._n):
+            r_vals = [per_env_per_key_reward[k][i] for k in self.reward_relabeling_keys]
+            s_vals = [per_env_per_key_success[k][i] for k in self.reward_relabeling_keys]
+            pred_rewards_abs[i] = np.mean(r_vals) if len(r_vals) else 0.0
+            success_probs[i] = np.mean(s_vals) if len(s_vals) else 0.0
+
+        pred_rewards_out = pred_rewards_abs.copy()
+        if self.use_relative_rewards:
+            for i in range(self._n):
+                cur = float(pred_rewards_abs[i])
+                pred_rewards_out[i] = cur - self._prev_rewards[i]
+                self._prev_rewards[i] = cur
+                if terminateds_np[i] or truncateds_np[i]:
+                    self._prev_rewards[i] = 0.0
+
+        # Success detection
+        if self.use_success_detection:
+            for i in range(self._n):
+                self._success_windows[i].append(float(success_probs[i]))
+                if len(self._success_windows[i]) == self.success_detection_duration:
+                    votes = sum(1 for p in self._success_windows[i] if p >= self.success_detection_threshold)
+                    if votes > (self.success_detection_duration / 2):
+                        terminateds_np[i] = True
+
+        # Determine reward output
+        if self.add_estimated_reward:
+            out_rewards = env_rewards_shifted + pred_rewards_out
+        else:
+            out_rewards = env_rewards_shifted if not self.replace_reward else pred_rewards_out
+
+        # Inject info
+        # Gymnasium vector env `info` is typically a dict of arrays; keep it dict-like.
+        if info is None:
+            info = {}
+        if isinstance(info, dict):
+            info = dict(info)
+            if "success" not in info:
+                info["success"] = terminateds_np.copy()
+            info["env_reward"] = env_rewards_shifted.astype(np.float64)
+            info["predicted_reward"] = pred_rewards_out.astype(np.float64)
+            info["predicted_reward_abs"] = pred_rewards_abs.astype(np.float64)
+            info["success_prob"] = success_probs.astype(np.float64)
+            info["step_in_episode"] = np.asarray(self._step_in_episode, dtype=np.int32)
+        # Also provide per-key arrays
+        if isinstance(info, dict):
+            for k in self.reward_relabeling_keys:
+                info[f"predicted_reward/{k}"] = np.asarray(per_env_per_key_reward[k], dtype=np.float64)
+                info[f"success_prob/{k}"] = np.asarray(per_env_per_key_success[k], dtype=np.float64)
+
+        # Advance step counters and clear per-env state on episode end (to support auto-reset vector envs)
+        for i in range(self._n):
+            self._step_in_episode[i] += 1
+            if terminateds_np[i] or truncateds_np[i]:
+                self._frames[i] = {k: deque(maxlen=self.max_frames) for k in self.reward_relabeling_keys}
+                self._language_instructions[i] = reset_instrs[i]
+                self._step_in_episode[i] = 0
+                self._success_windows[i] = deque(maxlen=self.success_detection_duration)
+                self._episode_ids[i] += 1
+
+                # If SAME_STEP autoreset happened, seed next episode history with reset obs immediately.
+                if isinstance(obs, dict) and final_obs is not None and i < len(final_obs) and final_obs[i] is not None:
+                    for k in self.reward_relabeling_keys:
+                        if k not in obs:
+                            continue
+                        arr_reset = t2n(obs[k])
+                        if arr_reset is not None and arr_reset.shape[0] == self._n:
+                            self._frames[i][k].append(arr_reset[i])
+
+        return obs, out_rewards.astype(np.float64), terminateds_np, truncateds_np, info
+
+def main():
+    try:
+        from libero.libero.envs import OffScreenRenderEnv, DummyVectorEnv
+        from libero.libero import benchmark, get_libero_path
+    except ImportError:
+        print("LIBERO not found. Please install LIBERO.")
+        sys.exit(1)
+
+    parser = argparse.ArgumentParser(
+        description="Run RBM inference locally: load model from HuggingFace and compute per-frame progress and success.",
+        epilog="Outputs: <out>.npy (rewards), <out>_success_probs.npy, <out>_progress_success.png",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument("--model-path", default="aliangdw/Robometer-4B-LIBERO", help="HuggingFace model id or local checkpoint path")
+    parser.add_argument("--task-suite-name", default="libero_90", help="LIBERO task suite name")
+    parser.add_argument("--task-id", default=28, type=int, help="LIBERO task id")
+    parser.add_argument("--vectorized", action="store_true", help="Run in vectorized mode")
+    parser.add_argument("--num-envs", default=2, type=int, help="Number of environments to run in parallel")
+    args = parser.parse_args()
+
+    if not args.vectorized:
+        print("Testing Single LIBERO Robometer Reward Wrapper")
+        seed = np.random.randint(0, 1000000)
+        # Get task info
+        benchmark_dict = benchmark.get_benchmark_dict()
+        task_suite = benchmark_dict[args.task_suite_name]()
+        task = task_suite.get_task(args.task_id)
+
+        task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
+
+        env_args = {"bddl_file_name": task_bddl_file, "camera_heights": 256, "camera_widths": 256}
+        base_env = OffScreenRenderEnv(**env_args)
+        base_env.seed(seed)
+
+        robometer_libero_env = LiberoRobometerRewardWrapper(base_env, 
+                                                            model_path=args.model_path, 
+                                                            device="cuda", 
+                                                            reward_relabeling_keys=["agentview_image"], 
+                                                            add_estimated_reward=True, 
+                                                            )
+        obs, info = robometer_libero_env.reset()
+        for i in range(10):
+            action = np.random.uniform(-1, 1, 7)
+            obs, reward, terminated, truncated, info = robometer_libero_env.step(action)
+            print(f"Reward at step {i}: {reward}")
+        
+        robometer_libero_env.close()
+
+    else:
+        print("Testing Vectorized LIBERO Robometer Reward Wrapper")
+        def make_env():
+            seed = np.random.randint(0, 1000000)
+            # Get task info
+            benchmark_dict = benchmark.get_benchmark_dict()
+            task_suite = benchmark_dict[args.task_suite_name]()
+            task = task_suite.get_task(args.task_id)
+
+            task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
+
+            env_args = {"bddl_file_name": task_bddl_file, "camera_heights": 256, "camera_widths": 256}
+            base_env = OffScreenRenderEnv(**env_args)
+            base_env.seed(seed)
+            sample_obs = base_env.reset()
+            env = GymToGymnasiumWrapper(base_env, time_limit=400)
+            # Action space remains the same
+            if not hasattr(env, "action_space"):
+                env.action_space = gym.spaces.Box(low=-1.0, high=1.0, shape=(7,), dtype=np.float32)
+            if not hasattr(env, "observation_space"):
+                # Create observation space from sample_obs, which is a dict of arrays
+                obs_space_dict = {}
+                for k, v in sample_obs.items():
+                    # Common LIBERO obs dict may include images (uint8) and sometimes text prompts.
+                    if isinstance(v, (str, bytes, bytearray)) or (isinstance(v, np.ndarray) and v.dtype.kind in {"U", "S"}):
+                        obs_space_dict[k] = gym.spaces.Text(max_length=2048)
+                        continue
+                    v_arr = np.asarray(v)
+                    dt = v_arr.dtype
+                    if np.issubdtype(dt, np.uint8):
+                        # Images: bounded [0, 255]
+                        obs_space_dict[k] = gym.spaces.Box(
+                            low=np.zeros(v_arr.shape, dtype=np.uint8),
+                            high=np.full(v_arr.shape, 255, dtype=np.uint8),
+                            shape=v_arr.shape,
+                            dtype=np.uint8,
+                        )
+                    elif np.issubdtype(dt, np.integer):
+                        ii = np.iinfo(dt)
+                        obs_space_dict[k] = gym.spaces.Box(
+                            low=np.full(v_arr.shape, ii.min, dtype=dt),
+                            high=np.full(v_arr.shape, ii.max, dtype=dt),
+                            shape=v_arr.shape,
+                            dtype=dt,
+                        )
+                    else:
+                        # Floats/other numeric: unbounded
+                        obs_space_dict[k] = gym.spaces.Box(
+                            low=np.full(v_arr.shape, -np.inf, dtype=np.float32),
+                            high=np.full(v_arr.shape, np.inf, dtype=np.float32),
+                            shape=v_arr.shape,
+                            dtype=np.float32,
+                        )
+                env.observation_space = gym.spaces.Dict(obs_space_dict)
+
+            return env
+
+        env_fns = [make_env for _ in range(args.num_envs)]
+        env = gym.vector.SyncVectorEnv(env_fns)
+        robometer_libero_env = VectorLiberoRobometerRewardWrapper(env, 
+                                                                model_path=args.model_path, 
+                                                                device="cuda", 
+                                                                reward_relabeling_keys=["agentview_image"], 
+                                                                add_estimated_reward=True, 
+                                                                )
+        obs, info = robometer_libero_env.reset()
+        for i in range(10):
+            actions = np.random.uniform(-1, 1, (args.num_envs, 7))
+            obs, rewards, terminateds, truncateds, infos = robometer_libero_env.step(actions)
+            print(f"Rewards at step {i}: {rewards}")
+        
+        robometer_libero_env.close()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/example_videos/berkeley_rpt_stack_cup.mp4 b/robometer/scripts/example_videos/berkeley_rpt_stack_cup.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8c865f70081da9a53718d3f43e404b55487c7a8a
--- /dev/null
+++ b/robometer/scripts/example_videos/berkeley_rpt_stack_cup.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:859c984717eaf2f37d7f6fd1dc3c8dc50b2e78e6f2f35c50b5c6d61d648c4206
+size 72511
diff --git a/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards.npy b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards.npy
new file mode 100644
index 0000000000000000000000000000000000000000..98d7879efc6830ef98403fee61accf7236194e7e
--- /dev/null
+++ b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf565c58ca367878cc5ca536cad5a144cf9d68be5c734d1de1b2cb6a1eec2439
+size 168
diff --git a/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_progress_success.png b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_progress_success.png
new file mode 100644
index 0000000000000000000000000000000000000000..af9fcba55309515c5ee0f118328ba3734b431dc0
--- /dev/null
+++ b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_progress_success.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33b6b41757edb2b4c8c5bb1831206861469641c56e743eede257d1aac1de4b1c
+size 95340
diff --git a/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_success_probs.npy b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_success_probs.npy
new file mode 100644
index 0000000000000000000000000000000000000000..263b5b8de04badd57fd86cedd16df27f73b4bfde
--- /dev/null
+++ b/robometer/scripts/example_videos/berkeley_rpt_stack_cup_rewards_success_probs.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed413c17c9840d31421b01e404b0bea2f04bd27c3b18edf798970c08235f5fe5
+size 168
diff --git a/robometer/scripts/example_videos/jaco_play_pick_up_green_cup.mp4 b/robometer/scripts/example_videos/jaco_play_pick_up_green_cup.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b7a57d8dbd3c3997467cb5acec99bddb32a1dd50
--- /dev/null
+++ b/robometer/scripts/example_videos/jaco_play_pick_up_green_cup.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91e2968f0548e8a8460a467de780c4828086f647033368b228bf2e4a3f4f4988
+size 38706
diff --git a/robometer/scripts/example_videos/pick_up_green_cup_rewards.npy b/robometer/scripts/example_videos/pick_up_green_cup_rewards.npy
new file mode 100644
index 0000000000000000000000000000000000000000..2b5039d7216871bf3386140dc6e44ef7fe888537
--- /dev/null
+++ b/robometer/scripts/example_videos/pick_up_green_cup_rewards.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:406821fae0f44f59afbd01aa35bc7e56baab09708f9f70652edb0592f4dd8948
+size 180
diff --git a/robometer/scripts/example_videos/pick_up_green_cup_rewards_progress_success.png b/robometer/scripts/example_videos/pick_up_green_cup_rewards_progress_success.png
new file mode 100644
index 0000000000000000000000000000000000000000..013581f37bf256eaa3371feb5b7349d0d5a00af9
--- /dev/null
+++ b/robometer/scripts/example_videos/pick_up_green_cup_rewards_progress_success.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b9815e7387976aa675738c857c68f71333a19e8c6cf6cc4ac50415a270c5a8a
+size 92382
diff --git a/robometer/scripts/example_videos/pick_up_green_cup_rewards_success_probs.npy b/robometer/scripts/example_videos/pick_up_green_cup_rewards_success_probs.npy
new file mode 100644
index 0000000000000000000000000000000000000000..84ff4720966c5eb97c41d788318fc414fc9622d6
--- /dev/null
+++ b/robometer/scripts/example_videos/pick_up_green_cup_rewards_success_probs.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70ed9ac0d35de7956a647f262c1773f5ae67e788d9f4624b352713c8c3f7f6c5
+size 180
diff --git a/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl.mp4 b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8ce3305176de83689e10e5384362df3a761abe17
--- /dev/null
+++ b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1c2315bf14d3e9d37b82693a635f633af0fabf056bdc17c93a8ce1d53b7c30f
+size 54806
diff --git a/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards.npy b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards.npy
new file mode 100644
index 0000000000000000000000000000000000000000..9c804cab05ea4ce9810fcfcbdad1cf44d9d9907a
--- /dev/null
+++ b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df2d64fd3a93fa52ffaf92ad56d8a03229ee918a9de89c9c2e9f2bf63a89beea
+size 168
diff --git a/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_progress_success.png b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_progress_success.png
new file mode 100644
index 0000000000000000000000000000000000000000..669ca82a1422f10fea2cf9dca6f98a3545d615a5
--- /dev/null
+++ b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_progress_success.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5599a40baaceec7527b7622a3b86d86e9a6ec6b46f2d40d688daa422bb28703c
+size 101670
diff --git a/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_success_probs.npy b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_success_probs.npy
new file mode 100644
index 0000000000000000000000000000000000000000..e9d7a927b4a6ece467cd21f7e413c61db173c394
--- /dev/null
+++ b/robometer/scripts/example_videos/soar_put_green_stick_in_brown_bowl_rewards_success_probs.npy
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ff73041ed206b4d3827a85722fbd9eab787c33c0398fab3ab54a3a3e0e9c55c
+size 168
diff --git a/robometer/scripts/robotics_demo_video_scraper.py b/robometer/scripts/robotics_demo_video_scraper.py
new file mode 100644
index 0000000000000000000000000000000000000000..3dc455d2073a3bcb5a97faa0449bfdf8a4f7cb34
--- /dev/null
+++ b/robometer/scripts/robotics_demo_video_scraper.py
@@ -0,0 +1,1039 @@
+#!/usr/bin/env python3
+"""
+Scrape robotics company websites for demo videos and download them locally.
+
+Focus: robot manipulation (arms, grasping, assembly, surgical, warehouse picking).
+1. Loads (or creates) robotics_companies.txt with company names and URLs (manipulation-focused by default).
+2. Visits each URL, finds video links (direct .mp4/.webm, YouTube, Vimeo, video src).
+3. Extracts language captions when available (title, aria-label, figcaption, og:title/og:description).
+4. Optionally runs a captioning model (e.g. BLIP-2) on sampled frames to describe what the robot is doing.
+5. Downloads videos into online_demo_videos/ and saves a companion .txt with instruction, company, url, and model_caption.
+6. Keeps a state file (scraped_state.json in out-dir) of processed URLs; reruns skip URLs already downloaded or skipped.
+
+Usage:
+  uv run python scripts/robotics_demo_video_scraper.py
+  uv run python scripts/robotics_demo_video_scraper.py --companies robotics_companies.txt --out-dir online_demo_videos
+  uv run python scripts/robotics_demo_video_scraper.py --caption-model blip2   # add VLM caption per video
+  uv run python scripts/robotics_demo_video_scraper.py --crawl-sublinks --max-pages-per-site 20   # follow same-domain links to find more videos
+  uv run python scripts/robotics_demo_video_scraper.py --download-youtube   # download YouTube/Vimeo via yt-dlp (pip install yt-dlp)
+  uv run python scripts/robotics_demo_video_scraper.py --require-robot --caption-model blip2   # keep only videos where VLM detects a robot
+
+Requires: requests. Captioning: pip install transformers torch (or uv with [robometer]). YouTube/Vimeo download: pip install yt-dlp, then use --download-youtube.
+Respect robots.txt and site terms of service; use for personal/educational purposes.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import shutil
+import subprocess
+import time
+from pathlib import Path
+from urllib.parse import parse_qs, urljoin, urlparse
+
+import requests
+
+STATE_FILENAME = "scraped_state.json"
+MAX_VIDEO_DURATION_SECONDS = 20.0
+
+OUT_DIR_DEFAULT = Path("online_demo_videos")
+COMPANIES_FILE_DEFAULT = Path("robotics_companies.txt")
+USER_AGENT = "Mozilla/5.0 (Windows NT 10.0; rv:91.0) Gecko/20100101 Firefox/91.0"
+
+# Starter list: robotics companies and their main / media / news pages
+ROBOTICS_COMPANIES = [
+    # Humanoid & legged
+    ("Boston Dynamics", "https://www.bostondynamics.com"),
+    ("Boston Dynamics YouTube", "https://www.youtube.com/@BostonDynamics"),
+    ("Agility Robotics", "https://www.agilityrobotics.com"),
+    ("Figure", "https://figure.ai"),
+    ("1X Technologies", "https://www.1x.tech"),
+    ("Unitree Robotics", "https://www.unitree.com"),
+    ("Sanctuary AI", "https://www.sanctuary.ai"),
+    ("Apptronik", "https://www.apptronik.com"),
+    ("Fourier Intelligence", "https://www.fftai.com"),
+    ("Tesla Optimus", "https://www.tesla.com"),
+    ("Honda Robotics", "https://global.honda/innovation/robotics"),
+    ("Hyundai Robotics", "https://www.hyundai-robotics.com"),
+    ("Engineered Arts (Ameca)", "https://www.engineeredarts.co.uk"),
+    ("PAL Robotics", "https://pal-robotics.com"),
+    ("UBTECH", "https://www.ubtrobot.com"),
+    ("Flexiv", "https://www.flexiv.com"),
+    ("Stretch (Hello Robot)", "https://www.hello-robot.com"),
+    # Industrial arms & automation
+    ("ABB Robotics", "https://global.abb/group/en/technologies/robotics"),
+    ("FANUC", "https://www.fanuc.com"),
+    ("KUKA", "https://www.kuka.com"),
+    ("Universal Robots", "https://www.universal-robots.com"),
+    ("Yaskawa Motoman", "https://www.yaskawa.com"),
+    ("Kawasaki Robotics", "https://robotics.kawasaki.com"),
+    ("Epson Robots", "https://robots.epson.com"),
+    ("DENSO Robotics", "https://www.densorobotics.com"),
+    ("Stäubli", "https://www.staubli.com"),
+    ("Comau", "https://www.comau.com"),
+    ("Nachi Robotics", "https://www.nachirobotics.com"),
+    ("Doosan Robotics", "https://www.doosanrobotics.com"),
+    ("Techman Robot", "https://www.techmanrobot.com"),
+    ("Rethink Robotics (legacy)", "https://www.rethinkrobotics.com"),
+    ("Productive Robotics", "https://www.productiverobotics.com"),
+    ("Vention", "https://www.vention.io"),
+    ("Formic", "https://formic.co"),
+    # Medical & surgical
+    ("Intuitive (da Vinci)", "https://www.intuitive.com"),
+    ("Medtronic (Hugo, Mazor)", "https://www.medtronic.com"),
+    ("Stryker (Mako)", "https://www.stryker.com"),
+    ("Johnson & Johnson (Ottava)", "https://www.jnj.com"),
+    ("CMR Surgical (Versius)", "https://www.cmrsurgical.com"),
+    ("Asensus Surgical", "https://www.asensus.com"),
+    ("Verb Surgical", "https://www.verbsurgical.com"),
+    ("Accuray", "https://www.accuray.com"),
+    # Consumer & home
+    ("iRobot", "https://www.irobot.com"),
+    ("Ecovacs", "https://www.ecovacs.com"),
+    ("Roborock", "https://www.roborock.com"),
+    ("SharkNinja (robotics)", "https://www.sharkninja.com"),
+    ("LG Robotics", "https://www.lg.com"),
+    ("Samsung Robotics", "https://www.samsung.com"),
+    # Logistics & warehouse
+    ("Clearpath Robotics", "https://clearpathrobotics.com"),
+    ("Fetch (Zebra)", "https://www.zebra.com/us/en/about-zebra/partners/fetch-robotics.html"),
+    ("Locus Robotics", "https://www.locusrobotics.com"),
+    ("6 River Systems", "https://6river.com"),
+    ("Exotec", "https://www.exotec.com"),
+    ("Berkshire Grey", "https://www.berkshiregrey.com"),
+    ("AutoStore", "https://www.autostore.com"),
+    ("Geek+", "https://www.geekplus.com"),
+    ("GreyOrange", "https://www.greyorange.com"),
+    ("inVia Robotics", "https://www.inviarobotics.com"),
+    ("RightHand Robotics", "https://www.righthandrobotics.com"),
+    ("Plus One Robotics", "https://www.plusone.ai"),
+    ("Covariant", "https://covariant.ai"),
+    # Drones & aerial
+    ("DJI", "https://www.dji.com"),
+    ("Skydio", "https://www.skydio.com"),
+    ("Parrot", "https://www.parrot.com"),
+    ("Autel Robotics", "https://www.autelrobotics.com"),
+    ("AgEagle", "https://www.ageagle.com"),
+    ("Iris Automation", "https://www.irisautomation.com"),
+    # Autonomous vehicles & mobility
+    ("Waymo", "https://waymo.com"),
+    ("Cruise", "https://www.getcruise.com"),
+    ("Aurora", "https://aurora.tech"),
+    ("Zoox", "https://zoox.com"),
+    ("Nuro", "https://www.nuro.ai"),
+    ("Starship Technologies", "https://www.starship.xyz"),
+    ("KiwiBot", "https://www.kiwicampus.com"),
+    ("Ottonomy", "https://www.ottonomy.io"),
+    # Research & open source
+    ("Open Robotics (ROS)", "https://www.openrobotics.org"),
+    ("Willow Garage (legacy)", "https://www.willowgarage.com"),
+    ("MIT CSAIL", "https://www.csail.mit.edu"),
+    ("CMU Robotics", "https://www.ri.cmu.edu"),
+    ("Berkeley AI / RAIL", "https://rail.eecs.berkeley.edu"),
+]
+
+# Robot manipulation only: arms, grasping, assembly, surgical, warehouse picking, manipulation research.
+# Excludes drones, AVs, consumer vacuums, pure locomotion humanoids.
+ROBOTICS_MANIPULATION_COMPANIES = [
+    # Industrial arms & automation
+    # ("ABB Robotics", "https://global.abb/group/en/technologies/robotics"),
+    # ("FANUC", "https://www.fanuc.com"),
+    # ("KUKA", "https://www.kuka.com"),
+    # ("Universal Robots", "https://www.universal-robots.com"),
+    # ("Yaskawa Motoman", "https://www.yaskawa.com"),
+    # ("Kawasaki Robotics", "https://robotics.kawasaki.com"),
+    # ("Epson Robots", "https://robots.epson.com"),
+    # ("DENSO Robotics", "https://www.densorobotics.com"),
+    # ("Stäubli", "https://www.staubli.com"),
+    # ("Comau", "https://www.comau.com"),
+    # ("Nachi Robotics", "https://www.nachirobotics.com"),
+    # ("Doosan Robotics", "https://www.doosanrobotics.com"),
+    # ("Techman Robot", "https://www.techmanrobot.com"),
+    # ("Rethink Robotics (legacy)", "https://www.rethinkrobotics.com"),
+    # ("Productive Robotics", "https://www.productiverobotics.com"),
+    # ("Vention", "https://www.vention.io"),
+    # ("Formic", "https://formic.co"),
+    # # Medical & surgical
+    # ("Intuitive (da Vinci)", "https://www.intuitive.com"),
+    # ("Medtronic (Hugo, Mazor)", "https://www.medtronic.com"),
+    # ("Stryker (Mako)", "https://www.stryker.com"),
+    # ("Johnson & Johnson (Ottava)", "https://www.jnj.com"),
+    # ("CMR Surgical (Versius)", "https://www.cmrsurgical.com"),
+    # ("Asensus Surgical", "https://www.asensus.com"),
+    # ("Verb Surgical", "https://www.verbsurgical.com"),
+    # ("Accuray", "https://www.accuray.com"),
+    # # Warehouse picking & manipulation
+    # ("RightHand Robotics", "https://www.righthandrobotics.com"),
+    # ("Plus One Robotics", "https://www.plusone.ai"),
+    # ("Covariant", "https://covariant.ai"),
+    # ("Berkshire Grey", "https://www.berkshiregrey.com"),
+    # ("inVia Robotics", "https://www.inviarobotics.com"),
+    # # Research
+    # ("Open Robotics (ROS)", "https://www.openrobotics.org"),
+    # ("Willow Garage (legacy)", "https://www.willowgarage.com"),
+    # ("MIT CSAIL", "https://www.csail.mit.edu"),
+    # ("CMU Robotics", "https://www.ri.cmu.edu"),
+    # ("Berkeley AI / RAIL", "https://rail.eecs.berkeley.edu"),
+    # # Manipulation-focused arms / humanoids
+    # ("Stretch (Hello Robot)", "https://www.hello-robot.com"),
+    # ("Flexiv", "https://www.flexiv.com"),
+    ("Figure", "https://figure.ai"),
+    # ("1X Technologies", "https://www.1x.tech"),
+    # ("Apptronik", "https://www.apptronik.com"),
+    # ("Sanctuary AI", "https://www.sanctuary.ai"),
+]
+
+
+def ensure_companies_file(path: Path, manipulation_only: bool = True) -> None:
+    """Create robotics_companies.txt with default list if it doesn't exist.
+    By default uses manipulation-only companies (arms, surgical, picking, research).
+    """
+    if path.exists():
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    companies = ROBOTICS_MANIPULATION_COMPANIES if manipulation_only else ROBOTICS_COMPANIES
+    with open(path, "w", encoding="utf-8") as f:
+        f.write("# Robot manipulation companies: one per line as 'Name<TAB>URL' or 'Name, URL'\n")
+        for name, url in companies:
+            f.write(f"{name}\t{url}\n")
+    print(f"Wrote {path} with {len(companies)} entries (manipulation_only={manipulation_only}). Edit to add more.")
+
+
+def load_companies(path: Path) -> list[tuple[str, str]]:
+    """Load (name, url) pairs from file. Lines: 'Name\\tURL' or 'Name, URL'; # ignored."""
+    pairs: list[tuple[str, str]] = []
+    with open(path, encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line or line.startswith("#"):
+                continue
+            if "\t" in line:
+                name, url = line.split("\t", 1)
+            else:
+                name, _, url = line.partition(",")
+                name, url = name.strip(), url.strip()
+            if name and url and url.startswith("http"):
+                pairs.append((name, url))
+    return pairs
+
+
+def _extract_page_caption(html: str) -> str:
+    """Extract page-level caption from og:title and og:description."""
+    parts: list[str] = []
+    for prop, name in [("og:title", "title"), ("og:description", "description")]:
+        m = re.search(
+            rf'<meta[^>]+property=["\']{re.escape(prop)}["\'][^>]+content=["\']([^"\']+)["\']',
+            html,
+            re.I,
+        )
+        if not m:
+            m = re.search(
+                rf'<meta[^>]+content=["\']([^"\']+)["\'][^>]+property=["\']{re.escape(prop)}["\']',
+                html,
+                re.I,
+            )
+        if m:
+            parts.append(m.group(1).strip())
+    return " | ".join(parts)[:2000] if parts else ""
+
+
+def _extract_caption_near(html: str, start: int, end: int) -> str:
+    """Extract caption from HTML segment (title=, aria-label=, data-caption=, figcaption)."""
+    segment = html[max(0, start - 600) : end + 200]
+    # title="..." or aria-label="..."
+    for attr in ("title", "aria-label", "data-caption", "data-title", "data-description"):
+        m = re.search(rf'{attr}=["\']([^"\']+)["\']', segment, re.I)
+        if m:
+            s = m.group(1).strip()
+            if len(s) > 10 and "script" not in s.lower():
+                return re.sub(r"\s+", " ", s)[:2000]
+    # <figcaption>...</figcaption>
+    m = re.search(r"<figcaption[^>]*>([^<]+(?:<[^>]+>[^<]*)*)</figcaption>", segment, re.I | re.DOTALL)
+    if m:
+        text = re.sub(r"<[^>]+>", " ", m.group(1)).strip()
+        text = re.sub(r"\s+", " ", text)[:2000]
+        if len(text) > 5:
+            return text
+    return ""
+
+
+def find_video_urls_in_html(html: str, base_url: str) -> list[tuple[str, str]]:
+    """Extract video URLs and optional captions from HTML. Returns list of (url, caption)."""
+    base = base_url.rstrip("/")
+    page_caption = _extract_page_caption(html)
+    results: list[tuple[str, str]] = []
+    seen: set[str] = set()
+
+    def add(url: str, caption: str = "") -> None:
+        if url not in seen and url.startswith("http"):
+            seen.add(url)
+            cap = caption.strip() or page_caption
+            results.append((url, cap))
+
+    # Direct video extensions (with optional caption from surrounding HTML)
+    ext_pat = re.compile(
+        r"\b(href|src|content)=[\"']([^\"']+?\.(?:mp4|webm|mov|m4v|ogv))[\"']",
+        re.I,
+    )
+    for m in re.finditer(ext_pat, html):
+        url = urljoin(base + "/", m.group(2))
+        cap = _extract_caption_near(html, m.start(), m.end())
+        add(url, cap)
+
+    # Any href or src with video-like path
+    url_in_attr = re.compile(
+        r'(?:href|src|content)=["\']([^"\']+)["\']',
+        re.I,
+    )
+    for m in re.finditer(url_in_attr, html):
+        raw = m.group(1).strip()
+        if not raw or raw.startswith("#") or raw.startswith("javascript:"):
+            continue
+        full = urljoin(base + "/", raw)
+        if full.startswith("http") and (
+            ".mp4" in full
+            or ".webm" in full
+            or ".mov" in full
+            or "youtube.com" in full
+            or "youtu.be" in full
+            or "vimeo.com" in full
+            or "video" in full.lower()
+            or "/v/" in full
+        ):
+            cap = _extract_caption_near(html, m.start(), m.end())
+            add(full, cap)
+
+    # YouTube embed (caption from iframe title if present)
+    yt = re.compile(
+        r"(?:youtube\.com/(?:embed/|watch\?v=)|youtu\.be/)([a-zA-Z0-9_-]{11})"
+    )
+    for m in re.finditer(yt, html):
+        url = f"https://www.youtube.com/watch?v={m.group(1)}"
+        cap = _extract_caption_near(html, m.start(), m.end())
+        add(url, cap)
+
+    # Vimeo
+    vimeo = re.compile(r"vimeo\.com/(?:video/)?(\d+)")
+    for m in re.finditer(vimeo, html):
+        url = f"https://vimeo.com/{m.group(1)}"
+        cap = _extract_caption_near(html, m.start(), m.end())
+        add(url, cap)
+
+    return results
+
+
+def _same_domain_links(html: str, base_url: str, netloc: str) -> set[str]:
+    """Extract href URLs that belong to the same domain as base_url. Returns absolute URLs."""
+    base = base_url.rstrip("/")
+    seen: set[str] = set()
+    # href="..."
+    for m in re.finditer(r'href\s*=\s*["\']([^"\']+)["\']', html, re.I):
+        raw = m.group(1).strip()
+        if not raw or raw.startswith("#") or raw.startswith("javascript:") or raw.startswith("mailto:"):
+            continue
+        full = urljoin(base + "/", raw)
+        parsed = urlparse(full)
+        if parsed.netloc != netloc or not full.startswith("http"):
+            continue
+        # Skip obvious non-HTML (files we don't want to fetch as pages)
+        path_lower = parsed.path.lower()
+        if any(path_lower.endswith(ext) for ext in (".pdf", ".zip", ".mp4", ".webm", ".mov", ".jpg", ".jpeg", ".png", ".gif", ".svg", ".css", ".js")):
+            continue
+        seen.add(full)
+    return seen
+
+
+def _crawl_site_for_videos(
+    start_url: str,
+    name: str,
+    session: requests.Session,
+    delay: float,
+    max_pages: int,
+) -> list[tuple[str, str]]:
+    """Crawl same-domain pages starting from start_url and collect all (video_url, caption) pairs."""
+    parsed_start = urlparse(start_url)
+    netloc = parsed_start.netloc
+    results: list[tuple[str, str]] = []
+    seen_pages: set[str] = set()
+    queue: list[str] = [start_url.rstrip("/")]
+    pages_fetched = 0
+
+    while queue and pages_fetched < max_pages:
+        url = queue.pop(0)
+        if url in seen_pages:
+            continue
+        seen_pages.add(url)
+        pages_fetched += 1
+        try:
+            r = session.get(url, timeout=15)
+            r.raise_for_status()
+            html = r.text
+        except Exception as e:
+            print(f"  Skip subpage {url[:60]}...: {e}")
+            time.sleep(delay)
+            continue
+        for u, caption in find_video_urls_in_html(html, url):
+            results.append((u, caption))
+        # Enqueue same-domain links we haven't visited
+        for link in _same_domain_links(html, url, netloc):
+            if link not in seen_pages and link not in queue:
+                queue.append(link)
+        time.sleep(delay)
+
+    return results
+
+
+def _get_video_duration_seconds(path: Path) -> float | None:
+    """Return video duration in seconds via ffprobe, or None if unknown/unavailable."""
+    try:
+        out = subprocess.run(
+            [
+                "ffprobe", "-v", "error", "-show_entries", "format=duration",
+                "-of", "default=noprint_wrappers=1:nokey=1", str(path),
+            ],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+        if out.returncode != 0 or not out.stdout.strip():
+            return None
+        return float(out.stdout.strip())
+    except (subprocess.TimeoutExpired, FileNotFoundError, OSError, ValueError):
+        return None
+
+
+def _is_mp4_file(path: Path) -> bool:
+    """Return True if file looks like an MP4 (ftyp at offset 4)."""
+    try:
+        with open(path, "rb") as f:
+            f.seek(4)
+            return f.read(4) == b"ftyp"
+    except Exception:
+        return False
+
+
+def _is_likely_html(path: Path) -> bool:
+    """Return True if file looks like HTML (common when server returns error page)."""
+    try:
+        with open(path, "rb") as f:
+            head = f.read(512)
+        return head.lstrip().startswith((b"<", b"\r\n", b"\n")) or b"<!DOCTYPE" in head[:200]
+    except Exception:
+        return False
+
+
+def _convert_to_mp4(path: Path) -> bool:
+    """Convert video to MP4 with ffmpeg (remux or re-encode). Returns True if successful."""
+    if _is_mp4_file(path):
+        return True
+    tmp = path.with_suffix(".tmp.mp4")
+    try:
+        # Try remux first (fast); then re-encode if needed
+        out = subprocess.run(
+            [
+                "ffmpeg", "-y", "-i", str(path),
+                "-c", "copy", "-movflags", "+faststart",
+                str(tmp),
+            ],
+            capture_output=True,
+            timeout=120,
+        )
+        if out.returncode != 0:
+            out = subprocess.run(
+                [
+                    "ffmpeg", "-y", "-i", str(path),
+                    "-c:v", "libx264", "-preset", "fast", "-c:a", "aac",
+                    "-movflags", "+faststart", str(tmp),
+                ],
+                capture_output=True,
+                timeout=300,
+            )
+        if out.returncode != 0:
+            return False
+        path.unlink(missing_ok=True)
+        tmp.rename(path)
+        return True
+    except (subprocess.TimeoutExpired, FileNotFoundError, OSError):
+        tmp.unlink(missing_ok=True)
+        return False
+    finally:
+        if tmp.exists():
+            tmp.unlink(missing_ok=True)
+
+
+def download_direct(url: str, path: Path, session: requests.Session) -> bool:
+    """Download a direct video URL and ensure output is valid .mp4. Returns True if successful.
+    Caller should pass path with .mp4 extension; non-MP4 content is converted via ffmpeg.
+    """
+    try:
+        r = session.get(url, stream=True, timeout=30)
+        r.raise_for_status()
+        path.parent.mkdir(parents=True, exist_ok=True)
+        with open(path, "wb") as f:
+            for chunk in r.iter_content(chunk_size=65536):
+                if chunk:
+                    f.write(chunk)
+    except Exception as e:
+        print(f"  Skip {url[:60]}...: {e}")
+        return False
+
+    # Reject HTML (error pages)
+    if _is_likely_html(path):
+        path.unlink(missing_ok=True)
+        print(f"  Skip (not video, got HTML): {path.name}")
+        return False
+
+    # Ensure valid MP4: convert with ffmpeg if not already MP4
+    if not _is_mp4_file(path):
+        if not _convert_to_mp4(path):
+            path.unlink(missing_ok=True)
+            print(f"  Skip (ffmpeg convert failed): {path.name}")
+            return False
+    return True
+
+
+def _youtube_vimeo_id(url: str) -> str | None:
+    """Extract video id from YouTube or Vimeo URL for stable filenames. Returns None if not recognized."""
+    if "youtube.com" in url or "youtu.be" in url:
+        if "youtu.be/" in url:
+            m = re.search(r"youtu\.be/([a-zA-Z0-9_-]{11})", url)
+            return m.group(1) if m else None
+        parsed = urlparse(url)
+        if parsed.netloc and "youtube" in parsed.netloc:
+            q = parse_qs(parsed.query)
+            v = q.get("v", [])
+            return v[0] if v else None
+    if "vimeo.com" in url:
+        m = re.search(r"vimeo\.com/(?:video/)?(\d+)", url)
+        return m.group(1) if m else None
+    return None
+
+
+def download_youtube_or_vimeo(url: str, output_path: Path) -> Path | None:
+    """Download a YouTube or Vimeo URL with yt-dlp; output is converted to .mp4. Returns final path or None."""
+    if not shutil.which("yt-dlp"):
+        print("  yt-dlp not found; install with: pip install yt-dlp")
+        return None
+    output_path = output_path.resolve()
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    # yt-dlp writes to stem.%(ext)s; we'll get stem.mkv or stem.webm etc., then convert to stem.mp4
+    stem = output_path.with_suffix("")
+    out_tpl = str(stem) + ".%(ext)s"
+    try:
+        out = subprocess.run(
+            [
+                "yt-dlp",
+                "--no-warnings",
+                "-o",
+                out_tpl,
+                "--no-playlist",
+                "--max-downloads", "1",
+                url,
+            ],
+            capture_output=True,
+            text=True,
+            timeout=300,
+        )
+        if out.returncode != 0:
+            print(f"  yt-dlp failed: {out.stderr[:200] if out.stderr else out.stdout[:200]}")
+            return None
+    except (subprocess.TimeoutExpired, FileNotFoundError, OSError) as e:
+        print(f"  yt-dlp error: {e}")
+        return None
+    # Find the file yt-dlp wrote (stem with some extension)
+    downloaded: Path | None = None
+    for f in output_path.parent.iterdir():
+        if f.stem == stem.name and f.suffix.lower() in (".mp4", ".webm", ".mkv", ".mov", ".m4a"):
+            downloaded = f
+            break
+    if not downloaded or not downloaded.is_file():
+        return None
+    # Ensure final file is .mp4
+    final = stem.with_suffix(".mp4")
+    if downloaded.suffix.lower() != ".mp4" or not _is_mp4_file(downloaded):
+        if not _convert_to_mp4(downloaded):
+            downloaded.unlink(missing_ok=True)
+            return None
+        if downloaded.suffix.lower() != ".mp4":
+            downloaded.rename(final)
+        else:
+            final = downloaded
+    else:
+        if downloaded != final:
+            downloaded.rename(final)
+    return final if final.exists() else None
+
+
+def _sample_frames(video_path: Path, num_frames: int = 3) -> list:
+    """Sample num_frames from video (start, middle, end). Returns list of PIL Images."""
+    try:
+        from decord import VideoReader  # type: ignore
+    except ImportError:
+        raise RuntimeError("decord is required for captioning; pip install decord")
+    import numpy as np
+    from PIL import Image
+
+    vr = VideoReader(str(video_path), num_threads=1)
+    n = len(vr)
+    if n == 0:
+        return []
+    indices = [
+        int(i * (n - 1) / max(1, num_frames - 1))
+        for i in range(num_frames)
+    ]
+    frames = vr.get_batch(indices).asnumpy()  # (N, H, W, 3) uint8
+    return [Image.fromarray(f).convert("RGB") for f in frames]
+
+
+# Lazy-loaded caption model (processor, model) keyed by model_id
+_caption_model_cache: dict[str, tuple] = {}
+
+
+def _caption_video_with_model(video_path: Path, model_id: str, prompt: str = "What is the robot doing in this image?") -> str:
+    """Run a vision-language model on sampled frames and return a single caption for the video.
+    model_id: e.g. 'blip2' (Salesforce/blip2-opt-2.7b) or any HuggingFace model id for Blip2ForConditionalGeneration.
+    """
+    try:
+        from transformers import Blip2Processor, Blip2ForConditionalGeneration  # type: ignore
+        import torch  # type: ignore
+    except ImportError as e:
+        raise RuntimeError("transformers and torch required for captioning; pip install transformers torch") from e
+
+    if model_id.lower() == "blip2":
+        model_id = "Salesforce/blip2-opt-2.7b"
+
+    if model_id not in _caption_model_cache:
+        processor = Blip2Processor.from_pretrained(model_id)
+        model = Blip2ForConditionalGeneration.from_pretrained(model_id)
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        model = model.to(device)
+        _caption_model_cache[model_id] = (processor, model, device)
+
+    processor, model, device = _caption_model_cache[model_id]
+    images = _sample_frames(video_path, num_frames=3)
+    if not images:
+        return ""
+
+    captions: list[str] = []
+    for img in images:
+        inputs = processor(images=img, text=prompt, return_tensors="pt")
+        inputs = {k: v.to(device) for k, v in inputs.items()}
+        out = model.generate(**inputs, max_new_tokens=80)
+        cap = processor.decode(out[0], skip_special_tokens=True).strip()
+        if cap and cap != prompt:
+            captions.append(cap)
+    if not captions:
+        return ""
+    # Prefer a single representative caption; if we have several, take the longest (often most descriptive)
+    return max(captions, key=len)
+
+
+def _query_vlm_yes_no(video_path: Path, model_id: str, question: str) -> bool:
+    """Run VLM on sampled frames with a yes/no question; return True if answer suggests yes.
+    Uses same BLIP-2 pipeline as captioning. Parses first few tokens for yes/no/robot.
+    """
+    raw = _caption_video_with_model(video_path, model_id, prompt=question)
+    if not raw:
+        return False
+    s = raw.strip().lower()
+    if s.startswith("yes") or s.startswith("yeah"):
+        return True
+    if s.startswith("no") or s.startswith("nope"):
+        return False
+    # Answer describes something; treat as yes if it mentions a robot (and doesn't negate)
+    if "robot" in s and not s.startswith("no ") and "no robot" not in s[:30]:
+        return True
+    return False
+
+
+def _video_has_robot(video_path: Path, model_id: str) -> bool:
+    """Return True if VLM indicates a robot is present in the video (sample frames)."""
+    question = "Is there a robot in this image? Answer yes or no."
+    return _query_vlm_yes_no(video_path, model_id, question)
+
+
+def _caption_to_instruction(caption: str, company: str) -> str:
+    """Turn scraped caption into a language instruction for the video (task description)."""
+    if caption and len(caption.strip()) > 5:
+        # Use as-is or normalize to imperative/task form
+        s = caption.strip()
+        s = re.sub(r"\s+", " ", s)
+        # If it looks like a title (no verb), prefix with "Demonstrate: " or use as task
+        if len(s) < 200 and not any(
+            s.strip().lower().startswith(p)
+            for p in ("the robot", "robot", "a ", "demonstrate", "show", "perform")
+        ):
+            return f"Demonstrate: {s}" if not s.endswith(".") else s
+        return s[:2000]
+    if company:
+        return f"Robot demonstration from {company}."
+    return "Robot demonstration."
+
+
+def save_caption(
+    path: Path,
+    caption: str,
+    company: str,
+    url: str,
+    model_caption: str | None = None,
+) -> None:
+    """Write companion .txt with instruction (language task for the video), company, url, and optional model_caption."""
+    # Prefer model-generated caption for instruction when available
+    if model_caption and len(model_caption.strip()) > 5:
+        instruction = model_caption.strip()[:2000]
+        if not instruction.endswith("."):
+            instruction += "."
+    else:
+        instruction = _caption_to_instruction(caption, company)
+    txt_path = path.with_suffix(".txt")
+    lines = [
+        "instruction: " + instruction,
+        "",
+        "company: " + (company or ""),
+        "url: " + (url or ""),
+    ]
+    if model_caption and model_caption.strip():
+        lines.append("")
+        lines.append("model_caption: " + model_caption.strip()[:2000])
+    if caption and caption.strip() != instruction:
+        lines.append("")
+        lines.append("caption: " + caption.strip()[:2000])
+    txt_path.parent.mkdir(parents=True, exist_ok=True)
+    with open(txt_path, "w", encoding="utf-8") as f:
+        f.write("\n".join(lines))
+
+
+def safe_filename(name: str, max_len: int = 80) -> str:
+    """Make a safe filename from a string."""
+    s = re.sub(r"[^\w\s\-\.]", "", name)
+    s = re.sub(r"\s+", "_", s).strip("_")
+    return s[:max_len] or "video"
+
+
+def _state_path(out_dir: Path) -> Path:
+    return out_dir / STATE_FILENAME
+
+
+def load_state(out_dir: Path) -> dict[str, str]:
+    """Load processed URL -> status from scraped_state.json. Status: downloaded, skipped_robot, skipped_fail, skipped_too_long."""
+    path = _state_path(out_dir)
+    if not path.exists():
+        return {}
+    try:
+        with open(path, encoding="utf-8") as f:
+            data = json.load(f)
+        return data.get("by_url", data) if isinstance(data, dict) else {}
+    except (json.JSONDecodeError, OSError):
+        return {}
+
+
+def save_state(out_dir: Path, state: dict[str, str]) -> None:
+    """Write processed URL -> status to scraped_state.json."""
+    path = _state_path(out_dir)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with open(path, "w", encoding="utf-8") as f:
+        json.dump({"by_url": state, "version": 1}, f, indent=0)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Scrape robotics company sites for demo videos and download them."
+    )
+    parser.add_argument(
+        "--companies",
+        type=Path,
+        default=COMPANIES_FILE_DEFAULT,
+        help="Path to .txt with company names and URLs (Name\\tURL per line)",
+    )
+    parser.add_argument(
+        "--out-dir",
+        type=Path,
+        default=OUT_DIR_DEFAULT,
+        help="Directory to save videos",
+    )
+    parser.add_argument(
+        "--delay",
+        type=float,
+        default=1.0,
+        help="Seconds between requests to the same site",
+    )
+    parser.add_argument(
+        "--skip-youtube-vimeo",
+        action="store_true",
+        help="Do not list YouTube/Vimeo (only download direct links unless yt-dlp used)",
+    )
+    parser.add_argument(
+        "--all-companies",
+        action="store_true",
+        help="Use full robotics list (drones, AVs, etc.) when creating companies file; default is manipulation-only",
+    )
+    parser.add_argument(
+        "--caption-model",
+        type=str,
+        default=None,
+        metavar="ID",
+        help="Run a vision model to caption each video (e.g. blip2 or Salesforce/blip2-opt-2.7b). Requires transformers, torch, decord.",
+    )
+    parser.add_argument(
+        "--crawl-sublinks",
+        action="store_true",
+        help="Follow same-domain links on each site to find videos on subpages (not just the main URL).",
+    )
+    parser.add_argument(
+        "--max-pages-per-site",
+        type=int,
+        default=20,
+        metavar="N",
+        help="When using --crawl-sublinks, limit to N pages per company site (default 20).",
+    )
+    parser.add_argument(
+        "--download-youtube",
+        action="store_true",
+        help="Download YouTube/Vimeo videos via yt-dlp into out-dir (pip install yt-dlp).",
+    )
+    parser.add_argument(
+        "--require-robot",
+        action="store_true",
+        help="After download, run VLM to check if a robot is in the video; keep only if yes (uses --caption-model or blip2).",
+    )
+    args = parser.parse_args()
+
+    ensure_companies_file(args.companies, manipulation_only=not args.all_companies)
+    companies = load_companies(args.companies)
+    if not companies:
+        print("No companies found in", args.companies)
+        return
+
+    print(f"Loaded {len(companies)} companies. Output dir: {args.out_dir.absolute()}")
+    if args.crawl_sublinks:
+        print(f"  Crawl sublinks: up to {args.max_pages_per_site} pages per site")
+    if args.download_youtube:
+        print("  Download YouTube/Vimeo: yes (yt-dlp)")
+    if args.require_robot:
+        print("  Require robot in video: yes (VLM filter)")
+    if args.caption_model:
+        print(f"  Caption model: {args.caption_model}")
+    print(f"  Max video duration: {MAX_VIDEO_DURATION_SECONDS:.0f}s (longer videos skipped)")
+    print()
+
+    args.out_dir.mkdir(parents=True, exist_ok=True)
+    state = load_state(args.out_dir)
+    if state:
+        print(f"Loaded state: {len(state)} URL(s) already processed (will skip)")
+    session = requests.Session()
+    session.headers["User-Agent"] = USER_AGENT
+
+    all_direct: list[tuple[str, str, str, str]] = []  # (company_name, url, suggested_path, caption)
+    yt_vimeo: list[tuple[str, str, str]] = []  # (company_name, url, caption)
+
+    for name, url in companies:
+        if args.crawl_sublinks:
+            print(f"Crawling ({args.max_pages_per_site} pages max): {name} — {url}")
+            try:
+                page_results = _crawl_site_for_videos(
+                    url, name, session, args.delay, args.max_pages_per_site
+                )
+            except Exception as e:
+                print(f"  Error: {e}")
+                continue
+        else:
+            print(f"Fetching: {name} — {url}")
+            try:
+                r = session.get(url, timeout=15)
+                r.raise_for_status()
+                html = r.text
+            except Exception as e:
+                print(f"  Error: {e}")
+                continue
+            page_results = find_video_urls_in_html(html, url)
+            time.sleep(args.delay)
+        n_direct = 0
+        n_yt = 0
+        for u, caption in page_results:
+            if "youtube.com" in u or "youtu.be" in u or "vimeo.com" in u:
+                yt_vimeo.append((name, u, caption))
+                n_yt += 1
+            else:
+                slug = safe_filename(name)
+                url_hash = str(abs(hash(u)) % 10**8)[:8]
+                all_direct.append((name, u, f"{slug}_{url_hash}.mp4", caption))
+                n_direct += 1
+        print(f"  Found {n_direct} direct video(s), {n_yt} YouTube/Vimeo link(s)")
+        if args.crawl_sublinks:
+            time.sleep(args.delay)
+
+    # Deduplicate by URL for direct (keep first occurrence and its caption)
+    n_unique_direct = len({u for _, u, _, _ in all_direct})
+    print(f"\nTotal: {len(all_direct)} direct video URL(s) (before dedup), {n_unique_direct} unique; {len(yt_vimeo)} YouTube/Vimeo URL(s)")
+    robot_check_model: str | None = (args.caption_model or "blip2") if args.require_robot else None
+    if robot_check_model:
+        print(f"Robot check enabled (model: {robot_check_model})")
+    print("\n--- Direct videos ---")
+    seen = set()
+    n_direct_kept = 0
+    n_direct_skipped_robot = 0
+    n_direct_skipped_fail = 0
+    n_direct_skipped_too_long = 0
+    n_direct_existed = 0
+    n_direct_skipped_state = 0
+    for name, u, fname, caption in all_direct:
+        if u in seen:
+            continue
+        seen.add(u)
+        if u in state:
+            n_direct_skipped_state += 1
+            continue
+        path = args.out_dir / fname
+        model_caption: str | None = None
+        if args.caption_model and path.exists():
+            try:
+                print(f"Captioning: {path.name}")
+                model_caption = _caption_video_with_model(path, args.caption_model)
+            except Exception as e:
+                print(f"  Caption error: {e}")
+        if path.exists():
+            print(f"Exists: {path.name}")
+            save_caption(path, caption, name, u, model_caption=model_caption)
+            state[u] = "downloaded"
+            save_state(args.out_dir, state)
+            n_direct_existed += 1
+            time.sleep(args.delay)
+            continue
+        print(f"Downloading: {path.name}")
+        if download_direct(u, path, session):
+            duration = _get_video_duration_seconds(path)
+            if duration is not None and duration > MAX_VIDEO_DURATION_SECONDS:
+                path.unlink(missing_ok=True)
+                print(f"  Skip (video > {MAX_VIDEO_DURATION_SECONDS:.0f}s): {path.name} ({duration:.1f}s)")
+                state[u] = "skipped_too_long"
+                save_state(args.out_dir, state)
+                n_direct_skipped_too_long += 1
+                time.sleep(args.delay)
+                continue
+            if robot_check_model:
+                try:
+                    print(f"  Checking for robot: {path.name}")
+                    if not _video_has_robot(path, robot_check_model):
+                        path.unlink(missing_ok=True)
+                        print(f"  Skip (no robot detected): {path.name}")
+                        state[u] = "skipped_robot"
+                        save_state(args.out_dir, state)
+                        n_direct_skipped_robot += 1
+                        time.sleep(args.delay)
+                        continue
+                    print(f"  Robot detected, keeping.")
+                except Exception as e:
+                    print(f"  Robot check error: {e}")
+            if args.caption_model:
+                try:
+                    print(f"  Captioning: {path.name}")
+                    model_caption = _caption_video_with_model(path, args.caption_model)
+                except Exception as e:
+                    print(f"  Caption error: {e}")
+            save_caption(path, caption, name, u, model_caption=model_caption)
+            print(f"  Saved: {path.name} (+ .txt)")
+            state[u] = "downloaded"
+            save_state(args.out_dir, state)
+            n_direct_kept += 1
+        else:
+            state[u] = "skipped_fail"
+            save_state(args.out_dir, state)
+            n_direct_skipped_fail += 1
+        time.sleep(args.delay)
+
+    # YouTube/Vimeo: optionally download with yt-dlp, or just list
+    if yt_vimeo:
+        if args.download_youtube:
+            n_yt_unique = len({u for _, u, _ in yt_vimeo})
+            print(f"\n--- YouTube/Vimeo ({n_yt_unique} unique) ---")
+            seen_yt = set()
+            n_yt_kept = 0
+            n_yt_skipped_robot = 0
+            n_yt_skipped_too_long = 0
+            n_yt_existed = 0
+            n_yt_skipped_state = 0
+            for name, u, caption in yt_vimeo:
+                if u in seen_yt:
+                    continue
+                seen_yt.add(u)
+                if u in state:
+                    n_yt_skipped_state += 1
+                    continue
+                vid = _youtube_vimeo_id(u)
+                fname = f"{safe_filename(name)}_{vid or abs(hash(u)) % 10**8}.mp4"
+                path = args.out_dir / fname
+                if path.exists():
+                    print(f"Exists (YT/Vimeo): {path.name}")
+                    save_caption(path, caption, name, u)
+                    state[u] = "downloaded"
+                    save_state(args.out_dir, state)
+                    n_yt_existed += 1
+                    time.sleep(args.delay)
+                    continue
+                print(f"Downloading (yt-dlp): {path.name}")
+                final = download_youtube_or_vimeo(u, path)
+                if final:
+                    print(f"  Downloaded: {final.name}")
+                    duration = _get_video_duration_seconds(final)
+                    if duration is not None and duration > MAX_VIDEO_DURATION_SECONDS:
+                        final.unlink(missing_ok=True)
+                        print(f"  Skip (video > {MAX_VIDEO_DURATION_SECONDS:.0f}s): {final.name} ({duration:.1f}s)")
+                        state[u] = "skipped_too_long"
+                        save_state(args.out_dir, state)
+                        n_yt_skipped_too_long += 1
+                        time.sleep(args.delay)
+                        continue
+                    if robot_check_model:
+                        try:
+                            print(f"  Checking for robot: {final.name}")
+                            if not _video_has_robot(final, robot_check_model):
+                                final.unlink(missing_ok=True)
+                                print(f"  Skip (no robot detected): {final.name}")
+                                state[u] = "skipped_robot"
+                                save_state(args.out_dir, state)
+                                n_yt_skipped_robot += 1
+                                time.sleep(args.delay)
+                                continue
+                            print(f"  Robot detected, keeping.")
+                        except Exception as e:
+                            print(f"  Robot check error: {e}")
+                    model_caption = None
+                    if args.caption_model:
+                        try:
+                            print(f"  Captioning: {final.name}")
+                            model_caption = _caption_video_with_model(final, args.caption_model)
+                        except Exception as e:
+                            print(f"  Caption error: {e}")
+                    save_caption(final, caption, name, u, model_caption=model_caption)
+                    print(f"  Saved: {final.name} (+ .txt)")
+                    state[u] = "downloaded"
+                    save_state(args.out_dir, state)
+                    n_yt_kept += 1
+                else:
+                    state[u] = "skipped_fail"
+                    save_state(args.out_dir, state)
+                time.sleep(args.delay)
+        elif not args.skip_youtube_vimeo:
+            print("\nYouTube/Vimeo URLs (use --download-youtube to download via yt-dlp):")
+            for name, u, cap in yt_vimeo[:30]:
+                print(f"  {name}: {u}")
+                if cap:
+                    print(f"    caption: {(cap[:80] + '...') if len(cap) > 80 else cap}")
+            if len(yt_vimeo) > 30:
+                print(f"  ... and {len(yt_vimeo) - 30} more")
+
+    # Summary
+    print("\n" + "=" * 60)
+    print("Summary")
+    print("=" * 60)
+    print(f"  Direct: {n_direct_kept} new, {n_direct_existed} already existed, {n_direct_skipped_state} skipped (in state), {n_direct_skipped_too_long} skipped (>{MAX_VIDEO_DURATION_SECONDS:.0f}s), {n_direct_skipped_robot} skipped (no robot), {n_direct_skipped_fail} download failed")
+    if yt_vimeo and args.download_youtube:
+        print(f"  YouTube/Vimeo: {n_yt_kept} new, {n_yt_existed} already existed, {n_yt_skipped_state} skipped (in state), {n_yt_skipped_too_long} skipped (>{MAX_VIDEO_DURATION_SECONDS:.0f}s), {n_yt_skipped_robot} skipped (no robot)")
+    print(f"  Output: {args.out_dir.absolute()}")
+    print("Done.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/single_video_benchmark_style.py b/robometer/scripts/single_video_benchmark_style.py
new file mode 100644
index 0000000000000000000000000000000000000000..79a2cf571696ebf36bcf41adfaba84eed2f5b145
--- /dev/null
+++ b/robometer/scripts/single_video_benchmark_style.py
@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+
+from benchmark_progress_mark_local import (
+    DEFAULT_INFERENCE_MODE,
+    DEFAULT_MAX_FRAMES,
+    DEFAULT_MIN_FRAMES,
+    DEFAULT_MODEL_PATH,
+    DEFAULT_PREFIX_BATCH_SIZE,
+    DEFAULT_PREFIX_SAMPLE_FRAMES,
+    RobometerLocalRunner,
+    build_frame_retry_schedule,
+    is_cuda_oom_error,
+    load_all_video_frames,
+    sample_video_frames_with_indices,
+    write_overview_plot,
+)
+
+
+def save_results(output_dir: Path, payload: dict) -> None:
+    with (output_dir / "results.json").open("w", encoding="utf-8") as f:
+        json.dump(payload, f, indent=2, ensure_ascii=False)
+        f.write("\n")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Run Robometer on a single video with the same benchmark-style inference logic, without progress marks."
+    )
+    parser.add_argument("--video", required=True, help="Path to the input video")
+    parser.add_argument("--task", required=True, help="Task instruction")
+    parser.add_argument("--model-path", default=str(DEFAULT_MODEL_PATH))
+    parser.add_argument("--fps", type=float, default=3.0)
+    parser.add_argument("--max-frames", type=int, default=DEFAULT_MAX_FRAMES)
+    parser.add_argument("--min-frames", type=int, default=DEFAULT_MIN_FRAMES)
+    parser.add_argument(
+        "--inference-mode",
+        choices=["frame_steps", "whole"],
+        default=DEFAULT_INFERENCE_MODE,
+        help="frame_steps matches benchmark behavior; whole does one forward pass on the full sampled trajectory",
+    )
+    parser.add_argument("--prefix-sample-frames", type=int, default=DEFAULT_PREFIX_SAMPLE_FRAMES)
+    parser.add_argument("--prefix-batch-size", type=int, default=DEFAULT_PREFIX_BATCH_SIZE)
+    parser.add_argument(
+        "--adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_true",
+        help="On CUDA OOM in whole mode, retry with smaller frame budgets",
+    )
+    parser.add_argument(
+        "--no-adaptive-max-frames",
+        dest="adaptive_max_frames",
+        action="store_false",
+        help="Disable frame-budget retry on CUDA OOM",
+    )
+    parser.add_argument(
+        "--output-root",
+        type=Path,
+        default=Path(__file__).parent / "outputs_benchmark_style",
+        help="Root directory for outputs",
+    )
+    parser.set_defaults(adaptive_max_frames=True)
+    args = parser.parse_args()
+
+    video_path = Path(args.video).expanduser().resolve()
+    if not video_path.exists():
+        raise FileNotFoundError(f"Video not found: {video_path}")
+
+    output_dir = args.output_root / video_path.stem
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    runner = RobometerLocalRunner(model_path=str(args.model_path))
+
+    print(f"[RUN] {video_path.name}")
+    print(f"[TASK] {args.task}")
+    print(f"[VIDEO] {video_path}")
+    print(f"[OUT] {output_dir}")
+
+    all_frames, native_fps = load_all_video_frames(video_path)
+    total_frames = len(all_frames)
+    retry_schedule = (
+        build_frame_retry_schedule(args.max_frames, args.min_frames, bool(args.adaptive_max_frames))
+        if args.inference_mode == "whole"
+        else [int(args.max_frames)]
+    )
+
+    progress_pred = None
+    success_probs = None
+    sampled_indices = None
+    used_max_frames = retry_schedule[0]
+
+    for attempt_idx, frame_budget in enumerate(retry_schedule, start=1):
+        frames, sampled_indices = sample_video_frames_with_indices(
+            all_frames,
+            native_fps=native_fps,
+            fps=float(args.fps),
+            max_frames=int(frame_budget),
+            required_frames=[],
+        )
+        print(
+            f"Loaded {total_frames} total frames; sampled {len(frames)} frames at fps={float(args.fps):g} "
+            f"(max_frames={int(frame_budget)}, try {attempt_idx}/{len(retry_schedule)})"
+        )
+        try:
+            progress_pred, success_probs = runner.compute_rewards_per_frame(
+                video_frames=frames,
+                task=args.task,
+                inference_mode=args.inference_mode,
+                prefix_sample_frames=int(args.prefix_sample_frames),
+                prefix_batch_size=int(args.prefix_batch_size),
+            )
+            used_max_frames = int(frame_budget)
+            break
+        except RuntimeError as exc:
+            if args.inference_mode != "whole" or not is_cuda_oom_error(exc) or attempt_idx == len(retry_schedule):
+                raise
+            next_budget = retry_schedule[attempt_idx]
+            print(
+                f"[OOM] whole inference hit CUDA OOM at max_frames={int(frame_budget)}; "
+                f"retrying with max_frames={int(next_budget)}"
+            )
+            runner.reload_model()
+
+    if progress_pred is None or success_probs is None or sampled_indices is None:
+        raise RuntimeError("Failed to compute Robometer outputs.")
+    if progress_pred.size == 0:
+        raise RuntimeError("Robometer returned empty progress predictions.")
+    if progress_pred.size != len(sampled_indices):
+        raise RuntimeError(
+            f"Progress length mismatch: got {progress_pred.size} predictions for {len(sampled_indices)} sampled frames"
+        )
+
+    np.save(str(output_dir / "progress.npy"), progress_pred)
+    np.save(str(output_dir / "success_probs.npy"), success_probs)
+
+    payload = {
+        "video": str(video_path),
+        "instruction": args.task,
+        "num_total_frames": total_frames,
+        "native_fps": native_fps,
+        "sample_fps": float(args.fps),
+        "num_sampled_frames": len(sampled_indices),
+        "sampled_original_frame_indices": sampled_indices,
+        "max_frames_used": used_max_frames,
+        "inference_mode": args.inference_mode,
+        "prefix_sample_frames": int(args.prefix_sample_frames),
+        "prefix_batch_size": int(args.prefix_batch_size),
+        "progress_min": float(np.min(progress_pred)),
+        "progress_max": float(np.max(progress_pred)),
+        "progress_mean": float(np.mean(progress_pred)),
+    }
+    save_results(output_dir, payload)
+    write_overview_plot(output_dir, sampled_indices, progress_pred, [])
+    print(f"[OK] Saved results to {output_dir}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/scripts/untar_processed_datasets.sh b/robometer/scripts/untar_processed_datasets.sh
new file mode 100755
index 0000000000000000000000000000000000000000..8dbcca2af573a4a7a32c875995912d6d1b63eaf7
--- /dev/null
+++ b/robometer/scripts/untar_processed_datasets.sh
@@ -0,0 +1,99 @@
+#!/bin/bash
+if [ -z "${ROBOMETER_PROCESSED_DATASETS_PATH:-$RBM_PROCESSED_DATASETS_PATH}" ]; then
+    echo "ROBOMETER_PROCESSED_DATASETS_PATH (or RBM_PROCESSED_DATASETS_PATH) is not set"
+    exit 1
+fi
+
+cd "${ROBOMETER_PROCESSED_DATASETS_PATH:-$RBM_PROCESSED_DATASETS_PATH}" || exit 1
+
+# Track already processed archives to avoid duplicates
+declare -A processed_archives
+
+# First, handle split archives (.tar.partaa, .tar.partab, etc.)
+echo "Processing split archives..."
+for file in *.tar.partaa; do
+    if [ -f "$file" ]; then
+        # Get the base name without .partaa
+        base_name="${file%.partaa}"
+        
+        echo "Extracting split archive: $base_name"
+        # Concatenate all parts and extract
+        cat "${base_name}.part"* | tar -xvf -
+
+        # remove the parts if successfully extracted
+        if [ $? -eq 0 ]; then
+            rm "${base_name}.part"*
+        else
+            echo "Failed to extract $base_name, will need to retry and remove the failed parts"
+            continue
+        fi
+        # Mark this base archive as processed
+        processed_archives["$base_name"]=1
+
+    fi
+done
+
+# Now handle split archives that look like .tar.part-aa, .tar.part-ab, etc.
+echo "Processing split archives..."
+for file in *.tar.part-aa; do
+    if [ -f "$file" ]; then
+        # Get the base name without .part-aa
+        base_name="${file%.part-aa}"
+        echo "Extracting split archive: $base_name"
+        # Concatenate all parts and extract
+        cat "${base_name}.part"* | tar -xvf -
+
+        # remove the parts if successfully extracted
+        if [ $? -eq 0 ]; then
+            rm "${base_name}.part"*
+        else
+            echo "Failed to extract $base_name, will need to retry and remove the failed parts"
+            continue
+        fi
+        # Mark this base archive as processed
+        processed_archives["$base_name"]=1
+    fi
+done
+
+# Then, handle regular tar files (skip those that were split archives)
+echo "Processing regular tar files..."
+for file in *.tar; do
+    if [ -f "$file" ]; then
+        # Skip if this was already processed as a split archive
+        if [ -z "${processed_archives[$file]}" ]; then
+            echo "Extracting: $file"
+            tar -xvf "$file"
+
+            # remove the tar file only if it was successfully extracted
+            if [ $? -eq 0 ]; then
+                processed_archives["$file"]=1
+                rm "$file"
+            else
+                echo "Failed to extract $file, will need to retry and remove the failed tar file"
+                continue
+            fi
+        fi
+    fi
+done
+
+# Now, some of the datasets are moved into a `processed_datasets` subdirectory, so move them out and delete 
+# the overall processed_datasets directory if it exists.
+if [ -d "processed_datasets" ]; then
+    echo "Moving datasets out of processed_datasets subdirectory..."
+    for dir in processed_datasets/*; do
+        if [ -d "$dir" ]; then
+            mv "$dir" .
+        fi
+    done
+    rm -rf processed_datasets
+    echo "Done moving datasets out of processed_datasets subdirectory!"
+fi
+
+# print which datasets might've failed
+for file in *.tar; do
+    if [ -z "${processed_archives[$file]}" ]; then
+        echo "Failed to extract $file"
+    fi
+done
+cd ..
+echo "Done extracting all archives!"
\ No newline at end of file
diff --git a/robometer/setup.py b/robometer/setup.py
new file mode 100755
index 0000000000000000000000000000000000000000..8a18f21c510eb70506c4432e9b7806f084e1a795
--- /dev/null
+++ b/robometer/setup.py
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""
+Setup script for the Robometer package.
+"""
+
+from setuptools import find_packages, setup
+
+# Read the README file
+with open("README.md", encoding="utf-8") as fh:
+    long_description = fh.read()
+
+setup(
+    name="robometer",
+    version="0.0.1",
+    author="Anthony Liang",
+    author_email="aliang80@usc.edu",
+    description="Robometer: Scaling General-purpose Robotic Reward Models via Trajectory Comparisons",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+    url="https://github.com/aliang8/robometer",
+    packages=find_packages(),
+    classifiers=[
+        "Development Status :: 3 - Alpha",
+        "Intended Audience :: Developers",
+        "Intended Audience :: Science/Research",
+        "License :: OSI Approved :: MIT License",
+        "Operating System :: OS Independent",
+        "Programming Language :: Python :: 3",
+        "Programming Language :: Python :: 3.10",
+        "Programming Language :: Python :: 3.11",
+        "Topic :: Scientific/Engineering :: Artificial Intelligence",
+        "Topic :: Software Development :: Libraries :: Python Modules",
+    ],
+    python_requires=">=3.10",
+    entry_points={
+        "console_scripts": [
+            "robometer-train=train:main",
+        ],
+    },
+    include_package_data=True,
+    zip_safe=False,
+)
diff --git a/robometer/train.py b/robometer/train.py
new file mode 100755
index 0000000000000000000000000000000000000000..a761f80eacc45bb29592384c985e85d4d7e0e611
--- /dev/null
+++ b/robometer/train.py
@@ -0,0 +1,358 @@
+import json
+import os
+from dataclasses import asdict
+import shutil
+
+import torch
+import torch.distributed as dist
+import yaml
+from rich import print as rprint
+from rich.panel import Panel
+from omegaconf import DictConfig
+from hydra.core.config_store import ConfigStore
+from hydra import main as hydra_main
+
+from peft import prepare_model_for_kbit_training
+from robometer.configs.experiment_configs import (
+    ExperimentConfig,
+    ModelConfig,
+    PEFTConfig,
+    DataConfig,
+    TrainingConfig,
+    LossConfig,
+    LoggingConfig,
+    SaveBestConfig,
+    CustomEvaluationConfig,
+)
+from robometer.trainers import ReWiNDTrainer, RBMHeadsTrainer
+from robometer.data.datasets.helpers import show_available_datasets
+from robometer.utils.distributed import is_rank_0
+from robometer.utils.logger import rank_0_info
+from robometer.utils.timer import _timer
+from robometer.utils.save import SaveBestCallback, resolve_checkpoint_path, update_cfg_with_pretrained_ckpt
+from robometer.utils.setup_utils import (
+    create_training_arguments,
+    setup_batch_collator,
+    setup_dataset,
+    setup_model_and_processor,
+    setup_peft_model,
+)
+from robometer.data.datasets.base import resolve_dataset_keys
+from robometer.utils.logger import Logger
+from robometer.utils.distributed import banner
+from robometer.utils.config_utils import display_config, convert_hydra_to_dataclass
+import datasets
+
+datasets.logging.set_verbosity_error()
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+torch.autograd.set_detect_anomaly(True)
+
+# Register structured configs with Hydra
+cs = ConfigStore.instance()
+cs.store(name="base_config", node=ExperimentConfig)
+cs.store(group="model", name="model_config", node=ModelConfig)
+cs.store(group="peft", name="peft_config", node=PEFTConfig)
+cs.store(group="data", name="data_config", node=DataConfig)
+cs.store(group="training", name="training_config", node=TrainingConfig)
+cs.store(group="loss", name="loss_config", node=LossConfig)
+cs.store(group="logging", name="logging_config", node=LoggingConfig)
+cs.store(group="logging/save_best", name="save_best_config", node=SaveBestConfig)
+cs.store(group="custom_eval", name="custom_eval_config", node=CustomEvaluationConfig)
+
+
+import torch
+
+torch.set_num_threads(64)
+torch.set_num_interop_threads(8)
+
+
+def train(cfg: ExperimentConfig):
+    timing_raw = {}
+
+    run_name = cfg.training.exp_name
+    if cfg.debug:
+        run_name += "_debug"
+        cfg.training.logging_steps = 1
+        cfg.training.eval_steps = 5
+        # cfg.data.eval_subset_size = 100
+        cfg.training.custom_eval_steps = 5
+        cfg.logging.save_best.save_every = 5
+        cfg.data.dataloader_num_workers = 0
+        cfg.data.dataloader_persistent_workers = False
+
+        # cfg.custom_eval.num_examples_per_quality_pr = 1
+        # cfg.custom_eval.policy_ranking_max_tasks = 10
+
+    # Set memory management
+    torch.backends.cudnn.benchmark = True
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    checkpoint_to_load = cfg.training.load_from_checkpoint or cfg.training.resume_from_checkpoint
+    if checkpoint_to_load:
+        rank_0_info(f"Loading model from checkpoint: {checkpoint_to_load}")
+    update_cfg_with_pretrained_ckpt(cfg, checkpoint_to_load)
+
+    banner("Setting up model and processor")
+    with _timer("time/setup_model_and_processor", timing_raw=timing_raw):
+        tokenizer, processor, rbm_model = setup_model_and_processor(
+            cfg.model,
+            hf_model_id=checkpoint_to_load or "",
+            peft_config=cfg.peft,
+        )
+
+    # Apply PEFT if enabled
+    if cfg.model.use_peft:
+        peft_rbm_model = setup_peft_model(rbm_model, cfg.peft)
+    else:
+        peft_rbm_model = rbm_model
+        rank_0_info("PEFT not enabled, using full model")
+
+    if cfg.model.quantization:
+        peft_rbm_model = prepare_model_for_kbit_training(peft_rbm_model)
+
+    output_dir = os.path.join(cfg.training.output_dir, run_name)
+
+    training_args = create_training_arguments(cfg.training, output_dir)
+
+    # Handle output directory existence (works with accelerate/distributed training)
+    overwrite_output_dir = getattr(cfg.training, "overwrite_output_dir", False)
+
+    # Check if distributed training is initialized (for proper synchronization)
+    # This is important for accelerate/FSDP setups where multiple processes run
+    dist_initialized = dist.is_available() and dist.is_initialized()
+
+    # Check if output directory exists (only on rank 0 to avoid race conditions)
+    if is_rank_0() and os.path.exists(output_dir):
+        if overwrite_output_dir:
+            rank_0_info(f"Output directory {output_dir} already exists. Overwriting (overwrite_output_dir=True)...")
+            shutil.rmtree(output_dir)
+        else:
+            raise ValueError(
+                f"Output directory {output_dir} already exists. "
+                f"Set overwrite_output_dir=True in config to overwrite it, or use a different output directory."
+            )
+
+    # Synchronize all processes before creating directory (important for distributed training)
+    # This ensures rank 0 finishes checking/removing before other processes try to create it
+    if dist_initialized:
+        dist.barrier()
+
+    banner("Creating output directory", f"Logging to: {output_dir}")
+    # Create output directory (all processes need to do this for distributed training)
+    # os.makedirs is safe to call multiple times (exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Synchronize after directory creation to ensure all processes see it
+    if dist_initialized:
+        dist.barrier()
+
+    # Initialize logger (works with wandb/tensorboard)
+    log_to = cfg.logging.log_to
+    log_level = cfg.logging.log_level
+    logger = Logger(log_to=log_to, output_dir=output_dir, is_main_process=is_rank_0(), log_level=log_level)
+    config_save_path = os.path.join(output_dir, "config.yaml")
+    config_dict = asdict(cfg)
+    with open(config_save_path, "w") as f:
+        yaml.dump(config_dict, f, default_flow_style=False, indent=2)
+    rank_0_info(f"Saved training config to: {config_save_path}")
+
+    # Try to load existing wandb info if resuming training
+    wandb_info_path = os.path.join(output_dir, "wandb_info.json")
+    resume_id = None
+    if os.path.exists(wandb_info_path):
+        try:
+            with open(wandb_info_path) as f:
+                wandb_info = json.load(f)
+            resume_id = wandb_info.get("wandb_id")
+            if resume_id:
+                rank_0_info(f"Found existing wandb run ID: {resume_id}, will resume run")
+        except Exception as e:
+            rank_0_info(f"Could not load wandb info: {e}")
+
+    # Initialize wandb via logger if requested
+    if "wandb" in (cfg.logging.log_to or []) and is_rank_0():
+        # Convert config to dict for wandb using dataclass asdict
+        config_dict = asdict(cfg)
+        logger.init_wandb(
+            project=cfg.logging.wandb_project,
+            entity=cfg.logging.wandb_entity,
+            name=run_name,
+            config=config_dict,
+            notes=cfg.logging.wandb_notes,
+            mode=cfg.logging.wandb_mode,
+            resume_id=resume_id,
+        )
+        if resume_id:
+            rank_0_info(f"Wandb resumed run: {run_name} (ID: {resume_id})")
+        else:
+            rank_0_info(f"Wandb initialized: {run_name}")
+        if cfg.logging.wandb_notes:
+            rank_0_info(f"Wandb notes: {cfg.logging.wandb_notes}")
+
+    logger.write_wandb_info(output_dir, run_name)
+
+    # Use the shared utilities for batch collator and dataset
+
+    if is_rank_0():
+        show_available_datasets()
+
+    banner("Resolving dataset keys")
+    cfg.data.train_datasets = resolve_dataset_keys(cfg.data.train_datasets, split="train")
+    rank_0_info(f"Resolved train datasets: {cfg.data.train_datasets}")
+
+    if cfg.data.eval_datasets:
+        cfg.data.eval_datasets = resolve_dataset_keys(cfg.data.eval_datasets, split="eval")
+        rank_0_info(f"Resolved eval datasets: {cfg.data.eval_datasets}")
+
+    # Resolve custom evaluation dataset keys once (replace in place)
+    for eval_type in cfg.custom_eval.eval_types:
+        datasets = getattr(cfg.custom_eval, eval_type, None)
+        if datasets:
+            resolved = resolve_dataset_keys(datasets, split="eval")
+            setattr(cfg.custom_eval, eval_type, resolved)
+            rank_0_info(f"Resolved {eval_type} datasets: {resolved}")
+
+    rank_0_info("Dataset keys resolved")
+
+    banner("Setting up training and evaluation datasets and collator")
+    with _timer("time/setup_data", timing_raw=timing_raw):
+        batch_collator = setup_batch_collator(processor, tokenizer, cfg, is_eval=False)
+        train_dataset = setup_dataset(cfg.data)
+        num_train_samples = len(train_dataset)
+        rank_0_info(f"Training dataset created with {num_train_samples} samples")
+        rank_0_info(f"=" * 100)
+
+    # Set up evaluation dataset if evaluation is enabled
+    eval_dataset = None
+    if cfg.training.do_eval:
+        if cfg.data.eval_subset_size is not None:
+            dataset_kwargs = {"max_samples": cfg.data.eval_subset_size}
+        else:
+            dataset_kwargs = {}
+
+        eval_dataset = setup_dataset(cfg.data, is_eval=True, **dataset_kwargs)
+        num_eval_samples = len(eval_dataset)
+        rank_0_info(f"Evaluation dataset created with {num_eval_samples} samples")
+
+    banner("Setting up trainer", f"Trainer class: {cfg.trainer_cls}")
+    trainer_cls = {
+        "rbm_heads": RBMHeadsTrainer,
+        "rewind_transformer": ReWiNDTrainer,
+        "rewind_scale_transformer": ReWiNDTrainer,
+    }[cfg.trainer_cls]
+
+    # Add SaveBestCallback to automatically save and upload best models
+    save_best_cfg = cfg.logging.save_best
+    save_callback = SaveBestCallback(
+        **asdict(save_best_cfg),
+        base_model=cfg.model.base_model_id,
+    )
+
+    trainer = trainer_cls(
+        model=peft_rbm_model,
+        args=training_args,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        data_collator=batch_collator,
+        config=cfg,
+        logger=logger,
+        callbacks=[save_callback],
+    )
+
+    # Set trainer reference in the callback so it can access trainer methods
+    save_callback.setup_trainer_reference(trainer)
+
+    # Debug: Check if callback was added
+    rank_0_info(f"🔧 DEBUG: Trainer callbacks: {[type(cb).__name__ for cb in trainer.callback_handler.callbacks]}")
+
+    metrics_info = []
+    for name, is_better in zip(save_best_cfg.metric_names, save_best_cfg.greater_is_better):
+        direction = "↗️ higher" if is_better else "↘️ lower"
+        metrics_info.append(f"{name} ({direction})")
+
+    rank_0_info(f"💾 SaveBest monitoring: {', '.join(metrics_info)}")
+    rank_0_info(f"📁 Keeping top {save_best_cfg.keep_top_k} checkpoint(s) and upload(s)")
+
+    if is_rank_0():
+        print("\n" + "=" * 80)
+        print("--- PRE-TRAINING FSDP DIAGNOSTICS ---")
+        # The Trainer creates its own Accelerator instance. Let's check its state.
+        if hasattr(trainer, "accelerator"):
+            print("Trainer's Accelerator object found.")
+            fsdp_plugin = getattr(trainer.accelerator.state, "fsdp_plugin", None)
+            if fsdp_plugin:
+                print("FSDP Plugin found in Accelerator state.")
+                # This is the configuration the accelerator will ACTUALLY use for wrapping.
+                print(f"VERIFY: Actual FSDP plugin config being used: {fsdp_plugin}")
+            else:
+                print("ERROR: FSDP Plugin NOT found in the Trainer's accelerator state!")
+        else:
+            print("ERROR: Trainer has no 'accelerator' attribute yet. This check needs to be later.")
+        print("=" * 80 + "\n")
+
+    # log timing_raw via logger
+    if is_rank_0():
+        logger.log_scalars(timing_raw)
+
+    rank_0_info(f"Timing raw: {timing_raw}")
+
+    # Full resume: restore optimizer state and step counter (load_from_checkpoint only loads weights at setup)
+    hub_token = (save_best_cfg.hub_token if save_best_cfg else None) or os.environ.get("HF_TOKEN")
+    resume_path = (
+        resolve_checkpoint_path(cfg.training.resume_from_checkpoint, hub_token=hub_token)
+        if cfg.training.resume_from_checkpoint
+        else None
+    )
+    if resume_path:
+        rank_0_info(f"Resuming training from checkpoint: {resume_path}")
+    else:
+        rank_0_info("Training from step 0 (no resume)")
+
+    # Restore random state from checkpoint only when doing full resume
+    if resume_path and os.path.isdir(resume_path):
+        random_state_file = os.path.join(resume_path, "dataset_random_state.json")
+        if os.path.exists(random_state_file):
+            try:
+                with open(random_state_file, "r") as f:
+                    random_state = json.load(f)
+                # Handle RepeatedDataset wrapper if present
+                train_dataset = train_dataset.dataset if hasattr(train_dataset, "dataset") else train_dataset
+                if hasattr(train_dataset, "set_random_state"):
+                    train_dataset.set_random_state(random_state)
+                    rank_0_info(f"Restored dataset random state from {random_state_file}")
+                else:
+                    rank_0_info(f"Dataset does not support random state restoration")
+            except Exception as e:
+                rank_0_info(f"Could not restore random state: {e}")
+        else:
+            rank_0_info(f"No dataset_random_state.json found in checkpoint, starting with fresh random state")
+
+    if cfg.debug:
+        rank_0_info("🐛 DEBUG MODE: eval_steps=2, custom_eval_steps=2, eval_subset_size=10")
+
+    trainer.train(resume_from_checkpoint=resume_path)
+    trainer.save_model(cfg.training.output_dir)
+    rank_0_info(f"Training complete! Check {cfg.training.output_dir} for checkpoints and final model.")
+
+
+@hydra_main(version_base=None, config_path="robometer/configs", config_name="config")
+def main(cfg: DictConfig):
+    banner("Starting Robometer Training")
+
+    # Convert Hydra config to dataclass
+    exp_cfg = convert_hydra_to_dataclass(cfg, ExperimentConfig)
+
+    # Display the configuration in a nice Rich format
+    display_config(exp_cfg)
+
+    if exp_cfg.mode == "train":
+        if is_rank_0():
+            rprint(Panel.fit("🚀 Starting Robometer Training", style="bold green"))
+        train(exp_cfg)
+    else:
+        raise ValueError(f"Unknown mode: {exp_cfg.mode}. Must be 'train' or 'evaluate'")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/robometer/uv.lock b/robometer/uv.lock
new file mode 100644
index 0000000000000000000000000000000000000000..01c3e032e4dea53e898d902375c11d01f8642e36
--- /dev/null
+++ b/robometer/uv.lock
@@ -0,0 +1,4679 @@
+version = 1
+revision = 3
+requires-python = "==3.10.*"
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+conflicts = [[
+    { package = "robometer", extra = "robometer" },
+    { package = "robometer", extra = "vlac" },
+]]
+
+[[package]]
+name = "absl-py"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/64/c7/8de93764ad66968d19329a7e0c147a2bb3c7054c554d4a119111b8f9440f/absl_py-2.4.0.tar.gz", hash = "sha256:8c6af82722b35cf71e0f4d1d47dcaebfff286e27110a99fc359349b247dfb5d4", size = 116543, upload-time = "2026-01-28T10:17:05.322Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/18/a6/907a406bb7d359e6a63f99c313846d9eec4f7e6f7437809e03aa00fa3074/absl_py-2.4.0-py3-none-any.whl", hash = "sha256:88476fd881ca8aab94ffa78b7b6c632a782ab3ba1cd19c9bd423abc4fb4cd28d", size = 135750, upload-time = "2026-01-28T10:17:04.19Z" },
+]
+
+[[package]]
+name = "accelerate"
+version = "1.13.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "huggingface-hub" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "psutil" },
+    { name = "pyyaml" },
+    { name = "safetensors" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ca/14/787e5498cd062640f0f3d92ef4ae4063174f76f9afd29d13fc52a319daae/accelerate-1.13.0.tar.gz", hash = "sha256:d631b4e0f5b3de4aff2d7e9e6857d164810dfc3237d54d017f075122d057b236", size = 402835, upload-time = "2026-03-04T19:34:12.359Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7e/46/02ac5e262d4af18054b3e922b2baedbb2a03289ee792162de60a865defc5/accelerate-1.13.0-py3-none-any.whl", hash = "sha256:cf1a3efb96c18f7b152eb0fa7490f3710b19c3f395699358f08decca2b8b62e0", size = 383744, upload-time = "2026-03-04T19:34:10.313Z" },
+]
+
+[[package]]
+name = "addict"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/85/ef/fd7649da8af11d93979831e8f1f8097e85e82d5bfeabc8c68b39175d8e75/addict-2.4.0.tar.gz", hash = "sha256:b3b2210e0e067a281f5646c8c5db92e99b7231ea8b0eb5f74dbdf9e259d4e494", size = 9186, upload-time = "2020-11-21T16:21:31.416Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6a/00/b08f23b7d7e1e14ce01419a467b583edbb93c6cdb8654e54a9cc579cd61f/addict-2.4.0-py3-none-any.whl", hash = "sha256:249bb56bbfd3cdc2a004ea0ff4c2b6ddc84d53bc2194761636eb314d5cfa5dfc", size = 3832, upload-time = "2020-11-21T16:21:29.588Z" },
+]
+
+[[package]]
+name = "aiofiles"
+version = "24.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/0b/03/a88171e277e8caa88a4c77808c20ebb04ba74cc4681bf1e9416c862de237/aiofiles-24.1.0.tar.gz", hash = "sha256:22a075c9e5a3810f0c2e48f3008c94d68c65d763b9b03857924c99e57355166c", size = 30247, upload-time = "2024-06-24T11:02:03.584Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a5/45/30bb92d442636f570cb5651bc661f52b610e2eec3f891a5dc3a4c3667db0/aiofiles-24.1.0-py3-none-any.whl", hash = "sha256:b4ec55f4195e3eb5d7abd1bf7e061763e864dd4954231fb8539a0ef8bb8260e5", size = 15896, upload-time = "2024-06-24T11:02:01.529Z" },
+]
+
+[[package]]
+name = "aiohappyeyeballs"
+version = "2.6.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/26/30/f84a107a9c4331c14b2b586036f40965c128aa4fee4dda5d3d51cb14ad54/aiohappyeyeballs-2.6.1.tar.gz", hash = "sha256:c3f9d0113123803ccadfdf3f0faa505bc78e6a72d1cc4806cbd719826e943558", size = 22760, upload-time = "2025-03-12T01:42:48.764Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0f/15/5bf3b99495fb160b63f95972b81750f18f7f4e02ad051373b669d17d44f2/aiohappyeyeballs-2.6.1-py3-none-any.whl", hash = "sha256:f349ba8f4b75cb25c99c5c2d84e997e485204d2902a9597802b0371f09331fb8", size = 15265, upload-time = "2025-03-12T01:42:47.083Z" },
+]
+
+[[package]]
+name = "aiohttp"
+version = "3.13.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aiohappyeyeballs" },
+    { name = "aiosignal" },
+    { name = "async-timeout" },
+    { name = "attrs" },
+    { name = "frozenlist" },
+    { name = "multidict" },
+    { name = "propcache" },
+    { name = "yarl" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/50/42/32cf8e7704ceb4481406eb87161349abb46a57fee3f008ba9cb610968646/aiohttp-3.13.3.tar.gz", hash = "sha256:a949eee43d3782f2daae4f4a2819b2cb9b0c5d3b7f7a927067cc84dafdbb9f88", size = 7844556, upload-time = "2026-01-03T17:33:05.204Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/36/d6/5aec9313ee6ea9c7cde8b891b69f4ff4001416867104580670a31daeba5b/aiohttp-3.13.3-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d5a372fd5afd301b3a89582817fdcdb6c34124787c70dbcc616f259013e7eef7", size = 738950, upload-time = "2026-01-03T17:29:13.002Z" },
+    { url = "https://files.pythonhosted.org/packages/68/03/8fa90a7e6d11ff20a18837a8e2b5dd23db01aabc475aa9271c8ad33299f5/aiohttp-3.13.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:147e422fd1223005c22b4fe080f5d93ced44460f5f9c105406b753612b587821", size = 496099, upload-time = "2026-01-03T17:29:15.268Z" },
+    { url = "https://files.pythonhosted.org/packages/d2/23/b81f744d402510a8366b74eb420fc0cc1170d0c43daca12d10814df85f10/aiohttp-3.13.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:859bd3f2156e81dd01432f5849fc73e2243d4a487c4fd26609b1299534ee1845", size = 491072, upload-time = "2026-01-03T17:29:16.922Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/e1/56d1d1c0dd334cd203dd97706ce004c1aa24b34a813b0b8daf3383039706/aiohttp-3.13.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dca68018bf48c251ba17c72ed479f4dafe9dbd5a73707ad8d28a38d11f3d42af", size = 1671588, upload-time = "2026-01-03T17:29:18.539Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/34/8d7f962604f4bc2b4e39eb1220dac7d4e4cba91fb9ba0474b4ecd67db165/aiohttp-3.13.3-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:fee0c6bc7db1de362252affec009707a17478a00ec69f797d23ca256e36d5940", size = 1640334, upload-time = "2026-01-03T17:29:21.028Z" },
+    { url = "https://files.pythonhosted.org/packages/94/1d/fcccf2c668d87337ddeef9881537baee13c58d8f01f12ba8a24215f2b804/aiohttp-3.13.3-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c048058117fd649334d81b4b526e94bde3ccaddb20463a815ced6ecbb7d11160", size = 1722656, upload-time = "2026-01-03T17:29:22.531Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/98/c6f3b081c4c606bc1e5f2ec102e87d6411c73a9ef3616fea6f2d5c98c062/aiohttp-3.13.3-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:215a685b6fbbfcf71dfe96e3eba7a6f58f10da1dfdf4889c7dd856abe430dca7", size = 1817625, upload-time = "2026-01-03T17:29:24.276Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/c0/cfcc3d2e11b477f86e1af2863f3858c8850d751ce8dc39c4058a072c9e54/aiohttp-3.13.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:de2c184bb1fe2cbd2cefba613e9db29a5ab559323f994b6737e370d3da0ac455", size = 1672604, upload-time = "2026-01-03T17:29:26.099Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/77/6b4ffcbcac4c6a5d041343a756f34a6dd26174ae07f977a64fe028dda5b0/aiohttp-3.13.3-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:75ca857eba4e20ce9f546cd59c7007b33906a4cd48f2ff6ccf1ccfc3b646f279", size = 1554370, upload-time = "2026-01-03T17:29:28.121Z" },
+    { url = "https://files.pythonhosted.org/packages/f2/f0/e3ddfa93f17d689dbe014ba048f18e0c9f9b456033b70e94349a2e9048be/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:81e97251d9298386c2b7dbeb490d3d1badbdc69107fb8c9299dd04eb39bddc0e", size = 1642023, upload-time = "2026-01-03T17:29:30.002Z" },
+    { url = "https://files.pythonhosted.org/packages/eb/45/c14019c9ec60a8e243d06d601b33dcc4fd92379424bde3021725859d7f99/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:c0e2d366af265797506f0283487223146af57815b388623f0357ef7eac9b209d", size = 1649680, upload-time = "2026-01-03T17:29:31.782Z" },
+    { url = "https://files.pythonhosted.org/packages/9c/fd/09c9451dae5aa5c5ed756df95ff9ef549d45d4be663bafd1e4954fd836f0/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:4e239d501f73d6db1522599e14b9b321a7e3b1de66ce33d53a765d975e9f4808", size = 1692407, upload-time = "2026-01-03T17:29:33.392Z" },
+    { url = "https://files.pythonhosted.org/packages/a6/81/938bc2ec33c10efd6637ccb3d22f9f3160d08e8f3aa2587a2c2d5ab578eb/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:0db318f7a6f065d84cb1e02662c526294450b314a02bd9e2a8e67f0d8564ce40", size = 1543047, upload-time = "2026-01-03T17:29:34.855Z" },
+    { url = "https://files.pythonhosted.org/packages/f7/23/80488ee21c8d567c83045e412e1d9b7077d27171591a4eb7822586e8c06a/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:bfc1cc2fe31a6026a8a88e4ecfb98d7f6b1fec150cfd708adbfd1d2f42257c29", size = 1715264, upload-time = "2026-01-03T17:29:36.389Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/83/259a8da6683182768200b368120ab3deff5370bed93880fb9a3a86299f34/aiohttp-3.13.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:af71fff7bac6bb7508956696dce8f6eec2bbb045eceb40343944b1ae62b5ef11", size = 1657275, upload-time = "2026-01-03T17:29:38.162Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/4f/2c41f800a0b560785c10fb316216ac058c105f9be50bdc6a285de88db625/aiohttp-3.13.3-cp310-cp310-win32.whl", hash = "sha256:37da61e244d1749798c151421602884db5270faf479cf0ef03af0ff68954c9dd", size = 434053, upload-time = "2026-01-03T17:29:40.074Z" },
+    { url = "https://files.pythonhosted.org/packages/80/df/29cd63c7ecfdb65ccc12f7d808cac4fa2a19544660c06c61a4a48462de0c/aiohttp-3.13.3-cp310-cp310-win_amd64.whl", hash = "sha256:7e63f210bc1b57ef699035f2b4b6d9ce096b5914414a49b0997c839b2bd2223c", size = 456687, upload-time = "2026-01-03T17:29:41.819Z" },
+]
+
+[[package]]
+name = "aiosignal"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "frozenlist" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fb/76/641ae371508676492379f16e2fa48f4e2c11741bd63c48be4b12a6b09cba/aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e", size = 7490, upload-time = "2025-07-03T22:54:42.156Z" },
+]
+
+[[package]]
+name = "aliyun-python-sdk-core"
+version = "2.16.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cryptography" },
+    { name = "jmespath" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3e/09/da9f58eb38b4fdb97ba6523274fbf445ef6a06be64b433693da8307b4bec/aliyun-python-sdk-core-2.16.0.tar.gz", hash = "sha256:651caad597eb39d4fad6cf85133dffe92837d53bdf62db9d8f37dab6508bb8f9", size = 449555, upload-time = "2024-10-09T06:01:01.762Z" }
+
+[[package]]
+name = "aliyun-python-sdk-kms"
+version = "2.16.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aliyun-python-sdk-core" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a8/2c/9877d0e6b18ecf246df671ac65a5d1d9fecbf85bdcb5d43efbde0d4662eb/aliyun-python-sdk-kms-2.16.5.tar.gz", hash = "sha256:f328a8a19d83ecbb965ffce0ec1e9930755216d104638cd95ecd362753b813b3", size = 12018, upload-time = "2024-08-30T09:01:20.104Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/11/5c/0132193d7da2c735669a1ed103b142fd63c9455984d48c5a88a1a516efaa/aliyun_python_sdk_kms-2.16.5-py2.py3-none-any.whl", hash = "sha256:24b6cdc4fd161d2942619479c8d050c63ea9cd22b044fe33b60bbb60153786f0", size = 99495, upload-time = "2024-08-30T09:01:18.462Z" },
+]
+
+[[package]]
+name = "annotated-doc"
+version = "0.0.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/57/ba/046ceea27344560984e26a590f90bc7f4a75b06701f653222458922b558c/annotated_doc-0.0.4.tar.gz", hash = "sha256:fbcda96e87e9c92ad167c2e53839e57503ecfda18804ea28102353485033faa4", size = 7288, upload-time = "2025-11-10T22:07:42.062Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1e/d3/26bf1008eb3d2daa8ef4cacc7f3bfdc11818d111f7e2d0201bc6e3b49d45/annotated_doc-0.0.4-py3-none-any.whl", hash = "sha256:571ac1dc6991c450b25a9c2d84a3705e2ae7a53467b5d111c24fa8baabbed320", size = 5303, upload-time = "2025-11-10T22:07:40.673Z" },
+]
+
+[[package]]
+name = "annotated-types"
+version = "0.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ee/67/531ea369ba64dcff5ec9c3402f9f51bf748cec26dde048a2f973a4eea7f5/annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89", size = 16081, upload-time = "2024-05-20T21:33:25.928Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/78/b6/6307fbef88d9b5ee7421e68d78a9f162e0da4900bc5f5793f6d3d0e34fb8/annotated_types-0.7.0-py3-none-any.whl", hash = "sha256:1f02e8b43a8fbbc3f3e0d4f0f4bfc8131bcb4eebe8849b8e5c773f3a1c582a53", size = 13643, upload-time = "2024-05-20T21:33:24.1Z" },
+]
+
+[[package]]
+name = "antlr4-python3-runtime"
+version = "4.9.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/3e/38/7859ff46355f76f8d19459005ca000b6e7012f2f1ca597746cbcd1fbfe5e/antlr4-python3-runtime-4.9.3.tar.gz", hash = "sha256:f224469b4168294902bb1efa80a8bf7855f24c99aef99cbefc1bcd3cce77881b", size = 117034, upload-time = "2021-11-06T17:52:23.524Z" }
+
+[[package]]
+name = "anyio"
+version = "4.12.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "exceptiongroup" },
+    { name = "idna" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/96/f0/5eb65b2bb0d09ac6776f2eb54adee6abe8228ea05b20a5ad0e4945de8aac/anyio-4.12.1.tar.gz", hash = "sha256:41cfcc3a4c85d3f05c932da7c26d0201ac36f72abd4435ba90d0464a3ffed703", size = 228685, upload-time = "2026-01-06T11:45:21.246Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/38/0e/27be9fdef66e72d64c0cdc3cc2823101b80585f8119b5c112c2e8f5f7dab/anyio-4.12.1-py3-none-any.whl", hash = "sha256:d405828884fc140aa80a3c667b8beed277f1dfedec42ba031bd6ac3db606ab6c", size = 113592, upload-time = "2026-01-06T11:45:19.497Z" },
+]
+
+[[package]]
+name = "appnope"
+version = "0.1.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/35/5d/752690df9ef5b76e169e68d6a129fa6d08a7100ca7f754c89495db3c6019/appnope-0.1.4.tar.gz", hash = "sha256:1de3860566df9caf38f01f86f65e0e13e379af54f9e4bee1e66b48f2efffd1ee", size = 4170, upload-time = "2024-02-06T09:43:11.258Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/81/29/5ecc3a15d5a33e31b26c11426c45c501e439cb865d0bff96315d86443b78/appnope-0.1.4-py2.py3-none-any.whl", hash = "sha256:502575ee11cd7a28c0205f379b525beefebab9d161b7c964670864014ed7213c", size = 4321, upload-time = "2024-02-06T09:43:09.663Z" },
+]
+
+[[package]]
+name = "array-record"
+version = "0.8.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "etils", extra = ["epath"] },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2c/14/32623b8d758f86febb63d0a19871f09a06d659485a54cd77a266045e0ec5/array_record-0.8.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:1bf3fcf0a23591667dbd1fe3d32983e8c864781cef5bfd070885ebd3e39e4594", size = 1965586, upload-time = "2025-08-21T19:06:13.699Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/1d/25333484f9c7fc92dcfce8ca77f2b5d9a57b2678200ffc391685d3aee3fa/array_record-0.8.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:39a69d8b8259bce043a8218f2defd46eb44cba0c2f60b7254fd1daaf26e0ea67", size = 2406222, upload-time = "2025-08-21T19:06:16.015Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/94/03b64645354537fe54d4647f86ae91f9625fec330210ed1dceebf22617fa/array_record-0.8.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:b63f02209bdd3acbb175248d04e0f3f827950eae72f010e6ec14e2732aaf008e", size = 2463811, upload-time = "2025-08-21T19:06:17.143Z" },
+]
+
+[[package]]
+name = "asttokens"
+version = "3.0.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/be/a5/8e3f9b6771b0b408517c82d97aed8f2036509bc247d46114925e32fe33f0/asttokens-3.0.1.tar.gz", hash = "sha256:71a4ee5de0bde6a31d64f6b13f2293ac190344478f081c3d1bccfcf5eacb0cb7", size = 62308, upload-time = "2025-11-15T16:43:48.578Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d2/39/e7eaf1799466a4aef85b6a4fe7bd175ad2b1c6345066aa33f1f58d4b18d0/asttokens-3.0.1-py3-none-any.whl", hash = "sha256:15a3ebc0f43c2d0a50eeafea25e19046c68398e487b9f1f5b517f7c0f40f976a", size = 27047, upload-time = "2025-11-15T16:43:16.109Z" },
+]
+
+[[package]]
+name = "astunparse"
+version = "1.6.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+    { name = "wheel" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f3/af/4182184d3c338792894f34a62672919db7ca008c89abee9b564dd34d8029/astunparse-1.6.3.tar.gz", hash = "sha256:5ad93a8456f0d084c3456d059fd9a92cce667963232cbf763eac3bc5b7940872", size = 18290, upload-time = "2019-12-22T18:12:13.129Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2b/03/13dde6512ad7b4557eb792fbcf0c653af6076b81e5941d36ec61f7ce6028/astunparse-1.6.3-py2.py3-none-any.whl", hash = "sha256:c2652417f2c8b5bb325c885ae329bdf3f86424075c4fd1a128674bc6fba4b8e8", size = 12732, upload-time = "2019-12-22T18:12:11.297Z" },
+]
+
+[[package]]
+name = "async-timeout"
+version = "5.0.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a5/ae/136395dfbfe00dfc94da3f3e136d0b13f394cba8f4841120e34226265780/async_timeout-5.0.1.tar.gz", hash = "sha256:d9321a7a3d5a6a5e187e824d2fa0793ce379a202935782d555d6e9d2735677d3", size = 9274, upload-time = "2024-11-06T16:41:39.6Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fe/ba/e2081de779ca30d473f21f5b30e0e737c438205440784c7dfc81efc2b029/async_timeout-5.0.1-py3-none-any.whl", hash = "sha256:39e3809566ff85354557ec2398b55e096c8364bacac9405a7a1fa429e77fe76c", size = 6233, upload-time = "2024-11-06T16:41:37.9Z" },
+]
+
+[[package]]
+name = "attrdict"
+version = "2.0.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3f/72/614aae677d28e81a5bf830fadcf580803876ef76e0306902d3ca5790cd9a/attrdict-2.0.1.tar.gz", hash = "sha256:35c90698b55c683946091177177a9e9c0713a0860f0e049febd72649ccd77b70", size = 9593, upload-time = "2019-02-01T22:33:58.493Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ef/97/28fe7e68bc7adfce67d4339756e85e9fcf3c6fd7f0c0781695352b70472c/attrdict-2.0.1-py2.py3-none-any.whl", hash = "sha256:9432e3498c74ff7e1b20b3d93b45d766b71cbffa90923496f82c4ae38b92be34", size = 9946, upload-time = "2019-02-01T22:33:56.508Z" },
+]
+
+[[package]]
+name = "attrs"
+version = "25.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6b/5c/685e6633917e101e5dcb62b9dd76946cbb57c26e133bae9e0cd36033c0a9/attrs-25.4.0.tar.gz", hash = "sha256:16d5969b87f0859ef33a48b35d55ac1be6e42ae49d5e853b597db70c35c57e11", size = 934251, upload-time = "2025-10-06T13:54:44.725Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3a/2a/7cc015f5b9f5db42b7d48157e23356022889fc354a2813c15934b7cb5c0e/attrs-25.4.0-py3-none-any.whl", hash = "sha256:adcf7e2a1fb3b36ac48d97835bb6d8ade15b8dcce26aba8bf1d14847b57a3373", size = 67615, upload-time = "2025-10-06T13:54:43.17Z" },
+]
+
+[[package]]
+name = "av"
+version = "16.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/78/cd/3a83ffbc3cc25b39721d174487fb0d51a76582f4a1703f98e46170ce83d4/av-16.1.0.tar.gz", hash = "sha256:a094b4fd87a3721dacf02794d3d2c82b8d712c85b9534437e82a8a978c175ffd", size = 4285203, upload-time = "2026-01-11T07:31:33.772Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/97/51/2217a9249409d2e88e16e3f16f7c0def9fd3e7ffc4238b2ec211f9935bdb/av-16.1.0-cp310-cp310-macosx_11_0_x86_64.whl", hash = "sha256:2395748b0c34fe3a150a1721e4f3d4487b939520991b13e7b36f8926b3b12295", size = 26942590, upload-time = "2026-01-09T20:17:58.588Z" },
+    { url = "https://files.pythonhosted.org/packages/bf/cd/a7070f4febc76a327c38808e01e2ff6b94531fe0b321af54ea3915165338/av-16.1.0-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:72d7ac832710a158eeb7a93242370aa024a7646516291c562ee7f14a7ea881fd", size = 21507910, upload-time = "2026-01-09T20:18:02.309Z" },
+    { url = "https://files.pythonhosted.org/packages/ae/30/ec812418cd9b297f0238fe20eb0747d8a8b68d82c5f73c56fe519a274143/av-16.1.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:6cbac833092e66b6b0ac4d81ab077970b8ca874951e9c3974d41d922aaa653ed", size = 38738309, upload-time = "2026-01-09T20:18:04.701Z" },
+    { url = "https://files.pythonhosted.org/packages/3a/b8/6c5795bf1f05f45c5261f8bce6154e0e5e86b158a6676650ddd77c28805e/av-16.1.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:eb990672d97c18f99c02f31c8d5750236f770ffe354b5a52c5f4d16c5e65f619", size = 40293006, upload-time = "2026-01-09T20:18:07.238Z" },
+    { url = "https://files.pythonhosted.org/packages/a7/44/5e183bcb9333fc3372ee6e683be8b0c9b515a506894b2d32ff465430c074/av-16.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:05ad70933ac3b8ef896a820ea64b33b6cca91a5fac5259cb9ba7fa010435be15", size = 40123516, upload-time = "2026-01-09T20:18:09.955Z" },
+    { url = "https://files.pythonhosted.org/packages/12/1d/b5346d582a3c3d958b4d26a2cc63ce607233582d956121eb20d2bbe55c2e/av-16.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:d831a1062a3c47520bf99de6ec682bd1d64a40dfa958e5457bb613c5270e7ce3", size = 41463289, upload-time = "2026-01-09T20:18:12.459Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/31/acc946c0545f72b8d0d74584cb2a0ade9b7dfe2190af3ef9aa52a2e3c0b1/av-16.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:358ab910fef3c5a806c55176f2b27e5663b33c4d0a692dafeb049c6ed71f8aff", size = 31754959, upload-time = "2026-01-09T20:18:14.718Z" },
+]
+
+[[package]]
+name = "binpacking"
+version = "2.0.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/20/6b/2fd636bdb0e66a5bf2f87244d15065a4afad387cfe103786f57c143d85b2/binpacking-2.0.1.tar.gz", hash = "sha256:c3e95c6cdc661d3d00e48da5f317a111531267069f2b4a288d74401b83779ffd", size = 23715, upload-time = "2026-01-22T10:27:36.763Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/53/89/665702cce9192ef6db2dc469e4933acb3695cc20c69bbfbf00a5b18ed1f3/binpacking-2.0.1-py3-none-any.whl", hash = "sha256:f274a60486543e2def108d05e7ae81a83e045d7bf872ae830f1582629e53ebdf", size = 28291, upload-time = "2026-01-22T10:27:35.202Z" },
+]
+
+[[package]]
+name = "bitsandbytes"
+version = "0.49.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d8/7d/f1fe0992334b18cd8494f89aeec1dcc674635584fcd9f115784fea3a1d05/bitsandbytes-0.49.2-py3-none-macosx_14_0_arm64.whl", hash = "sha256:87be5975edeac5396d699ecbc39dfc47cf2c026daaf2d5852a94368611a6823f", size = 131940, upload-time = "2026-02-16T21:26:04.572Z" },
+    { url = "https://files.pythonhosted.org/packages/29/71/acff7af06c818664aa87ff73e17a52c7788ad746b72aea09d3cb8e424348/bitsandbytes-0.49.2-py3-none-manylinux_2_24_aarch64.whl", hash = "sha256:2fc0830c5f7169be36e60e11f2be067c8f812dfcb829801a8703735842450750", size = 31442815, upload-time = "2026-02-16T21:26:06.783Z" },
+    { url = "https://files.pythonhosted.org/packages/19/57/3443d6f183436fbdaf5000aac332c4d5ddb056665d459244a5608e98ae92/bitsandbytes-0.49.2-py3-none-manylinux_2_24_x86_64.whl", hash = "sha256:54b771f06e1a3c73af5c7f16ccf0fc23a846052813d4b008d10cb6e017dd1c8c", size = 60651714, upload-time = "2026-02-16T21:26:11.579Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/d4/501655842ad6771fb077f576d78cbedb5445d15b1c3c91343ed58ca46f0e/bitsandbytes-0.49.2-py3-none-win_amd64.whl", hash = "sha256:2e0ddd09cd778155388023cbe81f00afbb7c000c214caef3ce83386e7144df7d", size = 55372289, upload-time = "2026-02-16T21:26:16.267Z" },
+]
+
+[[package]]
+name = "bokeh"
+version = "3.8.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "contourpy" },
+    { name = "jinja2" },
+    { name = "narwhals" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "pandas" },
+    { name = "pillow" },
+    { name = "pyyaml" },
+    { name = "tornado", marker = "sys_platform != 'emscripten' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "xyzservices" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e4/31/7ee0c4dfd0255631b0624ce01be178704f91f763f02a1879368eb109befd/bokeh-3.8.2.tar.gz", hash = "sha256:8e7dcacc21d53905581b54328ad2705954f72f2997f99fc332c1de8da53aa3cc", size = 6529251, upload-time = "2026-01-06T00:20:06.568Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f6/a8/877f306720bc114c612579c5af36bcb359026b83d051226945499b306b1a/bokeh-3.8.2-py3-none-any.whl", hash = "sha256:5e2c0d84f75acb25d60efb9e4d2f434a791c4639b47d685534194c4e07bd0111", size = 7207131, upload-time = "2026-01-06T00:20:04.917Z" },
+]
+
+[[package]]
+name = "brotli"
+version = "1.2.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f7/16/c92ca344d646e71a43b8bb353f0a6490d7f6e06210f8554c8f874e454285/brotli-1.2.0.tar.gz", hash = "sha256:e310f77e41941c13340a95976fe66a8a95b01e783d430eeaf7a2f87e0a57dd0a", size = 7388632, upload-time = "2025-11-05T18:39:42.86Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/64/10/a090475284fc4a71aed40a96f32e44a7fe5bda39687353dd977720b211b6/brotli-1.2.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:3b90b767916ac44e93a8e28ce6adf8d551e43affb512f2377c732d486ac6514e", size = 863089, upload-time = "2025-11-05T18:38:01.181Z" },
+    { url = "https://files.pythonhosted.org/packages/03/41/17416630e46c07ac21e378c3464815dd2e120b441e641bc516ac32cc51d2/brotli-1.2.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:6be67c19e0b0c56365c6a76e393b932fb0e78b3b56b711d180dd7013cb1fd984", size = 445442, upload-time = "2025-11-05T18:38:02.434Z" },
+    { url = "https://files.pythonhosted.org/packages/24/31/90cc06584deb5d4fcafc0985e37741fc6b9717926a78674bbb3ce018957e/brotli-1.2.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0bbd5b5ccd157ae7913750476d48099aaf507a79841c0d04a9db4415b14842de", size = 1532658, upload-time = "2025-11-05T18:38:03.588Z" },
+    { url = "https://files.pythonhosted.org/packages/62/17/33bf0c83bcbc96756dfd712201d87342732fad70bb3472c27e833a44a4f9/brotli-1.2.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3f3c908bcc404c90c77d5a073e55271a0a498f4e0756e48127c35d91cf155947", size = 1631241, upload-time = "2025-11-05T18:38:04.582Z" },
+    { url = "https://files.pythonhosted.org/packages/48/10/f47854a1917b62efe29bc98ac18e5d4f71df03f629184575b862ef2e743b/brotli-1.2.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1b557b29782a643420e08d75aea889462a4a8796e9a6cf5621ab05a3f7da8ef2", size = 1424307, upload-time = "2025-11-05T18:38:05.587Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/b7/f88eb461719259c17483484ea8456925ee057897f8e64487d76e24e5e38d/brotli-1.2.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:81da1b229b1889f25adadc929aeb9dbc4e922bd18561b65b08dd9343cfccca84", size = 1488208, upload-time = "2025-11-05T18:38:06.613Z" },
+    { url = "https://files.pythonhosted.org/packages/26/59/41bbcb983a0c48b0b8004203e74706c6b6e99a04f3c7ca6f4f41f364db50/brotli-1.2.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:ff09cd8c5eec3b9d02d2408db41be150d8891c5566addce57513bf546e3d6c6d", size = 1597574, upload-time = "2025-11-05T18:38:07.838Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/e6/8c89c3bdabbe802febb4c5c6ca224a395e97913b5df0dff11b54f23c1788/brotli-1.2.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:a1778532b978d2536e79c05dac2d8cd857f6c55cd0c95ace5b03740824e0e2f1", size = 1492109, upload-time = "2025-11-05T18:38:08.816Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/9a/4b19d4310b2dbd545c0c33f176b0528fa68c3cd0754e34b2f2bcf56548ae/brotli-1.2.0-cp310-cp310-win32.whl", hash = "sha256:b232029d100d393ae3c603c8ffd7e3fe6f798c5e28ddca5feabb8e8fdb732997", size = 334461, upload-time = "2025-11-05T18:38:10.729Z" },
+    { url = "https://files.pythonhosted.org/packages/ac/39/70981d9f47705e3c2b95c0847dfa3e7a37aa3b7c6030aedc4873081ed005/brotli-1.2.0-cp310-cp310-win_amd64.whl", hash = "sha256:ef87b8ab2704da227e83a246356a2b179ef826f550f794b2c52cddb4efbd0196", size = 369035, upload-time = "2025-11-05T18:38:11.827Z" },
+]
+
+[[package]]
+name = "certifi"
+version = "2026.2.25"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/af/2d/7bf41579a8986e348fa033a31cdd0e4121114f6bce2457e8876010b092dd/certifi-2026.2.25.tar.gz", hash = "sha256:e887ab5cee78ea814d3472169153c2d12cd43b14bd03329a39a9c6e2e80bfba7", size = 155029, upload-time = "2026-02-25T02:54:17.342Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9a/3c/c17fb3ca2d9c3acff52e30b309f538586f9f5b9c9cf454f3845fc9af4881/certifi-2026.2.25-py3-none-any.whl", hash = "sha256:027692e4402ad994f1c42e52a4997a9763c646b73e4096e4d5d6db8af1d6f0fa", size = 153684, upload-time = "2026-02-25T02:54:15.766Z" },
+]
+
+[[package]]
+name = "cffi"
+version = "2.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pycparser", marker = "implementation_name != 'PyPy' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/eb/56/b1ba7935a17738ae8453301356628e8147c79dbb825bcbc73dc7401f9846/cffi-2.0.0.tar.gz", hash = "sha256:44d1b5909021139fe36001ae048dbdde8214afa20200eda0f64c068cac5d5529", size = 523588, upload-time = "2025-09-08T23:24:04.541Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/93/d7/516d984057745a6cd96575eea814fe1edd6646ee6efd552fb7b0921dec83/cffi-2.0.0-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:0cf2d91ecc3fcc0625c2c530fe004f82c110405f101548512cce44322fa8ac44", size = 184283, upload-time = "2025-09-08T23:22:08.01Z" },
+    { url = "https://files.pythonhosted.org/packages/9e/84/ad6a0b408daa859246f57c03efd28e5dd1b33c21737c2db84cae8c237aa5/cffi-2.0.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:f73b96c41e3b2adedc34a7356e64c8eb96e03a3782b535e043a986276ce12a49", size = 180504, upload-time = "2025-09-08T23:22:10.637Z" },
+    { url = "https://files.pythonhosted.org/packages/50/bd/b1a6362b80628111e6653c961f987faa55262b4002fcec42308cad1db680/cffi-2.0.0-cp310-cp310-manylinux1_i686.manylinux2014_i686.manylinux_2_17_i686.manylinux_2_5_i686.whl", hash = "sha256:53f77cbe57044e88bbd5ed26ac1d0514d2acf0591dd6bb02a3ae37f76811b80c", size = 208811, upload-time = "2025-09-08T23:22:12.267Z" },
+    { url = "https://files.pythonhosted.org/packages/4f/27/6933a8b2562d7bd1fb595074cf99cc81fc3789f6a6c05cdabb46284a3188/cffi-2.0.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:3e837e369566884707ddaf85fc1744b47575005c0a229de3327f8f9a20f4efeb", size = 216402, upload-time = "2025-09-08T23:22:13.455Z" },
+    { url = "https://files.pythonhosted.org/packages/05/eb/b86f2a2645b62adcfff53b0dd97e8dfafb5c8aa864bd0d9a2c2049a0d551/cffi-2.0.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:5eda85d6d1879e692d546a078b44251cdd08dd1cfb98dfb77b670c97cee49ea0", size = 203217, upload-time = "2025-09-08T23:22:14.596Z" },
+    { url = "https://files.pythonhosted.org/packages/9f/e0/6cbe77a53acf5acc7c08cc186c9928864bd7c005f9efd0d126884858a5fe/cffi-2.0.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:9332088d75dc3241c702d852d4671613136d90fa6881da7d770a483fd05248b4", size = 203079, upload-time = "2025-09-08T23:22:15.769Z" },
+    { url = "https://files.pythonhosted.org/packages/98/29/9b366e70e243eb3d14a5cb488dfd3a0b6b2f1fb001a203f653b93ccfac88/cffi-2.0.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:fc7de24befaeae77ba923797c7c87834c73648a05a4bde34b3b7e5588973a453", size = 216475, upload-time = "2025-09-08T23:22:17.427Z" },
+    { url = "https://files.pythonhosted.org/packages/21/7a/13b24e70d2f90a322f2900c5d8e1f14fa7e2a6b3332b7309ba7b2ba51a5a/cffi-2.0.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:cf364028c016c03078a23b503f02058f1814320a56ad535686f90565636a9495", size = 218829, upload-time = "2025-09-08T23:22:19.069Z" },
+    { url = "https://files.pythonhosted.org/packages/60/99/c9dc110974c59cc981b1f5b66e1d8af8af764e00f0293266824d9c4254bc/cffi-2.0.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:e11e82b744887154b182fd3e7e8512418446501191994dbf9c9fc1f32cc8efd5", size = 211211, upload-time = "2025-09-08T23:22:20.588Z" },
+    { url = "https://files.pythonhosted.org/packages/49/72/ff2d12dbf21aca1b32a40ed792ee6b40f6dc3a9cf1644bd7ef6e95e0ac5e/cffi-2.0.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:8ea985900c5c95ce9db1745f7933eeef5d314f0565b27625d9a10ec9881e1bfb", size = 218036, upload-time = "2025-09-08T23:22:22.143Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/cc/027d7fb82e58c48ea717149b03bcadcbdc293553edb283af792bd4bcbb3f/cffi-2.0.0-cp310-cp310-win32.whl", hash = "sha256:1f72fb8906754ac8a2cc3f9f5aaa298070652a0ffae577e0ea9bd480dc3c931a", size = 172184, upload-time = "2025-09-08T23:22:23.328Z" },
+    { url = "https://files.pythonhosted.org/packages/33/fa/072dd15ae27fbb4e06b437eb6e944e75b068deb09e2a2826039e49ee2045/cffi-2.0.0-cp310-cp310-win_amd64.whl", hash = "sha256:b18a3ed7d5b3bd8d9ef7a8cb226502c6bf8308df1525e1cc676c3680e7176739", size = 182790, upload-time = "2025-09-08T23:22:24.752Z" },
+]
+
+[[package]]
+name = "cfgv"
+version = "3.5.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/4e/b5/721b8799b04bf9afe054a3899c6cf4e880fcf8563cc71c15610242490a0c/cfgv-3.5.0.tar.gz", hash = "sha256:d5b1034354820651caa73ede66a6294d6e95c1b00acc5e9b098e917404669132", size = 7334, upload-time = "2025-11-19T20:55:51.612Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/db/3c/33bac158f8ab7f89b2e59426d5fe2e4f63f7ed25df84c036890172b412b5/cfgv-3.5.0-py2.py3-none-any.whl", hash = "sha256:a8dc6b26ad22ff227d2634a65cb388215ce6cc96bbcc5cfde7641ae87e8dacc0", size = 7445, upload-time = "2025-11-19T20:55:50.744Z" },
+]
+
+[[package]]
+name = "charset-normalizer"
+version = "3.4.5"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1d/35/02daf95b9cd686320bb622eb148792655c9412dbb9b67abb5694e5910a24/charset_normalizer-3.4.5.tar.gz", hash = "sha256:95adae7b6c42a6c5b5b559b1a99149f090a57128155daeea91732c8d970d8644", size = 134804, upload-time = "2026-03-06T06:03:19.46Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a7/21/a2b1505639008ba2e6ef03733a81fc6cfd6a07ea6139a2b76421230b8dad/charset_normalizer-3.4.5-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:4167a621a9a1a986c73777dbc15d4b5eac8ac5c10393374109a343d4013ec765", size = 283319, upload-time = "2026-03-06T06:00:26.433Z" },
+    { url = "https://files.pythonhosted.org/packages/70/67/df234c29b68f4e1e095885c9db1cb4b69b8aba49cf94fac041db4aaf1267/charset_normalizer-3.4.5-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3f64c6bf8f32f9133b668c7f7a7cbdbc453412bc95ecdbd157f3b1e377a92990", size = 189974, upload-time = "2026-03-06T06:00:28.222Z" },
+    { url = "https://files.pythonhosted.org/packages/df/7f/fc66af802961c6be42e2c7b69c58f95cbd1f39b0e81b3365d8efe2a02a04/charset_normalizer-3.4.5-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:568e3c34b58422075a1b49575a6abc616d9751b4d61b23f712e12ebb78fe47b2", size = 207866, upload-time = "2026-03-06T06:00:29.769Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/23/404eb36fac4e95b833c50e305bba9a241086d427bb2167a42eac7c4f7da4/charset_normalizer-3.4.5-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:036c079aa08a6a592b82487f97c60b439428320ed1b2ea0b3912e99d30c77765", size = 203239, upload-time = "2026-03-06T06:00:31.086Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/2f/8a1d989bfadd120c90114ab33e0d2a0cbde05278c1fc15e83e62d570f50a/charset_normalizer-3.4.5-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:340810d34ef83af92148e96e3e44cb2d3f910d2bf95e5618a5c467d9f102231d", size = 196529, upload-time = "2026-03-06T06:00:32.608Z" },
+    { url = "https://files.pythonhosted.org/packages/a5/0c/c75f85ff7ca1f051958bb518cd43922d86f576c03947a050fbedfdfb4f15/charset_normalizer-3.4.5-cp310-cp310-manylinux_2_31_armv7l.whl", hash = "sha256:cd2d0f0ec9aa977a27731a3209ebbcacebebaf41f902bd453a928bfd281cf7f8", size = 184152, upload-time = "2026-03-06T06:00:33.93Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/20/4ed37f6199af5dde94d4aeaf577f3813a5ec6635834cda1d957013a09c76/charset_normalizer-3.4.5-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0b362bcd27819f9c07cbf23db4e0e8cd4b44c5ecd900c2ff907b2b92274a7412", size = 195226, upload-time = "2026-03-06T06:00:35.469Z" },
+    { url = "https://files.pythonhosted.org/packages/28/31/7ba1102178cba7c34dcc050f43d427172f389729e356038f0726253dd914/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:77be992288f720306ab4108fe5c74797de327f3248368dfc7e1a916d6ed9e5a2", size = 192933, upload-time = "2026-03-06T06:00:36.83Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/23/f86443ab3921e6a60b33b93f4a1161222231f6c69bc24fb18f3bee7b8518/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:8b78d8a609a4b82c273257ee9d631ded7fac0d875bdcdccc109f3ee8328cfcb1", size = 185647, upload-time = "2026-03-06T06:00:38.367Z" },
+    { url = "https://files.pythonhosted.org/packages/82/44/08b8be891760f1f5a6d23ce11d6d50c92981603e6eb740b4f72eea9424e2/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:ba20bdf69bd127f66d0174d6f2a93e69045e0b4036dc1ca78e091bcc765830c4", size = 209533, upload-time = "2026-03-06T06:00:41.931Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/5f/df114f23406199f8af711ddccfbf409ffbc5b7cdc18fa19644997ff0c9bb/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:76a9d0de4d0eab387822e7b35d8f89367dd237c72e82ab42b9f7bf5e15ada00f", size = 195901, upload-time = "2026-03-06T06:00:43.978Z" },
+    { url = "https://files.pythonhosted.org/packages/07/83/71ef34a76fe8aa05ff8f840244bda2d61e043c2ef6f30d200450b9f6a1be/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:8fff79bf5978c693c9b1a4d71e4a94fddfb5fe744eb062a318e15f4a2f63a550", size = 204950, upload-time = "2026-03-06T06:00:45.202Z" },
+    { url = "https://files.pythonhosted.org/packages/58/40/0253be623995365137d7dc68e45245036207ab2227251e69a3d93ce43183/charset_normalizer-3.4.5-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c7e84e0c0005e3bdc1a9211cd4e62c78ba80bc37b2365ef4410cd2007a9047f2", size = 198546, upload-time = "2026-03-06T06:00:46.481Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/5c/5f3cb5b259a130895ef5ae16b38eaf141430fa3f7af50cd06c5d67e4f7b2/charset_normalizer-3.4.5-cp310-cp310-win32.whl", hash = "sha256:58ad8270cfa5d4bef1bc85bd387217e14ff154d6630e976c6f56f9a040757475", size = 132516, upload-time = "2026-03-06T06:00:47.924Z" },
+    { url = "https://files.pythonhosted.org/packages/a5/c3/84fb174e7770f2df2e1a2115090771bfbc2227fb39a765c6d00568d1aab4/charset_normalizer-3.4.5-cp310-cp310-win_amd64.whl", hash = "sha256:02a9d1b01c1e12c27883b0c9349e0bcd9ae92e727ff1a277207e1a262b1cbf05", size = 142906, upload-time = "2026-03-06T06:00:49.389Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/b2/6f852f8b969f2cbd0d4092d2e60139ab1af95af9bb651337cae89ec0f684/charset_normalizer-3.4.5-cp310-cp310-win_arm64.whl", hash = "sha256:039215608ac7b358c4da0191d10fc76868567fbf276d54c14721bdedeb6de064", size = 133258, upload-time = "2026-03-06T06:00:51.051Z" },
+    { url = "https://files.pythonhosted.org/packages/c5/60/3a621758945513adfd4db86827a5bafcc615f913dbd0b4c2ed64a65731be/charset_normalizer-3.4.5-py3-none-any.whl", hash = "sha256:9db5e3fcdcee89a78c04dffb3fe33c79f77bd741a624946db2591c81b2fc85b0", size = 55455, upload-time = "2026-03-06T06:03:17.827Z" },
+]
+
+[[package]]
+name = "click"
+version = "8.3.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3d/fa/656b739db8587d7b5dfa22e22ed02566950fbfbcdc20311993483657a5c0/click-8.3.1.tar.gz", hash = "sha256:12ff4785d337a1bb490bb7e9c2b1ee5da3112e94a8622f26a6c77f5d2fc6842a", size = 295065, upload-time = "2025-11-15T20:45:42.706Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/98/78/01c019cdb5d6498122777c1a43056ebb3ebfeef2076d9d026bfe15583b2b/click-8.3.1-py3-none-any.whl", hash = "sha256:981153a64e25f12d547d3426c367a4857371575ee7ad18df2a6183ab0545b2a6", size = 108274, upload-time = "2025-11-15T20:45:41.139Z" },
+]
+
+[[package]]
+name = "cloudpickle"
+version = "3.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/27/fb/576f067976d320f5f0114a8d9fa1215425441bb35627b1993e5afd8111e5/cloudpickle-3.1.2.tar.gz", hash = "sha256:7fda9eb655c9c230dab534f1983763de5835249750e85fbcef43aaa30a9a2414", size = 22330, upload-time = "2025-11-03T09:25:26.604Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/88/39/799be3f2f0f38cc727ee3b4f1445fe6d5e4133064ec2e4115069418a5bb6/cloudpickle-3.1.2-py3-none-any.whl", hash = "sha256:9acb47f6afd73f60dc1df93bb801b472f05ff42fa6c84167d25cb206be1fbf4a", size = 22228, upload-time = "2025-11-03T09:25:25.534Z" },
+]
+
+[[package]]
+name = "codetiming"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ad/4e/c40bf151af20ba2748bd6ea24e484d7b6196b1056ba3a1a4ee33b6939c37/codetiming-1.4.0.tar.gz", hash = "sha256:4937bf913a2814258b87eaaa43d9a1bb24711ffd3557a9ab6934fa1fe3ba0dbc", size = 14899, upload-time = "2022-11-08T07:12:29.361Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/bc/91/e4a2b7c64e738beefddfa24b409d6eecb16c378bde01578918b6ea722a09/codetiming-1.4.0-py3-none-any.whl", hash = "sha256:3b80f409bef00941a9755c5524071ce2f72eaa4520f4bc35b33869cde024ccbd", size = 7165, upload-time = "2022-11-08T07:12:23.553Z" },
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d8/53/6f443c9a4a8358a93a6792e2acffb9d9d5cb0a5cfd8802644b7b1c9a02e4/colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44", size = 27697, upload-time = "2022-10-25T02:36:22.414Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" },
+]
+
+[[package]]
+name = "comm"
+version = "0.2.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/4c/13/7d740c5849255756bc17888787313b61fd38a0a8304fc4f073dfc46122aa/comm-0.2.3.tar.gz", hash = "sha256:2dc8048c10962d55d7ad693be1e7045d891b7ce8d999c97963a5e3e99c055971", size = 6319, upload-time = "2025-07-25T14:02:04.452Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/60/97/891a0971e1e4a8c5d2b20bbe0e524dc04548d2307fee33cdeba148fd4fc7/comm-0.2.3-py3-none-any.whl", hash = "sha256:c615d91d75f7f04f095b30d1c1711babd43bdc6419c1be9886a85f2f4e489417", size = 7294, upload-time = "2025-07-25T14:02:02.896Z" },
+]
+
+[[package]]
+name = "contourpy"
+version = "1.3.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/66/54/eb9bfc647b19f2009dd5c7f5ec51c4e6ca831725f1aea7a993034f483147/contourpy-1.3.2.tar.gz", hash = "sha256:b6945942715a034c671b7fc54f9588126b0b8bf23db2696e3ca8328f3ff0ab54", size = 13466130, upload-time = "2025-04-15T17:47:53.79Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/12/a3/da4153ec8fe25d263aa48c1a4cbde7f49b59af86f0b6f7862788c60da737/contourpy-1.3.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ba38e3f9f330af820c4b27ceb4b9c7feee5fe0493ea53a8720f4792667465934", size = 268551, upload-time = "2025-04-15T17:34:46.581Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/6c/330de89ae1087eb622bfca0177d32a7ece50c3ef07b28002de4757d9d875/contourpy-1.3.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dc41ba0714aa2968d1f8674ec97504a8f7e334f48eeacebcaa6256213acb0989", size = 253399, upload-time = "2025-04-15T17:34:51.427Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/bd/20c6726b1b7f81a8bee5271bed5c165f0a8e1f572578a9d27e2ccb763cb2/contourpy-1.3.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9be002b31c558d1ddf1b9b415b162c603405414bacd6932d031c5b5a8b757f0d", size = 312061, upload-time = "2025-04-15T17:34:55.961Z" },
+    { url = "https://files.pythonhosted.org/packages/22/fc/a9665c88f8a2473f823cf1ec601de9e5375050f1958cbb356cdf06ef1ab6/contourpy-1.3.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8d2e74acbcba3bfdb6d9d8384cdc4f9260cae86ed9beee8bd5f54fee49a430b9", size = 351956, upload-time = "2025-04-15T17:35:00.992Z" },
+    { url = "https://files.pythonhosted.org/packages/25/eb/9f0a0238f305ad8fb7ef42481020d6e20cf15e46be99a1fcf939546a177e/contourpy-1.3.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e259bced5549ac64410162adc973c5e2fb77f04df4a439d00b478e57a0e65512", size = 320872, upload-time = "2025-04-15T17:35:06.177Z" },
+    { url = "https://files.pythonhosted.org/packages/32/5c/1ee32d1c7956923202f00cf8d2a14a62ed7517bdc0ee1e55301227fc273c/contourpy-1.3.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ad687a04bc802cbe8b9c399c07162a3c35e227e2daccf1668eb1f278cb698631", size = 325027, upload-time = "2025-04-15T17:35:11.244Z" },
+    { url = "https://files.pythonhosted.org/packages/83/bf/9baed89785ba743ef329c2b07fd0611d12bfecbedbdd3eeecf929d8d3b52/contourpy-1.3.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:cdd22595308f53ef2f891040ab2b93d79192513ffccbd7fe19be7aa773a5e09f", size = 1306641, upload-time = "2025-04-15T17:35:26.701Z" },
+    { url = "https://files.pythonhosted.org/packages/d4/cc/74e5e83d1e35de2d28bd97033426b450bc4fd96e092a1f7a63dc7369b55d/contourpy-1.3.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b4f54d6a2defe9f257327b0f243612dd051cc43825587520b1bf74a31e2f6ef2", size = 1374075, upload-time = "2025-04-15T17:35:43.204Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/42/17f3b798fd5e033b46a16f8d9fcb39f1aba051307f5ebf441bad1ecf78f8/contourpy-1.3.2-cp310-cp310-win32.whl", hash = "sha256:f939a054192ddc596e031e50bb13b657ce318cf13d264f095ce9db7dc6ae81c0", size = 177534, upload-time = "2025-04-15T17:35:46.554Z" },
+    { url = "https://files.pythonhosted.org/packages/54/ec/5162b8582f2c994721018d0c9ece9dc6ff769d298a8ac6b6a652c307e7df/contourpy-1.3.2-cp310-cp310-win_amd64.whl", hash = "sha256:c440093bbc8fc21c637c03bafcbef95ccd963bc6e0514ad887932c18ca2a759a", size = 221188, upload-time = "2025-04-15T17:35:50.064Z" },
+    { url = "https://files.pythonhosted.org/packages/33/05/b26e3c6ecc05f349ee0013f0bb850a761016d89cec528a98193a48c34033/contourpy-1.3.2-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:fd93cc7f3139b6dd7aab2f26a90dde0aa9fc264dbf70f6740d498a70b860b82c", size = 265681, upload-time = "2025-04-15T17:44:59.314Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/25/ac07d6ad12affa7d1ffed11b77417d0a6308170f44ff20fa1d5aa6333f03/contourpy-1.3.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:107ba8a6a7eec58bb475329e6d3b95deba9440667c4d62b9b6063942b61d7f16", size = 315101, upload-time = "2025-04-15T17:45:04.165Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/4d/5bb3192bbe9d3f27e3061a6a8e7733c9120e203cb8515767d30973f71030/contourpy-1.3.2-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:ded1706ed0c1049224531b81128efbd5084598f18d8a2d9efae833edbd2b40ad", size = 220599, upload-time = "2025-04-15T17:45:08.456Z" },
+]
+
+[[package]]
+name = "cpm-kernels"
+version = "1.0.11"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/af/84/1831ce6ffa87b8fd4d9673c3595d0fc4e6631c0691eb43f406d3bf89b951/cpm_kernels-1.0.11-py3-none-any.whl", hash = "sha256:eab7f211f3b3f6a0686ded4c15cd7d9158393cdf69a931fa5b96a5fbcd366822", size = 416640, upload-time = "2022-03-07T05:34:49.81Z" },
+]
+
+[[package]]
+name = "crcmod"
+version = "1.7"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6b/b0/e595ce2a2527e169c3bcd6c33d2473c1918e0b7f6826a043ca1245dd4e5b/crcmod-1.7.tar.gz", hash = "sha256:dc7051a0db5f2bd48665a990d3ec1cc305a466a77358ca4492826f41f283601e", size = 89670, upload-time = "2010-06-27T14:35:29.538Z" }
+
+[[package]]
+name = "cryptography"
+version = "46.0.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cffi", marker = "platform_python_implementation != 'PyPy' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/60/04/ee2a9e8542e4fa2773b81771ff8349ff19cdd56b7258a0cc442639052edb/cryptography-46.0.5.tar.gz", hash = "sha256:abace499247268e3757271b2f1e244b36b06f8515cf27c4d49468fc9eb16e93d", size = 750064, upload-time = "2026-02-10T19:18:38.255Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f7/81/b0bb27f2ba931a65409c6b8a8b358a7f03c0e46eceacddff55f7c84b1f3b/cryptography-46.0.5-cp311-abi3-macosx_10_9_universal2.whl", hash = "sha256:351695ada9ea9618b3500b490ad54c739860883df6c1f555e088eaf25b1bbaad", size = 7176289, upload-time = "2026-02-10T19:17:08.274Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/9e/6b4397a3e3d15123de3b1806ef342522393d50736c13b20ec4c9ea6693a6/cryptography-46.0.5-cp311-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:c18ff11e86df2e28854939acde2d003f7984f721eba450b56a200ad90eeb0e6b", size = 4275637, upload-time = "2026-02-10T19:17:10.53Z" },
+    { url = "https://files.pythonhosted.org/packages/63/e7/471ab61099a3920b0c77852ea3f0ea611c9702f651600397ac567848b897/cryptography-46.0.5-cp311-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d7e3d356b8cd4ea5aff04f129d5f66ebdc7b6f8eae802b93739ed520c47c79b", size = 4424742, upload-time = "2026-02-10T19:17:12.388Z" },
+    { url = "https://files.pythonhosted.org/packages/37/53/a18500f270342d66bf7e4d9f091114e31e5ee9e7375a5aba2e85a91e0044/cryptography-46.0.5-cp311-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:50bfb6925eff619c9c023b967d5b77a54e04256c4281b0e21336a130cd7fc263", size = 4277528, upload-time = "2026-02-10T19:17:13.853Z" },
+    { url = "https://files.pythonhosted.org/packages/22/29/c2e812ebc38c57b40e7c583895e73c8c5adb4d1e4a0cc4c5a4fdab2b1acc/cryptography-46.0.5-cp311-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:803812e111e75d1aa73690d2facc295eaefd4439be1023fefc4995eaea2af90d", size = 4947993, upload-time = "2026-02-10T19:17:15.618Z" },
+    { url = "https://files.pythonhosted.org/packages/6b/e7/237155ae19a9023de7e30ec64e5d99a9431a567407ac21170a046d22a5a3/cryptography-46.0.5-cp311-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:3ee190460e2fbe447175cda91b88b84ae8322a104fc27766ad09428754a618ed", size = 4456855, upload-time = "2026-02-10T19:17:17.221Z" },
+    { url = "https://files.pythonhosted.org/packages/2d/87/fc628a7ad85b81206738abbd213b07702bcbdada1dd43f72236ef3cffbb5/cryptography-46.0.5-cp311-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:f145bba11b878005c496e93e257c1e88f154d278d2638e6450d17e0f31e558d2", size = 3984635, upload-time = "2026-02-10T19:17:18.792Z" },
+    { url = "https://files.pythonhosted.org/packages/84/29/65b55622bde135aedf4565dc509d99b560ee4095e56989e815f8fd2aa910/cryptography-46.0.5-cp311-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:e9251e3be159d1020c4030bd2e5f84d6a43fe54b6c19c12f51cde9542a2817b2", size = 4277038, upload-time = "2026-02-10T19:17:20.256Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/36/45e76c68d7311432741faf1fbf7fac8a196a0a735ca21f504c75d37e2558/cryptography-46.0.5-cp311-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:47fb8a66058b80e509c47118ef8a75d14c455e81ac369050f20ba0d23e77fee0", size = 4912181, upload-time = "2026-02-10T19:17:21.825Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/1a/c1ba8fead184d6e3d5afcf03d569acac5ad063f3ac9fb7258af158f7e378/cryptography-46.0.5-cp311-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:4c3341037c136030cb46e4b1e17b7418ea4cbd9dd207e4a6f3b2b24e0d4ac731", size = 4456482, upload-time = "2026-02-10T19:17:25.133Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/e5/3fb22e37f66827ced3b902cf895e6a6bc1d095b5b26be26bd13c441fdf19/cryptography-46.0.5-cp311-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:890bcb4abd5a2d3f852196437129eb3667d62630333aacc13dfd470fad3aaa82", size = 4405497, upload-time = "2026-02-10T19:17:26.66Z" },
+    { url = "https://files.pythonhosted.org/packages/1a/df/9d58bb32b1121a8a2f27383fabae4d63080c7ca60b9b5c88be742be04ee7/cryptography-46.0.5-cp311-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:80a8d7bfdf38f87ca30a5391c0c9ce4ed2926918e017c29ddf643d0ed2778ea1", size = 4667819, upload-time = "2026-02-10T19:17:28.569Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/ed/325d2a490c5e94038cdb0117da9397ece1f11201f425c4e9c57fe5b9f08b/cryptography-46.0.5-cp311-abi3-win32.whl", hash = "sha256:60ee7e19e95104d4c03871d7d7dfb3d22ef8a9b9c6778c94e1c8fcc8365afd48", size = 3028230, upload-time = "2026-02-10T19:17:30.518Z" },
+    { url = "https://files.pythonhosted.org/packages/e9/5a/ac0f49e48063ab4255d9e3b79f5def51697fce1a95ea1370f03dc9db76f6/cryptography-46.0.5-cp311-abi3-win_amd64.whl", hash = "sha256:38946c54b16c885c72c4f59846be9743d699eee2b69b6988e0a00a01f46a61a4", size = 3480909, upload-time = "2026-02-10T19:17:32.083Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/fa/a66aa722105ad6a458bebd64086ca2b72cdd361fed31763d20390f6f1389/cryptography-46.0.5-cp38-abi3-macosx_10_9_universal2.whl", hash = "sha256:4108d4c09fbbf2789d0c926eb4152ae1760d5a2d97612b92d508d96c861e4d31", size = 7170514, upload-time = "2026-02-10T19:17:56.267Z" },
+    { url = "https://files.pythonhosted.org/packages/0f/04/c85bdeab78c8bc77b701bf0d9bdcf514c044e18a46dcff330df5448631b0/cryptography-46.0.5-cp38-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7d1f30a86d2757199cb2d56e48cce14deddf1f9c95f1ef1b64ee91ea43fe2e18", size = 4275349, upload-time = "2026-02-10T19:17:58.419Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/32/9b87132a2f91ee7f5223b091dc963055503e9b442c98fc0b8a5ca765fab0/cryptography-46.0.5-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:039917b0dc418bb9f6edce8a906572d69e74bd330b0b3fea4f79dab7f8ddd235", size = 4420667, upload-time = "2026-02-10T19:18:00.619Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/a6/a7cb7010bec4b7c5692ca6f024150371b295ee1c108bdc1c400e4c44562b/cryptography-46.0.5-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:ba2a27ff02f48193fc4daeadf8ad2590516fa3d0adeeb34336b96f7fa64c1e3a", size = 4276980, upload-time = "2026-02-10T19:18:02.379Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/7c/c4f45e0eeff9b91e3f12dbd0e165fcf2a38847288fcfd889deea99fb7b6d/cryptography-46.0.5-cp38-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:61aa400dce22cb001a98014f647dc21cda08f7915ceb95df0c9eaf84b4b6af76", size = 4939143, upload-time = "2026-02-10T19:18:03.964Z" },
+    { url = "https://files.pythonhosted.org/packages/37/19/e1b8f964a834eddb44fa1b9a9976f4e414cbb7aa62809b6760c8803d22d1/cryptography-46.0.5-cp38-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:3ce58ba46e1bc2aac4f7d9290223cead56743fa6ab94a5d53292ffaac6a91614", size = 4453674, upload-time = "2026-02-10T19:18:05.588Z" },
+    { url = "https://files.pythonhosted.org/packages/db/ed/db15d3956f65264ca204625597c410d420e26530c4e2943e05a0d2f24d51/cryptography-46.0.5-cp38-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:420d0e909050490d04359e7fdb5ed7e667ca5c3c402b809ae2563d7e66a92229", size = 3978801, upload-time = "2026-02-10T19:18:07.167Z" },
+    { url = "https://files.pythonhosted.org/packages/41/e2/df40a31d82df0a70a0daf69791f91dbb70e47644c58581d654879b382d11/cryptography-46.0.5-cp38-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:582f5fcd2afa31622f317f80426a027f30dc792e9c80ffee87b993200ea115f1", size = 4276755, upload-time = "2026-02-10T19:18:09.813Z" },
+    { url = "https://files.pythonhosted.org/packages/33/45/726809d1176959f4a896b86907b98ff4391a8aa29c0aaaf9450a8a10630e/cryptography-46.0.5-cp38-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:bfd56bb4b37ed4f330b82402f6f435845a5f5648edf1ad497da51a8452d5d62d", size = 4901539, upload-time = "2026-02-10T19:18:11.263Z" },
+    { url = "https://files.pythonhosted.org/packages/99/0f/a3076874e9c88ecb2ecc31382f6e7c21b428ede6f55aafa1aa272613e3cd/cryptography-46.0.5-cp38-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:a3d507bb6a513ca96ba84443226af944b0f7f47dcc9a399d110cd6146481d24c", size = 4452794, upload-time = "2026-02-10T19:18:12.914Z" },
+    { url = "https://files.pythonhosted.org/packages/02/ef/ffeb542d3683d24194a38f66ca17c0a4b8bf10631feef44a7ef64e631b1a/cryptography-46.0.5-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:9f16fbdf4da055efb21c22d81b89f155f02ba420558db21288b3d0035bafd5f4", size = 4404160, upload-time = "2026-02-10T19:18:14.375Z" },
+    { url = "https://files.pythonhosted.org/packages/96/93/682d2b43c1d5f1406ed048f377c0fc9fc8f7b0447a478d5c65ab3d3a66eb/cryptography-46.0.5-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:ced80795227d70549a411a4ab66e8ce307899fad2220ce5ab2f296e687eacde9", size = 4667123, upload-time = "2026-02-10T19:18:15.886Z" },
+    { url = "https://files.pythonhosted.org/packages/45/2d/9c5f2926cb5300a8eefc3f4f0b3f3df39db7f7ce40c8365444c49363cbda/cryptography-46.0.5-cp38-abi3-win32.whl", hash = "sha256:02f547fce831f5096c9a567fd41bc12ca8f11df260959ecc7c3202555cc47a72", size = 3010220, upload-time = "2026-02-10T19:18:17.361Z" },
+    { url = "https://files.pythonhosted.org/packages/48/ef/0c2f4a8e31018a986949d34a01115dd057bf536905dca38897bacd21fac3/cryptography-46.0.5-cp38-abi3-win_amd64.whl", hash = "sha256:556e106ee01aa13484ce9b0239bca667be5004efb0aabbed28d353df86445595", size = 3467050, upload-time = "2026-02-10T19:18:18.899Z" },
+]
+
+[[package]]
+name = "cut-cross-entropy"
+version = "25.1.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "triton", marker = "sys_platform == 'linux' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7e/97/45ff09cfcda7b200389204daa0125168e6544fba257adbbcdf728501d4f9/cut_cross_entropy-25.1.1.tar.gz", hash = "sha256:5fe5924509248b1aea5c890f8887c6a7759f7c8b1ebc0490e42c247c4f7c1e34", size = 22972, upload-time = "2025-01-07T12:21:53.896Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/5f/62fdb048f84d19e2123b6bbd722fe09c8c79b4964c50094d1e979db808e2/cut_cross_entropy-25.1.1-py3-none-any.whl", hash = "sha256:e46f26d348f6a67927d17e65c5a212e795be13dcad5b10a77a200d6b8102d9d1", size = 22672, upload-time = "2025-01-07T12:21:51.678Z" },
+]
+
+[[package]]
+name = "cycler"
+version = "0.12.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a9/95/a3dbbb5028f35eafb79008e7522a75244477d2838f38cbb722248dabc2a8/cycler-0.12.1.tar.gz", hash = "sha256:88bb128f02ba341da8ef447245a9e138fae777f6a23943da4540077d3601eb1c", size = 7615, upload-time = "2023-10-07T05:32:18.335Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e7/05/c19819d5e3d95294a6f5947fb9b9629efb316b96de511b418c53d245aae6/cycler-0.12.1-py3-none-any.whl", hash = "sha256:85cef7cff222d8644161529808465972e51340599459b8ac3ccbac5a854e0d30", size = 8321, upload-time = "2023-10-07T05:32:16.783Z" },
+]
+
+[[package]]
+name = "dacite"
+version = "1.9.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/55/a0/7ca79796e799a3e782045d29bf052b5cde7439a2bbb17f15ff44f7aacc63/dacite-1.9.2.tar.gz", hash = "sha256:6ccc3b299727c7aa17582f0021f6ae14d5de47c7227932c47fec4cdfefd26f09", size = 22420, upload-time = "2025-02-05T09:27:29.757Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/94/35/386550fd60316d1e37eccdda609b074113298f23cef5bddb2049823fe666/dacite-1.9.2-py3-none-any.whl", hash = "sha256:053f7c3f5128ca2e9aceb66892b1a3c8936d02c686e707bee96e19deef4bc4a0", size = 16600, upload-time = "2025-02-05T09:27:24.345Z" },
+]
+
+[[package]]
+name = "datasets"
+version = "4.1.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "dill" },
+    { name = "filelock" },
+    { name = "fsspec", extra = ["http"] },
+    { name = "huggingface-hub" },
+    { name = "multiprocess" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "pandas" },
+    { name = "pyarrow" },
+    { name = "pyyaml" },
+    { name = "requests" },
+    { name = "tqdm" },
+    { name = "xxhash" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/91/a4/73f8e6ef52c535e1d20d5b2ca83bfe6de399d8b8b8a61ccc8d63d60735aa/datasets-4.1.1.tar.gz", hash = "sha256:7d8d5ba8b12861d2c44bfff9c83484ebfafff1ff553371e5901a8d3aab5450e2", size = 579324, upload-time = "2025-09-18T13:14:27.108Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f4/c8/09012ac195a0aab58755800d2efdc0e7d5905053509f12cb5d136c911cda/datasets-4.1.1-py3-none-any.whl", hash = "sha256:62e4f6899a36be9ec74a7e759a6951253cc85b3fcfa0a759b0efa8353b149dac", size = 503623, upload-time = "2025-09-18T13:14:25.111Z" },
+]
+
+[[package]]
+name = "debugpy"
+version = "1.8.20"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e0/b7/cd8080344452e4874aae67c40d8940e2b4d47b01601a8fd9f44786c757c7/debugpy-1.8.20.tar.gz", hash = "sha256:55bc8701714969f1ab89a6d5f2f3d40c36f91b2cbe2f65d98bf8196f6a6a2c33", size = 1645207, upload-time = "2026-01-29T23:03:28.199Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/71/be/8bd693a0b9d53d48c8978fa5d889e06f3b5b03e45fd1ea1e78267b4887cb/debugpy-1.8.20-cp310-cp310-macosx_15_0_x86_64.whl", hash = "sha256:157e96ffb7f80b3ad36d808646198c90acb46fdcfd8bb1999838f0b6f2b59c64", size = 2099192, upload-time = "2026-01-29T23:03:29.707Z" },
+    { url = "https://files.pythonhosted.org/packages/77/1b/85326d07432086a06361d493d2743edd0c4fc2ef62162be7f8618441ac37/debugpy-1.8.20-cp310-cp310-manylinux_2_34_x86_64.whl", hash = "sha256:c1178ae571aff42e61801a38b007af504ec8e05fde1c5c12e5a7efef21009642", size = 3088568, upload-time = "2026-01-29T23:03:31.467Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/60/3e08462ee3eccd10998853eb35947c416e446bfe2bc37dbb886b9044586c/debugpy-1.8.20-cp310-cp310-win32.whl", hash = "sha256:c29dd9d656c0fbd77906a6e6a82ae4881514aa3294b94c903ff99303e789b4a2", size = 5284399, upload-time = "2026-01-29T23:03:33.678Z" },
+    { url = "https://files.pythonhosted.org/packages/72/43/09d49106e770fe558ced5e80df2e3c2ebee10e576eda155dcc5670473663/debugpy-1.8.20-cp310-cp310-win_amd64.whl", hash = "sha256:3ca85463f63b5dd0aa7aaa933d97cbc47c174896dcae8431695872969f981893", size = 5316388, upload-time = "2026-01-29T23:03:35.095Z" },
+    { url = "https://files.pythonhosted.org/packages/e0/c3/7f67dea8ccf8fdcb9c99033bbe3e90b9e7395415843accb81428c441be2d/debugpy-1.8.20-py2.py3-none-any.whl", hash = "sha256:5be9bed9ae3be00665a06acaa48f8329d2b9632f15fd09f6a9a8c8d9907e54d7", size = 5337658, upload-time = "2026-01-29T23:04:17.404Z" },
+]
+
+[[package]]
+name = "decorator"
+version = "5.2.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/43/fa/6d96a0978d19e17b68d634497769987b16c8f4cd0a7a05048bec693caa6b/decorator-5.2.1.tar.gz", hash = "sha256:65f266143752f734b0a7cc83c46f4618af75b8c5911b00ccb61d0ac9b6da0360", size = 56711, upload-time = "2025-02-24T04:41:34.073Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4e/8c/f3147f5c4b73e7550fe5f9352eaa956ae838d5c51eb58e7a25b9f3e2643b/decorator-5.2.1-py3-none-any.whl", hash = "sha256:d316bb415a2d9e2d2b3abcc4084c6502fc09240e292cd76a76afc106a1c8e04a", size = 9190, upload-time = "2025-02-24T04:41:32.565Z" },
+]
+
+[[package]]
+name = "decord"
+version = "0.6.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/11/79/936af42edf90a7bd4e41a6cac89c913d4b47fa48a26b042d5129a9242ee3/decord-0.6.0-py3-none-manylinux2010_x86_64.whl", hash = "sha256:51997f20be8958e23b7c4061ba45d0efcd86bffd5fe81c695d0befee0d442976", size = 13602299, upload-time = "2021-06-14T21:30:55.486Z" },
+    { url = "https://files.pythonhosted.org/packages/6c/be/e15b5b866da452e62635a7b27513f31cb581fa2ea9cc9b768b535d62a955/decord-0.6.0-py3-none-win_amd64.whl", hash = "sha256:02665d7c4f1193a330205a791bc128f7e108eb6ae5b67144437a02f700943bad", size = 24733380, upload-time = "2021-06-14T21:30:57.766Z" },
+]
+
+[[package]]
+name = "diffusers"
+version = "0.37.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "filelock" },
+    { name = "httpx" },
+    { name = "huggingface-hub" },
+    { name = "importlib-metadata" },
+    { name = "numpy" },
+    { name = "pillow" },
+    { name = "regex" },
+    { name = "requests" },
+    { name = "safetensors" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/58/3b/01d0ff800b811c5ad8bba682f4c6abf1d7071cd81464c01724333fefb7ba/diffusers-0.37.0.tar.gz", hash = "sha256:408789af73898585f525afd07ca72b3955affea4216a669558e9f59b5b1fe704", size = 4141136, upload-time = "2026-03-05T14:58:39.704Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f7/55/586a3a2b9c95f371c9c3cb048c3cac15aedcce8d6d53ebd6bbc46860722d/diffusers-0.37.0-py3-none-any.whl", hash = "sha256:7eab74bf896974250b5e1027cae813aba1004f02d97c9b44891b83713386aa08", size = 5000449, upload-time = "2026-03-05T14:58:37.361Z" },
+]
+
+[[package]]
+name = "dill"
+version = "0.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/12/80/630b4b88364e9a8c8c5797f4602d0f76ef820909ee32f0bacb9f90654042/dill-0.4.0.tar.gz", hash = "sha256:0633f1d2df477324f53a895b02c901fb961bdbf65a17122586ea7019292cbcf0", size = 186976, upload-time = "2025-04-16T00:41:48.867Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/50/3d/9373ad9c56321fdab5b41197068e1d8c25883b3fea29dd361f9b55116869/dill-0.4.0-py3-none-any.whl", hash = "sha256:44f54bf6412c2c8464c14e8243eb163690a9800dbe2c367330883b19c7561049", size = 119668, upload-time = "2025-04-16T00:41:47.671Z" },
+]
+
+[[package]]
+name = "distlib"
+version = "0.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/96/8e/709914eb2b5749865801041647dc7f4e6d00b549cfe88b65ca192995f07c/distlib-0.4.0.tar.gz", hash = "sha256:feec40075be03a04501a973d81f633735b4b69f98b05450592310c0f401a4e0d", size = 614605, upload-time = "2025-07-17T16:52:00.465Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/33/6b/e0547afaf41bf2c42e52430072fa5658766e3d65bd4b03a563d1b6336f57/distlib-0.4.0-py2.py3-none-any.whl", hash = "sha256:9659f7d87e46584a30b5780e43ac7a2143098441670ff0a49d5f9034c54a6c16", size = 469047, upload-time = "2025-07-17T16:51:58.613Z" },
+]
+
+[[package]]
+name = "distro"
+version = "1.9.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fc/f8/98eea607f65de6527f8a2e8885fc8015d3e6f5775df186e443e0964a11c3/distro-1.9.0.tar.gz", hash = "sha256:2fa77c6fd8940f116ee1d6b94a2f90b13b5ea8d019b98bc8bafdcabcdd9bdbed", size = 60722, upload-time = "2023-12-24T09:54:32.31Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/12/b3/231ffd4ab1fc9d679809f356cebee130ac7daa00d6d6f3206dd4fd137e9e/distro-1.9.0-py3-none-any.whl", hash = "sha256:7bffd925d65168f85027d8da9af6bddab658135b840670a223589bc0c8ef02b2", size = 20277, upload-time = "2023-12-24T09:54:30.421Z" },
+]
+
+[[package]]
+name = "dm-tree"
+version = "0.1.9"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "attrs" },
+    { name = "numpy" },
+    { name = "wrapt" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a6/83/ce29720ccf934c6cfa9b9c95ebbe96558386e66886626066632b5e44afed/dm_tree-0.1.9.tar.gz", hash = "sha256:a4c7db3d3935a5a2d5e4b383fc26c6b0cd6f78c6d4605d3e7b518800ecd5342b", size = 35623, upload-time = "2025-01-30T20:45:37.13Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/77/d2/88f685534d87072a5174fe229e77aab6b7da50092d5151ebc172f6270b5c/dm_tree-0.1.9-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:5d5b28ee2e461b6af65330c143806a6d0945dcabbb8d22d2ba863e6dabd9254e", size = 173568, upload-time = "2025-03-31T08:35:38.425Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/6a/64924e102f559c1380263a28a751f20a1bdd18e85ea599e216feead84adf/dm_tree-0.1.9-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:54d5616015412311df154908069fcf2c2d8786f6088a2ae3554d186cdf2b1e15", size = 146935, upload-time = "2025-01-30T20:45:16.505Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/79/ba0f7274164eb6bd06a36c2f8cb21b0debc32fd9ba8e73a7c9e50c90041b/dm_tree-0.1.9-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:831699d2c60a1b38776a193b7143ae0acad0a687d87654e6d3342584166816bc", size = 152892, upload-time = "2025-01-30T20:45:18.021Z" },
+    { url = "https://files.pythonhosted.org/packages/bf/20/8b96a34a15c5c4d1d6af44795963fa44381716975aabac83beab4fe80974/dm_tree-0.1.9-cp310-cp310-win_amd64.whl", hash = "sha256:1ae3cbff592bb3f2e197f5a8030de4a94e292e6cdd85adeea0b971d07a1b85f2", size = 101469, upload-time = "2025-01-30T20:45:19.197Z" },
+]
+
+[[package]]
+name = "docopt"
+version = "0.6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a2/55/8f8cab2afd404cf578136ef2cc5dfb50baa1761b68c9da1fb1e4eed343c9/docopt-0.6.2.tar.gz", hash = "sha256:49b3a825280bd66b3aa83585ef59c4a8c82f2c8a522dbe754a8bc8d08c85c491", size = 25901, upload-time = "2014-06-16T11:18:57.406Z" }
+
+[[package]]
+name = "docstring-parser"
+version = "0.17.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b2/9d/c3b43da9515bd270df0f80548d9944e389870713cc1fe2b8fb35fe2bcefd/docstring_parser-0.17.0.tar.gz", hash = "sha256:583de4a309722b3315439bb31d64ba3eebada841f2e2cee23b99df001434c912", size = 27442, upload-time = "2025-07-21T07:35:01.868Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/55/e2/2537ebcff11c1ee1ff17d8d0b6f4db75873e3b0fb32c2d4a2ee31ecb310a/docstring_parser-0.17.0-py3-none-any.whl", hash = "sha256:cf2569abd23dce8099b300f9b4fa8191e9582dda731fd533daf54c4551658708", size = 36896, upload-time = "2025-07-21T07:35:00.684Z" },
+]
+
+[[package]]
+name = "einops"
+version = "0.8.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2c/77/850bef8d72ffb9219f0b1aac23fbc1bf7d038ee6ea666f331fa273031aa2/einops-0.8.2.tar.gz", hash = "sha256:609da665570e5e265e27283aab09e7f279ade90c4f01bcfca111f3d3e13f2827", size = 56261, upload-time = "2026-01-26T04:13:17.638Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2a/09/f8d8f8f31e4483c10a906437b4ce31bdf3d6d417b73fe33f1a8b59e34228/einops-0.8.2-py3-none-any.whl", hash = "sha256:54058201ac7087911181bfec4af6091bb59380360f069276601256a76af08193", size = 65638, upload-time = "2026-01-26T04:13:18.546Z" },
+]
+
+[[package]]
+name = "etils"
+version = "1.13.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9b/a0/522bbff0f3cdd37968f90dd7f26c7aa801ed87f5ba335f156de7f2b88a48/etils-1.13.0.tar.gz", hash = "sha256:a5b60c71f95bcd2d43d4e9fb3dc3879120c1f60472bb5ce19f7a860b1d44f607", size = 106368, upload-time = "2025-07-15T10:29:10.563Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e7/98/87b5946356095738cb90a6df7b35ff69ac5750f6e783d5fbcc5cb3b6cbd7/etils-1.13.0-py3-none-any.whl", hash = "sha256:d9cd4f40fbe77ad6613b7348a18132cc511237b6c076dbb89105c0b520a4c6bb", size = 170603, upload-time = "2025-07-15T10:29:09.076Z" },
+]
+
+[package.optional-dependencies]
+enp = [
+    { name = "einops" },
+    { name = "numpy" },
+    { name = "typing-extensions" },
+]
+epath = [
+    { name = "fsspec" },
+    { name = "importlib-resources" },
+    { name = "typing-extensions" },
+    { name = "zipp" },
+]
+
+[[package]]
+name = "evaluate"
+version = "0.4.6"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "datasets" },
+    { name = "dill" },
+    { name = "fsspec", extra = ["http"] },
+    { name = "huggingface-hub" },
+    { name = "multiprocess" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "pandas" },
+    { name = "requests" },
+    { name = "tqdm" },
+    { name = "xxhash" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ad/d0/0c17a8e6e8dc7245f22dea860557c32bae50fc4d287ae030cb0e8ab8720f/evaluate-0.4.6.tar.gz", hash = "sha256:e07036ca12b3c24331f83ab787f21cc2dbf3631813a1631e63e40897c69a3f21", size = 65716, upload-time = "2025-09-18T13:06:30.581Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3e/af/3e990d8d4002bbc9342adb4facd59506e653da93b2417de0fa6027cb86b1/evaluate-0.4.6-py3-none-any.whl", hash = "sha256:bca85bc294f338377b7ac2f861e21c308b11b2a285f510d7d5394d5df437db29", size = 84069, upload-time = "2025-09-18T13:06:29.265Z" },
+]
+
+[[package]]
+name = "evo-vlac"
+version = "1.0.0"
+source = { git = "https://github.com/InternRobotics/VLAC.git#951225643fa798e3a8e96695e3c37f2b9f6cb29e" }
+dependencies = [
+    { name = "flash-attn" },
+    { name = "loguru" },
+    { name = "ms-swift" },
+    { name = "opencv-python" },
+    { name = "peft", version = "0.15.2", source = { registry = "https://pypi.org/simple" } },
+    { name = "timm" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" } },
+]
+
+[[package]]
+name = "exceptiongroup"
+version = "1.3.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/50/79/66800aadf48771f6b62f7eb014e352e5d06856655206165d775e675a02c9/exceptiongroup-1.3.1.tar.gz", hash = "sha256:8b412432c6055b0b7d14c310000ae93352ed6754f70fa8f7c34141f91c4e3219", size = 30371, upload-time = "2025-11-21T23:01:54.787Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8a/0e/97c33bf5009bdbac74fd2beace167cab3f978feb69cc36f1ef79360d6c4e/exceptiongroup-1.3.1-py3-none-any.whl", hash = "sha256:a7a39a3bd276781e98394987d3a5701d0c4edffb633bb7a5144577f82c773598", size = 16740, upload-time = "2025-11-21T23:01:53.443Z" },
+]
+
+[[package]]
+name = "executing"
+version = "2.2.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cc/28/c14e053b6762b1044f34a13aab6859bbf40456d37d23aa286ac24cfd9a5d/executing-2.2.1.tar.gz", hash = "sha256:3632cc370565f6648cc328b32435bd120a1e4ebb20c77e3fdde9a13cd1e533c4", size = 1129488, upload-time = "2025-09-01T09:48:10.866Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c1/ea/53f2148663b321f21b5a606bd5f191517cf40b7072c0497d3c92c4a13b1e/executing-2.2.1-py2.py3-none-any.whl", hash = "sha256:760643d3452b4d777d295bb167ccc74c64a81df23fb5e08eff250c425a4b2017", size = 28317, upload-time = "2025-09-01T09:48:08.5Z" },
+]
+
+[[package]]
+name = "farama-notifications"
+version = "0.0.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2e/2c/8384832b7a6b1fd6ba95bbdcae26e7137bb3eedc955c42fd5cdcc086cfbf/Farama-Notifications-0.0.4.tar.gz", hash = "sha256:13fceff2d14314cf80703c8266462ebf3733c7d165336eee998fc58e545efd18", size = 2131, upload-time = "2023-02-27T18:28:41.047Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/05/2c/ffc08c54c05cdce6fbed2aeebc46348dbe180c6d2c541c7af7ba0aa5f5f8/Farama_Notifications-0.0.4-py3-none-any.whl", hash = "sha256:14de931035a41961f7c056361dc7f980762a143d05791ef5794a751a2caf05ae", size = 2511, upload-time = "2023-02-27T18:28:39.447Z" },
+]
+
+[[package]]
+name = "fastapi"
+version = "0.135.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "annotated-doc" },
+    { name = "pydantic" },
+    { name = "starlette" },
+    { name = "typing-extensions" },
+    { name = "typing-inspection" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e7/7b/f8e0211e9380f7195ba3f3d40c292594fd81ba8ec4629e3854c353aaca45/fastapi-0.135.1.tar.gz", hash = "sha256:d04115b508d936d254cea545b7312ecaa58a7b3a0f84952535b4c9afae7668cd", size = 394962, upload-time = "2026-03-01T18:18:29.369Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e4/72/42e900510195b23a56bde950d26a51f8b723846bfcaa0286e90287f0422b/fastapi-0.135.1-py3-none-any.whl", hash = "sha256:46e2fc5745924b7c840f71ddd277382af29ce1cdb7d5eab5bf697e3fb9999c9e", size = 116999, upload-time = "2026-03-01T18:18:30.831Z" },
+]
+
+[[package]]
+name = "ffmpeg"
+version = "1.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f0/cc/3b7408b8ecf7c1d20ad480c3eaed7619857bf1054b690226e906fdf14258/ffmpeg-1.4.tar.gz", hash = "sha256:6931692c890ff21d39938433c2189747815dca0c60ddc7f9bb97f199dba0b5b9", size = 5055, upload-time = "2018-10-08T07:50:05.748Z" }
+
+[[package]]
+name = "ffmpy"
+version = "1.0.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/7d/d2/1c4c582d71bcc65c76fa69fab85de6257d50fdf6fd4a2317c53917e9a581/ffmpy-1.0.0.tar.gz", hash = "sha256:b12932e95435c8820f1cd041024402765f821971e4bae753b327fc02a6e12f8b", size = 5101, upload-time = "2025-11-11T06:24:23.856Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/55/56/dd3669eccebb6d8ac81e624542ebd53fe6f08e1b8f2f8d50aeb7e3b83f99/ffmpy-1.0.0-py3-none-any.whl", hash = "sha256:5640e5f0fd03fb6236d0e119b16ccf6522db1c826fdf35dcb87087b60fd7504f", size = 5614, upload-time = "2025-11-11T06:24:22.818Z" },
+]
+
+[[package]]
+name = "filelock"
+version = "3.25.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/77/18/a1fd2231c679dcb9726204645721b12498aeac28e1ad0601038f94b42556/filelock-3.25.0.tar.gz", hash = "sha256:8f00faf3abf9dc730a1ffe9c354ae5c04e079ab7d3a683b7c32da5dd05f26af3", size = 40158, upload-time = "2026-03-01T15:08:45.916Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f9/0b/de6f54d4a8bedfe8645c41497f3c18d749f0bd3218170c667bf4b81d0cdd/filelock-3.25.0-py3-none-any.whl", hash = "sha256:5ccf8069f7948f494968fc0713c10e5c182a9c9d9eef3a636307a20c2490f047", size = 26427, upload-time = "2026-03-01T15:08:44.593Z" },
+]
+
+[[package]]
+name = "flash-attn"
+version = "2.8.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "einops" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3b/b2/8d76c41ad7974ee264754709c22963447f7f8134613fd9ce80984ed0dab7/flash_attn-2.8.3.tar.gz", hash = "sha256:1e71dd64a9e0280e0447b8a0c2541bad4bf6ac65bdeaa2f90e51a9e57de0370d", size = 8447812, upload-time = "2025-08-15T08:28:12.911Z" }
+
+[[package]]
+name = "flatbuffers"
+version = "25.12.19"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e8/2d/d2a548598be01649e2d46231d151a6c56d10b964d94043a335ae56ea2d92/flatbuffers-25.12.19-py2.py3-none-any.whl", hash = "sha256:7634f50c427838bb021c2d66a3d1168e9d199b0607e6329399f04846d42e20b4", size = 26661, upload-time = "2025-12-19T23:16:13.622Z" },
+]
+
+[[package]]
+name = "fonttools"
+version = "4.61.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ec/ca/cf17b88a8df95691275a3d77dc0a5ad9907f328ae53acbe6795da1b2f5ed/fonttools-4.61.1.tar.gz", hash = "sha256:6675329885c44657f826ef01d9e4fb33b9158e9d93c537d84ad8399539bc6f69", size = 3565756, upload-time = "2025-12-12T17:31:24.246Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5b/94/8a28707adb00bed1bf22dac16ccafe60faf2ade353dcb32c3617ee917307/fonttools-4.61.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:7c7db70d57e5e1089a274cbb2b1fd635c9a24de809a231b154965d415d6c6d24", size = 2854799, upload-time = "2025-12-12T17:29:27.5Z" },
+    { url = "https://files.pythonhosted.org/packages/94/93/c2e682faaa5ee92034818d8f8a8145ae73eb83619600495dcf8503fa7771/fonttools-4.61.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:5fe9fd43882620017add5eabb781ebfbc6998ee49b35bd7f8f79af1f9f99a958", size = 2403032, upload-time = "2025-12-12T17:29:30.115Z" },
+    { url = "https://files.pythonhosted.org/packages/f1/62/1748f7e7e1ee41aa52279fd2e3a6d0733dc42a673b16932bad8e5d0c8b28/fonttools-4.61.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d8db08051fc9e7d8bc622f2112511b8107d8f27cd89e2f64ec45e9825e8288da", size = 4897863, upload-time = "2025-12-12T17:29:32.535Z" },
+    { url = "https://files.pythonhosted.org/packages/69/69/4ca02ee367d2c98edcaeb83fc278d20972502ee071214ad9d8ca85e06080/fonttools-4.61.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:a76d4cb80f41ba94a6691264be76435e5f72f2cb3cab0b092a6212855f71c2f6", size = 4859076, upload-time = "2025-12-12T17:29:34.907Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/f5/660f9e3cefa078861a7f099107c6d203b568a6227eef163dd173bfc56bdc/fonttools-4.61.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:a13fc8aeb24bad755eea8f7f9d409438eb94e82cf86b08fe77a03fbc8f6a96b1", size = 4875623, upload-time = "2025-12-12T17:29:37.33Z" },
+    { url = "https://files.pythonhosted.org/packages/63/d1/9d7c5091d2276ed47795c131c1bf9316c3c1ab2789c22e2f59e0572ccd38/fonttools-4.61.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b846a1fcf8beadeb9ea4f44ec5bdde393e2f1569e17d700bfc49cd69bde75881", size = 4993327, upload-time = "2025-12-12T17:29:39.781Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/2d/28def73837885ae32260d07660a052b99f0aa00454867d33745dfe49dbf0/fonttools-4.61.1-cp310-cp310-win32.whl", hash = "sha256:78a7d3ab09dc47ac1a363a493e6112d8cabed7ba7caad5f54dbe2f08676d1b47", size = 1502180, upload-time = "2025-12-12T17:29:42.217Z" },
+    { url = "https://files.pythonhosted.org/packages/63/fa/bfdc98abb4dd2bd491033e85e3ba69a2313c850e759a6daa014bc9433b0f/fonttools-4.61.1-cp310-cp310-win_amd64.whl", hash = "sha256:eff1ac3cc66c2ac7cda1e64b4e2f3ffef474b7335f92fc3833fc632d595fcee6", size = 1550654, upload-time = "2025-12-12T17:29:44.564Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/4e/ce75a57ff3aebf6fc1f4e9d508b8e5810618a33d900ad6c19eb30b290b97/fonttools-4.61.1-py3-none-any.whl", hash = "sha256:17d2bf5d541add43822bcf0c43d7d847b160c9bb01d15d5007d84e2217aaa371", size = 1148996, upload-time = "2025-12-12T17:31:21.03Z" },
+]
+
+[[package]]
+name = "frozenlist"
+version = "1.8.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2d/f5/c831fac6cc817d26fd54c7eaccd04ef7e0288806943f7cc5bbf69f3ac1f0/frozenlist-1.8.0.tar.gz", hash = "sha256:3ede829ed8d842f6cd48fc7081d7a41001a56f1f38603f9d49bf3020d59a31ad", size = 45875, upload-time = "2025-10-06T05:38:17.865Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/83/4a/557715d5047da48d54e659203b9335be7bfaafda2c3f627b7c47e0b3aaf3/frozenlist-1.8.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b37f6d31b3dcea7deb5e9696e529a6aa4a898adc33db82da12e4c60a7c4d2011", size = 86230, upload-time = "2025-10-06T05:35:23.699Z" },
+    { url = "https://files.pythonhosted.org/packages/a2/fb/c85f9fed3ea8fe8740e5b46a59cc141c23b842eca617da8876cfce5f760e/frozenlist-1.8.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ef2b7b394f208233e471abc541cc6991f907ffd47dc72584acee3147899d6565", size = 49621, upload-time = "2025-10-06T05:35:25.341Z" },
+    { url = "https://files.pythonhosted.org/packages/63/70/26ca3f06aace16f2352796b08704338d74b6d1a24ca38f2771afbb7ed915/frozenlist-1.8.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a88f062f072d1589b7b46e951698950e7da00442fc1cacbe17e19e025dc327ad", size = 49889, upload-time = "2025-10-06T05:35:26.797Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/ed/c7895fd2fde7f3ee70d248175f9b6cdf792fb741ab92dc59cd9ef3bd241b/frozenlist-1.8.0-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:f57fb59d9f385710aa7060e89410aeb5058b99e62f4d16b08b91986b9a2140c2", size = 219464, upload-time = "2025-10-06T05:35:28.254Z" },
+    { url = "https://files.pythonhosted.org/packages/6b/83/4d587dccbfca74cb8b810472392ad62bfa100bf8108c7223eb4c4fa2f7b3/frozenlist-1.8.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:799345ab092bee59f01a915620b5d014698547afd011e691a208637312db9186", size = 221649, upload-time = "2025-10-06T05:35:29.454Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/c6/fd3b9cd046ec5fff9dab66831083bc2077006a874a2d3d9247dea93ddf7e/frozenlist-1.8.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c23c3ff005322a6e16f71bf8692fcf4d5a304aaafe1e262c98c6d4adc7be863e", size = 219188, upload-time = "2025-10-06T05:35:30.951Z" },
+    { url = "https://files.pythonhosted.org/packages/ce/80/6693f55eb2e085fc8afb28cf611448fb5b90e98e068fa1d1b8d8e66e5c7d/frozenlist-1.8.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8a76ea0f0b9dfa06f254ee06053d93a600865b3274358ca48a352ce4f0798450", size = 231748, upload-time = "2025-10-06T05:35:32.101Z" },
+    { url = "https://files.pythonhosted.org/packages/97/d6/e9459f7c5183854abd989ba384fe0cc1a0fb795a83c033f0571ec5933ca4/frozenlist-1.8.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c7366fe1418a6133d5aa824ee53d406550110984de7637d65a178010f759c6ef", size = 236351, upload-time = "2025-10-06T05:35:33.834Z" },
+    { url = "https://files.pythonhosted.org/packages/97/92/24e97474b65c0262e9ecd076e826bfd1d3074adcc165a256e42e7b8a7249/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:13d23a45c4cebade99340c4165bd90eeb4a56c6d8a9d8aa49568cac19a6d0dc4", size = 218767, upload-time = "2025-10-06T05:35:35.205Z" },
+    { url = "https://files.pythonhosted.org/packages/ee/bf/dc394a097508f15abff383c5108cb8ad880d1f64a725ed3b90d5c2fbf0bb/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:e4a3408834f65da56c83528fb52ce7911484f0d1eaf7b761fc66001db1646eff", size = 235887, upload-time = "2025-10-06T05:35:36.354Z" },
+    { url = "https://files.pythonhosted.org/packages/40/90/25b201b9c015dbc999a5baf475a257010471a1fa8c200c843fd4abbee725/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:42145cd2748ca39f32801dad54aeea10039da6f86e303659db90db1c4b614c8c", size = 228785, upload-time = "2025-10-06T05:35:37.949Z" },
+    { url = "https://files.pythonhosted.org/packages/84/f4/b5bc148df03082f05d2dd30c089e269acdbe251ac9a9cf4e727b2dbb8a3d/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:e2de870d16a7a53901e41b64ffdf26f2fbb8917b3e6ebf398098d72c5b20bd7f", size = 230312, upload-time = "2025-10-06T05:35:39.178Z" },
+    { url = "https://files.pythonhosted.org/packages/db/4b/87e95b5d15097c302430e647136b7d7ab2398a702390cf4c8601975709e7/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:20e63c9493d33ee48536600d1a5c95eefc870cd71e7ab037763d1fbb89cc51e7", size = 217650, upload-time = "2025-10-06T05:35:40.377Z" },
+    { url = "https://files.pythonhosted.org/packages/e5/70/78a0315d1fea97120591a83e0acd644da638c872f142fd72a6cebee825f3/frozenlist-1.8.0-cp310-cp310-win32.whl", hash = "sha256:adbeebaebae3526afc3c96fad434367cafbfd1b25d72369a9e5858453b1bb71a", size = 39659, upload-time = "2025-10-06T05:35:41.863Z" },
+    { url = "https://files.pythonhosted.org/packages/66/aa/3f04523fb189a00e147e60c5b2205126118f216b0aa908035c45336e27e4/frozenlist-1.8.0-cp310-cp310-win_amd64.whl", hash = "sha256:667c3777ca571e5dbeb76f331562ff98b957431df140b54c85fd4d52eea8d8f6", size = 43837, upload-time = "2025-10-06T05:35:43.205Z" },
+    { url = "https://files.pythonhosted.org/packages/39/75/1135feecdd7c336938bd55b4dc3b0dfc46d85b9be12ef2628574b28de776/frozenlist-1.8.0-cp310-cp310-win_arm64.whl", hash = "sha256:80f85f0a7cc86e7a54c46d99c9e1318ff01f4687c172ede30fd52d19d1da1c8e", size = 39989, upload-time = "2025-10-06T05:35:44.596Z" },
+    { url = "https://files.pythonhosted.org/packages/9a/9a/e35b4a917281c0b8419d4207f4334c8e8c5dbf4f3f5f9ada73958d937dcc/frozenlist-1.8.0-py3-none-any.whl", hash = "sha256:0c18a16eab41e82c295618a77502e17b195883241c563b00f0aa5106fc4eaa0d", size = 13409, upload-time = "2025-10-06T05:38:16.721Z" },
+]
+
+[[package]]
+name = "fsspec"
+version = "2025.9.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/de/e0/bab50af11c2d75c9c4a2a26a5254573c0bd97cea152254401510950486fa/fsspec-2025.9.0.tar.gz", hash = "sha256:19fd429483d25d28b65ec68f9f4adc16c17ea2c7c7bf54ec61360d478fb19c19", size = 304847, upload-time = "2025-09-02T19:10:49.215Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/47/71/70db47e4f6ce3e5c37a607355f80da8860a33226be640226ac52cb05ef2e/fsspec-2025.9.0-py3-none-any.whl", hash = "sha256:530dc2a2af60a414a832059574df4a6e10cce927f6f4a78209390fe38955cfb7", size = 199289, upload-time = "2025-09-02T19:10:47.708Z" },
+]
+
+[package.optional-dependencies]
+http = [
+    { name = "aiohttp" },
+]
+
+[[package]]
+name = "gast"
+version = "0.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/91/f6/e73969782a2ecec280f8a176f2476149dd9dba69d5f8779ec6108a7721e6/gast-0.7.0.tar.gz", hash = "sha256:0bb14cd1b806722e91ddbab6fb86bba148c22b40e7ff11e248974e04c8adfdae", size = 33630, upload-time = "2025-11-29T15:30:05.266Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1d/33/f1c6a276de27b7d7339a34749cc33fa87f077f921969c47185d34a887ae2/gast-0.7.0-py3-none-any.whl", hash = "sha256:99cbf1365633a74099f69c59bd650476b96baa5ef196fec88032b00b31ba36f7", size = 22966, upload-time = "2025-11-29T15:30:03.983Z" },
+]
+
+[[package]]
+name = "gitdb"
+version = "4.0.12"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "smmap" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/72/94/63b0fc47eb32792c7ba1fe1b694daec9a63620db1e313033d18140c2320a/gitdb-4.0.12.tar.gz", hash = "sha256:5ef71f855d191a3326fcfbc0d5da835f26b13fbcba60c32c21091c349ffdb571", size = 394684, upload-time = "2025-01-02T07:20:46.413Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/61/5c78b91c3143ed5c14207f463aecfc8f9dbb5092fb2869baf37c273b2705/gitdb-4.0.12-py3-none-any.whl", hash = "sha256:67073e15955400952c6565cc3e707c554a4eea2e428946f7a4c162fab9bd9bcf", size = 62794, upload-time = "2025-01-02T07:20:43.624Z" },
+]
+
+[[package]]
+name = "gitpython"
+version = "3.1.46"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "gitdb" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/df/b5/59d16470a1f0dfe8c793f9ef56fd3826093fc52b3bd96d6b9d6c26c7e27b/gitpython-3.1.46.tar.gz", hash = "sha256:400124c7d0ef4ea03f7310ac2fbf7151e09ff97f2a3288d64a440c584a29c37f", size = 215371, upload-time = "2026-01-01T15:37:32.073Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6a/09/e21df6aef1e1ffc0c816f0522ddc3f6dcded766c3261813131c78a704470/gitpython-3.1.46-py3-none-any.whl", hash = "sha256:79812ed143d9d25b6d176a10bb511de0f9c67b1fa641d82097b0ab90398a2058", size = 208620, upload-time = "2026-01-01T15:37:30.574Z" },
+]
+
+[[package]]
+name = "glfw"
+version = "2.10.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/96/72/642d4f12f61816ac96777f7360d413e3977a7dd08237d196f02da681b186/glfw-2.10.0.tar.gz", hash = "sha256:801e55d8581b34df9aa2cfea43feb06ff617576e2a8cc5dac23ee75b26d10abe", size = 31475, upload-time = "2025-09-12T08:54:38.871Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3d/1f/a9ce08b1173b0ab625ee92f0c47a5278b3e76fd367699880d8ee7d56c338/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-macosx_10_6_intel.whl", hash = "sha256:5f365a8c94bcea71ec91327e7c16e7cf739128479a18b8c1241b004b40acc412", size = 105329, upload-time = "2025-09-12T08:54:27.938Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/96/5a2220abcbd027eebcf8bedd28207a2de168899e51be13ba01ebdd4147a1/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-macosx_11_0_arm64.whl", hash = "sha256:5328db1a92d07abd988730517ec02aa8390d3e6ef7ce98c8b57ecba2f43a39ba", size = 102179, upload-time = "2025-09-12T08:54:29.163Z" },
+    { url = "https://files.pythonhosted.org/packages/9d/41/a5bd1d9e1808f400102bd7d328c4ac17b65fb2fc8014014ec6f23d02f662/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-manylinux2014_aarch64.whl", hash = "sha256:312c4c1dd5509613ed6bc1e95a8dbb75a36b6dcc4120f50dc3892b40172e9053", size = 230039, upload-time = "2025-09-12T08:54:30.201Z" },
+    { url = "https://files.pythonhosted.org/packages/80/aa/3b503c448609dee6cb4e7138b4109338f0e65b97be107ab85562269d378d/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-manylinux2014_x86_64.whl", hash = "sha256:59c53387dc08c62e8bed86bbe3a8d53ab1b27161281ffa0e7f27b64284e2627c", size = 241984, upload-time = "2025-09-12T08:54:31.347Z" },
+    { url = "https://files.pythonhosted.org/packages/ac/2d/bfe39a42cad8e80b02bf5f7cae19ba67832c1810bbd3624a8e83153d74a4/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-manylinux_2_28_aarch64.whl", hash = "sha256:c6f292fdaf3f9a99e598ede6582d21c523a6f51f8f5e66213849101a6bcdc699", size = 231052, upload-time = "2025-09-12T08:54:32.859Z" },
+    { url = "https://files.pythonhosted.org/packages/f7/02/6e639e90f181dc9127046e00d0528f9f7ad12d428972e3a5378b9aefdb0b/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-manylinux_2_28_x86_64.whl", hash = "sha256:7916034efa867927892635733a3b6af8cd95ceb10566fd7f1e0d2763c2ee8b12", size = 243525, upload-time = "2025-09-12T08:54:34.006Z" },
+    { url = "https://files.pythonhosted.org/packages/84/06/cb588ca65561defe0fc48d1df4c2ac12569b81231ae4f2b52ab37007d0bd/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-win32.whl", hash = "sha256:6c9549da71b93e367b4d71438798daae1da2592039fd14204a80a1a2348ae127", size = 552685, upload-time = "2025-09-12T08:54:35.723Z" },
+    { url = "https://files.pythonhosted.org/packages/86/27/00c9c96af18ac0a5eac2ff61cbe306551a2d770d7173f396d0792ee1a59e/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.p39.p310.p311.p312.p313-none-win_amd64.whl", hash = "sha256:6292d5d6634d668cd23d337e6089491d3945a9aa4ac6e1667b0003520d7caa51", size = 559466, upload-time = "2025-09-12T08:54:37.661Z" },
+    { url = "https://files.pythonhosted.org/packages/b3/87/de0b33f6f00687499ca1371f22aa73396341b85bf88f1a284f9da8842493/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-macosx_10_6_intel.whl", hash = "sha256:2aab89d2d9535635ba011fc7303390685169a1aa6731ad580d08d043524b8899", size = 105326, upload-time = "2026-01-28T05:57:56.083Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/a6/6ea2f73ad4474896d9e38b3ffbe6ffd5a802c738392269e99e8c6621a461/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-macosx_11_0_arm64.whl", hash = "sha256:23936202a107039b5372f0b88ae1d11080746aa1c78910a45d4a0c4cf408cfaa", size = 102180, upload-time = "2026-01-28T05:57:57.787Z" },
+    { url = "https://files.pythonhosted.org/packages/58/19/d81b19e8261b9cb51b81d1402167791fef81088dfe91f0c4e9d136fdc5ca/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-manylinux2014_aarch64.whl", hash = "sha256:7be06d0838f61df67bd54cb6266a6193d54083acb3624ff3c3812a6358406fa4", size = 230038, upload-time = "2026-01-28T05:57:59.105Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/fa/b035636cd82198b97b51a93efe9cfc4343d6b15cefbd336a3f2be871d848/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-manylinux2014_x86_64.whl", hash = "sha256:91d36b3582a766512eff8e3b5dcc2d3ffcbf10b7cf448551085a08a10f1b8244", size = 241983, upload-time = "2026-01-28T05:58:00.352Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/b4/f7b6cc022dd7c68b6c702d19da5d591f978f89c958b9bd3090615db0c739/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-manylinux_2_28_aarch64.whl", hash = "sha256:27c9e9a2d5e1dc3c9e3996171d844d9df9a5a101e797cb94cce217b7afcf8fd9", size = 231053, upload-time = "2026-01-28T05:58:01.683Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/3f/efeb7c6801c46e11bd666a5180f0d615f74f72264212f74f39586c6fda9d/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-manylinux_2_28_x86_64.whl", hash = "sha256:ce6724bb7cb3d0543dcba17206dce909f94176e68220b8eafee72e9f92bcf542", size = 243522, upload-time = "2026-01-28T05:58:03.517Z" },
+    { url = "https://files.pythonhosted.org/packages/cf/b9/b04c3aa0aad2870cfe799f32f8b59789c98e1816bbce9e83f4823c5b840b/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-win32.whl", hash = "sha256:fca724a21a372731edb290841edd28a9fb1ee490f833392752844ac807c0086a", size = 552682, upload-time = "2026-01-28T05:58:05.649Z" },
+    { url = "https://files.pythonhosted.org/packages/bd/e1/6d6816b296a529ac9b897ad228b1e084eb1f92319e96371880eebdc874a6/glfw-2.10.0-py2.py27.py3.py30.py31.py32.py33.py34.py35.py36.py37.py38.py39.py310.py311.py312.py313.py314-none-win_amd64.whl", hash = "sha256:823c0bd7770977d4b10e0ed0aef2f3682276b7c88b8b65cfc540afce5951392f", size = 559464, upload-time = "2026-01-28T05:58:07.261Z" },
+]
+
+[[package]]
+name = "google-ai-generativelanguage"
+version = "0.6.15"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "google-api-core", extra = ["grpc"] },
+    { name = "google-auth" },
+    { name = "proto-plus" },
+    { name = "protobuf" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/11/d1/48fe5d7a43d278e9f6b5ada810b0a3530bbeac7ed7fcbcd366f932f05316/google_ai_generativelanguage-0.6.15.tar.gz", hash = "sha256:8f6d9dc4c12b065fe2d0289026171acea5183ebf2d0b11cefe12f3821e159ec3", size = 1375443, upload-time = "2025-01-13T21:50:47.459Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7c/a3/67b8a6ff5001a1d8864922f2d6488dc2a14367ceb651bc3f09a947f2f306/google_ai_generativelanguage-0.6.15-py3-none-any.whl", hash = "sha256:5a03ef86377aa184ffef3662ca28f19eeee158733e45d7947982eb953c6ebb6c", size = 1327356, upload-time = "2025-01-13T21:50:44.174Z" },
+]
+
+[[package]]
+name = "google-api-core"
+version = "2.29.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "google-auth" },
+    { name = "googleapis-common-protos" },
+    { name = "proto-plus" },
+    { name = "protobuf" },
+    { name = "requests" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0d/10/05572d33273292bac49c2d1785925f7bc3ff2fe50e3044cf1062c1dde32e/google_api_core-2.29.0.tar.gz", hash = "sha256:84181be0f8e6b04006df75ddfe728f24489f0af57c96a529ff7cf45bc28797f7", size = 177828, upload-time = "2026-01-08T22:21:39.269Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/77/b6/85c4d21067220b9a78cfb81f516f9725ea6befc1544ec9bd2c1acd97c324/google_api_core-2.29.0-py3-none-any.whl", hash = "sha256:d30bc60980daa36e314b5d5a3e5958b0200cb44ca8fa1be2b614e932b75a3ea9", size = 173906, upload-time = "2026-01-08T22:21:36.093Z" },
+]
+
+[package.optional-dependencies]
+grpc = [
+    { name = "grpcio" },
+    { name = "grpcio-status" },
+]
+
+[[package]]
+name = "google-api-python-client"
+version = "2.192.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "google-api-core" },
+    { name = "google-auth" },
+    { name = "google-auth-httplib2" },
+    { name = "httplib2" },
+    { name = "uritemplate" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/85/d8/489052a40935e45b9b5b3d6accc14b041360c1507bdc659c2e1a19aaa3ff/google_api_python_client-2.192.0.tar.gz", hash = "sha256:d48cfa6078fadea788425481b007af33fe0ab6537b78f37da914fb6fc112eb27", size = 14209505, upload-time = "2026-03-05T15:17:01.598Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e0/76/ec4128f00fefb9011635ae2abc67d7dacd05c8559378f8f05f0c907c38d8/google_api_python_client-2.192.0-py3-none-any.whl", hash = "sha256:63a57d4457cd97df1d63eb89c5fda03c5a50588dcbc32c0115dd1433c08f4b62", size = 14783267, upload-time = "2026-03-05T15:16:58.804Z" },
+]
+
+[[package]]
+name = "google-auth"
+version = "2.49.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cryptography" },
+    { name = "pyasn1-modules" },
+    { name = "rsa" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7d/59/7371175bfd949abfb1170aa076352131d7281bd9449c0f978604fc4431c3/google_auth-2.49.0.tar.gz", hash = "sha256:9cc2d9259d3700d7a257681f81052db6737495a1a46b610597f4b8bafe5286ae", size = 333444, upload-time = "2026-03-06T21:53:06.07Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/37/45/de64b823b639103de4b63dd193480dce99526bd36be6530c2dba85bf7817/google_auth-2.49.0-py3-none-any.whl", hash = "sha256:f893ef7307f19cf53700b7e2f61b5a6affe3aa0edf9943b13788920ab92d8d87", size = 240676, upload-time = "2026-03-06T21:52:38.304Z" },
+]
+
+[[package]]
+name = "google-auth-httplib2"
+version = "0.3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "google-auth" },
+    { name = "httplib2" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d5/ad/c1f2b1175096a8d04cf202ad5ea6065f108d26be6fc7215876bde4a7981d/google_auth_httplib2-0.3.0.tar.gz", hash = "sha256:177898a0175252480d5ed916aeea183c2df87c1f9c26705d74ae6b951c268b0b", size = 11134, upload-time = "2025-12-15T22:13:51.825Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/99/d5/3c97526c8796d3caf5f4b3bed2b05e8a7102326f00a334e7a438237f3b22/google_auth_httplib2-0.3.0-py3-none-any.whl", hash = "sha256:426167e5df066e3f5a0fc7ea18768c08e7296046594ce4c8c409c2457dd1f776", size = 9529, upload-time = "2025-12-15T22:13:51.048Z" },
+]
+
+[[package]]
+name = "google-generativeai"
+version = "0.8.6"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "google-ai-generativelanguage" },
+    { name = "google-api-core" },
+    { name = "google-api-python-client" },
+    { name = "google-auth" },
+    { name = "protobuf" },
+    { name = "pydantic" },
+    { name = "tqdm" },
+    { name = "typing-extensions" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/97/0f/ef33b5bb71437966590c6297104c81051feae95d54b11ece08533ef937d3/google_generativeai-0.8.6-py3-none-any.whl", hash = "sha256:37a0eaaa95e5bbf888828e20a4a1b2c196cc9527d194706e58a68ff388aeb0fa", size = 155098, upload-time = "2025-12-16T17:53:58.61Z" },
+]
+
+[[package]]
+name = "google-pasta"
+version = "0.2.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/35/4a/0bd53b36ff0323d10d5f24ebd67af2de10a1117f5cf4d7add90df92756f1/google-pasta-0.2.0.tar.gz", hash = "sha256:c9f2c8dfc8f96d0d5808299920721be30c9eec37f2389f28904f454565c8a16e", size = 40430, upload-time = "2020-03-13T18:57:50.34Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a3/de/c648ef6835192e6e2cc03f40b19eeda4382c49b5bafb43d88b931c4c74ac/google_pasta-0.2.0-py3-none-any.whl", hash = "sha256:b32482794a366b5366a32c92a9a9201b107821889935a02b3e51f6b432ea84ed", size = 57471, upload-time = "2020-03-13T18:57:48.872Z" },
+]
+
+[[package]]
+name = "googleapis-common-protos"
+version = "1.73.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "protobuf" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/99/96/a0205167fa0154f4a542fd6925bdc63d039d88dab3588b875078107e6f06/googleapis_common_protos-1.73.0.tar.gz", hash = "sha256:778d07cd4fbeff84c6f7c72102f0daf98fa2bfd3fa8bea426edc545588da0b5a", size = 147323, upload-time = "2026-03-06T21:53:09.727Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/69/28/23eea8acd65972bbfe295ce3666b28ac510dfcb115fac089d3edb0feb00a/googleapis_common_protos-1.73.0-py3-none-any.whl", hash = "sha256:dfdaaa2e860f242046be561e6d6cb5c5f1541ae02cfbcb034371aadb2942b4e8", size = 297578, upload-time = "2026-03-06T21:52:33.933Z" },
+]
+
+[[package]]
+name = "gradio"
+version = "6.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aiofiles" },
+    { name = "anyio" },
+    { name = "brotli" },
+    { name = "fastapi" },
+    { name = "ffmpy" },
+    { name = "gradio-client" },
+    { name = "groovy" },
+    { name = "httpx" },
+    { name = "huggingface-hub" },
+    { name = "jinja2" },
+    { name = "markupsafe" },
+    { name = "numpy" },
+    { name = "orjson" },
+    { name = "packaging" },
+    { name = "pandas" },
+    { name = "pillow" },
+    { name = "pydantic" },
+    { name = "pydub" },
+    { name = "python-multipart" },
+    { name = "pytz" },
+    { name = "pyyaml" },
+    { name = "safehttpx" },
+    { name = "semantic-version" },
+    { name = "starlette" },
+    { name = "tomlkit" },
+    { name = "typer" },
+    { name = "typing-extensions" },
+    { name = "uvicorn" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/bd/83/29bdbf94b212512e3c775482d390f5b699a72d71a2c431dea367a6e45a37/gradio-6.9.0.tar.gz", hash = "sha256:593e60e33233f3586452ebfa9f741817c5ae849a98cc70945f3ccb8dc895eb22", size = 57904480, upload-time = "2026-03-06T17:44:26.025Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b3/8b/dc357ab966544e4dc898a2fee326d755c5f54da82af71a1a802e3476e78e/gradio-6.9.0-py3-none-any.whl", hash = "sha256:c173dd330c9247002a42222c85d76c0ecee65437eff808084e360862e7bbd24f", size = 42940853, upload-time = "2026-03-06T17:44:22.009Z" },
+]
+
+[[package]]
+name = "gradio-client"
+version = "2.3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "fsspec" },
+    { name = "httpx" },
+    { name = "huggingface-hub" },
+    { name = "packaging" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/97/d2/de2037f5eff13a5145cdf6982fd34c9735f0806e8a2ee5d4bfe9a7d25a54/gradio_client-2.3.0.tar.gz", hash = "sha256:1c700dc60e65bae4386ba7cf3732b9f9d5bcf5fb8eb451df3944fe092d7d9a29", size = 57552, upload-time = "2026-03-06T17:44:38.247Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/99/6a/41752781399811afbf8ac858f63c20eff354ed35169daa39604aefced4e8/gradio_client-2.3.0-py3-none-any.whl", hash = "sha256:9ec51a927888fc188e123a0ac5ad341d9265b325539a399554d1fc2604942e74", size = 58531, upload-time = "2026-03-06T17:44:36.961Z" },
+]
+
+[[package]]
+name = "groovy"
+version = "0.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/52/36/bbdede67400277bef33d3ec0e6a31750da972c469f75966b4930c753218f/groovy-0.1.2.tar.gz", hash = "sha256:25c1dc09b3f9d7e292458aa762c6beb96ea037071bf5e917fc81fb78d2231083", size = 17325, upload-time = "2025-02-28T20:24:56.068Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/28/27/3d6dcadc8a3214d8522c1e7f6a19554e33659be44546d44a2f7572ac7d2a/groovy-0.1.2-py3-none-any.whl", hash = "sha256:7f7975bab18c729a257a8b1ae9dcd70b7cafb1720481beae47719af57c35fa64", size = 14090, upload-time = "2025-02-28T20:24:55.152Z" },
+]
+
+[[package]]
+name = "grpcio"
+version = "1.78.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/06/8a/3d098f35c143a89520e568e6539cc098fcd294495910e359889ce8741c84/grpcio-1.78.0.tar.gz", hash = "sha256:7382b95189546f375c174f53a5fa873cef91c4b8005faa05cc5b3beea9c4f1c5", size = 12852416, upload-time = "2026-02-06T09:57:18.093Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5a/a8/690a085b4d1fe066130de97a87de32c45062cf2ecd218df9675add895550/grpcio-1.78.0-cp310-cp310-linux_armv7l.whl", hash = "sha256:7cc47943d524ee0096f973e1081cb8f4f17a4615f2116882a5f1416e4cfe92b5", size = 5946986, upload-time = "2026-02-06T09:54:34.043Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/1b/e5213c5c0ced9d2d92778d30529ad5bb2dcfb6c48c4e2d01b1f302d33d64/grpcio-1.78.0-cp310-cp310-macosx_11_0_universal2.whl", hash = "sha256:c3f293fdc675ccba4db5a561048cca627b5e7bd1c8a6973ffedabe7d116e22e2", size = 11816533, upload-time = "2026-02-06T09:54:37.04Z" },
+    { url = "https://files.pythonhosted.org/packages/18/37/1ba32dccf0a324cc5ace744c44331e300b000a924bf14840f948c559ede7/grpcio-1.78.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:10a9a644b5dd5aec3b82b5b0b90d41c0fa94c85ef42cb42cf78a23291ddb5e7d", size = 6519964, upload-time = "2026-02-06T09:54:40.268Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/f5/c0e178721b818072f2e8b6fde13faaba942406c634009caf065121ce246b/grpcio-1.78.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:4c5533d03a6cbd7f56acfc9cfb44ea64f63d29091e40e44010d34178d392d7eb", size = 7198058, upload-time = "2026-02-06T09:54:42.389Z" },
+    { url = "https://files.pythonhosted.org/packages/5b/b2/40d43c91ae9cd667edc960135f9f08e58faa1576dc95af29f66ec912985f/grpcio-1.78.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ff870aebe9a93a85283837801d35cd5f8814fe2ad01e606861a7fb47c762a2b7", size = 6727212, upload-time = "2026-02-06T09:54:44.91Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/88/9da42eed498f0efcfcd9156e48ae63c0cde3bea398a16c99fb5198c885b6/grpcio-1.78.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:391e93548644e6b2726f1bb84ed60048d4bcc424ce5e4af0843d28ca0b754fec", size = 7300845, upload-time = "2026-02-06T09:54:47.562Z" },
+    { url = "https://files.pythonhosted.org/packages/23/3f/1c66b7b1b19a8828890e37868411a6e6925df5a9030bfa87ab318f34095d/grpcio-1.78.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:df2c8f3141f7cbd112a6ebbd760290b5849cda01884554f7c67acc14e7b1758a", size = 8284605, upload-time = "2026-02-06T09:54:50.475Z" },
+    { url = "https://files.pythonhosted.org/packages/94/c4/ca1bd87394f7b033e88525384b4d1e269e8424ab441ea2fba1a0c5b50986/grpcio-1.78.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:bd8cb8026e5f5b50498a3c4f196f57f9db344dad829ffae16b82e4fdbaea2813", size = 7726672, upload-time = "2026-02-06T09:54:53.11Z" },
+    { url = "https://files.pythonhosted.org/packages/41/09/f16e487d4cc65ccaf670f6ebdd1a17566b965c74fc3d93999d3b2821e052/grpcio-1.78.0-cp310-cp310-win32.whl", hash = "sha256:f8dff3d9777e5d2703a962ee5c286c239bf0ba173877cc68dc02c17d042e29de", size = 4076715, upload-time = "2026-02-06T09:54:55.549Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/32/4ce60d94e242725fd3bcc5673c04502c82a8e87b21ea411a63992dc39f8f/grpcio-1.78.0-cp310-cp310-win_amd64.whl", hash = "sha256:94f95cf5d532d0e717eed4fc1810e8e6eded04621342ec54c89a7c2f14b581bf", size = 4799157, upload-time = "2026-02-06T09:54:59.838Z" },
+]
+
+[[package]]
+name = "grpcio-status"
+version = "1.48.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "googleapis-common-protos" },
+    { name = "grpcio" },
+    { name = "protobuf" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/22/62/a86443ec8f7bf635fe0b48abe56cd699816bdc0b29d24e0bcb5cada42d4a/grpcio-status-1.48.2.tar.gz", hash = "sha256:53695f45da07437b7c344ee4ef60d370fd2850179f5a28bb26d8e2aa1102ec11", size = 13485, upload-time = "2022-09-22T02:50:32.913Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/32/754cd4474790239c7436a7a9490bc0c4a0a2ed604cb9a940151a3b1055b9/grpcio_status-1.48.2-py3-none-any.whl", hash = "sha256:2c33bbdbe20188b2953f46f31af669263b6ee2a9b2d38fa0d36ee091532e21bf", size = 14441, upload-time = "2022-09-22T02:47:51.828Z" },
+]
+
+[[package]]
+name = "gymnasium"
+version = "1.2.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cloudpickle" },
+    { name = "farama-notifications" },
+    { name = "numpy" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/76/59/653a9417d98ed3e29ef9734ba52c3495f6c6823b8d5c0c75369f25111708/gymnasium-1.2.3.tar.gz", hash = "sha256:2b2cb5b5fbbbdf3afb9f38ca952cc48aa6aa3e26561400d940747fda3ad42509", size = 829230, upload-time = "2025-12-18T16:51:10.234Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/56/d3/ea5f088e3638dbab12e5c20d6559d5b3bdaeaa1f2af74e526e6815836285/gymnasium-1.2.3-py3-none-any.whl", hash = "sha256:e6314bba8f549c7fdcc8677f7cd786b64908af6e79b57ddaa5ce1825bffb5373", size = 952113, upload-time = "2025-12-18T16:51:08.445Z" },
+]
+
+[[package]]
+name = "h11"
+version = "0.16.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/01/ee/02a2c011bdab74c6fb3c75474d40b3052059d95df7e73351460c8588d963/h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1", size = 101250, upload-time = "2025-04-24T03:35:25.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/04/4b/29cac41a4d98d144bf5f6d33995617b185d14b22401f75ca86f384e87ff1/h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86", size = 37515, upload-time = "2025-04-24T03:35:24.344Z" },
+]
+
+[[package]]
+name = "h5py"
+version = "3.16.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/db/33/acd0ce6863b6c0d7735007df01815403f5589a21ff8c2e1ee2587a38f548/h5py-3.16.0.tar.gz", hash = "sha256:a0dbaad796840ccaa67a4c144a0d0c8080073c34c76d5a6941d6818678ef2738", size = 446526, upload-time = "2026-03-06T13:49:08.07Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3a/6b/231413e58a787a89b316bb0d1777da3c62257e4797e09afd8d17ad3549dc/h5py-3.16.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e06f864bedb2c8e7c1358e6c73af48519e317457c444d6f3d332bb4e8fa6d7d9", size = 3724137, upload-time = "2026-03-06T13:47:35.242Z" },
+    { url = "https://files.pythonhosted.org/packages/74/f9/557ce3aad0fe8471fb5279bab0fc56ea473858a022c4ce8a0b8f303d64e9/h5py-3.16.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ec86d4fffd87a0f4cb3d5796ceb5a50123a2a6d99b43e616e5504e66a953eca3", size = 3090112, upload-time = "2026-03-06T13:47:37.634Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/f5/e15b3d0dc8a18e56409a839e6468d6fb589bc5207c917399c2e0706eeb44/h5py-3.16.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:86385ea895508220b8a7e45efa428aeafaa586bd737c7af9ee04661d8d84a10d", size = 4844847, upload-time = "2026-03-06T13:47:39.811Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/92/a8851d936547efe30cc0ce5245feac01f3ec6171f7899bc3f775c72030b3/h5py-3.16.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:8975273c2c5921c25700193b408e28d6bdd0111c37468b2d4e25dcec4cd1d84d", size = 5065352, upload-time = "2026-03-06T13:47:41.489Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/ae/f2adc5d0ca9626db3277a3d87516e124cbc5d0eea0bd79bc085702d04f2c/h5py-3.16.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:1677ad48b703f44efc9ea0c3ab284527f81bc4f318386aaaebc5fede6bbae56f", size = 4839173, upload-time = "2026-03-06T13:47:43.586Z" },
+    { url = "https://files.pythonhosted.org/packages/64/0b/e0c8c69da1d8838da023a50cd3080eae5d475691f7636b35eff20bb6ef20/h5py-3.16.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7c4dd4cf5f0a4e36083f73172f6cfc25a5710789269547f132a20975bfe2434c", size = 5076216, upload-time = "2026-03-06T13:47:45.315Z" },
+    { url = "https://files.pythonhosted.org/packages/66/35/d88fd6718832133c885004c61ceeeb24dbd6397ef877dbed6b3a64d6a286/h5py-3.16.0-cp310-cp310-win_amd64.whl", hash = "sha256:bdef06507725b455fccba9c16529121a5e1fbf56aa375f7d9713d9e8ff42454d", size = 3183639, upload-time = "2026-03-06T13:47:47.041Z" },
+]
+
+[[package]]
+name = "hf-transfer"
+version = "0.1.9"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1a/eb/8fc64f40388c29ce8ce3b2b180a089d4d6b25b1d0d232d016704cb852104/hf_transfer-0.1.9.tar.gz", hash = "sha256:035572865dab29d17e783fbf1e84cf1cb24f3fcf8f1b17db1cfc7fdf139f02bf", size = 25201, upload-time = "2025-01-07T10:05:12.947Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/81/f5/461d2e5f307e5048289b1168d5c642ae3bb2504e88dff1a38b92ed990a21/hf_transfer-0.1.9-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:e66acf91df4a8b72f60223059df3003062a5ae111757187ed1a06750a30e911b", size = 1393046, upload-time = "2025-01-07T10:04:51.003Z" },
+    { url = "https://files.pythonhosted.org/packages/41/ba/8d9fd9f1083525edfcb389c93738c802f3559cb749324090d7109c8bf4c2/hf_transfer-0.1.9-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:8669dbcc7a3e2e8d61d42cd24da9c50d57770bd74b445c65123291ca842a7e7a", size = 1348126, upload-time = "2025-01-07T10:04:45.712Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/a2/cd7885bc9959421065a6fae0fe67b6c55becdeda4e69b873e52976f9a9f0/hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8fd0167c4407a3bc4cdd0307e65ada2294ec04f1813d8a69a5243e379b22e9d8", size = 3728604, upload-time = "2025-01-07T10:04:14.173Z" },
+    { url = "https://files.pythonhosted.org/packages/f6/2e/a072cf196edfeda3310c9a5ade0a0fdd785e6154b3ce24fc738c818da2a7/hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ee8b10afedcb75f71091bcc197c526a6ebf5c58bbbadb34fdeee6160f55f619f", size = 3064995, upload-time = "2025-01-07T10:04:18.663Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/84/aec9ef4c0fab93c1ea2b1badff38c78b4b2f86f0555b26d2051dbc920cde/hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5828057e313de59300dd1abb489444bc452efe3f479d3c55b31a8f680936ba42", size = 3580908, upload-time = "2025-01-07T10:04:32.834Z" },
+    { url = "https://files.pythonhosted.org/packages/29/63/b560d39651a56603d64f1a0212d0472a44cbd965db2fa62b99d99cb981bf/hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:fc6bd19e1cc177c66bdef15ef8636ad3bde79d5a4f608c158021153b4573509d", size = 3400839, upload-time = "2025-01-07T10:04:26.122Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/d8/f87ea6f42456254b48915970ed98e993110521e9263472840174d32c880d/hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cdca9bfb89e6f8f281890cc61a8aff2d3cecaff7e1a4d275574d96ca70098557", size = 3552664, upload-time = "2025-01-07T10:04:40.123Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/56/1267c39b65fc8f4e2113b36297320f102718bf5799b544a6cbe22013aa1d/hf_transfer-0.1.9-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:89a23f58b7b7effbc047b8ca286f131b17728c99a9f972723323003ffd1bb916", size = 4073732, upload-time = "2025-01-07T10:04:55.624Z" },
+    { url = "https://files.pythonhosted.org/packages/82/1a/9c748befbe3decf7cb415e34f8a0c3789a0a9c55910dea73d581e48c0ce5/hf_transfer-0.1.9-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:dc7fff1345980d6c0ebb92c811d24afa4b98b3e07ed070c8e38cc91fd80478c5", size = 3390096, upload-time = "2025-01-07T10:04:59.98Z" },
+    { url = "https://files.pythonhosted.org/packages/72/85/4c03da147b6b4b7cb12e074d3d44eee28604a387ed0eaf7eaaead5069c57/hf_transfer-0.1.9-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:1a6bd16c667ebe89a069ca163060127a794fa3a3525292c900b8c8cc47985b0d", size = 3664743, upload-time = "2025-01-07T10:05:05.416Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/6e/e597b04f753f1b09e6893075d53a82a30c13855cbaa791402695b01e369f/hf_transfer-0.1.9-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:d2fde99d502093ade3ab1b53f80da18480e9902aa960dab7f74fb1b9e5bc5746", size = 3695243, upload-time = "2025-01-07T10:05:11.411Z" },
+    { url = "https://files.pythonhosted.org/packages/09/89/d4e234727a26b2546c8fb70a276cd924260d60135f2165bf8b9ed67bb9a4/hf_transfer-0.1.9-cp38-abi3-win32.whl", hash = "sha256:435cc3cdc8524ce57b074032b8fd76eed70a4224d2091232fa6a8cef8fd6803e", size = 1086605, upload-time = "2025-01-07T10:05:18.873Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/14/f1e15b851d1c2af5b0b1a82bf8eb10bda2da62d98180220ba6fd8879bb5b/hf_transfer-0.1.9-cp38-abi3-win_amd64.whl", hash = "sha256:16f208fc678911c37e11aa7b586bc66a37d02e636208f18b6bc53d29b5df40ad", size = 1160240, upload-time = "2025-01-07T10:05:14.324Z" },
+]
+
+[[package]]
+name = "hf-xet"
+version = "1.3.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/8b/cb/9bb543bd987ffa1ee48202cc96a756951b734b79a542335c566148ade36c/hf_xet-1.3.2.tar.gz", hash = "sha256:e130ee08984783d12717444e538587fa2119385e5bd8fc2bb9f930419b73a7af", size = 643646, upload-time = "2026-02-27T17:26:08.051Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d8/28/dbb024e2e3907f6f3052847ca7d1a2f7a3972fafcd53ff79018977fcb3e4/hf_xet-1.3.2-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:f93b7595f1d8fefddfede775c18b5c9256757824f7f6832930b49858483cd56f", size = 3763961, upload-time = "2026-02-27T17:25:52.537Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/71/b99aed3823c9d1795e4865cf437d651097356a3f38c7d5877e4ac544b8e4/hf_xet-1.3.2-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:a85d3d43743174393afe27835bde0cd146e652b5fcfdbcd624602daef2ef3259", size = 3526171, upload-time = "2026-02-27T17:25:50.968Z" },
+    { url = "https://files.pythonhosted.org/packages/9d/ca/907890ce6ef5598b5920514f255ed0a65f558f820515b18db75a51b2f878/hf_xet-1.3.2-cp37-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:7c2a054a97c44e136b1f7f5a78f12b3efffdf2eed3abc6746fc5ea4b39511633", size = 4180750, upload-time = "2026-02-27T17:25:43.125Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/ad/bc7f41f87173d51d0bce497b171c4ee0cbde1eed2d7b4216db5d0ada9f50/hf_xet-1.3.2-cp37-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:06b724a361f670ae557836e57801b82c75b534812e351a87a2c739f77d1e0635", size = 3961035, upload-time = "2026-02-27T17:25:41.837Z" },
+    { url = "https://files.pythonhosted.org/packages/73/38/600f4dda40c4a33133404d9fe644f1d35ff2d9babb4d0435c646c63dd107/hf_xet-1.3.2-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:305f5489d7241a47e0458ef49334be02411d1d0f480846363c1c8084ed9916f7", size = 4161378, upload-time = "2026-02-27T17:26:00.365Z" },
+    { url = "https://files.pythonhosted.org/packages/00/b3/7bc1ff91d1ac18420b7ad1e169b618b27c00001b96310a89f8a9294fe509/hf_xet-1.3.2-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:06cdbde243c85f39a63b28e9034321399c507bcd5e7befdd17ed2ccc06dfe14e", size = 4398020, upload-time = "2026-02-27T17:26:03.977Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/0b/99bfd948a3ed3620ab709276df3ad3710dcea61976918cce8706502927af/hf_xet-1.3.2-cp37-abi3-win_amd64.whl", hash = "sha256:9298b47cce6037b7045ae41482e703c471ce36b52e73e49f71226d2e8e5685a1", size = 3641624, upload-time = "2026-02-27T17:26:13.542Z" },
+    { url = "https://files.pythonhosted.org/packages/cc/02/9a6e4ca1f3f73a164c0cd48e41b3cc56585dcc37e809250de443d673266f/hf_xet-1.3.2-cp37-abi3-win_arm64.whl", hash = "sha256:83d8ec273136171431833a6957e8f3af496bee227a0fe47c7b8b39c106d1749a", size = 3503976, upload-time = "2026-02-27T17:26:12.123Z" },
+]
+
+[[package]]
+name = "httpcore"
+version = "1.0.9"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi" },
+    { name = "h11" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/06/94/82699a10bca87a5556c9c59b5963f2d039dbd239f25bc2a63907a05a14cb/httpcore-1.0.9.tar.gz", hash = "sha256:6e34463af53fd2ab5d807f399a9b45ea31c3dfa2276f15a2c3f00afff6e176e8", size = 85484, upload-time = "2025-04-24T22:06:22.219Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7e/f5/f66802a942d491edb555dd61e3a9961140fd64c90bce1eafd741609d334d/httpcore-1.0.9-py3-none-any.whl", hash = "sha256:2d400746a40668fc9dec9810239072b40b4484b640a8c38fd654a024c7a1bf55", size = 78784, upload-time = "2025-04-24T22:06:20.566Z" },
+]
+
+[[package]]
+name = "httplib2"
+version = "0.31.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyparsing" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c1/1f/e86365613582c027dda5ddb64e1010e57a3d53e99ab8a72093fa13d565ec/httplib2-0.31.2.tar.gz", hash = "sha256:385e0869d7397484f4eab426197a4c020b606edd43372492337c0b4010ae5d24", size = 250800, upload-time = "2026-01-23T11:04:44.165Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2f/90/fd509079dfcab01102c0fdd87f3a9506894bc70afcf9e9785ef6b2b3aff6/httplib2-0.31.2-py3-none-any.whl", hash = "sha256:dbf0c2fa3862acf3c55c078ea9c0bc4481d7dc5117cae71be9514912cf9f8349", size = 91099, upload-time = "2026-01-23T11:04:42.78Z" },
+]
+
+[[package]]
+name = "httpx"
+version = "0.28.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+    { name = "certifi" },
+    { name = "httpcore" },
+    { name = "idna" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b1/df/48c586a5fe32a0f01324ee087459e112ebb7224f646c0b5023f5e79e9956/httpx-0.28.1.tar.gz", hash = "sha256:75e98c5f16b0f35b567856f597f06ff2270a374470a5c2392242528e3e3e42fc", size = 141406, upload-time = "2024-12-06T15:37:23.222Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2a/39/e50c7c3a983047577ee07d2a9e53faf5a69493943ec3f6a384bdc792deb2/httpx-0.28.1-py3-none-any.whl", hash = "sha256:d909fcccc110f8c7faf814ca82a9a4d816bc5a6dbfea25d6591d6985b8ba59ad", size = 73517, upload-time = "2024-12-06T15:37:21.509Z" },
+]
+
+[[package]]
+name = "huggingface-hub"
+version = "0.36.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "filelock" },
+    { name = "fsspec" },
+    { name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "packaging" },
+    { name = "pyyaml" },
+    { name = "requests" },
+    { name = "tqdm" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7c/b7/8cb61d2eece5fb05a83271da168186721c450eb74e3c31f7ef3169fa475b/huggingface_hub-0.36.2.tar.gz", hash = "sha256:1934304d2fb224f8afa3b87007d58501acfda9215b334eed53072dd5e815ff7a", size = 649782, upload-time = "2026-02-06T09:24:13.098Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a8/af/48ac8483240de756d2438c380746e7130d1c6f75802ef22f3c6d49982787/huggingface_hub-0.36.2-py3-none-any.whl", hash = "sha256:48f0c8eac16145dfce371e9d2d7772854a4f591bcb56c9cf548accf531d54270", size = 566395, upload-time = "2026-02-06T09:24:11.133Z" },
+]
+
+[package.optional-dependencies]
+hf-xet = [
+    { name = "hf-xet" },
+]
+
+[[package]]
+name = "hydra-core"
+version = "1.3.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "antlr4-python3-runtime" },
+    { name = "omegaconf" },
+    { name = "packaging" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/6d/8e/07e42bc434a847154083b315779b0a81d567154504624e181caf2c71cd98/hydra-core-1.3.2.tar.gz", hash = "sha256:8a878ed67216997c3e9d88a8e72e7b4767e81af37afb4ea3334b269a4390a824", size = 3263494, upload-time = "2023-02-23T18:33:43.03Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c6/50/e0edd38dcd63fb26a8547f13d28f7a008bc4a3fd4eb4ff030673f22ad41a/hydra_core-1.3.2-py3-none-any.whl", hash = "sha256:fa0238a9e31df3373b35b0bfb672c34cc92718d21f81311d8996a16de1141d8b", size = 154547, upload-time = "2023-02-23T18:33:40.801Z" },
+]
+
+[[package]]
+name = "identify"
+version = "2.6.17"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/57/84/376a3b96e5a8d33a7aa2c5b3b31a4b3c364117184bf0b17418055f6ace66/identify-2.6.17.tar.gz", hash = "sha256:f816b0b596b204c9fdf076ded172322f2723cf958d02f9c3587504834c8ff04d", size = 99579, upload-time = "2026-03-01T20:04:12.702Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/40/66/71c1227dff78aaeb942fed29dd5651f2aec166cc7c9aeea3e8b26a539b7d/identify-2.6.17-py2.py3-none-any.whl", hash = "sha256:be5f8412d5ed4b20f2bd41a65f920990bdccaa6a4a18a08f1eefdcd0bdd885f0", size = 99382, upload-time = "2026-03-01T20:04:11.439Z" },
+]
+
+[[package]]
+name = "idna"
+version = "3.11"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6f/6d/0703ccc57f3a7233505399edb88de3cbd678da106337b9fcde432b65ed60/idna-3.11.tar.gz", hash = "sha256:795dafcc9c04ed0c1fb032c2aa73654d8e8c5023a7df64a53f39190ada629902", size = 194582, upload-time = "2025-10-12T14:55:20.501Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0e/61/66938bbb5fc52dbdf84594873d5b51fb1f7c7794e9c0f5bd885f30bc507b/idna-3.11-py3-none-any.whl", hash = "sha256:771a87f49d9defaf64091e6e6fe9c18d4833f140bd19464795bc32d966ca37ea", size = 71008, upload-time = "2025-10-12T14:55:18.883Z" },
+]
+
+[[package]]
+name = "imageio"
+version = "2.37.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "pillow" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a3/6f/606be632e37bf8d05b253e8626c2291d74c691ddc7bcdf7d6aaf33b32f6a/imageio-2.37.2.tar.gz", hash = "sha256:0212ef2727ac9caa5ca4b2c75ae89454312f440a756fcfc8ef1993e718f50f8a", size = 389600, upload-time = "2025-11-04T14:29:39.898Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fb/fe/301e0936b79bcab4cacc7548bf2853fc28dced0a578bab1f7ef53c9aa75b/imageio-2.37.2-py3-none-any.whl", hash = "sha256:ad9adfb20335d718c03de457358ed69f141021a333c40a53e57273d8a5bd0b9b", size = 317646, upload-time = "2025-11-04T14:29:37.948Z" },
+]
+
+[[package]]
+name = "imageio-ffmpeg"
+version = "0.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/44/bd/c3343c721f2a1b0c9fc71c1aebf1966a3b7f08c2eea8ed5437a2865611d6/imageio_ffmpeg-0.6.0.tar.gz", hash = "sha256:e2556bed8e005564a9f925bb7afa4002d82770d6b08825078b7697ab88ba1755", size = 25210, upload-time = "2025-01-16T21:34:32.747Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/da/58/87ef68ac83f4c7690961bce288fd8e382bc5f1513860fc7f90a9c1c1c6bf/imageio_ffmpeg-0.6.0-py3-none-macosx_10_9_intel.macosx_10_9_x86_64.whl", hash = "sha256:9d2baaf867088508d4a3458e61eeb30e945c4ad8016025545f66c4b5aaef0a61", size = 24932969, upload-time = "2025-01-16T21:34:20.464Z" },
+    { url = "https://files.pythonhosted.org/packages/40/5c/f3d8a657d362cc93b81aab8feda487317da5b5d31c0e1fdfd5e986e55d17/imageio_ffmpeg-0.6.0-py3-none-macosx_11_0_arm64.whl", hash = "sha256:b1ae3173414b5fc5f538a726c4e48ea97edc0d2cdc11f103afee655c463fa742", size = 21113891, upload-time = "2025-01-16T21:34:00.277Z" },
+    { url = "https://files.pythonhosted.org/packages/33/e7/1925bfbc563c39c1d2e82501d8372734a5c725e53ac3b31b4c2d081e895b/imageio_ffmpeg-0.6.0-py3-none-manylinux2014_aarch64.whl", hash = "sha256:1d47bebd83d2c5fc770720d211855f208af8a596c82d17730aa51e815cdee6dc", size = 25632706, upload-time = "2025-01-16T21:33:53.475Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/2d/43c8522a2038e9d0e7dbdf3a61195ecc31ca576fb1527a528c877e87d973/imageio_ffmpeg-0.6.0-py3-none-manylinux2014_x86_64.whl", hash = "sha256:c7e46fcec401dd990405049d2e2f475e2b397779df2519b544b8aab515195282", size = 29498237, upload-time = "2025-01-16T21:34:13.726Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/13/59da54728351883c3c1d9fca1710ab8eee82c7beba585df8f25ca925f08f/imageio_ffmpeg-0.6.0-py3-none-win32.whl", hash = "sha256:196faa79366b4a82f95c0f4053191d2013f4714a715780f0ad2a68ff37483cc2", size = 19652251, upload-time = "2025-01-16T21:34:06.812Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/c6/fa760e12a2483469e2bf5058c5faff664acf66cadb4df2ad6205b016a73d/imageio_ffmpeg-0.6.0-py3-none-win_amd64.whl", hash = "sha256:02fa47c83703c37df6bfe4896aab339013f62bf02c5ebf2dce6da56af04ffc0a", size = 31246824, upload-time = "2025-01-16T21:34:28.6Z" },
+]
+
+[[package]]
+name = "importlib-metadata"
+version = "8.7.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "zipp" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f3/49/3b30cad09e7771a4982d9975a8cbf64f00d4a1ececb53297f1d9a7be1b10/importlib_metadata-8.7.1.tar.gz", hash = "sha256:49fef1ae6440c182052f407c8d34a68f72efc36db9ca90dc0113398f2fdde8bb", size = 57107, upload-time = "2025-12-21T10:00:19.278Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/5e/f8e9a1d23b9c20a551a8a02ea3637b4642e22c2626e3a13a9a29cdea99eb/importlib_metadata-8.7.1-py3-none-any.whl", hash = "sha256:5a1f80bf1daa489495071efbb095d75a634cf28a8bc299581244063b53176151", size = 27865, upload-time = "2025-12-21T10:00:18.329Z" },
+]
+
+[[package]]
+name = "importlib-resources"
+version = "6.5.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cf/8c/f834fbf984f691b4f7ff60f50b514cc3de5cc08abfc3295564dd89c5e2e7/importlib_resources-6.5.2.tar.gz", hash = "sha256:185f87adef5bcc288449d98fb4fba07cea78bc036455dd44c5fc4a2fe78fed2c", size = 44693, upload-time = "2025-01-03T18:51:56.698Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a4/ed/1f1afb2e9e7f38a545d628f864d562a5ae64fe6f7a10e28ffb9b185b4e89/importlib_resources-6.5.2-py3-none-any.whl", hash = "sha256:789cfdc3ed28c78b67a06acb8126751ced69a3d5f79c095a98298cd8a760ccec", size = 37461, upload-time = "2025-01-03T18:51:54.306Z" },
+]
+
+[[package]]
+name = "iniconfig"
+version = "2.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/72/34/14ca021ce8e5dfedc35312d08ba8bf51fdd999c576889fc2c24cb97f4f10/iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730", size = 20503, upload-time = "2025-10-18T21:55:43.219Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/cb/b1/3846dd7f199d53cb17f49cba7e651e9ce294d8497c8c150530ed11865bb8/iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12", size = 7484, upload-time = "2025-10-18T21:55:41.639Z" },
+]
+
+[[package]]
+name = "ipdb"
+version = "0.13.13"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "decorator" },
+    { name = "ipython" },
+    { name = "tomli" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3d/1b/7e07e7b752017f7693a0f4d41c13e5ca29ce8cbcfdcc1fd6c4ad8c0a27a0/ipdb-0.13.13.tar.gz", hash = "sha256:e3ac6018ef05126d442af680aad863006ec19d02290561ac88b8b1c0b0cfc726", size = 17042, upload-time = "2023-03-09T15:40:57.487Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0c/4c/b075da0092003d9a55cf2ecc1cae9384a1ca4f650d51b00fc59875fe76f6/ipdb-0.13.13-py3-none-any.whl", hash = "sha256:45529994741c4ab6d2388bfa5d7b725c2cf7fe9deffabdb8a6113aa5ed449ed4", size = 12130, upload-time = "2023-03-09T15:40:55.021Z" },
+]
+
+[[package]]
+name = "ipykernel"
+version = "7.2.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "appnope", marker = "sys_platform == 'darwin' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "comm" },
+    { name = "debugpy" },
+    { name = "ipython" },
+    { name = "jupyter-client" },
+    { name = "jupyter-core" },
+    { name = "matplotlib-inline" },
+    { name = "nest-asyncio" },
+    { name = "packaging" },
+    { name = "psutil" },
+    { name = "pyzmq" },
+    { name = "tornado" },
+    { name = "traitlets" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ca/8d/b68b728e2d06b9e0051019640a40a9eb7a88fcd82c2e1b5ce70bef5ff044/ipykernel-7.2.0.tar.gz", hash = "sha256:18ed160b6dee2cbb16e5f3575858bc19d8f1fe6046a9a680c708494ce31d909e", size = 176046, upload-time = "2026-02-06T16:43:27.403Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/82/b9/e73d5d9f405cba7706c539aa8b311b49d4c2f3d698d9c12f815231169c71/ipykernel-7.2.0-py3-none-any.whl", hash = "sha256:3bbd4420d2b3cc105cbdf3756bfc04500b1e52f090a90716851f3916c62e1661", size = 118788, upload-time = "2026-02-06T16:43:25.149Z" },
+]
+
+[[package]]
+name = "ipython"
+version = "8.38.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "decorator" },
+    { name = "exceptiongroup" },
+    { name = "jedi" },
+    { name = "matplotlib-inline" },
+    { name = "pexpect", marker = "(sys_platform != 'emscripten' and sys_platform != 'win32') or (sys_platform == 'emscripten' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "prompt-toolkit" },
+    { name = "pygments" },
+    { name = "stack-data" },
+    { name = "traitlets" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e5/61/1810830e8b93c72dcd3c0f150c80a00c3deb229562d9423807ec92c3a539/ipython-8.38.0.tar.gz", hash = "sha256:9cfea8c903ce0867cc2f23199ed8545eb741f3a69420bfcf3743ad1cec856d39", size = 5513996, upload-time = "2026-01-05T10:59:06.901Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9f/df/db59624f4c71b39717c423409950ac3f2c8b2ce4b0aac843112c7fb3f721/ipython-8.38.0-py3-none-any.whl", hash = "sha256:750162629d800ac65bb3b543a14e7a74b0e88063eac9b92124d4b2aa3f6d8e86", size = 831813, upload-time = "2026-01-05T10:59:04.239Z" },
+]
+
+[[package]]
+name = "ipywidgets"
+version = "8.1.8"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "comm" },
+    { name = "ipython" },
+    { name = "jupyterlab-widgets" },
+    { name = "traitlets" },
+    { name = "widgetsnbextension" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/4c/ae/c5ce1edc1afe042eadb445e95b0671b03cee61895264357956e61c0d2ac0/ipywidgets-8.1.8.tar.gz", hash = "sha256:61f969306b95f85fba6b6986b7fe45d73124d1d9e3023a8068710d47a22ea668", size = 116739, upload-time = "2025-11-01T21:18:12.393Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/56/6d/0d9848617b9f753b87f214f1c682592f7ca42de085f564352f10f0843026/ipywidgets-8.1.8-py3-none-any.whl", hash = "sha256:ecaca67aed704a338f88f67b1181b58f821ab5dc89c1f0f5ef99db43c1c2921e", size = 139808, upload-time = "2025-11-01T21:18:10.956Z" },
+]
+
+[[package]]
+name = "jedi"
+version = "0.19.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "parso" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/72/3a/79a912fbd4d8dd6fbb02bf69afd3bb72cf0c729bb3063c6f4498603db17a/jedi-0.19.2.tar.gz", hash = "sha256:4770dc3de41bde3966b02eb84fbcf557fb33cce26ad23da12c742fb50ecb11f0", size = 1231287, upload-time = "2024-11-11T01:41:42.873Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c0/5a/9cac0c82afec3d09ccd97c8b6502d48f165f9124db81b4bcb90b4af974ee/jedi-0.19.2-py2.py3-none-any.whl", hash = "sha256:a8ef22bde8490f57fe5c7681a3c83cb58874daf72b4784de3cce5b6ef6edb5b9", size = 1572278, upload-time = "2024-11-11T01:41:40.175Z" },
+]
+
+[[package]]
+name = "jieba"
+version = "0.42.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/c6/cb/18eeb235f833b726522d7ebed54f2278ce28ba9438e3135ab0278d9792a2/jieba-0.42.1.tar.gz", hash = "sha256:055ca12f62674fafed09427f176506079bc135638a14e23e25be909131928db2", size = 19214172, upload-time = "2020-01-20T14:27:23.5Z" }
+
+[[package]]
+name = "jinja2"
+version = "3.1.6"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "markupsafe" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/df/bf/f7da0350254c0ed7c72f3e33cef02e048281fec7ecec5f032d4aac52226b/jinja2-3.1.6.tar.gz", hash = "sha256:0137fb05990d35f1275a587e9aee6d56da821fc83491a0fb838183be43f66d6d", size = 245115, upload-time = "2025-03-05T20:05:02.478Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/62/a1/3d680cbfd5f4b8f15abc1d571870c5fc3e594bb582bc3b64ea099db13e56/jinja2-3.1.6-py3-none-any.whl", hash = "sha256:85ece4451f492d0c13c5dd7c13a64681a86afae63a5f347908daf103ce6d2f67", size = 134899, upload-time = "2025-03-05T20:05:00.369Z" },
+]
+
+[[package]]
+name = "jiter"
+version = "0.13.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/0d/5e/4ec91646aee381d01cdb9974e30882c9cd3b8c5d1079d6b5ff4af522439a/jiter-0.13.0.tar.gz", hash = "sha256:f2839f9c2c7e2dffc1bc5929a510e14ce0a946be9365fd1219e7ef342dae14f4", size = 164847, upload-time = "2026-02-02T12:37:56.441Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d0/5a/41da76c5ea07bec1b0472b6b2fdb1b651074d504b19374d7e130e0cdfb25/jiter-0.13.0-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:2ffc63785fd6c7977defe49b9824ae6ce2b2e2b77ce539bdaf006c26da06342e", size = 311164, upload-time = "2026-02-02T12:35:17.688Z" },
+    { url = "https://files.pythonhosted.org/packages/40/cb/4a1bf994a3e869f0d39d10e11efb471b76d0ad70ecbfb591427a46c880c2/jiter-0.13.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4a638816427006c1e3f0013eb66d391d7a3acda99a7b0cf091eff4497ccea33a", size = 320296, upload-time = "2026-02-02T12:35:19.828Z" },
+    { url = "https://files.pythonhosted.org/packages/09/82/acd71ca9b50ecebadc3979c541cd717cce2fe2bc86236f4fa597565d8f1a/jiter-0.13.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:19928b5d1ce0ff8c1ee1b9bdef3b5bfc19e8304f1b904e436caf30bc15dc6cf5", size = 352742, upload-time = "2026-02-02T12:35:21.258Z" },
+    { url = "https://files.pythonhosted.org/packages/71/03/d1fc996f3aecfd42eb70922edecfb6dd26421c874503e241153ad41df94f/jiter-0.13.0-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:309549b778b949d731a2f0e1594a3f805716be704a73bf3ad9a807eed5eb5721", size = 363145, upload-time = "2026-02-02T12:35:24.653Z" },
+    { url = "https://files.pythonhosted.org/packages/f1/61/a30492366378cc7a93088858f8991acd7d959759fe6138c12a4644e58e81/jiter-0.13.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:bcdabaea26cb04e25df3103ce47f97466627999260290349a88c8136ecae0060", size = 487683, upload-time = "2026-02-02T12:35:26.162Z" },
+    { url = "https://files.pythonhosted.org/packages/20/4e/4223cffa9dbbbc96ed821c5aeb6bca510848c72c02086d1ed3f1da3d58a7/jiter-0.13.0-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a3a377af27b236abbf665a69b2bdd680e3b5a0bd2af825cd3b81245279a7606c", size = 373579, upload-time = "2026-02-02T12:35:27.582Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/c9/b0489a01329ab07a83812d9ebcffe7820a38163c6d9e7da644f926ff877c/jiter-0.13.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fe49d3ff6db74321f144dff9addd4a5874d3105ac5ba7c5b77fac099cfae31ae", size = 362904, upload-time = "2026-02-02T12:35:28.925Z" },
+    { url = "https://files.pythonhosted.org/packages/05/af/53e561352a44afcba9a9bc67ee1d320b05a370aed8df54eafe714c4e454d/jiter-0.13.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2113c17c9a67071b0f820733c0893ed1d467b5fcf4414068169e5c2cabddb1e2", size = 392380, upload-time = "2026-02-02T12:35:30.385Z" },
+    { url = "https://files.pythonhosted.org/packages/76/2a/dd805c3afb8ed5b326c5ae49e725d1b1255b9754b1b77dbecdc621b20773/jiter-0.13.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:ab1185ca5c8b9491b55ebf6c1e8866b8f68258612899693e24a92c5fdb9455d5", size = 517939, upload-time = "2026-02-02T12:35:31.865Z" },
+    { url = "https://files.pythonhosted.org/packages/20/2a/7b67d76f55b8fe14c937e7640389612f05f9a4145fc28ae128aaa5e62257/jiter-0.13.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:9621ca242547edc16400981ca3231e0c91c0c4c1ab8573a596cd9bb3575d5c2b", size = 551696, upload-time = "2026-02-02T12:35:33.306Z" },
+    { url = "https://files.pythonhosted.org/packages/85/9c/57cdd64dac8f4c6ab8f994fe0eb04dc9fd1db102856a4458fcf8a99dfa62/jiter-0.13.0-cp310-cp310-win32.whl", hash = "sha256:a7637d92b1c9d7a771e8c56f445c7f84396d48f2e756e5978840ecba2fac0894", size = 204592, upload-time = "2026-02-02T12:35:34.58Z" },
+    { url = "https://files.pythonhosted.org/packages/a7/38/f4f3ea5788b8a5bae7510a678cdc747eda0c45ffe534f9878ff37e7cf3b3/jiter-0.13.0-cp310-cp310-win_amd64.whl", hash = "sha256:c1b609e5cbd2f52bb74fb721515745b407df26d7b800458bd97cb3b972c29e7d", size = 206016, upload-time = "2026-02-02T12:35:36.435Z" },
+]
+
+[[package]]
+name = "jmespath"
+version = "0.10.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/3c/56/3f325b1eef9791759784aa5046a8f6a1aff8f7c898a2e34506771d3b99d8/jmespath-0.10.0.tar.gz", hash = "sha256:b85d0567b8666149a93172712e68920734333c0ce7e89b78b3e987f71e5ed4f9", size = 21607, upload-time = "2020-05-12T22:03:47.267Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/07/cb/5f001272b6faeb23c1c9e0acc04d48eaaf5c862c17709d20e3469c6e0139/jmespath-0.10.0-py2.py3-none-any.whl", hash = "sha256:cdf6525904cc597730141d61b36f2e4b8ecc257c420fa2f4549bac2c2d0cb72f", size = 24489, upload-time = "2020-05-12T22:03:45.643Z" },
+]
+
+[[package]]
+name = "joblib"
+version = "1.5.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/41/f2/d34e8b3a08a9cc79a50b2208a93dce981fe615b64d5a4d4abee421d898df/joblib-1.5.3.tar.gz", hash = "sha256:8561a3269e6801106863fd0d6d84bb737be9e7631e33aaed3fb9ce5953688da3", size = 331603, upload-time = "2025-12-15T08:41:46.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7b/91/984aca2ec129e2757d1e4e3c81c3fcda9d0f85b74670a094cc443d9ee949/joblib-1.5.3-py3-none-any.whl", hash = "sha256:5fc3c5039fc5ca8c0276333a188bbd59d6b7ab37fe6632daa76bc7f9ec18e713", size = 309071, upload-time = "2025-12-15T08:41:44.973Z" },
+]
+
+[[package]]
+name = "jupyter-client"
+version = "8.8.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "jupyter-core" },
+    { name = "python-dateutil" },
+    { name = "pyzmq" },
+    { name = "tornado" },
+    { name = "traitlets" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/05/e4/ba649102a3bc3fbca54e7239fb924fd434c766f855693d86de0b1f2bec81/jupyter_client-8.8.0.tar.gz", hash = "sha256:d556811419a4f2d96c869af34e854e3f059b7cc2d6d01a9cd9c85c267691be3e", size = 348020, upload-time = "2026-01-08T13:55:47.938Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2d/0b/ceb7694d864abc0a047649aec263878acb9f792e1fec3e676f22dc9015e3/jupyter_client-8.8.0-py3-none-any.whl", hash = "sha256:f93a5b99c5e23a507b773d3a1136bd6e16c67883ccdbd9a829b0bbdb98cd7d7a", size = 107371, upload-time = "2026-01-08T13:55:45.562Z" },
+]
+
+[[package]]
+name = "jupyter-core"
+version = "5.9.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "platformdirs" },
+    { name = "traitlets" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/02/49/9d1284d0dc65e2c757b74c6687b6d319b02f822ad039e5c512df9194d9dd/jupyter_core-5.9.1.tar.gz", hash = "sha256:4d09aaff303b9566c3ce657f580bd089ff5c91f5f89cf7d8846c3cdf465b5508", size = 89814, upload-time = "2025-10-16T19:19:18.444Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e7/e7/80988e32bf6f73919a113473a604f5a8f09094de312b9d52b79c2df7612b/jupyter_core-5.9.1-py3-none-any.whl", hash = "sha256:ebf87fdc6073d142e114c72c9e29a9d7ca03fad818c5d300ce2adc1fb0743407", size = 29032, upload-time = "2025-10-16T19:19:16.783Z" },
+]
+
+[[package]]
+name = "jupyterlab-widgets"
+version = "3.0.16"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/26/2d/ef58fed122b268c69c0aa099da20bc67657cdfb2e222688d5731bd5b971d/jupyterlab_widgets-3.0.16.tar.gz", hash = "sha256:423da05071d55cf27a9e602216d35a3a65a3e41cdf9c5d3b643b814ce38c19e0", size = 897423, upload-time = "2025-11-01T21:11:29.724Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ab/b5/36c712098e6191d1b4e349304ef73a8d06aed77e56ceaac8c0a306c7bda1/jupyterlab_widgets-3.0.16-py3-none-any.whl", hash = "sha256:45fa36d9c6422cf2559198e4db481aa243c7a32d9926b500781c830c80f7ecf8", size = 914926, upload-time = "2025-11-01T21:11:28.008Z" },
+]
+
+[[package]]
+name = "keras"
+version = "3.12.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "h5py" },
+    { name = "ml-dtypes" },
+    { name = "namex" },
+    { name = "numpy" },
+    { name = "optree" },
+    { name = "packaging" },
+    { name = "rich" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/85/1d/b64986120f5de68b921aa9bc2b69cb53ad20c1a6ebe5431a73def28525ef/keras-3.12.1.tar.gz", hash = "sha256:3cb760b3fec105db4d893dd717daafdd0e35457a8201502c1ba8bedfaf334a71", size = 1127003, upload-time = "2026-01-30T18:28:57.087Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/13/7a/40096b1e1c29ffa830497a80800775ada5ba93f15cd52e0f1a84b4f22cd4/keras-3.12.1-py3-none-any.whl", hash = "sha256:c340f8a25362398b20500c64e290f6ee280c6aeec51e1044eb8d759b32dc272a", size = 1475784, upload-time = "2026-01-30T18:28:54.383Z" },
+]
+
+[[package]]
+name = "kiwisolver"
+version = "1.4.9"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/5c/3c/85844f1b0feb11ee581ac23fe5fce65cd049a200c1446708cc1b7f922875/kiwisolver-1.4.9.tar.gz", hash = "sha256:c3b22c26c6fd6811b0ae8363b95ca8ce4ea3c202d3d0975b2914310ceb1bcc4d", size = 97564, upload-time = "2025-08-10T21:27:49.279Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c6/5d/8ce64e36d4e3aac5ca96996457dcf33e34e6051492399a3f1fec5657f30b/kiwisolver-1.4.9-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b4b4d74bda2b8ebf4da5bd42af11d02d04428b2c32846e4c2c93219df8a7987b", size = 124159, upload-time = "2025-08-10T21:25:35.472Z" },
+    { url = "https://files.pythonhosted.org/packages/96/1e/22f63ec454874378175a5f435d6ea1363dd33fb2af832c6643e4ccea0dc8/kiwisolver-1.4.9-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:fb3b8132019ea572f4611d770991000d7f58127560c4889729248eb5852a102f", size = 66578, upload-time = "2025-08-10T21:25:36.73Z" },
+    { url = "https://files.pythonhosted.org/packages/41/4c/1925dcfff47a02d465121967b95151c82d11027d5ec5242771e580e731bd/kiwisolver-1.4.9-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:84fd60810829c27ae375114cd379da1fa65e6918e1da405f356a775d49a62bcf", size = 65312, upload-time = "2025-08-10T21:25:37.658Z" },
+    { url = "https://files.pythonhosted.org/packages/d4/42/0f333164e6307a0687d1eb9ad256215aae2f4bd5d28f4653d6cd319a3ba3/kiwisolver-1.4.9-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:b78efa4c6e804ecdf727e580dbb9cba85624d2e1c6b5cb059c66290063bd99a9", size = 1628458, upload-time = "2025-08-10T21:25:39.067Z" },
+    { url = "https://files.pythonhosted.org/packages/86/b6/2dccb977d651943995a90bfe3495c2ab2ba5cd77093d9f2318a20c9a6f59/kiwisolver-1.4.9-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d4efec7bcf21671db6a3294ff301d2fc861c31faa3c8740d1a94689234d1b415", size = 1225640, upload-time = "2025-08-10T21:25:40.489Z" },
+    { url = "https://files.pythonhosted.org/packages/50/2b/362ebd3eec46c850ccf2bfe3e30f2fc4c008750011f38a850f088c56a1c6/kiwisolver-1.4.9-cp310-cp310-manylinux_2_24_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:90f47e70293fc3688b71271100a1a5453aa9944a81d27ff779c108372cf5567b", size = 1244074, upload-time = "2025-08-10T21:25:42.221Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/bb/f09a1e66dab8984773d13184a10a29fe67125337649d26bdef547024ed6b/kiwisolver-1.4.9-cp310-cp310-manylinux_2_24_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8fdca1def57a2e88ef339de1737a1449d6dbf5fab184c54a1fca01d541317154", size = 1293036, upload-time = "2025-08-10T21:25:43.801Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/01/11ecf892f201cafda0f68fa59212edaea93e96c37884b747c181303fccd1/kiwisolver-1.4.9-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:9cf554f21be770f5111a1690d42313e140355e687e05cf82cb23d0a721a64a48", size = 2175310, upload-time = "2025-08-10T21:25:45.045Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/5f/bfe11d5b934f500cc004314819ea92427e6e5462706a498c1d4fc052e08f/kiwisolver-1.4.9-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:fc1795ac5cd0510207482c3d1d3ed781143383b8cfd36f5c645f3897ce066220", size = 2270943, upload-time = "2025-08-10T21:25:46.393Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/de/259f786bf71f1e03e73d87e2db1a9a3bcab64d7b4fd780167123161630ad/kiwisolver-1.4.9-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:ccd09f20ccdbbd341b21a67ab50a119b64a403b09288c27481575105283c1586", size = 2440488, upload-time = "2025-08-10T21:25:48.074Z" },
+    { url = "https://files.pythonhosted.org/packages/1b/76/c989c278faf037c4d3421ec07a5c452cd3e09545d6dae7f87c15f54e4edf/kiwisolver-1.4.9-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:540c7c72324d864406a009d72f5d6856f49693db95d1fbb46cf86febef873634", size = 2246787, upload-time = "2025-08-10T21:25:49.442Z" },
+    { url = "https://files.pythonhosted.org/packages/a2/55/c2898d84ca440852e560ca9f2a0d28e6e931ac0849b896d77231929900e7/kiwisolver-1.4.9-cp310-cp310-win_amd64.whl", hash = "sha256:ede8c6d533bc6601a47ad4046080d36b8fc99f81e6f1c17b0ac3c2dc91ac7611", size = 73730, upload-time = "2025-08-10T21:25:51.102Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/09/486d6ac523dd33b80b368247f238125d027964cfacb45c654841e88fb2ae/kiwisolver-1.4.9-cp310-cp310-win_arm64.whl", hash = "sha256:7b4da0d01ac866a57dd61ac258c5607b4cd677f63abaec7b148354d2b2cdd536", size = 65036, upload-time = "2025-08-10T21:25:52.063Z" },
+    { url = "https://files.pythonhosted.org/packages/a2/63/fde392691690f55b38d5dd7b3710f5353bf7a8e52de93a22968801ab8978/kiwisolver-1.4.9-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:4d1d9e582ad4d63062d34077a9a1e9f3c34088a2ec5135b1f7190c07cf366527", size = 60183, upload-time = "2025-08-10T21:27:37.669Z" },
+    { url = "https://files.pythonhosted.org/packages/27/b1/6aad34edfdb7cced27f371866f211332bba215bfd918ad3322a58f480d8b/kiwisolver-1.4.9-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:deed0c7258ceb4c44ad5ec7d9918f9f14fd05b2be86378d86cf50e63d1e7b771", size = 58675, upload-time = "2025-08-10T21:27:39.031Z" },
+    { url = "https://files.pythonhosted.org/packages/9d/1a/23d855a702bb35a76faed5ae2ba3de57d323f48b1f6b17ee2176c4849463/kiwisolver-1.4.9-pp310-pypy310_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:0a590506f303f512dff6b7f75fd2fd18e16943efee932008fe7140e5fa91d80e", size = 80277, upload-time = "2025-08-10T21:27:40.129Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/5b/5239e3c2b8fb5afa1e8508f721bb77325f740ab6994d963e61b2b7abcc1e/kiwisolver-1.4.9-pp310-pypy310_pp73-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e09c2279a4d01f099f52d5c4b3d9e208e91edcbd1a175c9662a8b16e000fece9", size = 77994, upload-time = "2025-08-10T21:27:41.181Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/1c/5d4d468fb16f8410e596ed0eac02d2c68752aa7dc92997fe9d60a7147665/kiwisolver-1.4.9-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:c9e7cdf45d594ee04d5be1b24dd9d49f3d1590959b2271fb30b5ca2b262c00fb", size = 73744, upload-time = "2025-08-10T21:27:42.254Z" },
+]
+
+[[package]]
+name = "libclang"
+version = "18.1.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6e/5c/ca35e19a4f142adffa27e3d652196b7362fa612243e2b916845d801454fc/libclang-18.1.1.tar.gz", hash = "sha256:a1214966d08d73d971287fc3ead8dfaf82eb07fb197680d8b3859dbbbbf78250", size = 39612, upload-time = "2024-03-17T16:04:37.434Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/49/f5e3e7e1419872b69f6f5e82ba56e33955a74bd537d8a1f5f1eff2f3668a/libclang-18.1.1-1-py2.py3-none-macosx_11_0_arm64.whl", hash = "sha256:0b2e143f0fac830156feb56f9231ff8338c20aecfe72b4ffe96f19e5a1dbb69a", size = 25836045, upload-time = "2024-06-30T17:40:31.646Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/e5/fc61bbded91a8830ccce94c5294ecd6e88e496cc85f6704bf350c0634b70/libclang-18.1.1-py2.py3-none-macosx_10_9_x86_64.whl", hash = "sha256:6f14c3f194704e5d09769108f03185fce7acaf1d1ae4bbb2f30a72c2400cb7c5", size = 26502641, upload-time = "2024-03-18T15:52:26.722Z" },
+    { url = "https://files.pythonhosted.org/packages/db/ed/1df62b44db2583375f6a8a5e2ca5432bbdc3edb477942b9b7c848c720055/libclang-18.1.1-py2.py3-none-macosx_11_0_arm64.whl", hash = "sha256:83ce5045d101b669ac38e6da8e58765f12da2d3aafb3b9b98d88b286a60964d8", size = 26420207, upload-time = "2024-03-17T15:00:26.63Z" },
+    { url = "https://files.pythonhosted.org/packages/1d/fc/716c1e62e512ef1c160e7984a73a5fc7df45166f2ff3f254e71c58076f7c/libclang-18.1.1-py2.py3-none-manylinux2010_x86_64.whl", hash = "sha256:c533091d8a3bbf7460a00cb6c1a71da93bffe148f172c7d03b1c31fbf8aa2a0b", size = 24515943, upload-time = "2024-03-17T16:03:45.942Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/3d/f0ac1150280d8d20d059608cf2d5ff61b7c3b7f7bcf9c0f425ab92df769a/libclang-18.1.1-py2.py3-none-manylinux2014_aarch64.whl", hash = "sha256:54dda940a4a0491a9d1532bf071ea3ef26e6dbaf03b5000ed94dd7174e8f9592", size = 23784972, upload-time = "2024-03-17T16:12:47.677Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/2f/d920822c2b1ce9326a4c78c0c2b4aa3fde610c7ee9f631b600acb5376c26/libclang-18.1.1-py2.py3-none-manylinux2014_armv7l.whl", hash = "sha256:cf4a99b05376513717ab5d82a0db832c56ccea4fd61a69dbb7bccf2dfb207dbe", size = 20259606, upload-time = "2024-03-17T16:17:42.437Z" },
+    { url = "https://files.pythonhosted.org/packages/2d/c2/de1db8c6d413597076a4259cea409b83459b2db997c003578affdd32bf66/libclang-18.1.1-py2.py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:69f8eb8f65c279e765ffd28aaa7e9e364c776c17618af8bff22a8df58677ff4f", size = 24921494, upload-time = "2024-03-17T16:14:20.132Z" },
+    { url = "https://files.pythonhosted.org/packages/0b/2d/3f480b1e1d31eb3d6de5e3ef641954e5c67430d5ac93b7fa7e07589576c7/libclang-18.1.1-py2.py3-none-win_amd64.whl", hash = "sha256:4dd2d3b82fab35e2bf9ca717d7b63ac990a3519c7e312f19fa8e86dcc712f7fb", size = 26415083, upload-time = "2024-03-17T16:42:21.703Z" },
+    { url = "https://files.pythonhosted.org/packages/71/cf/e01dc4cc79779cd82d77888a88ae2fa424d93b445ad4f6c02bfc18335b70/libclang-18.1.1-py2.py3-none-win_arm64.whl", hash = "sha256:3f0e1f49f04d3cd198985fea0511576b0aee16f9ff0e0f0cad7f9c57ec3c20e8", size = 22361112, upload-time = "2024-03-17T16:42:59.565Z" },
+]
+
+[[package]]
+name = "llvmlite"
+version = "0.46.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/74/cd/08ae687ba099c7e3d21fe2ea536500563ef1943c5105bf6ab4ee3829f68e/llvmlite-0.46.0.tar.gz", hash = "sha256:227c9fd6d09dce2783c18b754b7cd9d9b3b3515210c46acc2d3c5badd9870ceb", size = 193456, upload-time = "2025-12-08T18:15:36.295Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3d/a4/3959e1c61c5ca9db7921e5fd115b344c29b9d57a5dadd87bef97963ca1a5/llvmlite-0.46.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4323177e936d61ae0f73e653e2e614284d97d14d5dd12579adc92b6c2b0597b0", size = 37232766, upload-time = "2025-12-08T18:14:34.765Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/a5/a4d916f1015106e1da876028606a8e87fd5d5c840f98c87bc2d5153b6a2f/llvmlite-0.46.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:0a2d461cb89537b7c20feb04c46c32e12d5ad4f0896c9dfc0f60336219ff248e", size = 56275176, upload-time = "2025-12-08T18:14:37.944Z" },
+    { url = "https://files.pythonhosted.org/packages/79/7f/a7f2028805dac8c1a6fae7bda4e739b7ebbcd45b29e15bf6d21556fcd3d5/llvmlite-0.46.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b1f6595a35b7b39c3518b85a28bf18f45e075264e4b2dce3f0c2a4f232b4a910", size = 55128629, upload-time = "2025-12-08T18:14:41.674Z" },
+    { url = "https://files.pythonhosted.org/packages/b2/bc/4689e1ba0c073c196b594471eb21be0aa51d9e64b911728aa13cd85ef0ae/llvmlite-0.46.0-cp310-cp310-win_amd64.whl", hash = "sha256:e7a34d4aa6f9a97ee006b504be6d2b8cb7f755b80ab2f344dda1ef992f828559", size = 38138651, upload-time = "2025-12-08T18:14:45.845Z" },
+]
+
+[[package]]
+name = "loguru"
+version = "0.7.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "win32-setctime", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3a/05/a1dae3dffd1116099471c643b8924f5aa6524411dc6c63fdae648c4f1aca/loguru-0.7.3.tar.gz", hash = "sha256:19480589e77d47b8d85b2c827ad95d49bf31b0dcde16593892eb51dd18706eb6", size = 63559, upload-time = "2024-12-06T11:20:56.608Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0c/29/0348de65b8cc732daa3e33e67806420b2ae89bdce2b04af740289c5c6c8c/loguru-0.7.3-py3-none-any.whl", hash = "sha256:31a33c10c8e1e10422bfd431aeb5d351c7cf7fa671e3c4df004162264b28220c", size = 61595, upload-time = "2024-12-06T11:20:54.538Z" },
+]
+
+[[package]]
+name = "markdown"
+version = "3.10.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2b/f4/69fa6ed85ae003c2378ffa8f6d2e3234662abd02c10d216c0ba96081a238/markdown-3.10.2.tar.gz", hash = "sha256:994d51325d25ad8aa7ce4ebaec003febcce822c3f8c911e3b17c52f7f589f950", size = 368805, upload-time = "2026-02-09T14:57:26.942Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/de/1f/77fa3081e4f66ca3576c896ae5d31c3002ac6607f9747d2e3aa49227e464/markdown-3.10.2-py3-none-any.whl", hash = "sha256:e91464b71ae3ee7afd3017d9f358ef0baf158fd9a298db92f1d4761133824c36", size = 108180, upload-time = "2026-02-09T14:57:25.787Z" },
+]
+
+[[package]]
+name = "markdown-it-py"
+version = "4.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "mdurl" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/5b/f5/4ec618ed16cc4f8fb3b701563655a69816155e79e24a17b651541804721d/markdown_it_py-4.0.0.tar.gz", hash = "sha256:cb0a2b4aa34f932c007117b194e945bd74e0ec24133ceb5bac59009cda1cb9f3", size = 73070, upload-time = "2025-08-11T12:57:52.854Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/94/54/e7d793b573f298e1c9013b8c4dade17d481164aa517d1d7148619c2cedbf/markdown_it_py-4.0.0-py3-none-any.whl", hash = "sha256:87327c59b172c5011896038353a81343b6754500a08cd7a4973bb48c6d578147", size = 87321, upload-time = "2025-08-11T12:57:51.923Z" },
+]
+
+[[package]]
+name = "markupsafe"
+version = "3.0.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/7e/99/7690b6d4034fffd95959cbe0c02de8deb3098cc577c67bb6a24fe5d7caa7/markupsafe-3.0.3.tar.gz", hash = "sha256:722695808f4b6457b320fdc131280796bdceb04ab50fe1795cd540799ebe1698", size = 80313, upload-time = "2025-09-27T18:37:40.426Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e8/4b/3541d44f3937ba468b75da9eebcae497dcf67adb65caa16760b0a6807ebb/markupsafe-3.0.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2f981d352f04553a7171b8e44369f2af4055f888dfb147d55e42d29e29e74559", size = 11631, upload-time = "2025-09-27T18:36:05.558Z" },
+    { url = "https://files.pythonhosted.org/packages/98/1b/fbd8eed11021cabd9226c37342fa6ca4e8a98d8188a8d9b66740494960e4/markupsafe-3.0.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:e1c1493fb6e50ab01d20a22826e57520f1284df32f2d8601fdd90b6304601419", size = 12057, upload-time = "2025-09-27T18:36:07.165Z" },
+    { url = "https://files.pythonhosted.org/packages/40/01/e560d658dc0bb8ab762670ece35281dec7b6c1b33f5fbc09ebb57a185519/markupsafe-3.0.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1ba88449deb3de88bd40044603fafffb7bc2b055d626a330323a9ed736661695", size = 22050, upload-time = "2025-09-27T18:36:08.005Z" },
+    { url = "https://files.pythonhosted.org/packages/af/cd/ce6e848bbf2c32314c9b237839119c5a564a59725b53157c856e90937b7a/markupsafe-3.0.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f42d0984e947b8adf7dd6dde396e720934d12c506ce84eea8476409563607591", size = 20681, upload-time = "2025-09-27T18:36:08.881Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/2a/b5c12c809f1c3045c4d580b035a743d12fcde53cf685dbc44660826308da/markupsafe-3.0.3-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c0c0b3ade1c0b13b936d7970b1d37a57acde9199dc2aecc4c336773e1d86049c", size = 20705, upload-time = "2025-09-27T18:36:10.131Z" },
+    { url = "https://files.pythonhosted.org/packages/cf/e3/9427a68c82728d0a88c50f890d0fc072a1484de2f3ac1ad0bfc1a7214fd5/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:0303439a41979d9e74d18ff5e2dd8c43ed6c6001fd40e5bf2e43f7bd9bbc523f", size = 21524, upload-time = "2025-09-27T18:36:11.324Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/36/23578f29e9e582a4d0278e009b38081dbe363c5e7165113fad546918a232/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:d2ee202e79d8ed691ceebae8e0486bd9a2cd4794cec4824e1c99b6f5009502f6", size = 20282, upload-time = "2025-09-27T18:36:12.573Z" },
+    { url = "https://files.pythonhosted.org/packages/56/21/dca11354e756ebd03e036bd8ad58d6d7168c80ce1fe5e75218e4945cbab7/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:177b5253b2834fe3678cb4a5f0059808258584c559193998be2601324fdeafb1", size = 20745, upload-time = "2025-09-27T18:36:13.504Z" },
+    { url = "https://files.pythonhosted.org/packages/87/99/faba9369a7ad6e4d10b6a5fbf71fa2a188fe4a593b15f0963b73859a1bbd/markupsafe-3.0.3-cp310-cp310-win32.whl", hash = "sha256:2a15a08b17dd94c53a1da0438822d70ebcd13f8c3a95abe3a9ef9f11a94830aa", size = 14571, upload-time = "2025-09-27T18:36:14.779Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/25/55dc3ab959917602c96985cb1253efaa4ff42f71194bddeb61eb7278b8be/markupsafe-3.0.3-cp310-cp310-win_amd64.whl", hash = "sha256:c4ffb7ebf07cfe8931028e3e4c85f0357459a3f9f9490886198848f4fa002ec8", size = 15056, upload-time = "2025-09-27T18:36:16.125Z" },
+    { url = "https://files.pythonhosted.org/packages/d0/9e/0a02226640c255d1da0b8d12e24ac2aa6734da68bff14c05dd53b94a0fc3/markupsafe-3.0.3-cp310-cp310-win_arm64.whl", hash = "sha256:e2103a929dfa2fcaf9bb4e7c091983a49c9ac3b19c9061b6d5427dd7d14d81a1", size = 13932, upload-time = "2025-09-27T18:36:17.311Z" },
+]
+
+[[package]]
+name = "matplotlib"
+version = "3.10.8"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "contourpy" },
+    { name = "cycler" },
+    { name = "fonttools" },
+    { name = "kiwisolver" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "pillow" },
+    { name = "pyparsing" },
+    { name = "python-dateutil" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/8a/76/d3c6e3a13fe484ebe7718d14e269c9569c4eb0020a968a327acb3b9a8fe6/matplotlib-3.10.8.tar.gz", hash = "sha256:2299372c19d56bcd35cf05a2738308758d32b9eaed2371898d8f5bd33f084aa3", size = 34806269, upload-time = "2025-12-10T22:56:51.155Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/58/be/a30bd917018ad220c400169fba298f2bb7003c8ccbc0c3e24ae2aacad1e8/matplotlib-3.10.8-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:00270d217d6b20d14b584c521f810d60c5c78406dc289859776550df837dcda7", size = 8239828, upload-time = "2025-12-10T22:55:02.313Z" },
+    { url = "https://files.pythonhosted.org/packages/58/27/ca01e043c4841078e82cf6e80a6993dfecd315c3d79f5f3153afbb8e1ec6/matplotlib-3.10.8-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:37b3c1cc42aa184b3f738cfa18c1c1d72fd496d85467a6cf7b807936d39aa656", size = 8128050, upload-time = "2025-12-10T22:55:04.997Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/aa/7ab67f2b729ae6a91bcf9dcac0affb95fb8c56f7fd2b2af894ae0b0cf6fa/matplotlib-3.10.8-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ee40c27c795bda6a5292e9cff9890189d32f7e3a0bf04e0e3c9430c4a00c37df", size = 8700452, upload-time = "2025-12-10T22:55:07.47Z" },
+    { url = "https://files.pythonhosted.org/packages/73/ae/2d5817b0acee3c49b7e7ccfbf5b273f284957cc8e270adf36375db353190/matplotlib-3.10.8-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a48f2b74020919552ea25d222d5cc6af9ca3f4eb43a93e14d068457f545c2a17", size = 9534928, upload-time = "2025-12-10T22:55:10.566Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/5b/8e66653e9f7c39cb2e5cab25fce4810daffa2bff02cbf5f3077cea9e942c/matplotlib-3.10.8-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:f254d118d14a7f99d616271d6c3c27922c092dac11112670b157798b89bf4933", size = 9586377, upload-time = "2025-12-10T22:55:12.362Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/e2/fd0bbadf837f81edb0d208ba8f8cb552874c3b16e27cb91a31977d90875d/matplotlib-3.10.8-cp310-cp310-win_amd64.whl", hash = "sha256:f9b587c9c7274c1613a30afabf65a272114cd6cdbe67b3406f818c79d7ab2e2a", size = 8128127, upload-time = "2025-12-10T22:55:14.436Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/43/31d59500bb950b0d188e149a2e552040528c13d6e3d6e84d0cccac593dcd/matplotlib-3.10.8-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:f97aeb209c3d2511443f8797e3e5a569aebb040d4f8bc79aa3ee78a8fb9e3dd8", size = 8237252, upload-time = "2025-12-10T22:56:39.529Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/2c/615c09984f3c5f907f51c886538ad785cf72e0e11a3225de2c0f9442aecc/matplotlib-3.10.8-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:fb061f596dad3a0f52b60dc6a5dec4a0c300dec41e058a7efe09256188d170b7", size = 8124693, upload-time = "2025-12-10T22:56:41.758Z" },
+    { url = "https://files.pythonhosted.org/packages/91/e1/2757277a1c56041e1fc104b51a0f7b9a4afc8eb737865d63cababe30bc61/matplotlib-3.10.8-pp310-pypy310_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:12d90df9183093fcd479f4172ac26b322b1248b15729cb57f42f71f24c7e37a3", size = 8702205, upload-time = "2025-12-10T22:56:43.415Z" },
+]
+
+[[package]]
+name = "matplotlib-inline"
+version = "0.2.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "traitlets" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c7/74/97e72a36efd4ae2bccb3463284300f8953f199b5ffbc04cbbb0ec78f74b1/matplotlib_inline-0.2.1.tar.gz", hash = "sha256:e1ee949c340d771fc39e241ea75683deb94762c8fa5f2927ec57c83c4dffa9fe", size = 8110, upload-time = "2025-10-23T09:00:22.126Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/af/33/ee4519fa02ed11a94aef9559552f3b17bb863f2ecfe1a35dc7f548cde231/matplotlib_inline-0.2.1-py3-none-any.whl", hash = "sha256:d56ce5156ba6085e00a9d54fead6ed29a9c47e215cd1bba2e976ef39f5710a76", size = 9516, upload-time = "2025-10-23T09:00:20.675Z" },
+]
+
+[[package]]
+name = "mdurl"
+version = "0.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d6/54/cfe61301667036ec958cb99bd3efefba235e65cdeb9c84d24a8293ba1d90/mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba", size = 8729, upload-time = "2022-08-14T12:40:10.846Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b3/38/89ba8ad64ae25be8de66a6d463314cf1eb366222074cfda9ee839c56a4b4/mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8", size = 9979, upload-time = "2022-08-14T12:40:09.779Z" },
+]
+
+[[package]]
+name = "metaworld"
+version = "3.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "gymnasium" },
+    { name = "imageio" },
+    { name = "mujoco" },
+    { name = "numpy" },
+    { name = "scipy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/8e/ea/ee112af5a58de09a1fdf33e8ba11cd38327c9b7a11049325a62055a66e37/metaworld-3.0.0.tar.gz", hash = "sha256:5f30950a5c71e4054c3d6d0293da13eddda9806a43f85cf576cf9632efd2cc87", size = 36452617, upload-time = "2025-06-14T01:44:35.873Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/29/e1/c758357e82c2deb2401a7214fbfbc6ddf09b5453371dee1b7a2da0aab274/metaworld-3.0.0-py3-none-any.whl", hash = "sha256:f1dd9f8a1bcceab34a5f3c20113724dd90d21984ae89df98c21c842eb3ece137", size = 36660341, upload-time = "2025-06-14T01:44:32.171Z" },
+]
+
+[[package]]
+name = "ml-dtypes"
+version = "0.5.4"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0e/4a/c27b42ed9b1c7d13d9ba8b6905dece787d6259152f2309338aed29b2447b/ml_dtypes-0.5.4.tar.gz", hash = "sha256:8ab06a50fb9bf9666dd0fe5dfb4676fa2b0ac0f31ecff72a6c3af8e22c063453", size = 692314, upload-time = "2025-11-17T22:32:31.031Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fe/3a/c5b855752a70267ff729c349e650263adb3c206c29d28cc8ea7ace30a1d5/ml_dtypes-0.5.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b95e97e470fe60ed493fd9ae3911d8da4ebac16bd21f87ffa2b7c588bf22ea2c", size = 679735, upload-time = "2025-11-17T22:31:31.367Z" },
+    { url = "https://files.pythonhosted.org/packages/41/79/7433f30ee04bd4faa303844048f55e1eb939131c8e5195a00a96a0939b64/ml_dtypes-0.5.4-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b4b801ebe0b477be666696bda493a9be8356f1f0057a57f1e35cd26928823e5a", size = 5051883, upload-time = "2025-11-17T22:31:33.658Z" },
+    { url = "https://files.pythonhosted.org/packages/10/b1/8938e8830b0ee2e167fc75a094dea766a1152bde46752cd9bfc57ee78a82/ml_dtypes-0.5.4-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:388d399a2152dd79a3f0456a952284a99ee5c93d3e2f8dfe25977511e0515270", size = 5030369, upload-time = "2025-11-17T22:31:35.595Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/a3/51886727bd16e2f47587997b802dd56398692ce8c6c03c2e5bb32ecafe26/ml_dtypes-0.5.4-cp310-cp310-win_amd64.whl", hash = "sha256:4ff7f3e7ca2972e7de850e7b8fcbb355304271e2933dd90814c1cb847414d6e2", size = 210738, upload-time = "2025-11-17T22:31:37.43Z" },
+]
+
+[[package]]
+name = "modelscope"
+version = "1.19.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "requests" },
+    { name = "tqdm" },
+    { name = "urllib3" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/58/5e/239075a17c5f4f89c54edef7846c1cd34e7f8a93eb70c93820f7b0519a64/modelscope-1.19.1.tar.gz", hash = "sha256:14443e5e08325adff9fafaa07196aff2dc36722bbc4a334dde6fce0e73ed0131", size = 4315959, upload-time = "2024-10-22T09:25:18.789Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/96/fbcdd3015d6a34f78ffbf870ced519bdb961c260e81e90888684f4e768e2/modelscope-1.19.1-py3-none-any.whl", hash = "sha256:f7df219fc93f2290eebcf00d65dedb4b606c9554a94458a6976284caa974c327", size = 5768360, upload-time = "2024-10-22T09:25:14.813Z" },
+]
+
+[package.optional-dependencies]
+datasets = [
+    { name = "addict" },
+    { name = "attrs" },
+    { name = "datasets" },
+    { name = "einops" },
+    { name = "oss2" },
+    { name = "pillow" },
+    { name = "python-dateutil" },
+    { name = "scipy" },
+    { name = "setuptools" },
+    { name = "simplejson" },
+    { name = "sortedcontainers" },
+    { name = "urllib3" },
+]
+
+[[package]]
+name = "moviepy"
+version = "2.2.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "decorator" },
+    { name = "imageio" },
+    { name = "imageio-ffmpeg" },
+    { name = "numpy" },
+    { name = "pillow" },
+    { name = "proglog" },
+    { name = "python-dotenv" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/de/61/15f9476e270f64c78a834e7459ca045d669f869cec24eed26807b8cd479d/moviepy-2.2.1.tar.gz", hash = "sha256:c80cb56815ece94e5e3e2d361aa40070eeb30a09d23a24c4e684d03e16deacb1", size = 58431438, upload-time = "2025-05-21T19:31:52.601Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9a/73/7d3b2010baa0b5eb1e4dfa9e4385e89b6716be76f2fa21a6c0fe34b68e5a/moviepy-2.2.1-py3-none-any.whl", hash = "sha256:6b56803fec2ac54b557404126ac1160e65448e03798fa282bd23e8fab3795060", size = 129871, upload-time = "2025-05-21T19:31:50.11Z" },
+]
+
+[[package]]
+name = "mpmath"
+version = "1.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e0/47/dd32fa426cc72114383ac549964eecb20ecfd886d1e5ccf5340b55b02f57/mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f", size = 508106, upload-time = "2023-03-07T16:47:11.061Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/43/e3/7d92a15f894aa0c9c4b49b8ee9ac9850d6e63b03c9c32c0367a13ae62209/mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c", size = 536198, upload-time = "2023-03-07T16:47:09.197Z" },
+]
+
+[[package]]
+name = "ms-swift"
+version = "3.3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "accelerate" },
+    { name = "addict" },
+    { name = "aiohttp" },
+    { name = "attrdict" },
+    { name = "binpacking" },
+    { name = "charset-normalizer" },
+    { name = "cpm-kernels" },
+    { name = "dacite" },
+    { name = "datasets" },
+    { name = "einops" },
+    { name = "fastapi" },
+    { name = "gradio" },
+    { name = "importlib-metadata" },
+    { name = "jieba" },
+    { name = "matplotlib" },
+    { name = "modelscope", extra = ["datasets"], marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "nltk" },
+    { name = "numpy" },
+    { name = "openai" },
+    { name = "oss2" },
+    { name = "pandas" },
+    { name = "peft", version = "0.15.2", source = { registry = "https://pypi.org/simple" } },
+    { name = "pillow" },
+    { name = "requests" },
+    { name = "rouge" },
+    { name = "safetensors" },
+    { name = "sentencepiece" },
+    { name = "tensorboard" },
+    { name = "tiktoken" },
+    { name = "tqdm" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" } },
+    { name = "transformers-stream-generator" },
+    { name = "trl", version = "0.16.1", source = { registry = "https://pypi.org/simple" } },
+    { name = "uvicorn" },
+    { name = "zstandard" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/5c/c6/8f8753a3151e2f1a9130f3eac2e436721e87fbe8973f45f809822e858b38/ms_swift-3.3.0.tar.gz", hash = "sha256:e33d2abb613b666fa329005ee533d8bd45b02eca0b30f9076532f4695303607e", size = 459331, upload-time = "2025-04-11T06:25:02.591Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/0d/3821b7e3a7a4ce9bd3c84a71215b524186bb32cc8276ee125fa344fa8474/ms_swift-3.3.0-py3-none-any.whl", hash = "sha256:77dfe925cf885f0710745660262f931dd8d812f8c36a812cd40d106ecc5593cc", size = 604706, upload-time = "2025-04-11T06:25:00.085Z" },
+]
+
+[[package]]
+name = "msgspec"
+version = "0.20.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ea/9c/bfbd12955a49180cbd234c5d29ec6f74fe641698f0cd9df154a854fc8a15/msgspec-0.20.0.tar.gz", hash = "sha256:692349e588fde322875f8d3025ac01689fead5901e7fb18d6870a44519d62a29", size = 317862, upload-time = "2025-11-24T03:56:28.934Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e3/5e/151883ba2047cca9db8ed2f86186b054ad200bc231352df15b0c1dd75b1f/msgspec-0.20.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:23a6ec2a3b5038c233b04740a545856a068bc5cb8db184ff493a58e08c994fbf", size = 195191, upload-time = "2025-11-24T03:55:08.549Z" },
+    { url = "https://files.pythonhosted.org/packages/50/88/a795647672f547c983eff0823b82aaa35db922c767e1b3693e2dcf96678d/msgspec-0.20.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:cde2c41ed3eaaef6146365cb0d69580078a19f974c6cb8165cc5dcd5734f573e", size = 188513, upload-time = "2025-11-24T03:55:10.008Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/91/eb0abb0e0de142066cebfe546dc9140c5972ea824aa6ff507ad0b6a126ac/msgspec-0.20.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5da0daa782f95d364f0d95962faed01e218732aa1aa6cad56b25a5d2092e75a4", size = 216370, upload-time = "2025-11-24T03:55:11.566Z" },
+    { url = "https://files.pythonhosted.org/packages/15/2a/48e41d9ef0a24b1c6e67cbd94a676799e0561bfbc163be1aaaff5ca853f5/msgspec-0.20.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9369d5266144bef91be2940a3821e03e51a93c9080fde3ef72728c3f0a3a8bb7", size = 222653, upload-time = "2025-11-24T03:55:13.159Z" },
+    { url = "https://files.pythonhosted.org/packages/90/c9/14b825df203d980f82a623450d5f39e7f7a09e6e256c52b498ea8f29d923/msgspec-0.20.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:90fb865b306ca92c03964a5f3d0cd9eb1adda14f7e5ac7943efd159719ea9f10", size = 222337, upload-time = "2025-11-24T03:55:14.777Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/d7/39a5c3ddd294f587d6fb8efccc8361b6aa5089974015054071e665c9d24b/msgspec-0.20.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:e8112cd48b67dfc0cfa49fc812b6ce7eb37499e1d95b9575061683f3428975d3", size = 225565, upload-time = "2025-11-24T03:55:16.4Z" },
+    { url = "https://files.pythonhosted.org/packages/98/bd/5db3c14d675ee12842afb9b70c94c64f2c873f31198c46cbfcd7dffafab0/msgspec-0.20.0-cp310-cp310-win_amd64.whl", hash = "sha256:666b966d503df5dc27287675f525a56b6e66a2b8e8ccd2877b0c01328f19ae6c", size = 188412, upload-time = "2025-11-24T03:55:17.747Z" },
+    { url = "https://files.pythonhosted.org/packages/76/c7/06cc218bc0c86f0c6c6f34f7eeea6cfb8b835070e8031e3b0ef00f6c7c69/msgspec-0.20.0-cp310-cp310-win_arm64.whl", hash = "sha256:099e3e85cd5b238f2669621be65f0728169b8c7cb7ab07f6137b02dc7feea781", size = 173951, upload-time = "2025-11-24T03:55:19.335Z" },
+]
+
+[[package]]
+name = "mujoco"
+version = "3.5.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "etils", extra = ["epath"] },
+    { name = "glfw" },
+    { name = "numpy" },
+    { name = "pyopengl" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/56/0d/005f0d49ad5878f0611a7c018550b8504d480a7a17ad7e6773ff47d8627a/mujoco-3.5.0.tar.gz", hash = "sha256:5c85a6fc7560ab5fa4534f35ff459e12dc3609681f307e457dbb49b6217f4d73", size = 912543, upload-time = "2026-02-13T01:02:51.554Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e8/20/9e0595e653543df3e4233bc3ad7e50b371b81dbe48d45ffbc867ed7c379d/mujoco-3.5.0-cp310-cp310-macosx_10_16_x86_64.whl", hash = "sha256:c4324161cb4f334dd984fbb4a4f7d7db9f914f40d06174b02dcf05463d8275e4", size = 7088320, upload-time = "2026-02-13T01:02:06.745Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/6b/fdac8ed97086e12ac930fb44e419eda1626e339010df73678cb1f22527d7/mujoco-3.5.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5f3803ff0dd7bc04d6c47d53a794343843bde06f0aeefeac28bb62b4cf2baab3", size = 7093261, upload-time = "2026-02-13T01:02:09.857Z" },
+    { url = "https://files.pythonhosted.org/packages/19/ce/abcd9cc6ee7802f97c729ae0ccd517c68f04882f5db755b178e199511dc2/mujoco-3.5.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e13560991c779a139b53151733a0a6f3420ef09459b32d90302c2661c1b20992", size = 6637850, upload-time = "2026-02-13T01:02:11.808Z" },
+    { url = "https://files.pythonhosted.org/packages/ca/d6/a5a7b615b257867b7c97db6b3ce07dec9351d5d9d5a5aca881cbb583d7a3/mujoco-3.5.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:01b12896ae906f157e18d8b1b7c24a8b72d2576fffa09869047150f186e92b33", size = 7079429, upload-time = "2026-02-13T01:02:13.738Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/91/d82dd3c16892e1b0e27a2f537eec8aad54d91d939cb3cd37db2e8c09ecc2/mujoco-3.5.0-cp310-cp310-win_amd64.whl", hash = "sha256:2328358d2f0031175897092560dd6d04b14bab1cc22caa145ce99b843c17daa2", size = 5624454, upload-time = "2026-02-13T01:02:15.714Z" },
+]
+
+[[package]]
+name = "multidict"
+version = "6.7.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/1a/c2/c2d94cbe6ac1753f3fc980da97b3d930efe1da3af3c9f5125354436c073d/multidict-6.7.1.tar.gz", hash = "sha256:ec6652a1bee61c53a3e5776b6049172c53b6aaba34f18c9ad04f82712bac623d", size = 102010, upload-time = "2026-01-26T02:46:45.979Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/84/0b/19348d4c98980c4851d2f943f8ebafdece2ae7ef737adcfa5994ce8e5f10/multidict-6.7.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:c93c3db7ea657dd4637d57e74ab73de31bccefe144d3d4ce370052035bc85fb5", size = 77176, upload-time = "2026-01-26T02:42:59.784Z" },
+    { url = "https://files.pythonhosted.org/packages/ef/04/9de3f8077852e3d438215c81e9b691244532d2e05b4270e89ce67b7d103c/multidict-6.7.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:974e72a2474600827abaeda71af0c53d9ebbc3c2eb7da37b37d7829ae31232d8", size = 44996, upload-time = "2026-01-26T02:43:01.674Z" },
+    { url = "https://files.pythonhosted.org/packages/31/5c/08c7f7fe311f32e83f7621cd3f99d805f45519cd06fafb247628b861da7d/multidict-6.7.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:cdea2e7b2456cfb6694fb113066fd0ec7ea4d67e3a35e1f4cbeea0b448bf5872", size = 44631, upload-time = "2026-01-26T02:43:03.169Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/7f/0e3b1390ae772f27501199996b94b52ceeb64fe6f9120a32c6c3f6b781be/multidict-6.7.1-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:17207077e29342fdc2c9a82e4b306f1127bf1ea91f8b71e02d4798a70bb99991", size = 242561, upload-time = "2026-01-26T02:43:04.733Z" },
+    { url = "https://files.pythonhosted.org/packages/dd/f4/8719f4f167586af317b69dd3e90f913416c91ca610cac79a45c53f590312/multidict-6.7.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d4f49cb5661344764e4c7c7973e92a47a59b8fc19b6523649ec9dc4960e58a03", size = 242223, upload-time = "2026-01-26T02:43:06.695Z" },
+    { url = "https://files.pythonhosted.org/packages/47/ab/7c36164cce64a6ad19c6d9a85377b7178ecf3b89f8fd589c73381a5eedfd/multidict-6.7.1-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:a9fc4caa29e2e6ae408d1c450ac8bf19892c5fca83ee634ecd88a53332c59981", size = 222322, upload-time = "2026-01-26T02:43:08.472Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/79/a25add6fb38035b5337bc5734f296d9afc99163403bbcf56d4170f97eb62/multidict-6.7.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c5f0c21549ab432b57dcc82130f388d84ad8179824cc3f223d5e7cfbfd4143f6", size = 254005, upload-time = "2026-01-26T02:43:10.127Z" },
+    { url = "https://files.pythonhosted.org/packages/4a/7b/64a87cf98e12f756fc8bd444b001232ffff2be37288f018ad0d3f0aae931/multidict-6.7.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:7dfb78d966b2c906ae1d28ccf6e6712a3cd04407ee5088cd276fe8cb42186190", size = 251173, upload-time = "2026-01-26T02:43:11.731Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/ac/b605473de2bb404e742f2cc3583d12aedb2352a70e49ae8fce455b50c5aa/multidict-6.7.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9b0d9b91d1aa44db9c1f1ecd0d9d2ae610b2f4f856448664e01a3b35899f3f92", size = 243273, upload-time = "2026-01-26T02:43:13.063Z" },
+    { url = "https://files.pythonhosted.org/packages/03/65/11492d6a0e259783720f3bc1d9ea55579a76f1407e31ed44045c99542004/multidict-6.7.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:dd96c01a9dcd4889dcfcf9eb5544ca0c77603f239e3ffab0524ec17aea9a93ee", size = 238956, upload-time = "2026-01-26T02:43:14.843Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/a7/7ee591302af64e7c196fb63fe856c788993c1372df765102bd0448e7e165/multidict-6.7.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:067343c68cd6612d375710f895337b3a98a033c94f14b9a99eff902f205424e2", size = 233477, upload-time = "2026-01-26T02:43:16.025Z" },
+    { url = "https://files.pythonhosted.org/packages/9c/99/c109962d58756c35fd9992fed7f2355303846ea2ff054bb5f5e9d6b888de/multidict-6.7.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5884a04f4ff56c6120f6ccf703bdeb8b5079d808ba604d4d53aec0d55dc33568", size = 243615, upload-time = "2026-01-26T02:43:17.84Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/5f/1973e7c771c86e93dcfe1c9cc55a5481b610f6614acfc28c0d326fe6bfad/multidict-6.7.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:8affcf1c98b82bc901702eb73b6947a1bfa170823c153fe8a47b5f5f02e48e40", size = 249930, upload-time = "2026-01-26T02:43:19.06Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/a5/f170fc2268c3243853580203378cd522446b2df632061e0a5409817854c7/multidict-6.7.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:0d17522c37d03e85c8098ec8431636309b2682cf12e58f4dbc76121fb50e4962", size = 243807, upload-time = "2026-01-26T02:43:20.286Z" },
+    { url = "https://files.pythonhosted.org/packages/de/01/73856fab6d125e5bc652c3986b90e8699a95e84b48d72f39ade6c0e74a8c/multidict-6.7.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:24c0cf81544ca5e17cfcb6e482e7a82cd475925242b308b890c9452a074d4505", size = 239103, upload-time = "2026-01-26T02:43:21.508Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/46/f1220bd9944d8aa40d8ccff100eeeee19b505b857b6f603d6078cb5315b0/multidict-6.7.1-cp310-cp310-win32.whl", hash = "sha256:d82dd730a95e6643802f4454b8fdecdf08667881a9c5670db85bc5a56693f122", size = 41416, upload-time = "2026-01-26T02:43:22.703Z" },
+    { url = "https://files.pythonhosted.org/packages/68/00/9b38e272a770303692fc406c36e1a4c740f401522d5787691eb38a8925a8/multidict-6.7.1-cp310-cp310-win_amd64.whl", hash = "sha256:cf37cbe5ced48d417ba045aca1b21bafca67489452debcde94778a576666a1df", size = 46022, upload-time = "2026-01-26T02:43:23.77Z" },
+    { url = "https://files.pythonhosted.org/packages/64/65/d8d42490c02ee07b6bbe00f7190d70bb4738b3cce7629aaf9f213ef730dd/multidict-6.7.1-cp310-cp310-win_arm64.whl", hash = "sha256:59bc83d3f66b41dac1e7460aac1d196edc70c9ba3094965c467715a70ecb46db", size = 43238, upload-time = "2026-01-26T02:43:24.882Z" },
+    { url = "https://files.pythonhosted.org/packages/81/08/7036c080d7117f28a4af526d794aab6a84463126db031b007717c1a6676e/multidict-6.7.1-py3-none-any.whl", hash = "sha256:55d97cc6dae627efa6a6e548885712d4864b81110ac76fa4e534c03819fa4a56", size = 12319, upload-time = "2026-01-26T02:46:44.004Z" },
+]
+
+[[package]]
+name = "multiprocess"
+version = "0.70.16"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "dill" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b5/ae/04f39c5d0d0def03247c2893d6f2b83c136bf3320a2154d7b8858f2ba72d/multiprocess-0.70.16.tar.gz", hash = "sha256:161af703d4652a0e1410be6abccecde4a7ddffd19341be0a7011b94aeb171ac1", size = 1772603, upload-time = "2024-01-28T18:52:34.85Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ef/76/6e712a2623d146d314f17598df5de7224c85c0060ef63fd95cc15a25b3fa/multiprocess-0.70.16-pp310-pypy310_pp73-macosx_10_13_x86_64.whl", hash = "sha256:476887be10e2f59ff183c006af746cb6f1fd0eadcfd4ef49e605cbe2659920ee", size = 134980, upload-time = "2024-01-28T18:52:15.731Z" },
+    { url = "https://files.pythonhosted.org/packages/0f/ab/1e6e8009e380e22254ff539ebe117861e5bdb3bff1fc977920972237c6c7/multiprocess-0.70.16-pp310-pypy310_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:d951bed82c8f73929ac82c61f01a7b5ce8f3e5ef40f5b52553b4f547ce2b08ec", size = 134982, upload-time = "2024-01-28T18:52:17.783Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/f7/7ec7fddc92e50714ea3745631f79bd9c96424cb2702632521028e57d3a36/multiprocess-0.70.16-py310-none-any.whl", hash = "sha256:c4a9944c67bd49f823687463660a2d6daae94c289adff97e0f9d696ba6371d02", size = 134824, upload-time = "2024-01-28T18:52:26.062Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/89/38df130f2c799090c978b366cfdf5b96d08de5b29a4a293df7f7429fa50b/multiprocess-0.70.16-py38-none-any.whl", hash = "sha256:a71d82033454891091a226dfc319d0cfa8019a4e888ef9ca910372a446de4435", size = 132628, upload-time = "2024-01-28T18:52:30.853Z" },
+    { url = "https://files.pythonhosted.org/packages/da/d9/f7f9379981e39b8c2511c9e0326d212accacb82f12fbfdc1aa2ce2a7b2b6/multiprocess-0.70.16-py39-none-any.whl", hash = "sha256:a0bafd3ae1b732eac64be2e72038231c1ba97724b60b09400d68f229fcc2fbf3", size = 133351, upload-time = "2024-01-28T18:52:31.981Z" },
+]
+
+[[package]]
+name = "mypy-extensions"
+version = "1.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a2/6e/371856a3fb9d31ca8dac321cda606860fa4548858c0cc45d9d1d4ca2628b/mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558", size = 6343, upload-time = "2025-04-22T14:54:24.164Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/79/7b/2c79738432f5c924bef5071f933bcc9efd0473bac3b4aa584a6f7c1c8df8/mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505", size = 4963, upload-time = "2025-04-22T14:54:22.983Z" },
+]
+
+[[package]]
+name = "namex"
+version = "0.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/0c/c0/ee95b28f029c73f8d49d8f52edaed02a1d4a9acb8b69355737fdb1faa191/namex-0.1.0.tar.gz", hash = "sha256:117f03ccd302cc48e3f5c58a296838f6b89c83455ab8683a1e85f2a430aa4306", size = 6649, upload-time = "2025-05-26T23:17:38.918Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b2/bc/465daf1de06409cdd4532082806770ee0d8d7df434da79c76564d0f69741/namex-0.1.0-py3-none-any.whl", hash = "sha256:e2012a474502f1e2251267062aae3114611f07df4224b6e06334c57b0f2ce87c", size = 5905, upload-time = "2025-05-26T23:17:37.695Z" },
+]
+
+[[package]]
+name = "narwhals"
+version = "2.17.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/75/59/81d0f4cad21484083466f278e6b392addd9f4205b48d45b5c8771670ebf8/narwhals-2.17.0.tar.gz", hash = "sha256:ebd5bc95bcfa2f8e89a8ac09e2765a63055162837208e67b42d6eeb6651d5e67", size = 620306, upload-time = "2026-02-23T09:44:34.142Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/27/20770bd6bf8fbe1e16f848ba21da9df061f38d2e6483952c29d2bb5d1d8b/narwhals-2.17.0-py3-none-any.whl", hash = "sha256:2ac5307b7c2b275a7d66eeda906b8605e3d7a760951e188dcfff86e8ebe083dd", size = 444897, upload-time = "2026-02-23T09:44:32.006Z" },
+]
+
+[[package]]
+name = "nest-asyncio"
+version = "1.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/83/f8/51569ac65d696c8ecbee95938f89d4abf00f47d58d48f6fbabfe8f0baefe/nest_asyncio-1.6.0.tar.gz", hash = "sha256:6f172d5449aca15afd6c646851f4e31e02c598d553a667e38cafa997cfec55fe", size = 7418, upload-time = "2024-01-21T14:25:19.227Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/c4/c2971a3ba4c6103a3d10c4b0f24f461ddc027f0f09763220cf35ca1401b3/nest_asyncio-1.6.0-py3-none-any.whl", hash = "sha256:87af6efd6b5e897c81050477ef65c62e2b2f35d51703cae01aff2905b1852e1c", size = 5195, upload-time = "2024-01-21T14:25:17.223Z" },
+]
+
+[[package]]
+name = "networkx"
+version = "3.4.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fd/1d/06475e1cd5264c0b870ea2cc6fdb3e37177c1e565c43f56ff17a10e3937f/networkx-3.4.2.tar.gz", hash = "sha256:307c3669428c5362aab27c8a1260aa8f47c4e91d3891f48be0141738d8d053e1", size = 2151368, upload-time = "2024-10-21T12:39:38.695Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b9/54/dd730b32ea14ea797530a4479b2ed46a6fb250f682a9cfb997e968bf0261/networkx-3.4.2-py3-none-any.whl", hash = "sha256:df5d4365b724cf81b8c6a7312509d0c22386097011ad1abe274afd5e9d3bbc5f", size = 1723263, upload-time = "2024-10-21T12:39:36.247Z" },
+]
+
+[[package]]
+name = "nltk"
+version = "3.9.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "joblib" },
+    { name = "regex" },
+    { name = "tqdm" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e1/8f/915e1c12df07c70ed779d18ab83d065718a926e70d3ea33eb0cd66ffb7c0/nltk-3.9.3.tar.gz", hash = "sha256:cb5945d6424a98d694c2b9a0264519fab4363711065a46aa0ae7a2195b92e71f", size = 2923673, upload-time = "2026-02-24T12:05:53.833Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c2/7e/9af5a710a1236e4772de8dfcc6af942a561327bb9f42b5b4a24d0cf100fd/nltk-3.9.3-py3-none-any.whl", hash = "sha256:60b3db6e9995b3dd976b1f0fa7dec22069b2677e759c28eb69b62ddd44870522", size = 1525385, upload-time = "2026-02-24T12:05:46.54Z" },
+]
+
+[[package]]
+name = "nodeenv"
+version = "1.10.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/24/bf/d1bda4f6168e0b2e9e5958945e01910052158313224ada5ce1fb2e1113b8/nodeenv-1.10.0.tar.gz", hash = "sha256:996c191ad80897d076bdfba80a41994c2b47c68e224c542b48feba42ba00f8bb", size = 55611, upload-time = "2025-12-20T14:08:54.006Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/88/b2/d0896bdcdc8d28a7fc5717c305f1a861c26e18c05047949fb371034d98bd/nodeenv-1.10.0-py2.py3-none-any.whl", hash = "sha256:5bb13e3eed2923615535339b3c620e76779af4cb4c6a90deccc9e36b274d3827", size = 23438, upload-time = "2025-12-20T14:08:52.782Z" },
+]
+
+[[package]]
+name = "num2words"
+version = "0.5.14"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "docopt" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f6/58/ad645bd38b4b648eb2fc2ba1b909398e54eb0cbb6a7dbd2b4953e38c9621/num2words-0.5.14.tar.gz", hash = "sha256:b066ec18e56b6616a3b38086b5747daafbaa8868b226a36127e0451c0cf379c6", size = 218213, upload-time = "2024-12-17T20:17:10.191Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d6/5b/545e9267a1cc080c8a1be2746113a063e34bcdd0f5173fd665a5c13cb234/num2words-0.5.14-py3-none-any.whl", hash = "sha256:1c8e5b00142fc2966fd8d685001e36c4a9911e070d1b120e1beb721fa1edb33d", size = 163525, upload-time = "2024-12-17T20:17:06.074Z" },
+]
+
+[[package]]
+name = "numba"
+version = "0.64.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "llvmlite" },
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/23/c9/a0fb41787d01d621046138da30f6c2100d80857bf34b3390dd68040f27a3/numba-0.64.0.tar.gz", hash = "sha256:95e7300af648baa3308127b1955b52ce6d11889d16e8cfe637b4f85d2fca52b1", size = 2765679, upload-time = "2026-02-18T18:41:20.974Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4c/5e/604fed821cd7e3426bb3bc99a7ed6ac0bcb489f4cd93052256437d082f95/numba-0.64.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:cc09b79440952e3098eeebea4bf6e8d2355fb7f12734fcd9fc5039f0dca90727", size = 2683250, upload-time = "2026-02-18T18:40:45.829Z" },
+    { url = "https://files.pythonhosted.org/packages/4f/9f/9275a723d050b5f1a9b1c7fb7dbfce324fef301a8e50c5f88338569db06c/numba-0.64.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1afe3a80b8c2f376b211fb7a49e536ef9eafc92436afc95a2f41ea5392f8cc65", size = 3742168, upload-time = "2026-02-18T18:40:48.066Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/d1/97ca7dddaa36b16f4c46319bdb6b4913ba15d0245317d0d8ccde7b2d7d92/numba-0.64.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:23804194b93b8cd416c6444b5fbc4956082a45fed2d25436ef49c594666e7f7e", size = 3449103, upload-time = "2026-02-18T18:40:49.905Z" },
+    { url = "https://files.pythonhosted.org/packages/52/0a/b9e137ad78415373e3353564500e8bf29dbce3c0d73633bb384d4e5d7537/numba-0.64.0-cp310-cp310-win_amd64.whl", hash = "sha256:e2a9fe998bb2cf848960b34db02c2c3b5e02cf82c07a26d9eef3494069740278", size = 2749950, upload-time = "2026-02-18T18:40:51.536Z" },
+]
+
+[[package]]
+name = "numpy"
+version = "1.26.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/65/6e/09db70a523a96d25e115e71cc56a6f9031e7b8cd166c1ac8438307c14058/numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010", size = 15786129, upload-time = "2024-02-06T00:26:44.495Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a7/94/ace0fdea5241a27d13543ee117cbc65868e82213fb31a8eb7fe9ff23f313/numpy-1.26.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9ff0f4f29c51e2803569d7a51c2304de5554655a60c5d776e35b4a41413830d0", size = 20631468, upload-time = "2024-02-05T23:48:01.194Z" },
+    { url = "https://files.pythonhosted.org/packages/20/f7/b24208eba89f9d1b58c1668bc6c8c4fd472b20c45573cb767f59d49fb0f6/numpy-1.26.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2e4ee3380d6de9c9ec04745830fd9e2eccb3e6cf790d39d7b98ffd19b0dd754a", size = 13966411, upload-time = "2024-02-05T23:48:29.038Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/a5/4beee6488160798683eed5bdb7eead455892c3b4e1f78d79d8d3f3b084ac/numpy-1.26.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d209d8969599b27ad20994c8e41936ee0964e6da07478d6c35016bc386b66ad4", size = 14219016, upload-time = "2024-02-05T23:48:54.098Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/d7/ecf66c1cd12dc28b4040b15ab4d17b773b87fa9d29ca16125de01adb36cd/numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ffa75af20b44f8dba823498024771d5ac50620e6915abac414251bd971b4529f", size = 18240889, upload-time = "2024-02-05T23:49:25.361Z" },
+    { url = "https://files.pythonhosted.org/packages/24/03/6f229fe3187546435c4f6f89f6d26c129d4f5bed40552899fcf1f0bf9e50/numpy-1.26.4-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:62b8e4b1e28009ef2846b4c7852046736bab361f7aeadeb6a5b89ebec3c7055a", size = 13876746, upload-time = "2024-02-05T23:49:51.983Z" },
+    { url = "https://files.pythonhosted.org/packages/39/fe/39ada9b094f01f5a35486577c848fe274e374bbf8d8f472e1423a0bbd26d/numpy-1.26.4-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a4abb4f9001ad2858e7ac189089c42178fcce737e4169dc61321660f1a96c7d2", size = 18078620, upload-time = "2024-02-05T23:50:22.515Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/ef/6ad11d51197aad206a9ad2286dc1aac6a378059e06e8cf22cd08ed4f20dc/numpy-1.26.4-cp310-cp310-win32.whl", hash = "sha256:bfe25acf8b437eb2a8b2d49d443800a5f18508cd811fea3181723922a8a82b07", size = 5972659, upload-time = "2024-02-05T23:50:35.834Z" },
+    { url = "https://files.pythonhosted.org/packages/19/77/538f202862b9183f54108557bfda67e17603fc560c384559e769321c9d92/numpy-1.26.4-cp310-cp310-win_amd64.whl", hash = "sha256:b97fe8060236edf3662adfc2c633f56a08ae30560c56310562cb4f95500022d5", size = 15808905, upload-time = "2024-02-05T23:51:03.701Z" },
+]
+
+[[package]]
+name = "nvidia-cublas-cu12"
+version = "12.8.4.1"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/29/99/db44d685f0e257ff0e213ade1964fc459b4a690a73293220e98feb3307cf/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b86f6dd8935884615a0683b663891d43781b819ac4f2ba2b0c9604676af346d0", size = 590537124, upload-time = "2025-03-07T01:43:53.556Z" },
+    { url = "https://files.pythonhosted.org/packages/dc/61/e24b560ab2e2eaeb3c839129175fb330dfcfc29e5203196e5541a4c44682/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142", size = 594346921, upload-time = "2025-03-07T01:44:31.254Z" },
+    { url = "https://files.pythonhosted.org/packages/70/61/7d7b3c70186fb651d0fbd35b01dbfc8e755f69fd58f817f3d0f642df20c3/nvidia_cublas_cu12-12.8.4.1-py3-none-win_amd64.whl", hash = "sha256:47e9b82132fa8d2b4944e708049229601448aaad7e6f296f630f2d1a32de35af", size = 567544208, upload-time = "2025-03-07T01:53:30.535Z" },
+]
+
+[[package]]
+name = "nvidia-cuda-cupti-cu12"
+version = "12.8.90"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d5/1f/b3bd73445e5cb342727fd24fe1f7b748f690b460acadc27ea22f904502c8/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:4412396548808ddfed3f17a467b104ba7751e6b58678a4b840675c56d21cf7ed", size = 9533318, upload-time = "2025-03-07T01:40:10.421Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/02/2adcaa145158bf1a8295d83591d22e4103dbfd821bcaf6f3f53151ca4ffa/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182", size = 10248621, upload-time = "2025-03-07T01:40:21.213Z" },
+    { url = "https://files.pythonhosted.org/packages/41/bc/83f5426095d93694ae39fe1311431b5d5a9bb82e48bf0dd8e19be2765942/nvidia_cuda_cupti_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:bb479dcdf7e6d4f8b0b01b115260399bf34154a1a2e9fe11c85c517d87efd98e", size = 7015759, upload-time = "2025-03-07T01:51:11.355Z" },
+]
+
+[[package]]
+name = "nvidia-cuda-nvrtc-cu12"
+version = "12.8.93"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/05/6b/32f747947df2da6994e999492ab306a903659555dddc0fbdeb9d71f75e52/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994", size = 88040029, upload-time = "2025-03-07T01:42:13.562Z" },
+    { url = "https://files.pythonhosted.org/packages/eb/d1/e50d0acaab360482034b84b6e27ee83c6738f7d32182b987f9c7a4e32962/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fc1fec1e1637854b4c0a65fb9a8346b51dd9ee69e61ebaccc82058441f15bce8", size = 43106076, upload-time = "2025-03-07T01:41:59.817Z" },
+    { url = "https://files.pythonhosted.org/packages/45/51/52a3d84baa2136cc8df15500ad731d74d3a1114d4c123e043cb608d4a32b/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:7a4b6b2904850fe78e0bd179c4b655c404d4bb799ef03ddc60804247099ae909", size = 73586838, upload-time = "2025-03-07T01:52:13.483Z" },
+]
+
+[[package]]
+name = "nvidia-cuda-runtime-cu12"
+version = "12.8.90"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7c/75/f865a3b236e4647605ea34cc450900854ba123834a5f1598e160b9530c3a/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:52bf7bbee900262ffefe5e9d5a2a69a30d97e2bc5bb6cc866688caa976966e3d", size = 965265, upload-time = "2025-03-07T01:39:43.533Z" },
+    { url = "https://files.pythonhosted.org/packages/0d/9b/a997b638fcd068ad6e4d53b8551a7d30fe8b404d6f1804abf1df69838932/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90", size = 954765, upload-time = "2025-03-07T01:40:01.615Z" },
+    { url = "https://files.pythonhosted.org/packages/30/a5/a515b7600ad361ea14bfa13fb4d6687abf500adc270f19e89849c0590492/nvidia_cuda_runtime_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:c0c6027f01505bfed6c3b21ec546f69c687689aad5f1a377554bc6ca4aa993a8", size = 944318, upload-time = "2025-03-07T01:51:01.794Z" },
+]
+
+[[package]]
+name = "nvidia-cudnn-cu12"
+version = "9.10.2.21"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/41/e79269ce215c857c935fd86bcfe91a451a584dfc27f1e068f568b9ad1ab7/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c9132cc3f8958447b4910a1720036d9eff5928cc3179b0a51fb6d167c6cc87d8", size = 705026878, upload-time = "2025-06-06T21:52:51.348Z" },
+    { url = "https://files.pythonhosted.org/packages/ba/51/e123d997aa098c61d029f76663dedbfb9bc8dcf8c60cbd6adbe42f76d049/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8", size = 706758467, upload-time = "2025-06-06T21:54:08.597Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/90/0bd6e586701b3a890fd38aa71c387dab4883d619d6e5ad912ccbd05bfd67/nvidia_cudnn_cu12-9.10.2.21-py3-none-win_amd64.whl", hash = "sha256:c6288de7d63e6cf62988f0923f96dc339cea362decb1bf5b3141883392a7d65e", size = 692992268, upload-time = "2025-06-06T21:55:18.114Z" },
+]
+
+[[package]]
+name = "nvidia-cufft-cu12"
+version = "11.3.3.83"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/60/bc/7771846d3a0272026c416fbb7e5f4c1f146d6d80704534d0b187dd6f4800/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:848ef7224d6305cdb2a4df928759dca7b1201874787083b6e7550dd6765ce69a", size = 193109211, upload-time = "2025-03-07T01:44:56.873Z" },
+    { url = "https://files.pythonhosted.org/packages/1f/13/ee4e00f30e676b66ae65b4f08cb5bcbb8392c03f54f2d5413ea99a5d1c80/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74", size = 193118695, upload-time = "2025-03-07T01:45:27.821Z" },
+    { url = "https://files.pythonhosted.org/packages/7d/ec/ce1629f1e478bb5ccd208986b5f9e0316a78538dd6ab1d0484f012f8e2a1/nvidia_cufft_cu12-11.3.3.83-py3-none-win_amd64.whl", hash = "sha256:7a64a98ef2a7c47f905aaf8931b69a3a43f27c55530c698bb2ed7c75c0b42cb7", size = 192216559, upload-time = "2025-03-07T01:53:57.106Z" },
+]
+
+[[package]]
+name = "nvidia-cufile-cu12"
+version = "1.13.1.3"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/bb/fe/1bcba1dfbfb8d01be8d93f07bfc502c93fa23afa6fd5ab3fc7c1df71038a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc", size = 1197834, upload-time = "2025-03-07T01:45:50.723Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/f5/5607710447a6fe9fd9b3283956fceeee8a06cda1d2f56ce31371f595db2a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:4beb6d4cce47c1a0f1013d72e02b0994730359e17801d395bdcbf20cfb3bb00a", size = 1120705, upload-time = "2025-03-07T01:45:41.434Z" },
+]
+
+[[package]]
+name = "nvidia-curand-cu12"
+version = "10.3.9.90"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/45/5e/92aa15eca622a388b80fbf8375d4760738df6285b1e92c43d37390a33a9a/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:dfab99248034673b779bc6decafdc3404a8a6f502462201f2f31f11354204acd", size = 63625754, upload-time = "2025-03-07T01:46:10.735Z" },
+    { url = "https://files.pythonhosted.org/packages/fb/aa/6584b56dc84ebe9cf93226a5cde4d99080c8e90ab40f0c27bda7a0f29aa1/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9", size = 63619976, upload-time = "2025-03-07T01:46:23.323Z" },
+    { url = "https://files.pythonhosted.org/packages/b9/75/70c05b2f3ed5be3bb30b7102b6eb78e100da4bbf6944fd6725c012831cab/nvidia_curand_cu12-10.3.9.90-py3-none-win_amd64.whl", hash = "sha256:f149a8ca457277da854f89cf282d6ef43176861926c7ac85b2a0fbd237c587ec", size = 62765309, upload-time = "2025-03-07T01:54:20.478Z" },
+]
+
+[[package]]
+name = "nvidia-cusolver-cu12"
+version = "11.7.3.90"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c8/32/f7cd6ce8a7690544d084ea21c26e910a97e077c9b7f07bf5de623ee19981/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:db9ed69dbef9715071232caa9b69c52ac7de3a95773c2db65bdba85916e4e5c0", size = 267229841, upload-time = "2025-03-07T01:46:54.356Z" },
+    { url = "https://files.pythonhosted.org/packages/85/48/9a13d2975803e8cf2777d5ed57b87a0b6ca2cc795f9a4f59796a910bfb80/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450", size = 267506905, upload-time = "2025-03-07T01:47:16.273Z" },
+    { url = "https://files.pythonhosted.org/packages/13/c0/76ca8551b8a84146ffa189fec81c26d04adba4bc0dbe09cd6e6fd9b7de04/nvidia_cusolver_cu12-11.7.3.90-py3-none-win_amd64.whl", hash = "sha256:4a550db115fcabc4d495eb7d39ac8b58d4ab5d8e63274d3754df1c0ad6a22d34", size = 256720438, upload-time = "2025-03-07T01:54:39.898Z" },
+]
+
+[[package]]
+name = "nvidia-cusparse-cu12"
+version = "12.5.8.93"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/bc/f7/cd777c4109681367721b00a106f491e0d0d15cfa1fd59672ce580ce42a97/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9b6c161cb130be1a07a27ea6923df8141f3c295852f4b260c65f18f3e0a091dc", size = 288117129, upload-time = "2025-03-07T01:47:40.407Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/f5/e1854cb2f2bcd4280c44736c93550cc300ff4b8c95ebe370d0aa7d2b473d/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b", size = 288216466, upload-time = "2025-03-07T01:48:13.779Z" },
+    { url = "https://files.pythonhosted.org/packages/62/07/f3b2ad63f8e3d257a599f422ae34eb565e70c41031aecefa3d18b62cabd1/nvidia_cusparse_cu12-12.5.8.93-py3-none-win_amd64.whl", hash = "sha256:9a33604331cb2cac199f2e7f5104dfbb8a5a898c367a53dfda9ff2acb6b6b4dd", size = 284937404, upload-time = "2025-03-07T01:55:07.742Z" },
+]
+
+[[package]]
+name = "nvidia-cusparselt-cu12"
+version = "0.7.1"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/73/b9/598f6ff36faaece4b3c50d26f50e38661499ff34346f00e057760b35cc9d/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:8878dce784d0fac90131b6817b607e803c36e629ba34dc5b433471382196b6a5", size = 283835557, upload-time = "2025-02-26T00:16:54.265Z" },
+    { url = "https://files.pythonhosted.org/packages/56/79/12978b96bd44274fe38b5dde5cfb660b1d114f70a65ef962bcbbed99b549/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623", size = 287193691, upload-time = "2025-02-26T00:15:44.104Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/d8/a6b0d0d0c2435e9310f3e2bb0d9c9dd4c33daef86aa5f30b3681defd37ea/nvidia_cusparselt_cu12-0.7.1-py3-none-win_amd64.whl", hash = "sha256:f67fbb5831940ec829c9117b7f33807db9f9678dc2a617fbe781cac17b4e1075", size = 271020911, upload-time = "2025-02-26T00:14:47.204Z" },
+]
+
+[[package]]
+name = "nvidia-nccl-cu12"
+version = "2.27.3"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/7b/8354b784cf73b0ba51e566b4baba3ddd44fe8288a3d39ef1e06cd5417226/nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9ddf1a245abc36c550870f26d537a9b6087fb2e2e3d6e0ef03374c6fd19d984f", size = 322397768, upload-time = "2025-06-03T21:57:30.234Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/5b/4e4fff7bad39adf89f735f2bc87248c81db71205b62bcc0d5ca5b606b3c3/nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adf27ccf4238253e0b826bce3ff5fa532d65fc42322c8bfdfaf28024c0fbe039", size = 322364134, upload-time = "2025-06-03T21:58:04.013Z" },
+]
+
+[[package]]
+name = "nvidia-nvjitlink-cu12"
+version = "12.8.93"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f6/74/86a07f1d0f42998ca31312f998bd3b9a7eff7f52378f4f270c8679c77fb9/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88", size = 39254836, upload-time = "2025-03-07T01:49:55.661Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/a2/8cee5da30d13430e87bf99bb33455d2724d0a4a9cb5d7926d80ccb96d008/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:adccd7161ace7261e01bb91e44e88da350895c270d23f744f0820c818b7229e7", size = 38386204, upload-time = "2025-03-07T01:49:43.612Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/d7/34f02dad2e30c31b10a51f6b04e025e5dd60e5f936af9045a9b858a05383/nvidia_nvjitlink_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:bd93fbeeee850917903583587f4fc3a4eafa022e34572251368238ab5e6bd67f", size = 268553710, upload-time = "2025-03-07T01:56:24.13Z" },
+]
+
+[[package]]
+name = "nvidia-nvtx-cu12"
+version = "12.8.90"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/10/c0/1b303feea90d296f6176f32a2a70b5ef230f9bdeb3a72bddb0dc922dc137/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d7ad891da111ebafbf7e015d34879f7112832fc239ff0d7d776b6cb685274615", size = 91161, upload-time = "2025-03-07T01:42:23.922Z" },
+    { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" },
+    { url = "https://files.pythonhosted.org/packages/9f/99/4c9c0c329bf9fc125008c3b54c7c94c0023518d06fc025ae36431375e1fe/nvidia_nvtx_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:619c8304aedc69f02ea82dd244541a83c3d9d40993381b3b590f1adaed3db41e", size = 56492, upload-time = "2025-03-07T01:52:24.69Z" },
+]
+
+[[package]]
+name = "omegaconf"
+version = "2.3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "antlr4-python3-runtime" },
+    { name = "pyyaml" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/09/48/6388f1bb9da707110532cb70ec4d2822858ddfb44f1cdf1233c20a80ea4b/omegaconf-2.3.0.tar.gz", hash = "sha256:d5d4b6d29955cc50ad50c46dc269bcd92c6e00f5f90d23ab5fee7bfca4ba4cc7", size = 3298120, upload-time = "2022-12-08T20:59:22.753Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e3/94/1843518e420fa3ed6919835845df698c7e27e183cb997394e4a670973a65/omegaconf-2.3.0-py3-none-any.whl", hash = "sha256:7b4df175cdb08ba400f45cae3bdcae7ba8365db4d165fc65fd04b050ab63b46b", size = 79500, upload-time = "2022-12-08T20:59:19.686Z" },
+]
+
+[[package]]
+name = "openai"
+version = "2.26.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+    { name = "distro" },
+    { name = "httpx" },
+    { name = "jiter" },
+    { name = "pydantic" },
+    { name = "sniffio" },
+    { name = "tqdm" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d7/91/2a06c4e9597c338cac1e5e5a8dd6f29e1836fc229c4c523529dca387fda8/openai-2.26.0.tar.gz", hash = "sha256:b41f37c140ae0034a6e92b0c509376d907f3a66109935fba2c1b471a7c05a8fb", size = 666702, upload-time = "2026-03-05T23:17:35.874Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c6/2e/3f73e8ca53718952222cacd0cf7eecc9db439d020f0c1fe7ae717e4e199a/openai-2.26.0-py3-none-any.whl", hash = "sha256:6151bf8f83802f036117f06cc8a57b3a4da60da9926826cc96747888b57f394f", size = 1136409, upload-time = "2026-03-05T23:17:34.072Z" },
+]
+
+[[package]]
+name = "opencv-contrib-python-headless"
+version = "4.11.0.86"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/53/cc/295e9a4e783ca71ba1b8fbd34e51bc603eba4611afcfc7de1b09b2d6ed8d/opencv-contrib-python-headless-4.11.0.86.tar.gz", hash = "sha256:839319098a73264c580c97cb1ca835f7fce3d30e4fa9fa6d4d0618fff551be0b", size = 150579288, upload-time = "2025-01-16T13:54:11.763Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1d/fd/501948c96381bc94f76dde8357c2d7cf2bb98e7734366856d7b32d5e6c31/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_arm64.whl", hash = "sha256:b34485c0164057726eee8cb80d5cd5dedaab5edde716451fb4107dcc60cf40f0", size = 46276903, upload-time = "2025-01-16T13:53:11.509Z" },
+    { url = "https://files.pythonhosted.org/packages/29/fa/ba711201197bd0f9c9cda21f739393b9e14c2394b0d8ba886a877071a576/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_x86_64.whl", hash = "sha256:ec88044ecb426549b05a8a6187e68051299848c4556ef8cef310621ca8f4316b", size = 66524225, upload-time = "2025-01-16T13:55:56.728Z" },
+    { url = "https://files.pythonhosted.org/packages/83/ec/b3fb322e8bac7b797f98676c34599827920b3972e4d664bbdf8de84d7fca/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a8dc2f4109904ffa55967bf9ceb1521ce46d66c333e2f6261dfa1f957a1dbde0", size = 35122073, upload-time = "2025-01-16T13:52:07.72Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/80/26c4ad9459498fc9213dea7254c8d6cb7717b279306b070588a2781559d4/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3e7a86bf02e8157a2d9fce26d44eaafd31113fc21fc8b4f44ff56c28ab32fdba", size = 56078660, upload-time = "2025-01-16T13:53:56.436Z" },
+    { url = "https://files.pythonhosted.org/packages/c0/38/2ce4259eca6ca356e3757b596d7d583b4ab0be4a482c9f4dacaa3eb688d1/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-win32.whl", hash = "sha256:d2c10564c01f6c308ded345a3b37359714e694361e593e515c148465eda09c2a", size = 35092082, upload-time = "2025-01-16T13:53:06.296Z" },
+    { url = "https://files.pythonhosted.org/packages/49/c1/c7600136283a2d4d3327968bdd895ba917a033d5a5498b6c7ffcd78c772c/opencv_contrib_python_headless-4.11.0.86-cp37-abi3-win_amd64.whl", hash = "sha256:2671a828e5c8ec9d237dd8506a9e0268487d37e07625725f1a6de5fa973ea7fa", size = 46095689, upload-time = "2025-01-16T13:53:12.934Z" },
+]
+
+[[package]]
+name = "opencv-python"
+version = "4.11.0.86"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/17/06/68c27a523103dad5837dc5b87e71285280c4f098c60e4fe8a8db6486ab09/opencv-python-4.11.0.86.tar.gz", hash = "sha256:03d60ccae62304860d232272e4a4fda93c39d595780cb40b161b310244b736a4", size = 95171956, upload-time = "2025-01-16T13:52:24.737Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/05/4d/53b30a2a3ac1f75f65a59eb29cf2ee7207ce64867db47036ad61743d5a23/opencv_python-4.11.0.86-cp37-abi3-macosx_13_0_arm64.whl", hash = "sha256:432f67c223f1dc2824f5e73cdfcd9db0efc8710647d4e813012195dc9122a52a", size = 37326322, upload-time = "2025-01-16T13:52:25.887Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/84/0a67490741867eacdfa37bc18df96e08a9d579583b419010d7f3da8ff503/opencv_python-4.11.0.86-cp37-abi3-macosx_13_0_x86_64.whl", hash = "sha256:9d05ef13d23fe97f575153558653e2d6e87103995d54e6a35db3f282fe1f9c66", size = 56723197, upload-time = "2025-01-16T13:55:21.222Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/bd/29c126788da65c1fb2b5fb621b7fed0ed5f9122aa22a0868c5e2c15c6d23/opencv_python-4.11.0.86-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1b92ae2c8852208817e6776ba1ea0d6b1e0a1b5431e971a2a0ddd2a8cc398202", size = 42230439, upload-time = "2025-01-16T13:51:35.822Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/8b/90eb44a40476fa0e71e05a0283947cfd74a5d36121a11d926ad6f3193cc4/opencv_python-4.11.0.86-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6b02611523803495003bd87362db3e1d2a0454a6a63025dc6658a9830570aa0d", size = 62986597, upload-time = "2025-01-16T13:52:08.836Z" },
+    { url = "https://files.pythonhosted.org/packages/fb/d7/1d5941a9dde095468b288d989ff6539dd69cd429dbf1b9e839013d21b6f0/opencv_python-4.11.0.86-cp37-abi3-win32.whl", hash = "sha256:810549cb2a4aedaa84ad9a1c92fbfdfc14090e2749cedf2c1589ad8359aa169b", size = 29384337, upload-time = "2025-01-16T13:52:13.549Z" },
+    { url = "https://files.pythonhosted.org/packages/a4/7d/f1c30a92854540bf789e9cd5dde7ef49bbe63f855b85a2e6b3db8135c591/opencv_python-4.11.0.86-cp37-abi3-win_amd64.whl", hash = "sha256:085ad9b77c18853ea66283e98affefe2de8cc4c1f43eda4c100cf9b2721142ec", size = 39488044, upload-time = "2025-01-16T13:52:21.928Z" },
+]
+
+[[package]]
+name = "opencv-python-headless"
+version = "4.11.0.86"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/36/2f/5b2b3ba52c864848885ba988f24b7f105052f68da9ab0e693cc7c25b0b30/opencv-python-headless-4.11.0.86.tar.gz", hash = "sha256:996eb282ca4b43ec6a3972414de0e2331f5d9cda2b41091a49739c19fb843798", size = 95177929, upload-time = "2025-01-16T13:53:40.22Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/dc/53/2c50afa0b1e05ecdb4603818e85f7d174e683d874ef63a6abe3ac92220c8/opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_arm64.whl", hash = "sha256:48128188ade4a7e517237c8e1e11a9cdf5c282761473383e77beb875bb1e61ca", size = 37326460, upload-time = "2025-01-16T13:52:57.015Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/43/68555327df94bb9b59a1fd645f63fafb0762515344d2046698762fc19d58/opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_x86_64.whl", hash = "sha256:a66c1b286a9de872c343ee7c3553b084244299714ebb50fbdcd76f07ebbe6c81", size = 56723330, upload-time = "2025-01-16T13:55:45.731Z" },
+    { url = "https://files.pythonhosted.org/packages/45/be/1438ce43ebe65317344a87e4b150865c5585f4c0db880a34cdae5ac46881/opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6efabcaa9df731f29e5ea9051776715b1bdd1845d7c9530065c7951d2a2899eb", size = 29487060, upload-time = "2025-01-16T13:51:59.625Z" },
+    { url = "https://files.pythonhosted.org/packages/dd/5c/c139a7876099916879609372bfa513b7f1257f7f1a908b0bdc1c2328241b/opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0e0a27c19dd1f40ddff94976cfe43066fbbe9dfbb2ec1907d66c19caef42a57b", size = 49969856, upload-time = "2025-01-16T13:53:29.654Z" },
+    { url = "https://files.pythonhosted.org/packages/95/dd/ed1191c9dc91abcc9f752b499b7928aacabf10567bb2c2535944d848af18/opencv_python_headless-4.11.0.86-cp37-abi3-win32.whl", hash = "sha256:f447d8acbb0b6f2808da71fddd29c1cdd448d2bc98f72d9bb78a7a898fc9621b", size = 29324425, upload-time = "2025-01-16T13:52:49.048Z" },
+    { url = "https://files.pythonhosted.org/packages/86/8a/69176a64335aed183529207ba8bc3d329c2999d852b4f3818027203f50e6/opencv_python_headless-4.11.0.86-cp37-abi3-win_amd64.whl", hash = "sha256:6c304df9caa7a6a5710b91709dd4786bf20a74d57672b3c31f7033cc638174ca", size = 39402386, upload-time = "2025-01-16T13:52:56.418Z" },
+]
+
+[[package]]
+name = "opt-einsum"
+version = "3.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/8c/b9/2ac072041e899a52f20cf9510850ff58295003aa75525e58343591b0cbfb/opt_einsum-3.4.0.tar.gz", hash = "sha256:96ca72f1b886d148241348783498194c577fa30a8faac108586b14f1ba4473ac", size = 63004, upload-time = "2024-09-26T14:33:24.483Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/23/cd/066e86230ae37ed0be70aae89aabf03ca8d9f39c8aea0dec8029455b5540/opt_einsum-3.4.0-py3-none-any.whl", hash = "sha256:69bb92469f86a1565195ece4ac0323943e83477171b91d24c35afe028a90d7cd", size = 71932, upload-time = "2024-09-26T14:33:23.039Z" },
+]
+
+[[package]]
+name = "optree"
+version = "0.19.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3d/63/7b078bc36d5a206c21b03565a818ede38ff0fbf014e92085ec467ef10adb/optree-0.19.0.tar.gz", hash = "sha256:bc1991a948590756409e76be4e29efd4a487a185056d35db6c67619c19ea27a1", size = 175199, upload-time = "2026-02-23T01:56:37.752Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/93/7f/265404c1d3f8f6ea7ac9926e5d9cffb34e9e4650fc053df3b6157748c7d5/optree-0.19.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ebf33aa7487b1b8f7a9c6ac6298f41614955b30124133de72b82ae1714e2451a", size = 392338, upload-time = "2026-02-23T01:54:38.226Z" },
+    { url = "https://files.pythonhosted.org/packages/27/10/372f936347c1812de938377102d410f900a1e46fa770921b46013b4c0493/optree-0.19.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c5606d57a1601d8f2a9dcfa6ea774b91fe4eb4877a961040e8b61ef11244fafa", size = 363922, upload-time = "2026-02-23T01:54:39.907Z" },
+    { url = "https://files.pythonhosted.org/packages/11/17/b967abd72f5c0bc6f94146ba861114ef93e87bb45a340532a0c22f5254bd/optree-0.19.0-cp310-cp310-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c1d2059b0005ff0d3ceff9d35b777f82014cde7e841f981b3518d419d76e814b", size = 383890, upload-time = "2026-02-23T01:54:40.948Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/a9/bd693caf9f249a9616ad31beca80f25d7ba43caadb2cffe10d2048b1e68e/optree-0.19.0-cp310-cp310-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:66ac5ffb43e7db4d2035f6c7dff459601e047929dbee503cc83515ff07def236", size = 439934, upload-time = "2026-02-23T01:54:42.079Z" },
+    { url = "https://files.pythonhosted.org/packages/05/d4/ed9d13ff27bce911ed51dca99dd75ec774e55d6634b363bd08a2b7be6e18/optree-0.19.0-cp310-cp310-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:1058a0ece1dd8e2f1ed19e1b5f2a4195b08b248421e9a1ca643ce678cf78fdaf", size = 441453, upload-time = "2026-02-23T01:54:43.125Z" },
+    { url = "https://files.pythonhosted.org/packages/f0/df/2c238f389b02843f4088d40f0af5bbf6a888e77067841f49588da6ea7c4a/optree-0.19.0-cp310-cp310-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:36b81df071a4e6e9c4dcccf85405867bc12796ac270b665d7a1474e576f15b01", size = 436456, upload-time = "2026-02-23T01:54:44.143Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/08/eb95702a6ce8aba8cc5e58a8e47e4cd61c1c3639ac26914e8e84019edb7c/optree-0.19.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0230ed9ba60fbe97ee23c000850c54308009ba51a129a1060137f329f460e42e", size = 419682, upload-time = "2026-02-23T01:54:45.622Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/a3/e49cacd8a3b1ca32c98aff8c54bd37db99b8ffecd4967652dc84b4256dd5/optree-0.19.0-cp310-cp310-manylinux_2_39_riscv64.whl", hash = "sha256:1b0c90272aa51e5043e17455f0467eb3426b09d130187f901c9c0bab62b2490b", size = 384818, upload-time = "2026-02-23T01:54:46.958Z" },
+    { url = "https://files.pythonhosted.org/packages/46/0e/73ab9446b5e40e408ee3526cc25c7a7ff4f75a251ed815711e7aa184ee48/optree-0.19.0-cp310-cp310-win32.whl", hash = "sha256:4fad3d69591eedaf75a8acb7505db844b1a06c998d41a3a3e6dd4d3b764b1b55", size = 297879, upload-time = "2026-02-23T01:54:48.038Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/c7/54a69ca41ce4ed2f10cf2e980386a629c740572ae58db79a9babe60dc28c/optree-0.19.0-cp310-cp310-win_amd64.whl", hash = "sha256:cf381c451ae2dd7a88552f208890e5e36399510fe53718cc15e8ef10c7a95732", size = 321784, upload-time = "2026-02-23T01:54:49.282Z" },
+]
+
+[[package]]
+name = "orjson"
+version = "3.11.7"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/45/b268004f745ede84e5798b48ee12b05129d19235d0e15267aa57dcdb400b/orjson-3.11.7.tar.gz", hash = "sha256:9b1a67243945819ce55d24a30b59d6a168e86220452d2c96f4d1f093e71c0c49", size = 6144992, upload-time = "2026-02-02T15:38:49.29Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/de/1a/a373746fa6d0e116dd9e54371a7b54622c44d12296d5d0f3ad5e3ff33490/orjson-3.11.7-cp310-cp310-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:a02c833f38f36546ba65a452127633afce4cf0dd7296b753d3bb54e55e5c0174", size = 229140, upload-time = "2026-02-02T15:37:06.082Z" },
+    { url = "https://files.pythonhosted.org/packages/52/a2/fa129e749d500f9b183e8a3446a193818a25f60261e9ce143ad61e975208/orjson-3.11.7-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b63c6e6738d7c3470ad01601e23376aa511e50e1f3931395b9f9c722406d1a67", size = 128670, upload-time = "2026-02-02T15:37:08.002Z" },
+    { url = "https://files.pythonhosted.org/packages/08/93/1e82011cd1e0bd051ef9d35bed1aa7fb4ea1f0a055dc2c841b46b43a9ebd/orjson-3.11.7-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:043d3006b7d32c7e233b8cfb1f01c651013ea079e08dcef7189a29abd8befe11", size = 123832, upload-time = "2026-02-02T15:37:09.191Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/d8/a26b431ef962c7d55736674dddade876822f3e33223c1f47a36879350d04/orjson-3.11.7-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:57036b27ac8a25d81112eb0cc9835cd4833c5b16e1467816adc0015f59e870dc", size = 129171, upload-time = "2026-02-02T15:37:11.112Z" },
+    { url = "https://files.pythonhosted.org/packages/a7/19/f47819b84a580f490da260c3ee9ade214cf4cf78ac9ce8c1c758f80fdfc9/orjson-3.11.7-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:733ae23ada68b804b222c44affed76b39e30806d38660bf1eb200520d259cc16", size = 141967, upload-time = "2026-02-02T15:37:12.282Z" },
+    { url = "https://files.pythonhosted.org/packages/5b/cd/37ece39a0777ba077fdcdbe4cccae3be8ed00290c14bf8afdc548befc260/orjson-3.11.7-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5fdfad2093bdd08245f2e204d977facd5f871c88c4a71230d5bcbd0e43bf6222", size = 130991, upload-time = "2026-02-02T15:37:13.465Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/ed/f2b5d66aa9b6b5c02ff5f120efc7b38c7c4962b21e6be0f00fd99a5c348e/orjson-3.11.7-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cededd6738e1c153530793998e31c05086582b08315db48ab66649768f326baa", size = 133674, upload-time = "2026-02-02T15:37:14.694Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/6e/baa83e68d1aa09fa8c3e5b2c087d01d0a0bd45256de719ed7bc22c07052d/orjson-3.11.7-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:14f440c7268c8f8633d1b3d443a434bd70cb15686117ea6beff8fdc8f5917a1e", size = 138722, upload-time = "2026-02-02T15:37:16.501Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/47/7f8ef4963b772cd56999b535e553f7eb5cd27e9dd6c049baee6f18bfa05d/orjson-3.11.7-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:3a2479753bbb95b0ebcf7969f562cdb9668e6d12416a35b0dda79febf89cdea2", size = 409056, upload-time = "2026-02-02T15:37:17.895Z" },
+    { url = "https://files.pythonhosted.org/packages/38/eb/2df104dd2244b3618f25325a656f85cc3277f74bbd91224752410a78f3c7/orjson-3.11.7-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:71924496986275a737f38e3f22b4e0878882b3f7a310d2ff4dc96e812789120c", size = 144196, upload-time = "2026-02-02T15:37:19.349Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/2a/ee41de0aa3a6686598661eae2b4ebdff1340c65bfb17fcff8b87138aab21/orjson-3.11.7-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b4a9eefdc70bf8bf9857f0290f973dec534ac84c35cd6a7f4083be43e7170a8f", size = 134979, upload-time = "2026-02-02T15:37:20.906Z" },
+    { url = "https://files.pythonhosted.org/packages/4c/fa/92fc5d3d402b87a8b28277a9ed35386218a6a5287c7fe5ee9b9f02c53fb2/orjson-3.11.7-cp310-cp310-win32.whl", hash = "sha256:ae9e0b37a834cef7ce8f99de6498f8fad4a2c0bf6bfc3d02abd8ed56aa15b2de", size = 127968, upload-time = "2026-02-02T15:37:23.178Z" },
+    { url = "https://files.pythonhosted.org/packages/07/29/a576bf36d73d60df06904d3844a9df08e25d59eba64363aaf8ec2f9bff41/orjson-3.11.7-cp310-cp310-win_amd64.whl", hash = "sha256:d772afdb22555f0c58cfc741bdae44180122b3616faa1ecadb595cd526e4c993", size = 125128, upload-time = "2026-02-02T15:37:24.329Z" },
+]
+
+[[package]]
+name = "oss2"
+version = "2.19.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aliyun-python-sdk-core" },
+    { name = "aliyun-python-sdk-kms" },
+    { name = "crcmod" },
+    { name = "pycryptodome" },
+    { name = "requests" },
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/df/b5/f2cb1950dda46ac2284d6c950489fdacd0e743c2d79a347924d3cc44b86f/oss2-2.19.1.tar.gz", hash = "sha256:a8ab9ee7eb99e88a7e1382edc6ea641d219d585a7e074e3776e9dec9473e59c1", size = 298845, upload-time = "2024-10-25T11:37:46.638Z" }
+
+[[package]]
+name = "packaging"
+version = "26.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/65/ee/299d360cdc32edc7d2cf530f3accf79c4fca01e96ffc950d8a52213bd8e4/packaging-26.0.tar.gz", hash = "sha256:00243ae351a257117b6a241061796684b084ed1c516a08c48a3f7e147a9d80b4", size = 143416, upload-time = "2026-01-21T20:50:39.064Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b7/b9/c538f279a4e237a006a2c98387d081e9eb060d203d8ed34467cc0f0b9b53/packaging-26.0-py3-none-any.whl", hash = "sha256:b36f1fef9334a5588b4166f8bcd26a14e521f2b55e6b9de3aaa80d3ff7a37529", size = 74366, upload-time = "2026-01-21T20:50:37.788Z" },
+]
+
+[[package]]
+name = "pandas"
+version = "2.3.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "python-dateutil" },
+    { name = "pytz" },
+    { name = "tzdata" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/33/01/d40b85317f86cf08d853a4f495195c73815fdf205eef3993821720274518/pandas-2.3.3.tar.gz", hash = "sha256:e05e1af93b977f7eafa636d043f9f94c7ee3ac81af99c13508215942e64c993b", size = 4495223, upload-time = "2025-09-29T23:34:51.853Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3d/f7/f425a00df4fcc22b292c6895c6831c0c8ae1d9fac1e024d16f98a9ce8749/pandas-2.3.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:376c6446ae31770764215a6c937f72d917f214b43560603cd60da6408f183b6c", size = 11555763, upload-time = "2025-09-29T23:16:53.287Z" },
+    { url = "https://files.pythonhosted.org/packages/13/4f/66d99628ff8ce7857aca52fed8f0066ce209f96be2fede6cef9f84e8d04f/pandas-2.3.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:e19d192383eab2f4ceb30b412b22ea30690c9e618f78870357ae1d682912015a", size = 10801217, upload-time = "2025-09-29T23:17:04.522Z" },
+    { url = "https://files.pythonhosted.org/packages/1d/03/3fc4a529a7710f890a239cc496fc6d50ad4a0995657dccc1d64695adb9f4/pandas-2.3.3-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5caf26f64126b6c7aec964f74266f435afef1c1b13da3b0636c7518a1fa3e2b1", size = 12148791, upload-time = "2025-09-29T23:17:18.444Z" },
+    { url = "https://files.pythonhosted.org/packages/40/a8/4dac1f8f8235e5d25b9955d02ff6f29396191d4e665d71122c3722ca83c5/pandas-2.3.3-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:dd7478f1463441ae4ca7308a70e90b33470fa593429f9d4c578dd00d1fa78838", size = 12769373, upload-time = "2025-09-29T23:17:35.846Z" },
+    { url = "https://files.pythonhosted.org/packages/df/91/82cc5169b6b25440a7fc0ef3a694582418d875c8e3ebf796a6d6470aa578/pandas-2.3.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:4793891684806ae50d1288c9bae9330293ab4e083ccd1c5e383c34549c6e4250", size = 13200444, upload-time = "2025-09-29T23:17:49.341Z" },
+    { url = "https://files.pythonhosted.org/packages/10/ae/89b3283800ab58f7af2952704078555fa60c807fff764395bb57ea0b0dbd/pandas-2.3.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:28083c648d9a99a5dd035ec125d42439c6c1c525098c58af0fc38dd1a7a1b3d4", size = 13858459, upload-time = "2025-09-29T23:18:03.722Z" },
+    { url = "https://files.pythonhosted.org/packages/85/72/530900610650f54a35a19476eca5104f38555afccda1aa11a92ee14cb21d/pandas-2.3.3-cp310-cp310-win_amd64.whl", hash = "sha256:503cf027cf9940d2ceaa1a93cfb5f8c8c7e6e90720a2850378f0b3f3b1e06826", size = 11346086, upload-time = "2025-09-29T23:18:18.505Z" },
+]
+
+[[package]]
+name = "parso"
+version = "0.8.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/81/76/a1e769043c0c0c9fe391b702539d594731a4362334cdf4dc25d0c09761e7/parso-0.8.6.tar.gz", hash = "sha256:2b9a0332696df97d454fa67b81618fd69c35a7b90327cbe6ba5c92d2c68a7bfd", size = 401621, upload-time = "2026-02-09T15:45:24.425Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b6/61/fae042894f4296ec49e3f193aff5d7c18440da9e48102c3315e1bc4519a7/parso-0.8.6-py2.py3-none-any.whl", hash = "sha256:2c549f800b70a5c4952197248825584cb00f033b29c692671d3bf08bf380baff", size = 106894, upload-time = "2026-02-09T15:45:21.391Z" },
+]
+
+[[package]]
+name = "peft"
+version = "0.15.2"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "pyyaml", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "safetensors", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/33/65/faa18cd8ffbe0f742c3f2559770646cce2574b9cd28a2a05e8d36f64e968/peft-0.15.2.tar.gz", hash = "sha256:7059029f4d42a092ded1aa117dd366a46084aef638bdd593f6ab0195d5427fcd", size = 472952, upload-time = "2025-04-15T15:27:53.09Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/68/85/8e6ea3d1089f2b6de3c1cd34bbbd7560912af9d34b057be3b8b8fefe1da3/peft-0.15.2-py3-none-any.whl", hash = "sha256:0dfc942b03b7af4b7267cd4e30b15e3a4a1d277adc581ce6245fc13f1f93d0a0", size = 411051, upload-time = "2025-04-15T15:27:50.799Z" },
+]
+
+[[package]]
+name = "peft"
+version = "0.18.1"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "pyyaml", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "safetensors", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform != 'linux' and sys_platform != 'win32' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'linux' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d8/48/147b3ea999560b40a34fd78724c7777aa9d18409c2250bdcaf9c4f2db7fc/peft-0.18.1.tar.gz", hash = "sha256:2dd0d6bfce936d1850e48aaddbd250941c5c02fc8ef3237cd8fd5aac35e0bae2", size = 635030, upload-time = "2026-01-09T13:08:01.136Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b3/14/b4e3f574acf349ae6f61f9c000a77f97a3b315b4bb6ad03791e79ae4a568/peft-0.18.1-py3-none-any.whl", hash = "sha256:0bf06847a3551e3019fc58c440cffc9a6b73e6e2962c95b52e224f77bbdb50f1", size = 556960, upload-time = "2026-01-09T13:07:55.865Z" },
+]
+
+[[package]]
+name = "pexpect"
+version = "4.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "ptyprocess", marker = "sys_platform != 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/42/92/cc564bf6381ff43ce1f4d06852fc19a2f11d180f23dc32d9588bee2f149d/pexpect-4.9.0.tar.gz", hash = "sha256:ee7d41123f3c9911050ea2c2dac107568dc43b2d3b0c7557a33212c398ead30f", size = 166450, upload-time = "2023-11-25T09:07:26.339Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9e/c3/059298687310d527a58bb01f3b1965787ee3b40dce76752eda8b44e9a2c5/pexpect-4.9.0-py2.py3-none-any.whl", hash = "sha256:7236d1e080e4936be2dc3e326cec0af72acf9212a7e1d060210e70a47e253523", size = 63772, upload-time = "2023-11-25T06:56:14.81Z" },
+]
+
+[[package]]
+name = "pillow"
+version = "11.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f3/0d/d0d6dea55cd152ce3d6767bb38a8fc10e33796ba4ba210cbab9354b6d238/pillow-11.3.0.tar.gz", hash = "sha256:3828ee7586cd0b2091b6209e5ad53e20d0649bbe87164a459d0676e035e8f523", size = 47113069, upload-time = "2025-07-01T09:16:30.666Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4c/5d/45a3553a253ac8763f3561371432a90bdbe6000fbdcf1397ffe502aa206c/pillow-11.3.0-cp310-cp310-macosx_10_10_x86_64.whl", hash = "sha256:1b9c17fd4ace828b3003dfd1e30bff24863e0eb59b535e8f80194d9cc7ecf860", size = 5316554, upload-time = "2025-07-01T09:13:39.342Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/c8/67c12ab069ef586a25a4a79ced553586748fad100c77c0ce59bb4983ac98/pillow-11.3.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:65dc69160114cdd0ca0f35cb434633c75e8e7fad4cf855177a05bf38678f73ad", size = 4686548, upload-time = "2025-07-01T09:13:41.835Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/bd/6741ebd56263390b382ae4c5de02979af7f8bd9807346d068700dd6d5cf9/pillow-11.3.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7107195ddc914f656c7fc8e4a5e1c25f32e9236ea3ea860f257b0436011fddd0", size = 5859742, upload-time = "2025-07-03T13:09:47.439Z" },
+    { url = "https://files.pythonhosted.org/packages/ca/0b/c412a9e27e1e6a829e6ab6c2dca52dd563efbedf4c9c6aa453d9a9b77359/pillow-11.3.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:cc3e831b563b3114baac7ec2ee86819eb03caa1a2cef0b481a5675b59c4fe23b", size = 7633087, upload-time = "2025-07-03T13:09:51.796Z" },
+    { url = "https://files.pythonhosted.org/packages/59/9d/9b7076aaf30f5dd17e5e5589b2d2f5a5d7e30ff67a171eb686e4eecc2adf/pillow-11.3.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f1f182ebd2303acf8c380a54f615ec883322593320a9b00438eb842c1f37ae50", size = 5963350, upload-time = "2025-07-01T09:13:43.865Z" },
+    { url = "https://files.pythonhosted.org/packages/f0/16/1a6bf01fb622fb9cf5c91683823f073f053005c849b1f52ed613afcf8dae/pillow-11.3.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4445fa62e15936a028672fd48c4c11a66d641d2c05726c7ec1f8ba6a572036ae", size = 6631840, upload-time = "2025-07-01T09:13:46.161Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/e6/6ff7077077eb47fde78739e7d570bdcd7c10495666b6afcd23ab56b19a43/pillow-11.3.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:71f511f6b3b91dd543282477be45a033e4845a40278fa8dcdbfdb07109bf18f9", size = 6074005, upload-time = "2025-07-01T09:13:47.829Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/3a/b13f36832ea6d279a697231658199e0a03cd87ef12048016bdcc84131601/pillow-11.3.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:040a5b691b0713e1f6cbe222e0f4f74cd233421e105850ae3b3c0ceda520f42e", size = 6708372, upload-time = "2025-07-01T09:13:52.145Z" },
+    { url = "https://files.pythonhosted.org/packages/6c/e4/61b2e1a7528740efbc70b3d581f33937e38e98ef3d50b05007267a55bcb2/pillow-11.3.0-cp310-cp310-win32.whl", hash = "sha256:89bd777bc6624fe4115e9fac3352c79ed60f3bb18651420635f26e643e3dd1f6", size = 6277090, upload-time = "2025-07-01T09:13:53.915Z" },
+    { url = "https://files.pythonhosted.org/packages/a9/d3/60c781c83a785d6afbd6a326ed4d759d141de43aa7365725cbcd65ce5e54/pillow-11.3.0-cp310-cp310-win_amd64.whl", hash = "sha256:19d2ff547c75b8e3ff46f4d9ef969a06c30ab2d4263a9e287733aa8b2429ce8f", size = 6985988, upload-time = "2025-07-01T09:13:55.699Z" },
+    { url = "https://files.pythonhosted.org/packages/9f/28/4f4a0203165eefb3763939c6789ba31013a2e90adffb456610f30f613850/pillow-11.3.0-cp310-cp310-win_arm64.whl", hash = "sha256:819931d25e57b513242859ce1876c58c59dc31587847bf74cfe06b2e0cb22d2f", size = 2422899, upload-time = "2025-07-01T09:13:57.497Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/8b/209bd6b62ce8367f47e68a218bffac88888fdf2c9fcf1ecadc6c3ec1ebc7/pillow-11.3.0-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:3cee80663f29e3843b68199b9d6f4f54bd1d4a6b59bdd91bceefc51238bcb967", size = 5270556, upload-time = "2025-07-01T09:16:09.961Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/e6/231a0b76070c2cfd9e260a7a5b504fb72da0a95279410fa7afd99d9751d6/pillow-11.3.0-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:b5f56c3f344f2ccaf0dd875d3e180f631dc60a51b314295a3e681fe8cf851fbe", size = 4654625, upload-time = "2025-07-01T09:16:11.913Z" },
+    { url = "https://files.pythonhosted.org/packages/13/f4/10cf94fda33cb12765f2397fc285fa6d8eb9c29de7f3185165b702fc7386/pillow-11.3.0-pp310-pypy310_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e67d793d180c9df62f1f40aee3accca4829d3794c95098887edc18af4b8b780c", size = 4874207, upload-time = "2025-07-03T13:11:10.201Z" },
+    { url = "https://files.pythonhosted.org/packages/72/c9/583821097dc691880c92892e8e2d41fe0a5a3d6021f4963371d2f6d57250/pillow-11.3.0-pp310-pypy310_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d000f46e2917c705e9fb93a3606ee4a819d1e3aa7a9b442f6444f07e77cf5e25", size = 6583939, upload-time = "2025-07-03T13:11:15.68Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/8e/5c9d410f9217b12320efc7c413e72693f48468979a013ad17fd690397b9a/pillow-11.3.0-pp310-pypy310_pp73-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:527b37216b6ac3a12d7838dc3bd75208ec57c1c6d11ef01902266a5a0c14fc27", size = 4957166, upload-time = "2025-07-01T09:16:13.74Z" },
+    { url = "https://files.pythonhosted.org/packages/62/bb/78347dbe13219991877ffb3a91bf09da8317fbfcd4b5f9140aeae020ad71/pillow-11.3.0-pp310-pypy310_pp73-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:be5463ac478b623b9dd3937afd7fb7ab3d79dd290a28e2b6df292dc75063eb8a", size = 5581482, upload-time = "2025-07-01T09:16:16.107Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/28/1000353d5e61498aaeaaf7f1e4b49ddb05f2c6575f9d4f9f914a3538b6e1/pillow-11.3.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:8dc70ca24c110503e16918a658b869019126ecfe03109b754c402daff12b3d9f", size = 6984596, upload-time = "2025-07-01T09:16:18.07Z" },
+]
+
+[[package]]
+name = "platformdirs"
+version = "4.9.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/19/56/8d4c30c8a1d07013911a8fdbd8f89440ef9f08d07a1b50ab8ca8be5a20f9/platformdirs-4.9.4.tar.gz", hash = "sha256:1ec356301b7dc906d83f371c8f487070e99d3ccf9e501686456394622a01a934", size = 28737, upload-time = "2026-03-05T18:34:13.271Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/63/d7/97f7e3a6abb67d8080dd406fd4df842c2be0efaf712d1c899c32a075027c/platformdirs-4.9.4-py3-none-any.whl", hash = "sha256:68a9a4619a666ea6439f2ff250c12a853cd1cbd5158d258bd824a7df6be2f868", size = 21216, upload-time = "2026-03-05T18:34:12.172Z" },
+]
+
+[[package]]
+name = "plotly"
+version = "6.6.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "narwhals" },
+    { name = "packaging" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/24/fb/41efe84970cfddefd4ccf025e2cbfafe780004555f583e93dba3dac2cdef/plotly-6.6.0.tar.gz", hash = "sha256:b897f15f3b02028d69f755f236be890ba950d0a42d7dfc619b44e2d8cea8748c", size = 7027956, upload-time = "2026-03-02T21:10:25.321Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/52/d2/c6e44dba74f17c6216ce1b56044a9b93a929f1c2d5bdaff892512b260f5e/plotly-6.6.0-py3-none-any.whl", hash = "sha256:8d6daf0f87412e0c0bfe72e809d615217ab57cc715899a1e5145135a7800d1d0", size = 9910315, upload-time = "2026-03-02T21:10:18.131Z" },
+]
+
+[[package]]
+name = "pluggy"
+version = "1.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f9/e2/3e91f31a7d2b083fe6ef3fa267035b518369d9511ffab804f839851d2779/pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3", size = 69412, upload-time = "2025-05-15T12:30:07.975Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/54/20/4d324d65cc6d9205fabedc306948156824eb9f0ee1633355a8f7ec5c66bf/pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746", size = 20538, upload-time = "2025-05-15T12:30:06.134Z" },
+]
+
+[[package]]
+name = "pre-commit"
+version = "4.5.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cfgv" },
+    { name = "identify" },
+    { name = "nodeenv" },
+    { name = "pyyaml" },
+    { name = "virtualenv" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/40/f1/6d86a29246dfd2e9b6237f0b5823717f60cad94d47ddc26afa916d21f525/pre_commit-4.5.1.tar.gz", hash = "sha256:eb545fcff725875197837263e977ea257a402056661f09dae08e4b149b030a61", size = 198232, upload-time = "2025-12-16T21:14:33.552Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/19/fd3ef348460c80af7bb4669ea7926651d1f95c23ff2df18b9d24bab4f3fa/pre_commit-4.5.1-py2.py3-none-any.whl", hash = "sha256:3b3afd891e97337708c1674210f8eba659b52a38ea5f822ff142d10786221f77", size = 226437, upload-time = "2025-12-16T21:14:32.409Z" },
+]
+
+[[package]]
+name = "proglog"
+version = "0.1.12"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "tqdm" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c2/af/c108866c452eda1132f3d6b3cb6be2ae8430c97e9309f38ca9dbd430af37/proglog-0.1.12.tar.gz", hash = "sha256:361ee074721c277b89b75c061336cb8c5f287c92b043efa562ccf7866cda931c", size = 8794, upload-time = "2025-05-09T14:36:18.316Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c1/1b/f7ea6cde25621cd9236541c66ff018f4268012a534ec31032bcb187dc5e7/proglog-0.1.12-py3-none-any.whl", hash = "sha256:ccaafce51e80a81c65dc907a460c07ccb8ec1f78dc660cfd8f9ec3a22f01b84c", size = 6337, upload-time = "2025-05-09T14:36:16.798Z" },
+]
+
+[[package]]
+name = "promise"
+version = "2.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/cf/9c/fb5d48abfe5d791cd496e4242ebcf87a4bb2e0c3dcd6e0ae68c11426a528/promise-2.3.tar.gz", hash = "sha256:dfd18337c523ba4b6a58801c164c1904a9d4d1b1747c7d5dbf45b693a49d93d0", size = 19534, upload-time = "2019-12-18T07:31:43.07Z" }
+
+[[package]]
+name = "prompt-toolkit"
+version = "3.0.52"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "wcwidth" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a1/96/06e01a7b38dce6fe1db213e061a4602dd6032a8a97ef6c1a862537732421/prompt_toolkit-3.0.52.tar.gz", hash = "sha256:28cde192929c8e7321de85de1ddbe736f1375148b02f2e17edd840042b1be855", size = 434198, upload-time = "2025-08-27T15:24:02.057Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/84/03/0d3ce49e2505ae70cf43bc5bb3033955d2fc9f932163e84dc0779cc47f48/prompt_toolkit-3.0.52-py3-none-any.whl", hash = "sha256:9aac639a3bbd33284347de5ad8d68ecc044b91a762dc39b7c21095fcd6a19955", size = 391431, upload-time = "2025-08-27T15:23:59.498Z" },
+]
+
+[[package]]
+name = "propcache"
+version = "0.4.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9e/da/e9fc233cf63743258bff22b3dfa7ea5baef7b5bc324af47a0ad89b8ffc6f/propcache-0.4.1.tar.gz", hash = "sha256:f48107a8c637e80362555f37ecf49abe20370e557cc4ab374f04ec4423c97c3d", size = 46442, upload-time = "2025-10-08T19:49:02.291Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3c/0e/934b541323035566a9af292dba85a195f7b78179114f2c6ebb24551118a9/propcache-0.4.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:7c2d1fa3201efaf55d730400d945b5b3ab6e672e100ba0f9a409d950ab25d7db", size = 79534, upload-time = "2025-10-08T19:46:02.083Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/6b/db0d03d96726d995dc7171286c6ba9d8d14251f37433890f88368951a44e/propcache-0.4.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:1eb2994229cc8ce7fe9b3db88f5465f5fd8651672840b2e426b88cdb1a30aac8", size = 45526, upload-time = "2025-10-08T19:46:03.884Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/c3/82728404aea669e1600f304f2609cde9e665c18df5a11cdd57ed73c1dceb/propcache-0.4.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:66c1f011f45a3b33d7bcb22daed4b29c0c9e2224758b6be00686731e1b46f925", size = 47263, upload-time = "2025-10-08T19:46:05.405Z" },
+    { url = "https://files.pythonhosted.org/packages/df/1b/39313ddad2bf9187a1432654c38249bab4562ef535ef07f5eb6eb04d0b1b/propcache-0.4.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9a52009f2adffe195d0b605c25ec929d26b36ef986ba85244891dee3b294df21", size = 201012, upload-time = "2025-10-08T19:46:07.165Z" },
+    { url = "https://files.pythonhosted.org/packages/5b/01/f1d0b57d136f294a142acf97f4ed58c8e5b974c21e543000968357115011/propcache-0.4.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5d4e2366a9c7b837555cf02fb9be2e3167d333aff716332ef1b7c3a142ec40c5", size = 209491, upload-time = "2025-10-08T19:46:08.909Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/c8/038d909c61c5bb039070b3fb02ad5cccdb1dde0d714792e251cdb17c9c05/propcache-0.4.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9d2b6caef873b4f09e26ea7e33d65f42b944837563a47a94719cc3544319a0db", size = 215319, upload-time = "2025-10-08T19:46:10.7Z" },
+    { url = "https://files.pythonhosted.org/packages/08/57/8c87e93142b2c1fa2408e45695205a7ba05fb5db458c0bf5c06ba0e09ea6/propcache-0.4.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2b16ec437a8c8a965ecf95739448dd938b5c7f56e67ea009f4300d8df05f32b7", size = 196856, upload-time = "2025-10-08T19:46:12.003Z" },
+    { url = "https://files.pythonhosted.org/packages/42/df/5615fec76aa561987a534759b3686008a288e73107faa49a8ae5795a9f7a/propcache-0.4.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:296f4c8ed03ca7476813fe666c9ea97869a8d7aec972618671b33a38a5182ef4", size = 193241, upload-time = "2025-10-08T19:46:13.495Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/21/62949eb3a7a54afe8327011c90aca7e03547787a88fb8bd9726806482fea/propcache-0.4.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:1f0978529a418ebd1f49dad413a2b68af33f85d5c5ca5c6ca2a3bed375a7ac60", size = 190552, upload-time = "2025-10-08T19:46:14.938Z" },
+    { url = "https://files.pythonhosted.org/packages/30/ee/ab4d727dd70806e5b4de96a798ae7ac6e4d42516f030ee60522474b6b332/propcache-0.4.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:fd138803047fb4c062b1c1dd95462f5209456bfab55c734458f15d11da288f8f", size = 200113, upload-time = "2025-10-08T19:46:16.695Z" },
+    { url = "https://files.pythonhosted.org/packages/8a/0b/38b46208e6711b016aa8966a3ac793eee0d05c7159d8342aa27fc0bc365e/propcache-0.4.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:8c9b3cbe4584636d72ff556d9036e0c9317fa27b3ac1f0f558e7e84d1c9c5900", size = 200778, upload-time = "2025-10-08T19:46:18.023Z" },
+    { url = "https://files.pythonhosted.org/packages/cf/81/5abec54355ed344476bee711e9f04815d4b00a311ab0535599204eecc257/propcache-0.4.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:f93243fdc5657247533273ac4f86ae106cc6445a0efacb9a1bfe982fcfefd90c", size = 193047, upload-time = "2025-10-08T19:46:19.449Z" },
+    { url = "https://files.pythonhosted.org/packages/ec/b6/1f237c04e32063cb034acd5f6ef34ef3a394f75502e72703545631ab1ef6/propcache-0.4.1-cp310-cp310-win32.whl", hash = "sha256:a0ee98db9c5f80785b266eb805016e36058ac72c51a064040f2bc43b61101cdb", size = 38093, upload-time = "2025-10-08T19:46:20.643Z" },
+    { url = "https://files.pythonhosted.org/packages/a6/67/354aac4e0603a15f76439caf0427781bcd6797f370377f75a642133bc954/propcache-0.4.1-cp310-cp310-win_amd64.whl", hash = "sha256:1cdb7988c4e5ac7f6d175a28a9aa0c94cb6f2ebe52756a3c0cda98d2809a9e37", size = 41638, upload-time = "2025-10-08T19:46:21.935Z" },
+    { url = "https://files.pythonhosted.org/packages/e0/e1/74e55b9fd1a4c209ff1a9a824bf6c8b3d1fc5a1ac3eabe23462637466785/propcache-0.4.1-cp310-cp310-win_arm64.whl", hash = "sha256:d82ad62b19645419fe79dd63b3f9253e15b30e955c0170e5cebc350c1844e581", size = 38229, upload-time = "2025-10-08T19:46:23.368Z" },
+    { url = "https://files.pythonhosted.org/packages/5b/5a/bc7b4a4ef808fa59a816c17b20c4bef6884daebbdf627ff2a161da67da19/propcache-0.4.1-py3-none-any.whl", hash = "sha256:af2a6052aeb6cf17d3e46ee169099044fd8224cbaf75c76a2ef596e8163e2237", size = 13305, upload-time = "2025-10-08T19:49:00.792Z" },
+]
+
+[[package]]
+name = "proto-plus"
+version = "1.27.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "protobuf" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3a/02/8832cde80e7380c600fbf55090b6ab7b62bd6825dbedde6d6657c15a1f8e/proto_plus-1.27.1.tar.gz", hash = "sha256:912a7460446625b792f6448bade9e55cd4e41e6ac10e27009ef71a7f317fa147", size = 56929, upload-time = "2026-02-02T17:34:49.035Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/79/ac273cbbf744691821a9cca88957257f41afe271637794975ca090b9588b/proto_plus-1.27.1-py3-none-any.whl", hash = "sha256:e4643061f3a4d0de092d62aa4ad09fa4756b2cbb89d4627f3985018216f9fefc", size = 50480, upload-time = "2026-02-02T17:34:47.339Z" },
+]
+
+[[package]]
+name = "protobuf"
+version = "3.20.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/55/5b/e3d951e34f8356e5feecacd12a8e3b258a1da6d9a03ad1770f28925f29bc/protobuf-3.20.3.tar.gz", hash = "sha256:2e3427429c9cffebf259491be0af70189607f365c2f41c7c3764af6f337105f2", size = 216768, upload-time = "2022-09-29T22:39:47.592Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/28/55/b80e8567ec327c060fa39b242392e25690c8899c489ecd7bb65b46b7bb55/protobuf-3.20.3-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:f4bd856d702e5b0d96a00ec6b307b0f51c1982c2bf9c0052cf9019e9a544ba99", size = 918427, upload-time = "2022-09-29T22:10:52.591Z" },
+    { url = "https://files.pythonhosted.org/packages/31/be/80a9c6f16dfa4d41be3edbe655349778ae30882407fa8275eb46b4d34854/protobuf-3.20.3-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:9aae4406ea63d825636cc11ffb34ad3379335803216ee3a856787bcf5ccc751e", size = 1051042, upload-time = "2022-09-29T22:10:54.629Z" },
+    { url = "https://files.pythonhosted.org/packages/db/96/948d3fcc1fa816e7ae1d27af59b9d8c5c5e582f3994fd14394f31da95b99/protobuf-3.20.3-cp310-cp310-win32.whl", hash = "sha256:28545383d61f55b57cf4df63eebd9827754fd2dc25f80c5253f9184235db242c", size = 780167, upload-time = "2022-09-29T22:10:56.276Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/5e/fc6feb366b0a9f28e0a2de3b062667c521cd9517d4ff55077b8f351ba2f3/protobuf-3.20.3-cp310-cp310-win_amd64.whl", hash = "sha256:67a3598f0a2dcbc58d02dd1928544e7d88f764b47d4a286202913f0b2801c2e7", size = 904029, upload-time = "2022-09-29T22:10:57.866Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/14/619e24a4c70df2901e1f4dbc50a6291eb63a759172558df326347dce1f0d/protobuf-3.20.3-py2.py3-none-any.whl", hash = "sha256:a7ca6d488aa8ff7f329d4c545b2dbad8ac31464f1d8b1c87ad1346717731e4db", size = 162128, upload-time = "2022-09-29T22:39:44.547Z" },
+]
+
+[[package]]
+name = "psutil"
+version = "7.2.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/aa/c6/d1ddf4abb55e93cebc4f2ed8b5d6dbad109ecb8d63748dd2b20ab5e57ebe/psutil-7.2.2.tar.gz", hash = "sha256:0746f5f8d406af344fd547f1c8daa5f5c33dbc293bb8d6a16d80b4bb88f59372", size = 493740, upload-time = "2026-01-28T18:14:54.428Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e7/36/5ee6e05c9bd427237b11b3937ad82bb8ad2752d72c6969314590dd0c2f6e/psutil-7.2.2-cp36-abi3-macosx_10_9_x86_64.whl", hash = "sha256:ed0cace939114f62738d808fdcecd4c869222507e266e574799e9c0faa17d486", size = 129090, upload-time = "2026-01-28T18:15:22.168Z" },
+    { url = "https://files.pythonhosted.org/packages/80/c4/f5af4c1ca8c1eeb2e92ccca14ce8effdeec651d5ab6053c589b074eda6e1/psutil-7.2.2-cp36-abi3-macosx_11_0_arm64.whl", hash = "sha256:1a7b04c10f32cc88ab39cbf606e117fd74721c831c98a27dc04578deb0c16979", size = 129859, upload-time = "2026-01-28T18:15:23.795Z" },
+    { url = "https://files.pythonhosted.org/packages/b5/70/5d8df3b09e25bce090399cf48e452d25c935ab72dad19406c77f4e828045/psutil-7.2.2-cp36-abi3-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:076a2d2f923fd4821644f5ba89f059523da90dc9014e85f8e45a5774ca5bc6f9", size = 155560, upload-time = "2026-01-28T18:15:25.976Z" },
+    { url = "https://files.pythonhosted.org/packages/63/65/37648c0c158dc222aba51c089eb3bdfa238e621674dc42d48706e639204f/psutil-7.2.2-cp36-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b0726cecd84f9474419d67252add4ac0cd9811b04d61123054b9fb6f57df6e9e", size = 156997, upload-time = "2026-01-28T18:15:27.794Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/13/125093eadae863ce03c6ffdbae9929430d116a246ef69866dad94da3bfbc/psutil-7.2.2-cp36-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:fd04ef36b4a6d599bbdb225dd1d3f51e00105f6d48a28f006da7f9822f2606d8", size = 148972, upload-time = "2026-01-28T18:15:29.342Z" },
+    { url = "https://files.pythonhosted.org/packages/04/78/0acd37ca84ce3ddffaa92ef0f571e073faa6d8ff1f0559ab1272188ea2be/psutil-7.2.2-cp36-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:b58fabe35e80b264a4e3bb23e6b96f9e45a3df7fb7eed419ac0e5947c61e47cc", size = 148266, upload-time = "2026-01-28T18:15:31.597Z" },
+    { url = "https://files.pythonhosted.org/packages/b4/90/e2159492b5426be0c1fef7acba807a03511f97c5f86b3caeda6ad92351a7/psutil-7.2.2-cp37-abi3-win_amd64.whl", hash = "sha256:eb7e81434c8d223ec4a219b5fc1c47d0417b12be7ea866e24fb5ad6e84b3d988", size = 137737, upload-time = "2026-01-28T18:15:33.849Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/c7/7bb2e321574b10df20cbde462a94e2b71d05f9bbda251ef27d104668306a/psutil-7.2.2-cp37-abi3-win_arm64.whl", hash = "sha256:8c233660f575a5a89e6d4cb65d9f938126312bca76d8fe087b947b3a1aaac9ee", size = 134617, upload-time = "2026-01-28T18:15:36.514Z" },
+]
+
+[[package]]
+name = "ptyprocess"
+version = "0.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/20/e5/16ff212c1e452235a90aeb09066144d0c5a6a8c0834397e03f5224495c4e/ptyprocess-0.7.0.tar.gz", hash = "sha256:5c5d0a3b48ceee0b48485e0c26037c0acd7d29765ca3fbb5cb3831d347423220", size = 70762, upload-time = "2020-12-28T15:15:30.155Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/22/a6/858897256d0deac81a172289110f31629fc4cee19b6f01283303e18c8db3/ptyprocess-0.7.0-py2.py3-none-any.whl", hash = "sha256:4b41f3967fce3af57cc7e94b888626c18bf37a083e3651ca8feeb66d492fef35", size = 13993, upload-time = "2020-12-28T15:15:28.35Z" },
+]
+
+[[package]]
+name = "pure-eval"
+version = "0.2.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/05/0a34433a064256a578f1783a10da6df098ceaa4a57bbeaa96a6c0352786b/pure_eval-0.2.3.tar.gz", hash = "sha256:5f4e983f40564c576c7c8635ae88db5956bb2229d7e9237d03b3c0b0190eaf42", size = 19752, upload-time = "2024-07-21T12:58:21.801Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8e/37/efad0257dc6e593a18957422533ff0f87ede7c9c6ea010a2177d738fb82f/pure_eval-0.2.3-py3-none-any.whl", hash = "sha256:1db8e35b67b3d218d818ae653e27f06c3aa420901fa7b081ca98cbedc874e0d0", size = 11842, upload-time = "2024-07-21T12:58:20.04Z" },
+]
+
+[[package]]
+name = "pyarrow"
+version = "23.0.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/88/22/134986a4cc224d593c1afde5494d18ff629393d74cc2eddb176669f234a4/pyarrow-23.0.1.tar.gz", hash = "sha256:b8c5873e33440b2bc2f4a79d2b47017a89c5a24116c055625e6f2ee50523f019", size = 1167336, upload-time = "2026-02-16T10:14:12.39Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/bc/a8/24e5dc6855f50a62936ceb004e6e9645e4219a8065f304145d7fb8a79d5d/pyarrow-23.0.1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:3fab8f82571844eb3c460f90a75583801d14ca0cc32b1acc8c361650e006fd56", size = 34307390, upload-time = "2026-02-16T10:08:08.654Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/8e/4be5617b4aaae0287f621ad31c6036e5f63118cfca0dc57d42121ff49b51/pyarrow-23.0.1-cp310-cp310-macosx_12_0_x86_64.whl", hash = "sha256:3f91c038b95f71ddfc865f11d5876c42f343b4495535bd262c7b321b0b94507c", size = 35853761, upload-time = "2026-02-16T10:08:17.811Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/08/3e56a18819462210432ae37d10f5c8eed3828be1d6c751b6e6a2e93c286a/pyarrow-23.0.1-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:d0744403adabef53c985a7f8a082b502a368510c40d184df349a0a8754533258", size = 44493116, upload-time = "2026-02-16T10:08:25.792Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/82/c40b68001dbec8a3faa4c08cd8c200798ac732d2854537c5449dc859f55a/pyarrow-23.0.1-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:c33b5bf406284fd0bba436ed6f6c3ebe8e311722b441d89397c54f871c6863a2", size = 47564532, upload-time = "2026-02-16T10:08:34.27Z" },
+    { url = "https://files.pythonhosted.org/packages/20/bc/73f611989116b6f53347581b02177f9f620efdf3cd3f405d0e83cdf53a83/pyarrow-23.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ddf743e82f69dcd6dbbcb63628895d7161e04e56794ef80550ac6f3315eeb1d5", size = 48183685, upload-time = "2026-02-16T10:08:42.889Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/cc/6c6b3ecdae2a8c3aced99956187e8302fc954cc2cca2a37cf2111dad16ce/pyarrow-23.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:e052a211c5ac9848ae15d5ec875ed0943c0221e2fcfe69eee80b604b4e703222", size = 50605582, upload-time = "2026-02-16T10:08:51.641Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/94/d359e708672878d7638a04a0448edf7c707f9e5606cee11e15aaa5c7535a/pyarrow-23.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:5abde149bb3ce524782d838eb67ac095cd3fd6090eba051130589793f1a7f76d", size = 27521148, upload-time = "2026-02-16T10:08:58.077Z" },
+]
+
+[[package]]
+name = "pyasn1"
+version = "0.6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fe/b6/6e630dff89739fcd427e3f72b3d905ce0acb85a45d4ec3e2678718a3487f/pyasn1-0.6.2.tar.gz", hash = "sha256:9b59a2b25ba7e4f8197db7686c09fb33e658b98339fadb826e9512629017833b", size = 146586, upload-time = "2026-01-16T18:04:18.534Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/44/b5/a96872e5184f354da9c84ae119971a0a4c221fe9b27a4d94bd43f2596727/pyasn1-0.6.2-py3-none-any.whl", hash = "sha256:1eb26d860996a18e9b6ed05e7aae0e9fc21619fcee6af91cca9bad4fbea224bf", size = 83371, upload-time = "2026-01-16T18:04:17.174Z" },
+]
+
+[[package]]
+name = "pyasn1-modules"
+version = "0.4.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyasn1" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e9/e6/78ebbb10a8c8e4b61a59249394a4a594c1a7af95593dc933a349c8d00964/pyasn1_modules-0.4.2.tar.gz", hash = "sha256:677091de870a80aae844b1ca6134f54652fa2c8c5a52aa396440ac3106e941e6", size = 307892, upload-time = "2025-03-28T02:41:22.17Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/47/8d/d529b5d697919ba8c11ad626e835d4039be708a35b0d22de83a269a6682c/pyasn1_modules-0.4.2-py3-none-any.whl", hash = "sha256:29253a9207ce32b64c3ac6600edc75368f98473906e8fd1043bd6b5b1de2c14a", size = 181259, upload-time = "2025-03-28T02:41:19.028Z" },
+]
+
+[[package]]
+name = "pycparser"
+version = "3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1b/7d/92392ff7815c21062bea51aa7b87d45576f649f16458d78b7cf94b9ab2e6/pycparser-3.0.tar.gz", hash = "sha256:600f49d217304a5902ac3c37e1281c9fe94e4d0489de643a9504c5cdfdfc6b29", size = 103492, upload-time = "2026-01-21T14:26:51.89Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0c/c3/44f3fbbfa403ea2a7c779186dc20772604442dde72947e7d01069cbe98e3/pycparser-3.0-py3-none-any.whl", hash = "sha256:b727414169a36b7d524c1c3e31839a521725078d7b2ff038656844266160a992", size = 48172, upload-time = "2026-01-21T14:26:50.693Z" },
+]
+
+[[package]]
+name = "pycryptodome"
+version = "3.23.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/8e/a6/8452177684d5e906854776276ddd34eca30d1b1e15aa1ee9cefc289a33f5/pycryptodome-3.23.0.tar.gz", hash = "sha256:447700a657182d60338bab09fdb27518f8856aecd80ae4c6bdddb67ff5da44ef", size = 4921276, upload-time = "2025-05-17T17:21:45.242Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/db/6c/a1f71542c969912bb0e106f64f60a56cc1f0fabecf9396f45accbe63fa68/pycryptodome-3.23.0-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:187058ab80b3281b1de11c2e6842a357a1f71b42cb1e15bce373f3d238135c27", size = 2495627, upload-time = "2025-05-17T17:20:47.139Z" },
+    { url = "https://files.pythonhosted.org/packages/6e/4e/a066527e079fc5002390c8acdd3aca431e6ea0a50ffd7201551175b47323/pycryptodome-3.23.0-cp37-abi3-macosx_10_9_x86_64.whl", hash = "sha256:cfb5cd445280c5b0a4e6187a7ce8de5a07b5f3f897f235caa11f1f435f182843", size = 1640362, upload-time = "2025-05-17T17:20:50.392Z" },
+    { url = "https://files.pythonhosted.org/packages/50/52/adaf4c8c100a8c49d2bd058e5b551f73dfd8cb89eb4911e25a0c469b6b4e/pycryptodome-3.23.0-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:67bd81fcbe34f43ad9422ee8fd4843c8e7198dd88dd3d40e6de42ee65fbe1490", size = 2182625, upload-time = "2025-05-17T17:20:52.866Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/e9/a09476d436d0ff1402ac3867d933c61805ec2326c6ea557aeeac3825604e/pycryptodome-3.23.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c8987bd3307a39bc03df5c8e0e3d8be0c4c3518b7f044b0f4c15d1aa78f52575", size = 2268954, upload-time = "2025-05-17T17:20:55.027Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/c5/ffe6474e0c551d54cab931918127c46d70cab8f114e0c2b5a3c071c2f484/pycryptodome-3.23.0-cp37-abi3-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:aa0698f65e5b570426fc31b8162ed4603b0c2841cbb9088e2b01641e3065915b", size = 2308534, upload-time = "2025-05-17T17:20:57.279Z" },
+    { url = "https://files.pythonhosted.org/packages/18/28/e199677fc15ecf43010f2463fde4c1a53015d1fe95fb03bca2890836603a/pycryptodome-3.23.0-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:53ecbafc2b55353edcebd64bf5da94a2a2cdf5090a6915bcca6eca6cc452585a", size = 2181853, upload-time = "2025-05-17T17:20:59.322Z" },
+    { url = "https://files.pythonhosted.org/packages/ce/ea/4fdb09f2165ce1365c9eaefef36625583371ee514db58dc9b65d3a255c4c/pycryptodome-3.23.0-cp37-abi3-musllinux_1_2_i686.whl", hash = "sha256:156df9667ad9f2ad26255926524e1c136d6664b741547deb0a86a9acf5ea631f", size = 2342465, upload-time = "2025-05-17T17:21:03.83Z" },
+    { url = "https://files.pythonhosted.org/packages/22/82/6edc3fc42fe9284aead511394bac167693fb2b0e0395b28b8bedaa07ef04/pycryptodome-3.23.0-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:dea827b4d55ee390dc89b2afe5927d4308a8b538ae91d9c6f7a5090f397af1aa", size = 2267414, upload-time = "2025-05-17T17:21:06.72Z" },
+    { url = "https://files.pythonhosted.org/packages/59/fe/aae679b64363eb78326c7fdc9d06ec3de18bac68be4b612fc1fe8902693c/pycryptodome-3.23.0-cp37-abi3-win32.whl", hash = "sha256:507dbead45474b62b2bbe318eb1c4c8ee641077532067fec9c1aa82c31f84886", size = 1768484, upload-time = "2025-05-17T17:21:08.535Z" },
+    { url = "https://files.pythonhosted.org/packages/54/2f/e97a1b8294db0daaa87012c24a7bb714147c7ade7656973fd6c736b484ff/pycryptodome-3.23.0-cp37-abi3-win_amd64.whl", hash = "sha256:c75b52aacc6c0c260f204cbdd834f76edc9fb0d8e0da9fbf8352ef58202564e2", size = 1799636, upload-time = "2025-05-17T17:21:10.393Z" },
+    { url = "https://files.pythonhosted.org/packages/18/3d/f9441a0d798bf2b1e645adc3265e55706aead1255ccdad3856dbdcffec14/pycryptodome-3.23.0-cp37-abi3-win_arm64.whl", hash = "sha256:11eeeb6917903876f134b56ba11abe95c0b0fd5e3330def218083c7d98bbcb3c", size = 1703675, upload-time = "2025-05-17T17:21:13.146Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/12/e33935a0709c07de084d7d58d330ec3f4daf7910a18e77937affdb728452/pycryptodome-3.23.0-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:ddb95b49df036ddd264a0ad246d1be5b672000f12d6961ea2c267083a5e19379", size = 1623886, upload-time = "2025-05-17T17:21:20.614Z" },
+    { url = "https://files.pythonhosted.org/packages/22/0b/aa8f9419f25870889bebf0b26b223c6986652bdf071f000623df11212c90/pycryptodome-3.23.0-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d8e95564beb8782abfd9e431c974e14563a794a4944c29d6d3b7b5ea042110b4", size = 1672151, upload-time = "2025-05-17T17:21:22.666Z" },
+    { url = "https://files.pythonhosted.org/packages/d4/5e/63f5cbde2342b7f70a39e591dbe75d9809d6338ce0b07c10406f1a140cdc/pycryptodome-3.23.0-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:14e15c081e912c4b0d75632acd8382dfce45b258667aa3c67caf7a4d4c13f630", size = 1664461, upload-time = "2025-05-17T17:21:25.225Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/92/608fbdad566ebe499297a86aae5f2a5263818ceeecd16733006f1600403c/pycryptodome-3.23.0-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a7fc76bf273353dc7e5207d172b83f569540fc9a28d63171061c42e361d22353", size = 1702440, upload-time = "2025-05-17T17:21:27.991Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/92/2eadd1341abd2989cce2e2740b4423608ee2014acb8110438244ee97d7ff/pycryptodome-3.23.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:45c69ad715ca1a94f778215a11e66b7ff989d792a4d63b68dc586a1da1392ff5", size = 1803005, upload-time = "2025-05-17T17:21:31.37Z" },
+]
+
+[[package]]
+name = "pydantic"
+version = "2.12.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "annotated-types" },
+    { name = "pydantic-core" },
+    { name = "typing-extensions" },
+    { name = "typing-inspection" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/69/44/36f1a6e523abc58ae5f928898e4aca2e0ea509b5aa6f6f392a5d882be928/pydantic-2.12.5.tar.gz", hash = "sha256:4d351024c75c0f085a9febbb665ce8c0c6ec5d30e903bdb6394b7ede26aebb49", size = 821591, upload-time = "2025-11-26T15:11:46.471Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5a/87/b70ad306ebb6f9b585f114d0ac2137d792b48be34d732d60e597c2f8465a/pydantic-2.12.5-py3-none-any.whl", hash = "sha256:e561593fccf61e8a20fc46dfc2dfe075b8be7d0188df33f221ad1f0139180f9d", size = 463580, upload-time = "2025-11-26T15:11:44.605Z" },
+]
+
+[[package]]
+name = "pydantic-core"
+version = "2.41.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/71/70/23b021c950c2addd24ec408e9ab05d59b035b39d97cdc1130e1bce647bb6/pydantic_core-2.41.5.tar.gz", hash = "sha256:08daa51ea16ad373ffd5e7606252cc32f07bc72b28284b6bc9c6df804816476e", size = 460952, upload-time = "2025-11-04T13:43:49.098Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c6/90/32c9941e728d564b411d574d8ee0cf09b12ec978cb22b294995bae5549a5/pydantic_core-2.41.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:77b63866ca88d804225eaa4af3e664c5faf3568cea95360d21f4725ab6e07146", size = 2107298, upload-time = "2025-11-04T13:39:04.116Z" },
+    { url = "https://files.pythonhosted.org/packages/fb/a8/61c96a77fe28993d9a6fb0f4127e05430a267b235a124545d79fea46dd65/pydantic_core-2.41.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dfa8a0c812ac681395907e71e1274819dec685fec28273a28905df579ef137e2", size = 1901475, upload-time = "2025-11-04T13:39:06.055Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/b6/338abf60225acc18cdc08b4faef592d0310923d19a87fba1faf05af5346e/pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5921a4d3ca3aee735d9fd163808f5e8dd6c6972101e4adbda9a4667908849b97", size = 1918815, upload-time = "2025-11-04T13:39:10.41Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/1c/2ed0433e682983d8e8cba9c8d8ef274d4791ec6a6f24c58935b90e780e0a/pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e25c479382d26a2a41b7ebea1043564a937db462816ea07afa8a44c0866d52f9", size = 2065567, upload-time = "2025-11-04T13:39:12.244Z" },
+    { url = "https://files.pythonhosted.org/packages/b3/24/cf84974ee7d6eae06b9e63289b7b8f6549d416b5c199ca2d7ce13bbcf619/pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f547144f2966e1e16ae626d8ce72b4cfa0caedc7fa28052001c94fb2fcaa1c52", size = 2230442, upload-time = "2025-11-04T13:39:13.962Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/21/4e287865504b3edc0136c89c9c09431be326168b1eb7841911cbc877a995/pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6f52298fbd394f9ed112d56f3d11aabd0d5bd27beb3084cc3d8ad069483b8941", size = 2350956, upload-time = "2025-11-04T13:39:15.889Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/76/7727ef2ffa4b62fcab916686a68a0426b9b790139720e1934e8ba797e238/pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:100baa204bb412b74fe285fb0f3a385256dad1d1879f0a5cb1499ed2e83d132a", size = 2068253, upload-time = "2025-11-04T13:39:17.403Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/8c/a4abfc79604bcb4c748e18975c44f94f756f08fb04218d5cb87eb0d3a63e/pydantic_core-2.41.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:05a2c8852530ad2812cb7914dc61a1125dc4e06252ee98e5638a12da6cc6fb6c", size = 2177050, upload-time = "2025-11-04T13:39:19.351Z" },
+    { url = "https://files.pythonhosted.org/packages/67/b1/de2e9a9a79b480f9cb0b6e8b6ba4c50b18d4e89852426364c66aa82bb7b3/pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:29452c56df2ed968d18d7e21f4ab0ac55e71dc59524872f6fc57dcf4a3249ed2", size = 2147178, upload-time = "2025-11-04T13:39:21Z" },
+    { url = "https://files.pythonhosted.org/packages/16/c1/dfb33f837a47b20417500efaa0378adc6635b3c79e8369ff7a03c494b4ac/pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:d5160812ea7a8a2ffbe233d8da666880cad0cbaf5d4de74ae15c313213d62556", size = 2341833, upload-time = "2025-11-04T13:39:22.606Z" },
+    { url = "https://files.pythonhosted.org/packages/47/36/00f398642a0f4b815a9a558c4f1dca1b4020a7d49562807d7bc9ff279a6c/pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:df3959765b553b9440adfd3c795617c352154e497a4eaf3752555cfb5da8fc49", size = 2321156, upload-time = "2025-11-04T13:39:25.843Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/70/cad3acd89fde2010807354d978725ae111ddf6d0ea46d1ea1775b5c1bd0c/pydantic_core-2.41.5-cp310-cp310-win32.whl", hash = "sha256:1f8d33a7f4d5a7889e60dc39856d76d09333d8a6ed0f5f1190635cbec70ec4ba", size = 1989378, upload-time = "2025-11-04T13:39:27.92Z" },
+    { url = "https://files.pythonhosted.org/packages/76/92/d338652464c6c367e5608e4488201702cd1cbb0f33f7b6a85a60fe5f3720/pydantic_core-2.41.5-cp310-cp310-win_amd64.whl", hash = "sha256:62de39db01b8d593e45871af2af9e497295db8d73b085f6bfd0b18c83c70a8f9", size = 2013622, upload-time = "2025-11-04T13:39:29.848Z" },
+    { url = "https://files.pythonhosted.org/packages/e6/b0/1a2aa41e3b5a4ba11420aba2d091b2d17959c8d1519ece3627c371951e73/pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b5819cd790dbf0c5eb9f82c73c16b39a65dd6dd4d1439dcdea7816ec9adddab8", size = 2103351, upload-time = "2025-11-04T13:43:02.058Z" },
+    { url = "https://files.pythonhosted.org/packages/a4/ee/31b1f0020baaf6d091c87900ae05c6aeae101fa4e188e1613c80e4f1ea31/pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:5a4e67afbc95fa5c34cf27d9089bca7fcab4e51e57278d710320a70b956d1b9a", size = 1925363, upload-time = "2025-11-04T13:43:05.159Z" },
+    { url = "https://files.pythonhosted.org/packages/e1/89/ab8e86208467e467a80deaca4e434adac37b10a9d134cd2f99b28a01e483/pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ece5c59f0ce7d001e017643d8d24da587ea1f74f6993467d85ae8a5ef9d4f42b", size = 2135615, upload-time = "2025-11-04T13:43:08.116Z" },
+    { url = "https://files.pythonhosted.org/packages/99/0a/99a53d06dd0348b2008f2f30884b34719c323f16c3be4e6cc1203b74a91d/pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:16f80f7abe3351f8ea6858914ddc8c77e02578544a0ebc15b4c2e1a0e813b0b2", size = 2175369, upload-time = "2025-11-04T13:43:12.49Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/94/30ca3b73c6d485b9bb0bc66e611cff4a7138ff9736b7e66bcf0852151636/pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:33cb885e759a705b426baada1fe68cbb0a2e68e34c5d0d0289a364cf01709093", size = 2144218, upload-time = "2025-11-04T13:43:15.431Z" },
+    { url = "https://files.pythonhosted.org/packages/87/57/31b4f8e12680b739a91f472b5671294236b82586889ef764b5fbc6669238/pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:c8d8b4eb992936023be7dee581270af5c6e0697a8559895f527f5b7105ecd36a", size = 2329951, upload-time = "2025-11-04T13:43:18.062Z" },
+    { url = "https://files.pythonhosted.org/packages/7d/73/3c2c8edef77b8f7310e6fb012dbc4b8551386ed575b9eb6fb2506e28a7eb/pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:242a206cd0318f95cd21bdacff3fcc3aab23e79bba5cac3db5a841c9ef9c6963", size = 2318428, upload-time = "2025-11-04T13:43:20.679Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/02/8559b1f26ee0d502c74f9cca5c0d2fd97e967e083e006bbbb4e97f3a043a/pydantic_core-2.41.5-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d3a978c4f57a597908b7e697229d996d77a6d3c94901e9edee593adada95ce1a", size = 2147009, upload-time = "2025-11-04T13:43:23.286Z" },
+]
+
+[[package]]
+name = "pydub"
+version = "0.25.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fe/9a/e6bca0eed82db26562c73b5076539a4a08d3cffd19c3cc5913a3e61145fd/pydub-0.25.1.tar.gz", hash = "sha256:980a33ce9949cab2a569606b65674d748ecbca4f0796887fd6f46173a7b0d30f", size = 38326, upload-time = "2021-03-10T02:09:54.659Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a6/53/d78dc063216e62fc55f6b2eebb447f6a4b0a59f55c8406376f76bf959b08/pydub-0.25.1-py2.py3-none-any.whl", hash = "sha256:65617e33033874b59d87db603aa1ed450633288aefead953b30bded59cb599a6", size = 32327, upload-time = "2021-03-10T02:09:53.503Z" },
+]
+
+[[package]]
+name = "pygments"
+version = "2.19.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b0/77/a5b8c569bf593b0140bde72ea885a803b82086995367bf2037de0159d924/pygments-2.19.2.tar.gz", hash = "sha256:636cb2477cec7f8952536970bc533bc43743542f70392ae026374600add5b887", size = 4968631, upload-time = "2025-06-21T13:39:12.283Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c7/21/705964c7812476f378728bdf590ca4b771ec72385c533964653c68e86bdc/pygments-2.19.2-py3-none-any.whl", hash = "sha256:86540386c03d588bb81d44bc3928634ff26449851e99741617ecb9037ee5ec0b", size = 1225217, upload-time = "2025-06-21T13:39:07.939Z" },
+]
+
+[[package]]
+name = "pyopengl"
+version = "3.1.10"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6f/16/912b7225d56284859cd9a672827f18be43f8012f8b7b932bc4bd959a298e/pyopengl-3.1.10.tar.gz", hash = "sha256:c4a02d6866b54eb119c8e9b3fb04fa835a95ab802dd96607ab4cdb0012df8335", size = 1915580, upload-time = "2025-08-18T02:33:01.76Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/de/e4/1ba6f44e491c4eece978685230dde56b14d51a0365bc1b774ddaa94d14cd/pyopengl-3.1.10-py3-none-any.whl", hash = "sha256:794a943daced39300879e4e47bd94525280685f42dbb5a998d336cfff151d74f", size = 3194996, upload-time = "2025-08-18T02:32:59.902Z" },
+]
+
+[[package]]
+name = "pyparsing"
+version = "3.3.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f3/91/9c6ee907786a473bf81c5f53cf703ba0957b23ab84c264080fb5a450416f/pyparsing-3.3.2.tar.gz", hash = "sha256:c777f4d763f140633dcb6d8a3eda953bf7a214dc4eff598413c070bcdc117cbc", size = 6851574, upload-time = "2026-01-21T03:57:59.36Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/10/bd/c038d7cc38edc1aa5bf91ab8068b63d4308c66c4c8bb3cbba7dfbc049f9c/pyparsing-3.3.2-py3-none-any.whl", hash = "sha256:850ba148bd908d7e2411587e247a1e4f0327839c40e2e5e6d05a007ecc69911d", size = 122781, upload-time = "2026-01-21T03:57:55.912Z" },
+]
+
+[[package]]
+name = "pyrallis"
+version = "0.3.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyyaml" },
+    { name = "typing-inspect" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b3/ff/4d865cd0166b3e0b0b09b1f5399114fe27c836b74f45f580d61efecf28dd/pyrallis-0.3.1.tar.gz", hash = "sha256:ab7298f31c633d4858ec3b045a30e6cc9b9f7ab50f21ea93aa3fb28532ac4b6e", size = 33799, upload-time = "2022-03-14T08:23:25.805Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7e/21/896ce8acd58566f38ffa179b833fd52b4c285fd4e90842ad6b8362372639/pyrallis-0.3.1-py3-none-any.whl", hash = "sha256:632370b563486495f5f9e7caf86cddffab8351214a3bdc60ae0d23b261ebdb59", size = 33419, upload-time = "2022-03-14T08:23:24.193Z" },
+]
+
+[[package]]
+name = "pytest"
+version = "9.0.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "exceptiongroup" },
+    { name = "iniconfig" },
+    { name = "packaging" },
+    { name = "pluggy" },
+    { name = "pygments" },
+    { name = "tomli" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d1/db/7ef3487e0fb0049ddb5ce41d3a49c235bf9ad299b6a25d5780a89f19230f/pytest-9.0.2.tar.gz", hash = "sha256:75186651a92bd89611d1d9fc20f0b4345fd827c41ccd5c299a868a05d70edf11", size = 1568901, upload-time = "2025-12-06T21:30:51.014Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3b/ab/b3226f0bd7cdcf710fbede2b3548584366da3b19b5021e74f5bde2a8fa3f/pytest-9.0.2-py3-none-any.whl", hash = "sha256:711ffd45bf766d5264d487b917733b453d917afd2b0ad65223959f59089f875b", size = 374801, upload-time = "2025-12-06T21:30:49.154Z" },
+]
+
+[[package]]
+name = "python-dateutil"
+version = "2.9.0.post0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/66/c0/0c8b6ad9f17a802ee498c46e004a0eb49bc148f2fd230864601a86dcf6db/python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3", size = 342432, upload-time = "2024-03-01T18:36:20.211Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" },
+]
+
+[[package]]
+name = "python-discovery"
+version = "1.1.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "filelock" },
+    { name = "platformdirs" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ec/67/09765eacf4e44413c4f8943ba5a317fcb9c7b447c3b8b0b7fce7e3090b0b/python_discovery-1.1.1.tar.gz", hash = "sha256:584c08b141c5b7029f206b4e8b78b1a1764b22121e21519b89dec56936e95b0a", size = 56016, upload-time = "2026-03-07T00:00:56.354Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/75/0f/2bf7e3b5a4a65f623cb820feb5793e243fad58ae561015ee15a6152f67a2/python_discovery-1.1.1-py3-none-any.whl", hash = "sha256:69f11073fa2392251e405d4e847d60ffffd25fd762a0dc4d1a7d6b9c3f79f1a3", size = 30732, upload-time = "2026-03-07T00:00:55.143Z" },
+]
+
+[[package]]
+name = "python-dotenv"
+version = "1.2.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/82/ed/0301aeeac3e5353ef3d94b6ec08bbcabd04a72018415dcb29e588514bba8/python_dotenv-1.2.2.tar.gz", hash = "sha256:2c371a91fbd7ba082c2c1dc1f8bf89ca22564a087c2c287cd9b662adde799cf3", size = 50135, upload-time = "2026-03-01T16:00:26.196Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0b/d7/1959b9648791274998a9c3526f6d0ec8fd2233e4d4acce81bbae76b44b2a/python_dotenv-1.2.2-py3-none-any.whl", hash = "sha256:1d8214789a24de455a8b8bd8ae6fe3c6b69a5e3d64aa8a8e5d68e694bbcb285a", size = 22101, upload-time = "2026-03-01T16:00:25.09Z" },
+]
+
+[[package]]
+name = "python-multipart"
+version = "0.0.22"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/94/01/979e98d542a70714b0cb2b6728ed0b7c46792b695e3eaec3e20711271ca3/python_multipart-0.0.22.tar.gz", hash = "sha256:7340bef99a7e0032613f56dc36027b959fd3b30a787ed62d310e951f7c3a3a58", size = 37612, upload-time = "2026-01-25T10:15:56.219Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1b/d0/397f9626e711ff749a95d96b7af99b9c566a9bb5129b8e4c10fc4d100304/python_multipart-0.0.22-py3-none-any.whl", hash = "sha256:2b2cd894c83d21bf49d702499531c7bafd057d730c201782048f7945d82de155", size = 24579, upload-time = "2026-01-25T10:15:54.811Z" },
+]
+
+[[package]]
+name = "pytz"
+version = "2026.1.post1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/56/db/b8721d71d945e6a8ac63c0fc900b2067181dbb50805958d4d4661cf7d277/pytz-2026.1.post1.tar.gz", hash = "sha256:3378dde6a0c3d26719182142c56e60c7f9af7e968076f31aae569d72a0358ee1", size = 321088, upload-time = "2026-03-03T07:47:50.683Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/10/99/781fe0c827be2742bcc775efefccb3b048a3a9c6ce9aec0cbf4a101677e5/pytz-2026.1.post1-py2.py3-none-any.whl", hash = "sha256:f2fd16142fda348286a75e1a524be810bb05d444e5a081f37f7affc635035f7a", size = 510489, upload-time = "2026-03-03T07:47:49.167Z" },
+]
+
+[[package]]
+name = "pyyaml"
+version = "6.0.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/05/8e/961c0007c59b8dd7729d542c61a4d537767a59645b82a0b521206e1e25c2/pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f", size = 130960, upload-time = "2025-09-25T21:33:16.546Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f4/a0/39350dd17dd6d6c6507025c0e53aef67a9293a6d37d3511f23ea510d5800/pyyaml-6.0.3-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:214ed4befebe12df36bcc8bc2b64b396ca31be9304b8f59e25c11cf94a4c033b", size = 184227, upload-time = "2025-09-25T21:31:46.04Z" },
+    { url = "https://files.pythonhosted.org/packages/05/14/52d505b5c59ce73244f59c7a50ecf47093ce4765f116cdb98286a71eeca2/pyyaml-6.0.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:02ea2dfa234451bbb8772601d7b8e426c2bfa197136796224e50e35a78777956", size = 174019, upload-time = "2025-09-25T21:31:47.706Z" },
+    { url = "https://files.pythonhosted.org/packages/43/f7/0e6a5ae5599c838c696adb4e6330a59f463265bfa1e116cfd1fbb0abaaae/pyyaml-6.0.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b30236e45cf30d2b8e7b3e85881719e98507abed1011bf463a8fa23e9c3e98a8", size = 740646, upload-time = "2025-09-25T21:31:49.21Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/3a/61b9db1d28f00f8fd0ae760459a5c4bf1b941baf714e207b6eb0657d2578/pyyaml-6.0.3-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:66291b10affd76d76f54fad28e22e51719ef9ba22b29e1d7d03d6777a9174198", size = 840793, upload-time = "2025-09-25T21:31:50.735Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/1e/7acc4f0e74c4b3d9531e24739e0ab832a5edf40e64fbae1a9c01941cabd7/pyyaml-6.0.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9c7708761fccb9397fe64bbc0395abcae8c4bf7b0eac081e12b809bf47700d0b", size = 770293, upload-time = "2025-09-25T21:31:51.828Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/ef/abd085f06853af0cd59fa5f913d61a8eab65d7639ff2a658d18a25d6a89d/pyyaml-6.0.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:418cf3f2111bc80e0933b2cd8cd04f286338bb88bdc7bc8e6dd775ebde60b5e0", size = 732872, upload-time = "2025-09-25T21:31:53.282Z" },
+    { url = "https://files.pythonhosted.org/packages/1f/15/2bc9c8faf6450a8b3c9fc5448ed869c599c0a74ba2669772b1f3a0040180/pyyaml-6.0.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5e0b74767e5f8c593e8c9b5912019159ed0533c70051e9cce3e8b6aa699fcd69", size = 758828, upload-time = "2025-09-25T21:31:54.807Z" },
+    { url = "https://files.pythonhosted.org/packages/a3/00/531e92e88c00f4333ce359e50c19b8d1de9fe8d581b1534e35ccfbc5f393/pyyaml-6.0.3-cp310-cp310-win32.whl", hash = "sha256:28c8d926f98f432f88adc23edf2e6d4921ac26fb084b028c733d01868d19007e", size = 142415, upload-time = "2025-09-25T21:31:55.885Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/fa/926c003379b19fca39dd4634818b00dec6c62d87faf628d1394e137354d4/pyyaml-6.0.3-cp310-cp310-win_amd64.whl", hash = "sha256:bdb2c67c6c1390b63c6ff89f210c8fd09d9a1217a465701eac7316313c915e4c", size = 158561, upload-time = "2025-09-25T21:31:57.406Z" },
+]
+
+[[package]]
+name = "pyzmq"
+version = "27.1.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cffi", marker = "implementation_name == 'pypy' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/04/0b/3c9baedbdf613ecaa7aa07027780b8867f57b6293b6ee50de316c9f3222b/pyzmq-27.1.0.tar.gz", hash = "sha256:ac0765e3d44455adb6ddbf4417dcce460fc40a05978c08efdf2948072f6db540", size = 281750, upload-time = "2025-09-08T23:10:18.157Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/67/b9/52aa9ec2867528b54f1e60846728d8b4d84726630874fee3a91e66c7df81/pyzmq-27.1.0-cp310-cp310-macosx_10_15_universal2.whl", hash = "sha256:508e23ec9bc44c0005c4946ea013d9317ae00ac67778bd47519fdf5a0e930ff4", size = 1329850, upload-time = "2025-09-08T23:07:26.274Z" },
+    { url = "https://files.pythonhosted.org/packages/99/64/5653e7b7425b169f994835a2b2abf9486264401fdef18df91ddae47ce2cc/pyzmq-27.1.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:507b6f430bdcf0ee48c0d30e734ea89ce5567fd7b8a0f0044a369c176aa44556", size = 906380, upload-time = "2025-09-08T23:07:29.78Z" },
+    { url = "https://files.pythonhosted.org/packages/73/78/7d713284dbe022f6440e391bd1f3c48d9185673878034cfb3939cdf333b2/pyzmq-27.1.0-cp310-cp310-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:bf7b38f9fd7b81cb6d9391b2946382c8237fd814075c6aa9c3b746d53076023b", size = 666421, upload-time = "2025-09-08T23:07:31.263Z" },
+    { url = "https://files.pythonhosted.org/packages/30/76/8f099f9d6482450428b17c4d6b241281af7ce6a9de8149ca8c1c649f6792/pyzmq-27.1.0-cp310-cp310-manylinux_2_26_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:03ff0b279b40d687691a6217c12242ee71f0fba28bf8626ff50e3ef0f4410e1e", size = 854149, upload-time = "2025-09-08T23:07:33.17Z" },
+    { url = "https://files.pythonhosted.org/packages/59/f0/37fbfff06c68016019043897e4c969ceab18bde46cd2aca89821fcf4fb2e/pyzmq-27.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:677e744fee605753eac48198b15a2124016c009a11056f93807000ab11ce6526", size = 1655070, upload-time = "2025-09-08T23:07:35.205Z" },
+    { url = "https://files.pythonhosted.org/packages/47/14/7254be73f7a8edc3587609554fcaa7bfd30649bf89cd260e4487ca70fdaa/pyzmq-27.1.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:dd2fec2b13137416a1c5648b7009499bcc8fea78154cd888855fa32514f3dad1", size = 2033441, upload-time = "2025-09-08T23:07:37.432Z" },
+    { url = "https://files.pythonhosted.org/packages/22/dc/49f2be26c6f86f347e796a4d99b19167fc94503f0af3fd010ad262158822/pyzmq-27.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:08e90bb4b57603b84eab1d0ca05b3bbb10f60c1839dc471fc1c9e1507bef3386", size = 1891529, upload-time = "2025-09-08T23:07:39.047Z" },
+    { url = "https://files.pythonhosted.org/packages/a3/3e/154fb963ae25be70c0064ce97776c937ecc7d8b0259f22858154a9999769/pyzmq-27.1.0-cp310-cp310-win32.whl", hash = "sha256:a5b42d7a0658b515319148875fcb782bbf118dd41c671b62dae33666c2213bda", size = 567276, upload-time = "2025-09-08T23:07:40.695Z" },
+    { url = "https://files.pythonhosted.org/packages/62/b2/f4ab56c8c595abcb26b2be5fd9fa9e6899c1e5ad54964e93ae8bb35482be/pyzmq-27.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:c0bb87227430ee3aefcc0ade2088100e528d5d3298a0a715a64f3d04c60ba02f", size = 632208, upload-time = "2025-09-08T23:07:42.298Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/e3/be2cc7ab8332bdac0522fdb64c17b1b6241a795bee02e0196636ec5beb79/pyzmq-27.1.0-cp310-cp310-win_arm64.whl", hash = "sha256:9a916f76c2ab8d045b19f2286851a38e9ac94ea91faf65bd64735924522a8b32", size = 559766, upload-time = "2025-09-08T23:07:43.869Z" },
+    { url = "https://files.pythonhosted.org/packages/92/e7/038aab64a946d535901103da16b953c8c9cc9c961dadcbf3609ed6428d23/pyzmq-27.1.0-cp312-abi3-macosx_10_15_universal2.whl", hash = "sha256:452631b640340c928fa343801b0d07eb0c3789a5ffa843f6e1a9cee0ba4eb4fc", size = 1306279, upload-time = "2025-09-08T23:08:03.807Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/5e/c3c49fdd0f535ef45eefcc16934648e9e59dace4a37ee88fc53f6cd8e641/pyzmq-27.1.0-cp312-abi3-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:1c179799b118e554b66da67d88ed66cd37a169f1f23b5d9f0a231b4e8d44a113", size = 895645, upload-time = "2025-09-08T23:08:05.301Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/e5/b0b2504cb4e903a74dcf1ebae157f9e20ebb6ea76095f6cfffea28c42ecd/pyzmq-27.1.0-cp312-abi3-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3837439b7f99e60312f0c926a6ad437b067356dc2bc2ec96eb395fd0fe804233", size = 652574, upload-time = "2025-09-08T23:08:06.828Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/9b/c108cdb55560eaf253f0cbdb61b29971e9fb34d9c3499b0e96e4e60ed8a5/pyzmq-27.1.0-cp312-abi3-manylinux_2_26_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:43ad9a73e3da1fab5b0e7e13402f0b2fb934ae1c876c51d0afff0e7c052eca31", size = 840995, upload-time = "2025-09-08T23:08:08.396Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/bb/b79798ca177b9eb0825b4c9998c6af8cd2a7f15a6a1a4272c1d1a21d382f/pyzmq-27.1.0-cp312-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:0de3028d69d4cdc475bfe47a6128eb38d8bc0e8f4d69646adfbcd840facbac28", size = 1642070, upload-time = "2025-09-08T23:08:09.989Z" },
+    { url = "https://files.pythonhosted.org/packages/9c/80/2df2e7977c4ede24c79ae39dcef3899bfc5f34d1ca7a5b24f182c9b7a9ca/pyzmq-27.1.0-cp312-abi3-musllinux_1_2_i686.whl", hash = "sha256:cf44a7763aea9298c0aa7dbf859f87ed7012de8bda0f3977b6fb1d96745df856", size = 2021121, upload-time = "2025-09-08T23:08:11.907Z" },
+    { url = "https://files.pythonhosted.org/packages/46/bd/2d45ad24f5f5ae7e8d01525eb76786fa7557136555cac7d929880519e33a/pyzmq-27.1.0-cp312-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:f30f395a9e6fbca195400ce833c731e7b64c3919aa481af4d88c3759e0cb7496", size = 1878550, upload-time = "2025-09-08T23:08:13.513Z" },
+    { url = "https://files.pythonhosted.org/packages/e6/2f/104c0a3c778d7c2ab8190e9db4f62f0b6957b53c9d87db77c284b69f33ea/pyzmq-27.1.0-cp312-abi3-win32.whl", hash = "sha256:250e5436a4ba13885494412b3da5d518cd0d3a278a1ae640e113c073a5f88edd", size = 559184, upload-time = "2025-09-08T23:08:15.163Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/7f/a21b20d577e4100c6a41795842028235998a643b1ad406a6d4163ea8f53e/pyzmq-27.1.0-cp312-abi3-win_amd64.whl", hash = "sha256:9ce490cf1d2ca2ad84733aa1d69ce6855372cb5ce9223802450c9b2a7cba0ccf", size = 619480, upload-time = "2025-09-08T23:08:17.192Z" },
+    { url = "https://files.pythonhosted.org/packages/78/c2/c012beae5f76b72f007a9e91ee9401cb88c51d0f83c6257a03e785c81cc2/pyzmq-27.1.0-cp312-abi3-win_arm64.whl", hash = "sha256:75a2f36223f0d535a0c919e23615fc85a1e23b71f40c7eb43d7b1dedb4d8f15f", size = 552993, upload-time = "2025-09-08T23:08:18.926Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/81/a65e71c1552f74dec9dff91d95bafb6e0d33338a8dfefbc88aa562a20c92/pyzmq-27.1.0-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:c17e03cbc9312bee223864f1a2b13a99522e0dc9f7c5df0177cd45210ac286e6", size = 836266, upload-time = "2025-09-08T23:09:40.048Z" },
+    { url = "https://files.pythonhosted.org/packages/58/ed/0202ca350f4f2b69faa95c6d931e3c05c3a397c184cacb84cb4f8f42f287/pyzmq-27.1.0-pp310-pypy310_pp73-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:f328d01128373cb6763823b2b4e7f73bdf767834268c565151eacb3b7a392f90", size = 800206, upload-time = "2025-09-08T23:09:41.902Z" },
+    { url = "https://files.pythonhosted.org/packages/47/42/1ff831fa87fe8f0a840ddb399054ca0009605d820e2b44ea43114f5459f4/pyzmq-27.1.0-pp310-pypy310_pp73-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9c1790386614232e1b3a40a958454bdd42c6d1811837b15ddbb052a032a43f62", size = 567747, upload-time = "2025-09-08T23:09:43.741Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/db/5c4d6807434751e3f21231bee98109aa57b9b9b55e058e450d0aef59b70f/pyzmq-27.1.0-pp310-pypy310_pp73-manylinux_2_26_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:448f9cb54eb0cee4732b46584f2710c8bc178b0e5371d9e4fc8125201e413a74", size = 747371, upload-time = "2025-09-08T23:09:45.575Z" },
+    { url = "https://files.pythonhosted.org/packages/26/af/78ce193dbf03567eb8c0dc30e3df2b9e56f12a670bf7eb20f9fb532c7e8a/pyzmq-27.1.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:05b12f2d32112bf8c95ef2e74ec4f1d4beb01f8b5e703b38537f8849f92cb9ba", size = 544862, upload-time = "2025-09-08T23:09:47.448Z" },
+]
+
+[[package]]
+name = "qwen-vl-utils"
+version = "0.0.14"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "av" },
+    { name = "packaging" },
+    { name = "pillow" },
+    { name = "requests" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b6/b1/ad4fc2260a3badd278b38d642f3b987412f1f6682f0ef2b31b0572d5caa8/qwen_vl_utils-0.0.14.tar.gz", hash = "sha256:9c7cad5ae803b3a10f8bb7194deb12aeacdd032f92f4224e880c73587a7346ad", size = 8453, upload-time = "2025-09-23T09:38:57.532Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c4/43/80f67e0336cb2fc725f8e06f7fe35c1d0fe946f4d2b8b2175e797e07349e/qwen_vl_utils-0.0.14-py3-none-any.whl", hash = "sha256:5e28657bfd031e56bd447c5901b58ddfc3835285ed100f4c56580e0ade054e96", size = 8120, upload-time = "2025-09-23T09:38:56.297Z" },
+]
+
+[package.optional-dependencies]
+decord = [
+    { name = "decord" },
+]
+
+[[package]]
+name = "rdkit"
+version = "2025.9.6"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "pillow" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/28/cd/864924bb0606c8d90869f1f460f7a5b6911e98d67d56ad404b5eab1852b3/rdkit-2025.9.6-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ce558b4f8e63a52ec932b59e867a8e1b2e3cde304e17b652cd54e4f73a6f5607", size = 29513135, upload-time = "2026-03-02T18:08:24.822Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/9d/3d4d367e5ad6c995d043adc6382ea82480b1b8b4aba2c6a730ddd5d21abf/rdkit-2025.9.6-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:5aedf8a713998553e17b75a09990561a7824a542c85b48987d9bc88dbd1329f0", size = 35244087, upload-time = "2026-03-02T18:08:29.612Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/ba/761a33df54a935a20d5ee85e6a9c6334f509a2d4e1b1fcb7d382ab731835/rdkit-2025.9.6-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:6d6746cea16ce048259bf463a5beef92cf130fd0f8f62a999d4eb28384e7831a", size = 36730482, upload-time = "2026-03-02T18:08:33.683Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/1d/1851932f47b08bc203c373e4102b552ed031eaf596444a5a98eff5f67a83/rdkit-2025.9.6-cp310-cp310-win_amd64.whl", hash = "sha256:a961a99f91f1d4366d8eaf53e0648c36e3f6dd96a040be9fc6b90b4a7133dae8", size = 24297586, upload-time = "2026-03-02T18:08:37.728Z" },
+]
+
+[[package]]
+name = "regex"
+version = "2026.2.28"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/8b/71/41455aa99a5a5ac1eaf311f5d8efd9ce6433c03ac1e0962de163350d0d97/regex-2026.2.28.tar.gz", hash = "sha256:a729e47d418ea11d03469f321aaf67cdee8954cde3ff2cf8403ab87951ad10f2", size = 415184, upload-time = "2026-02-28T02:19:42.792Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/70/b8/845a927e078f5e5cc55d29f57becbfde0003d52806544531ab3f2da4503c/regex-2026.2.28-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:fc48c500838be6882b32748f60a15229d2dea96e59ef341eaa96ec83538f498d", size = 488461, upload-time = "2026-02-28T02:15:48.405Z" },
+    { url = "https://files.pythonhosted.org/packages/32/f9/8a0034716684e38a729210ded6222249f29978b24b684f448162ef21f204/regex-2026.2.28-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2afa673660928d0b63d84353c6c08a8a476ddfc4a47e11742949d182e6863ce8", size = 290774, upload-time = "2026-02-28T02:15:51.738Z" },
+    { url = "https://files.pythonhosted.org/packages/a6/ba/b27feefffbb199528dd32667cd172ed484d9c197618c575f01217fbe6103/regex-2026.2.28-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:7ab218076eb0944549e7fe74cf0e2b83a82edb27e81cc87411f76240865e04d5", size = 288737, upload-time = "2026-02-28T02:15:53.534Z" },
+    { url = "https://files.pythonhosted.org/packages/18/c5/65379448ca3cbfe774fcc33774dc8295b1ee97dc3237ae3d3c7b27423c9d/regex-2026.2.28-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:94d63db12e45a9b9f064bfe4800cefefc7e5f182052e4c1b774d46a40ab1d9bb", size = 782675, upload-time = "2026-02-28T02:15:55.488Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/30/6fa55bef48090f900fbd4649333791fc3e6467380b9e775e741beeb3231f/regex-2026.2.28-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:195237dc327858a7721bf8b0bbbef797554bc13563c3591e91cd0767bacbe359", size = 850514, upload-time = "2026-02-28T02:15:57.509Z" },
+    { url = "https://files.pythonhosted.org/packages/a9/28/9ca180fb3787a54150209754ac06a42409913571fa94994f340b3bba4e1e/regex-2026.2.28-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b387a0d092dac157fb026d737dde35ff3e49ef27f285343e7c6401851239df27", size = 896612, upload-time = "2026-02-28T02:15:59.682Z" },
+    { url = "https://files.pythonhosted.org/packages/46/b5/f30d7d3936d6deecc3ea7bea4f7d3c5ee5124e7c8de372226e436b330a55/regex-2026.2.28-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3935174fa4d9f70525a4367aaff3cb8bc0548129d114260c29d9dfa4a5b41692", size = 791691, upload-time = "2026-02-28T02:16:01.752Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/34/96631bcf446a56ba0b2a7f684358a76855dfe315b7c2f89b35388494ede0/regex-2026.2.28-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2b2b23587b26496ff5fd40df4278becdf386813ec00dc3533fa43a4cf0e2ad3c", size = 783111, upload-time = "2026-02-28T02:16:03.651Z" },
+    { url = "https://files.pythonhosted.org/packages/39/54/f95cb7a85fe284d41cd2f3625e0f2ae30172b55dfd2af1d9b4eaef6259d7/regex-2026.2.28-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3b24bd7e9d85dc7c6a8bd2aa14ecd234274a0248335a02adeb25448aecdd420d", size = 767512, upload-time = "2026-02-28T02:16:05.616Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/af/a650f64a79c02a97f73f64d4e7fc4cc1984e64affab14075e7c1f9a2db34/regex-2026.2.28-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:bd477d5f79920338107f04aa645f094032d9e3030cc55be581df3d1ef61aa318", size = 773920, upload-time = "2026-02-28T02:16:08.325Z" },
+    { url = "https://files.pythonhosted.org/packages/72/f8/3f9c2c2af37aedb3f5a1e7227f81bea065028785260d9cacc488e43e6997/regex-2026.2.28-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:b49eb78048c6354f49e91e4b77da21257fecb92256b6d599ae44403cab30b05b", size = 846681, upload-time = "2026-02-28T02:16:10.381Z" },
+    { url = "https://files.pythonhosted.org/packages/54/12/8db04a334571359f4d127d8f89550917ec6561a2fddfd69cd91402b47482/regex-2026.2.28-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:a25c7701e4f7a70021db9aaf4a4a0a67033c6318752146e03d1b94d32006217e", size = 755565, upload-time = "2026-02-28T02:16:11.972Z" },
+    { url = "https://files.pythonhosted.org/packages/da/bc/91c22f384d79324121b134c267a86ca90d11f8016aafb1dc5bee05890ee3/regex-2026.2.28-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:9dd450db6458387167e033cfa80887a34c99c81d26da1bf8b0b41bf8c9cac88e", size = 835789, upload-time = "2026-02-28T02:16:14.036Z" },
+    { url = "https://files.pythonhosted.org/packages/46/a7/4cc94fd3af01dcfdf5a9ed75c8e15fd80fcd62cc46da7592b1749e9c35db/regex-2026.2.28-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:2954379dd20752e82d22accf3ff465311cbb2bac6c1f92c4afd400e1757f7451", size = 780094, upload-time = "2026-02-28T02:16:15.468Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/21/e5a38f420af3c77cab4a65f0c3a55ec02ac9babf04479cfd282d356988a6/regex-2026.2.28-cp310-cp310-win32.whl", hash = "sha256:1f8b17be5c27a684ea6759983c13506bd77bfc7c0347dff41b18ce5ddd2ee09a", size = 266025, upload-time = "2026-02-28T02:16:16.828Z" },
+    { url = "https://files.pythonhosted.org/packages/4d/0a/205c4c1466a36e04d90afcd01d8908bac327673050c7fe316b2416d99d3d/regex-2026.2.28-cp310-cp310-win_amd64.whl", hash = "sha256:dd8847c4978bc3c7e6c826fb745f5570e518b8459ac2892151ce6627c7bc00d5", size = 277965, upload-time = "2026-02-28T02:16:18.752Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/4d/29b58172f954b6ec2c5ed28529a65e9026ab96b4b7016bcd3858f1c31d3c/regex-2026.2.28-cp310-cp310-win_arm64.whl", hash = "sha256:73cdcdbba8028167ea81490c7f45280113e41db2c7afb65a276f4711fa3bcbff", size = 270336, upload-time = "2026-02-28T02:16:20.735Z" },
+]
+
+[[package]]
+name = "requests"
+version = "2.32.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi" },
+    { name = "charset-normalizer" },
+    { name = "idna" },
+    { name = "urllib3" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c9/74/b3ff8e6c8446842c3f5c837e9c3dfcfe2018ea6ecef224c710c85ef728f4/requests-2.32.5.tar.gz", hash = "sha256:dbba0bac56e100853db0ea71b82b4dfd5fe2bf6d3754a8893c3af500cec7d7cf", size = 134517, upload-time = "2025-08-18T20:46:02.573Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1e/db/4254e3eabe8020b458f1a747140d32277ec7a271daf1d235b70dc0b4e6e3/requests-2.32.5-py3-none-any.whl", hash = "sha256:2462f94637a34fd532264295e186976db0f5d453d1cdd31473c85a6a161affb6", size = 64738, upload-time = "2025-08-18T20:46:00.542Z" },
+]
+
+[[package]]
+name = "rich"
+version = "14.3.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "markdown-it-py" },
+    { name = "pygments" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b3/c6/f3b320c27991c46f43ee9d856302c70dc2d0fb2dba4842ff739d5f46b393/rich-14.3.3.tar.gz", hash = "sha256:b8daa0b9e4eef54dd8cf7c86c03713f53241884e814f4e2f5fb342fe520f639b", size = 230582, upload-time = "2026-02-19T17:23:12.474Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/25/b208c5683343959b670dc001595f2f3737e051da617f66c31f7c4fa93abc/rich-14.3.3-py3-none-any.whl", hash = "sha256:793431c1f8619afa7d3b52b2cdec859562b950ea0d4b6b505397612db8d5362d", size = 310458, upload-time = "2026-02-19T17:23:13.732Z" },
+]
+
+[[package]]
+name = "robometer"
+version = "0.0.1"
+source = { editable = "." }
+dependencies = [
+    { name = "accelerate" },
+    { name = "bitsandbytes" },
+    { name = "codetiming" },
+    { name = "datasets" },
+    { name = "decord" },
+    { name = "einops" },
+    { name = "evaluate" },
+    { name = "fastapi" },
+    { name = "ffmpeg" },
+    { name = "google-generativeai" },
+    { name = "gradio" },
+    { name = "gymnasium" },
+    { name = "h5py" },
+    { name = "hf-transfer" },
+    { name = "huggingface-hub", extra = ["hf-xet"] },
+    { name = "hydra-core" },
+    { name = "imageio" },
+    { name = "ipdb" },
+    { name = "loguru" },
+    { name = "matplotlib" },
+    { name = "metaworld" },
+    { name = "moviepy" },
+    { name = "num2words" },
+    { name = "numpy" },
+    { name = "omegaconf" },
+    { name = "opencv-contrib-python-headless" },
+    { name = "opencv-python-headless" },
+    { name = "peft", version = "0.15.2", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "peft", version = "0.18.1", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "pillow" },
+    { name = "protobuf" },
+    { name = "pyrallis" },
+    { name = "pyyaml" },
+    { name = "qwen-vl-utils", extra = ["decord"] },
+    { name = "requests" },
+    { name = "rich" },
+    { name = "robosuite" },
+    { name = "safetensors" },
+    { name = "scipy" },
+    { name = "seaborn" },
+    { name = "sentence-transformers" },
+    { name = "soundfile" },
+    { name = "tensorboard" },
+    { name = "tensorflow" },
+    { name = "tensorflow-datasets" },
+    { name = "termcolor" },
+    { name = "tf-keras" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchcodec" },
+    { name = "torchvision", version = "0.23.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm" },
+    { name = "unsloth", version = "2025.10.8", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "unsloth", version = "2026.3.4", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "uvicorn" },
+    { name = "wandb", extra = ["media"] },
+    { name = "xformers" },
+]
+
+[package.optional-dependencies]
+dev = [
+    { name = "ipykernel" },
+    { name = "ipywidgets" },
+    { name = "pre-commit" },
+    { name = "pytest" },
+    { name = "ruff" },
+]
+quantization = [
+    { name = "bitsandbytes" },
+]
+robometer = [
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" } },
+    { name = "trl", version = "0.20.0", source = { registry = "https://pypi.org/simple" } },
+]
+vlac = [
+    { name = "evo-vlac" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" } },
+    { name = "trl", version = "0.16.1", source = { registry = "https://pypi.org/simple" } },
+]
+
+[package.metadata]
+requires-dist = [
+    { name = "accelerate", specifier = ">=1.9.0" },
+    { name = "bitsandbytes" },
+    { name = "bitsandbytes", marker = "extra == 'quantization'", specifier = ">=0.46.1" },
+    { name = "codetiming", specifier = ">=1.4.0" },
+    { name = "datasets", specifier = "==4.1.1" },
+    { name = "decord", specifier = ">=0.6.0" },
+    { name = "einops", specifier = ">=0.8.0" },
+    { name = "evaluate", specifier = ">=0.4.3" },
+    { name = "evo-vlac", marker = "extra == 'vlac'", git = "https://github.com/InternRobotics/VLAC.git" },
+    { name = "fastapi", specifier = ">=0.116.1" },
+    { name = "ffmpeg" },
+    { name = "google-generativeai" },
+    { name = "gradio", specifier = ">=4.0.0" },
+    { name = "gymnasium", specifier = ">=1.2.2" },
+    { name = "h5py", specifier = ">=3.10.0" },
+    { name = "hf-transfer", specifier = ">=0.1.0" },
+    { name = "huggingface-hub", extras = ["hf-xet"], specifier = ">0.35" },
+    { name = "hydra-core" },
+    { name = "imageio", specifier = ">=2.37.0" },
+    { name = "ipdb", specifier = ">=0.13.13" },
+    { name = "ipykernel", marker = "extra == 'dev'" },
+    { name = "ipywidgets", marker = "extra == 'dev'" },
+    { name = "loguru", specifier = ">=0.7.3" },
+    { name = "matplotlib", specifier = ">=3.5.0" },
+    { name = "metaworld", specifier = ">=3.0.0" },
+    { name = "moviepy" },
+    { name = "num2words" },
+    { name = "numpy", specifier = ">=1.21.0" },
+    { name = "omegaconf" },
+    { name = "omegaconf", specifier = ">=2.3.0" },
+    { name = "opencv-contrib-python-headless", specifier = ">=4.8.0" },
+    { name = "opencv-python-headless", specifier = ">=4.8.0" },
+    { name = "peft" },
+    { name = "pillow", specifier = ">=9.0.0" },
+    { name = "pre-commit", marker = "extra == 'dev'", specifier = ">=4.0.0" },
+    { name = "protobuf", specifier = "<3.21" },
+    { name = "pyrallis", specifier = ">=0.3.0" },
+    { name = "pytest", marker = "extra == 'dev'", specifier = ">=8.0.0" },
+    { name = "pyyaml", specifier = ">=6.0.0" },
+    { name = "qwen-vl-utils", extras = ["decord"], specifier = "==0.0.14" },
+    { name = "requests" },
+    { name = "requests", specifier = ">=2.32.4" },
+    { name = "rich", specifier = ">=14.0.0" },
+    { name = "robosuite", specifier = "==1.4.0" },
+    { name = "ruff", marker = "extra == 'dev'", specifier = ">=0.8.0" },
+    { name = "safetensors", specifier = ">=0.6.0rc0" },
+    { name = "scipy", specifier = ">=1.15.3" },
+    { name = "seaborn" },
+    { name = "sentence-transformers", specifier = ">=2.0.0" },
+    { name = "soundfile", specifier = ">=0.13.1" },
+    { name = "tensorboard" },
+    { name = "tensorflow", specifier = "==2.19" },
+    { name = "tensorflow-datasets", specifier = "==4.9.2" },
+    { name = "termcolor", specifier = ">=3.1.0" },
+    { name = "tf-keras", specifier = ">=2.19.0" },
+    { name = "torch", marker = "sys_platform != 'linux' and sys_platform != 'win32'", specifier = "==2.8.0" },
+    { name = "torch", marker = "sys_platform == 'linux' or sys_platform == 'win32'", specifier = "==2.8.0", index = "https://download.pytorch.org/whl/cu128" },
+    { name = "torchcodec" },
+    { name = "torchvision", marker = "sys_platform != 'linux' and sys_platform != 'win32'" },
+    { name = "torchvision", marker = "sys_platform == 'linux' or sys_platform == 'win32'", index = "https://download.pytorch.org/whl/cu128" },
+    { name = "tqdm", specifier = ">=4.67.1" },
+    { name = "transformers", marker = "extra == 'robometer'", specifier = ">=4.57" },
+    { name = "transformers", marker = "extra == 'vlac'", specifier = "<4.52" },
+    { name = "trl", marker = "extra == 'robometer'", specifier = "==0.20.0" },
+    { name = "trl", marker = "extra == 'vlac'", specifier = ">=0.13,<0.17" },
+    { name = "unsloth", specifier = ">=2025.10" },
+    { name = "uvicorn", specifier = ">=0.35.0" },
+    { name = "wandb", specifier = ">=0.20.1" },
+    { name = "wandb", extras = ["media"] },
+    { name = "xformers", specifier = "==0.0.32.post2" },
+]
+provides-extras = ["dev", "quantization", "robometer", "vlac"]
+
+[[package]]
+name = "robosuite"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "mujoco" },
+    { name = "numba" },
+    { name = "numpy" },
+    { name = "opencv-python" },
+    { name = "pillow" },
+    { name = "scipy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/25/a1/9dd07a9a5e09c6aa032faf531da985808b34437cbf6c8f358fe8f7c47118/robosuite-1.4.0.tar.gz", hash = "sha256:a8a6233d7458dbd91bf00a86cab15aa1c178bd9d1b28d515db2cf3d152cb48e6", size = 192182147, upload-time = "2022-12-01T07:31:55.791Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2e/08/fe231064caaf2766d47818ca12fd8acf10dc4e762a33dabc6293e83bfead/robosuite-1.4.0-py3-none-any.whl", hash = "sha256:aba065e7b36745738cede259457b2cb349427f3608728d867ef3a2034cb62994", size = 193477875, upload-time = "2022-12-01T07:28:53.457Z" },
+]
+
+[[package]]
+name = "rouge"
+version = "1.0.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/db/e4/3420a1ab1e82a280fb6107f7ae99e88eb12383c978fe573c0c64d0327d6b/rouge-1.0.1.tar.gz", hash = "sha256:12b48346ca47d6bcf3c45061f315452b9ccec0620ee895ec85b7efc3d54aae34", size = 14292, upload-time = "2021-07-20T08:45:56.53Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/7c/650ae86f92460e9e8ef969cc5008b24798dcf56a9a8947d04c78f550b3f5/rouge-1.0.1-py3-none-any.whl", hash = "sha256:28d118536e8c774dc47d1d15ec266479b4dd0914c4672ce117d4002789bdc644", size = 13725, upload-time = "2021-07-20T08:45:54.605Z" },
+]
+
+[[package]]
+name = "rsa"
+version = "4.9.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyasn1" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/da/8a/22b7beea3ee0d44b1916c0c1cb0ee3af23b700b6da9f04991899d0c555d4/rsa-4.9.1.tar.gz", hash = "sha256:e7bdbfdb5497da4c07dfd35530e1a902659db6ff241e39d9953cad06ebd0ae75", size = 29034, upload-time = "2025-04-16T09:51:18.218Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/64/8d/0133e4eb4beed9e425d9a98ed6e081a55d195481b7632472be1af08d2f6b/rsa-4.9.1-py3-none-any.whl", hash = "sha256:68635866661c6836b8d39430f97a996acbd61bfa49406748ea243539fe239762", size = 34696, upload-time = "2025-04-16T09:51:17.142Z" },
+]
+
+[[package]]
+name = "ruff"
+version = "0.15.5"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/77/9b/840e0039e65fcf12758adf684d2289024d6140cde9268cc59887dc55189c/ruff-0.15.5.tar.gz", hash = "sha256:7c3601d3b6d76dce18c5c824fc8d06f4eef33d6df0c21ec7799510cde0f159a2", size = 4574214, upload-time = "2026-03-05T20:06:34.946Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/47/20/5369c3ce21588c708bcbe517a8fbe1a8dfdb5dfd5137e14790b1da71612c/ruff-0.15.5-py3-none-linux_armv6l.whl", hash = "sha256:4ae44c42281f42e3b06b988e442d344a5b9b72450ff3c892e30d11b29a96a57c", size = 10478185, upload-time = "2026-03-05T20:06:29.093Z" },
+    { url = "https://files.pythonhosted.org/packages/44/ed/e81dd668547da281e5dce710cf0bc60193f8d3d43833e8241d006720e42b/ruff-0.15.5-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:6edd3792d408ebcf61adabc01822da687579a1a023f297618ac27a5b51ef0080", size = 10859201, upload-time = "2026-03-05T20:06:32.632Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/8f/533075f00aaf19b07c5cd6aa6e5d89424b06b3b3f4583bfa9c640a079059/ruff-0.15.5-py3-none-macosx_11_0_arm64.whl", hash = "sha256:89f463f7c8205a9f8dea9d658d59eff49db05f88f89cc3047fb1a02d9f344010", size = 10184752, upload-time = "2026-03-05T20:06:40.312Z" },
+    { url = "https://files.pythonhosted.org/packages/66/0e/ba49e2c3fa0395b3152bad634c7432f7edfc509c133b8f4529053ff024fb/ruff-0.15.5-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ba786a8295c6574c1116704cf0b9e6563de3432ac888d8f83685654fe528fd65", size = 10534857, upload-time = "2026-03-05T20:06:19.581Z" },
+    { url = "https://files.pythonhosted.org/packages/59/71/39234440f27a226475a0659561adb0d784b4d247dfe7f43ffc12dd02e288/ruff-0.15.5-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:fd4b801e57955fe9f02b31d20375ab3a5c4415f2e5105b79fb94cf2642c91440", size = 10309120, upload-time = "2026-03-05T20:06:00.435Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/87/4140aa86a93df032156982b726f4952aaec4a883bb98cb6ef73c347da253/ruff-0.15.5-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:391f7c73388f3d8c11b794dbbc2959a5b5afe66642c142a6effa90b45f6f5204", size = 11047428, upload-time = "2026-03-05T20:05:51.867Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/f7/4953e7e3287676f78fbe85e3a0ca414c5ca81237b7575bdadc00229ac240/ruff-0.15.5-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8dc18f30302e379fe1e998548b0f5e9f4dff907f52f73ad6da419ea9c19d66c8", size = 11914251, upload-time = "2026-03-05T20:06:22.887Z" },
+    { url = "https://files.pythonhosted.org/packages/77/46/0f7c865c10cf896ccf5a939c3e84e1cfaeed608ff5249584799a74d33835/ruff-0.15.5-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:1cc6e7f90087e2d27f98dc34ed1b3ab7c8f0d273cc5431415454e22c0bd2a681", size = 11333801, upload-time = "2026-03-05T20:05:57.168Z" },
+    { url = "https://files.pythonhosted.org/packages/d3/01/a10fe54b653061585e655f5286c2662ebddb68831ed3eaebfb0eb08c0a16/ruff-0.15.5-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c1cb7169f53c1ddb06e71a9aebd7e98fc0fea936b39afb36d8e86d36ecc2636a", size = 11206821, upload-time = "2026-03-05T20:06:03.441Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/0d/2132ceaf20c5e8699aa83da2706ecb5c5dcdf78b453f77edca7fb70f8a93/ruff-0.15.5-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:9b037924500a31ee17389b5c8c4d88874cc6ea8e42f12e9c61a3d754ff72f1ca", size = 11133326, upload-time = "2026-03-05T20:06:25.655Z" },
+    { url = "https://files.pythonhosted.org/packages/72/cb/2e5259a7eb2a0f87c08c0fe5bf5825a1e4b90883a52685524596bfc93072/ruff-0.15.5-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:65bb414e5b4eadd95a8c1e4804f6772bbe8995889f203a01f77ddf2d790929dd", size = 10510820, upload-time = "2026-03-05T20:06:37.79Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/20/b67ce78f9e6c59ffbdb5b4503d0090e749b5f2d31b599b554698a80d861c/ruff-0.15.5-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:d20aa469ae3b57033519c559e9bc9cd9e782842e39be05b50e852c7c981fa01d", size = 10302395, upload-time = "2026-03-05T20:05:54.504Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/e5/719f1acccd31b720d477751558ed74e9c88134adcc377e5e886af89d3072/ruff-0.15.5-py3-none-musllinux_1_2_i686.whl", hash = "sha256:15388dd28c9161cdb8eda68993533acc870aa4e646a0a277aa166de9ad5a8752", size = 10754069, upload-time = "2026-03-05T20:06:06.422Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/9c/d1db14469e32d98f3ca27079dbd30b7b44dbb5317d06ab36718dee3baf03/ruff-0.15.5-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:b30da330cbd03bed0c21420b6b953158f60c74c54c5f4c1dabbdf3a57bf355d2", size = 11304315, upload-time = "2026-03-05T20:06:10.867Z" },
+    { url = "https://files.pythonhosted.org/packages/28/3a/950367aee7c69027f4f422059227b290ed780366b6aecee5de5039d50fa8/ruff-0.15.5-py3-none-win32.whl", hash = "sha256:732e5ee1f98ba5b3679029989a06ca39a950cced52143a0ea82a2102cb592b74", size = 10551676, upload-time = "2026-03-05T20:06:13.705Z" },
+    { url = "https://files.pythonhosted.org/packages/b8/00/bf077a505b4e649bdd3c47ff8ec967735ce2544c8e4a43aba42ee9bf935d/ruff-0.15.5-py3-none-win_amd64.whl", hash = "sha256:821d41c5fa9e19117616c35eaa3f4b75046ec76c65e7ae20a333e9a8696bc7fe", size = 11678972, upload-time = "2026-03-05T20:06:45.379Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/4e/cd76eca6db6115604b7626668e891c9dd03330384082e33662fb0f113614/ruff-0.15.5-py3-none-win_arm64.whl", hash = "sha256:b498d1c60d2fe5c10c45ec3f698901065772730b411f164ae270bb6bfcc4740b", size = 10965572, upload-time = "2026-03-05T20:06:16.984Z" },
+]
+
+[[package]]
+name = "safehttpx"
+version = "0.1.7"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "httpx" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/89/d1/4282284d9cf1ee873607a46442da977fc3c985059315ab23610be31d5885/safehttpx-0.1.7.tar.gz", hash = "sha256:db201c0978c41eddb8bb480f3eee59dd67304fdd91646035e9d9a720049a9d23", size = 10385, upload-time = "2025-10-24T18:30:09.783Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2e/a3/0f0b7d78e2f1eb9e8e1afbff1d2bff8d60144aee17aca51c065b516743dd/safehttpx-0.1.7-py3-none-any.whl", hash = "sha256:c4f4a162db6993464d7ca3d7cc4af0ffc6515a606dfd220b9f82c6945d869cde", size = 8959, upload-time = "2025-10-24T18:30:08.733Z" },
+]
+
+[[package]]
+name = "safetensors"
+version = "0.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/29/9c/6e74567782559a63bd040a236edca26fd71bc7ba88de2ef35d75df3bca5e/safetensors-0.7.0.tar.gz", hash = "sha256:07663963b67e8bd9f0b8ad15bb9163606cd27cc5a1b96235a50d8369803b96b0", size = 200878, upload-time = "2025-11-19T15:18:43.199Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/47/aef6c06649039accf914afef490268e1067ed82be62bcfa5b7e886ad15e8/safetensors-0.7.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c82f4d474cf725255d9e6acf17252991c3c8aac038d6ef363a4bf8be2f6db517", size = 467781, upload-time = "2025-11-19T15:18:35.84Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/00/374c0c068e30cd31f1e1b46b4b5738168ec79e7689ca82ee93ddfea05109/safetensors-0.7.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:94fd4858284736bb67a897a41608b5b0c2496c9bdb3bf2af1fa3409127f20d57", size = 447058, upload-time = "2025-11-19T15:18:34.416Z" },
+    { url = "https://files.pythonhosted.org/packages/f1/06/578ffed52c2296f93d7fd2d844cabfa92be51a587c38c8afbb8ae449ca89/safetensors-0.7.0-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e07d91d0c92a31200f25351f4acb2bc6aff7f48094e13ebb1d0fb995b54b6542", size = 491748, upload-time = "2025-11-19T15:18:09.79Z" },
+    { url = "https://files.pythonhosted.org/packages/ae/33/1debbbb70e4791dde185edb9413d1fe01619255abb64b300157d7f15dddd/safetensors-0.7.0-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8469155f4cb518bafb4acf4865e8bb9d6804110d2d9bdcaa78564b9fd841e104", size = 503881, upload-time = "2025-11-19T15:18:16.145Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/1c/40c2ca924d60792c3be509833df711b553c60effbd91da6f5284a83f7122/safetensors-0.7.0-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:54bef08bf00a2bff599982f6b08e8770e09cc012d7bba00783fc7ea38f1fb37d", size = 623463, upload-time = "2025-11-19T15:18:21.11Z" },
+    { url = "https://files.pythonhosted.org/packages/9b/3a/13784a9364bd43b0d61eef4bea2845039bc2030458b16594a1bd787ae26e/safetensors-0.7.0-cp38-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:42cb091236206bb2016d245c377ed383aa7f78691748f3bb6ee1bfa51ae2ce6a", size = 532855, upload-time = "2025-11-19T15:18:25.719Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/60/429e9b1cb3fc651937727befe258ea24122d9663e4d5709a48c9cbfceecb/safetensors-0.7.0-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:dac7252938f0696ddea46f5e855dd3138444e82236e3be475f54929f0c510d48", size = 507152, upload-time = "2025-11-19T15:18:33.023Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/a8/4b45e4e059270d17af60359713ffd83f97900d45a6afa73aaa0d737d48b6/safetensors-0.7.0-cp38-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1d060c70284127fa805085d8f10fbd0962792aed71879d00864acda69dbab981", size = 541856, upload-time = "2025-11-19T15:18:31.075Z" },
+    { url = "https://files.pythonhosted.org/packages/06/87/d26d8407c44175d8ae164a95b5a62707fcc445f3c0c56108e37d98070a3d/safetensors-0.7.0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:cdab83a366799fa730f90a4ebb563e494f28e9e92c4819e556152ad55e43591b", size = 674060, upload-time = "2025-11-19T15:18:37.211Z" },
+    { url = "https://files.pythonhosted.org/packages/11/f5/57644a2ff08dc6325816ba7217e5095f17269dada2554b658442c66aed51/safetensors-0.7.0-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:672132907fcad9f2aedcb705b2d7b3b93354a2aec1b2f706c4db852abe338f85", size = 771715, upload-time = "2025-11-19T15:18:38.689Z" },
+    { url = "https://files.pythonhosted.org/packages/86/31/17883e13a814bd278ae6e266b13282a01049b0c81341da7fd0e3e71a80a3/safetensors-0.7.0-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:5d72abdb8a4d56d4020713724ba81dac065fedb7f3667151c4a637f1d3fb26c0", size = 714377, upload-time = "2025-11-19T15:18:40.162Z" },
+    { url = "https://files.pythonhosted.org/packages/4a/d8/0c8a7dc9b41dcac53c4cbf9df2b9c83e0e0097203de8b37a712b345c0be5/safetensors-0.7.0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:b0f6d66c1c538d5a94a73aa9ddca8ccc4227e6c9ff555322ea40bdd142391dd4", size = 677368, upload-time = "2025-11-19T15:18:41.627Z" },
+    { url = "https://files.pythonhosted.org/packages/05/e5/cb4b713c8a93469e3c5be7c3f8d77d307e65fe89673e731f5c2bfd0a9237/safetensors-0.7.0-cp38-abi3-win32.whl", hash = "sha256:c74af94bf3ac15ac4d0f2a7c7b4663a15f8c2ab15ed0fc7531ca61d0835eccba", size = 326423, upload-time = "2025-11-19T15:18:45.74Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/e6/ec8471c8072382cb91233ba7267fd931219753bb43814cbc71757bfd4dab/safetensors-0.7.0-cp38-abi3-win_amd64.whl", hash = "sha256:d1239932053f56f3456f32eb9625590cc7582e905021f94636202a864d470755", size = 341380, upload-time = "2025-11-19T15:18:44.427Z" },
+    { url = "https://files.pythonhosted.org/packages/a7/6a/4d08d89a6fcbe905c5ae68b8b34f0791850882fc19782d0d02c65abbdf3b/safetensors-0.7.0-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f4729811a6640d019a4b7ba8638ee2fd21fa5ca8c7e7bdf0fed62068fcaac737", size = 492430, upload-time = "2025-11-19T15:18:11.884Z" },
+    { url = "https://files.pythonhosted.org/packages/dd/29/59ed8152b30f72c42d00d241e58eaca558ae9dbfa5695206e2e0f54c7063/safetensors-0.7.0-pp310-pypy310_pp73-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:12f49080303fa6bb424b362149a12949dfbbf1e06811a88f2307276b0c131afd", size = 503977, upload-time = "2025-11-19T15:18:17.523Z" },
+    { url = "https://files.pythonhosted.org/packages/d3/0b/4811bfec67fa260e791369b16dab105e4bae82686120554cc484064e22b4/safetensors-0.7.0-pp310-pypy310_pp73-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:0071bffba4150c2f46cae1432d31995d77acfd9f8db598b5d1a2ce67e8440ad2", size = 623890, upload-time = "2025-11-19T15:18:22.666Z" },
+    { url = "https://files.pythonhosted.org/packages/58/5b/632a58724221ef03d78ab65062e82a1010e1bef8e8e0b9d7c6d7b8044841/safetensors-0.7.0-pp310-pypy310_pp73-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:473b32699f4200e69801bf5abf93f1a4ecd432a70984df164fc22ccf39c4a6f3", size = 531885, upload-time = "2025-11-19T15:18:27.146Z" },
+]
+
+[[package]]
+name = "scikit-learn"
+version = "1.7.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "joblib" },
+    { name = "numpy" },
+    { name = "scipy" },
+    { name = "threadpoolctl" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/98/c2/a7855e41c9d285dfe86dc50b250978105dce513d6e459ea66a6aeb0e1e0c/scikit_learn-1.7.2.tar.gz", hash = "sha256:20e9e49ecd130598f1ca38a1d85090e1a600147b9c02fa6f15d69cb53d968fda", size = 7193136, upload-time = "2025-09-09T08:21:29.075Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ba/3e/daed796fd69cce768b8788401cc464ea90b306fb196ae1ffed0b98182859/scikit_learn-1.7.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:6b33579c10a3081d076ab403df4a4190da4f4432d443521674637677dc91e61f", size = 9336221, upload-time = "2025-09-09T08:20:19.328Z" },
+    { url = "https://files.pythonhosted.org/packages/1c/ce/af9d99533b24c55ff4e18d9b7b4d9919bbc6cd8f22fe7a7be01519a347d5/scikit_learn-1.7.2-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:36749fb62b3d961b1ce4fedf08fa57a1986cd409eff2d783bca5d4b9b5fce51c", size = 8653834, upload-time = "2025-09-09T08:20:22.073Z" },
+    { url = "https://files.pythonhosted.org/packages/58/0e/8c2a03d518fb6bd0b6b0d4b114c63d5f1db01ff0f9925d8eb10960d01c01/scikit_learn-1.7.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:7a58814265dfc52b3295b1900cfb5701589d30a8bb026c7540f1e9d3499d5ec8", size = 9660938, upload-time = "2025-09-09T08:20:24.327Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/75/4311605069b5d220e7cf5adabb38535bd96f0079313cdbb04b291479b22a/scikit_learn-1.7.2-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4a847fea807e278f821a0406ca01e387f97653e284ecbd9750e3ee7c90347f18", size = 9477818, upload-time = "2025-09-09T08:20:26.845Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/9b/87961813c34adbca21a6b3f6b2bea344c43b30217a6d24cc437c6147f3e8/scikit_learn-1.7.2-cp310-cp310-win_amd64.whl", hash = "sha256:ca250e6836d10e6f402436d6463d6c0e4d8e0234cfb6a9a47835bd392b852ce5", size = 8886969, upload-time = "2025-09-09T08:20:29.329Z" },
+]
+
+[[package]]
+name = "scipy"
+version = "1.15.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0f/37/6964b830433e654ec7485e45a00fc9a27cf868d622838f6b6d9c5ec0d532/scipy-1.15.3.tar.gz", hash = "sha256:eae3cf522bc7df64b42cad3925c876e1b0b6c35c1337c93e12c0f366f55b0eaf", size = 59419214, upload-time = "2025-05-08T16:13:05.955Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/78/2f/4966032c5f8cc7e6a60f1b2e0ad686293b9474b65246b0c642e3ef3badd0/scipy-1.15.3-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:a345928c86d535060c9c2b25e71e87c39ab2f22fc96e9636bd74d1dbf9de448c", size = 38702770, upload-time = "2025-05-08T16:04:20.849Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/6e/0c3bf90fae0e910c274db43304ebe25a6b391327f3f10b5dcc638c090795/scipy-1.15.3-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:ad3432cb0f9ed87477a8d97f03b763fd1d57709f1bbde3c9369b1dff5503b253", size = 30094511, upload-time = "2025-05-08T16:04:27.103Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/b1/4deb37252311c1acff7f101f6453f0440794f51b6eacb1aad4459a134081/scipy-1.15.3-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:aef683a9ae6eb00728a542b796f52a5477b78252edede72b8327a886ab63293f", size = 22368151, upload-time = "2025-05-08T16:04:31.731Z" },
+    { url = "https://files.pythonhosted.org/packages/38/7d/f457626e3cd3c29b3a49ca115a304cebb8cc6f31b04678f03b216899d3c6/scipy-1.15.3-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:1c832e1bd78dea67d5c16f786681b28dd695a8cb1fb90af2e27580d3d0967e92", size = 25121732, upload-time = "2025-05-08T16:04:36.596Z" },
+    { url = "https://files.pythonhosted.org/packages/db/0a/92b1de4a7adc7a15dcf5bddc6e191f6f29ee663b30511ce20467ef9b82e4/scipy-1.15.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:263961f658ce2165bbd7b99fa5135195c3a12d9bef045345016b8b50c315cb82", size = 35547617, upload-time = "2025-05-08T16:04:43.546Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/6d/41991e503e51fc1134502694c5fa7a1671501a17ffa12716a4a9151af3df/scipy-1.15.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9e2abc762b0811e09a0d3258abee2d98e0c703eee49464ce0069590846f31d40", size = 37662964, upload-time = "2025-05-08T16:04:49.431Z" },
+    { url = "https://files.pythonhosted.org/packages/25/e1/3df8f83cb15f3500478c889be8fb18700813b95e9e087328230b98d547ff/scipy-1.15.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ed7284b21a7a0c8f1b6e5977ac05396c0d008b89e05498c8b7e8f4a1423bba0e", size = 37238749, upload-time = "2025-05-08T16:04:55.215Z" },
+    { url = "https://files.pythonhosted.org/packages/93/3e/b3257cf446f2a3533ed7809757039016b74cd6f38271de91682aa844cfc5/scipy-1.15.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5380741e53df2c566f4d234b100a484b420af85deb39ea35a1cc1be84ff53a5c", size = 40022383, upload-time = "2025-05-08T16:05:01.914Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/84/55bc4881973d3f79b479a5a2e2df61c8c9a04fcb986a213ac9c02cfb659b/scipy-1.15.3-cp310-cp310-win_amd64.whl", hash = "sha256:9d61e97b186a57350f6d6fd72640f9e99d5a4a2b8fbf4b9ee9a841eab327dc13", size = 41259201, upload-time = "2025-05-08T16:05:08.166Z" },
+]
+
+[[package]]
+name = "seaborn"
+version = "0.13.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "matplotlib" },
+    { name = "numpy" },
+    { name = "pandas" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/86/59/a451d7420a77ab0b98f7affa3a1d78a313d2f7281a57afb1a34bae8ab412/seaborn-0.13.2.tar.gz", hash = "sha256:93e60a40988f4d65e9f4885df477e2fdaff6b73a9ded434c1ab356dd57eefff7", size = 1457696, upload-time = "2024-01-25T13:21:52.551Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/83/11/00d3c3dfc25ad54e731d91449895a79e4bf2384dc3ac01809010ba88f6d5/seaborn-0.13.2-py3-none-any.whl", hash = "sha256:636f8336facf092165e27924f223d3c62ca560b1f2bb5dff7ab7fad265361987", size = 294914, upload-time = "2024-01-25T13:21:49.598Z" },
+]
+
+[[package]]
+name = "semantic-version"
+version = "2.10.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/7d/31/f2289ce78b9b473d582568c234e104d2a342fd658cc288a7553d83bb8595/semantic_version-2.10.0.tar.gz", hash = "sha256:bdabb6d336998cbb378d4b9db3a4b56a1e3235701dc05ea2690d9a997ed5041c", size = 52289, upload-time = "2022-05-26T13:35:23.454Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6a/23/8146aad7d88f4fcb3a6218f41a60f6c2d4e3a72de72da1825dc7c8f7877c/semantic_version-2.10.0-py2.py3-none-any.whl", hash = "sha256:de78a3b8e0feda74cabc54aab2da702113e33ac9d9eb9d2389bcf1f58b7d9177", size = 15552, upload-time = "2022-05-26T13:35:21.206Z" },
+]
+
+[[package]]
+name = "sentence-transformers"
+version = "5.2.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "huggingface-hub" },
+    { name = "numpy" },
+    { name = "scikit-learn" },
+    { name = "scipy" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/5b/30/21664028fc0776eb1ca024879480bbbab36f02923a8ff9e4cae5a150fa35/sentence_transformers-5.2.3.tar.gz", hash = "sha256:3cd3044e1f3fe859b6a1b66336aac502eaae5d3dd7d5c8fc237f37fbf58137c7", size = 381623, upload-time = "2026-02-17T14:05:20.238Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/46/9f/dba4b3e18ebbe1eaa29d9f1764fbc7da0cd91937b83f2b7928d15c5d2d36/sentence_transformers-5.2.3-py3-none-any.whl", hash = "sha256:6437c62d4112b615ddebda362dfc16a4308d604c5b68125ed586e3e95d5b2e30", size = 494225, upload-time = "2026-02-17T14:05:18.596Z" },
+]
+
+[[package]]
+name = "sentencepiece"
+version = "0.2.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/15/15/2e7a025fc62d764b151ae6d0f2a92f8081755ebe8d4a64099accc6f77ba6/sentencepiece-0.2.1.tar.gz", hash = "sha256:8138cec27c2f2282f4a34d9a016e3374cd40e5c6e9cb335063db66a0a3b71fad", size = 3228515, upload-time = "2025-08-12T07:00:51.718Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/af/31/5b7cccb307b485db1a2372d6d2980b0a65d067f8be5ca943a103b4acd5b3/sentencepiece-0.2.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e10fa50bdbaa5e2445dbd387979980d391760faf0ec99a09bd7780ff37eaec44", size = 1942557, upload-time = "2025-08-12T06:59:12.379Z" },
+    { url = "https://files.pythonhosted.org/packages/1f/41/0ac923a8e685ad290c5afc8ae55c5844977b8d75076fcc04302b9a324274/sentencepiece-0.2.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2f27ae6deea72efdb6f361750c92f6c21fd0ad087445082770cc34015213c526", size = 1325384, upload-time = "2025-08-12T06:59:14.334Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/ef/3751555d67daf9003384978f169d31c775cb5c7baf28633caaf1eb2b2b4d/sentencepiece-0.2.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:60937c959e6f44159fdd9f56fbdd302501f96114a5ba436829496d5f32d8de3f", size = 1253317, upload-time = "2025-08-12T06:59:16.247Z" },
+    { url = "https://files.pythonhosted.org/packages/46/a5/742c69b7bd144eb32b6e5fd50dbd8abbbc7a95fce2fe16e50156fa400e3b/sentencepiece-0.2.1-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d8b1d91545578852f128650b8cce4ec20f93d39b378ff554ebe66290f2dabb92", size = 1316379, upload-time = "2025-08-12T06:59:17.825Z" },
+    { url = "https://files.pythonhosted.org/packages/c8/89/8deeafbba2871e8fa10f20f17447786f4ac38085925335728d360eaf4cae/sentencepiece-0.2.1-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:27e38eee653abc3d387862e67bc5c8b6f428cd604e688b85d29170b7e725c26c", size = 1387926, upload-time = "2025-08-12T06:59:19.395Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/ca/67fe73005f0ab617c6a970b199754e28e524b6873aa7025224fad3cda252/sentencepiece-0.2.1-cp310-cp310-win32.whl", hash = "sha256:251874d720ac7f28024a168501f3c7bb15d1802245f6e66de565f18bbb9b5eaa", size = 999550, upload-time = "2025-08-12T06:59:20.844Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/33/dc5b54042050d2dda4229c3ce1f862541c99966390b6aa20f54d520d2dc2/sentencepiece-0.2.1-cp310-cp310-win_amd64.whl", hash = "sha256:e52144670738b4b477fade6c2a9b6af71a8d0094514c9853ac9f6fc1fcfabae7", size = 1054613, upload-time = "2025-08-12T06:59:22.255Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/19/1ea47f46ff97fe04422b78997da1a37cd632f414aae042d27a9009c5b733/sentencepiece-0.2.1-cp310-cp310-win_arm64.whl", hash = "sha256:9076430ac25dfa7147d9d05751dbc66a04bc1aaac371c07f84952979ea59f0d0", size = 1033884, upload-time = "2025-08-12T06:59:24.194Z" },
+]
+
+[[package]]
+name = "sentry-sdk"
+version = "2.54.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi" },
+    { name = "urllib3" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c8/e9/2e3a46c304e7fa21eaa70612f60354e32699c7102eb961f67448e222ad7c/sentry_sdk-2.54.0.tar.gz", hash = "sha256:2620c2575128d009b11b20f7feb81e4e4e8ae08ec1d36cbc845705060b45cc1b", size = 413813, upload-time = "2026-03-02T15:12:41.355Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/53/39/be412cc86bc6247b8f69e9383d7950711bd86f8d0a4a4b0fe8fad685bc21/sentry_sdk-2.54.0-py2.py3-none-any.whl", hash = "sha256:fd74e0e281dcda63afff095d23ebcd6e97006102cdc8e78a29f19ecdf796a0de", size = 439198, upload-time = "2026-03-02T15:12:39.546Z" },
+]
+
+[[package]]
+name = "setuptools"
+version = "69.5.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d6/4f/b10f707e14ef7de524fe1f8988a294fb262a29c9b5b12275c7e188864aed/setuptools-69.5.1.tar.gz", hash = "sha256:6c1fccdac05a97e598fb0ae3bbed5904ccb317337a51139dcd51453611bbb987", size = 2291314, upload-time = "2024-04-13T21:06:28.589Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f7/29/13965af254e3373bceae8fb9a0e6ea0d0e571171b80d6646932131d6439b/setuptools-69.5.1-py3-none-any.whl", hash = "sha256:c636ac361bc47580504644275c9ad802c50415c7522212252c033bd15f301f32", size = 894566, upload-time = "2024-04-13T21:06:23.256Z" },
+]
+
+[[package]]
+name = "shellingham"
+version = "1.5.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/58/15/8b3609fd3830ef7b27b655beb4b4e9c62313a4e8da8c676e142cc210d58e/shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de", size = 10310, upload-time = "2023-10-24T04:13:40.426Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e0/f9/0595336914c5619e5f28a1fb793285925a8cd4b432c9da0a987836c7f822/shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686", size = 9755, upload-time = "2023-10-24T04:13:38.866Z" },
+]
+
+[[package]]
+name = "simplejson"
+version = "3.20.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/41/f4/a1ac5ed32f7ed9a088d62a59d410d4c204b3b3815722e2ccfb491fa8251b/simplejson-3.20.2.tar.gz", hash = "sha256:5fe7a6ce14d1c300d80d08695b7f7e633de6cd72c80644021874d985b3393649", size = 85784, upload-time = "2025-09-26T16:29:36.64Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/78/09/2bf3761de89ea2d91bdce6cf107dcd858892d0adc22c995684878826cc6b/simplejson-3.20.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:6d7286dc11af60a2f76eafb0c2acde2d997e87890e37e24590bb513bec9f1bc5", size = 94039, upload-time = "2025-09-26T16:27:29.283Z" },
+    { url = "https://files.pythonhosted.org/packages/0f/33/c3277db8931f0ae9e54b9292668863365672d90fb0f632f4cf9829cb7d68/simplejson-3.20.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c01379b4861c3b0aa40cba8d44f2b448f5743999aa68aaa5d3ef7049d4a28a2d", size = 75894, upload-time = "2025-09-26T16:27:30.378Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/ea/ae47b04d03c7c8a7b7b1a8b39a6e27c3bd424e52f4988d70aca6293ff5e5/simplejson-3.20.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a16b029ca25645b3bc44e84a4f941efa51bf93c180b31bd704ce6349d1fc77c1", size = 76116, upload-time = "2025-09-26T16:27:31.42Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/42/6c9af551e5a8d0f171d6dce3d9d1260068927f7b80f1f09834e07887c8c4/simplejson-3.20.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3e22a5fb7b1437ffb057e02e1936a3bfb19084ae9d221ec5e9f4cf85f69946b6", size = 138827, upload-time = "2025-09-26T16:27:32.486Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/22/5e268bbcbe9f75577491e406ec0a5536f5b2fa91a3b52031fea51cd83e1d/simplejson-3.20.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d8b6ff02fc7b8555c906c24735908854819b0d0dc85883d453e23ca4c0445d01", size = 146772, upload-time = "2025-09-26T16:27:34.036Z" },
+    { url = "https://files.pythonhosted.org/packages/71/b4/800f14728e2ad666f420dfdb57697ca128aeae7f991b35759c09356b829a/simplejson-3.20.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:2bfc1c396ad972ba4431130b42307b2321dba14d988580c1ac421ec6a6b7cee3", size = 134497, upload-time = "2025-09-26T16:27:35.211Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/b9/c54eef4226c6ac8e9a389bbe5b21fef116768f97a2dc1a683c716ffe66ef/simplejson-3.20.2-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3a97249ee1aee005d891b5a211faf58092a309f3d9d440bc269043b08f662eda", size = 138172, upload-time = "2025-09-26T16:27:36.44Z" },
+    { url = "https://files.pythonhosted.org/packages/09/36/4e282f5211b34620f1b2e4b51d9ddaab5af82219b9b7b78360a33f7e5387/simplejson-3.20.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:f1036be00b5edaddbddbb89c0f80ed229714a941cfd21e51386dc69c237201c2", size = 140272, upload-time = "2025-09-26T16:27:37.605Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/b0/94ad2cf32f477c449e1f63c863d8a513e2408d651c4e58fe4b6a7434e168/simplejson-3.20.2-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5d6f5bacb8cdee64946b45f2680afa3f54cd38e62471ceda89f777693aeca4e4", size = 140468, upload-time = "2025-09-26T16:27:39.015Z" },
+    { url = "https://files.pythonhosted.org/packages/e5/46/827731e4163be3f987cb8ee90f5d444161db8f540b5e735355faa098d9bc/simplejson-3.20.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:8db6841fb796ec5af632f677abf21c6425a1ebea0d9ac3ef1a340b8dc69f52b8", size = 148700, upload-time = "2025-09-26T16:27:40.171Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/28/c32121064b1ec2fb7b5d872d9a1abda62df064d35e0160eddfa907118343/simplejson-3.20.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c0a341f7cc2aae82ee2b31f8a827fd2e51d09626f8b3accc441a6907c88aedb7", size = 141323, upload-time = "2025-09-26T16:27:41.324Z" },
+    { url = "https://files.pythonhosted.org/packages/46/b6/c897c54326fe86dd12d101981171a49361949f4728294f418c3b86a1af77/simplejson-3.20.2-cp310-cp310-win32.whl", hash = "sha256:27f9c01a6bc581d32ab026f515226864576da05ef322d7fc141cd8a15a95ce53", size = 74377, upload-time = "2025-09-26T16:27:42.533Z" },
+    { url = "https://files.pythonhosted.org/packages/ad/87/a6e03d4d80cca99c1fee4e960f3440e2f21be9470e537970f960ca5547f1/simplejson-3.20.2-cp310-cp310-win_amd64.whl", hash = "sha256:c0a63ec98a4547ff366871bf832a7367ee43d047bcec0b07b66c794e2137b476", size = 76081, upload-time = "2025-09-26T16:27:43.945Z" },
+    { url = "https://files.pythonhosted.org/packages/05/5b/83e1ff87eb60ca706972f7e02e15c0b33396e7bdbd080069a5d1b53cf0d8/simplejson-3.20.2-py3-none-any.whl", hash = "sha256:3b6bb7fb96efd673eac2e4235200bfffdc2353ad12c54117e1e4e2fc485ac017", size = 57309, upload-time = "2025-09-26T16:29:35.312Z" },
+]
+
+[[package]]
+name = "six"
+version = "1.17.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/94/e7/b2c673351809dca68a0e064b6af791aa332cf192da575fd474ed7d6f16a2/six-1.17.0.tar.gz", hash = "sha256:ff70335d468e7eb6ec65b95b99d3a2836546063f63acc5171de367e834932a81", size = 34031, upload-time = "2024-12-04T17:35:28.174Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b7/ce/149a00dd41f10bc29e5921b496af8b574d8413afcd5e30dfa0ed46c2cc5e/six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274", size = 11050, upload-time = "2024-12-04T17:35:26.475Z" },
+]
+
+[[package]]
+name = "smmap"
+version = "5.0.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/44/cd/a040c4b3119bbe532e5b0732286f805445375489fceaec1f48306068ee3b/smmap-5.0.2.tar.gz", hash = "sha256:26ea65a03958fa0c8a1c7e8c7a58fdc77221b8910f6be2131affade476898ad5", size = 22329, upload-time = "2025-01-02T07:14:40.909Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/04/be/d09147ad1ec7934636ad912901c5fd7667e1c858e19d355237db0d0cd5e4/smmap-5.0.2-py3-none-any.whl", hash = "sha256:b30115f0def7d7531d22a0fb6502488d879e75b260a9db4d0819cfb25403af5e", size = 24303, upload-time = "2025-01-02T07:14:38.724Z" },
+]
+
+[[package]]
+name = "sniffio"
+version = "1.3.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a2/87/a6771e1546d97e7e041b6ae58d80074f81b7d5121207425c964ddf5cfdbd/sniffio-1.3.1.tar.gz", hash = "sha256:f4324edc670a0f49750a81b895f35c3adb843cca46f0530f79fc1babb23789dc", size = 20372, upload-time = "2024-02-25T23:20:04.057Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e9/44/75a9c9421471a6c4805dbf2356f7c181a29c1879239abab1ea2cc8f38b40/sniffio-1.3.1-py3-none-any.whl", hash = "sha256:2f6da418d1f1e0fddd844478f41680e794e6051915791a034ff65e5f100525a2", size = 10235, upload-time = "2024-02-25T23:20:01.196Z" },
+]
+
+[[package]]
+name = "sortedcontainers"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e8/c4/ba2f8066cceb6f23394729afe52f3bf7adec04bf9ed2c820b39e19299111/sortedcontainers-2.4.0.tar.gz", hash = "sha256:25caa5a06cc30b6b83d11423433f65d1f9d76c4c6a0c90e3379eaa43b9bfdb88", size = 30594, upload-time = "2021-05-16T22:03:42.897Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/46/9cb0e58b2deb7f82b84065f37f3bffeb12413f947f9388e4cac22c4621ce/sortedcontainers-2.4.0-py2.py3-none-any.whl", hash = "sha256:a163dcaede0f1c021485e957a39245190e74249897e2ae4b2aa38595db237ee0", size = 29575, upload-time = "2021-05-16T22:03:41.177Z" },
+]
+
+[[package]]
+name = "soundfile"
+version = "0.13.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cffi" },
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e1/41/9b873a8c055582859b239be17902a85339bec6a30ad162f98c9b0288a2cc/soundfile-0.13.1.tar.gz", hash = "sha256:b2c68dab1e30297317080a5b43df57e302584c49e2942defdde0acccc53f0e5b", size = 46156, upload-time = "2025-01-25T09:17:04.831Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/64/28/e2a36573ccbcf3d57c00626a21fe51989380636e821b341d36ccca0c1c3a/soundfile-0.13.1-py2.py3-none-any.whl", hash = "sha256:a23c717560da2cf4c7b5ae1142514e0fd82d6bbd9dfc93a50423447142f2c445", size = 25751, upload-time = "2025-01-25T09:16:44.235Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/ab/73e97a5b3cc46bba7ff8650a1504348fa1863a6f9d57d7001c6b67c5f20e/soundfile-0.13.1-py2.py3-none-macosx_10_9_x86_64.whl", hash = "sha256:82dc664d19831933fe59adad199bf3945ad06d84bc111a5b4c0d3089a5b9ec33", size = 1142250, upload-time = "2025-01-25T09:16:47.583Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/e5/58fd1a8d7b26fc113af244f966ee3aecf03cb9293cb935daaddc1e455e18/soundfile-0.13.1-py2.py3-none-macosx_11_0_arm64.whl", hash = "sha256:743f12c12c4054921e15736c6be09ac26b3b3d603aef6fd69f9dde68748f2593", size = 1101406, upload-time = "2025-01-25T09:16:49.662Z" },
+    { url = "https://files.pythonhosted.org/packages/58/ae/c0e4a53d77cf6e9a04179535766b3321b0b9ced5f70522e4caf9329f0046/soundfile-0.13.1-py2.py3-none-manylinux_2_28_aarch64.whl", hash = "sha256:9c9e855f5a4d06ce4213f31918653ab7de0c5a8d8107cd2427e44b42df547deb", size = 1235729, upload-time = "2025-01-25T09:16:53.018Z" },
+    { url = "https://files.pythonhosted.org/packages/57/5e/70bdd9579b35003a489fc850b5047beeda26328053ebadc1fb60f320f7db/soundfile-0.13.1-py2.py3-none-manylinux_2_28_x86_64.whl", hash = "sha256:03267c4e493315294834a0870f31dbb3b28a95561b80b134f0bd3cf2d5f0e618", size = 1313646, upload-time = "2025-01-25T09:16:54.872Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/df/8c11dc4dfceda14e3003bb81a0d0edcaaf0796dd7b4f826ea3e532146bba/soundfile-0.13.1-py2.py3-none-win32.whl", hash = "sha256:c734564fab7c5ddf8e9be5bf70bab68042cd17e9c214c06e365e20d64f9a69d5", size = 899881, upload-time = "2025-01-25T09:16:56.663Z" },
+    { url = "https://files.pythonhosted.org/packages/14/e9/6b761de83277f2f02ded7e7ea6f07828ec78e4b229b80e4ca55dd205b9dc/soundfile-0.13.1-py2.py3-none-win_amd64.whl", hash = "sha256:1e70a05a0626524a69e9f0f4dd2ec174b4e9567f4d8b6c11d38b5c289be36ee9", size = 1019162, upload-time = "2025-01-25T09:16:59.573Z" },
+]
+
+[[package]]
+name = "stack-data"
+version = "0.6.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "asttokens" },
+    { name = "executing" },
+    { name = "pure-eval" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/28/e3/55dcc2cfbc3ca9c29519eb6884dd1415ecb53b0e934862d3559ddcb7e20b/stack_data-0.6.3.tar.gz", hash = "sha256:836a778de4fec4dcd1dcd89ed8abff8a221f58308462e1c4aa2a3cf30148f0b9", size = 44707, upload-time = "2023-09-30T13:58:05.479Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f1/7b/ce1eafaf1a76852e2ec9b22edecf1daa58175c090266e9f6c64afcd81d91/stack_data-0.6.3-py3-none-any.whl", hash = "sha256:d5558e0c25a4cb0853cddad3d77da9891a08cb85dd9f9f91b9f8cd66e511e695", size = 24521, upload-time = "2023-09-30T13:58:03.53Z" },
+]
+
+[[package]]
+name = "starlette"
+version = "0.52.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c4/68/79977123bb7be889ad680d79a40f339082c1978b5cfcf62c2d8d196873ac/starlette-0.52.1.tar.gz", hash = "sha256:834edd1b0a23167694292e94f597773bc3f89f362be6effee198165a35d62933", size = 2653702, upload-time = "2026-01-18T13:34:11.062Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/81/0d/13d1d239a25cbfb19e740db83143e95c772a1fe10202dda4b76792b114dd/starlette-0.52.1-py3-none-any.whl", hash = "sha256:0029d43eb3d273bc4f83a08720b4912ea4b071087a3b48db01b7c839f7954d74", size = 74272, upload-time = "2026-01-18T13:34:09.188Z" },
+]
+
+[[package]]
+name = "sympy"
+version = "1.14.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "mpmath" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/83/d3/803453b36afefb7c2bb238361cd4ae6125a569b4db67cd9e79846ba2d68c/sympy-1.14.0.tar.gz", hash = "sha256:d3d3fe8df1e5a0b42f0e7bdf50541697dbe7d23746e894990c030e2b05e72517", size = 7793921, upload-time = "2025-04-27T18:05:01.611Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a2/09/77d55d46fd61b4a135c444fc97158ef34a095e5681d0a6c10b75bf356191/sympy-1.14.0-py3-none-any.whl", hash = "sha256:e091cc3e99d2141a0ba2847328f5479b05d94a6635cb96148ccb3f34671bd8f5", size = 6299353, upload-time = "2025-04-27T18:04:59.103Z" },
+]
+
+[[package]]
+name = "tensorboard"
+version = "2.19.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "grpcio" },
+    { name = "markdown" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "protobuf" },
+    { name = "setuptools" },
+    { name = "six" },
+    { name = "tensorboard-data-server" },
+    { name = "werkzeug" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/12/4f70e8e2ba0dbe72ea978429d8530b0333f0ed2140cc571a48802878ef99/tensorboard-2.19.0-py3-none-any.whl", hash = "sha256:5e71b98663a641a7ce8a6e70b0be8e1a4c0c45d48760b076383ac4755c35b9a0", size = 5503412, upload-time = "2025-02-12T08:17:27.21Z" },
+]
+
+[[package]]
+name = "tensorboard-data-server"
+version = "0.7.2"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7a/13/e503968fefabd4c6b2650af21e110aa8466fe21432cd7c43a84577a89438/tensorboard_data_server-0.7.2-py3-none-any.whl", hash = "sha256:7e0610d205889588983836ec05dc098e80f97b7e7bbff7e994ebb78f578d0ddb", size = 2356, upload-time = "2023-10-23T21:23:32.16Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/85/dabeaf902892922777492e1d253bb7e1264cadce3cea932f7ff599e53fea/tensorboard_data_server-0.7.2-py3-none-macosx_10_9_x86_64.whl", hash = "sha256:9fe5d24221b29625dbc7328b0436ca7fc1c23de4acf4d272f1180856e32f9f60", size = 4823598, upload-time = "2023-10-23T21:23:33.714Z" },
+    { url = "https://files.pythonhosted.org/packages/73/c6/825dab04195756cf8ff2e12698f22513b3db2f64925bdd41671bfb33aaa5/tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl", hash = "sha256:ef687163c24185ae9754ed5650eb5bc4d84ff257aabdc33f0cc6f74d8ba54530", size = 6590363, upload-time = "2023-10-23T21:23:35.583Z" },
+]
+
+[[package]]
+name = "tensorflow"
+version = "2.19.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "astunparse" },
+    { name = "flatbuffers" },
+    { name = "gast" },
+    { name = "google-pasta" },
+    { name = "grpcio" },
+    { name = "h5py" },
+    { name = "keras" },
+    { name = "libclang" },
+    { name = "ml-dtypes" },
+    { name = "numpy" },
+    { name = "opt-einsum" },
+    { name = "packaging" },
+    { name = "protobuf" },
+    { name = "requests" },
+    { name = "setuptools" },
+    { name = "six" },
+    { name = "tensorboard" },
+    { name = "tensorflow-io-gcs-filesystem" },
+    { name = "termcolor" },
+    { name = "typing-extensions" },
+    { name = "wrapt" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f5/49/9e39dc714629285ef421fc986c082409833bf86ec0bdf8cbcc6702949922/tensorflow-2.19.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:c95604f25c3032e9591c7e01e457fdd442dde48e9cc1ce951078973ab1b4ca34", size = 252464253, upload-time = "2025-03-12T01:04:13.652Z" },
+    { url = "https://files.pythonhosted.org/packages/45/cf/96dfffd7b04398cf0fe74c228972ba275b8f5867a6a0d4a472005d3469c4/tensorflow-2.19.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2b39293cae3aeee534dc4746dc6097b48c281e5e8b9a423efbd14d4495968e5c", size = 252498594, upload-time = "2025-03-12T01:04:25.295Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/b6/86f99528b3edca3c31cad43e79b15debc9124c7cbc772a8f8e82667fd427/tensorflow-2.19.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:83e2d6c748105488205d30e43093f28fc90e8da0176db9ddee12e2784cf435e8", size = 644752673, upload-time = "2025-03-12T01:04:42.14Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/03/8bf7bfb538fad40571b781a2aaa1ae905f617acef79d0aa8da7cc92390fb/tensorflow-2.19.0-cp310-cp310-win_amd64.whl", hash = "sha256:d3f47452246bd08902f0c865d3839fa715f1738d801d256934b943aa21c5a1d2", size = 375723719, upload-time = "2025-03-12T01:05:02.774Z" },
+]
+
+[[package]]
+name = "tensorflow-datasets"
+version = "4.9.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "array-record" },
+    { name = "click" },
+    { name = "dm-tree" },
+    { name = "etils", extra = ["enp", "epath"] },
+    { name = "numpy" },
+    { name = "promise" },
+    { name = "protobuf" },
+    { name = "psutil" },
+    { name = "requests" },
+    { name = "tensorflow-metadata" },
+    { name = "termcolor" },
+    { name = "toml" },
+    { name = "tqdm" },
+    { name = "wrapt" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/16/76/1fb7925616c4227907af5003e82a6da4cd0c77e576787cd5d1bfd94f91b9/tensorflow-datasets-4.9.2.tar.gz", hash = "sha256:09c6f59cf89b247328ce7b0c623ac37c0d10b7c23d53dfa39a58ecd0e88823e1", size = 4079257, upload-time = "2023-04-13T11:21:07.969Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/62/82/521e36235c7890b067415fa874f5074ff21f1005a189d79fd72b33b39ca1/tensorflow_datasets-4.9.2-py3-none-any.whl", hash = "sha256:aa23fdde1c661f2342fd11d9b373599055899a27b79b1d839adc77fc7daaa15a", size = 5387170, upload-time = "2023-04-13T11:21:05.539Z" },
+]
+
+[[package]]
+name = "tensorflow-io-gcs-filesystem"
+version = "0.37.1"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e9/a3/12d7e7326a707919b321e2d6e4c88eb61596457940fd2b8ff3e9b7fac8a7/tensorflow_io_gcs_filesystem-0.37.1-cp310-cp310-macosx_10_14_x86_64.whl", hash = "sha256:249c12b830165841411ba71e08215d0e94277a49c551e6dd5d72aab54fe5491b", size = 2470224, upload-time = "2024-07-01T23:44:15.341Z" },
+    { url = "https://files.pythonhosted.org/packages/1c/55/3849a188cc15e58fefde20e9524d124a629a67a06b4dc0f6c881cb3c6e39/tensorflow_io_gcs_filesystem-0.37.1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:257aab23470a0796978efc9c2bcf8b0bc80f22e6298612a4c0a50d3f4e88060c", size = 3479613, upload-time = "2024-07-01T23:44:17.445Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/19/9095c69e22c879cb3896321e676c69273a549a3148c4f62aa4bc5ebdb20f/tensorflow_io_gcs_filesystem-0.37.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8febbfcc67c61e542a5ac1a98c7c20a91a5e1afc2e14b1ef0cb7c28bc3b6aa70", size = 4842078, upload-time = "2024-07-01T23:44:18.977Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/48/47b7d25572961a48b1de3729b7a11e835b888e41e0203cca82df95d23b91/tensorflow_io_gcs_filesystem-0.37.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9679b36e3a80921876f31685ab6f7270f3411a4cc51bc2847e80d0e4b5291e27", size = 5085736, upload-time = "2024-07-01T23:44:21.034Z" },
+]
+
+[[package]]
+name = "tensorflow-metadata"
+version = "1.16.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "protobuf" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e0/57/393aa9dde72347cde9e01f665bac344f14adefd5c748be45b23aa5804f6d/tensorflow_metadata-1.16.1-py3-none-any.whl", hash = "sha256:2ce72ea31d78a00c0c74c6d465482335aa5cb2a3b2a104dedba0b258bc7bb18a", size = 28984, upload-time = "2024-10-09T19:57:12.683Z" },
+]
+
+[[package]]
+name = "termcolor"
+version = "3.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/46/79/cf31d7a93a8fdc6aa0fbb665be84426a8c5a557d9240b6239e9e11e35fc5/termcolor-3.3.0.tar.gz", hash = "sha256:348871ca648ec6a9a983a13ab626c0acce02f515b9e1983332b17af7979521c5", size = 14434, upload-time = "2025-12-29T12:55:21.882Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/33/d1/8bb87d21e9aeb323cc03034f5eaf2c8f69841e40e4853c2627edf8111ed3/termcolor-3.3.0-py3-none-any.whl", hash = "sha256:cf642efadaf0a8ebbbf4bc7a31cec2f9b5f21a9f726f4ccbb08192c9c26f43a5", size = 7734, upload-time = "2025-12-29T12:55:20.718Z" },
+]
+
+[[package]]
+name = "tf-keras"
+version = "2.19.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "tensorflow" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a9/71/a6e82b8e8ce04a8ab1ff0560791698b70d1b2dda1a0270f1d69a834e8880/tf_keras-2.19.0.tar.gz", hash = "sha256:b09a407d87a4571ce1e8ca985cfc68483e3d63b2518a5d79a97ad92cb64dbe9c", size = 1261172, upload-time = "2025-03-03T21:24:13.582Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/45/6b/d245122d108a94df5969ee7408ad343af1627730e91478e01ef098976bfa/tf_keras-2.19.0-py3-none-any.whl", hash = "sha256:4f339e800987b39d1548a8c76a7b33b6801a97ec7fcd89c299ec29741f7890bd", size = 1726787, upload-time = "2025-03-03T21:24:11.826Z" },
+]
+
+[[package]]
+name = "threadpoolctl"
+version = "3.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b7/4d/08c89e34946fce2aec4fbb45c9016efd5f4d7f24af8e5d93296e935631d8/threadpoolctl-3.6.0.tar.gz", hash = "sha256:8ab8b4aa3491d812b623328249fab5302a68d2d71745c8a4c719a2fcaba9f44e", size = 21274, upload-time = "2025-03-13T13:49:23.031Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/d5/f9a850d79b0851d1d4ef6456097579a9005b31fea68726a4ae5f2d82ddd9/threadpoolctl-3.6.0-py3-none-any.whl", hash = "sha256:43a0b8fd5a2928500110039e43a5eed8480b918967083ea48dc3ab9f13c4a7fb", size = 18638, upload-time = "2025-03-13T13:49:21.846Z" },
+]
+
+[[package]]
+name = "tiktoken"
+version = "0.12.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "regex" },
+    { name = "requests" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7d/ab/4d017d0f76ec3171d469d80fc03dfbb4e48a4bcaddaa831b31d526f05edc/tiktoken-0.12.0.tar.gz", hash = "sha256:b18ba7ee2b093863978fcb14f74b3707cdc8d4d4d3836853ce7ec60772139931", size = 37806, upload-time = "2025-10-06T20:22:45.419Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/89/b3/2cb7c17b6c4cf8ca983204255d3f1d95eda7213e247e6947a0ee2c747a2c/tiktoken-0.12.0-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:3de02f5a491cfd179aec916eddb70331814bd6bf764075d39e21d5862e533970", size = 1051991, upload-time = "2025-10-06T20:21:34.098Z" },
+    { url = "https://files.pythonhosted.org/packages/27/0f/df139f1df5f6167194ee5ab24634582ba9a1b62c6b996472b0277ec80f66/tiktoken-0.12.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:b6cfb6d9b7b54d20af21a912bfe63a2727d9cfa8fbda642fd8322c70340aad16", size = 995798, upload-time = "2025-10-06T20:21:35.579Z" },
+    { url = "https://files.pythonhosted.org/packages/ef/5d/26a691f28ab220d5edc09b9b787399b130f24327ef824de15e5d85ef21aa/tiktoken-0.12.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:cde24cdb1b8a08368f709124f15b36ab5524aac5fa830cc3fdce9c03d4fb8030", size = 1129865, upload-time = "2025-10-06T20:21:36.675Z" },
+    { url = "https://files.pythonhosted.org/packages/b2/94/443fab3d4e5ebecac895712abd3849b8da93b7b7dec61c7db5c9c7ebe40c/tiktoken-0.12.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:6de0da39f605992649b9cfa6f84071e3f9ef2cec458d08c5feb1b6f0ff62e134", size = 1152856, upload-time = "2025-10-06T20:21:37.873Z" },
+    { url = "https://files.pythonhosted.org/packages/54/35/388f941251b2521c70dd4c5958e598ea6d2c88e28445d2fb8189eecc1dfc/tiktoken-0.12.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:6faa0534e0eefbcafaccb75927a4a380463a2eaa7e26000f0173b920e98b720a", size = 1195308, upload-time = "2025-10-06T20:21:39.577Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/00/c6681c7f833dd410576183715a530437a9873fa910265817081f65f9105f/tiktoken-0.12.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:82991e04fc860afb933efb63957affc7ad54f83e2216fe7d319007dab1ba5892", size = 1255697, upload-time = "2025-10-06T20:21:41.154Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/d2/82e795a6a9bafa034bf26a58e68fe9a89eeaaa610d51dbeb22106ba04f0a/tiktoken-0.12.0-cp310-cp310-win_amd64.whl", hash = "sha256:6fb2995b487c2e31acf0a9e17647e3b242235a20832642bb7a9d1a181c0c1bb1", size = 879375, upload-time = "2025-10-06T20:21:43.201Z" },
+]
+
+[[package]]
+name = "timm"
+version = "1.0.25"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "huggingface-hub" },
+    { name = "pyyaml" },
+    { name = "safetensors" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d7/2c/593109822fe735e637382aca6640c1102c19797f7791f1fd1dab2d6c3cb1/timm-1.0.25.tar.gz", hash = "sha256:47f59fc2754725735cc81bb83bcbfce5bec4ebd5d4bb9e69da57daa92fcfa768", size = 2414743, upload-time = "2026-02-23T16:49:00.137Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ef/50/de09f69a74278a16f08f1d562047a2d6713783765ee3c6971881a2b21a3f/timm-1.0.25-py3-none-any.whl", hash = "sha256:bef7f61dd717cb2dbbb7e326f143e13d660a47ecbd84116e6fe33732bed5c484", size = 2565837, upload-time = "2026-02-23T16:48:58.324Z" },
+]
+
+[[package]]
+name = "tokenizers"
+version = "0.21.4"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c2/2f/402986d0823f8d7ca139d969af2917fefaa9b947d1fb32f6168c509f2492/tokenizers-0.21.4.tar.gz", hash = "sha256:fa23f85fbc9a02ec5c6978da172cdcbac23498c3ca9f3645c5c68740ac007880", size = 351253, upload-time = "2025-07-28T15:48:54.325Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/98/c6/fdb6f72bf6454f52eb4a2510be7fb0f614e541a2554d6210e370d85efff4/tokenizers-0.21.4-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:2ccc10a7c3bcefe0f242867dc914fc1226ee44321eb618cfe3019b5df3400133", size = 2863987, upload-time = "2025-07-28T15:48:44.877Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/a6/28975479e35ddc751dc1ddc97b9b69bf7fcf074db31548aab37f8116674c/tokenizers-0.21.4-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:5e2f601a8e0cd5be5cc7506b20a79112370b9b3e9cb5f13f68ab11acd6ca7d60", size = 2732457, upload-time = "2025-07-28T15:48:43.265Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/8f/24f39d7b5c726b7b0be95dca04f344df278a3fe3a4deb15a975d194cbb32/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:39b376f5a1aee67b4d29032ee85511bbd1b99007ec735f7f35c8a2eb104eade5", size = 3012624, upload-time = "2025-07-28T13:22:43.895Z" },
+    { url = "https://files.pythonhosted.org/packages/58/47/26358925717687a58cb74d7a508de96649544fad5778f0cd9827398dc499/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:2107ad649e2cda4488d41dfd031469e9da3fcbfd6183e74e4958fa729ffbf9c6", size = 2939681, upload-time = "2025-07-28T13:22:47.499Z" },
+    { url = "https://files.pythonhosted.org/packages/99/6f/cc300fea5db2ab5ddc2c8aea5757a27b89c84469899710c3aeddc1d39801/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3c73012da95afafdf235ba80047699df4384fdc481527448a078ffd00e45a7d9", size = 3247445, upload-time = "2025-07-28T15:48:39.711Z" },
+    { url = "https://files.pythonhosted.org/packages/be/bf/98cb4b9c3c4afd8be89cfa6423704337dc20b73eb4180397a6e0d456c334/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f23186c40395fc390d27f519679a58023f368a0aad234af145e0f39ad1212732", size = 3428014, upload-time = "2025-07-28T13:22:49.569Z" },
+    { url = "https://files.pythonhosted.org/packages/75/c7/96c1cc780e6ca7f01a57c13235dd05b7bc1c0f3588512ebe9d1331b5f5ae/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:cc88bb34e23a54cc42713d6d98af5f1bf79c07653d24fe984d2d695ba2c922a2", size = 3193197, upload-time = "2025-07-28T13:22:51.471Z" },
+    { url = "https://files.pythonhosted.org/packages/f2/90/273b6c7ec78af547694eddeea9e05de771278bd20476525ab930cecaf7d8/tokenizers-0.21.4-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:51b7eabb104f46c1c50b486520555715457ae833d5aee9ff6ae853d1130506ff", size = 3115426, upload-time = "2025-07-28T15:48:41.439Z" },
+    { url = "https://files.pythonhosted.org/packages/91/43/c640d5a07e95f1cf9d2c92501f20a25f179ac53a4f71e1489a3dcfcc67ee/tokenizers-0.21.4-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:714b05b2e1af1288bd1bc56ce496c4cebb64a20d158ee802887757791191e6e2", size = 9089127, upload-time = "2025-07-28T15:48:46.472Z" },
+    { url = "https://files.pythonhosted.org/packages/44/a1/dd23edd6271d4dca788e5200a807b49ec3e6987815cd9d0a07ad9c96c7c2/tokenizers-0.21.4-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:1340ff877ceedfa937544b7d79f5b7becf33a4cfb58f89b3b49927004ef66f78", size = 9055243, upload-time = "2025-07-28T15:48:48.539Z" },
+    { url = "https://files.pythonhosted.org/packages/21/2b/b410d6e9021c4b7ddb57248304dc817c4d4970b73b6ee343674914701197/tokenizers-0.21.4-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:3c1f4317576e465ac9ef0d165b247825a2a4078bcd01cba6b54b867bdf9fdd8b", size = 9298237, upload-time = "2025-07-28T15:48:50.443Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/0a/42348c995c67e2e6e5c89ffb9cfd68507cbaeb84ff39c49ee6e0a6dd0fd2/tokenizers-0.21.4-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:c212aa4e45ec0bb5274b16b6f31dd3f1c41944025c2358faaa5782c754e84c24", size = 9461980, upload-time = "2025-07-28T15:48:52.325Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/d3/dacccd834404cd71b5c334882f3ba40331ad2120e69ded32cf5fda9a7436/tokenizers-0.21.4-cp39-abi3-win32.whl", hash = "sha256:6c42a930bc5f4c47f4ea775c91de47d27910881902b0f20e4990ebe045a415d0", size = 2329871, upload-time = "2025-07-28T15:48:56.841Z" },
+    { url = "https://files.pythonhosted.org/packages/41/f2/fd673d979185f5dcbac4be7d09461cbb99751554ffb6718d0013af8604cb/tokenizers-0.21.4-cp39-abi3-win_amd64.whl", hash = "sha256:475d807a5c3eb72c59ad9b5fcdb254f6e17f53dfcbb9903233b0dfa9c943b597", size = 2507568, upload-time = "2025-07-28T15:48:55.456Z" },
+]
+
+[[package]]
+name = "tokenizers"
+version = "0.22.2"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/73/6f/f80cfef4a312e1fb34baf7d85c72d4411afde10978d4657f8cdd811d3ccc/tokenizers-0.22.2.tar.gz", hash = "sha256:473b83b915e547aa366d1eee11806deaf419e17be16310ac0a14077f1e28f917", size = 372115, upload-time = "2026-01-05T10:45:15.988Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/92/97/5dbfabf04c7e348e655e907ed27913e03db0923abb5dfdd120d7b25630e1/tokenizers-0.22.2-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:544dd704ae7238755d790de45ba8da072e9af3eea688f698b137915ae959281c", size = 3100275, upload-time = "2026-01-05T10:41:02.158Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/47/174dca0502ef88b28f1c9e06b73ce33500eedfac7a7692108aec220464e7/tokenizers-0.22.2-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:1e418a55456beedca4621dbab65a318981467a2b188e982a23e117f115ce5001", size = 2981472, upload-time = "2026-01-05T10:41:00.276Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/84/7990e799f1309a8b87af6b948f31edaa12a3ed22d11b352eaf4f4b2e5753/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2249487018adec45d6e3554c71d46eb39fa8ea67156c640f7513eb26f318cec7", size = 3290736, upload-time = "2026-01-05T10:40:32.165Z" },
+    { url = "https://files.pythonhosted.org/packages/78/59/09d0d9ba94dcd5f4f1368d4858d24546b4bdc0231c2354aa31d6199f0399/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:25b85325d0815e86e0bac263506dd114578953b7b53d7de09a6485e4a160a7dd", size = 3168835, upload-time = "2026-01-05T10:40:38.847Z" },
+    { url = "https://files.pythonhosted.org/packages/47/50/b3ebb4243e7160bda8d34b731e54dd8ab8b133e50775872e7a434e524c28/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:bfb88f22a209ff7b40a576d5324bf8286b519d7358663db21d6246fb17eea2d5", size = 3521673, upload-time = "2026-01-05T10:40:56.614Z" },
+    { url = "https://files.pythonhosted.org/packages/e0/fa/89f4cb9e08df770b57adb96f8cbb7e22695a4cb6c2bd5f0c4f0ebcf33b66/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1c774b1276f71e1ef716e5486f21e76333464f47bece56bbd554485982a9e03e", size = 3724818, upload-time = "2026-01-05T10:40:44.507Z" },
+    { url = "https://files.pythonhosted.org/packages/64/04/ca2363f0bfbe3b3d36e95bf67e56a4c88c8e3362b658e616d1ac185d47f2/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:df6c4265b289083bf710dff49bc51ef252f9d5be33a45ee2bed151114a56207b", size = 3379195, upload-time = "2026-01-05T10:40:51.139Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/76/932be4b50ef6ccedf9d3c6639b056a967a86258c6d9200643f01269211ca/tokenizers-0.22.2-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:369cc9fc8cc10cb24143873a0d95438bb8ee257bb80c71989e3ee290e8d72c67", size = 3274982, upload-time = "2026-01-05T10:40:58.331Z" },
+    { url = "https://files.pythonhosted.org/packages/1d/28/5f9f5a4cc211b69e89420980e483831bcc29dade307955cc9dc858a40f01/tokenizers-0.22.2-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:29c30b83d8dcd061078b05ae0cb94d3c710555fbb44861139f9f83dcca3dc3e4", size = 9478245, upload-time = "2026-01-05T10:41:04.053Z" },
+    { url = "https://files.pythonhosted.org/packages/6c/fb/66e2da4704d6aadebf8cb39f1d6d1957df667ab24cff2326b77cda0dcb85/tokenizers-0.22.2-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:37ae80a28c1d3265bb1f22464c856bd23c02a05bb211e56d0c5301a435be6c1a", size = 9560069, upload-time = "2026-01-05T10:45:10.673Z" },
+    { url = "https://files.pythonhosted.org/packages/16/04/fed398b05caa87ce9b1a1bb5166645e38196081b225059a6edaff6440fac/tokenizers-0.22.2-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:791135ee325f2336f498590eb2f11dc5c295232f288e75c99a36c5dbce63088a", size = 9899263, upload-time = "2026-01-05T10:45:12.559Z" },
+    { url = "https://files.pythonhosted.org/packages/05/a1/d62dfe7376beaaf1394917e0f8e93ee5f67fea8fcf4107501db35996586b/tokenizers-0.22.2-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:38337540fbbddff8e999d59970f3c6f35a82de10053206a7562f1ea02d046fa5", size = 10033429, upload-time = "2026-01-05T10:45:14.333Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/18/a545c4ea42af3df6effd7d13d250ba77a0a86fb20393143bbb9a92e434d4/tokenizers-0.22.2-cp39-abi3-win32.whl", hash = "sha256:a6bf3f88c554a2b653af81f3204491c818ae2ac6fbc09e76ef4773351292bc92", size = 2502363, upload-time = "2026-01-05T10:45:20.593Z" },
+    { url = "https://files.pythonhosted.org/packages/65/71/0670843133a43d43070abeb1949abfdef12a86d490bea9cd9e18e37c5ff7/tokenizers-0.22.2-cp39-abi3-win_amd64.whl", hash = "sha256:c9ea31edff2968b44a88f97d784c2f16dc0729b8b143ed004699ebca91f05c48", size = 2747786, upload-time = "2026-01-05T10:45:18.411Z" },
+    { url = "https://files.pythonhosted.org/packages/72/f4/0de46cfa12cdcbcd464cc59fde36912af405696f687e53a091fb432f694c/tokenizers-0.22.2-cp39-abi3-win_arm64.whl", hash = "sha256:9ce725d22864a1e965217204946f830c37876eee3b2ba6fc6255e8e903d5fcbc", size = 2612133, upload-time = "2026-01-05T10:45:17.232Z" },
+    { url = "https://files.pythonhosted.org/packages/84/04/655b79dbcc9b3ac5f1479f18e931a344af67e5b7d3b251d2dcdcd7558592/tokenizers-0.22.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:753d47ebd4542742ef9261d9da92cd545b2cacbb48349a1225466745bb866ec4", size = 3282301, upload-time = "2026-01-05T10:40:34.858Z" },
+    { url = "https://files.pythonhosted.org/packages/46/cd/e4851401f3d8f6f45d8480262ab6a5c8cb9c4302a790a35aa14eeed6d2fd/tokenizers-0.22.2-pp310-pypy310_pp73-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e10bf9113d209be7cd046d40fbabbaf3278ff6d18eb4da4c500443185dc1896c", size = 3161308, upload-time = "2026-01-05T10:40:40.737Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/6e/55553992a89982cd12d4a66dddb5e02126c58677ea3931efcbe601d419db/tokenizers-0.22.2-pp310-pypy310_pp73-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:64d94e84f6660764e64e7e0b22baa72f6cd942279fdbb21d46abd70d179f0195", size = 3718964, upload-time = "2026-01-05T10:40:46.56Z" },
+    { url = "https://files.pythonhosted.org/packages/59/8c/b1c87148aa15e099243ec9f0cf9d0e970cc2234c3257d558c25a2c5304e6/tokenizers-0.22.2-pp310-pypy310_pp73-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f01a9c019878532f98927d2bacb79bbb404b43d3437455522a00a30718cdedb5", size = 3373542, upload-time = "2026-01-05T10:40:52.803Z" },
+]
+
+[[package]]
+name = "toml"
+version = "0.10.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/be/ba/1f744cdc819428fc6b5084ec34d9b30660f6f9daaf70eead706e3203ec3c/toml-0.10.2.tar.gz", hash = "sha256:b3bda1d108d5dd99f4a20d24d9c348e91c4db7ab1b749200bded2f839ccbe68f", size = 22253, upload-time = "2020-11-01T01:40:22.204Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/44/6f/7120676b6d73228c96e17f1f794d8ab046fc910d781c8d151120c3f1569e/toml-0.10.2-py2.py3-none-any.whl", hash = "sha256:806143ae5bfb6a3c6e736a764057db0e6a0e05e338b5630894a5f779cabb4f9b", size = 16588, upload-time = "2020-11-01T01:40:20.672Z" },
+]
+
+[[package]]
+name = "tomli"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/82/30/31573e9457673ab10aa432461bee537ce6cef177667deca369efb79df071/tomli-2.4.0.tar.gz", hash = "sha256:aa89c3f6c277dd275d8e243ad24f3b5e701491a860d5121f2cdd399fbb31fc9c", size = 17477, upload-time = "2026-01-11T11:22:38.165Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/23/d1/136eb2cb77520a31e1f64cbae9d33ec6df0d78bdf4160398e86eec8a8754/tomli-2.4.0-py3-none-any.whl", hash = "sha256:1f776e7d669ebceb01dee46484485f43a4048746235e683bcdffacdf1fb4785a", size = 14477, upload-time = "2026-01-11T11:22:37.446Z" },
+]
+
+[[package]]
+name = "tomlkit"
+version = "0.13.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cc/18/0bbf3884e9eaa38819ebe46a7bd25dcd56b67434402b66a58c4b8e552575/tomlkit-0.13.3.tar.gz", hash = "sha256:430cf247ee57df2b94ee3fbe588e71d362a941ebb545dec29b53961d61add2a1", size = 185207, upload-time = "2025-06-05T07:13:44.947Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/bd/75/8539d011f6be8e29f339c42e633aae3cb73bffa95dd0f9adec09b9c58e85/tomlkit-0.13.3-py3-none-any.whl", hash = "sha256:c89c649d79ee40629a9fda55f8ace8c6a1b42deb912b2a8fd8d942ddadb606b0", size = 38901, upload-time = "2025-06-05T07:13:43.546Z" },
+]
+
+[[package]]
+name = "torch"
+version = "2.8.0"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "filelock", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "fsspec", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "jinja2", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "networkx", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "sympy", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "typing-extensions", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/63/28/110f7274254f1b8476c561dada127173f994afa2b1ffc044efb773c15650/torch-2.8.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:0be92c08b44009d4131d1ff7a8060d10bafdb7ddcb7359ef8d8c5169007ea905", size = 102052793, upload-time = "2025-08-06T14:53:15.852Z" },
+    { url = "https://files.pythonhosted.org/packages/70/1c/58da560016f81c339ae14ab16c98153d51c941544ae568da3cb5b1ceb572/torch-2.8.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:89aa9ee820bb39d4d72b794345cccef106b574508dd17dbec457949678c76011", size = 888025420, upload-time = "2025-08-06T14:54:18.014Z" },
+    { url = "https://files.pythonhosted.org/packages/70/87/f69752d0dd4ba8218c390f0438130c166fa264a33b7025adb5014b92192c/torch-2.8.0-cp310-cp310-win_amd64.whl", hash = "sha256:e8e5bf982e87e2b59d932769938b698858c64cc53753894be25629bdf5cf2f46", size = 241363614, upload-time = "2025-08-06T14:53:31.496Z" },
+    { url = "https://files.pythonhosted.org/packages/ef/d6/e6d4c57e61c2b2175d3aafbfb779926a2cfd7c32eeda7c543925dceec923/torch-2.8.0-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:a3f16a58a9a800f589b26d47ee15aca3acf065546137fc2af039876135f4c760", size = 73611154, upload-time = "2025-08-06T14:53:10.919Z" },
+]
+
+[[package]]
+name = "torch"
+version = "2.8.0+cu128"
+source = { registry = "https://download.pytorch.org/whl/cu128" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+]
+dependencies = [
+    { name = "filelock", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "fsspec", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "jinja2", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "networkx", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cublas-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cuda-cupti-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cuda-nvrtc-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cuda-runtime-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cudnn-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cufft-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cufile-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-curand-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cusolver-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cusparse-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-cusparselt-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-nccl-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-nvjitlink-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "nvidia-nvtx-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "sympy", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "triton", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "typing-extensions", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:0c96999d15cf1f13dd7c913e0b21a9a355538e6cfc10861a17158320292f5954" },
+    { url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-win_amd64.whl", hash = "sha256:43938e9a174c90e5eb9e906532b2f1e21532bbfa5a61b65193b4f54714d34f9e" },
+]
+
+[[package]]
+name = "torchao"
+version = "0.16.0"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8d/7f/0acda8a429ac9cfabd142d30af624d7958bf828c438be5a54ca87bbe16d7/torchao-0.16.0-cp310-abi3-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2d6293a0c57c9dd505efb025a7189459d154965fbed000efd638cf299f9362dd", size = 3160415, upload-time = "2026-02-10T22:12:12.32Z" },
+    { url = "https://files.pythonhosted.org/packages/d0/3d/0c5a5833a135a045510e06c06b3d4cf316b06d59415bc21e0b021a000cc8/torchao-0.16.0-py3-none-any.whl", hash = "sha256:d0a8d773351fd17b95fee81dfbcbf98577b567dcdbec47d221b0ee258432101d", size = 1164150, upload-time = "2026-02-10T22:12:15.28Z" },
+]
+
+[[package]]
+name = "torchcodec"
+version = "0.10.0"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d0/1c/549e79c0f5a7e48c1f767b648c3ea3301a8cd702f66c9cb32c86eb347d63/torchcodec-0.10.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a2ecb3e38414a9326ffca56e49a05a9f513813aa118b7f540110d09ab725d1f7", size = 3406242, upload-time = "2026-01-22T15:41:42.056Z" },
+    { url = "https://files.pythonhosted.org/packages/f0/47/041180c095e4dbc0cff8e847974bf400114379c8f114aa8c3c96e9d6bd4e/torchcodec-0.10.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:61f9b12fcd5b89d5e7e874c5feeb7c2c99821868a32f6ebbf6e8692409b2b6f7", size = 2062569, upload-time = "2026-01-22T15:41:34.99Z" },
+    { url = "https://files.pythonhosted.org/packages/dc/51/b7c339cabf375789fc74b56006fe6bc5e029ededf83bc7629864cb6ed083/torchcodec-0.10.0-cp310-cp310-win_amd64.whl", hash = "sha256:a49e3ebab65e559a98af6f63371a837f69b8542059da8def30bd6e658c5e86d3", size = 2208079, upload-time = "2026-01-22T15:41:49.866Z" },
+]
+
+[[package]]
+name = "torchvision"
+version = "0.23.0"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "numpy", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "pillow", marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4d/49/5ad5c3ff4920be0adee9eb4339b4fb3b023a0fc55b9ed8dbc73df92946b8/torchvision-0.23.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:7266871daca00ad46d1c073e55d972179d12a58fa5c9adec9a3db9bbed71284a", size = 1856885, upload-time = "2025-08-06T14:57:55.024Z" },
+    { url = "https://files.pythonhosted.org/packages/25/44/ddd56d1637bac42a8c5da2c8c440d8a28c431f996dd9790f32dd9a96ca6e/torchvision-0.23.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:31c583ba27426a3a04eca8c05450524105c1564db41be6632f7536ef405a6de2", size = 2394251, upload-time = "2025-08-06T14:58:01.725Z" },
+    { url = "https://files.pythonhosted.org/packages/93/f3/3cdf55bbf0f737304d997561c34ab0176222e0496b6743b0feab5995182c/torchvision-0.23.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:3932bf67256f2d095ce90a9f826f6033694c818856f4bb26794cf2ce64253e53", size = 8627497, upload-time = "2025-08-06T14:58:09.317Z" },
+    { url = "https://files.pythonhosted.org/packages/97/90/02afe57c3ef4284c5cf89d3b7ae203829b3a981f72b93a7dd2a3fd2c83c1/torchvision-0.23.0-cp310-cp310-win_amd64.whl", hash = "sha256:83ee5bf827d61a8af14620c0a61d8608558638ac9c3bac8adb7b27138e2147d1", size = 1600760, upload-time = "2025-08-06T14:57:56.783Z" },
+]
+
+[[package]]
+name = "torchvision"
+version = "0.23.0+cu128"
+source = { registry = "https://download.pytorch.org/whl/cu128" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+]
+dependencies = [
+    { name = "numpy", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "pillow", marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+wheels = [
+    { url = "https://download.pytorch.org/whl/cu128/torchvision-0.23.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:460bc8d70f63bdb433a7351decc2c1ae1903f7f378e4a7614fc8e8c97a5c36aa" },
+    { url = "https://download.pytorch.org/whl/cu128/torchvision-0.23.0%2Bcu128-cp310-cp310-win_amd64.whl", hash = "sha256:8ec6f2281ef5d52471b01b99eb04243d0c2cccb1972ba43217085025fe5a6c3f" },
+]
+
+[[package]]
+name = "tornado"
+version = "6.5.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/37/1d/0a336abf618272d53f62ebe274f712e213f5a03c0b2339575430b8362ef2/tornado-6.5.4.tar.gz", hash = "sha256:a22fa9047405d03260b483980635f0b041989d8bcc9a313f8fe18b411d84b1d7", size = 513632, upload-time = "2025-12-15T19:21:03.836Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ab/a9/e94a9d5224107d7ce3cc1fab8d5dc97f5ea351ccc6322ee4fb661da94e35/tornado-6.5.4-cp39-abi3-macosx_10_9_universal2.whl", hash = "sha256:d6241c1a16b1c9e4cc28148b1cda97dd1c6cb4fb7068ac1bedc610768dff0ba9", size = 443909, upload-time = "2025-12-15T19:20:48.382Z" },
+    { url = "https://files.pythonhosted.org/packages/db/7e/f7b8d8c4453f305a51f80dbb49014257bb7d28ccb4bbb8dd328ea995ecad/tornado-6.5.4-cp39-abi3-macosx_10_9_x86_64.whl", hash = "sha256:2d50f63dda1d2cac3ae1fa23d254e16b5e38153758470e9956cbc3d813d40843", size = 442163, upload-time = "2025-12-15T19:20:49.791Z" },
+    { url = "https://files.pythonhosted.org/packages/ba/b5/206f82d51e1bfa940ba366a8d2f83904b15942c45a78dd978b599870ab44/tornado-6.5.4-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d1cf66105dc6acb5af613c054955b8137e34a03698aa53272dbda4afe252be17", size = 445746, upload-time = "2025-12-15T19:20:51.491Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/9d/1a3338e0bd30ada6ad4356c13a0a6c35fbc859063fa7eddb309183364ac1/tornado-6.5.4-cp39-abi3-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:50ff0a58b0dc97939d29da29cd624da010e7f804746621c78d14b80238669335", size = 445083, upload-time = "2025-12-15T19:20:52.778Z" },
+    { url = "https://files.pythonhosted.org/packages/50/d4/e51d52047e7eb9a582da59f32125d17c0482d065afd5d3bc435ff2120dc5/tornado-6.5.4-cp39-abi3-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e5fb5e04efa54cf0baabdd10061eb4148e0be137166146fff835745f59ab9f7f", size = 445315, upload-time = "2025-12-15T19:20:53.996Z" },
+    { url = "https://files.pythonhosted.org/packages/27/07/2273972f69ca63dbc139694a3fc4684edec3ea3f9efabf77ed32483b875c/tornado-6.5.4-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:9c86b1643b33a4cd415f8d0fe53045f913bf07b4a3ef646b735a6a86047dda84", size = 446003, upload-time = "2025-12-15T19:20:56.101Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/83/41c52e47502bf7260044413b6770d1a48dda2f0246f95ee1384a3cd9c44a/tornado-6.5.4-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:6eb82872335a53dd063a4f10917b3efd28270b56a33db69009606a0312660a6f", size = 445412, upload-time = "2025-12-15T19:20:57.398Z" },
+    { url = "https://files.pythonhosted.org/packages/10/c7/bc96917f06cbee182d44735d4ecde9c432e25b84f4c2086143013e7b9e52/tornado-6.5.4-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:6076d5dda368c9328ff41ab5d9dd3608e695e8225d1cd0fd1e006f05da3635a8", size = 445392, upload-time = "2025-12-15T19:20:58.692Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/1a/d7592328d037d36f2d2462f4bc1fbb383eec9278bc786c1b111cbbd44cfa/tornado-6.5.4-cp39-abi3-win32.whl", hash = "sha256:1768110f2411d5cd281bac0a090f707223ce77fd110424361092859e089b38d1", size = 446481, upload-time = "2025-12-15T19:21:00.008Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/6d/c69be695a0a64fd37a97db12355a035a6d90f79067a3cf936ec2b1dc38cd/tornado-6.5.4-cp39-abi3-win_amd64.whl", hash = "sha256:fa07d31e0cd85c60713f2b995da613588aa03e1303d75705dca6af8babc18ddc", size = 446886, upload-time = "2025-12-15T19:21:01.287Z" },
+    { url = "https://files.pythonhosted.org/packages/50/49/8dc3fd90902f70084bd2cd059d576ddb4f8bb44c2c7c0e33a11422acb17e/tornado-6.5.4-cp39-abi3-win_arm64.whl", hash = "sha256:053e6e16701eb6cbe641f308f4c1a9541f91b6261991160391bfc342e8a551a1", size = 445910, upload-time = "2025-12-15T19:21:02.571Z" },
+]
+
+[[package]]
+name = "tqdm"
+version = "4.67.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/09/a9/6ba95a270c6f1fbcd8dac228323f2777d886cb206987444e4bce66338dd4/tqdm-4.67.3.tar.gz", hash = "sha256:7d825f03f89244ef73f1d4ce193cb1774a8179fd96f31d7e1dcde62092b960bb", size = 169598, upload-time = "2026-02-03T17:35:53.048Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/16/e1/3079a9ff9b8e11b846c6ac5c8b5bfb7ff225eee721825310c91b3b50304f/tqdm-4.67.3-py3-none-any.whl", hash = "sha256:ee1e4c0e59148062281c49d80b25b67771a127c85fc9676d3be5f243206826bf", size = 78374, upload-time = "2026-02-03T17:35:50.982Z" },
+]
+
+[[package]]
+name = "traitlets"
+version = "5.14.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/eb/79/72064e6a701c2183016abbbfedaba506d81e30e232a68c9f0d6f6fcd1574/traitlets-5.14.3.tar.gz", hash = "sha256:9ed0579d3502c94b4b3732ac120375cda96f923114522847de4b3bb98b96b6b7", size = 161621, upload-time = "2024-04-19T11:11:49.746Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/00/c0/8f5d070730d7836adc9c9b6408dec68c6ced86b304a9b26a14df072a6e8c/traitlets-5.14.3-py3-none-any.whl", hash = "sha256:b74e89e397b1ed28cc831db7aea759ba6640cb3de13090ca145426688ff1ac4f", size = 85359, upload-time = "2024-04-19T11:11:46.763Z" },
+]
+
+[[package]]
+name = "transformers"
+version = "4.51.3"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "filelock", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "pyyaml", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "regex", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "requests", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "safetensors", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "tokenizers", version = "0.21.4", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f1/11/7414d5bc07690002ce4d7553602107bf969af85144bbd02830f9fb471236/transformers-4.51.3.tar.gz", hash = "sha256:e292fcab3990c6defe6328f0f7d2004283ca81a7a07b2de9a46d67fd81ea1409", size = 8941266, upload-time = "2025-04-14T08:15:00.485Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a9/b6/5257d04ae327b44db31f15cce39e6020cc986333c715660b1315a9724d82/transformers-4.51.3-py3-none-any.whl", hash = "sha256:fd3279633ceb2b777013234bbf0b4f5c2d23c4626b05497691f00cfda55e8a83", size = 10383940, upload-time = "2025-04-14T08:13:43.023Z" },
+]
+
+[[package]]
+name = "transformers"
+version = "4.57.6"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "filelock", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "pyyaml", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "regex", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "requests", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "safetensors", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "tokenizers", version = "0.22.2", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c4/35/67252acc1b929dc88b6602e8c4a982e64f31e733b804c14bc24b47da35e6/transformers-4.57.6.tar.gz", hash = "sha256:55e44126ece9dc0a291521b7e5492b572e6ef2766338a610b9ab5afbb70689d3", size = 10134912, upload-time = "2026-01-16T10:38:39.284Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/03/b8/e484ef633af3887baeeb4b6ad12743363af7cce68ae51e938e00aaa0529d/transformers-4.57.6-py3-none-any.whl", hash = "sha256:4c9e9de11333ddfe5114bc872c9f370509198acf0b87a832a0ab9458e2bd0550", size = 11993498, upload-time = "2026-01-16T10:38:31.289Z" },
+]
+
+[[package]]
+name = "transformers-stream-generator"
+version = "0.0.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" } },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/42/c2/65f13aec253100e1916e9bd7965fe17bde796ebabeb1265f45191ab4ddc0/transformers-stream-generator-0.0.5.tar.gz", hash = "sha256:271deace0abf9c0f83b36db472c8ba61fdc7b04d1bf89d845644acac2795ed57", size = 13033, upload-time = "2024-03-11T14:18:02.079Z" }
+
+[[package]]
+name = "triton"
+version = "3.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "setuptools" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/62/ee/0ee5f64a87eeda19bbad9bc54ae5ca5b98186ed00055281fd40fb4beb10e/triton-3.4.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7ff2785de9bc02f500e085420273bb5cc9c9bb767584a4aa28d6e360cec70128", size = 155430069, upload-time = "2025-07-30T19:58:21.715Z" },
+]
+
+[[package]]
+name = "triton-windows"
+version = "3.6.0.post25"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ba/ca/6d38c374a427a360dc4c7687f15fdb217dd4ca3bf87d0ad31f9818e22188/triton_windows-3.6.0.post25-cp310-cp310-win_amd64.whl", hash = "sha256:8c45b7f83eecb71c3aeded1da7914af0050bddda710f47a2cfae936d55fae0ca", size = 47380144, upload-time = "2026-01-26T03:21:14.942Z" },
+]
+
+[[package]]
+name = "trl"
+version = "0.16.1"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "rich", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/09/de/ba9095a2c5b9150921802727323356d47d755e62ea700dd6788eae734262/trl-0.16.1.tar.gz", hash = "sha256:166acab64811acfbaaa3839688547d78fec806f9b12b99497b168431e13968be", size = 353198, upload-time = "2025-04-04T18:21:49.62Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a5/e0/02e129c156d38a408274276def68524fd080b1242780cc23d7e261571157/trl-0.16.1-py3-none-any.whl", hash = "sha256:2b893bbed77dce79c88d6e864ce1216641f6587f4a28988281befba1d2a13ae9", size = 336439, upload-time = "2025-04-04T18:21:47.824Z" },
+]
+
+[[package]]
+name = "trl"
+version = "0.20.0"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/60/11/95cf1210df9f241b7b1084abe1032e322374f667c4587c09af8d14a1d76f/trl-0.20.0.tar.gz", hash = "sha256:3f949b009b79dc609cd8f5469d67209ab8f71c5cb4d8d979f7b568ef054922fa", size = 461791, upload-time = "2025-07-29T04:10:06.305Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5c/25/fb8509e9206a2ebfa89f9e1c1b595c1c7d68852716ac64f9c4845aa4133e/trl-0.20.0-py3-none-any.whl", hash = "sha256:925703a4b6117c45c64625768669e0cccb2cc71011b3d52163e47c0e7d6eb8af", size = 504561, upload-time = "2025-07-29T04:10:04.697Z" },
+]
+
+[[package]]
+name = "typeguard"
+version = "4.5.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/2b/e8/66e25efcc18542d58706ce4e50415710593721aae26e794ab1dec34fb66f/typeguard-4.5.1.tar.gz", hash = "sha256:f6f8ecbbc819c9bc749983cc67c02391e16a9b43b8b27f15dc70ed7c4a007274", size = 80121, upload-time = "2026-02-19T16:09:03.392Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/91/88/b55b3117287a8540b76dbdd87733808d4d01c8067a3b339408c250bb3600/typeguard-4.5.1-py3-none-any.whl", hash = "sha256:44d2bf329d49a244110a090b55f5f91aa82d9a9834ebfd30bcc73651e4a8cc40", size = 36745, upload-time = "2026-02-19T16:09:01.6Z" },
+]
+
+[[package]]
+name = "typer"
+version = "0.24.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "annotated-doc" },
+    { name = "click" },
+    { name = "rich" },
+    { name = "shellingham" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f5/24/cb09efec5cc954f7f9b930bf8279447d24618bb6758d4f6adf2574c41780/typer-0.24.1.tar.gz", hash = "sha256:e39b4732d65fbdcde189ae76cf7cd48aeae72919dea1fdfc16593be016256b45", size = 118613, upload-time = "2026-02-21T16:54:40.609Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4a/91/48db081e7a63bb37284f9fbcefda7c44c277b18b0e13fbc36ea2335b71e6/typer-0.24.1-py3-none-any.whl", hash = "sha256:112c1f0ce578bfb4cab9ffdabc68f031416ebcc216536611ba21f04e9aa84c9e", size = 56085, upload-time = "2026-02-21T16:54:41.616Z" },
+]
+
+[[package]]
+name = "typing-extensions"
+version = "4.15.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/72/94/1a15dd82efb362ac84269196e94cf00f187f7ed21c242792a923cdb1c61f/typing_extensions-4.15.0.tar.gz", hash = "sha256:0cea48d173cc12fa28ecabc3b837ea3cf6f38c6d1136f85cbaaf598984861466", size = 109391, upload-time = "2025-08-25T13:49:26.313Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/18/67/36e9267722cc04a6b9f15c7f3441c2363321a3ea07da7ae0c0707beb2a9c/typing_extensions-4.15.0-py3-none-any.whl", hash = "sha256:f0fa19c6845758ab08074a0cfa8b7aecb71c999ca73d62883bc25cc018c4e548", size = 44614, upload-time = "2025-08-25T13:49:24.86Z" },
+]
+
+[[package]]
+name = "typing-inspect"
+version = "0.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "mypy-extensions" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/dc/74/1789779d91f1961fa9438e9a8710cdae6bd138c80d7303996933d117264a/typing_inspect-0.9.0.tar.gz", hash = "sha256:b23fc42ff6f6ef6954e4852c1fb512cdd18dbea03134f91f856a95ccc9461f78", size = 13825, upload-time = "2023-05-24T20:25:47.612Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/65/f3/107a22063bf27bdccf2024833d3445f4eea42b2e598abfbd46f6a63b6cb0/typing_inspect-0.9.0-py3-none-any.whl", hash = "sha256:9ee6fc59062311ef8547596ab6b955e1b8aa46242d854bfc78f4f6b0eff35f9f", size = 8827, upload-time = "2023-05-24T20:25:45.287Z" },
+]
+
+[[package]]
+name = "typing-inspection"
+version = "0.4.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/55/e3/70399cb7dd41c10ac53367ae42139cf4b1ca5f36bb3dc6c9d33acdb43655/typing_inspection-0.4.2.tar.gz", hash = "sha256:ba561c48a67c5958007083d386c3295464928b01faa735ab8547c5692e87f464", size = 75949, upload-time = "2025-10-01T02:14:41.687Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/dc/9b/47798a6c91d8bdb567fe2698fe81e0c6b7cb7ef4d13da4114b41d239f65d/typing_inspection-0.4.2-py3-none-any.whl", hash = "sha256:4ed1cacbdc298c220f1bd249ed5287caa16f34d44ef4e9c3d0cbad5b521545e7", size = 14611, upload-time = "2025-10-01T02:14:40.154Z" },
+]
+
+[[package]]
+name = "tyro"
+version = "1.0.8"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "docstring-parser" },
+    { name = "typeguard" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/9e/89/f4d76ba5ac4ca627c3a0413fdc8d90f202929c45d9c25deec29c9a86a0ae/tyro-1.0.8.tar.gz", hash = "sha256:c22bf238ce029b8cc262759129fccfd968300c3761f00ce186570af3891e14bb", size = 460626, upload-time = "2026-02-25T00:18:45.388Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/61/7c/9c5a4ab9e24f1428fd5ab2728b7f68878f54d549a11fa431ead76d3d127b/tyro-1.0.8-py3-none-any.whl", hash = "sha256:abb6a4054f20616d9ca5181ba0c6ae0e35cf8f35ae0c78d3d4dbd1f678e0643d", size = 181935, upload-time = "2026-02-25T00:18:43.689Z" },
+]
+
+[[package]]
+name = "tzdata"
+version = "2025.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/5e/a7/c202b344c5ca7daf398f3b8a477eeb205cf3b6f32e7ec3a6bac0629ca975/tzdata-2025.3.tar.gz", hash = "sha256:de39c2ca5dc7b0344f2eba86f49d614019d29f060fc4ebc8a417896a620b56a7", size = 196772, upload-time = "2025-12-13T17:45:35.667Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c7/b0/003792df09decd6849a5e39c28b513c06e84436a54440380862b5aeff25d/tzdata-2025.3-py2.py3-none-any.whl", hash = "sha256:06a47e5700f3081aab02b2e513160914ff0694bce9947d6b76ebd6bf57cfc5d1", size = 348521, upload-time = "2025-12-13T17:45:33.889Z" },
+]
+
+[[package]]
+name = "unsloth"
+version = "2025.10.8"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "bitsandbytes", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "diffusers", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "hf-transfer", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "peft", version = "0.15.2", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "protobuf", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "sentencepiece", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "triton", marker = "('linux' in sys_platform and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "triton-windows", marker = "(platform_machine != 'AMD64' and platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform != 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "trl", version = "0.16.1", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "tyro", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "unsloth-zoo", version = "2025.10.9", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "wheel", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "xformers", marker = "(platform_machine == 'AMD64' and 'linux' in sys_platform and extra == 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and 'linux' in sys_platform and extra == 'extra-9-robometer-vlac') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ae/c6/cdeec138f6640bca687337291c90636ef7a43cf1a7147438f22d082a27fe/unsloth-2025.10.8.tar.gz", hash = "sha256:a35abeaf780c0bcfab7ee3f07a1fe3771f98b1a732ca937b71f3dffc7c929285", size = 4742041, upload-time = "2025-10-22T14:59:35.934Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c0/e2/65bebfc7e5881fd2531d3613b6a3585cb1262cad9bcd0d084597bc9b4159/unsloth-2025.10.8-py3-none-any.whl", hash = "sha256:0f86570c024d865493f05b4754fb2188c32da2132619c66f8b85b5c4e7a17428", size = 347307, upload-time = "2025-10-22T14:59:29.061Z" },
+]
+
+[[package]]
+name = "unsloth"
+version = "2026.3.4"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "bitsandbytes", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "diffusers", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "hf-transfer", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "peft", version = "0.18.1", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "protobuf", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "sentencepiece", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform != 'linux' and sys_platform != 'win32' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'linux' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform != 'linux' and sys_platform != 'win32' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchvision", version = "0.23.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'linux' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "triton", marker = "('linux' in sys_platform and extra == 'extra-9-robometer-robometer') or ('linux' in sys_platform and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "triton-windows", marker = "(platform_machine != 'AMD64' and platform_machine != 'x86_64' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (sys_platform != 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "trl", version = "0.20.0", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "tyro", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "unsloth-zoo", version = "2026.3.2", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "wheel", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "xformers", marker = "(platform_machine == 'AMD64' and 'linux' in sys_platform and extra == 'extra-9-robometer-robometer') or (platform_machine == 'AMD64' and 'linux' in sys_platform and extra != 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and 'linux' in sys_platform and extra == 'extra-9-robometer-robometer') or (platform_machine == 'x86_64' and 'linux' in sys_platform and extra != 'extra-9-robometer-vlac') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (platform_machine == 'AMD64' and sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (platform_machine == 'x86_64' and sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/32/1e/978b8524f0153768340cf23a0b3e29941f86b3e6dc762a6c0b2fb3cdebb0/unsloth-2026.3.4.tar.gz", hash = "sha256:7b260d4a66fe492e63509ccfbd279c1f8824e26426433e38a603f119bc0c9054", size = 4855549, upload-time = "2026-03-08T13:41:58.844Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3e/84/0de034f96b67328a7674aca872e803072c5d27de4569cfe0c3e5894fb17a/unsloth-2026.3.4-py3-none-any.whl", hash = "sha256:f5d40405e61d23018f61c1fd44c19b28b0a4a3c95c84c594c4d3313735db2c4d", size = 447154, upload-time = "2026-03-08T13:41:45.976Z" },
+]
+
+[[package]]
+name = "unsloth-zoo"
+version = "2025.10.9"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "cut-cross-entropy", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "filelock", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "hf-transfer", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "msgspec", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "peft", version = "0.15.2", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "pillow", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "protobuf", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "regex", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "sentencepiece", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchao", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.51.3", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "triton", marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "triton-windows", marker = "(sys_platform == 'win32' and extra == 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "trl", version = "0.16.1", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "typing-extensions", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "tyro", marker = "extra == 'extra-9-robometer-vlac'" },
+    { name = "wheel", marker = "extra == 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7c/e9/c5e1ffd3a8cb5516387c72f86a5aa5b4f0536b6f9e45a899ce83406a3fcf/unsloth_zoo-2025.10.9.tar.gz", hash = "sha256:dbfb3094068a6dd488a8c786a0c681a5b6a26167e49ca990c1a04aebe1304cb2", size = 250615, upload-time = "2025-10-22T14:57:17.069Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/75/bd/984a148b147881d7833ae67bd1dc05b730b904833d5446f4b25d74bd64b5/unsloth_zoo-2025.10.9-py3-none-any.whl", hash = "sha256:06c1642af57cca5ec9b022d58972be1b41fc644f88efbd70b06c11471e75c54f", size = 269385, upload-time = "2025-10-22T14:57:15.389Z" },
+]
+
+[[package]]
+name = "unsloth-zoo"
+version = "2026.3.2"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+    "platform_machine == 'aarch64' and sys_platform == 'linux'",
+    "platform_machine != 'aarch64' and sys_platform == 'linux'",
+    "sys_platform == 'win32'",
+    "sys_platform == 'darwin'",
+    "sys_platform != 'darwin' and sys_platform != 'linux' and sys_platform != 'win32'",
+]
+dependencies = [
+    { name = "accelerate", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "cut-cross-entropy", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "datasets", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "filelock", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "hf-transfer", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "huggingface-hub", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "msgspec", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "numpy", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "packaging", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "peft", version = "0.18.1", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "pillow", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "protobuf", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "psutil", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "regex", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "sentencepiece", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform != 'linux' and sys_platform != 'win32' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "(sys_platform == 'linux' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'linux' and extra != 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer') or (sys_platform == 'win32' and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torchao", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "tqdm", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "transformers", version = "4.57.6", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "triton", marker = "('linux' in sys_platform and extra == 'extra-9-robometer-robometer') or ('linux' in sys_platform and extra != 'extra-9-robometer-vlac') or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "trl", version = "0.20.0", source = { registry = "https://pypi.org/simple" }, marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "typing-extensions", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "tyro", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+    { name = "wheel", marker = "extra == 'extra-9-robometer-robometer' or extra != 'extra-9-robometer-vlac'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/9a/c9/771bfbc1a533a2434803b8b28ed2c9ff287d5cd504989917263e473cb724/unsloth_zoo-2026.3.2.tar.gz", hash = "sha256:4f4516d07079da020605ddc5fa2a1cbe62e290f5502d68c0b2ded20f60223845", size = 365174, upload-time = "2026-03-08T13:39:45.873Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2a/de/c35948382df1a56bb172d99fd6ee5007a79820bf747b7f5bb81523292698/unsloth_zoo-2026.3.2-py3-none-any.whl", hash = "sha256:d356615733a2e0070a50ea8e6c07efce1a485676c11c2314fb5cc4b6ea2568e7", size = 395174, upload-time = "2026-03-08T13:39:44.228Z" },
+]
+
+[[package]]
+name = "uritemplate"
+version = "4.2.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/98/60/f174043244c5306c9988380d2cb10009f91563fc4b31293d27e17201af56/uritemplate-4.2.0.tar.gz", hash = "sha256:480c2ed180878955863323eea31b0ede668795de182617fef9c6ca09e6ec9d0e", size = 33267, upload-time = "2025-06-02T15:12:06.318Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a9/99/3ae339466c9183ea5b8ae87b34c0b897eda475d2aec2307cae60e5cd4f29/uritemplate-4.2.0-py3-none-any.whl", hash = "sha256:962201ba1c4edcab02e60f9a0d3821e82dfc5d2d6662a21abd533879bdb8a686", size = 11488, upload-time = "2025-06-02T15:12:03.405Z" },
+]
+
+[[package]]
+name = "urllib3"
+version = "2.6.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/c7/24/5f1b3bdffd70275f6661c76461e25f024d5a38a46f04aaca912426a2b1d3/urllib3-2.6.3.tar.gz", hash = "sha256:1b62b6884944a57dbe321509ab94fd4d3b307075e0c2eae991ac71ee15ad38ed", size = 435556, upload-time = "2026-01-07T16:24:43.925Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
+]
+
+[[package]]
+name = "uvicorn"
+version = "0.41.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "h11" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/32/ce/eeb58ae4ac36fe09e3842eb02e0eb676bf2c53ae062b98f1b2531673efdd/uvicorn-0.41.0.tar.gz", hash = "sha256:09d11cf7008da33113824ee5a1c6422d89fbc2ff476540d69a34c87fab8b571a", size = 82633, upload-time = "2026-02-16T23:07:24.1Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/83/e4/d04a086285c20886c0daad0e026f250869201013d18f81d9ff5eada73a88/uvicorn-0.41.0-py3-none-any.whl", hash = "sha256:29e35b1d2c36a04b9e180d4007ede3bcb32a85fbdfd6c6aeb3f26839de088187", size = 68783, upload-time = "2026-02-16T23:07:22.357Z" },
+]
+
+[[package]]
+name = "virtualenv"
+version = "21.1.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "distlib" },
+    { name = "filelock" },
+    { name = "platformdirs" },
+    { name = "python-discovery" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/2f/c9/18d4b36606d6091844daa3bd93cf7dc78e6f5da21d9f21d06c221104b684/virtualenv-21.1.0.tar.gz", hash = "sha256:1990a0188c8f16b6b9cf65c9183049007375b26aad415514d377ccacf1e4fb44", size = 5840471, upload-time = "2026-02-27T08:49:29.702Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/78/55/896b06bf93a49bec0f4ae2a6f1ed12bd05c8860744ac3a70eda041064e4d/virtualenv-21.1.0-py3-none-any.whl", hash = "sha256:164f5e14c5587d170cf98e60378eb91ea35bf037be313811905d3a24ea33cc07", size = 5825072, upload-time = "2026-02-27T08:49:27.516Z" },
+]
+
+[[package]]
+name = "wandb"
+version = "0.25.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "gitpython" },
+    { name = "packaging" },
+    { name = "platformdirs" },
+    { name = "protobuf" },
+    { name = "pydantic" },
+    { name = "pyyaml" },
+    { name = "requests" },
+    { name = "sentry-sdk" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/fd/60/d94952549920469524b689479c864c692ca47eca4b8c2fe3389b64a58778/wandb-0.25.0.tar.gz", hash = "sha256:45840495a288e34245d69d07b5a0b449220fbc5b032e6b51c4f92ec9026d2ad1", size = 43951335, upload-time = "2026-02-13T00:17:45.515Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c1/7d/0c131db3ec9deaabbd32263d90863cbfbe07659527e11c35a5c738cecdc5/wandb-0.25.0-py3-none-macosx_12_0_arm64.whl", hash = "sha256:5eecb3c7b5e60d1acfa4b056bfbaa0b79a482566a9db58c9f99724b3862bc8e5", size = 23287536, upload-time = "2026-02-13T00:17:20.265Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/95/31bb7f76a966ec87495e5a72ac7570685be162494c41757ac871768dbc4f/wandb-0.25.0-py3-none-macosx_12_0_x86_64.whl", hash = "sha256:daeedaadb183dc466e634fba90ab2bab1d4e93000912be0dee95065a0624a3fd", size = 25196062, upload-time = "2026-02-13T00:17:23.356Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/a1/258cdedbf30cebc692198a774cf0ef945b7ed98ee64bdaf62621281c95d8/wandb-0.25.0-py3-none-manylinux_2_28_aarch64.whl", hash = "sha256:5e0127dbcef13eea48f4b84268da7004d34d3120ebc7b2fa9cefb72b49dbb825", size = 22799744, upload-time = "2026-02-13T00:17:26.437Z" },
+    { url = "https://files.pythonhosted.org/packages/de/91/ec9465d014cfd199c5b2083d271d31b3c2aedeae66f3d8a0712f7f54bdf3/wandb-0.25.0-py3-none-manylinux_2_28_x86_64.whl", hash = "sha256:6c4c38077836f9b7569a35b0e1dcf1f0c43616fcd936d182f475edbfea063665", size = 25262839, upload-time = "2026-02-13T00:17:28.8Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/95/cb2d1c7143f534544147fb53fe87944508b8cb9a058bc5b6f8a94adbee15/wandb-0.25.0-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:6edd8948d305cb73745bf564b807bd73da2ccbd47c548196b8a362f7df40aed8", size = 22853714, upload-time = "2026-02-13T00:17:31.68Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/94/68163f70c1669edcf130822aaaea782d8198b5df74443eca0085ec596774/wandb-0.25.0-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:ada6f08629bb014ad6e0a19d5dec478cdaa116431baa3f0a4bf4ab8d9893611f", size = 25358037, upload-time = "2026-02-13T00:17:34.676Z" },
+    { url = "https://files.pythonhosted.org/packages/cc/fb/9578eed2c01b2fc6c8b693da110aa9c73a33d7bb556480f5cfc42e48c94e/wandb-0.25.0-py3-none-win32.whl", hash = "sha256:020b42ca4d76e347709d65f59b30d4623a115edc28f462af1c92681cb17eae7c", size = 24604118, upload-time = "2026-02-13T00:17:37.641Z" },
+    { url = "https://files.pythonhosted.org/packages/25/97/460f6cb738aaa39b4eb2e6b4c630b2ae4321cdd70a79d5955ea75a878981/wandb-0.25.0-py3-none-win_amd64.whl", hash = "sha256:78307ac0b328f2dc334c8607bec772851215584b62c439eb320c4af4fb077a00", size = 24604122, upload-time = "2026-02-13T00:17:39.991Z" },
+    { url = "https://files.pythonhosted.org/packages/27/6c/5847b4dda1dfd52630dac08711d4348c69ed657f0698fc2d949c7f7a6622/wandb-0.25.0-py3-none-win_arm64.whl", hash = "sha256:c6174401fd6fb726295e98d57b4231c100eca96bd17de51bfc64038a57230aaf", size = 21785298, upload-time = "2026-02-13T00:17:42.475Z" },
+]
+
+[package.optional-dependencies]
+media = [
+    { name = "bokeh" },
+    { name = "imageio" },
+    { name = "moviepy" },
+    { name = "numpy" },
+    { name = "pillow" },
+    { name = "plotly" },
+    { name = "rdkit" },
+    { name = "soundfile" },
+]
+
+[[package]]
+name = "wcwidth"
+version = "0.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/35/a2/8e3becb46433538a38726c948d3399905a4c7cabd0df578ede5dc51f0ec2/wcwidth-0.6.0.tar.gz", hash = "sha256:cdc4e4262d6ef9a1a57e018384cbeb1208d8abbc64176027e2c2455c81313159", size = 159684, upload-time = "2026-02-06T19:19:40.919Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/68/5a/199c59e0a824a3db2b89c5d2dade7ab5f9624dbf6448dc291b46d5ec94d3/wcwidth-0.6.0-py3-none-any.whl", hash = "sha256:1a3a1e510b553315f8e146c54764f4fb6264ffad731b3d78088cdb1478ffbdad", size = 94189, upload-time = "2026-02-06T19:19:39.646Z" },
+]
+
+[[package]]
+name = "werkzeug"
+version = "3.1.6"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "markupsafe" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/61/f1/ee81806690a87dab5f5653c1f146c92bc066d7f4cebc603ef88eb9e13957/werkzeug-3.1.6.tar.gz", hash = "sha256:210c6bede5a420a913956b4791a7f4d6843a43b6fcee4dfa08a65e93007d0d25", size = 864736, upload-time = "2026-02-19T15:17:18.884Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4d/ec/d58832f89ede95652fd01f4f24236af7d32b70cab2196dfcc2d2fd13c5c2/werkzeug-3.1.6-py3-none-any.whl", hash = "sha256:7ddf3357bb9564e407607f988f683d72038551200c704012bb9a4c523d42f131", size = 225166, upload-time = "2026-02-19T15:17:17.475Z" },
+]
+
+[[package]]
+name = "wheel"
+version = "0.46.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "packaging" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/89/24/a2eb353a6edac9a0303977c4cb048134959dd2a51b48a269dfc9dde00c8a/wheel-0.46.3.tar.gz", hash = "sha256:e3e79874b07d776c40bd6033f8ddf76a7dad46a7b8aa1b2787a83083519a1803", size = 60605, upload-time = "2026-01-22T12:39:49.136Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/87/22/b76d483683216dde3d67cba61fb2444be8d5be289bf628c13fc0fd90e5f9/wheel-0.46.3-py3-none-any.whl", hash = "sha256:4b399d56c9d9338230118d705d9737a2a468ccca63d5e813e2a4fc7815d8bc4d", size = 30557, upload-time = "2026-01-22T12:39:48.099Z" },
+]
+
+[[package]]
+name = "widgetsnbextension"
+version = "4.0.15"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/bd/f4/c67440c7fb409a71b7404b7aefcd7569a9c0d6bd071299bf4198ae7a5d95/widgetsnbextension-4.0.15.tar.gz", hash = "sha256:de8610639996f1567952d763a5a41af8af37f2575a41f9852a38f947eb82a3b9", size = 1097402, upload-time = "2025-11-01T21:15:55.178Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3f/0e/fa3b193432cfc60c93b42f3be03365f5f909d2b3ea410295cf36df739e31/widgetsnbextension-4.0.15-py3-none-any.whl", hash = "sha256:8156704e4346a571d9ce73b84bee86a29906c9abfd7223b7228a28899ccf3366", size = 2196503, upload-time = "2025-11-01T21:15:53.565Z" },
+]
+
+[[package]]
+name = "win32-setctime"
+version = "1.2.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b3/8f/705086c9d734d3b663af0e9bb3d4de6578d08f46b1b101c2442fd9aecaa2/win32_setctime-1.2.0.tar.gz", hash = "sha256:ae1fdf948f5640aae05c511ade119313fb6a30d7eabe25fef9764dca5873c4c0", size = 4867, upload-time = "2024-12-07T15:28:28.314Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e1/07/c6fe3ad3e685340704d314d765b7912993bcb8dc198f0e7a89382d37974b/win32_setctime-1.2.0-py3-none-any.whl", hash = "sha256:95d644c4e708aba81dc3704a116d8cbc974d70b3bdb8be1d150e36be6e9d1390", size = 4083, upload-time = "2024-12-07T15:28:26.465Z" },
+]
+
+[[package]]
+name = "wrapt"
+version = "2.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2e/64/925f213fdcbb9baeb1530449ac71a4d57fc361c053d06bf78d0c5c7cd80c/wrapt-2.1.2.tar.gz", hash = "sha256:3996a67eecc2c68fd47b4e3c564405a5777367adfd9b8abb58387b63ee83b21e", size = 81678, upload-time = "2026-03-06T02:53:25.134Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/da/d2/387594fb592d027366645f3d7cc9b4d7ca7be93845fbaba6d835a912ef3c/wrapt-2.1.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:4b7a86d99a14f76facb269dc148590c01aaf47584071809a70da30555228158c", size = 60669, upload-time = "2026-03-06T02:52:40.671Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/18/3f373935bc5509e7ac444c8026a56762e50c1183e7061797437ca96c12ce/wrapt-2.1.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a819e39017f95bf7aede768f75915635aa8f671f2993c036991b8d3bfe8dbb6f", size = 61603, upload-time = "2026-03-06T02:54:21.032Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/7a/32758ca2853b07a887a4574b74e28843919103194bb47001a304e24af62f/wrapt-2.1.2-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:5681123e60aed0e64c7d44f72bbf8b4ce45f79d81467e2c4c728629f5baf06eb", size = 113632, upload-time = "2026-03-06T02:53:54.121Z" },
+    { url = "https://files.pythonhosted.org/packages/1d/d5/eeaa38f670d462e97d978b3b0d9ce06d5b91e54bebac6fbed867809216e7/wrapt-2.1.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2b8b28e97a44d21836259739ae76284e180b18abbb4dcfdff07a415cf1016c3e", size = 115644, upload-time = "2026-03-06T02:54:53.33Z" },
+    { url = "https://files.pythonhosted.org/packages/e3/09/2a41506cb17affb0bdf9d5e2129c8c19e192b388c4c01d05e1b14db23c00/wrapt-2.1.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:cef91c95a50596fcdc31397eb6955476f82ae8a3f5a8eabdc13611b60ee380ba", size = 112016, upload-time = "2026-03-06T02:54:43.274Z" },
+    { url = "https://files.pythonhosted.org/packages/64/15/0e6c3f5e87caadc43db279724ee36979246d5194fa32fed489c73643ba59/wrapt-2.1.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:dad63212b168de8569b1c512f4eac4b57f2c6934b30df32d6ee9534a79f1493f", size = 114823, upload-time = "2026-03-06T02:54:29.392Z" },
+    { url = "https://files.pythonhosted.org/packages/56/b2/0ad17c8248f4e57bedf44938c26ec3ee194715f812d2dbbd9d7ff4be6c06/wrapt-2.1.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:d307aa6888d5efab2c1cde09843d48c843990be13069003184b67d426d145394", size = 111244, upload-time = "2026-03-06T02:54:02.149Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/04/bcdba98c26f2c6522c7c09a726d5d9229120163493620205b2f76bd13c01/wrapt-2.1.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c87cf3f0c85e27b3ac7d9ad95da166bf8739ca215a8b171e8404a2d739897a45", size = 113307, upload-time = "2026-03-06T02:54:12.428Z" },
+    { url = "https://files.pythonhosted.org/packages/0e/1b/5e2883c6bc14143924e465a6fc5a92d09eeabe35310842a481fb0581f832/wrapt-2.1.2-cp310-cp310-win32.whl", hash = "sha256:d1c5fea4f9fe3762e2b905fdd67df51e4be7a73b7674957af2d2ade71a5c075d", size = 57986, upload-time = "2026-03-06T02:54:26.823Z" },
+    { url = "https://files.pythonhosted.org/packages/42/5a/4efc997bccadd3af5749c250b49412793bc41e13a83a486b2b54a33e240c/wrapt-2.1.2-cp310-cp310-win_amd64.whl", hash = "sha256:d8f7740e1af13dff2684e4d56fe604a7e04d6c94e737a60568d8d4238b9a0c71", size = 60336, upload-time = "2026-03-06T02:54:18Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/f5/a2bb833e20181b937e87c242645ed5d5aa9c373006b0467bfe1a35c727d0/wrapt-2.1.2-cp310-cp310-win_arm64.whl", hash = "sha256:1c6cc827c00dc839350155f316f1f8b4b0c370f52b6a19e782e2bda89600c7dc", size = 58757, upload-time = "2026-03-06T02:53:51.545Z" },
+    { url = "https://files.pythonhosted.org/packages/1a/c7/8528ac2dfa2c1e6708f647df7ae144ead13f0a31146f43c7264b4942bf12/wrapt-2.1.2-py3-none-any.whl", hash = "sha256:b8fd6fa2b2c4e7621808f8c62e8317f4aae56e59721ad933bac5239d913cf0e8", size = 43993, upload-time = "2026-03-06T02:53:12.905Z" },
+]
+
+[[package]]
+name = "xformers"
+version = "0.0.32.post2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "torch", version = "2.8.0", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform != 'linux' and sys_platform != 'win32') or (sys_platform == 'linux' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac') or (sys_platform == 'win32' and extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+    { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32' or (extra == 'extra-9-robometer-robometer' and extra == 'extra-9-robometer-vlac')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/47/99/68d2e7a17b65a08180d6b120d1d293d5855a0c5999cb30a79fb466b49642/xformers-0.0.32.post2.tar.gz", hash = "sha256:9538be803969c6e1ca16a3ece921e472c24f79970b10be1087a389dcb66e412a", size = 12105990, upload-time = "2025-08-15T11:48:56.584Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ce/16/4ed2fbf7c20ba20565d9a1fc544c2bef39d409c8eca1949654c6ff9f8452/xformers-0.0.32.post2-cp39-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:3d1fefe0006eb00a730b3b197ff41aab085f1209b50419baab6152445f09e508", size = 117196673, upload-time = "2025-08-15T11:48:46.726Z" },
+    { url = "https://files.pythonhosted.org/packages/50/19/ea64609a535f87a44cb177dc37c6d6a58f6d540adfff819d565fb0657cf7/xformers-0.0.32.post2-cp39-abi3-win_amd64.whl", hash = "sha256:87fa13f4b406ed640554cea6687d0020b496bd6b446ca17d24f9438a079e6f09", size = 100221079, upload-time = "2025-08-15T11:48:53.234Z" },
+]
+
+[[package]]
+name = "xxhash"
+version = "3.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/02/84/30869e01909fb37a6cc7e18688ee8bf1e42d57e7e0777636bd47524c43c7/xxhash-3.6.0.tar.gz", hash = "sha256:f0162a78b13a0d7617b2845b90c763339d1f1d82bb04a4b07f4ab535cc5e05d6", size = 85160, upload-time = "2025-10-02T14:37:08.097Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/34/ee/f9f1d656ad168681bb0f6b092372c1e533c4416b8069b1896a175c46e484/xxhash-3.6.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:87ff03d7e35c61435976554477a7f4cd1704c3596a89a8300d5ce7fc83874a71", size = 32845, upload-time = "2025-10-02T14:33:51.573Z" },
+    { url = "https://files.pythonhosted.org/packages/a3/b1/93508d9460b292c74a09b83d16750c52a0ead89c51eea9951cb97a60d959/xxhash-3.6.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:f572dfd3d0e2eb1a57511831cf6341242f5a9f8298a45862d085f5b93394a27d", size = 30807, upload-time = "2025-10-02T14:33:52.964Z" },
+    { url = "https://files.pythonhosted.org/packages/07/55/28c93a3662f2d200c70704efe74aab9640e824f8ce330d8d3943bf7c9b3c/xxhash-3.6.0-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:89952ea539566b9fed2bbd94e589672794b4286f342254fad28b149f9615fef8", size = 193786, upload-time = "2025-10-02T14:33:54.272Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/96/fec0be9bb4b8f5d9c57d76380a366f31a1781fb802f76fc7cda6c84893c7/xxhash-3.6.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:48e6f2ffb07a50b52465a1032c3cf1f4a5683f944acaca8a134a2f23674c2058", size = 212830, upload-time = "2025-10-02T14:33:55.706Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/a0/c706845ba77b9611f81fd2e93fad9859346b026e8445e76f8c6fd057cc6d/xxhash-3.6.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b5b848ad6c16d308c3ac7ad4ba6bede80ed5df2ba8ed382f8932df63158dd4b2", size = 211606, upload-time = "2025-10-02T14:33:57.133Z" },
+    { url = "https://files.pythonhosted.org/packages/67/1e/164126a2999e5045f04a69257eea946c0dc3e86541b400d4385d646b53d7/xxhash-3.6.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a034590a727b44dd8ac5914236a7b8504144447a9682586c3327e935f33ec8cc", size = 444872, upload-time = "2025-10-02T14:33:58.446Z" },
+    { url = "https://files.pythonhosted.org/packages/2d/4b/55ab404c56cd70a2cf5ecfe484838865d0fea5627365c6c8ca156bd09c8f/xxhash-3.6.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8a8f1972e75ebdd161d7896743122834fe87378160c20e97f8b09166213bf8cc", size = 193217, upload-time = "2025-10-02T14:33:59.724Z" },
+    { url = "https://files.pythonhosted.org/packages/45/e6/52abf06bac316db33aa269091ae7311bd53cfc6f4b120ae77bac1b348091/xxhash-3.6.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ee34327b187f002a596d7b167ebc59a1b729e963ce645964bbc050d2f1b73d07", size = 210139, upload-time = "2025-10-02T14:34:02.041Z" },
+    { url = "https://files.pythonhosted.org/packages/34/37/db94d490b8691236d356bc249c08819cbcef9273a1a30acf1254ff9ce157/xxhash-3.6.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:339f518c3c7a850dd033ab416ea25a692759dc7478a71131fe8869010d2b75e4", size = 197669, upload-time = "2025-10-02T14:34:03.664Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/36/c4f219ef4a17a4f7a64ed3569bc2b5a9c8311abdb22249ac96093625b1a4/xxhash-3.6.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:bf48889c9630542d4709192578aebbd836177c9f7a4a2778a7d6340107c65f06", size = 210018, upload-time = "2025-10-02T14:34:05.325Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/06/bfac889a374fc2fc439a69223d1750eed2e18a7db8514737ab630534fa08/xxhash-3.6.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:5576b002a56207f640636056b4160a378fe36a58db73ae5c27a7ec8db35f71d4", size = 413058, upload-time = "2025-10-02T14:34:06.925Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/d1/555d8447e0dd32ad0930a249a522bb2e289f0d08b6b16204cfa42c1f5a0c/xxhash-3.6.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:af1f3278bd02814d6dedc5dec397993b549d6f16c19379721e5a1d31e132c49b", size = 190628, upload-time = "2025-10-02T14:34:08.669Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/15/8751330b5186cedc4ed4b597989882ea05e0408b53fa47bcb46a6125bfc6/xxhash-3.6.0-cp310-cp310-win32.whl", hash = "sha256:aed058764db109dc9052720da65fafe84873b05eb8b07e5e653597951af57c3b", size = 30577, upload-time = "2025-10-02T14:34:10.234Z" },
+    { url = "https://files.pythonhosted.org/packages/bb/cc/53f87e8b5871a6eb2ff7e89c48c66093bda2be52315a8161ddc54ea550c4/xxhash-3.6.0-cp310-cp310-win_amd64.whl", hash = "sha256:e82da5670f2d0d98950317f82a0e4a0197150ff19a6df2ba40399c2a3b9ae5fb", size = 31487, upload-time = "2025-10-02T14:34:11.618Z" },
+    { url = "https://files.pythonhosted.org/packages/9f/00/60f9ea3bb697667a14314d7269956f58bf56bb73864f8f8d52a3c2535e9a/xxhash-3.6.0-cp310-cp310-win_arm64.whl", hash = "sha256:4a082ffff8c6ac07707fb6b671caf7c6e020c75226c561830b73d862060f281d", size = 27863, upload-time = "2025-10-02T14:34:12.619Z" },
+]
+
+[[package]]
+name = "xyzservices"
+version = "2025.11.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ee/0f/022795fc1201e7c29e742a509913badb53ce0b38f64b6db859e2f6339da9/xyzservices-2025.11.0.tar.gz", hash = "sha256:2fc72b49502b25023fd71e8f532fb4beddbbf0aa124d90ea25dba44f545e17ce", size = 1135703, upload-time = "2025-11-22T11:31:51.82Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ef/5c/2c189d18d495dd0fa3f27ccc60762bbc787eed95b9b0147266e72bb76585/xyzservices-2025.11.0-py3-none-any.whl", hash = "sha256:de66a7599a8d6dad63980b77defd1d8f5a5a9cb5fc8774ea1c6e89ca7c2a3d2f", size = 93916, upload-time = "2025-11-22T11:31:50.525Z" },
+]
+
+[[package]]
+name = "yarl"
+version = "1.23.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "idna" },
+    { name = "multidict" },
+    { name = "propcache" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/23/6e/beb1beec874a72f23815c1434518bfc4ed2175065173fb138c3705f658d4/yarl-1.23.0.tar.gz", hash = "sha256:53b1ea6ca88ebd4420379c330aea57e258408dd0df9af0992e5de2078dc9f5d5", size = 194676, upload-time = "2026-03-01T22:07:53.373Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8b/0d/9cc638702f6fc3c7a3685bcc8cf2a9ed7d6206e932a49f5242658047ef51/yarl-1.23.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:cff6d44cb13d39db2663a22b22305d10855efa0fa8015ddeacc40bc59b9d8107", size = 123764, upload-time = "2026-03-01T22:04:09.7Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/35/5a553687c5793df5429cd1db45909d4f3af7eee90014888c208d086a44f0/yarl-1.23.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e4c53f8347cd4200f0d70a48ad059cabaf24f5adc6ba08622a23423bc7efa10d", size = 86282, upload-time = "2026-03-01T22:04:11.892Z" },
+    { url = "https://files.pythonhosted.org/packages/68/2e/c5a2234238f8ce37a8312b52801ee74117f576b1539eec8404a480434acc/yarl-1.23.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2a6940a074fb3c48356ed0158a3ca5699c955ee4185b4d7d619be3c327143e05", size = 86053, upload-time = "2026-03-01T22:04:13.292Z" },
+    { url = "https://files.pythonhosted.org/packages/74/3f/bbd8ff36fb038622797ffbaf7db314918bb4d76f1cc8a4f9ca7a55fe5195/yarl-1.23.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ed5f69ce7be7902e5c70ea19eb72d20abf7d725ab5d49777d696e32d4fc1811d", size = 99395, upload-time = "2026-03-01T22:04:15.133Z" },
+    { url = "https://files.pythonhosted.org/packages/77/04/9516bc4e269d2a3ec9c6779fcdeac51ce5b3a9b0156f06ac7152e5bba864/yarl-1.23.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:389871e65468400d6283c0308e791a640b5ab5c83bcee02a2f51295f95e09748", size = 92143, upload-time = "2026-03-01T22:04:16.829Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/63/88802d1f6b1cb1fc67d67a58cd0cf8a1790de4ce7946e434240f1d60ab4a/yarl-1.23.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:dda608c88cf709b1d406bdfcd84d8d63cff7c9e577a403c6108ce8ce9dcc8764", size = 107643, upload-time = "2026-03-01T22:04:18.519Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/db/4f9b838f4d8bdd6f0f385aed8bbf21c71ed11a0b9983305c302cbd557815/yarl-1.23.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8c4fe09e0780c6c3bf2b7d4af02ee2394439d11a523bbcf095cf4747c2932007", size = 108700, upload-time = "2026-03-01T22:04:20.373Z" },
+    { url = "https://files.pythonhosted.org/packages/50/12/95a1d33f04a79c402664070d43b8b9f72dc18914e135b345b611b0b1f8cc/yarl-1.23.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:31c9921eb8bd12633b41ad27686bbb0b1a2a9b8452bfdf221e34f311e9942ed4", size = 102769, upload-time = "2026-03-01T22:04:23.055Z" },
+    { url = "https://files.pythonhosted.org/packages/86/65/91a0285f51321369fd1a8308aa19207520c5f0587772cfc2e03fc2467e90/yarl-1.23.0-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5f10fd85e4b75967468af655228fbfd212bdf66db1c0d135065ce288982eda26", size = 101114, upload-time = "2026-03-01T22:04:25.031Z" },
+    { url = "https://files.pythonhosted.org/packages/58/80/c7c8244fc3e5bc483dc71a09560f43b619fab29301a0f0a8f936e42865c7/yarl-1.23.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:dbf507e9ef5688bada447a24d68b4b58dd389ba93b7afc065a2ba892bea54769", size = 98883, upload-time = "2026-03-01T22:04:27.281Z" },
+    { url = "https://files.pythonhosted.org/packages/86/e7/71ca9cc9ca79c0b7d491216177d1aed559d632947b8ffb0ee60f7d8b23e3/yarl-1.23.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:85e9beda1f591bc73e77ea1c51965c68e98dafd0fec72cdd745f77d727466716", size = 94172, upload-time = "2026-03-01T22:04:28.554Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/3f/6c6c8a0fe29c26fb2db2e8d32195bb84ec1bfb8f1d32e7f73b787fcf349b/yarl-1.23.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:0e1fdaa14ef51366d7757b45bde294e95f6c8c049194e793eedb8387c86d5993", size = 107010, upload-time = "2026-03-01T22:04:30.385Z" },
+    { url = "https://files.pythonhosted.org/packages/56/38/12730c05e5ad40a76374d440ed8b0899729a96c250516d91c620a6e38fc2/yarl-1.23.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:75e3026ab649bf48f9a10c0134512638725b521340293f202a69b567518d94e0", size = 100285, upload-time = "2026-03-01T22:04:31.752Z" },
+    { url = "https://files.pythonhosted.org/packages/34/92/6a7be9239f2347234e027284e7a5f74b1140cc86575e7b469d13fba1ebfe/yarl-1.23.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:80e6d33a3d42a7549b409f199857b4fb54e2103fc44fb87605b6663b7a7ff750", size = 108230, upload-time = "2026-03-01T22:04:33.844Z" },
+    { url = "https://files.pythonhosted.org/packages/5e/81/4aebccfa9376bd98b9d8bfad20621a57d3e8cfc5b8631c1fa5f62cdd03f4/yarl-1.23.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5ec2f42d41ccbd5df0270d7df31618a8ee267bfa50997f5d720ddba86c4a83a6", size = 103008, upload-time = "2026-03-01T22:04:35.856Z" },
+    { url = "https://files.pythonhosted.org/packages/38/0f/0b4e3edcec794a86b853b0c6396c0a888d72dfce19b2d88c02ac289fb6c1/yarl-1.23.0-cp310-cp310-win32.whl", hash = "sha256:debe9c4f41c32990771be5c22b56f810659f9ddf3d63f67abfdcaa2c6c9c5c1d", size = 83073, upload-time = "2026-03-01T22:04:38.268Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/71/ad95c33da18897e4c636528bbc24a1dd23fe16797de8bc4ec667b8db0ba4/yarl-1.23.0-cp310-cp310-win_amd64.whl", hash = "sha256:ab5f043cb8a2d71c981c09c510da013bc79fd661f5c60139f00dd3c3cc4f2ffb", size = 87328, upload-time = "2026-03-01T22:04:39.558Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/14/dfa369523c79bccf9c9c746b0a63eb31f65db9418ac01275f7950962e504/yarl-1.23.0-cp310-cp310-win_arm64.whl", hash = "sha256:263cd4f47159c09b8b685890af949195b51d1aa82ba451c5847ca9bc6413c220", size = 82463, upload-time = "2026-03-01T22:04:41.454Z" },
+    { url = "https://files.pythonhosted.org/packages/69/68/c8739671f5699c7dc470580a4f821ef37c32c4cb0b047ce223a7f115757f/yarl-1.23.0-py3-none-any.whl", hash = "sha256:a2df6afe50dea8ae15fa34c9f824a3ee958d785fd5d089063d960bae1daa0a3f", size = 48288, upload-time = "2026-03-01T22:07:51.388Z" },
+]
+
+[[package]]
+name = "zipp"
+version = "3.23.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e3/02/0f2892c661036d50ede074e376733dca2ae7c6eb617489437771209d4180/zipp-3.23.0.tar.gz", hash = "sha256:a07157588a12518c9d4034df3fbbee09c814741a33ff63c05fa29d26a2404166", size = 25547, upload-time = "2025-06-08T17:06:39.4Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2e/54/647ade08bf0db230bfea292f893923872fd20be6ac6f53b2b936ba839d75/zipp-3.23.0-py3-none-any.whl", hash = "sha256:071652d6115ed432f5ce1d34c336c0adfd6a884660d1e9712a256d3d3bd4b14e", size = 10276, upload-time = "2025-06-08T17:06:38.034Z" },
+]
+
+[[package]]
+name = "zstandard"
+version = "0.25.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fd/aa/3e0508d5a5dd96529cdc5a97011299056e14c6505b678fd58938792794b1/zstandard-0.25.0.tar.gz", hash = "sha256:7713e1179d162cf5c7906da876ec2ccb9c3a9dcbdffef0cc7f70c3667a205f0b", size = 711513, upload-time = "2025-09-14T22:15:54.002Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/56/7a/28efd1d371f1acd037ac64ed1c5e2b41514a6cc937dd6ab6a13ab9f0702f/zstandard-0.25.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e59fdc271772f6686e01e1b3b74537259800f57e24280be3f29c8a0deb1904dd", size = 795256, upload-time = "2025-09-14T22:15:56.415Z" },
+    { url = "https://files.pythonhosted.org/packages/96/34/ef34ef77f1ee38fc8e4f9775217a613b452916e633c4f1d98f31db52c4a5/zstandard-0.25.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4d441506e9b372386a5271c64125f72d5df6d2a8e8a2a45a0ae09b03cb781ef7", size = 640565, upload-time = "2025-09-14T22:15:58.177Z" },
+    { url = "https://files.pythonhosted.org/packages/9d/1b/4fdb2c12eb58f31f28c4d28e8dc36611dd7205df8452e63f52fb6261d13e/zstandard-0.25.0-cp310-cp310-manylinux2010_i686.manylinux2014_i686.manylinux_2_12_i686.manylinux_2_17_i686.whl", hash = "sha256:ab85470ab54c2cb96e176f40342d9ed41e58ca5733be6a893b730e7af9c40550", size = 5345306, upload-time = "2025-09-14T22:16:00.165Z" },
+    { url = "https://files.pythonhosted.org/packages/73/28/a44bdece01bca027b079f0e00be3b6bd89a4df180071da59a3dd7381665b/zstandard-0.25.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e05ab82ea7753354bb054b92e2f288afb750e6b439ff6ca78af52939ebbc476d", size = 5055561, upload-time = "2025-09-14T22:16:02.22Z" },
+    { url = "https://files.pythonhosted.org/packages/e9/74/68341185a4f32b274e0fc3410d5ad0750497e1acc20bd0f5b5f64ce17785/zstandard-0.25.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:78228d8a6a1c177a96b94f7e2e8d012c55f9c760761980da16ae7546a15a8e9b", size = 5402214, upload-time = "2025-09-14T22:16:04.109Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/67/f92e64e748fd6aaffe01e2b75a083c0c4fd27abe1c8747fee4555fcee7dd/zstandard-0.25.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:2b6bd67528ee8b5c5f10255735abc21aa106931f0dbaf297c7be0c886353c3d0", size = 5449703, upload-time = "2025-09-14T22:16:06.312Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/e5/6d36f92a197c3c17729a2125e29c169f460538a7d939a27eaaa6dcfcba8e/zstandard-0.25.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4b6d83057e713ff235a12e73916b6d356e3084fd3d14ced499d84240f3eecee0", size = 5556583, upload-time = "2025-09-14T22:16:08.457Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/83/41939e60d8d7ebfe2b747be022d0806953799140a702b90ffe214d557638/zstandard-0.25.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:9174f4ed06f790a6869b41cba05b43eeb9a35f8993c4422ab853b705e8112bbd", size = 5045332, upload-time = "2025-09-14T22:16:10.444Z" },
+    { url = "https://files.pythonhosted.org/packages/b3/87/d3ee185e3d1aa0133399893697ae91f221fda79deb61adbe998a7235c43f/zstandard-0.25.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:25f8f3cd45087d089aef5ba3848cd9efe3ad41163d3400862fb42f81a3a46701", size = 5572283, upload-time = "2025-09-14T22:16:12.128Z" },
+    { url = "https://files.pythonhosted.org/packages/0a/1d/58635ae6104df96671076ac7d4ae7816838ce7debd94aecf83e30b7121b0/zstandard-0.25.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:3756b3e9da9b83da1796f8809dd57cb024f838b9eeafde28f3cb472012797ac1", size = 4959754, upload-time = "2025-09-14T22:16:14.225Z" },
+    { url = "https://files.pythonhosted.org/packages/75/d6/57e9cb0a9983e9a229dd8fd2e6e96593ef2aa82a3907188436f22b111ccd/zstandard-0.25.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:81dad8d145d8fd981b2962b686b2241d3a1ea07733e76a2f15435dfb7fb60150", size = 5266477, upload-time = "2025-09-14T22:16:16.343Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/a9/ee891e5edf33a6ebce0a028726f0bbd8567effe20fe3d5808c42323e8542/zstandard-0.25.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:a5a419712cf88862a45a23def0ae063686db3d324cec7edbe40509d1a79a0aab", size = 5440914, upload-time = "2025-09-14T22:16:18.453Z" },
+    { url = "https://files.pythonhosted.org/packages/58/08/a8522c28c08031a9521f27abc6f78dbdee7312a7463dd2cfc658b813323b/zstandard-0.25.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:e7360eae90809efd19b886e59a09dad07da4ca9ba096752e61a2e03c8aca188e", size = 5819847, upload-time = "2025-09-14T22:16:20.559Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/11/4c91411805c3f7b6f31c60e78ce347ca48f6f16d552fc659af6ec3b73202/zstandard-0.25.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:75ffc32a569fb049499e63ce68c743155477610532da1eb38e7f24bf7cd29e74", size = 5363131, upload-time = "2025-09-14T22:16:22.206Z" },
+    { url = "https://files.pythonhosted.org/packages/ef/d6/8c4bd38a3b24c4c7676a7a3d8de85d6ee7a983602a734b9f9cdefb04a5d6/zstandard-0.25.0-cp310-cp310-win32.whl", hash = "sha256:106281ae350e494f4ac8a80470e66d1fe27e497052c8d9c3b95dc4cf1ade81aa", size = 436469, upload-time = "2025-09-14T22:16:25.002Z" },
+    { url = "https://files.pythonhosted.org/packages/93/90/96d50ad417a8ace5f841b3228e93d1bb13e6ad356737f42e2dde30d8bd68/zstandard-0.25.0-cp310-cp310-win_amd64.whl", hash = "sha256:ea9d54cc3d8064260114a0bbf3479fc4a98b21dffc89b3459edd506b69262f6e", size = 506100, upload-time = "2025-09-14T22:16:23.569Z" },
+]