Finalize task-transfer catalogue and preserved dataset selections

README.md

@@ -1201,6 +1201,142 @@ configs:
     - evaluations/flappy/factor_sweep:flappy:observation_stride:uniform_u1_3:fs4:obs7p5:stride4:seed12/uniform_min_1_max_3/episode_metrics.jsonl
     - evaluations/flappy/factor_sweep:flappy:observation_stride:uniform_u1_3:fs4:obs7p5:stride4:seed13/uniform_min_1_max_3/episode_metrics.jsonl
     - evaluations/flappy/factor_sweep:flappy:observation_stride:uniform_u1_3:fs4:obs7p5:stride4:seed14/uniform_min_1_max_3/episode_metrics.jsonl
+- config_name: task-transfer--asterix_fixed_latency_2_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/train-00000-of-00001.parquet
+  - split: val
+    path:
+    - latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--asterix_fixed_latency_4_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/train-00000-of-00001.parquet
+  - split: val
+    path:
+    - latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--atlantis_fixed_latency_2_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00000-of-00002.parquet
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00001-of-00002.parquet
+  - split: val
+    path:
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--atlantis_fixed_latency_4_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet
+  - split: val
+    path:
+    - latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--deadly_corridor_fix_latency_2_1000ep
+  data_files:
+  - split: train
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/train.parquet
+  - split: val
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/val.parquet
+- config_name: task-transfer--deadly_corridor_fix_latency_4_1000ep
+  data_files:
+  - split: train
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/train.parquet
+  - split: val
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/val.parquet
+- config_name: task-transfer--deadly_corridor_fix_latency_6_1000ep
+  data_files:
+  - split: train
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/train.parquet
+  - split: val
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/val.parquet
+- config_name: task-transfer--deadly_corridor_fix_latency_8_1000ep
+  data_files:
+  - split: train
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/train.parquet
+  - split: val
+    path:
+    - latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/val.parquet
+- config_name: task-transfer--defend_the_line_fixed_latency_2_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00000-of-00003.parquet
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00001-of-00003.parquet
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00002-of-00003.parquet
+  - split: val
+    path:
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--defend_the_line_fixed_latency_4_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet
+  - split: val
+    path:
+    - latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--asterix_fixed_latency_0_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/train-00000-of-00001.parquet
+  - split: val
+    path:
+    - zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--atlantis_fixed_latency_0_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet
+    - zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet
+  - split: val
+    path:
+    - zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--deadly_corridor_fix_latency_0_1000ep
+  data_files:
+  - split: train
+    path:
+    - zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/train.parquet
+  - split: val
+    path:
+    - zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/val.parquet
+- config_name: task-transfer--defend_the_line_fixed_latency_0_200ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet
+    - zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet
+  - split: val
+    path:
+    - zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet
+- config_name: task-transfer--defend_the_line_fixed_latency_0_1000ep_7k2steps
+  data_files:
+  - split: train
+    path:
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00000-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00001-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00002-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00003-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00004-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00005-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00006-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00007-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00008-of-00010.parquet
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00009-of-00010.parquet
+  - split: val
+    path:
+    - zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/val-00000-of-00001.parquet
 dataset_info:
 - config_name: humanoidbench_balance_simple
   features:
@@ -3257,6 +3393,498 @@ dataset_info:
   - name: t
     dtype: int64
   download_size: 11243675137
+- config_name: task-transfer--asterix_fixed_latency_2_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 256820
+  - name: val
+    num_examples: 2598
+  download_size: 329687901
+- config_name: task-transfer--asterix_fixed_latency_4_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 222525
+  - name: val
+    num_examples: 2154
+  download_size: 283416109
+- config_name: task-transfer--atlantis_fixed_latency_2_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 356400
+  - name: val
+    num_examples: 3600
+  download_size: 577290972
+- config_name: task-transfer--atlantis_fixed_latency_4_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 356400
+  - name: val
+    num_examples: 3600
+  download_size: 588296037
+- config_name: task-transfer--deadly_corridor_fix_latency_2_1000ep
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 32500
+  - name: val
+    num_examples: 4000
+  download_size: 558841872
+- config_name: task-transfer--deadly_corridor_fix_latency_4_1000ep
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 28737
+  - name: val
+    num_examples: 3178
+  download_size: 501369715
+- config_name: task-transfer--deadly_corridor_fix_latency_6_1000ep
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 34290
+  - name: val
+    num_examples: 3890
+  download_size: 597152880
+- config_name: task-transfer--deadly_corridor_fix_latency_8_1000ep
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 28895
+  - name: val
+    num_examples: 3245
+  download_size: 517062764
+- config_name: task-transfer--defend_the_line_fixed_latency_2_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 63623
+  - name: val
+    num_examples: 644
+  download_size: 1002177725
+- config_name: task-transfer--defend_the_line_fixed_latency_4_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 60249
+  - name: val
+    num_examples: 657
+  download_size: 946114447
+- config_name: task-transfer--asterix_fixed_latency_0_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 272266
+  - name: val
+    num_examples: 2768
+  download_size: 355282674
+- config_name: task-transfer--atlantis_fixed_latency_0_200ep_7k2steps
+  features:
+  - name: image
+    dtype: image
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: decision_step
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: seed
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: latency_raw_frames
+    dtype: int64
+  - name: latency_ms
+    dtype: float64
+  - name: raw_reward
+    dtype: float64
+  - name: action_id
+    dtype: int64
+  - name: action
+    list: float32
+  - name: action_text
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 356400
+  - name: val
+    num_examples: 3600
+  download_size: 599995003
+- config_name: task-transfer--deadly_corridor_fix_latency_0_1000ep
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 38308
+  - name: val
+    num_examples: 4295
+  download_size: 661234054
+- config_name: task-transfer--defend_the_line_fixed_latency_0_200ep_7k2steps
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 55907
+  - name: val
+    num_examples: 626
+  download_size: 875557436
+- config_name: task-transfer--defend_the_line_fixed_latency_0_1000ep_7k2steps
+  features:
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: decision_step
+    dtype: int64
+  - name: env_name
+    dtype: string
+  - name: episode_idx
+    dtype: int64
+  - name: image
+    dtype: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: raw_reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 287148
+  - name: val
+    num_examples: 2844
+  download_size: 4492816871
 ---
 
 # LAGEN benchmark-datasets

archive/task-transfer/manifest.json

@@ -77,6 +77,185 @@
           "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
           "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/post_train/step_2000.json"
         }
+      ],
+      "training_data": [
+        {
+          "task": "defend_the_line",
+          "original_configuration": {
+            "name": "defend_the_line",
+            "role": "A",
+            "train_source_hf": "latency-sensitive-bench/memory-rollouts",
+            "prompt_source_hf": "latency-sensitive-bench/memory-rollouts",
+            "train_source_subdir": "defend_the_line_fixed_latency_0_1000ep_7k2steps",
+            "prompt_source_subdir": "defend_the_line_fixed_latency_0_1000ep_7k2steps",
+            "converted_name": "defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3",
+            "train_latency_filter": [
+              0
+            ],
+            "eval_latency_filter": [
+              0
+            ],
+            "episodes_per_latency": 1000,
+            "max_episodes": null,
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00000-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00001-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00002-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00003-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00004-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00005-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00006-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00007-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00008-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00009-of-00010.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/val-00000-of-00001.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "deadly_corridor",
+          "original_configuration": {
+            "name": "deadly_corridor",
+            "role": "B",
+            "train_source_hf": "latency-sensitive-bench/deadly_1000ep",
+            "prompt_source_hf": "latency-sensitive-bench/deadly_1000ep",
+            "converted_name": "deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3",
+            "train_latency_filter": [
+              0
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": 1000,
+            "max_episodes": null,
+            "action_layout": "multibinary_7",
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/val.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        }
       ]
     },
     {
@@ -335,6 +514,418 @@
           "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
           "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/run-7ujo2tgu.wandb"
         }
+      ],
+      "training_data": [
+        {
+          "task": "flappy",
+          "original_configuration": {
+            "name": "flappy",
+            "role": "A",
+            "train_source_hf": "latency-sensitive-bench/flappy_200ep",
+            "prompt_source_hf": "latency-sensitive-bench/flappy_200ep",
+            "train_source_subdir": "flappy_fix_latency_0_200ep",
+            "prompt_source_subdir": "flappy_fix_latency_0_200ep",
+            "converted_name": "flappy_mixed_024_200ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 200,
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep/val.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "demon_attack",
+          "original_configuration": {
+            "name": "demon_attack",
+            "role": "B",
+            "train_source_hf": "latency-sensitive-bench/demon_attack_200ep",
+            "prompt_source_hf": "latency-sensitive-bench/demon_attack_200ep",
+            "train_source_subdir": "demon_attack_fix_latency_0_200ep",
+            "prompt_source_subdir": "demon_attack_fix_latency_0_200ep",
+            "converted_name": "demon_attack_mixed_024_200ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 200,
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep/val.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "deadly_corridor",
+          "original_configuration": {
+            "name": "deadly_corridor",
+            "role": "C",
+            "train_source_hf": "latency-sensitive-bench/deadly_1000ep",
+            "prompt_source_hf": "latency-sensitive-bench/deadly_1000ep",
+            "train_source_subdir": "deadly_corridor_fix_latency_0_1000ep",
+            "prompt_source_subdir": "deadly_corridor_fix_latency_0_1000ep",
+            "converted_name": "deadly_corridor_mixed_024_1000ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 1000,
+            "action_layout": "multibinary_7",
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/val.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/train.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/val.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "asterix",
+          "original_configuration": {
+            "name": "asterix",
+            "role": "D",
+            "train_source_hf": "latency-sensitive-bench/asterix_200ep",
+            "prompt_source_hf": "latency-sensitive-bench/asterix_200ep",
+            "train_source_subdir": "asterix_fixed_latency_0_200ep_7k2steps",
+            "prompt_source_subdir": "asterix_fixed_latency_0_200ep_7k2steps",
+            "converted_name": "asterix_mixed_024_200ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 200,
+            "action_layout": "factorized_6",
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/train-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/train-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/train-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "atlantis",
+          "original_configuration": {
+            "name": "atlantis",
+            "role": "E",
+            "train_source_hf": "latency-sensitive-bench/atlantis_200ep",
+            "prompt_source_hf": "latency-sensitive-bench/atlantis_200ep",
+            "train_source_subdir": "atlantis_fixed_latency_0_200ep_7k2steps",
+            "prompt_source_subdir": "atlantis_fixed_latency_0_200ep_7k2steps",
+            "converted_name": "atlantis_mixed_024_200ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 200,
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00000-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00001-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        },
+        {
+          "task": "defend_the_line",
+          "original_configuration": {
+            "name": "defend_the_line",
+            "role": "F",
+            "train_source_hf": "latency-sensitive-bench/defend_the_line_200ep",
+            "prompt_source_hf": "latency-sensitive-bench/defend_the_line_200ep",
+            "train_source_subdir": "defend_the_line_fixed_latency_0_200ep_7k2steps",
+            "prompt_source_subdir": "defend_the_line_fixed_latency_0_200ep_7k2steps",
+            "converted_name": "defend_the_line_mixed_024_200ep_cross_train",
+            "train_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "eval_latency_filter": [
+              0,
+              2,
+              4
+            ],
+            "episodes_per_latency": null,
+            "max_episodes": 200,
+            "weight": 1.0
+          },
+          "available_dataset_snapshot": [
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00000-of-00003.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00001-of-00003.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00002-of-00003.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet"
+            },
+            {
+              "repo": "latency-sensitive-bench/benchmark-datasets",
+              "repo_type": "dataset",
+              "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+              "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet"
+            }
+          ],
+          "exact_training_dataset_revision_confirmed": false,
+          "source_limit": "The original run config does not pin a dataset revision. These references identify preserved available snapshots, not proof of exact training bytes."
+        }
       ]
     }
   ],

experiments/task-transfer/README.md

@@ -6,3 +6,5 @@
 | Archive | [Joint training and earlier evaluation versions](../../archive/task-transfer/README.md) |
 
 Shared checkpoints and datasets have one canonical location in benchmark-models and benchmark-datasets. Fixed references and source limits are in each manifest. Latency profiles use the existing profiles repository.
+
+Verification uses fixed Git/LFS object identities and sizes. All 46 Parquet files passed real sample reads; five representative images decoded. Complete framework loading was not repeated.

experiments/task-transfer/provenance/consumer_changes.json

@@ -0,0 +1,33 @@
+[
+  "scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/gr00t/bridge/single/deadly_corridor_multibinary_bce.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/gr00t/bridge/single/deadly_corridor_new_lr.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/deadly_zero_demon_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/deadly_zero_flappy_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/defendtheline_deadly/defendtheline_zero_deadly_zero.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_mixed_deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_mixed_deadly_zero.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_zero_deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_zero_deadly_zero.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_zero_deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_mixed_deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_mixed_deadly_zero.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_zero_deadly_mixed.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_zero_deadly_zero.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_demon_deadly_asterix_atlantis_defend_024.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/rollout_defend_asterix_atlantis_024_small_models.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/eval/demon_attack_mixed_024_latency_024_cross_games.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/eval/demon_attack_zero_latency_024_cross_games.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_2.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_4.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_6.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_8.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/mixed/demon_024.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/single/deadly_corridor_multibinary_bce.sh",
+  "third_party/starVLA/examples/rl_games/bash_scripts/pi_0.5/bridge/single/deadly_corridor_multibinary_bce.sh",
+  "third_party/starVLA/examples/rl_games/config/cross_task_setup/defendtheline_zero_deadly_zero.yaml",
+  "third_party/starVLA/examples/rl_games/config/cross_task_setup/flappy_demon_deadly_asterix_atlantis_defend_024.yaml",
+  "third_party/starVLA/examples/rl_games/scripts/setup_training_assets.py",
+  "third_party/starVLA/tests/rl_games/test_training_commands.py"
+]

experiments/task-transfer/provenance/current_asset_identity_verification.json

@@ -0,0 +1,1766 @@
+[
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/checkpoints/steps_2000_model.safetensors",
+    "identity": "38e0a39116fe43c53068d76929d93fb391c669002fd43c359174ab267f609f3c",
+    "size": 9784896838
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/config.full.yaml",
+    "identity": "70a38d604b853b7adbdc727834fdc3e3e6e109f8",
+    "size": 11501
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/config.yaml",
+    "identity": "2bdb1510f7bd60c7e907d143fddb0546434669cd",
+    "size": 4946
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/dataset_statistics.json",
+    "identity": "324d8d243e8a7676d2431a8aec094aa585ff2423",
+    "size": 1840
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/dataset_statistics_eval.json",
+    "identity": "bf2f8e6c66bb9c767849d10bd6bfd1865acc2104",
+    "size": 1840
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/steps_15000_model.safetensors",
+    "identity": "ccab5fb797b96ccde78035c1f1d9e83ebe5f2a4bc9e5b8134f0682bc4e979bdd",
+    "size": 9138230516
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/config.full.yaml",
+    "identity": "16dbc1b42747ad6bf503a81421a607e10229be4b",
+    "size": 14835
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/config.yaml",
+    "identity": "b312127c0e31b525dc655295c82c61e0bf08a188",
+    "size": 3858
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/dataset_statistics.json",
+    "identity": "65426e274f8bd3515fa572f76e416d26f49f3f1e",
+    "size": 1845
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/dataset_statistics_eval.json",
+    "identity": "353df6fa80f7ffde013ebd6f0b909b485baff180",
+    "size": 1839
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/config.yaml",
+    "identity": "c0b4214048c307bc86f2ff68a4d31bd792d4b8af",
+    "size": 14890
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/hydra.yaml",
+    "identity": "9184975a2e6b3dd774732bed6230247451c402c0",
+    "size": 44061
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/.hydra/overrides.yaml",
+    "identity": "f5fc2348a29d29430eee72c7a4830c2075098e24",
+    "size": 19984
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-models",
+    "revision": "81365b0f8691c4f80955fe94de68976a59c624b1",
+    "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_latency_mixed_024_exp1/checkpoints/steps_5000_model.safetensors",
+    "identity": "dc6f0c52cfc33ea2ae7d7c382abe762561693936045464b1c160b1d9fdf0f501",
+    "size": 9138230516
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/asterix_200ep/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/asterix_200ep/README.md",
+    "identity": "466f681d27b7ff329688d58fc97f694063ac3359",
+    "size": 2899
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/train-00000-of-00001.parquet",
+    "identity": "08965162575a9212603483aa744ad500009572407d274a40f82a8b000b3ce84a",
+    "size": 351689678
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "78fb4b1f7ecd55810501ff112d33a666f59385aa9a819c6843839e105679065b",
+    "size": 3592996
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/train-00000-of-00001.parquet",
+    "identity": "048589d78216b1da859011943e19be246dd250f0717bde883f4f988358c9de8f",
+    "size": 326368562
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "4ee6ab4b71064f29db50e01671940c26555f706870be3b242166899c6880f10b",
+    "size": 3319339
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/train-00000-of-00001.parquet",
+    "identity": "9617ce0ebda127d70d84db391c2c73ebfebd6dac8cc0ef0f14fda6786aaa9d66",
+    "size": 280637541
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "2fdbb8f742a72aed8c48180bc7846b69e34b319e0300a792b1f605cf858173c1",
+    "size": 2778568
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/defend_the_line_200ep/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/defend_the_line_200ep/README.md",
+    "identity": "184279f393b69ba02f5344d2582e6290a2cf7404",
+    "size": 2889
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/metadata.json",
+    "identity": "6e8c907b2eb04081978fd590a978022990f0a6b2",
+    "size": 2410
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet",
+    "identity": "5a690bfca5411b93a67621280def3d3608e111c2cc3c2d2fd499da0876f4ce97",
+    "size": 444775269
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet",
+    "identity": "eb62cdef51dab8fe68419e38d712598e4e103b21deb3733557299c323b6a0cf2",
+    "size": 421150593
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "3b5fdeabdd91257259dc7d5d2261e5b51357d7c5bcf5d7edc4e07fab73d7a1b0",
+    "size": 9631574
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00000-of-00003.parquet",
+    "identity": "3874191de85f932df2b921fe3213d92882027f69b092c413c87f7fc30267e700",
+    "size": 330460271
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00001-of-00003.parquet",
+    "identity": "ff16d87d75058aa08c88a829e1b4c0db1689954a508082e9544d15a629dc4f1c",
+    "size": 331066562
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00002-of-00003.parquet",
+    "identity": "527a12734dc4381cc45aa9e3132ce0af2a67700dd46097e43cc20990fc01ba24",
+    "size": 330568547
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "6b9bb598b82f8e443f506671a7cd5e4eae4f31d098928d265aa2e15fe3abc854",
+    "size": 10082345
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet",
+    "identity": "216097a7d050207e0e64cae4985df62e756855d9cf1bfa6271161c7d96ee9851",
+    "size": 468273970
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet",
+    "identity": "dac853430e8bf878b86f9263fb7bbd5940787dd2af8acef269613ee270494279",
+    "size": 467671674
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "d9c51f7a5fe4f1a290baa91528941b14c96caacccaef5f2742ab427fd750e738",
+    "size": 10168803
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/.gitattributes",
+    "identity": "a6344aac8c09253b3b630fb776ae94478aa0275b",
+    "size": 1519
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/mid_train/step_1000.json",
+    "identity": "9378807f51fc9b6e27e83e6b05ec6ca15c7f2c10",
+    "size": 9549
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/mid_train/step_1500.json",
+    "identity": "30f90a51e00b0ba0de75f704a82680838ba67eba",
+    "size": 9516
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/mid_train/step_2000.json",
+    "identity": "0f9fb50b629e58c120ab71f99441d8d94ad1f5ba",
+    "size": 9558
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/mid_train/step_500.json",
+    "identity": "5b1df0ba1b1baded7e340a366b45fc238f6ada76",
+    "size": 9661
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3/eval/post_train/step_2000.json",
+    "identity": "ab1965a332fc580664aa3caf472e622f88c134d8",
+    "size": 20593
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00000-of-00010.parquet",
+    "identity": "5de4bf1be2adeaa7379936ba2fa8ae3ecfc3c4b640fbd0826ed4d0cc5f0feebd",
+    "size": 444891548
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00001-of-00010.parquet",
+    "identity": "ea0034fd34539f4791d46fa628b11a12567edbc8ddc1bdea67945ff1a438e954",
+    "size": 444807064
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00002-of-00010.parquet",
+    "identity": "8bbdcc2012239c81c8481230b1ee11bc20ef720fb23147c66b406c71bc55f310",
+    "size": 444688952
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00003-of-00010.parquet",
+    "identity": "18619b859a1ff5c3ffd227c01422d61ad3a1b2d8a10459e5ce080999c32654f6",
+    "size": 444596599
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00004-of-00010.parquet",
+    "identity": "bf7bb43e02f437087105d3d2d417ccc7bd9a8c81743d59e9e3cc9f025327fb67",
+    "size": 444645418
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00005-of-00010.parquet",
+    "identity": "ca737abb08dc2553758914ca39d6aac661f2c50d375000c57614119d84a3ad95",
+    "size": 444958475
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00006-of-00010.parquet",
+    "identity": "02198e70aa3ae05324d1206584d53afaa922690db18ece8d8f5ad82397c78f7b",
+    "size": 444949834
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00007-of-00010.parquet",
+    "identity": "04ca060e0a5af9a935b18b5c3708373ff1f47ca3646ba592c63446785d6bb872",
+    "size": 444759817
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00008-of-00010.parquet",
+    "identity": "ca5aba0d0f87dd5facf3c2b04615acabf625e5b0522fc7aec1920cbf5f7aba2f",
+    "size": 444853837
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00009-of-00010.parquet",
+    "identity": "4d902a02c6cb188061e2bf69374bd010b4e03da383cb015cea36f919e6269dd9",
+    "size": 445472971
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "961f783822b9ed92ab1f2f2f13ca32d795968f0c40f810023e4c77043c60e0bf",
+    "size": 44192356
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/deadly_1000ep/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/filter_report.json",
+    "identity": "71c63b1e5e235ca4a3b96a0d6bc79615546a873a",
+    "size": 336
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/metadata.json",
+    "identity": "65e47eb459d872c68a8de844ac7b20aaa0e0b9e4",
+    "size": 13241
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/train.parquet",
+    "identity": "c8bf679fe02b619e0e718390abc2538ffa0448948294d21b00b3c1b8956822cc",
+    "size": 594610508
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/val.parquet",
+    "identity": "2817a0fc442c11ff2d672f6f7f93ecdd9e972b8ba63d53f4ea50f32fda0735cd",
+    "size": 66623546
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/filter_report.json",
+    "identity": "3fd27dd6ecf3e1c485713837cb24bf47d21e8a65",
+    "size": 337
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/metadata.json",
+    "identity": "74421a79294dc82436cc9c6ad81f6f08516bdeca",
+    "size": 13257
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/train.parquet",
+    "identity": "b6627ab830bb16f901841b1f3cfa7c4a6c727e323301a079cf7ffe0f1043089d",
+    "size": 497929218
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/val.parquet",
+    "identity": "681b9cd8330aefd7bfe74c9dfac4d6737aca4c48a5c54e4e6864a604ec67a0d4",
+    "size": 60912654
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/filter_report.json",
+    "identity": "0bd561d792d0b16593b3de8e0d592afdc7823603",
+    "size": 339
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/metadata.json",
+    "identity": "81db6ee056fb586bec4bfe82dd48c0bfecda8ffb",
+    "size": 13259
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/probe_state.json",
+    "identity": "013d2dab7bf7bbe4398feefc4aa4e132762588fb",
+    "size": 233
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/train.parquet",
+    "identity": "5ea5cc6ae47203a19b0a00e1ee4d130b7c8a0ee0c93e15240ba4dd75f0575153",
+    "size": 451550808
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/val.parquet",
+    "identity": "d1183efc614fb8e5769327a8d010a0a09536b9a1b01fe2b249d9f4eb8c32bc64",
+    "size": 49818907
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/filter_report.json",
+    "identity": "dfd47cd9da099db3ef0dc7760045196d43a23140",
+    "size": 337
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/metadata.json",
+    "identity": "f747c4427f5af85e219bd2bb2aaf0d4b205cfe11",
+    "size": 13256
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/train.parquet",
+    "identity": "0bca0ff3cf7a8866cc57c48dbb69f251b273655b1e615a8b9701e0e69a79dcb7",
+    "size": 535881935
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/val.parquet",
+    "identity": "d043c98ef817648eef1df10f6ce2e9a6a7262fb50575c5238317b5e0ebb87b0f",
+    "size": 61270945
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/filter_report.json",
+    "identity": "66f7f91bd793f183f9b7d653228b3dbce16398a1",
+    "size": 339
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/metadata.json",
+    "identity": "ca5468de22ad95fa70a1b1d0c891fa84d59e5109",
+    "size": 13258
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/train.parquet",
+    "identity": "5772070093e9dd86d98f8a2048fad09374d84b37c32b2685b895c22c7c6ad258",
+    "size": 464858411
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/val.parquet",
+    "identity": "071d97014411b0fcc797e1f4ef8f1bad39cd1eb0a58df33bb4cafb5e8ae6b740",
+    "size": 52204353
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/.gitattributes",
+    "identity": "4925e5503d90d10009a593819bb3d89e8603a5cd",
+    "size": 1711
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/_initialization/StarVLA_Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt",
+    "identity": "371cb744227687bb99bcad7f9ff2250cf06da75631359ad3eba4c6bc52570607",
+    "size": 9785060316
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/steps_15000_state/optimizer.bin",
+    "identity": "e8b2dc7f1a406b093db1b4a03d3ae0dc2d77415802548498921dd5b5668fe8ed",
+    "size": 18276885098
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/checkpoints/steps_15000_state/random_states_0.pkl",
+    "identity": "8f041f3f444886eb524567d2ff1df968b6f86a749c7ad6ad18b8d1478843cfb4",
+    "size": 14821
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/eval/mid_train/step_1000.json",
+    "identity": "a19f99cc0b96a4e8f5c8bd824181ba78919f4ca3",
+    "size": 1183
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/eval/post_train/step_15000.json",
+    "identity": "7f0c1a700c4dcad4c9eb67a745bac8ac233b17ee",
+    "size": 17780
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/hydra/train_starvla_hydra.log",
+    "identity": "430208c7c56f24bf789e8eb2779661c2daa5eb78",
+    "size": 5964411
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/summary.jsonl",
+    "identity": "60c2c92c6a52931fc5e3550f01dcb1c7339e6563",
+    "size": 17
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/debug-internal.log",
+    "identity": "ccd61cb4f8ccd50c945048c84b4cb34e5147b1bf",
+    "size": 2362127
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/debug.log",
+    "identity": "c17eae50e2cd298e151d707f71fa706d9ed04032",
+    "size": 3077
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_064644-r8358sud/files/wandb-summary.json",
+    "identity": "b0a620d0c1047a4dd8a400939b6da246ed8063a7",
+    "size": 37
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_064644-r8358sud/logs/debug-core.log",
+    "identity": "441f2d6bf9b9e4c4d852af1421b1428ac927f599",
+    "size": 1902
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_064644-r8358sud/logs/debug-internal.log",
+    "identity": "0176b96a511592ef1c3e0eb0e13a6b92d66f7012",
+    "size": 1875
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_064644-r8358sud/logs/debug.log",
+    "identity": "d78fda4d3689209111d383d1341078d82c8159e9",
+    "size": 1685
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_064644-r8358sud/run-r8358sud.wandb",
+    "identity": "27552a57891f6c5595f184f031dd32d8bd4168c0",
+    "size": 616
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/files/config.yaml",
+    "identity": "79eecf226b542e1f1f909d730dce93f72f3b7eb8",
+    "size": 30949
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/files/output.log",
+    "identity": "9ae3f268594d41310ad3882822a30e955f3f3599",
+    "size": 379209
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/files/requirements.txt",
+    "identity": "cbfc0ff2c5349a052911fcef1819dd9c1a3bb9af",
+    "size": 4289
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/files/wandb-metadata.json",
+    "identity": "df6c5b42cbb1ce11a36f7fda1518dba823c41232",
+    "size": 24813
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/files/wandb-summary.json",
+    "identity": "5ed3bdf66ef24da3d1c17496db88b5cd15d27bff",
+    "size": 375
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/logs/debug-core.log",
+    "identity": "12cad40af018961e26e405bce0fe6c3edeb70a4e",
+    "size": 2036
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/logs/debug-internal.log",
+    "identity": "27b1f7fa3a7e04612726b4bd665dd79cc72f3e60",
+    "size": 139662
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/logs/debug.log",
+    "identity": "d21941619508be73e8a2238af4bdef11289c4c7d",
+    "size": 2823
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_065340-67ykhdhp/run-67ykhdhp.wandb",
+    "identity": "51c7b17f1b22d417c4f38fbee97d787f97cd9ac52981b0a09dfb0b58552c6819",
+    "size": 5928658
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/files/config.yaml",
+    "identity": "04f0bd4ca4bd701248e01c2e07b9ef2a32ea094d",
+    "size": 30291
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/files/output.log",
+    "identity": "e90ba907ab38b0b4f47082fa7daff4dda49c7a22",
+    "size": 5525911
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/files/requirements.txt",
+    "identity": "cbfc0ff2c5349a052911fcef1819dd9c1a3bb9af",
+    "size": 4289
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/files/wandb-metadata.json",
+    "identity": "ec4edbe5f95984d7abc26b68f1e66b095ef2e8ec",
+    "size": 24159
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/files/wandb-summary.json",
+    "identity": "dd1c5d64e258ad3f9f254f4c0c567acbe7caede0",
+    "size": 27273
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/logs/debug-core.log",
+    "identity": "8d3ba7e6785e413ff5625731042e54dcab7800e9",
+    "size": 2576
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/logs/debug-internal.log",
+    "identity": "ccd61cb4f8ccd50c945048c84b4cb34e5147b1bf",
+    "size": 2362127
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/logs/debug.log",
+    "identity": "c17eae50e2cd298e151d707f71fa706d9ed04032",
+    "size": 3077
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1/wandb/wandb/run-20260923_105406-7ujo2tgu/run-7ujo2tgu.wandb",
+    "identity": "0c8e7caa217b97d88dfb6c0cc1138eae16c3183996e52dd464d3ed3df20439e5",
+    "size": 102100192
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/starvla-demon-attack-zero-latency-evals/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_zero_latency_latency024_cross_games/air_raid/eval/post_train/step_7000.json",
+    "identity": "0ca208bf731cf1fbc0703ee30ff89532e859c081",
+    "size": 5139
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_zero_latency_latency024_cross_games/asterix/eval/post_train/step_7000.json",
+    "identity": "2dd77d78ef7449b57184ac03e8186b9cbab88c66",
+    "size": 5122
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_zero_latency_latency024_cross_games/atlantis/eval/post_train/step_7000.json",
+    "identity": "61297f366012bf805f24ec4b0e8d9fdf6ed6c923",
+    "size": 5059
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_zero_latency_latency024_cross_games/demon_attack/eval/post_train/step_7000.json",
+    "identity": "d8f617eff558698aec6cf0f6e2db211f9c5b0abe",
+    "size": 5314
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_demon_attack_latency_mixed_024_exp1/.gitattributes",
+    "identity": "a6344aac8c09253b3b630fb776ae94478aa0275b",
+    "size": 1519
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_demon_attack_latency_mixed_024_exp1/eval/post_train/step_5000.json",
+    "identity": "c6a80acc0f3e10f212f03a20636dd17a144340ec",
+    "size": 5409
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_demon_attack_latency_mixed_024_exp1/steps_5000_state/optimizer.bin",
+    "identity": "1f232fb50fdd4dfe48e1ae8eed8aa5f937fef45da60d646493e0bf06a64ee6f6",
+    "size": 18276885098
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/openvla_bridge_demon_attack_latency_mixed_024_exp1/steps_5000_state/random_states_0.pkl",
+    "identity": "608b90a865655c4bd9bf57be0147891841890389d1b5b4d14494f4e90353f43e",
+    "size": 14821
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/atlantis_200ep/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/atlantis_200ep/README.md",
+    "identity": "05c9578dc2ae0e406b426d7a7af6782eb3f0438c",
+    "size": 2911
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet",
+    "identity": "d6bc52a95736b458609e6c9967476a8c6b161f05601fb4bff17e7c044f75419e",
+    "size": 297073109
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet",
+    "identity": "5317050e2bc381ea09a2fbb882667fbf6469b8e37fa75834e6c80c6e03145443",
+    "size": 296988310
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "e7b41aa1e6a97c174a84b687d657b9203ce209fe087c5e0c195a68d452fd4226",
+    "size": 5933584
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00000-of-00002.parquet",
+    "identity": "55d24975e5f46d8e63cbd13bf5a82d5ec28155fb7c1c5d8339d05881080d4a27",
+    "size": 286257711
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00001-of-00002.parquet",
+    "identity": "11453319182bfef00a590ba9378e798d51e21ddad99babc4eb3a0933057d0500",
+    "size": 285288978
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "604d8182bf37c78dd7394c7dfe4a8ae8be502a4830b244bd88e150842bbeaa33",
+    "size": 5744283
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet",
+    "identity": "236afd6fb0af6df3ae2be74ceb07d4f6101637f24638168450160a8166a27481",
+    "size": 291143956
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet",
+    "identity": "b6e0a66ee13299134fe7ead5e1464f3327376e6ea1e314c04764865b51b167cb",
+    "size": 291146804
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "identity": "285ce004ee1fbbfc0456b683cc7f407a70eda4920504ebe7b206d2d95b753854",
+    "size": 6005277
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "experiments/task-transfer/provenance/starvla-demon-attack-mixed-024-latency-evals/.gitattributes",
+    "identity": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a",
+    "size": 2504
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/eval/post_train/step_5000.json",
+    "identity": "b19bf892fc43b7f18804fe17ec9f698238d8d690",
+    "size": 5061
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/episode_metrics.jsonl",
+    "identity": "8f753a860409618c232ae8c4eb2cd7b475ca31b8",
+    "size": 21706
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/resolved_config.yaml",
+    "identity": "58a87af5e4268b393a39c751136b7dce583dd830",
+    "size": 3746
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/episode_metrics.jsonl",
+    "identity": "f323768c01aacfd758d37ffa9af25727bdc36690",
+    "size": 21822
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/resolved_config.yaml",
+    "identity": "e440ec4c8d5033de1644d5062291239d7fdfed9c",
+    "size": 3773
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/episode_metrics.jsonl",
+    "identity": "ec6086fc899f96324afe05b3511db8d2d4e5b74f",
+    "size": 21801
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/resolved_config.yaml",
+    "identity": "efed5fd76953b2b794135699e46c8f509a99aec2",
+    "size": 3773
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/eval/post_train/step_5000.json",
+    "identity": "02dedfd1e6a1abe1e85721a31da3c539fa883721",
+    "size": 5059
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/episode_metrics.jsonl",
+    "identity": "f507e5028125873ae9ee72fcc076d048e1cbacc8",
+    "size": 21834
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/resolved_config.yaml",
+    "identity": "7ccae245e9427f207dd939b7de9b40224b9b9c4b",
+    "size": 4004
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/episode_metrics.jsonl",
+    "identity": "0341789a58c5aa50abb81df793e12efb9ef23775",
+    "size": 22767
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/resolved_config.yaml",
+    "identity": "446323a63fb5a5befe2e7fe2b2ab47d064658512",
+    "size": 4044
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/episode_metrics.jsonl",
+    "identity": "b7eda6e93d1a0246ef4b0649bd65b590315e8584",
+    "size": 22689
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/resolved_config.yaml",
+    "identity": "f9eaa1023bf944c6d034c8213b6cbb7b496f87d5",
+    "size": 4043
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/eval/post_train/step_5000.json",
+    "identity": "0b6c723989718266b7fd7e23ae8f5469a2f3a6e9",
+    "size": 5111
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/episode_metrics.jsonl",
+    "identity": "b1c750abf03dcbf077ea34d0dd64af05e63b7418",
+    "size": 22030
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/resolved_config.yaml",
+    "identity": "b31bc6e35f1e4a6880b110a017825ebdec704f11",
+    "size": 3675
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/episode_metrics.jsonl",
+    "identity": "c931fd2fcd485fc57597a8a6e535d34db4f942a8",
+    "size": 22950
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/resolved_config.yaml",
+    "identity": "b055bf0950857edcf325ef7f498016c2c61ff050",
+    "size": 3715
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/episode_metrics.jsonl",
+    "identity": "1211e941d61258bccb9374e5e5d5c371305a773a",
+    "size": 22894
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/resolved_config.yaml",
+    "identity": "4a867822302c07eb90c6e3a2af9e9ca0e8c1f86a",
+    "size": 3714
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/eval/post_train/step_5000.json",
+    "identity": "a96086fd70367442f201ffae2424c35e5e652ef3",
+    "size": 5192
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/episode_metrics.jsonl",
+    "identity": "c443e5e6bd694758586ba17c07b3bc932224143d",
+    "size": 23582
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/resolved_config.yaml",
+    "identity": "b8d497468b99f1519bd103ca86d99a46d93b07cc",
+    "size": 2559
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/episode_metrics.jsonl",
+    "identity": "0bef968602772d2f550167f5bc83f012b8ff253b",
+    "size": 24525
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/resolved_config.yaml",
+    "identity": "7d544f0be1249393aae57b06df63c25b01c2d538",
+    "size": 2599
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/episode_metrics.jsonl",
+    "identity": "773b7866d30af4722f7448b10f07bc060af404cc",
+    "size": 24452
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "evaluations/task-transfer/demon-attack/demon_attack_mixed_024_horizonfix_full_20261001_002008/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/resolved_config.yaml",
+    "identity": "d4c8ff317132e4ce6add80a371630ca4991cdb58",
+    "size": 2598
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games/air_raid/eval/post_train/step_5000.json",
+    "identity": "72bb5de37c973ac818df47e0ede9d99acf9bca73",
+    "size": 5257
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games/asterix/eval/post_train/step_5000.json",
+    "identity": "e9b85e354966b9413d166ed156233322beb27671",
+    "size": 5318
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games/atlantis/eval/post_train/step_5000.json",
+    "identity": "267d4ee43ccbe3badeb3a36fc49015b9c14a6988",
+    "size": 5113
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games/demon_attack/eval/post_train/step_5000.json",
+    "identity": "3e76adbfc3d32a921c7f53cd440471137850bf2c",
+    "size": 5349
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/eval/post_train/step_5000.json",
+    "identity": "76083209256387804bfb2ffa61bfc08c7847829c",
+    "size": 5063
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/episode_metrics.jsonl",
+    "identity": "33a9f7ba24cc095e6ff46728831114400be1ec52",
+    "size": 22074
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/resolved_config.yaml",
+    "identity": "cb308d634f7725f4b70942cb534c85c305405a57",
+    "size": 3733
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/episode_metrics.jsonl",
+    "identity": "548345e958214d088eba2d71c2d62843af804742",
+    "size": 22217
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/resolved_config.yaml",
+    "identity": "2cdfedc9cd91a4f6a23b3100e28f7f287ef44f9d",
+    "size": 3760
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/episode_metrics.jsonl",
+    "identity": "5a94dc6743b97b2f49f650f4a52be30d12895eab",
+    "size": 22212
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/resolved_config.yaml",
+    "identity": "a0a33bb6477c25861f3921bcec6109d6c5fafad0",
+    "size": 3760
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/eval/post_train/step_5000.json",
+    "identity": "9e83d564db56e731bda47b93e9456934500e0431",
+    "size": 5068
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/episode_metrics.jsonl",
+    "identity": "480a026820255516c9918a751511eef16126288d",
+    "size": 22230
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/resolved_config.yaml",
+    "identity": "84400e1f52ef31c66632781e41b9cf6a74f5c818",
+    "size": 3991
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/episode_metrics.jsonl",
+    "identity": "809fe0be4e37608f3663d79a6f89fb5bbdcd712e",
+    "size": 23164
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/resolved_config.yaml",
+    "identity": "c6548c1b42f23e3fb2a9e5676065bdc467abf9da",
+    "size": 4031
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/episode_metrics.jsonl",
+    "identity": "5945c683ac3f2545629e8448c225f0ae6fc8b8c9",
+    "size": 23109
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/resolved_config.yaml",
+    "identity": "79853125bfc883ae521338f4ec20ea1198cd9a97",
+    "size": 4030
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/eval/post_train/step_5000.json",
+    "identity": "50cf79a1490d8577d3a43841861e9850e03db60c",
+    "size": 5084
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/episode_metrics.jsonl",
+    "identity": "60e26079e084639caaa0ddfb082b7f99a1315759",
+    "size": 22430
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/resolved_config.yaml",
+    "identity": "d75ac2d9124a92540af1c5fb1fee0bdb358deb58",
+    "size": 3662
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/episode_metrics.jsonl",
+    "identity": "efc89929eef160bb7cdd4b6fa27158e05608fd50",
+    "size": 23350
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/resolved_config.yaml",
+    "identity": "942ecfa155af572624c6bbfb82afa43f7e99a884",
+    "size": 3702
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/episode_metrics.jsonl",
+    "identity": "02b1bd21086d27648e29a61ac45590c5892f25eb",
+    "size": 23290
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/resolved_config.yaml",
+    "identity": "30b08edead4c80a1d1bf5787e1d65ed177462207",
+    "size": 3701
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/eval/post_train/step_5000.json",
+    "identity": "ca040a41c96dcc08ee4d1790630dfc924bebcebb",
+    "size": 5103
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/episode_metrics.jsonl",
+    "identity": "09916d7e76293f8b648c5e45412e1c65c543ea67",
+    "size": 23926
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/resolved_config.yaml",
+    "identity": "f4d9bbd2bdac5bfadf0d2f52342fd1f3deaa06e9",
+    "size": 2368
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/episode_metrics.jsonl",
+    "identity": "a7bd21815955784535a933ec23d66562a0a2dfb2",
+    "size": 24845
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/resolved_config.yaml",
+    "identity": "caa9e4703dc3d2075e60738306a89f5a70d35fe9",
+    "size": 2408
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/episode_metrics.jsonl",
+    "identity": "5762fd44f8e2afedebbfb54d8a42516be6623119",
+    "size": 24771
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_latency_cross_games_rerun_20260930_192352/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/resolved_config.yaml",
+    "identity": "25916621e72a3da2769dbcbb96fe0b18064385a5",
+    "size": 2407
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/eval/post_train/step_5000.json",
+    "identity": "cae2ec50e6d0a13636b9bbdfb1b5383f70f25f25",
+    "size": 5097
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/episode_metrics.jsonl",
+    "identity": "24974498f430e70aae88a2420574d08fe1b3a98c",
+    "size": 21648
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_0/resolved_config.yaml",
+    "identity": "9fa301001b06aaea318efb4d1ba63b9bf047bba0",
+    "size": 3711
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/episode_metrics.jsonl",
+    "identity": "86a5460a97f5ac65afefad8338f7c49b7916b7a2",
+    "size": 21782
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_2/resolved_config.yaml",
+    "identity": "58082a75e56c8549ce0ae49ba28c9e933718f764",
+    "size": 3738
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/episode_metrics.jsonl",
+    "identity": "9d18eaf8a0ef943ec886cf9c75a345e02543691c",
+    "size": 21757
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/air_raid/latency_bench_eval/post_train/step_5000/air_raid/latency_4/resolved_config.yaml",
+    "identity": "435f95193dc98ef32d8da8c7f663a892b06bd792",
+    "size": 3738
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/eval/post_train/step_5000.json",
+    "identity": "be99b5e38a8dbd3bbf28b3689e1212ce78b029d6",
+    "size": 5084
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/episode_metrics.jsonl",
+    "identity": "4a286dcb645a7b4f1a1cfa7dbd1742820324a7c4",
+    "size": 21794
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_0/resolved_config.yaml",
+    "identity": "1c89a21f62147d2fa743d0ae8a017fbbe3cb6771",
+    "size": 3969
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/episode_metrics.jsonl",
+    "identity": "4110baa72b58998fb52b59cf905233c9567ad241",
+    "size": 22727
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_2/resolved_config.yaml",
+    "identity": "9814af4e865e25960ac355e02d249dfc5d57c017",
+    "size": 4009
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/episode_metrics.jsonl",
+    "identity": "b935b2a3c2e4aa14b16de9cadd0aaf78c6d75d7a",
+    "size": 22649
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/asterix/latency_bench_eval/post_train/step_5000/asterix/latency_4/resolved_config.yaml",
+    "identity": "5d52573fa9b22178aa25b70c1822794015cd6f72",
+    "size": 4008
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/eval/post_train/step_5000.json",
+    "identity": "7c223c499d97b68b367e2556ecde8d52004c5ed6",
+    "size": 5165
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/episode_metrics.jsonl",
+    "identity": "4263b985afdebe3a2a3697092c091b5236fcb730",
+    "size": 21990
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_0/resolved_config.yaml",
+    "identity": "7b3c26b4026db82aeb0303adbe7b7717104ba045",
+    "size": 3640
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/episode_metrics.jsonl",
+    "identity": "ba767ebf6233847006b8bd7f832b2667d49fd5f9",
+    "size": 22910
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_2/resolved_config.yaml",
+    "identity": "5b29a81381fd6ed84d50b2b07a0964b2f54e3f12",
+    "size": 3680
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/episode_metrics.jsonl",
+    "identity": "c6a6f763015e9d8e01a17fab63c1cdd1b693a432",
+    "size": 22852
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/atlantis/latency_bench_eval/post_train/step_5000/atlantis/latency_4/resolved_config.yaml",
+    "identity": "79bcd4f0ca32899379261c5e792c5b4910f1e51c",
+    "size": 3679
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/eval/post_train/step_5000.json",
+    "identity": "112f8451196ae9298d580daf75986451aa535cbd",
+    "size": 5083
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/episode_metrics.jsonl",
+    "identity": "8743d5878d12883e06bb1cdc57181b1dbd7c4b68",
+    "size": 23478
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_0/resolved_config.yaml",
+    "identity": "ac7bb48f9999c93d4dc03b18c7b94a52a16b035b",
+    "size": 2524
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/episode_metrics.jsonl",
+    "identity": "b014157922f424e7d3c78511680b78e40d981997",
+    "size": 24410
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_2/resolved_config.yaml",
+    "identity": "4e82351a9e8ddeadafb08f22ca182a834ef6ca29",
+    "size": 2564
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/episode_metrics.jsonl",
+    "identity": "1fe1c529313d7c53371ceeb00193252793b1edbb",
+    "size": 24348
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/observation_attempts.jsonl",
+    "identity": "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391",
+    "size": 0
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "00ff82642633c6f0fe2e39f90cbd74308291d09e",
+    "path": "archive/task-transfer/evaluations/demon_attack_mixed_024_promptfix_full_20260930_212002/demon_attack/latency_bench_eval/post_train/step_5000/demon_attack/latency_4/resolved_config.yaml",
+    "identity": "27a235592dd04f5d03360935557fa334ad71c223",
+    "size": 2563
+  }
+]

experiments/task-transfer/provenance/data_readability_verification.json

@@ -0,0 +1,758 @@
+[
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/train-00000-of-00001.parquet",
+    "rows": 272266,
+    "row_groups": 5,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 2,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 2768,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 2,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/train-00000-of-00001.parquet",
+    "rows": 256820,
+    "row_groups": 5,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 6,
+      "latency_raw_frames": 2,
+      "latency_ms": 33.333333333333336,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 2 raw frames (33.33 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 2598,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 1,
+      "latency_raw_frames": 2,
+      "latency_ms": 33.333333333333336,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 2 raw frames (33.33 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/train-00000-of-00001.parquet",
+    "rows": 222525,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 5,
+      "latency_raw_frames": 4,
+      "latency_ms": 66.66666666666667,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 4 raw frames (66.67 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/asterix/asterix_200ep/asterix_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 2154,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 2,
+      "latency_raw_frames": 4,
+      "latency_ms": 66.66666666666667,
+      "prompt": "Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft. Current action latency is 4 raw frames (66.67 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": true,
+    "image_size": [
+      160,
+      210
+    ]
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet",
+    "rows": 28715,
+    "row_groups": 1,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet",
+    "rows": 27192,
+    "row_groups": 1,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 105,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 626,
+    "row_groups": 1,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 16,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": true,
+    "image_size": [
+      160,
+      120
+    ]
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00000-of-00003.parquet",
+    "rows": 21208,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 5,
+      "latency_raw_frames": 2,
+      "latency_ms": 57.142857142857146,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 2 raw frames (57.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00001-of-00003.parquet",
+    "rows": 21208,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 65,
+      "action_id": 1,
+      "latency_raw_frames": 2,
+      "latency_ms": 57.142857142857146,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 2 raw frames (57.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/train-00002-of-00003.parquet",
+    "rows": 21207,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 133,
+      "action_id": 5,
+      "latency_raw_frames": 2,
+      "latency_ms": 57.142857142857146,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 2 raw frames (57.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 644,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 5,
+      "latency_raw_frames": 2,
+      "latency_ms": 57.142857142857146,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 2 raw frames (57.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet",
+    "rows": 30125,
+    "row_groups": 5,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 5,
+      "latency_raw_frames": 4,
+      "latency_ms": 114.28571428571429,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 4 raw frames (114.29 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet",
+    "rows": 30124,
+    "row_groups": 5,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 96,
+      "action_id": 1,
+      "latency_raw_frames": 4,
+      "latency_ms": 114.28571428571429,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 4 raw frames (114.29 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 657,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 569",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 5,
+      "latency_raw_frames": 4,
+      "latency_ms": 114.28571428571429,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 4 raw frames (114.29 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00000-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00001-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 105,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00002-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 207,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00003-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 307,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00004-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 406,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00005-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 503,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00006-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 603,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00007-of-00010.parquet",
+    "rows": 28715,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 703,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00008-of-00010.parquet",
+    "rows": 28714,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 799,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/train-00009-of-00010.parquet",
+    "rows": 28714,
+    "row_groups": 5,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 903,
+      "action_id": 1,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 2844,
+    "row_groups": 1,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 16,
+      "action_id": 5,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": true,
+    "image_size": [
+      160,
+      120
+    ]
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/train.parquet",
+    "rows": 38308,
+    "row_groups": 150,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 22,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep/val.parquet",
+    "rows": 4295,
+    "row_groups": 17,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 2,
+      "action_id": 22,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/train.parquet",
+    "rows": 32500,
+    "row_groups": 127,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 29,
+      "latency_raw_frames": 8,
+      "latency_ms": 228.57142857142858,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 8 raw frames (228.57 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_2_1000ep/val.parquet",
+    "rows": 4000,
+    "row_groups": 16,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 2,
+      "action_id": 29,
+      "latency_raw_frames": 8,
+      "latency_ms": 228.57142857142858,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 8 raw frames (228.57 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/train.parquet",
+    "rows": 28737,
+    "row_groups": 113,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 24,
+      "latency_raw_frames": 16,
+      "latency_ms": 457.14285714285717,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 16 raw frames (457.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_4_1000ep/val.parquet",
+    "rows": 3178,
+    "row_groups": 13,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 2,
+      "action_id": 24,
+      "latency_raw_frames": 16,
+      "latency_ms": 457.14285714285717,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 16 raw frames (457.14 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": true,
+    "image_size": [
+      160,
+      120
+    ]
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/train.parquet",
+    "rows": 34290,
+    "row_groups": 134,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 26,
+      "latency_raw_frames": 24,
+      "latency_ms": 685.7142857142858,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 24 raw frames (685.71 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_6_1000ep/val.parquet",
+    "rows": 3890,
+    "row_groups": 16,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 2,
+      "action_id": 26,
+      "latency_raw_frames": 24,
+      "latency_ms": 685.7142857142858,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 24 raw frames (685.71 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/train.parquet",
+    "rows": 28895,
+    "row_groups": 113,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 26,
+      "latency_raw_frames": 32,
+      "latency_ms": 914.2857142857143,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 32 raw frames (914.29 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_8_1000ep/val.parquet",
+    "rows": 3245,
+    "row_groups": 13,
+    "schema": "action_id: int64\naction_text: string\ndecision_step: int64\nenv_name: string\nepisode_idx: int64\nimage: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nlatency_ms: double\nlatency_raw_frames: int64\nprompt: string\nraw_reward: double\nseed: int64\nsplit: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"action_id\": {\"dtype\": \"int64\", \"_ty' + 569",
+    "sample": {
+      "episode_idx": 2,
+      "action_id": 26,
+      "latency_raw_frames": 32,
+      "latency_ms": 914.2857142857143,
+      "prompt": "You are playing Deadly Corridor in VizDoom. Your goal is to survive the corridor, fight enemies, and reach the green armor vest at the far end. The available actions are: MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation. Current action latency is 32 raw frames (914.29 ms).  The environment runs at 35 FPS and observations are emitted at 8.75 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00000-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 0,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/train-00001-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 99,
+      "action_id": 0,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 3600,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 0,
+      "latency_raw_frames": 0,
+      "latency_ms": 0.0,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00000-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 2,
+      "latency_raw_frames": 2,
+      "latency_ms": 33.333333333333336,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 2 raw frames (33.33 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/train-00001-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 99,
+      "action_id": 0,
+      "latency_raw_frames": 2,
+      "latency_ms": 33.333333333333336,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 2 raw frames (33.33 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_2_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 3600,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 2,
+      "latency_raw_frames": 2,
+      "latency_ms": 33.333333333333336,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 2 raw frames (33.33 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": true,
+    "image_size": [
+      160,
+      210
+    ]
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00000-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 0,
+      "action_id": 0,
+      "latency_raw_frames": 4,
+      "latency_ms": 66.66666666666667,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 4 raw frames (66.67 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/train-00001-of-00002.parquet",
+    "rows": 178200,
+    "row_groups": 4,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 99,
+      "action_id": 0,
+      "latency_raw_frames": 4,
+      "latency_ms": 66.66666666666667,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 4 raw frames (66.67 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "d6f0ad0995f6e4b095f315920e8de18e04297048",
+    "path": "latency-aware/atlantis/atlantis_200ep/atlantis_fixed_latency_4_200ep_7k2steps/val-00000-of-00001.parquet",
+    "rows": 3600,
+    "row_groups": 1,
+    "schema": "image: struct<bytes: binary, path: string>\n  child 0, bytes: binary\n  child 1, path: string\nenv_name: string\nepisode_idx: int64\ndecision_step: int64\nsplit: string\nseed: int64\nprompt: string\nlatency_raw_frames: int64\nlatency_ms: double\nraw_reward: double\naction_id: int64\naction: list<element: float>\n  child 0, element: float\naction_text: string\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"env_na' + 649",
+    "sample": {
+      "episode_idx": 127,
+      "action_id": 0,
+      "latency_raw_frames": 4,
+      "latency_ms": 66.66666666666667,
+      "prompt": "Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire. Current action latency is 4 raw frames (66.67 ms).  The environment runs at 60 FPS and observations are emitted at 15 FPS. Choose the best next action."
+    },
+    "image_decoded": false
+  }
+]

experiments/task-transfer/provenance/fixed_reference_verification.json

@@ -0,0 +1,4 @@
+{
+  "references": 126,
+  "all_resolve": true
+}

experiments/task-transfer/provenance/release_code/scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh

@@ -0,0 +1,629 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Generate/upload fixed-latency rollout datasets for the latency-0/2/4
+# small-model teachers used to build mixed-latency data.
+#
+# New machine usage:
+#   cd /path/to/latency-sensitive-bench
+#   conda activate latency
+#   export HF_TOKEN=hf_...
+#   bash scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh all
+#
+# Or source and call individual functions:
+#   source scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh
+#   rollout_asterix_l0
+#   rollout_asterix_l2
+#   rollout_asterix_l4
+#   rollout_atlantis_l0
+#   rollout_atlantis_l2
+#   rollout_atlantis_l4
+#   rollout_defend_the_line_l0
+#   rollout_defend_the_line_l2
+#   rollout_defend_the_line_l4
+#
+# Common overrides:
+#   MAX_EPISODES=200
+#   MAX_STEPS=7200
+#   NUM_ENVS=64
+#   REPLAY_NUM_ENVS=64
+#   DEVICE=cuda
+#   FILTER_PRESET=none
+#   EPISODE_RETURN_GT=<override per-task gate>
+#   HF_MODEL_REPO=latency-sensitive-bench/paper-experiment-models
+#   HF_ROLLOUT_REPO=<override one shared upload repo>
+#   HF_ROLLOUT_REPO_PREFIX=latency-sensitive-bench
+#
+# Default teacher seeds are one selected small model per env/latency:
+#   asterix:  l0=s11, l2=s14, l4=s14
+#   atlantis: l0=s12, l2=s14, l4=s10
+#   defend:   l2=s0,  l4=s0
+# Defend-the-Line l0 is copied from the existing 1000ep memory-rollouts
+# dataset because no l0 small-model checkpoint is uploaded in HF_MODEL_REPO.
+
+export HF_MODEL_REPO="${HF_MODEL_REPO:-latency-sensitive-bench/paper-experiment-models}"
+export HF_ROLLOUT_REPO="${HF_ROLLOUT_REPO:-latency-sensitive-bench/benchmark-datasets}"
+export HF_ROLLOUT_REPO_PREFIX="${HF_ROLLOUT_REPO_PREFIX:-latency-sensitive-bench}"
+export SMALL_MODEL_CKPT_ROOT="${SMALL_MODEL_CKPT_ROOT:-/mnt/local/talha/latency024_small_model_ckpts}"
+export ROLLOUT_DATA_ROOT="${ROLLOUT_DATA_ROOT:-/mnt/local/talha/latency024_rollout_datasets}"
+export DEFEND_L0_SOURCE_REPO="${DEFEND_L0_SOURCE_REPO:-latency-sensitive-bench/benchmark-datasets}"
+export DEFEND_L0_SOURCE_CONFIG="${DEFEND_L0_SOURCE_CONFIG:-zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps}"
+export DEFEND_L0_SOURCE_REVISION="${DEFEND_L0_SOURCE_REVISION:-d6f0ad0995f6e4b095f315920e8de18e04297048}"
+
+need_repo_root() {
+  if [[ ! -f scripts/bash_scripts/run_sf_teacher_dataset_export.sh ]]; then
+    echo "[error] run this from the latency-sensitive-bench repo root." >&2
+    return 1
+  fi
+}
+
+need_cmd() {
+  local cmd="$1"
+  if ! command -v "${cmd}" >/dev/null 2>&1; then
+    echo "[error] missing command: ${cmd}" >&2
+    return 1
+  fi
+}
+
+setup_cuda_env() {
+  if [[ -z "${CUDA_HOME:-}" ]]; then
+    if [[ -n "${CONDA_PREFIX:-}" && -x "${CONDA_PREFIX}/bin/nvcc" ]]; then
+      export CUDA_HOME="${CONDA_PREFIX}"
+    elif [[ -d /usr/local/cuda ]]; then
+      export CUDA_HOME=/usr/local/cuda
+    fi
+  fi
+
+  if [[ -n "${CUDA_HOME:-}" ]]; then
+    export PATH="${CUDA_HOME}/bin:${PATH}"
+    export LD_LIBRARY_PATH="${CUDA_HOME}/lib:${CUDA_HOME}/lib64:${CUDA_HOME}/targets/x86_64-linux/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+  fi
+
+  export TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-8.6;9.0;12.0}"
+}
+
+tag_for_env() {
+  case "$1" in
+    asterix|atlantis) printf '%s\n' "10m_fs4_1min_eval" ;;
+    defend_the_line) printf '%s\n' "25m_fs4_1min_eval" ;;
+    *)
+      echo "[error] unknown env: $1" >&2
+      return 2
+      ;;
+  esac
+}
+
+model_family_for_env_latency() {
+  local env_name="$1"
+  local latency="$2"
+  local tag
+  tag="$(tag_for_env "${env_name}")" || return 1
+
+  if [[ "${latency}" == "0" ]]; then
+    if [[ "${env_name}" == "defend_the_line" ]]; then
+      echo "[error] no uploaded defend_the_line zero-latency small-model checkpoint was found in ${HF_MODEL_REPO}." >&2
+      return 1
+    fi
+    printf '%s\n' "zero_latency_${tag}"
+    return 0
+  fi
+
+  printf '%s\n' "latency_024_${tag}"
+}
+
+selected_seed_for_env_latency() {
+  local env_name="$1"
+  local latency="$2"
+  local override_key
+
+  override_key="$(printf '%s_L%s_SEED' "${env_name}" "${latency}" | tr '[:lower:]' '[:upper:]')"
+  if [[ -n "${!override_key:-}" ]]; then
+    printf '%s\n' "${!override_key}"
+    return 0
+  fi
+
+  case "${env_name}:${latency}" in
+    asterix:0) printf '%s\n' 11 ;;
+    asterix:2) printf '%s\n' 14 ;;
+    asterix:4) printf '%s\n' 14 ;;
+    atlantis:0) printf '%s\n' 12 ;;
+    atlantis:2) printf '%s\n' 14 ;;
+    atlantis:4) printf '%s\n' 10 ;;
+    defend_the_line:2) printf '%s\n' 0 ;;
+    defend_the_line:4) printf '%s\n' 0 ;;
+    defend_the_line:0)
+      echo "[error] no selected defend_the_line l0 teacher; no uploaded l0 checkpoint was found." >&2
+      return 1
+      ;;
+    *)
+      echo "[error] no selected seed for ${env_name} latency ${latency}" >&2
+      return 1
+      ;;
+  esac
+}
+
+default_episode_return_gt_for_env_latency() {
+  local env_name="$1"
+  local latency="$2"
+  case "${env_name}:${latency}" in
+    asterix:0) printf '%s\n' 6000 ;;
+    asterix:2) printf '%s\n' 6000 ;;
+    asterix:4) printf '%s\n' 6000 ;;
+    atlantis:0) printf '%s\n' 30000 ;;
+    atlantis:2) printf '%s\n' 30000 ;;
+    atlantis:4) printf '%s\n' 30000 ;;
+    defend_the_line:2) printf '%s\n' 29 ;;
+    defend_the_line:4) printf '%s\n' 29 ;;
+    defend_the_line:0)
+      echo "[error] no default return gate for defend_the_line l0; no uploaded l0 teacher was found." >&2
+      return 1
+      ;;
+    *)
+      echo "[error] no default return gate for ${env_name} latency ${latency}" >&2
+      return 1
+      ;;
+  esac
+}
+
+hf_rollout_repo_for_env() {
+  local env_name="$1"
+  if [[ -n "${HF_ROLLOUT_REPO:-}" ]]; then
+    printf '%s\n' "${HF_ROLLOUT_REPO}"
+    return 0
+  fi
+  printf '%s\n' "${HF_ROLLOUT_REPO_PREFIX}/${env_name}_200ep"
+}
+
+best_remote_ckpt() {
+  local env_name="$1"
+  local latency="$2"
+  local family seed
+  family="$(model_family_for_env_latency "${env_name}" "${latency}")" || return 1
+  seed="$(selected_seed_for_env_latency "${env_name}" "${latency}")" || return 1
+
+  local api_url="https://huggingface.co/api/models/${HF_MODEL_REPO}/tree/main/memory/small_model/${env_name}?recursive=true"
+
+  curl -fsSL "${api_url}" |
+    ENV_NAME="${env_name}" LATENCY="${latency}" FAMILY="${family}" SEED="${seed}" python3 -c '
+import json
+import os
+import re
+import sys
+
+env_name = os.environ["ENV_NAME"]
+latency = os.environ["LATENCY"]
+family = os.environ["FAMILY"]
+seed = os.environ["SEED"]
+
+pattern = re.compile(
+    r"^memory/small_model/" + re.escape(env_name) + r"/" + re.escape(family)
+    + r"/train/memory_small_model:" + re.escape(env_name)
+    + r":frame_stack:fixed_l" + re.escape(latency)
+    + r":fs4:obs[^:]+:stride4:seed" + re.escape(seed) + r"/checkpoint_p0/"
+    + r"best_.*reward_(-?\d+(?:\.\d+)?)\.pth$"
+)
+
+rows = []
+for item in json.load(sys.stdin):
+    path = item.get("path", "")
+    match = pattern.match(path)
+    if match:
+        rows.append((float(match.group(1)), path))
+
+if not rows:
+    raise SystemExit(1)
+
+rows.sort(key=lambda item: item[0])
+print(rows[-1][1])
+'
+}
+
+download_best_latency024_ckpt() {
+  local env_name="$1"
+  local latency="$2"
+
+  need_cmd curl || return 1
+  need_cmd hf || return 1
+
+  local remote_ckpt remote_exp_dir remote_config local_root ckpt_name ckpt
+  remote_ckpt="$(best_remote_ckpt "${env_name}" "${latency}")" || {
+    echo "[error] no HF checkpoint found for ${env_name} fixed_l${latency} in ${HF_MODEL_REPO}" >&2
+    return 1
+  }
+  remote_exp_dir="${remote_ckpt%/checkpoint_p0/*}"
+  remote_config="${remote_exp_dir}/config.json"
+  local_root="${SMALL_MODEL_CKPT_ROOT}/${env_name}/fixed_l${latency}"
+  mkdir -p "${local_root}"
+
+  echo "[download] ${HF_MODEL_REPO}:${remote_config}" >&2
+  hf download "${HF_MODEL_REPO}" "${remote_config}" \
+    --repo-type model \
+    --local-dir "${local_root}" >/dev/null
+
+  echo "[download] ${HF_MODEL_REPO}:${remote_ckpt}" >&2
+  hf download "${HF_MODEL_REPO}" "${remote_ckpt}" \
+    --repo-type model \
+    --local-dir "${local_root}" >/dev/null
+
+  ckpt_name="$(basename "${remote_ckpt}")"
+  ckpt="$(find "${local_root}" -type f -path "*/checkpoint_p0/${ckpt_name}" -print -quit)"
+  if [[ -z "${ckpt}" || ! -f "${ckpt}" ]]; then
+    echo "[error] downloaded checkpoint missing under ${local_root}" >&2
+    find "${local_root}" -type f -name "*.pth" -print >&2
+    return 1
+  fi
+
+  printf '%s\n' "${ckpt}"
+}
+
+rollout_latency024_one() {
+  local env_name="$1"
+  local latency="$2"
+
+  need_repo_root || return 1
+  setup_cuda_env
+
+  local ckpt config_name out_dir steps_label hf_rollout_repo
+  ckpt="$(download_best_latency024_ckpt "${env_name}" "${latency}")" || return 1
+  hf_rollout_repo="$(hf_rollout_repo_for_env "${env_name}")"
+
+  case "${MAX_STEPS:-7200}" in
+    7200) steps_label="7k2steps" ;;
+    3600) steps_label="3k6steps" ;;
+    *) steps_label="${MAX_STEPS:-7200}steps" ;;
+  esac
+
+  config_name="${env_name}_fixed_latency_${latency}_${MAX_EPISODES:-200}ep_${steps_label}"
+  out_dir="${ROLLOUT_DATA_ROOT}/${config_name}"
+  episode_return_gt="${EPISODE_RETURN_GT:-$(default_episode_return_gt_for_env_latency "${env_name}" "${latency}")}"
+
+  echo "[run] ${env_name} fixed_l${latency}"
+  echo "[ckpt] ${ckpt}"
+  echo "[out]  ${out_dir}"
+  echo "[hf]   ${hf_rollout_repo}/${config_name}"
+  echo "[gate] FILTER_PRESET=${FILTER_PRESET:-none} EPISODE_RETURN_GT=${episode_return_gt}"
+
+  MAX_EPISODES="${MAX_EPISODES:-200}" \
+  MAX_STEPS="${MAX_STEPS:-7200}" \
+  NUM_ENVS="${NUM_ENVS:-64}" \
+  REPLAY_NUM_ENVS="${REPLAY_NUM_ENVS:-64}" \
+  DEVICE="${DEVICE:-cuda}" \
+  CONTEXT_WINDOW="${CONTEXT_WINDOW:-1}" \
+  FILTER_PRESET="${FILTER_PRESET:-none}" \
+  EPISODE_RETURN_GT="${episode_return_gt}" \
+  IMAGE_WRITER_WORKERS="${IMAGE_WRITER_WORKERS:-16}" \
+  HF_REPO_ID="${hf_rollout_repo}" \
+  HF_CONFIG_NAME="task-transfer--${config_name}" \
+  HF_DATA_DIR="$(if [[ "${latency}" == "0" ]]; then echo zero-latency; else echo latency-aware; fi)/${env_name//_/-}/${env_name}_200ep/${config_name}" \
+  HF_PRIVATE="${HF_PRIVATE:-0}" \
+    bash scripts/bash_scripts/run_sf_teacher_dataset_export.sh \
+      "${ckpt}" \
+      "${out_dir}"
+}
+
+rollout_defend_the_line_l0_from_memory_rollouts() {
+  need_repo_root || return 1
+
+  local env_name="defend_the_line"
+  local latency="0"
+  local steps_label config_name out_dir hf_rollout_repo max_episodes val_fraction train_episodes val_episodes
+
+  case "${MAX_STEPS:-7200}" in
+    7200) steps_label="7k2steps" ;;
+    3600) steps_label="3k6steps" ;;
+    *) steps_label="${MAX_STEPS:-7200}steps" ;;
+  esac
+
+  max_episodes="${MAX_EPISODES:-200}"
+  val_fraction="${VAL_FRACTION:-0.01}"
+  val_episodes="$(
+    MAX_EPISODES="${max_episodes}" VAL_FRACTION="${val_fraction}" python3 - <<'PY'
+import math
+import os
+
+max_episodes = int(os.environ["MAX_EPISODES"])
+val_fraction = float(os.environ["VAL_FRACTION"])
+print(min(max_episodes, max(1, int(math.floor(max_episodes * val_fraction)))))
+PY
+  )"
+  train_episodes="$((max_episodes - val_episodes))"
+  if [[ "${train_episodes}" -le 0 ]]; then
+    echo "[error] MAX_EPISODES=${max_episodes} and VAL_FRACTION=${val_fraction} leave no train episodes." >&2
+    return 1
+  fi
+
+  config_name="${env_name}_fixed_latency_${latency}_${max_episodes}ep_${steps_label}"
+  out_dir="${ROLLOUT_DATA_ROOT}/${config_name}"
+  hf_rollout_repo="$(hf_rollout_repo_for_env "${env_name}")"
+
+  mkdir -p "${out_dir}"
+
+  echo "[run] ${env_name} fixed_l${latency} from existing memory-rollouts dataset"
+  echo "[source] ${DEFEND_L0_SOURCE_REPO}/${DEFEND_L0_SOURCE_CONFIG}@${DEFEND_L0_SOURCE_REVISION}"
+  echo "[out]    ${out_dir}"
+  echo "[hf]     ${hf_rollout_repo}/${config_name}"
+  echo "[split]  train=${train_episodes} val=${val_episodes} VAL_FRACTION=${val_fraction}"
+
+  SOURCE_REPO="${DEFEND_L0_SOURCE_REPO}" \
+  SOURCE_CONFIG="${DEFEND_L0_SOURCE_CONFIG}" \
+  SOURCE_REVISION="${DEFEND_L0_SOURCE_REVISION}" \
+  DEST_REPO="${hf_rollout_repo}" \
+  DEST_CONFIG="zero-latency/${env_name//_/-}/${env_name}_200ep/${config_name}" \
+  OUTPUT_DIR="${out_dir}" \
+  TRAIN_EPISODES="${train_episodes}" \
+  VAL_EPISODES="${val_episodes}" \
+  HF_PRIVATE="${HF_PRIVATE:-0}" \
+  HF_TOKEN="${HF_TOKEN:-}" \
+  HF_MAX_SHARD_SIZE="${HF_MAX_SHARD_SIZE:-500MB}" \
+    python3 - <<'PY'
+import json
+import os
+from pathlib import Path
+
+import pyarrow as pa
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+from huggingface_hub import HfApi, hf_hub_download
+
+
+def _token() -> str | None:
+    return os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN") or None
+
+
+def _bool_env(name: str, default: str = "0") -> bool:
+    return str(os.environ.get(name, default)).lower() in {"1", "true", "yes", "on"}
+
+
+def _source_split_files(api: HfApi, *, repo: str, config: str, revision: str, split: str, token: str | None) -> list[str]:
+    prefix = f"{config}/{split}-"
+    files = api.list_repo_files(repo_id=repo, repo_type="dataset", revision=revision, token=token)
+    matches = sorted(path for path in files if path.startswith(prefix) and path.endswith(".parquet"))
+    if not matches:
+        raise RuntimeError(f"no source parquet files found for {repo}/{config} split={split}")
+    return matches
+
+
+def _download(repo: str, revision: str, path: str, token: str | None) -> str:
+    print(f"[download] {path}", flush=True)
+    return hf_hub_download(
+        repo_id=repo,
+        repo_type="dataset",
+        revision=revision,
+        filename=path,
+        token=token,
+    )
+
+
+def _select_episode_ids(
+    *,
+    repo: str,
+    revision: str,
+    files: list[str],
+    budget: int,
+    token: str | None,
+) -> tuple[list[int], list[str]]:
+    selected: list[int] = []
+    seen: set[int] = set()
+    downloaded: list[str] = []
+    for path in files:
+        local = _download(repo, revision, path, token)
+        downloaded.append(local)
+        table = pq.read_table(local, columns=["episode_idx"])
+        for value in table.column("episode_idx").to_pylist():
+            episode_idx = int(value)
+            if episode_idx in seen:
+                continue
+            seen.add(episode_idx)
+            selected.append(episode_idx)
+            if len(selected) >= budget:
+                print(
+                    f"[select] selected {len(selected)} episodes "
+                    f"first={selected[0]} last={selected[-1]}",
+                    flush=True,
+                )
+                return selected, downloaded
+    raise RuntimeError(f"source only exposed {len(selected)} unique episodes; requested {budget}")
+
+
+def _write_filtered_split(
+    *,
+    split: str,
+    selected: list[int],
+    local_files: list[str],
+    output_dir: Path,
+) -> dict[str, int | list[str]]:
+    selected_array = pa.array(selected, type=pa.int64())
+    tmp_paths: list[Path] = []
+    total_rows = 0
+    for local in local_files:
+        episode_idx_col = pq.read_table(local, columns=["episode_idx"]).column("episode_idx")
+        mask = pc.is_in(episode_idx_col, value_set=selected_array)
+        if not bool(pc.any(mask).as_py()):
+            continue
+        table = pq.read_table(local)
+        filtered = table.filter(mask)
+        if filtered.num_rows == 0:
+            continue
+        tmp_path = output_dir / f"{split}-tmp-{len(tmp_paths):05d}.parquet"
+        pq.write_table(filtered, tmp_path)
+        tmp_paths.append(tmp_path)
+        total_rows += filtered.num_rows
+        print(f"[write] {split}: {tmp_path.name} rows={filtered.num_rows}", flush=True)
+
+    final_names: list[str] = []
+    total_files = len(tmp_paths)
+    if total_files == 0:
+        raise RuntimeError(f"no rows written for split={split}")
+    for idx, tmp_path in enumerate(tmp_paths):
+        final_path = output_dir / f"{split}-{idx:05d}-of-{total_files:05d}.parquet"
+        tmp_path.replace(final_path)
+        final_names.append(final_path.name)
+    print(f"[write] {split}: rows={total_rows} files={total_files}", flush=True)
+    return {"rows": total_rows, "files": final_names}
+
+
+repo = os.environ["SOURCE_REPO"]
+config = os.environ["SOURCE_CONFIG"]
+revision = os.environ["SOURCE_REVISION"]
+dest_repo = os.environ["DEST_REPO"]
+dest_config = os.environ["DEST_CONFIG"]
+output_dir = Path(os.environ["OUTPUT_DIR"])
+train_budget = int(os.environ["TRAIN_EPISODES"])
+val_budget = int(os.environ["VAL_EPISODES"])
+token = _token()
+api = HfApi()
+
+output_dir.mkdir(parents=True, exist_ok=True)
+for stale in output_dir.glob("*.parquet"):
+    stale.unlink()
+
+print(f"[source] {repo}/{config}@{revision}", flush=True)
+train_files = _source_split_files(api, repo=repo, config=config, revision=revision, split="train", token=token)
+val_files = _source_split_files(api, repo=repo, config=config, revision=revision, split="val", token=token)
+
+print(f"[select] train budget={train_budget}", flush=True)
+train_ids, train_local_files = _select_episode_ids(
+    repo=repo,
+    revision=revision,
+    files=train_files,
+    budget=train_budget,
+    token=token,
+)
+print(f"[select] val budget={val_budget}", flush=True)
+val_ids, val_local_files = _select_episode_ids(
+    repo=repo,
+    revision=revision,
+    files=val_files,
+    budget=val_budget,
+    token=token,
+)
+
+train_stats = _write_filtered_split(split="train", selected=train_ids, local_files=train_local_files, output_dir=output_dir)
+val_stats = _write_filtered_split(split="val", selected=val_ids, local_files=val_local_files, output_dir=output_dir)
+
+metadata = {
+    "source_repo": repo,
+    "source_config": config,
+    "source_revision": revision,
+    "derived_config": dest_config,
+    "train_episodes": len(train_ids),
+    "val_episodes": len(val_ids),
+    "total_episodes": len(train_ids) + len(val_ids),
+    "train_episode_idx": train_ids,
+    "val_episode_idx": val_ids,
+    "train_rows": int(train_stats["rows"]),
+    "val_rows": int(val_stats["rows"]),
+    "total_rows": int(train_stats["rows"]) + int(val_stats["rows"]),
+    "train_files": train_stats["files"],
+    "val_files": val_stats["files"],
+}
+(output_dir / "metadata.json").write_text(json.dumps(metadata, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+
+print(f"[upload] repo={dest_repo} path={dest_config}", flush=True)
+api.create_repo(repo_id=dest_repo, repo_type="dataset", private=_bool_env("HF_PRIVATE"), exist_ok=True, token=token)
+commit = api.upload_folder(
+    folder_path=str(output_dir),
+    path_in_repo=dest_config,
+    repo_id=dest_repo,
+    repo_type="dataset",
+    allow_patterns=["*.parquet", "metadata.json"],
+    commit_message=f"Upload {dest_config} from {config}",
+    token=token,
+)
+metadata["commit_url"] = getattr(commit, "commit_url", None)
+(output_dir / "metadata.json").write_text(json.dumps(metadata, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+api.upload_file(
+    path_or_fileobj=str(output_dir / "metadata.json"),
+    path_in_repo=f"{dest_config}/metadata.json",
+    repo_id=dest_repo,
+    repo_type="dataset",
+    token=token,
+)
+print("[upload] complete", flush=True)
+print(json.dumps(metadata, indent=2, sort_keys=True))
+PY
+}
+
+rollout_asterix_l0() { rollout_latency024_one asterix 0; }
+rollout_asterix_l2() { rollout_latency024_one asterix 2; }
+rollout_asterix_l4() { rollout_latency024_one asterix 4; }
+
+rollout_atlantis_l0() { rollout_latency024_one atlantis 0; }
+rollout_atlantis_l2() { rollout_latency024_one atlantis 2; }
+rollout_atlantis_l4() { rollout_latency024_one atlantis 4; }
+
+rollout_defend_the_line_l0() { rollout_defend_the_line_l0_from_memory_rollouts; }
+rollout_defend_the_line_l2() { rollout_latency024_one defend_the_line 2; }
+rollout_defend_the_line_l4() { rollout_latency024_one defend_the_line 4; }
+
+rollout_asterix_all() {
+  rollout_asterix_l0
+  rollout_asterix_l2
+  rollout_asterix_l4
+}
+
+rollout_atlantis_all() {
+  rollout_atlantis_l0
+  rollout_atlantis_l2
+  rollout_atlantis_l4
+}
+
+rollout_defend_the_line_all() {
+  rollout_defend_the_line_l0
+  rollout_defend_the_line_l2
+  rollout_defend_the_line_l4
+}
+
+rollout_all_latency024() {
+  rollout_asterix_all
+  rollout_atlantis_all
+  rollout_defend_the_line_all
+}
+
+print_help() {
+  cat <<'EOF'
+Usage:
+  bash scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh <target>
+
+Targets:
+  asterix-l0 | asterix-l2 | asterix-l4 | asterix | asterix-all
+  atlantis-l0 | atlantis-l2 | atlantis-l4 | atlantis | atlantis-all
+  defend-the-line-l0 | defend-the-line-l2 | defend-the-line-l4
+  defend-the-line | defend-the-line-all
+  all
+
+Note:
+  defend-the-line-l0 is derived from
+  latency-sensitive-bench/benchmark-datasets/zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps
+  by streaming the requested number of episodes into this repo's standard 200ep
+  rollout destination.
+
+Source mode:
+  source scripts/bash_scripts/rollout_data/rollout_latency024_three_games.sh
+  rollout_asterix_l0
+EOF
+}
+
+if [[ "${BASH_SOURCE[0]}" == "$0" ]]; then
+  case "${1:-help}" in
+    asterix-l0) rollout_asterix_l0 ;;
+    asterix-l2) rollout_asterix_l2 ;;
+    asterix-l4) rollout_asterix_l4 ;;
+    asterix|asterix-all) rollout_asterix_all ;;
+    atlantis-l0) rollout_atlantis_l0 ;;
+    atlantis-l2) rollout_atlantis_l2 ;;
+    atlantis-l4) rollout_atlantis_l4 ;;
+    atlantis|atlantis-all) rollout_atlantis_all ;;
+    defend-the-line-l0|defend_the_line-l0) rollout_defend_the_line_l0 ;;
+    defend-the-line-l2|defend_the_line-l2) rollout_defend_the_line_l2 ;;
+    defend-the-line-l4|defend_the_line-l4) rollout_defend_the_line_l4 ;;
+    defend-the-line|defend_the_line|defend-the-line-all|defend_the_line-all) rollout_defend_the_line_all ;;
+    all) rollout_all_latency024 ;;
+    help|-h|--help) print_help ;;
+    *)
+      echo "[error] unknown target: ${1}" >&2
+      print_help >&2
+      exit 2
+      ;;
+  esac
+fi

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/gr00t/bridge/single/deadly_corridor_multibinary_bce.sh

@@ -0,0 +1,40 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh gr00t deadly_corridor
+
+conda activate starvla_rl_games_gr00t
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=gr00t \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="gr00t_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/gr00t_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/gr00t_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/gr00t/bridge/single/deadly_corridor_new_lr.sh

@@ -0,0 +1,45 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh gr00t deadly_corridor
+
+conda activate starvla_rl_games_gr00t
+
+python examples/rl_games/scripts/launch_train.py \
+    model=gr00t \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="gr00t_bridge_deadly_corridor_single_latency_clean_data_exp2_action_1e-4_backbone_1e-5" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/gr00t_bridge_deadly_corridor_single_latency_clean_data_exp2_action_1e-4_backbone_1e-5" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/gr00t_bridge_deadly_corridor_single_latency_clean_data_exp2_action_1e-4_backbone_1e-5" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=true \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    trainer.logging_frequency=1 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    trainer.learning_rate.base=1e-05 \
+    trainer.learning_rate.qwen_vl_interface=1e-05 \
+    trainer.learning_rate.action_model=1e-04 \
+    rl_games.env_eval.mid_train.enabled=true \
+    rl_games.env_eval.mid_train.latencies=[0] \
+    rl_games.env_eval.mid_train.interval_steps=50 \
+    rl_games.env_eval.post_train.num_episodes=4 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600 \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,2,4,6,8] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/deadly_zero_demon_mixed.sh

@@ -0,0 +1,67 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla cross_task
+
+conda activate starvla_rl_games_openvla
+
+pip install flash-attn --no-build-isolation
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=deadly_zero_demon_mixed \
+    run_id="openvla_bridge_cross_deadly_zero_demon_024_exp1" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha1503/openvla_bridge_cross_deadly_zero_demon_024_exp1" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha1503/openvla_bridge_cross_deadly_zero_demon_024_exp1" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=7000 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=1400 \
+    trainer.save_interval=7000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_200ep_cross_0_train \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/deadly_zero_flappy_mixed.sh

@@ -0,0 +1,71 @@
+bash /workspace/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh
+
+cd /workspace/starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla flappy
+
+conda activate starvla_rl_games_openvla
+
+bash /workspace/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh 
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=deadly_zero_flappy_mixed \
+    run_id="openvla_bridge_cross_deadly_zero_flappy_024_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha1503/openvla_bridge_cross_deadly_zero_flappy_024_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha1503/openvla_bridge_cross_deadly_zero_flappy_024_exp2" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=7000 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=1400 \
+    trainer.save_interval=7000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    trainer.eval_action_classification=true \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_200ep_cross_0_train \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/defendtheline_deadly/defendtheline_zero_deadly_zero.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla defend_the_line
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=defendtheline_zero_deadly_zero \
+    run_id="openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="latency-sensitive-bench/benchmark-models" \
+    checkpoint.sync.enabled=false \
+    checkpoint.sync.repo_id="latency-sensitive-bench/benchmark-models" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=4000 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=defend_the_line \
+    rl_games.cross_task.train_tasks.0.converted_name=defend_the_line_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_0_cross_train_openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.defend_the_line.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.defend_the_line.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.defend_the_line.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.defend_the_line.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_mixed_deadly_mixed.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla demon_attack
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=demon_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_demon_024_deadly_024_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_demon_024_deadly_024_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_demon_024_deadly_024_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2600 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2600 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_024_cross_train_openvla_bridge_cross_demon_024_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train_openvla_bridge_cross_demon_024_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_mixed_deadly_zero.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla demon_attack
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=demon_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_demon_024_deadly_zero_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_demon_024_deadly_zero_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_demon_024_deadly_zero_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2200 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2200 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_024_cross_train_openvla_bridge_cross_demon_024_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_0_cross_train_openvla_bridge_cross_demon_024_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_zero_deadly_mixed.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla demon_attack
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=demon_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_demon_zero_deadly_024_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_demon_zero_deadly_024_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_demon_zero_deadly_024_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2400 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2400 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_0_cross_train_openvla_bridge_cross_demon_zero_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train_openvla_bridge_cross_demon_zero_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=400 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_deadly/demon_zero_deadly_zero.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla demon_attack
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=demon_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_demon_zero_deadly_zero_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_demon_zero_deadly_zero_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_demon_zero_deadly_zero_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2000 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_0_cross_train_openvla_bridge_cross_demon_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_0_cross_train_openvla_bridge_cross_demon_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/demon_zero_deadly_mixed.sh

@@ -0,0 +1,67 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla cross_task
+
+conda activate starvla_rl_games_openvla
+
+pip install flash-attn --no-build-isolation
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=demon_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_demon_zero_deadly_024_exp1" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha1503/openvla_bridge_cross_demon_zero_deadly_024_exp1" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha1503/openvla_bridge_cross_demon_zero_deadly_024_exp1" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=7000 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=1400 \
+    trainer.save_interval=7000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=demon_attack \
+    rl_games.cross_task.train_tasks.0.converted_name=demon_attack_200ep_cross_0_train \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/deadly_mixed.sh

@@ -0,0 +1,57 @@
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=mixed \
+    run_id="openvla_bridge_deadly_corridor_latency_mixed_024_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_latency_mixed_024_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_latency_mixed_024_exp3" \
+    checkpoint.save_best_model=false \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    dataset.latency_filter=[0,2,4] \
+    dataset.episodes_per_latency=400 \
+    dataset.converted_name=deadly_corridor_024_openvla_bridge_mixed_exp3 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.max_train_steps=2200 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2200 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,2,4] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_mixed_deadly_mixed.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR=""
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla flappy
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=flappy_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_flappy_024_deadly_024_exp3 " \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_flappy_024_deadly_024_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_flappy_024_deadly_024_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2200 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2200 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_200ep_cross_024_train_openvla_bridge_cross_flappy_024_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train_openvla_bridge_cross_flappy_024_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_mixed_deadly_zero.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR=""
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla flappy
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=flappy_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_flappy_024_deadly_zero_exp3 " \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_flappy_024_deadly_zero_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_flappy_024_deadly_zero_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2200 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2200 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_200ep_cross_024_train_openvla_bridge_cross_flappy_024_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_0_cross_train_openvla_bridge_cross_flappy_024_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_zero_deadly_mixed.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR=""
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla flappy
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=flappy_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_flappy_zero_deadly_024_exp3" \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_flappy_zero_deadly_024_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_flappy_zero_deadly_024_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2400 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2400 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_200ep_cross_0_train_openvla_bridge_cross_flappy_zero_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_024_cross_train_openvla_bridge_cross_flappy_zero_deadly_024_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=400 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_deadly/flappy_zero_deadly_zero.sh

@@ -0,0 +1,79 @@
+WORKSPACE_DIR=""
+export WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla flappy
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=flappy_zero_deadly_mixed \
+    run_id="openvla_bridge_cross_flappy_zero_deadly_zero_exp3 " \
+    trainer.distributed_backend=none \
+    workspace_dir="$WORKSPACE_DIR" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_cross_flappy_zero_deadly_zero_exp3" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_cross_flappy_zero_deadly_zero_exp3" \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=2000 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=400 \
+    trainer.save_interval=2000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.per_latency_eval_num_batches=5 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_200ep_cross_0_train_openvla_bridge_cross_flappy_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=40 \
+    rl_games.cross_task.train_tasks.1.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.1.converted_name=deadly_corridor_0_cross_train_openvla_bridge_cross_flappy_zero_deadly_zero_exp3 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter=[0] \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter=[0,2,4] \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=1000 \
+    rl_games.cross_task.train_tasks.1.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0 \
+    rl_games.cross_task.train_tasks.0.max_episodes=null \
+    rl_games.cross_task.train_tasks.1.max_episodes=null \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.flappy.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.flappy.post_train.max_steps_per_episode=3600 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies=[0,2,4] \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.num_episodes=20 \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/flappy_demon_deadly_asterix_atlantis_defend_024.sh

@@ -0,0 +1,169 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export HF_ENDPOINT=https://huggingface.co
+
+WORKSPACE_DIR="${WORKSPACE_DIR:-/home/ubuntu/talha}"
+DATA_WORKSPACE_DIR="${DATA_WORKSPACE_DIR:-/mnt/local/talha}"
+export WORKSPACE_DIR DATA_WORKSPACE_DIR
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh"
+
+cd "${WORKSPACE_DIR}/starVLA"
+
+bash examples/rl_games/install/install_stack.sh openvla cross_task
+
+CONDA_BASE="$(conda info --base)"
+# shellcheck source=/dev/null
+source "${CONDA_BASE}/etc/profile.d/conda.sh"
+conda activate starvla_rl_games_openvla
+
+bash "${WORKSPACE_DIR}/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh"
+
+mkdir -p "${DATA_WORKSPACE_DIR}"
+mkdir -p "${DATA_WORKSPACE_DIR}/.cache/huggingface"
+export HF_HOME="${DATA_WORKSPACE_DIR}/.cache/huggingface"
+export HUGGINGFACE_HUB_CACHE="${DATA_WORKSPACE_DIR}/.cache/huggingface/hub"
+export HF_HUB_CACHE="${DATA_WORKSPACE_DIR}/.cache/huggingface/hub"
+export TRANSFORMERS_CACHE="${DATA_WORKSPACE_DIR}/.cache/huggingface/transformers"
+export PYTHONPATH="${WORKSPACE_DIR}/latency-sensitive-bench:${PYTHONPATH:-}"
+
+RUN_ID="${RUN_ID:-openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1}"
+HF_REPO_ID="${HF_REPO_ID:-latency-sensitive-bench/benchmark-models}"
+MAX_EPISODES_PER_ENV="${MAX_EPISODES_PER_ENV:-200}"
+MAX_EPISODES_DEADLY="${MAX_EPISODES_DEADLY:-1000}"
+MIX_LATENCIES="${MIX_LATENCIES:-[0,2,4]}"
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=cross_task \
+    init=bridge \
+    mode=cross_task \
+    cross_task_setup=flappy_demon_deadly_asterix_atlantis_defend_024 \
+    run_id="${RUN_ID}" \
+    trainer.distributed_backend=none \
+    workspace_dir="${DATA_WORKSPACE_DIR}" \
+    paths.run_root_dir="${DATA_WORKSPACE_DIR}/results/Checkpoints" \
+    paths.dataset_local_dir="${DATA_WORKSPACE_DIR}/playground/Datasets/rl_games" \
+    paths.base_model_dir="${DATA_WORKSPACE_DIR}/playground/Pretrained_models/Qwen3-VL-4B-Instruct" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="latency-sensitive-bench/benchmark-models" \
+    checkpoint.sync.enabled=false \
+    checkpoint.sync.repo_id="latency-sensitive-bench/benchmark-models" \
+    checkpoint.save_best_model=false \
+    checkpoint.local.keep_last_n=1 \
+    trainer.max_train_steps=15000 \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=15000 \
+    trainer.save_interval=15000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=8 \
+    trainer.per_latency_eval_num_batches=5 \
+    ++wandb_tags='["final_vla_profiling_training","demon_attack","openvla","rtx3090_profile"]' \
+    ++training_latency_condition=rtx3090_profile \
+    ++wandb_group="final_vla_profiling_training" \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies="${MIX_LATENCIES}" \
+    trainer.eval_action_classification=true \
+    datasets.vla_data.per_device_batch_size=32 \
+    datasets.vla_data.include_state=false \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    framework.action_model.loss_type=discrete_ce \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.action_layout=bridge_cross_task_7 \
+    rl_games.cross_task.train_tasks.0.name=flappy \
+    rl_games.cross_task.train_tasks.0.converted_name=flappy_mixed_024_200ep_cross_train \
+    rl_games.cross_task.train_tasks.0.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.0.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.0.train_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.prompt_source_subdir=zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.0.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.0.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.0.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.0.max_episodes="${MAX_EPISODES_PER_ENV}" \
+    rl_games.cross_task.train_tasks.1.name=demon_attack \
+    rl_games.cross_task.train_tasks.1.converted_name=demon_attack_mixed_024_200ep_cross_train \
+    rl_games.cross_task.train_tasks.1.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.train_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.1.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.1.prompt_revision=5305590625c28a6922e73f46d18d52738762c906 \
+    rl_games.cross_task.train_tasks.1.train_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.1.prompt_source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    rl_games.cross_task.train_tasks.1.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.1.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.1.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.1.max_episodes="${MAX_EPISODES_PER_ENV}" \
+    rl_games.cross_task.train_tasks.2.name=deadly_corridor \
+    rl_games.cross_task.train_tasks.2.converted_name=deadly_corridor_mixed_024_1000ep_cross_train \
+    rl_games.cross_task.train_tasks.2.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.2.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.2.prompt_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.2.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.2.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.2.train_source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    rl_games.cross_task.train_tasks.2.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.2.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.2.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.2.max_episodes="${MAX_EPISODES_DEADLY}" \
+    rl_games.cross_task.train_tasks.2.action_layout=multibinary_7 \
+    rl_games.cross_task.train_tasks.3.name=asterix \
+    rl_games.cross_task.train_tasks.3.converted_name=asterix_mixed_024_200ep_cross_train \
+    rl_games.cross_task.train_tasks.3.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.3.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.3.prompt_source_subdir=zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.3.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.3.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.3.train_source_subdir=zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.3.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.3.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.3.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.3.max_episodes="${MAX_EPISODES_PER_ENV}" \
+    rl_games.cross_task.train_tasks.3.action_layout=factorized_6 \
+    rl_games.cross_task.train_tasks.4.name=atlantis \
+    rl_games.cross_task.train_tasks.4.converted_name=atlantis_mixed_024_200ep_cross_train \
+    rl_games.cross_task.train_tasks.4.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.4.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.4.prompt_source_subdir=zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.4.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.4.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.4.train_source_subdir=zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.4.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.4.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.4.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.4.max_episodes="${MAX_EPISODES_PER_ENV}" \
+    rl_games.cross_task.train_tasks.5.name=defend_the_line \
+    rl_games.cross_task.train_tasks.5.converted_name=defend_the_line_mixed_024_200ep_cross_train \
+    rl_games.cross_task.train_tasks.5.prompt_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.5.prompt_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.5.prompt_source_subdir=zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.5.train_source_hf=latency-sensitive-bench/benchmark-datasets \
+    rl_games.cross_task.train_tasks.5.train_revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    rl_games.cross_task.train_tasks.5.train_source_subdir=zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps \
+    rl_games.cross_task.train_tasks.5.train_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.5.eval_latency_filter="${MIX_LATENCIES}" \
+    rl_games.cross_task.train_tasks.5.episodes_per_latency=null \
+    rl_games.cross_task.train_tasks.5.max_episodes="${MAX_EPISODES_PER_ENV}" \
+    rl_games.cross_task.eval_tasks.flappy.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.flappy.post_train.enabled=true \
+    rl_games.cross_task.eval_tasks.flappy.post_train.latencies="${MIX_LATENCIES}" \
+    rl_games.cross_task.eval_tasks.demon_attack.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.enabled=true \
+    rl_games.cross_task.eval_tasks.demon_attack.post_train.latencies="${MIX_LATENCIES}" \
+    rl_games.cross_task.eval_tasks.deadly_corridor.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.enabled=true \
+    rl_games.cross_task.eval_tasks.deadly_corridor.post_train.latencies="${MIX_LATENCIES}" \
+    rl_games.cross_task.eval_tasks.asterix.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.asterix.post_train.enabled=false \
+    rl_games.cross_task.eval_tasks.atlantis.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.atlantis.post_train.enabled=false \
+    rl_games.cross_task.eval_tasks.defend_the_line.mid_train.enabled=false \
+    rl_games.cross_task.eval_tasks.defend_the_line.post_train.enabled=false \
+    rl_games.env_eval.eval_backend=eval_core \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    rl_games.env_eval.deadly.multibinary_threshold=0.0

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/cross_task/rollout_defend_asterix_atlantis_024_small_models.sh

@@ -0,0 +1,246 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Roll out Sample Factory small models for the next cross-task data mix.
+#
+# Order:
+#   1. defend_the_line latencies 0,2,4
+#   2. asterix         latencies 0,2,4
+#   3. atlantis        latencies 0,2,4
+#
+# Each latency exports 200 accepted episodes, capped at 7200 raw env frames,
+# then publishes named configurations in benchmark-datasets.
+
+WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
+BENCH_DIR="${BENCH_DIR:-${WORKSPACE_DIR}/latency-sensitive-bench}"
+CONDA_ENV="${CONDA_ENV:-latency}"
+
+HF_MODEL_REPO_ID="${HF_MODEL_REPO_ID:-latency-sensitive-bench/paper-experiment-models}"
+DEFEND_THE_LINE_HF_REPO="${DEFEND_THE_LINE_HF_REPO:-latency-sensitive-bench/benchmark-datasets}"
+ASTERIX_HF_REPO="${ASTERIX_HF_REPO:-latency-sensitive-bench/benchmark-datasets}"
+ATLANTIS_HF_REPO="${ATLANTIS_HF_REPO:-latency-sensitive-bench/benchmark-datasets}"
+HF_PRIVATE="${HF_PRIVATE:-0}"
+HF_MAX_SHARD_SIZE="${HF_MAX_SHARD_SIZE:-500MB}"
+
+LOCAL_MODEL_ROOT="${LOCAL_MODEL_ROOT:-${WORKSPACE_DIR}/hf_small_models/paper-experiment-models}"
+OUTPUT_ROOT="${OUTPUT_ROOT:-${WORKSPACE_DIR}/small_model_rollouts/defend_asterix_atlantis_024}"
+
+EPISODES_PER_LATENCY="${EPISODES_PER_LATENCY:-200}"
+MAX_STEPS="${MAX_STEPS:-7200}"
+VAL_FRACTION="${VAL_FRACTION:-0.01}"
+NUM_ENVS="${NUM_ENVS:-12}"
+REPLAY_NUM_ENVS="${REPLAY_NUM_ENVS:-16}"
+DEVICE="${DEVICE:-cuda}"
+EXPORT_IMAGE_SIZE="${EXPORT_IMAGE_SIZE:-224,224}"
+IMAGE_WRITER_WORKERS="${IMAGE_WRITER_WORKERS:-16}"
+CONTEXT_WINDOW="${CONTEXT_WINDOW:-1}"
+FILTER_PRESET="${FILTER_PRESET:-none}"
+SPLIT_SEED="${SPLIT_SEED:-0}"
+
+# Asterix/Atlantis have five trained seeds in the generated manifest. Use the
+# first seed by default; override if you want another checkpoint.
+DEFEND_THE_LINE_SEED="${DEFEND_THE_LINE_SEED:-0}"
+ASTERIX_SEED="${ASTERIX_SEED:-10}"
+ATLANTIS_SEED="${ATLANTIS_SEED:-10}"
+
+LATENCIES=(0 2 4)
+ENVS=(defend_the_line asterix atlantis)
+
+usage() {
+  cat >&2 <<'EOF'
+Usage:
+  bash rollout_defend_asterix_atlantis_024_small_models.sh
+
+Expected on the Vast box:
+  WORKSPACE_DIR=/workspace
+  /workspace/latency-sensitive-bench exists
+  conda env "latency" exists
+  HF_TOKEN or HUGGINGFACE_HUB_TOKEN is set if the repos are private or uploads require auth
+
+Useful overrides:
+  BENCH_DIR=/path/to/latency-sensitive-bench
+  CONDA_ENV=latency
+  LOCAL_MODEL_ROOT=/mnt/models/paper-experiment-models
+  OUTPUT_ROOT=/mnt/data/small_model_rollouts
+  NUM_ENVS=24 REPLAY_NUM_ENVS=24 IMAGE_WRITER_WORKERS=24
+  FILTER_PRESET=none
+  EPISODE_RETURN_GT=1000
+  DEFEND_THE_LINE_EPISODE_RETURN_GT=...
+  ASTERIX_EPISODE_RETURN_GT=...
+  ATLANTIS_EPISODE_RETURN_GT=...
+  ASTERIX_SEED=11 ATLANTIS_SEED=14
+  HF_PRIVATE=1
+  DEFEND_THE_LINE_HF_REPO=latency-sensitive-bench/benchmark-datasets
+  ASTERIX_HF_REPO=latency-sensitive-bench/benchmark-datasets
+  ATLANTIS_HF_REPO=latency-sensitive-bench/benchmark-datasets
+EOF
+}
+
+if [[ "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+  usage
+  exit 0
+fi
+
+if [[ ! -d "${BENCH_DIR}" ]]; then
+  echo "latency-sensitive-bench not found at BENCH_DIR=${BENCH_DIR}" >&2
+  exit 2
+fi
+
+cd "${BENCH_DIR}"
+mkdir -p "${LOCAL_MODEL_ROOT}" "${OUTPUT_ROOT}"
+
+repo_for_env() {
+  case "$1" in
+    defend_the_line) echo "${DEFEND_THE_LINE_HF_REPO}" ;;
+    asterix) echo "${ASTERIX_HF_REPO}" ;;
+    atlantis) echo "${ATLANTIS_HF_REPO}" ;;
+    *) echo "Unknown env: $1" >&2; exit 2 ;;
+  esac
+}
+
+seed_for_env() {
+  case "$1" in
+    defend_the_line) echo "${DEFEND_THE_LINE_SEED}" ;;
+    asterix) echo "${ASTERIX_SEED}" ;;
+    atlantis) echo "${ATLANTIS_SEED}" ;;
+    *) echo "Unknown env: $1" >&2; exit 2 ;;
+  esac
+}
+
+manifest_for_env() {
+  echo "${BENCH_DIR}/configs/train/memory/small_model/$1/latency_024/manifest.tsv"
+}
+
+episode_return_gt_for_env() {
+  case "$1" in
+    defend_the_line) echo "${DEFEND_THE_LINE_EPISODE_RETURN_GT:-${EPISODE_RETURN_GT:-}}" ;;
+    asterix) echo "${ASTERIX_EPISODE_RETURN_GT:-${EPISODE_RETURN_GT:-}}" ;;
+    atlantis) echo "${ATLANTIS_EPISODE_RETURN_GT:-${EPISODE_RETURN_GT:-}}" ;;
+    *) echo "Unknown env: $1" >&2; exit 2 ;;
+  esac
+}
+
+manifest_value() {
+  local manifest="$1"
+  local latency="$2"
+  local seed="$3"
+  local field="$4"
+  python - "$manifest" "$latency" "$seed" "$field" <<'PY'
+import csv
+import sys
+
+manifest, latency, seed, field = sys.argv[1:5]
+with open(manifest, encoding="utf-8", newline="") as handle:
+    rows = [
+        row for row in csv.DictReader(handle, delimiter="\t")
+        if row["latency_fixed_frames"] == latency and row["seed"] == seed
+    ]
+if not rows:
+    raise SystemExit(
+        f"No manifest row in {manifest} for latency_fixed_frames={latency}, seed={seed}"
+    )
+print(rows[0][field])
+PY
+}
+
+download_model_artifact() {
+  local repo_path="$1"
+  conda run --no-capture-output -n "${CONDA_ENV}" python - "$HF_MODEL_REPO_ID" "$repo_path" "$LOCAL_MODEL_ROOT" <<'PY'
+import sys
+from pathlib import Path
+from huggingface_hub import snapshot_download
+
+repo_id, repo_path, local_root = sys.argv[1:4]
+Path(local_root).mkdir(parents=True, exist_ok=True)
+snapshot_download(
+    repo_id=repo_id,
+    repo_type="model",
+    allow_patterns=[f"{repo_path}/**"],
+    local_dir=local_root,
+)
+print(str(Path(local_root) / repo_path))
+PY
+}
+
+export_one_latency() {
+  local env_name="$1"
+  local latency="$2"
+  local seed repo_id manifest repo_path eval_config checkpoint_root output_dir hf_config_name return_gt hf_data_dir condition task_slug
+
+  seed="$(seed_for_env "${env_name}")"
+  repo_id="$(repo_for_env "${env_name}")"
+  manifest="$(manifest_for_env "${env_name}")"
+  repo_path="$(manifest_value "${manifest}" "${latency}" "${seed}" train_artifact_repo_path)"
+  eval_config="$(manifest_value "${manifest}" "${latency}" "${seed}" eval_config)"
+  checkpoint_root="${LOCAL_MODEL_ROOT}/${repo_path}"
+  output_dir="${OUTPUT_ROOT}/${env_name}_fixed_latency_${latency}_${EPISODES_PER_LATENCY}ep_7k2steps"
+  hf_config_name="${env_name}_fixed_latency_${latency}_${EPISODES_PER_LATENCY}ep_7k2steps"
+  return_gt="$(episode_return_gt_for_env "${env_name}")"
+  task_slug="${env_name//_/-}"
+  condition=latency-aware
+  if [[ "${latency}" == "0" ]]; then condition=zero-latency; fi
+  hf_data_dir="${condition}/${task_slug}/${env_name}_200ep/${hf_config_name}"
+
+  echo
+  echo "==> ${env_name} latency=${latency} seed=${seed}"
+  echo "    model repo: ${HF_MODEL_REPO_ID}/${repo_path}"
+  echo "    dataset:    ${repo_id} config=${hf_config_name}"
+
+  download_model_artifact "${repo_path}"
+
+  cmd=(
+    conda run --no-capture-output -n "${CONDA_ENV}"
+    python scripts/rollout_data/export_sf_teacher_dataset.py
+    --checkpoint-root "${checkpoint_root}"
+    --output-dir "${output_dir}"
+    --latency-config "${BENCH_DIR}/${eval_config}"
+    --max-episodes "${EPISODES_PER_LATENCY}"
+    --max-steps-per-episode "${MAX_STEPS}"
+    --val-fraction "${VAL_FRACTION}"
+    --split-seed "${SPLIT_SEED}"
+    --num-envs "${NUM_ENVS}"
+    --replay-num-envs "${REPLAY_NUM_ENVS}"
+    --device "${DEVICE}"
+    --export-image-size "${EXPORT_IMAGE_SIZE}"
+    --image-writer-workers "${IMAGE_WRITER_WORKERS}"
+    --context-window "${CONTEXT_WINDOW}"
+    --filter-preset "${FILTER_PRESET}"
+    --push-to-hub
+    --hf-repo-id "${repo_id}"
+    --hf-config-name "task-transfer--${hf_config_name}"
+    --hf-data-dir "${hf_data_dir}"
+    --hf-max-shard-size "${HF_MAX_SHARD_SIZE}"
+    --hf-token "${HF_TOKEN:-${HUGGINGFACE_HUB_TOKEN:-}}"
+  )
+
+  if [[ "${HF_PRIVATE}" == "1" || "${HF_PRIVATE}" == "true" ]]; then
+    cmd+=(--hf-private)
+  fi
+  if [[ -n "${return_gt}" ]]; then
+    cmd+=(--episode-return-gt "${return_gt}")
+  fi
+
+  "${cmd[@]}"
+}
+
+echo "==> Small-model rollout plan"
+echo "    bench:       ${BENCH_DIR}"
+echo "    conda env:   ${CONDA_ENV}"
+echo "    model repo:  ${HF_MODEL_REPO_ID}"
+echo "    output root: ${OUTPUT_ROOT}"
+echo "    latencies:   ${LATENCIES[*]}"
+echo "    episodes:    ${EPISODES_PER_LATENCY} per latency"
+echo "    max steps:   ${MAX_STEPS}"
+echo "    filter:      ${FILTER_PRESET}"
+
+for env_name in "${ENVS[@]}"; do
+  for latency in "${LATENCIES[@]}"; do
+    export_one_latency "${env_name}" "${latency}"
+  done
+done
+
+echo
+echo "==> Done. Pushed:"
+echo "    ${DEFEND_THE_LINE_HF_REPO}"
+echo "    ${ASTERIX_HF_REPO}"
+echo "    ${ATLANTIS_HF_REPO}"

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/eval/demon_attack_mixed_024_latency_024_cross_games.sh

@@ -0,0 +1,322 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export HF_ENDPOINT=https://huggingface.co
+
+# Eval-only sweep for the Demon Attack mixed-0/2/4-latency OpenVLA bridge policy.
+# The script generates a reproduction config. Original training metadata is missing.
+# Evaluation uses latency labels on Demon Attack, Asterix, Atlantis and Air Raid.
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../../../../../.." && pwd)"
+WORKSPACE_DIR="${WORKSPACE_DIR:-$(cd "${REPO_ROOT}/.." && pwd)}"
+export WORKSPACE_DIR
+
+MODEL_TAG="demon_attack_mixed_024_latency"
+HF_REPO_ID="${HF_REPO_ID:-latency-sensitive-bench/benchmark-models}"
+HF_REVISION="${HF_REVISION:-671e2a7451fda7f7dba5606c6d11183ebd804b6d}"
+HF_SUBDIR="${HF_SUBDIR:-latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_latency_mixed_024_exp1}"
+CHECKPOINT_REL="${CHECKPOINT_REL:-checkpoints/steps_5000_model.safetensors}"
+CHECKPOINT_STEP="${CHECKPOINT_STEP:-5000}"
+
+TASKS="${TASKS:-demon_attack,asterix,atlantis,air_raid}"
+LATENCIES="${LATENCIES:-0,2,4}"
+NUM_EPISODES="${NUM_EPISODES:-20}"
+MAX_STEPS_PER_EPISODE="${MAX_STEPS_PER_EPISODE:-3600}"
+GPUS="${GPUS:-}"
+WANDB_ENABLED="${WANDB_ENABLED:-false}"
+WANDB_PROJECT="${WANDB_PROJECT:-}"
+WANDB_ENTITY="${WANDB_ENTITY:-}"
+WANDB_RUN_PREFIX="${WANDB_RUN_PREFIX:-${MODEL_TAG}_latency024}"
+PRINT_PLAN_ONLY="${PRINT_PLAN_ONLY:-0}"
+RUN_INSTALL="${RUN_INSTALL:-0}"
+RUN_LATENCY_DEPS="${RUN_LATENCY_DEPS:-0}"
+ACTIVATE_CONDA="${ACTIVATE_CONDA:-1}"
+CONDA_ENV="${CONDA_ENV:-starvla_rl_games_openvla}"
+
+HF_CACHE_ROOT="${HF_CACHE_ROOT:-${REPO_ROOT}/results/HFCheckpoints}"
+RUN_ROOT="${RUN_ROOT:-${REPO_ROOT}/results/Evals/${MODEL_TAG}_cross_games}"
+
+cd "${REPO_ROOT}"
+
+detect_latency_bench_root() {
+  if [[ -n "${LATENCY_BENCH_ROOT:-}" && -d "${LATENCY_BENCH_ROOT}" ]]; then
+    return
+  fi
+  if [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench" ]]; then
+    LATENCY_BENCH_ROOT="${WORKSPACE_DIR}/latency-sensitive-bench"
+  elif [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench-realtime-human" ]]; then
+    LATENCY_BENCH_ROOT="${WORKSPACE_DIR}/latency-sensitive-bench-realtime-human"
+  else
+    LATENCY_BENCH_ROOT=""
+  fi
+  export LATENCY_BENCH_ROOT
+}
+
+activate_conda_env() {
+  if [[ "${ACTIVATE_CONDA}" != "1" ]]; then
+    return
+  fi
+  if command -v conda >/dev/null 2>&1; then
+    # shellcheck source=/dev/null
+    source "$(conda info --base)/etc/profile.d/conda.sh"
+  elif [[ -f "${HOME}/miniconda3/etc/profile.d/conda.sh" ]]; then
+    # shellcheck source=/dev/null
+    source "${HOME}/miniconda3/etc/profile.d/conda.sh"
+  elif [[ -f "${HOME}/miniconda/etc/profile.d/conda.sh" ]]; then
+    # shellcheck source=/dev/null
+    source "${HOME}/miniconda/etc/profile.d/conda.sh"
+  else
+    echo "[error] conda not found; set ACTIVATE_CONDA=0 if your shell is already activated" >&2
+    exit 2
+  fi
+  conda activate "${CONDA_ENV}"
+}
+
+download_hf_snapshot() {
+  local repo_id="$1"
+  local revision="$2"
+  local local_dir="$3"
+  shift 3
+  python - "$repo_id" "$revision" "$local_dir" "$@" <<'PY'
+import sys
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+repo_id, revision, local_dir, *patterns = sys.argv[1:]
+Path(local_dir).mkdir(parents=True, exist_ok=True)
+snapshot_download(
+    repo_id=repo_id,
+    repo_type="model",
+    revision=revision or None,
+    local_dir=local_dir,
+    allow_patterns=patterns or None,
+)
+PY
+}
+
+write_eval_config() {
+  local output_config="$1"
+  local output_dir
+  output_dir="$(dirname "${output_config}")"
+  mkdir -p "${output_dir}"
+  python - "${output_config}" "${WORKSPACE_DIR}" "${RUN_ROOT}" "${HF_REPO_ID}" <<'PY'
+import sys
+from pathlib import Path
+
+from hydra import compose, initialize_config_dir
+from omegaconf import OmegaConf
+
+output_config, workspace_dir, run_root, hf_repo_id = sys.argv[1:]
+repo_root = Path.cwd()
+config_dir = repo_root / "examples" / "rl_games" / "config"
+
+overrides = [
+    "model=openvla",
+    "env=demon_attack",
+    "init=bridge",
+    "mode=mixed_latency",
+    "run_id=openvla_bridge_demon_attack_latency_mixed_024_exp1",
+    f"run_root_dir={run_root}",
+    f"workspace_dir={workspace_dir}",
+    "trainer.distributed_backend=none",
+    "wandb_entity=talha1503",
+    f"checkpoint.hf_repo_id={hf_repo_id}",
+    "checkpoint.sync.enabled=false",
+    f"checkpoint.sync.repo_id={hf_repo_id}",
+    "dataset.source_hf=latency-sensitive-bench/benchmark-datasets",
+    "dataset.revision=15d17287246bd19940c837d6c8a5327564f969c5",
+    "dataset.source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep",
+    "dataset.latency_filter=[0,2,4]",
+    "datasets.vla_data.sequential_step_sampling=true",
+    "datasets.vla_data.shuffle=true",
+    "trainer.per_latency_eval_num_batches=5",
+    "dataset.episodes_per_latency=40",
+    "rl_games.env_eval.eval_backend=latency_bench",
+    "checkpoint.local.keep_last_n=1",
+    "trainer.max_train_steps=5000",
+    "trainer.num_warmup_steps=0",
+    "checkpoint.save_final_model=true",
+    "checkpoint.save_best_model=false",
+    "trainer.eval_interval=5000",
+    "trainer.save_interval=5000",
+    "trainer.logging_frequency=1",
+    "trainer.gradient_accumulation_steps=4",
+    "datasets.vla_data.per_device_batch_size=32",
+    "rl_games.env_eval.mid_train.enabled=false",
+    "rl_games.env_eval.post_train.enabled=true",
+    "rl_games.env_eval.post_train.latencies=[0,2,4]",
+    "rl_games.env_eval.post_train.num_episodes=20",
+    "rl_games.env_eval.post_train.max_steps_per_episode=3600",
+]
+
+with initialize_config_dir(config_dir=str(config_dir), version_base=None):
+    cfg = compose(config_name="train", overrides=overrides)
+
+OmegaConf.resolve(cfg)
+OmegaConf.save(config=cfg, f=output_config)
+PY
+}
+
+require_file() {
+  local path="$1"
+  if [[ ! -f "${path}" ]]; then
+    echo "[error] expected file is missing: ${path}" >&2
+    exit 2
+  fi
+}
+
+is_print_plan_only() {
+  case "${PRINT_PLAN_ONLY}" in
+    1|true|TRUE|yes|YES|on|ON) return 0 ;;
+    *) return 1 ;;
+  esac
+}
+
+add_common_eval_args() {
+  EVAL_ARGS=(
+    --step "${CHECKPOINT_STEP}"
+    --stage post_train
+    --latencies "${LATENCIES}"
+    --num-episodes "${NUM_EPISODES}"
+    --max-steps-per-episode "${MAX_STEPS_PER_EPISODE}"
+    --wandb-enabled "${WANDB_ENABLED}"
+    --workspace-dir "${WORKSPACE_DIR}"
+    --base-model-repo-id "Qwen/Qwen3-VL-4B-Instruct"
+    --override "rl_games.env_eval.eval_backend=latency_bench"
+    --override "rl_games.env_eval.mid_train.enabled=false"
+    --override "rl_games.env_eval.post_train.enabled=true"
+    --override "rl_games.env_eval.frameskip=4"
+    --override "rl_games.env_eval.latency.prompt_map_path=null"
+  )
+  if [[ -n "${GPUS}" ]]; then
+    EVAL_ARGS+=(--gpus "${GPUS}")
+  fi
+  if [[ -n "${WANDB_PROJECT}" ]]; then
+    EVAL_ARGS+=(--wandb-project "${WANDB_PROJECT}")
+  fi
+  if [[ -n "${WANDB_ENTITY}" ]]; then
+    EVAL_ARGS+=(--wandb-entity "${WANDB_ENTITY}")
+  fi
+  if is_print_plan_only; then
+    EVAL_ARGS+=(--print-plan-only)
+  fi
+}
+
+add_task_overrides() {
+  local task="$1"
+  TASK_OVERRIDES=(
+    --override "env=${task}"
+    --override "rl_games.task=${task}"
+  )
+  case "${task}" in
+    demon_attack)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'"
+        --override "rl_games.env_eval.demon_attack.noop_max=30"
+      )
+      ;;
+    asterix)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft.'"
+        --override "rl_games.env_eval.asterix.action_layout=factorized_6"
+        --override "rl_games.env_eval.atari.env_id=ALE/Asterix-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=4"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+      )
+      ;;
+    atlantis)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire.'"
+        --override "rl_games.env_eval.atari.env_id=ALE/Atlantis-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=4"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+      )
+      ;;
+    air_raid)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.'"
+        --override "rl_games.env_eval.atari.env_id=ALE/AirRaid-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=1"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.fire_reset=true"
+        --override "rl_games.env_eval.atari.mode=1"
+        --override "rl_games.env_eval.atari.difficulty=0"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+        --override "rl_games.env_eval.atari.max_num_frames_per_episode=108000"
+      )
+      ;;
+    *)
+      echo "[error] unsupported task '${task}'" >&2
+      exit 2
+      ;;
+  esac
+}
+
+detect_latency_bench_root
+
+if [[ "${RUN_INSTALL}" == "1" ]]; then
+  bash examples/rl_games/install/install_stack.sh openvla demon_attack
+fi
+
+activate_conda_env
+
+if [[ -n "${LATENCY_BENCH_ROOT}" ]]; then
+  export PYTHONPATH="${LATENCY_BENCH_ROOT}:${PYTHONPATH:-}"
+fi
+
+if [[ "${RUN_LATENCY_DEPS}" == "1" ]]; then
+  if [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench/third_party/flappy-bird-gymnasium" ]]; then
+    bash examples/rl_games/bash_scripts/install/latency_deps.sh
+  else
+    echo "[warn] skipping latency_deps.sh because ${WORKSPACE_DIR}/latency-sensitive-bench is not present" >&2
+  fi
+fi
+
+HF_LOCAL_DIR="${HF_CACHE_ROOT}/openvla_bridge_demon_attack_latency_mixed_024_exp1"
+if ! is_print_plan_only; then
+  download_hf_snapshot \
+    "${HF_REPO_ID}" \
+    "${HF_REVISION}" \
+    "${HF_LOCAL_DIR}" \
+    "${HF_SUBDIR}/${CHECKPOINT_REL}"
+fi
+
+CONFIG_PATH="${HF_LOCAL_DIR}/${HF_SUBDIR}/eval_config.generated.yaml"
+CHECKPOINT_PATH="${HF_LOCAL_DIR}/${HF_SUBDIR}/${CHECKPOINT_REL}"
+write_eval_config "${CONFIG_PATH}"
+require_file "${CONFIG_PATH}"
+if ! is_print_plan_only; then
+  require_file "${CHECKPOINT_PATH}"
+fi
+
+IFS=',' read -r -a TASK_ARRAY <<< "${TASKS}"
+add_common_eval_args
+
+for raw_task in "${TASK_ARRAY[@]}"; do
+  task="$(echo "${raw_task}" | xargs)"
+  [[ -n "${task}" ]] || continue
+  add_task_overrides "${task}"
+  TASK_RUN_DIR="${RUN_ROOT}/${task}"
+  mkdir -p "${TASK_RUN_DIR}"
+
+  CMD=(
+    bash examples/rl_games/scripts/run_eval.sh
+    --run-dir "${TASK_RUN_DIR}"
+    --config "${CONFIG_PATH}"
+    --checkpoint "${CHECKPOINT_PATH}"
+    --wandb-run-name "${WANDB_RUN_PREFIX}_${task}"
+    "${EVAL_ARGS[@]}"
+    "${TASK_OVERRIDES[@]}"
+  )
+
+  printf '[eval:%s]' "${task}"
+  printf ' %q' "${CMD[@]}"
+  echo
+  "${CMD[@]}"
+done

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/eval/demon_attack_zero_latency_024_cross_games.sh

@@ -0,0 +1,268 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export HF_ENDPOINT=https://huggingface.co
+
+# Eval-only sweep for the Demon Attack fixed-0-latency OpenVLA bridge policy.
+# Evaluates latency labels, not raw frame skips, on Demon Attack, Asterix,
+# Atlantis, and Air Raid.
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../../../../../.." && pwd)"
+WORKSPACE_DIR="${WORKSPACE_DIR:-$(cd "${REPO_ROOT}/.." && pwd)}"
+export WORKSPACE_DIR
+
+MODEL_TAG="demon_attack_zero_latency"
+HF_REPO_ID="${HF_REPO_ID:-latency-sensitive-bench/benchmark-models}"
+HF_REVISION="${HF_REVISION:-671e2a7451fda7f7dba5606c6d11183ebd804b6d}"
+HF_SUBDIR="${HF_SUBDIR:-zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep}"
+CHECKPOINT_REL="${CHECKPOINT_REL:-checkpoints/steps_7000_model.safetensors}"
+CHECKPOINT_STEP="${CHECKPOINT_STEP:-7000}"
+
+TASKS="${TASKS:-demon_attack,asterix,atlantis,air_raid}"
+LATENCIES="${LATENCIES:-0,2,4}"
+NUM_EPISODES="${NUM_EPISODES:-20}"
+MAX_STEPS_PER_EPISODE="${MAX_STEPS_PER_EPISODE:-3600}"
+GPUS="${GPUS:-}"
+WANDB_ENABLED="${WANDB_ENABLED:-false}"
+WANDB_PROJECT="${WANDB_PROJECT:-}"
+WANDB_ENTITY="${WANDB_ENTITY:-}"
+WANDB_RUN_PREFIX="${WANDB_RUN_PREFIX:-${MODEL_TAG}_latency024}"
+PRINT_PLAN_ONLY="${PRINT_PLAN_ONLY:-0}"
+RUN_INSTALL="${RUN_INSTALL:-0}"
+RUN_LATENCY_DEPS="${RUN_LATENCY_DEPS:-0}"
+ACTIVATE_CONDA="${ACTIVATE_CONDA:-1}"
+CONDA_ENV="${CONDA_ENV:-starvla_rl_games_openvla}"
+
+HF_CACHE_ROOT="${HF_CACHE_ROOT:-${REPO_ROOT}/results/HFCheckpoints}"
+RUN_ROOT="${RUN_ROOT:-${REPO_ROOT}/results/Evals/${MODEL_TAG}_latency024_cross_games}"
+
+cd "${REPO_ROOT}"
+
+detect_latency_bench_root() {
+  if [[ -n "${LATENCY_BENCH_ROOT:-}" && -d "${LATENCY_BENCH_ROOT}" ]]; then
+    return
+  fi
+  if [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench" ]]; then
+    LATENCY_BENCH_ROOT="${WORKSPACE_DIR}/latency-sensitive-bench"
+  elif [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench-realtime-human" ]]; then
+    LATENCY_BENCH_ROOT="${WORKSPACE_DIR}/latency-sensitive-bench-realtime-human"
+  else
+    LATENCY_BENCH_ROOT=""
+  fi
+  export LATENCY_BENCH_ROOT
+}
+
+activate_conda_env() {
+  if [[ "${ACTIVATE_CONDA}" != "1" ]]; then
+    return
+  fi
+  if command -v conda >/dev/null 2>&1; then
+    # shellcheck source=/dev/null
+    source "$(conda info --base)/etc/profile.d/conda.sh"
+  elif [[ -f "${HOME}/miniconda3/etc/profile.d/conda.sh" ]]; then
+    # shellcheck source=/dev/null
+    source "${HOME}/miniconda3/etc/profile.d/conda.sh"
+  elif [[ -f "${HOME}/miniconda/etc/profile.d/conda.sh" ]]; then
+    # shellcheck source=/dev/null
+    source "${HOME}/miniconda/etc/profile.d/conda.sh"
+  else
+    echo "[error] conda not found; set ACTIVATE_CONDA=0 if your shell is already activated" >&2
+    exit 2
+  fi
+  conda activate "${CONDA_ENV}"
+}
+
+download_hf_snapshot() {
+  local repo_id="$1"
+  local revision="$2"
+  local local_dir="$3"
+  shift 3
+  python - "$repo_id" "$revision" "$local_dir" "$@" <<'PY'
+import sys
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+repo_id, revision, local_dir, *patterns = sys.argv[1:]
+Path(local_dir).mkdir(parents=True, exist_ok=True)
+snapshot_download(
+    repo_id=repo_id,
+    repo_type="model",
+    revision=revision or None,
+    local_dir=local_dir,
+    allow_patterns=patterns or None,
+)
+PY
+}
+
+require_file() {
+  local path="$1"
+  if [[ ! -f "${path}" ]]; then
+    echo "[error] expected file is missing: ${path}" >&2
+    exit 2
+  fi
+}
+
+is_print_plan_only() {
+  case "${PRINT_PLAN_ONLY}" in
+    1|true|TRUE|yes|YES|on|ON) return 0 ;;
+    *) return 1 ;;
+  esac
+}
+
+add_common_eval_args() {
+  EVAL_ARGS=(
+    --step "${CHECKPOINT_STEP}"
+    --stage post_train
+    --latencies "${LATENCIES}"
+    --num-episodes "${NUM_EPISODES}"
+    --max-steps-per-episode "${MAX_STEPS_PER_EPISODE}"
+    --wandb-enabled "${WANDB_ENABLED}"
+    --workspace-dir "${WORKSPACE_DIR}"
+    --base-model-repo-id "Qwen/Qwen3-VL-4B-Instruct"
+    --override "rl_games.env_eval.eval_backend=latency_bench"
+    --override "rl_games.env_eval.mid_train.enabled=false"
+    --override "rl_games.env_eval.post_train.enabled=true"
+    --override "rl_games.env_eval.frameskip=4"
+    --override "rl_games.env_eval.latency.prompt_map_path=null"
+  )
+  if [[ -n "${GPUS}" ]]; then
+    EVAL_ARGS+=(--gpus "${GPUS}")
+  fi
+  if [[ -n "${WANDB_PROJECT}" ]]; then
+    EVAL_ARGS+=(--wandb-project "${WANDB_PROJECT}")
+  fi
+  if [[ -n "${WANDB_ENTITY}" ]]; then
+    EVAL_ARGS+=(--wandb-entity "${WANDB_ENTITY}")
+  fi
+  if is_print_plan_only; then
+    EVAL_ARGS+=(--print-plan-only)
+  fi
+}
+
+add_task_overrides() {
+  local task="$1"
+  TASK_OVERRIDES=(
+    --override "env=${task}"
+    --override "rl_games.task=${task}"
+  )
+  case "${task}" in
+    demon_attack)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'"
+        --override "rl_games.env_eval.demon_attack.noop_max=30"
+      )
+      ;;
+    asterix)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Collect useful objects and avoid lyres. Choose exactly one action from: noop, up, right, left, down, upright, upleft, downright, downleft.'"
+        --override "rl_games.env_eval.asterix.action_layout=factorized_6"
+        --override "rl_games.env_eval.atari.env_id=ALE/Asterix-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=4"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+      )
+      ;;
+    atlantis)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Defend Atlantis by firing at descending enemies. Choose exactly one action from: noop, fire, rightfire, leftfire.'"
+        --override "rl_games.env_eval.atari.env_id=ALE/Atlantis-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=4"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+      )
+      ;;
+    air_raid)
+      TASK_OVERRIDES+=(
+        --override "rl_games.env_eval.task_description='Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.'"
+        --override "rl_games.env_eval.atari.env_id=ALE/AirRaid-v5"
+        --override "rl_games.env_eval.atari.gym_frameskip=1"
+        --override "rl_games.env_eval.atari.noop_max=30"
+        --override "rl_games.env_eval.atari.fire_reset=true"
+        --override "rl_games.env_eval.atari.mode=1"
+        --override "rl_games.env_eval.atari.difficulty=0"
+        --override "rl_games.env_eval.atari.repeat_action_probability=0.0"
+        --override "rl_games.env_eval.atari.full_action_space=false"
+        --override "rl_games.env_eval.atari.max_num_frames_per_episode=108000"
+      )
+      ;;
+    *)
+      echo "[error] unsupported task '${task}'" >&2
+      exit 2
+      ;;
+  esac
+}
+
+detect_latency_bench_root
+
+if [[ "${RUN_INSTALL}" == "1" ]]; then
+  bash examples/rl_games/install/install_stack.sh openvla demon_attack
+fi
+
+activate_conda_env
+
+if [[ -n "${LATENCY_BENCH_ROOT}" ]]; then
+  export PYTHONPATH="${LATENCY_BENCH_ROOT}:${PYTHONPATH:-}"
+fi
+
+if [[ "${RUN_LATENCY_DEPS}" == "1" ]]; then
+  if [[ -d "${WORKSPACE_DIR}/latency-sensitive-bench/third_party/flappy-bird-gymnasium" ]]; then
+    bash examples/rl_games/bash_scripts/install/latency_deps.sh
+  else
+    echo "[warn] skipping latency_deps.sh because ${WORKSPACE_DIR}/latency-sensitive-bench is not present" >&2
+  fi
+fi
+
+HF_LOCAL_DIR="${HF_CACHE_ROOT}/openvla_demon_attack_200ep"
+HF_ALLOW_PATTERNS=(
+  "${HF_SUBDIR}/config.yaml"
+  "${HF_SUBDIR}/config.full.yaml"
+  "${HF_SUBDIR}/dataset_statistics.json"
+  "${HF_SUBDIR}/dataset_statistics_eval.json"
+  "${HF_SUBDIR}/hydra/.hydra/config.yaml"
+  "${HF_SUBDIR}/hydra/.hydra/hydra.yaml"
+  "${HF_SUBDIR}/hydra/.hydra/overrides.yaml"
+)
+if ! is_print_plan_only; then
+  HF_ALLOW_PATTERNS+=("${HF_SUBDIR}/${CHECKPOINT_REL}")
+fi
+download_hf_snapshot \
+  "${HF_REPO_ID}" \
+  "${HF_REVISION}" \
+  "${HF_LOCAL_DIR}" \
+  "${HF_ALLOW_PATTERNS[@]}"
+
+CONFIG_PATH="${HF_LOCAL_DIR}/${HF_SUBDIR}/config.full.yaml"
+CHECKPOINT_PATH="${HF_LOCAL_DIR}/${HF_SUBDIR}/${CHECKPOINT_REL}"
+require_file "${CONFIG_PATH}"
+if ! is_print_plan_only; then
+  require_file "${CHECKPOINT_PATH}"
+fi
+
+IFS=',' read -r -a TASK_ARRAY <<< "${TASKS}"
+add_common_eval_args
+
+for raw_task in "${TASK_ARRAY[@]}"; do
+  task="$(echo "${raw_task}" | xargs)"
+  [[ -n "${task}" ]] || continue
+  add_task_overrides "${task}"
+  TASK_RUN_DIR="${RUN_ROOT}/${task}"
+  mkdir -p "${TASK_RUN_DIR}"
+
+  CMD=(
+    bash examples/rl_games/scripts/run_eval.sh
+    --run-dir "${TASK_RUN_DIR}"
+    --config "${CONFIG_PATH}"
+    --checkpoint "${CHECKPOINT_PATH}"
+    --wandb-run-name "${WANDB_RUN_PREFIX}_${task}"
+    "${EVAL_ARGS[@]}"
+    "${TASK_OVERRIDES[@]}"
+  )
+
+  printf '[eval:%s]' "${task}"
+  printf ' %q' "${CMD[@]}"
+  echo
+  "${CMD[@]}"
+done

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_2.sh

@@ -0,0 +1,41 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="openvla_bridge_deadly_corridor_fix_latency_2_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_2_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_2_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_4.sh

@@ -0,0 +1,41 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="openvla_bridge_deadly_corridor_fix_latency_4_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_4_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_4_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_6.sh

@@ -0,0 +1,41 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="openvla_bridge_deadly_corridor_fix_latency_6_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_6_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_6_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/fixed/deadly_corridor/latency_8.sh

@@ -0,0 +1,41 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="openvla_bridge_deadly_corridor_fix_latency_8_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_8_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_fix_latency_8_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/mixed/demon_024.sh

@@ -0,0 +1,47 @@
+bash /workspace/starVLA/examples/rl_games/bash_scripts/install/pre_launch.sh
+
+cd /workspace/starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla demon_attack
+
+conda activate starvla_rl_games_openvla
+
+bash /workspace/starVLA/examples/rl_games/bash_scripts/install/latency_deps.sh 
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=demon_attack \
+    init=bridge \
+    mode=mixed_latency \
+    run_id="openvla_bridge_demon_attack_latency_mixed_024_exp1" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="latency-sensitive-bench/benchmark-models" \
+    checkpoint.sync.enabled=false \
+    checkpoint.sync.repo_id="latency-sensitive-bench/benchmark-models" \
+    dataset.source_hf="latency-sensitive-bench/benchmark-datasets" \
+    dataset.revision=15d17287246bd19940c837d6c8a5327564f969c5 \
+    dataset.source_subdir=zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep \
+    dataset.latency_filter=[0,2,4] \
+    datasets.vla_data.sequential_step_sampling=true \
+    datasets.vla_data.shuffle=true \
+    trainer.per_latency_eval_num_batches=5 \
+    dataset.episodes_per_latency=40 \
+    rl_games.env_eval.eval_backend=latency_bench \
+    datasets.vla_data.shuffle=true \
+    checkpoint.local.keep_last_n=1 \
+    trainer.max_train_steps=5000 \
+    trainer.num_warmup_steps=0 \
+    checkpoint.save_final_model=true \
+    checkpoint.save_best_model=false \
+    trainer.eval_interval=5000 \
+    trainer.save_interval=5000 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,2,4] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/openvla/bridge/single/deadly_corridor_multibinary_bce.sh

@@ -0,0 +1,43 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh openvla deadly_corridor
+
+conda activate starvla_rl_games_openvla
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=openvla \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="openvla_bridge_deadly_corridor_single_latency_clean_data_fl1_exp1" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/openvla_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/openvla_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    framework.action_model.loss_type=multibinary_bce \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.interval_steps=50 \
+    rl_games.env_eval.mid_train.max_steps_per_episode=3600 \
+    rl_games.env_eval.mid_train.num_episodes=5 \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/bash_scripts/pi_0.5/bridge/single/deadly_corridor_multibinary_bce.sh

@@ -0,0 +1,40 @@
+cd starVLA
+
+bash examples/rl_games/install/install_stack.sh pi05 deadly_corridor
+
+conda activate starvla_rl_games_pi05
+
+bash examples/rl_games/install/flash_attn.sh --check >/dev/null 2>&1 || bash examples/rl_games/install/flash_attn.sh
+
+python examples/rl_games/scripts/launch_train.py \
+    model=pi05 \
+    env=deadly_corridor \
+    init=bridge \
+    mode=single \
+    run_id="pi05_bridge_deadly_corridor_single_latency_clean_data_bce_exp2" \
+    trainer.distributed_backend=none \
+    workspace_dir="/workspace" \
+    wandb_entity="talha1503" \
+    checkpoint.hf_repo_id="talha15032/pi05_bridge_deadly_corridor_single_latency_clean_data_exp2" \
+    checkpoint.sync.enabled=true \
+    checkpoint.sync.repo_id="talha15032/pi05_bridge_deadly_corridor_single_latency_clean_data_exp2" \
+    dataset.source_hf=latency-sensitive-bench/benchmark-datasets \
+    dataset.revision=d6f0ad0995f6e4b095f315920e8de18e04297048 \
+    dataset.source_subdir=zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep \
+    checkpoint.local.keep_last_n=1 \
+    checkpoint.save_best_model=false \
+    trainer.max_train_steps=500 \
+    trainer.num_warmup_steps=0 \
+    trainer.eval_interval=50 \
+    trainer.save_interval=500 \
+    rl_games.env_eval.deadly.action_layout=multibinary_7 \
+    framework.action_model.action_dim=7 \
+    framework.action_model.action_env_dim=7 \
+    trainer.logging_frequency=1 \
+    trainer.gradient_accumulation_steps=4 \
+    datasets.vla_data.per_device_batch_size=32 \
+    rl_games.env_eval.mid_train.enabled=false \
+    rl_games.env_eval.post_train.enabled=true \
+    rl_games.env_eval.post_train.latencies=[0,1,2,3,4,5,6,7,8,9,10] \
+    rl_games.env_eval.post_train.num_episodes=20 \
+    rl_games.env_eval.post_train.max_steps_per_episode=3600

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/config/cross_task_setup/defendtheline_zero_deadly_zero.yaml

@@ -0,0 +1,77 @@
+# @package _global_
+
+rl_games:
+  env_eval:
+    mid_train:
+      interval_steps: 500
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+  cross_task:
+    loss_by_task:
+      defend_the_line: discrete_ce
+      deadly_corridor: multibinary_bce
+    loss_weight_by_task:
+      defend_the_line: 1.0
+      deadly_corridor: 1.0
+    train_tasks:
+      - name: defend_the_line
+        role: A
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps
+        prompt_source_subdir: zero-latency/defend-the-line/defend_the_line_fixed_latency_0_1000ep_7k2steps
+        converted_name: defend_the_line_zero_latency_train
+        train_latency_filter: [0]
+        eval_latency_filter: [0]
+        episodes_per_latency: 40
+        max_episodes: null
+        weight: 1.0
+      - name: deadly_corridor
+        role: B
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_subdir: zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep
+        converted_name: deadly_corridor_zero_latency_train
+        train_latency_filter: [0]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: 1000
+        max_episodes: null
+        action_layout: multibinary_7
+        weight: 1.0
+    eval_tasks:
+      defend_the_line:
+        frameskip: 4
+        image_size: 224
+        task_description: "You are playing Defend the Line in VizDoom. Your goal is to survive by turning toward enemies and shooting them before they reach you. The available actions are: TURN_LEFT, TURN_RIGHT, and ATTACK. Choose the best next action from the current visual observation."
+        mid_train:
+          latencies: [0]
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          latencies: [0, 2, 4]
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      deadly_corridor:
+        frameskip: 4
+        image_size: 224
+        task_description: "You are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK."
+        mid_train:
+          latencies: [0]
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          latencies: [0, 2, 4]
+          num_episodes: 20
+          max_steps_per_episode: 3600
+
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    loss_type: discrete_ce

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/config/cross_task_setup/flappy_demon_deadly_asterix_atlantis_defend_024.yaml

@@ -0,0 +1,164 @@
+# @package _global_
+
+rl_games:
+  env_eval:
+    mid_train:
+      enabled: false
+      interval_steps: 1000
+    post_train:
+      enabled: true
+      latencies: [0, 2, 4]
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+    asterix:
+      action_layout: factorized_6
+  cross_task:
+    loss_by_task:
+      flappy: discrete_ce
+      demon_attack: discrete_ce
+      deadly_corridor: multibinary_bce
+      asterix: asterix_factorized_ce
+      atlantis: discrete_ce
+      defend_the_line: discrete_ce
+    loss_weight_by_task:
+      flappy: 1.0
+      demon_attack: 1.0
+      deadly_corridor: 1.0
+      asterix: 1.0
+      atlantis: 1.0
+      defend_the_line: 1.0
+    train_tasks:
+      - name: flappy
+        role: A
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: 8a5af732702f31824156975ed248ce6534e736ee
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: 8a5af732702f31824156975ed248ce6534e736ee
+        train_source_subdir: zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep
+        prompt_source_subdir: zero-latency/flappy/flappy_200ep/flappy_fix_latency_0_200ep
+        converted_name: flappy_mixed_024_200ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 200
+        weight: 1.0
+      - name: demon_attack
+        role: B
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: 8a5af732702f31824156975ed248ce6534e736ee
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: 8a5af732702f31824156975ed248ce6534e736ee
+        train_source_subdir: zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep
+        prompt_source_subdir: zero-latency/demon-attack/demon_attack_200ep/demon_attack_fix_latency_0_200ep
+        converted_name: demon_attack_mixed_024_200ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 200
+        weight: 1.0
+      - name: deadly_corridor
+        role: C
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_subdir: zero-latency/deadly-corridor/deadly_1000ep/deadly_corridor_fix_latency_0_1000ep
+        converted_name: deadly_corridor_mixed_024_1000ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 1000
+        action_layout: multibinary_7
+        weight: 1.0
+      - name: asterix
+        role: D
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_subdir: zero-latency/asterix/asterix_200ep/asterix_fixed_latency_0_200ep_7k2steps
+        converted_name: asterix_mixed_024_200ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 200
+        action_layout: factorized_6
+        weight: 1.0
+      - name: atlantis
+        role: E
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_subdir: zero-latency/atlantis/atlantis_200ep/atlantis_fixed_latency_0_200ep_7k2steps
+        converted_name: atlantis_mixed_024_200ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 200
+        weight: 1.0
+      - name: defend_the_line
+        role: F
+        train_source_hf: latency-sensitive-bench/benchmark-datasets
+        train_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        train_source_subdir: zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps
+        prompt_source_hf: latency-sensitive-bench/benchmark-datasets
+        prompt_revision: d6f0ad0995f6e4b095f315920e8de18e04297048
+        prompt_source_subdir: zero-latency/defend-the-line/defend_the_line_200ep/defend_the_line_fixed_latency_0_200ep_7k2steps
+        converted_name: defend_the_line_mixed_024_200ep_cross_train
+        train_latency_filter: [0, 2, 4]
+        eval_latency_filter: [0, 2, 4]
+        episodes_per_latency: null
+        max_episodes: 200
+        weight: 1.0
+    eval_tasks:
+      flappy:
+        frameskip: 1
+        image_size: 224
+        task_description: "You are playing Flappy Bird. Pass through the pipe gaps and stay alive. Choose the action: NOOP, FLAP."
+        mid_train:
+          enabled: false
+          latencies: [0, 2, 4]
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          latencies: [0, 2, 4]
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      demon_attack:
+        frameskip: 4
+        image_size: 224
+        task_description: "You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE."
+        mid_train:
+          enabled: false
+          latencies: [0, 2, 4]
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          latencies: [0, 2, 4]
+          num_episodes: 20
+          max_steps_per_episode: 3600
+      deadly_corridor:
+        frameskip: 4
+        image_size: 224
+        task_description: "You are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK."
+        mid_train:
+          enabled: false
+          latencies: [0, 2, 4]
+          num_episodes: 5
+          max_steps_per_episode: 3600
+        post_train:
+          latencies: [0, 2, 4]
+          num_episodes: 20
+          max_steps_per_episode: 3600
+
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    loss_type: discrete_ce
+    action_layout: bridge_cross_task_7

experiments/task-transfer/provenance/release_code/third_party/starVLA/examples/rl_games/scripts/setup_training_assets.py

@@ -0,0 +1,1665 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import contextlib
+import hashlib
+import inspect
+import json
+import os
+import pickle
+import re
+import sys
+from pathlib import Path
+from typing import Any
+
+from huggingface_hub import snapshot_download
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+if str(REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(REPO_ROOT))
+
+from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_flappy_dataset import resolve_latency_subdirs
+
+
+STEP_FILE_RE = re.compile(r"steps_(\d+)_(?:pytorch_model\.pt|model\.safetensors)$")
+STEP_STATE_RE = re.compile(r"steps_(\d+)_state$")
+DEBUG_DATASET_RE = re.compile(r"^(?P<base>.+)(?P<debug>__debug(?:_[A-Za-z0-9_-]+)?_\d+ep)$")
+INITIALIZATION_SOURCE_MODES = {"bridge", "pre-trained", "pretrained", "backbone_bridge_factorized11"}
+
+
+def _str2bool(value: str | bool) -> bool:
+    if isinstance(value, bool):
+        return value
+    return str(value).lower() in {"1", "true", "yes", "y", "on"}
+
+
+def _safe_path_name(value: str) -> str:
+    return re.sub(r"[^A-Za-z0-9_.-]+", "_", str(value)).strip("_") or "checkpoint"
+
+
+def _safe_token(value: Any) -> str:
+    return re.sub(r"[^A-Za-z0-9_-]+", "_", str(value)).strip("_") or "value"
+
+
+def _fixed_dataset_source(dataset_name, revision, source_subdir, latencies, cache_dir):
+    """Materialize one fixed HF dataset revision for existing local converters."""
+    if revision is None:
+        return dataset_name
+
+    subdirs = resolve_latency_subdirs(source_subdir, latencies)
+    patterns = [f"{subdir}/**" for subdir in subdirs] if source_subdir is not None else None
+    return snapshot_download(
+        dataset_name, repo_type="dataset", revision=revision,
+        cache_dir=cache_dir, allow_patterns=patterns, endpoint="https://huggingface.co",
+    )
+
+
+def _record_fixed_dataset_source(dataset_path, source, revision):
+    """Preserve the HF origin when conversion reads its local fixed snapshot."""
+    if revision is None:
+        return
+    path = dataset_path / "manifest.json"
+    manifest = json.loads(path.read_text())
+    manifest["source"] = source
+    manifest["source_revision"] = revision
+    path.write_text(json.dumps(manifest, indent=2))
+
+
+def _latency_suffix(latencies: list[int] | None) -> str:
+    if not latencies:
+        return "all"
+    return "lat" + "_".join(str(int(value)) for value in latencies)
+
+
+def _derived_dataset_name(base_name: str, *, latency_filter: list[int] | None, episodes_per_latency: int | None, max_episodes: int | None) -> str:
+    pieces = [str(base_name), _latency_suffix(latency_filter)]
+    if episodes_per_latency is not None:
+        pieces.append(f"{int(episodes_per_latency)}ep_per_lat")
+    elif max_episodes is not None:
+        pieces.append(f"{int(max_episodes)}ep")
+    return "__".join(_safe_token(piece) for piece in pieces)
+
+
+def _has_files(path: Path) -> bool:
+    if not path.exists():
+        return False
+    if path.is_file():
+        return True
+    return any(path.iterdir())
+
+
+def _dataset_ready(dataset_dir: Path) -> bool:
+    required = [
+        dataset_dir / "meta/modality.json",
+        dataset_dir / "meta/info.json",
+        dataset_dir / "meta/episodes.jsonl",
+        dataset_dir / "meta/tasks.jsonl",
+    ]
+    return all(path.exists() for path in required) and any(dataset_dir.glob("data/*/*.parquet"))
+
+
+def _load_source_latency_prompt_map(
+    dataset_name: str,
+    *,
+    cache_dir: str | None = None,
+    dataset_config_name: str | None = None,
+    dataset_source_subdir: str | None = None,
+    source_latency_column: str,
+    target_latency_unit: str,
+    obs_stride_raw_frames: int,
+    latencies: list[int] | None = None,
+) -> dict[str, dict[str, Any]]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_flappy_dataset import (
+        _load_train_split,
+        build_latency_prompt_map,
+    )
+
+    column_variants: tuple[list[str], ...] = (
+        ["prompt", source_latency_column, "latency_ms", "split"],
+        ["prompt", source_latency_column, "latency_ms"],
+    )
+    last_exc: Exception | None = None
+    ds = None
+    for columns in column_variants:
+        try:
+            ds = _load_train_split(
+                dataset_name,
+                cache_dir,
+                columns=columns,
+                dataset_config_name=dataset_config_name,
+                dataset_source_subdir=dataset_source_subdir,
+                latencies=latencies,
+            )
+            break
+        except Exception as exc:
+            last_exc = exc
+            ds = None
+    if ds is None:
+        raise ValueError(
+            "could not load source dataset latency prompts "
+            f"(dataset={dataset_name!r}, config={dataset_config_name!r}, "
+            f"source_subdir={dataset_source_subdir!r}, latencies={latencies!r})"
+        ) from last_exc
+    return build_latency_prompt_map(
+        ds,
+        latency_column=source_latency_column,
+        target_latency_unit=target_latency_unit,
+        obs_stride_raw_frames=obs_stride_raw_frames,
+    )
+
+
+def _write_prompt_map(path: Path, prompt_map: dict[str, dict[str, Any]], *, latency_filter: list[int] | None = None) -> Path:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if latency_filter:
+        allowed = {int(value) for value in latency_filter}
+        prompt_map = {str(k): v for k, v in prompt_map.items() if int(k) in allowed}
+    path.write_text(json.dumps(prompt_map, indent=2), encoding="utf-8")
+    return path
+
+
+def _prompt_map_has_raw_frame_entries(path: Path) -> bool:
+    if not path.exists():
+        return False
+    try:
+        mapping = json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return False
+    if not isinstance(mapping, dict) or not mapping:
+        return False
+    return all(
+        isinstance(entry, dict) and "latency_raw_frames" in entry
+        for entry in mapping.values()
+    )
+
+
+def _read_prompt_map(path: Path) -> dict[str, dict[str, Any]]:
+    if not path.exists():
+        return {}
+    try:
+        mapping = json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return {}
+    if not isinstance(mapping, dict):
+        return {}
+    return {str(key): value for key, value in mapping.items() if isinstance(value, dict)}
+
+
+def _prompt_map_entry_count(path: Path) -> int:
+    return len(_read_prompt_map(path))
+
+
+def _extract_latency_ms_from_prompt(prompt: str) -> float | None:
+    match = re.search(r"\(([0-9]+(?:\.[0-9]+)?)\s*ms\)", prompt)
+    if match is None:
+        return None
+    return float(match.group(1))
+
+
+def _read_local_task_prompts(dataset_dir: Path) -> dict[int, str]:
+    tasks_path = dataset_dir / "meta/tasks.jsonl"
+    if not tasks_path.exists():
+        return {}
+    prompts: dict[int, str] = {}
+    for line in tasks_path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        record = json.loads(line)
+        task_index = int(record.get("task_index", len(prompts)))
+        prompt = record.get("task", record.get("prompt"))
+        if prompt is not None:
+            prompts[task_index] = str(prompt)
+    return prompts
+
+
+def _latency_raw_frames_from_target_latency(
+    latency: int,
+    *,
+    target_latency_unit: str,
+    obs_stride_raw_frames: int,
+) -> int:
+    if target_latency_unit == "raw_frames":
+        return int(latency)
+    if target_latency_unit == "observation_steps":
+        return int(latency) * int(obs_stride_raw_frames)
+    raise ValueError(f"unsupported target_latency_unit={target_latency_unit!r}")
+
+
+def _load_local_latency_prompt_map(
+    dataset_dir: Path,
+    *,
+    target_latency_unit: str,
+    obs_stride_raw_frames: int,
+) -> dict[str, dict[str, Any]]:
+    task_prompts = _read_local_task_prompts(dataset_dir)
+    if not task_prompts:
+        return {}
+
+    try:
+        import pyarrow.parquet as pq
+    except Exception:
+        return {}
+
+    prompt_map: dict[int, dict[str, Any]] = {}
+    for parquet_path in sorted(dataset_dir.glob("data/*/*.parquet")):
+        try:
+            parquet = pq.ParquetFile(parquet_path)
+            schema_names = set(parquet.schema_arrow.names)
+        except Exception:
+            continue
+        if "latency" not in schema_names or "task_index" not in schema_names:
+            continue
+        for batch in parquet.iter_batches(columns=["latency", "task_index"], batch_size=65536):
+            for row in batch.to_pylist():
+                latency = int(row["latency"])
+                prompt = task_prompts.get(int(row["task_index"]))
+                if prompt is None:
+                    continue
+                latency_raw_frames = _latency_raw_frames_from_target_latency(
+                    latency,
+                    target_latency_unit=target_latency_unit,
+                    obs_stride_raw_frames=obs_stride_raw_frames,
+                )
+                entry = {
+                    "latency": latency,
+                    "latency_raw_frames": latency_raw_frames,
+                    "latency_ms": _extract_latency_ms_from_prompt(prompt),
+                    "prompt": prompt,
+                }
+                existing = prompt_map.get(latency)
+                if existing is None:
+                    prompt_map[latency] = entry
+                elif existing["prompt"] != prompt or int(existing["latency_raw_frames"]) != latency_raw_frames:
+                    raise ValueError(
+                        f"local dataset has inconsistent prompt map rows for latency={latency}: "
+                        f"{existing!r} vs {entry!r}"
+                    )
+    return {str(k): prompt_map[k] for k in sorted(prompt_map)}
+
+
+def _normalize_prompt_map_entries(
+    prompt_map: dict[str, dict[str, Any]],
+    *,
+    target_latency_unit: str,
+    obs_stride_raw_frames: int,
+) -> dict[str, dict[str, Any]]:
+    normalized: dict[str, dict[str, Any]] = {}
+    for key, value in prompt_map.items():
+        if not isinstance(value, dict) or "prompt" not in value:
+            continue
+        latency = int(value.get("latency", key))
+        entry = dict(value)
+        entry["latency"] = latency
+        if entry.get("latency_raw_frames") is None:
+            entry["latency_raw_frames"] = _latency_raw_frames_from_target_latency(
+                latency,
+                target_latency_unit=target_latency_unit,
+                obs_stride_raw_frames=obs_stride_raw_frames,
+            )
+        else:
+            entry["latency_raw_frames"] = int(entry["latency_raw_frames"])
+        if "latency_ms" not in entry:
+            entry["latency_ms"] = _extract_latency_ms_from_prompt(str(entry["prompt"]))
+        normalized[str(latency)] = entry
+    return normalized
+
+
+def _write_sorted_prompt_map(path: Path, prompt_map: dict[str, dict[str, Any]]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(
+        json.dumps({str(k): prompt_map[str(k)] for k in sorted(int(k) for k in prompt_map)}, indent=2),
+        encoding="utf-8",
+    )
+
+
+def _find_latest_local_checkpoint(checkpoint_dir: Path) -> tuple[Path | None, int, str | None]:
+    if not checkpoint_dir.exists():
+        return None, 0, None
+    candidates: list[tuple[int, int, Path, str]] = []
+    for item in checkpoint_dir.iterdir():
+        if item.is_dir():
+            match = STEP_STATE_RE.match(item.name)
+            if match:
+                candidates.append((int(match.group(1)), 1, item, "state"))
+        elif item.is_file():
+            match = STEP_FILE_RE.match(item.name)
+            if match:
+                candidates.append((int(match.group(1)), 0, item, "model"))
+    if not candidates:
+        return None, 0, None
+    candidates.sort(key=lambda item: (item[0], item[1]))
+    step, _, path, kind = candidates[-1]
+    return path, step, kind
+
+
+def _read_best_checkpoint_step(checkpoint_dir: Path) -> int:
+    metadata_path = checkpoint_dir / "best_model_metadata.json"
+    if not metadata_path.exists():
+        return 0
+    try:
+        metadata = json.loads(metadata_path.read_text(encoding="utf-8"))
+    except Exception:
+        return 0
+    for key in ("best_step", "step"):
+        if metadata.get(key) is not None:
+            try:
+                return int(metadata[key])
+            except Exception:
+                return 0
+    return 0
+
+
+def _find_local_best_checkpoint(checkpoint_dir: Path) -> tuple[Path | None, int, str | None]:
+    best_state = checkpoint_dir / "best_state"
+    if not best_state.is_dir():
+        return None, 0, None
+    return best_state, _read_best_checkpoint_step(checkpoint_dir), "state"
+
+
+def _candidate_resume_paths(checkpoint: str, checkpoint_dir: Path) -> list[Path]:
+    checkpoint_path = Path(checkpoint).expanduser()
+    if checkpoint_path.is_absolute():
+        return [checkpoint_path]
+
+    run_dir = checkpoint_dir.parent
+    return [
+        run_dir / checkpoint_path,
+        checkpoint_dir / checkpoint_path.name,
+        checkpoint_dir / checkpoint_path,
+    ]
+
+
+def _resolve_existing_resume_path(checkpoint: str, checkpoint_dir: Path) -> Path:
+    candidates = _candidate_resume_paths(checkpoint, checkpoint_dir)
+    for candidate in candidates:
+        if candidate.exists():
+            return candidate.resolve()
+    checked = ", ".join(str(candidate.resolve()) for candidate in candidates)
+    raise FileNotFoundError(f"explicit resume checkpoint does not exist: {checkpoint}. Checked: {checked}")
+
+
+def _resolve_explicit_resume_checkpoint(checkpoint: str, checkpoint_dir: Path) -> tuple[Path, int, str]:
+    checkpoint_path = _resolve_existing_resume_path(checkpoint, checkpoint_dir)
+    if not checkpoint_path.exists():
+        raise FileNotFoundError(f"explicit resume checkpoint does not exist: {checkpoint_path}")
+    if checkpoint_path.is_dir():
+        state_match = STEP_STATE_RE.match(checkpoint_path.name)
+        if state_match:
+            return checkpoint_path, int(state_match.group(1)), "state"
+        if checkpoint_path.name == "best_state":
+            return checkpoint_path, _read_best_checkpoint_step(checkpoint_path.parent), "state"
+        raise ValueError(
+            f"unsupported explicit resume checkpoint directory: {checkpoint_path}. "
+            "Expected steps_<N>_state/ or best_state/."
+        )
+    if checkpoint_path.is_file():
+        file_match = STEP_FILE_RE.match(checkpoint_path.name)
+        if file_match:
+            return checkpoint_path, int(file_match.group(1)), "model"
+        if checkpoint_path.parent.name.endswith("_state"):
+            state_match = STEP_STATE_RE.match(checkpoint_path.parent.name)
+            if state_match and checkpoint_path.name in {"model.safetensors", "pytorch_model.bin"}:
+                return checkpoint_path.parent, int(state_match.group(1)), "state"
+        if checkpoint_path.parent.name == "best_state" and checkpoint_path.name in {"model.safetensors", "pytorch_model.bin"}:
+            return checkpoint_path.parent, _read_best_checkpoint_step(checkpoint_path.parent.parent), "state"
+    raise ValueError(
+        f"unsupported explicit resume checkpoint: {checkpoint_path}. "
+        "Expected steps_<N>_state/, steps_<N>_pytorch_model.pt, steps_<N>_model.safetensors, "
+        "best_state/, or a model file inside a state directory."
+    )
+
+
+def _download_latest_hf_checkpoint(repo_id: str, checkpoint_dir: Path) -> tuple[Path | None, int, str | None, str | None]:
+    try:
+        from huggingface_hub import HfApi, hf_hub_download, snapshot_download
+    except Exception as exc:
+        return None, 0, None, f"huggingface_hub import failed: {exc}"
+
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN")
+    try:
+        files = HfApi(token=hf_token).list_repo_files(repo_id=repo_id, repo_type="model")
+    except Exception as exc:
+        return None, 0, None, f"could not list HF checkpoint repo {repo_id}: {exc}"
+
+    candidates: list[tuple[int, int, str, str]] = []
+    seen_state_dirs = set()
+    for file_path in files:
+        first_part = file_path.split("/", 1)[0]
+        state_match = STEP_STATE_RE.match(first_part)
+        if state_match:
+            if first_part not in seen_state_dirs:
+                candidates.append((int(state_match.group(1)), 1, first_part, "state"))
+                seen_state_dirs.add(first_part)
+            continue
+        file_match = STEP_FILE_RE.match(os.path.basename(file_path))
+        if file_match:
+            candidates.append((int(file_match.group(1)), 0, file_path, "model"))
+    if not candidates:
+        return None, 0, None, None
+
+    candidates.sort(key=lambda item: (item[0], item[1]))
+    step, _, chosen, kind = candidates[-1]
+    checkpoint_dir.mkdir(parents=True, exist_ok=True)
+    try:
+        if kind == "state":
+            snapshot_download(
+                repo_id=repo_id,
+                repo_type="model",
+                allow_patterns=[f"{chosen}/**"],
+                local_dir=str(checkpoint_dir),
+                local_dir_use_symlinks=False,
+                token=hf_token,
+            )
+            local_path = checkpoint_dir / chosen
+        else:
+            local_path = hf_hub_download(
+                repo_id=repo_id,
+                repo_type="model",
+                filename=chosen,
+                local_dir=str(checkpoint_dir),
+                local_dir_use_symlinks=False,
+                token=hf_token,
+            )
+    except Exception as exc:
+        return None, 0, None, f"could not download HF checkpoint {chosen}: {exc}"
+    return Path(local_path), step, kind, None
+
+
+def _download_hf_best_checkpoint(repo_id: str, checkpoint_dir: Path) -> tuple[Path | None, int, str | None, str | None]:
+    try:
+        from huggingface_hub import HfApi, snapshot_download
+    except Exception as exc:
+        return None, 0, None, f"huggingface_hub import failed: {exc}"
+
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN")
+    try:
+        files = HfApi(token=hf_token).list_repo_files(repo_id=repo_id, repo_type="model")
+    except Exception as exc:
+        return None, 0, None, f"could not list HF checkpoint repo {repo_id}: {exc}"
+
+    if not any(file_path.startswith("best_state/") for file_path in files):
+        return None, 0, None, None
+
+    checkpoint_dir.mkdir(parents=True, exist_ok=True)
+    try:
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="model",
+            allow_patterns=["best_state/**", "best_model_metadata.json"],
+            local_dir=str(checkpoint_dir),
+            local_dir_use_symlinks=False,
+            token=hf_token,
+        )
+    except Exception as exc:
+        return None, 0, None, f"could not download HF best checkpoint from {repo_id}: {exc}"
+
+    best_state = checkpoint_dir / "best_state"
+    if not best_state.is_dir():
+        return None, 0, None, f"HF best checkpoint from {repo_id} did not contain best_state/"
+    return best_state, _read_best_checkpoint_step(checkpoint_dir), "state", None
+
+
+def _download_hf_checkpoint_file(repo_id: str, filename: str, checkpoint_dir: Path) -> tuple[Path | None, int, str | None]:
+    try:
+        from huggingface_hub import hf_hub_download
+    except Exception as exc:
+        return None, 0, f"huggingface_hub import failed: {exc}"
+
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN")
+    checkpoint_dir.mkdir(parents=True, exist_ok=True)
+    try:
+        local_path = hf_hub_download(
+            repo_id=repo_id,
+            repo_type="model",
+            filename=filename,
+            local_dir=str(checkpoint_dir),
+            local_dir_use_symlinks=False,
+            token=hf_token,
+        )
+    except Exception as exc:
+        return None, 0, f"could not download HF checkpoint {filename} from {repo_id}: {exc}"
+
+    file_match = STEP_FILE_RE.match(os.path.basename(filename))
+    step = int(file_match.group(1)) if file_match else 0
+    return Path(local_path), step, None
+
+
+def _resolve_local_initialization_checkpoint(local_dir: str, filename: str) -> tuple[Path | None, int]:
+    if not local_dir or not filename:
+        return None, 0
+    checkpoint_path = (Path(local_dir).expanduser() / filename).resolve()
+    if not checkpoint_path.exists():
+        return None, 0
+    file_match = STEP_FILE_RE.match(checkpoint_path.name)
+    step = int(file_match.group(1)) if file_match else 0
+    return checkpoint_path, step
+
+
+def _is_missing_hf_repo_error(error: str | None) -> bool:
+    if not error:
+        return False
+    return "Repository Not Found" in error or "404 Client Error" in error
+
+
+def _ensure_base_model(model: str, base_model_dir: Path, base_model_repo_id: str | None) -> dict[str, Any]:
+    info = {
+        "base_model_dir": str(base_model_dir),
+        "base_model_repo_id": base_model_repo_id,
+        "base_model_downloaded": False,
+    }
+    if _has_files(base_model_dir):
+        return info
+    if not base_model_repo_id:
+        raise ValueError(f"base model directory is missing and no repo id was provided: {base_model_dir}")
+
+    try:
+        from huggingface_hub import snapshot_download
+    except Exception as exc:
+        raise RuntimeError("huggingface_hub is required to download base model weights") from exc
+
+    base_model_dir.mkdir(parents=True, exist_ok=True)
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN")
+    snapshot_download(
+        repo_id=base_model_repo_id,
+        repo_type="model",
+        local_dir=str(base_model_dir),
+        token=hf_token,
+    )
+    info["base_model_downloaded"] = True
+    return info
+
+
+def _starvla_runtime_cache_paths(dataset_dir: Path) -> dict[str, Path]:
+    return {
+        "dataset_statistics": dataset_dir / "dataset_statistics.json",
+        "stats": dataset_dir / "meta" / "stats_gr00t.json",
+        "steps": dataset_dir / "meta" / "steps_data_index.pkl",
+    }
+
+
+def _starvla_runtime_cache_ready(dataset_dir: Path) -> bool:
+    paths = _starvla_runtime_cache_paths(dataset_dir)
+    return all(path.exists() for path in paths.values())
+
+
+def _read_starvla_runtime_cache_summary(data_root_dir: Path, data_mix: str) -> dict[str, Any]:
+    from starVLA.dataloader.gr00t_lerobot.registry import (
+        ROBOT_TYPE_TO_EMBODIMENT_TAG,
+        get_dataset_named_mixture,
+    )
+
+    total_steps = 0
+    total_trajectories = 0
+    first_stats: dict[str, Any] | None = None
+    for dataset_name, _, robot_type in get_dataset_named_mixture(data_mix):
+        paths = _starvla_runtime_cache_paths(data_root_dir / dataset_name)
+        with paths["steps"].open("rb") as stream:
+            steps_cache = pickle.load(stream)
+        total_steps += steps_cache["total_steps"]
+        total_trajectories += steps_cache["num_trajectories"]
+        if first_stats is None:
+            first_stats = {
+                "dataset_stats_path": str(paths["dataset_statistics"]),
+                "dataset_robot_type": robot_type,
+                "dataset_embodiment_tag": str(ROBOT_TYPE_TO_EMBODIMENT_TAG[robot_type]),
+            }
+    assert first_stats is not None
+    return {
+        **first_stats,
+        "dataset_num_steps": total_steps,
+        "dataset_num_trajectories": total_trajectories,
+    }
+
+
+def _materialize_starvla_runtime_cache(data_root_dir: Path, data_mix: str) -> dict[str, Any]:
+    from starVLA.dataloader.lerobot_datasets import make_LeRobotSingleDataset
+    from starVLA.dataloader.gr00t_lerobot.registry import (
+        ROBOT_TYPE_TO_EMBODIMENT_TAG,
+        get_dataset_named_mixture,
+    )
+
+    mixture = get_dataset_named_mixture(data_mix)
+    total_steps = 0
+    total_trajectories = 0
+    first_stats: dict[str, Any] | None = None
+    for dataset_name, _, robot_type in mixture:
+        dataset = make_LeRobotSingleDataset(
+            data_root_dir=data_root_dir,
+            data_name=dataset_name,
+            robot_type=robot_type,
+            data_cfg={
+                "include_state": False,
+                "video_backend": "torchvision_av",
+                "lerobot_version": "v2.0",
+            },
+        )
+        stats_path = data_root_dir / dataset_name / "dataset_statistics.json"
+        dataset._save_dataset_statistics_(stats_path)
+        total_steps += len(dataset)
+        total_trajectories += len(dataset.trajectory_ids)
+        if first_stats is None:
+            first_stats = {
+                "dataset_stats_path": str(stats_path),
+                "dataset_robot_type": robot_type,
+                "dataset_embodiment_tag": str(ROBOT_TYPE_TO_EMBODIMENT_TAG[robot_type]),
+            }
+    assert first_stats is not None
+    return {
+        **first_stats,
+        "dataset_num_steps": total_steps,
+        "dataset_num_trajectories": total_trajectories,
+    }
+
+
+def _validate_starvla_dataset(data_root_dir: Path, data_mix: str) -> dict[str, Any]:
+    from starVLA.dataloader.gr00t_lerobot.registry import get_dataset_named_mixture
+
+    mixture = get_dataset_named_mixture(data_mix)
+    if all(_starvla_runtime_cache_ready(data_root_dir / dataset_name) for dataset_name, _, _ in mixture):
+        return _read_starvla_runtime_cache_summary(data_root_dir=data_root_dir, data_mix=data_mix)
+    return _materialize_starvla_runtime_cache(data_root_dir=data_root_dir, data_mix=data_mix)
+
+
+def _initialization_mode(args) -> str:
+    return str(getattr(args, "initialization_mode", None) or "scratch").lower()
+
+
+def _action_carrier(args) -> str:
+    configured = str(getattr(args, "action_carrier", "") or "").lower()
+    if configured in {"native", "bridge"}:
+        return configured
+    if _initialization_mode(args) in {"pre-trained", "pretrained", "bridge"}:
+        return "bridge"
+    return "native"
+
+
+def _carrier_dataset_name(data_mix: str, action_carrier: str) -> str:
+    if action_carrier != "bridge" or "__bridge" in data_mix:
+        return data_mix
+    match = DEBUG_DATASET_RE.match(data_mix)
+    if match:
+        return f"{match.group('base')}__bridge{match.group('debug')}"
+    return f"{data_mix}__bridge"
+
+
+def _ensure_rl_games_lerobot_dataset(
+    args,
+    *,
+    convert_dataset,
+    verify_dataset,
+    env_name: str,
+    env_fps: float,
+    obs_fps: float,
+) -> dict[str, Any]:
+    data_root_dir = Path(args.dataset_local_dir).expanduser().resolve()
+    action_carrier = _action_carrier(args)
+    action_layout = str(
+        getattr(args, "action_layout", "")
+        or getattr(args, "deadly_action_layout", "")
+        or ""
+    )
+    source_dataset = str(getattr(args, "source_dataset_hf", "") or "")
+    source_config_name = getattr(args, "source_dataset_config_name", None)
+    source_config_name = None if source_config_name in (None, "") else str(source_config_name)
+    source_subdir = args.source_dataset_subdir
+    source_revision = getattr(args, "source_dataset_revision", None)
+    required_latencies = getattr(args, "latency_filter", None)
+    source_for_reading = _fixed_dataset_source(
+        source_dataset, source_revision, source_subdir, required_latencies, args.dataset_cache_dir,
+    )
+    data_mix = _carrier_dataset_name(args.converted_dataset_name, action_carrier)
+    dataset_dir = data_root_dir / data_mix
+    eval_data_mix = f"{data_mix}__val"
+    eval_dataset_dir = data_root_dir / eval_data_mix
+    force = _str2bool(args.setup_force) or _str2bool(args.dataset_force_download)
+    mixed_latency = args.mode == "mixed_latency" or str(args.latency_mode or "").lower() == "mixed"
+    source_latency_column = "latency_raw_frames"
+    target_latency_unit = args.target_latency_unit
+    obs_stride_raw_frames = round(env_fps / obs_fps)
+    prompt_map = dataset_dir / "latency_prompt_map.json"
+
+    def _manifest_mismatch_reasons(dataset_path: Path) -> list[str]:
+
+        reasons: list[str] = []
+        manifest_path = dataset_path / "manifest.json"
+        if not manifest_path.exists():
+            return reasons
+        try:
+            manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+        except Exception as exc:
+            return [f"{manifest_path} is unreadable: {exc}"]
+        if str(manifest.get("action_carrier", "native")) != action_carrier:
+            reasons.append(
+                f"action_carrier mismatch: manifest={manifest.get('action_carrier', 'native')!r} expected={action_carrier!r}"
+            )
+        if source_revision is not None and ("source_revision" not in manifest or manifest["source_revision"] != source_revision):
+            reasons.append("source_revision mismatch")
+        if (manifest.get("source_config") or None) != source_config_name:
+            reasons.append(
+                f"source_config mismatch: manifest={(manifest.get('source_config') or None)!r} expected={source_config_name!r}"
+            )
+        if (manifest.get("source_subdir") or None) != source_subdir:
+            reasons.append(
+                f"source_subdir mismatch: manifest={(manifest.get('source_subdir') or None)!r} expected={source_subdir!r}"
+            )
+        if action_layout and ("action_layout" not in manifest or str(manifest["action_layout"]) != action_layout):
+            reasons.append(
+                f"action_layout mismatch: manifest={manifest.get('action_layout')!r} expected={action_layout!r}"
+            )
+        if mixed_latency and manifest.get("latency_metadata") is not True:
+            reasons.append("latency_metadata missing/false for mixed-latency run")
+        if manifest.get("fps") != obs_fps:
+            reasons.append(f"fps mismatch: manifest={manifest.get('fps')!r} expected={obs_fps!r}")
+        if manifest.get("obs_stride_raw_frames") != obs_stride_raw_frames:
+            reasons.append(
+                f"obs_stride_raw_frames mismatch: manifest={manifest.get('obs_stride_raw_frames')!r} expected={obs_stride_raw_frames!r}"
+            )
+        if manifest.get("target_latency_unit") != target_latency_unit:
+            reasons.append(
+                f"target_latency_unit mismatch: manifest={manifest.get('target_latency_unit')!r} expected={target_latency_unit!r}"
+            )
+        if manifest.get("source_latency_column") != source_latency_column:
+            reasons.append(
+                f"source_latency_column mismatch: manifest={manifest.get('source_latency_column')!r} expected={source_latency_column!r}"
+            )
+        expected_latency_filter = getattr(args, "latency_filter", None)
+        if expected_latency_filter is not None:
+            manifest_latency_filter = manifest.get("latency_filter")
+            if manifest_latency_filter != [int(value) for value in expected_latency_filter]:
+                reasons.append(
+                    f"latency_filter mismatch: manifest={manifest_latency_filter!r} "
+                    f"expected={[int(value) for value in expected_latency_filter]!r}"
+                )
+            expected_subdirs = [str(s) for s in resolve_latency_subdirs(source_subdir, expected_latency_filter)]
+            if manifest.get("latency_subdirs") != expected_subdirs:
+                reasons.append(
+                    f"latency_subdirs mismatch: manifest={manifest.get('latency_subdirs')!r} expected={expected_subdirs!r}"
+                )
+        expected_episodes_per_latency = getattr(args, "episodes_per_latency", None)
+        if expected_episodes_per_latency is not None:
+            if manifest.get("episodes_per_latency") != int(expected_episodes_per_latency):
+                reasons.append(
+                    f"episodes_per_latency mismatch: manifest={manifest.get('episodes_per_latency')!r} "
+                    f"expected={int(expected_episodes_per_latency)!r}"
+                )
+        expected_max_episodes = getattr(args, "max_episodes", None)
+        if expected_max_episodes is not None:
+            if manifest.get("max_episodes") != int(expected_max_episodes):
+                reasons.append(
+                    f"max_episodes mismatch: manifest={manifest.get('max_episodes')!r} expected={int(expected_max_episodes)!r}"
+                )
+        return reasons
+
+    def _manifest_matches(dataset_path: Path) -> bool:
+        return not _manifest_mismatch_reasons(dataset_path)
+
+    def _mixed_prompt_map_ready() -> bool:
+        if not mixed_latency:
+            return True
+        if not prompt_map.exists():
+            return _dataset_ready(dataset_dir)
+        try:
+            mapping = json.loads(prompt_map.read_text(encoding="utf-8"))
+        except Exception:
+            return _dataset_ready(dataset_dir)
+        if len(mapping) > 1 and _prompt_map_has_raw_frame_entries(prompt_map):
+            return True
+        return _dataset_ready(dataset_dir)
+
+    def _source_prompt_map(latencies: list[int] | None = None) -> dict[str, dict[str, Any]]:
+        if not source_dataset:
+            return {}
+        prompt_source_for_reading = _fixed_dataset_source(
+            source_dataset, source_revision, source_subdir, latencies, args.dataset_cache_dir,
+        )
+        return _load_source_latency_prompt_map(
+            prompt_source_for_reading,
+            cache_dir=getattr(args, "dataset_cache_dir", None),
+            dataset_config_name=source_config_name,
+            dataset_source_subdir=source_subdir,
+            latencies=latencies,
+            source_latency_column=source_latency_column,
+            target_latency_unit=target_latency_unit,
+            obs_stride_raw_frames=obs_stride_raw_frames,
+        )
+
+    def _remake_prompt_map_if_needed(*, required_latencies: list[int] | None = None) -> None:
+        existing_map = _normalize_prompt_map_entries(
+            _read_prompt_map(prompt_map),
+            target_latency_unit=target_latency_unit,
+            obs_stride_raw_frames=obs_stride_raw_frames,
+        )
+        required_keys = {str(int(value)) for value in (required_latencies or [])}
+        prompt_map_required = mixed_latency or bool(required_keys) or prompt_map.exists()
+        if not prompt_map_required:
+            return
+        needs_remake = (
+            not _prompt_map_has_raw_frame_entries(prompt_map)
+            or (mixed_latency and _prompt_map_entry_count(prompt_map) <= 1)
+            or not required_keys.issubset(existing_map)
+        )
+        if not needs_remake:
+            return
+
+        local_map = _load_local_latency_prompt_map(
+            dataset_dir,
+            target_latency_unit=target_latency_unit,
+            obs_stride_raw_frames=obs_stride_raw_frames,
+        )
+        merged = {
+            **existing_map,
+            **_normalize_prompt_map_entries(
+                local_map,
+                target_latency_unit=target_latency_unit,
+                obs_stride_raw_frames=obs_stride_raw_frames,
+            ),
+        }
+        missing_required = [int(value) for value in sorted(required_keys, key=int) if value not in merged]
+        if (
+            (missing_required or (mixed_latency and len(merged) <= 1))
+            and source_dataset
+        ):
+            merged = {
+                **merged,
+                **_normalize_prompt_map_entries(
+                    _source_prompt_map(missing_required or None),
+                    target_latency_unit=target_latency_unit,
+                    obs_stride_raw_frames=obs_stride_raw_frames,
+                ),
+            }
+
+        if not merged:
+            if required_keys or mixed_latency:
+                raise ValueError(
+                    f"could not remake latency prompt map from local dataset: {prompt_map}. "
+                    "Expected converted parquet files with latency/task_index columns and meta/tasks.jsonl."
+                )
+            return
+        if mixed_latency and len(merged) <= 1:
+            raise ValueError(
+                f"remade latency prompt map has only {len(merged)} entry: {prompt_map}. "
+                "Mixed-latency training/eval needs more than one latency prompt."
+            )
+        missing_required = [int(value) for value in sorted(required_keys, key=int) if value not in merged]
+        if missing_required:
+            raise ValueError(
+                f"could not add required eval latencies {missing_required} to {prompt_map}; "
+                f"available latencies are {sorted(int(key) for key in merged)}"
+            )
+        _write_sorted_prompt_map(prompt_map, merged)
+
+    rebuild_reasons: list[str] = []
+    if force:
+        rebuild_reasons.append("dataset.setup_force or dataset.force_download is true")
+    if not _dataset_ready(dataset_dir):
+        rebuild_reasons.append(f"train dataset is not ready: {dataset_dir}")
+    if not _dataset_ready(eval_dataset_dir):
+        rebuild_reasons.append(f"eval dataset is not ready: {eval_dataset_dir}")
+    for reason in _manifest_mismatch_reasons(dataset_dir):
+        rebuild_reasons.append(f"train manifest: {reason}")
+    for reason in _manifest_mismatch_reasons(eval_dataset_dir):
+        rebuild_reasons.append(f"eval manifest: {reason}")
+    if not _mixed_prompt_map_ready():
+        rebuild_reasons.append(f"mixed latency prompt map is not ready: {prompt_map}")
+    rebuild = bool(rebuild_reasons)
+    converted = False
+    if rebuild:
+        print("[setup] rebuilding converted dataset because:", file=sys.stderr, flush=True)
+        for reason in rebuild_reasons:
+            print(f"[setup]   - {reason}", file=sys.stderr, flush=True)
+        if not args.source_dataset_hf:
+            raise ValueError(
+                f"{dataset_dir} is not ready; pass --source-dataset-hf so setup can verify and convert it"
+            )
+        if not _str2bool(getattr(args, "skip_verification", "false")):
+            verify_kwargs = {
+                "rows": args.verify_rows,
+                "cache_dir": args.dataset_cache_dir,
+                "strict": True,
+                "allow_mixed_latency_prompts": mixed_latency,
+                "source_latency_column": source_latency_column,
+                "target_latency_unit": target_latency_unit,
+                "obs_stride_raw_frames": obs_stride_raw_frames,
+            }
+            if "dataset_config_name" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["dataset_config_name"] = source_config_name
+            if "dataset_source_subdir" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["dataset_source_subdir"] = source_subdir
+            if action_layout and "action_layout" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["action_layout"] = action_layout
+            if "latencies" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["latencies"] = getattr(args, "latency_filter", None)
+            verify_dataset(source_for_reading, **verify_kwargs)
+        convert_kwargs = {
+            "cache_dir": args.dataset_cache_dir,
+            "max_episodes": args.max_episodes,
+            "force": rebuild,
+            "require_latency_prompt_map": mixed_latency,
+            "fps": obs_fps,
+            "obs_stride_raw_frames": obs_stride_raw_frames,
+            "source_latency_column": source_latency_column,
+            "target_latency_unit": target_latency_unit,
+            "source_rows_unit": "decision_step",
+        }
+        if "dataset_config_name" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["dataset_config_name"] = source_config_name
+        if "dataset_source_subdir" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["dataset_source_subdir"] = source_subdir
+        if "latency_filter" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["latency_filter"] = getattr(args, "latency_filter", None)
+        if "episodes_per_latency" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["episodes_per_latency"] = getattr(args, "episodes_per_latency", None)
+        if "action_carrier" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["action_carrier"] = action_carrier
+        if action_layout and "action_layout" in inspect.signature(convert_dataset).parameters:
+            convert_kwargs["action_layout"] = action_layout
+        convert_dataset(source_for_reading, dataset_dir, **convert_kwargs)
+        _record_fixed_dataset_source(dataset_dir, source_dataset, source_revision)
+        _record_fixed_dataset_source(eval_dataset_dir, source_dataset, source_revision)
+        converted = True
+        if mixed_latency and not _mixed_prompt_map_ready():
+            raise ValueError(
+                f"mixed-latency dataset conversion did not create a usable prompt map: {prompt_map}. "
+                "Check that the selected training episodes contain latency/prompt columns for more than one latency."
+            )
+        validation = _materialize_starvla_runtime_cache(data_root_dir=data_root_dir, data_mix=data_mix)
+        eval_validation = _materialize_starvla_runtime_cache(data_root_dir=data_root_dir, data_mix=eval_data_mix)
+    else:
+        validation = _validate_starvla_dataset(data_root_dir=data_root_dir, data_mix=data_mix)
+        eval_validation = _validate_starvla_dataset(data_root_dir=data_root_dir, data_mix=eval_data_mix)
+
+    # If training used a latency_filter the embedded prompt map only has the filtered
+    # latencies. Load the missing latency subdirs from HF (already cached for training
+    # latencies; downloads only the needed eval-only ones) and fill in the map so that
+    # post-train eval at non-training latencies gets the correct dataset prompt.
+    latency_filter = getattr(args, "latency_filter", None)
+    eval_latencies = getattr(args, "eval_latencies", None)
+    all_needed = sorted({
+        *(int(v) for v in (latency_filter or [])),
+        *(int(v) for v in (eval_latencies or [])),
+    })
+    _remake_prompt_map_if_needed(required_latencies=all_needed)
+
+    return {
+        "dataset_ready": True,
+        "dataset_converted": converted,
+        "dataset_local_dir": str(data_root_dir),
+        "dataset_dir": str(dataset_dir),
+        "data_mix": data_mix,
+        "eval_data_mix": eval_data_mix,
+        "eval_dataset_dir": str(eval_dataset_dir),
+        "action_carrier": action_carrier,
+        "latency_prompt_map_path": str(prompt_map) if prompt_map.exists() else None,
+        **validation,
+        "eval_dataset_stats_path": eval_validation["dataset_stats_path"],
+        "eval_dataset_num_steps": eval_validation["dataset_num_steps"],
+        "eval_dataset_num_trajectories": eval_validation["dataset_num_trajectories"],
+    }
+
+
+def _ensure_flappy_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_flappy_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_flappy_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="flappy",
+        env_fps=30.0,
+        obs_fps=30.0,
+    )
+
+
+def _ensure_demon_attack_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_demon_attack_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_demon_attack_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="demon_attack",
+        env_fps=60.0,
+        obs_fps=15.0,
+    )
+
+
+def _ensure_defend_the_line_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_defend_the_line_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_defend_the_line_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="defend_the_line",
+        env_fps=35.0,
+        obs_fps=8.75,
+    )
+
+
+def _ensure_asterix_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_asterix_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_asterix_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="asterix",
+        env_fps=60.0,
+        obs_fps=15.0,
+    )
+
+
+def _ensure_atlantis_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_atlantis_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_atlantis_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="atlantis",
+        env_fps=60.0,
+        obs_fps=15.0,
+    )
+
+
+def _task_converter_and_verifier(task: str):
+    if task == "flappy":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_flappy_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_flappy_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_flappy"
+    if task == "demon_attack":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_demon_attack_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_demon_attack_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_demon_attack"
+    if task == "defend_the_line":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_defend_the_line_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_defend_the_line_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_defend_the_line"
+    if task == "deadly_corridor":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_deadly_corridor_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_deadly_corridor_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_deadly_corridor"
+    if task == "asterix":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_asterix_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_asterix_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_asterix"
+    if task == "atlantis":
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_atlantis_to_starvla_lerobot import convert_dataset
+        from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_atlantis_dataset import verify_dataset
+        return convert_dataset, verify_dataset, "rl_games_atlantis"
+    raise ValueError(
+        "cross-task rl_games currently supports only flappy, demon_attack, "
+        f"defend_the_line, deadly_corridor, asterix, and atlantis, got {task!r}"
+    )
+
+
+def _get_task_value(task_cfg: dict[str, Any], *names: str, default: Any = None) -> Any:
+    for name in names:
+        if name in task_cfg and task_cfg[name] not in (None, ""):
+            return task_cfg[name]
+    return default
+
+
+def _ensure_cross_task_datasets(args) -> dict[str, Any]:
+    from starVLA.dataloader.gr00t_lerobot.registry import load_custom_mixtures
+
+    cross_cfg = getattr(args, "cross_task", None) or {}
+    train_tasks = cross_cfg.get("train_tasks") if isinstance(cross_cfg, dict) else None
+    if not train_tasks:
+        raise ValueError("rl_games.cross_task.train_tasks must contain at least flappy and demon_attack entries")
+
+    data_root_dir = Path(args.dataset_local_dir).expanduser().resolve()
+    action_carrier = _action_carrier(args)
+    if action_carrier != "bridge":
+        raise ValueError("cross-task OpenVLA RL-games currently requires action_carrier=bridge")
+
+    mixture_entries: list[list[Any]] = []
+    val_mixture_entries: list[list[Any]] = []
+    prompt_maps: dict[str, str] = {}
+    converted_names: dict[str, str] = {}
+    converted = False
+    force = _str2bool(args.setup_force) or _str2bool(args.dataset_force_download)
+
+    for raw_task_cfg in train_tasks:
+        task_cfg = dict(raw_task_cfg)
+        task_name = str(_get_task_value(task_cfg, "name", "task"))
+        convert_dataset, verify_dataset, robot_type = _task_converter_and_verifier(task_name)
+
+        train_source_value = _get_task_value(task_cfg, "train_source_hf", "source_hf")
+        prompt_source_value = _get_task_value(task_cfg, "prompt_source_hf", default=train_source_value)
+        train_config_value = _get_task_value(task_cfg, "train_config_name", "source_config_name", "config_name", default=None)
+        prompt_config_value = _get_task_value(task_cfg, "prompt_config_name", default=train_config_value)
+        train_subdir_value = _get_task_value(task_cfg, "train_source_subdir", "source_subdir", default=None)
+        prompt_subdir_value = _get_task_value(task_cfg, "prompt_source_subdir", default=train_subdir_value)
+        train_source = str(train_source_value or "")
+        prompt_source = str(prompt_source_value or "")
+        train_config_name = None if train_config_value in (None, "") else str(train_config_value)
+        prompt_config_name = None if prompt_config_value in (None, "") else str(prompt_config_value)
+        train_subdir = train_subdir_value
+        prompt_subdir = prompt_subdir_value
+        if not train_source:
+            raise ValueError(f"cross-task train task {task_name} is missing train_source_hf/source_hf")
+        if not prompt_source:
+            raise ValueError(f"cross-task train task {task_name} is missing prompt_source_hf")
+
+        latency_filter = _get_task_value(task_cfg, "train_latency_filter", "latency_filter", default=None)
+        latency_filter = [int(value) for value in latency_filter] if latency_filter not in (None, "") else None
+        eval_latency_filter = _get_task_value(task_cfg, "eval_latency_filter", default=latency_filter)
+        eval_latency_filter = (
+            [int(value) for value in eval_latency_filter]
+            if eval_latency_filter not in (None, "")
+            else None
+        )
+        required_latencies = sorted({int(v) for v in (latency_filter or []) + (eval_latency_filter or [])}) or None
+        episodes_per_latency = _get_task_value(task_cfg, "episodes_per_latency", default=None)
+        episodes_per_latency = int(episodes_per_latency) if episodes_per_latency not in (None, "") else None
+        eval_episodes_per_latency = _get_task_value(task_cfg, "eval_episodes_per_latency", default=episodes_per_latency)
+        eval_episodes_per_latency = (
+            int(eval_episodes_per_latency)
+            if eval_episodes_per_latency not in (None, "")
+            else None
+        )
+        max_episodes = _get_task_value(task_cfg, "max_episodes", default=None)
+        max_episodes = int(max_episodes) if max_episodes not in (None, "") else None
+        weight = float(_get_task_value(task_cfg, "weight", default=1.0))
+        action_layout = _get_task_value(task_cfg, "action_layout", default=None)
+        action_layout = None if action_layout in (None, "") else str(action_layout)
+        task_timing = {
+            "flappy": (30.0, 1),
+            "demon_attack": (15.0, 4),
+            "defend_the_line": (8.75, 4),
+            "deadly_corridor": (8.75, 4),
+            "asterix": (15.0, 4),
+            "atlantis": (15.0, 4),
+        }[task_name]
+        task_fps, task_obs_stride_raw_frames = task_timing
+        task_target_latency_unit = (
+            "raw_frames" if task_name in {"asterix", "atlantis", "defend_the_line"} else "observation_steps"
+        )
+
+        base_converted_name = str(_get_task_value(task_cfg, "converted_name", default=f"{task_name}_cross_task_train"))
+        converted_base = _derived_dataset_name(
+            base_converted_name,
+            latency_filter=latency_filter,
+            episodes_per_latency=episodes_per_latency,
+            max_episodes=max_episodes,
+        )
+        data_mix = _carrier_dataset_name(converted_base, action_carrier)
+        dataset_dir = data_root_dir / data_mix
+        eval_data_mix = f"{data_mix}__val"
+        eval_dataset_dir = data_root_dir / eval_data_mix
+
+        train_revision = _get_task_value(task_cfg, "train_revision", default=None)
+        prompt_revision = _get_task_value(task_cfg, "prompt_revision", default=None)
+        train_for_reading = _fixed_dataset_source(
+            train_source, train_revision, train_subdir, required_latencies, args.dataset_cache_dir,
+        )
+        prompt_for_reading = _fixed_dataset_source(
+            prompt_source, prompt_revision, prompt_subdir, required_latencies, args.dataset_cache_dir,
+        )
+        prompt_map = _load_source_latency_prompt_map(
+            prompt_for_reading,
+            cache_dir=args.dataset_cache_dir,
+            dataset_config_name=prompt_config_name,
+            dataset_source_subdir=prompt_subdir,
+            source_latency_column="latency_raw_frames",
+            target_latency_unit=task_target_latency_unit,
+            obs_stride_raw_frames=task_obs_stride_raw_frames,
+            latencies=required_latencies,
+        )
+        prompt_dir = data_root_dir / "_prompt_maps" / data_mix
+        eval_prompt_map_path = _write_prompt_map(prompt_dir / "eval_latency_prompt_map.json", prompt_map)
+        train_prompt_map_path = _write_prompt_map(
+            prompt_dir / "train_latency_prompt_map.json",
+            prompt_map,
+            latency_filter=latency_filter,
+        )
+
+        def _manifest_matches(
+            dataset_path: Path,
+            *,
+            expected_latency_filter: list[int] | None,
+            expected_episodes_per_latency: int | None,
+        ) -> bool:
+
+            manifest_path = dataset_path / "manifest.json"
+            if not manifest_path.exists():
+                return False
+            try:
+                manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+            except Exception:
+                return False
+            expected_subdirs = [str(s) for s in resolve_latency_subdirs(train_subdir, expected_latency_filter)]
+            return (
+                str(manifest.get("source", "")) == train_source
+                and (train_revision is None or ("source_revision" in manifest and manifest["source_revision"] == train_revision))
+                and (prompt_revision is None or ("prompt_source_revision" in manifest and manifest["prompt_source_revision"] == prompt_revision))
+                and str(manifest.get("prompt_source", "")) == prompt_source
+                and (manifest.get("source_config") or None) == train_config_name
+                and (manifest.get("prompt_source_config") or None) == prompt_config_name
+                and (manifest.get("source_subdir") or None) == train_subdir
+                and (manifest.get("prompt_source_subdir") or None) == prompt_subdir
+                and manifest.get("latency_subdirs") == expected_subdirs
+                and str(manifest.get("action_carrier", "")) == action_carrier
+                and (not action_layout or str(manifest.get("action_layout", "")) == action_layout)
+                and manifest.get("latency_filter") == expected_latency_filter
+                and manifest.get("episodes_per_latency") == expected_episodes_per_latency
+                and manifest.get("max_episodes") == max_episodes
+                and manifest.get("fps") == task_fps
+                and manifest.get("obs_stride_raw_frames") == task_obs_stride_raw_frames
+                and manifest.get("target_latency_unit") == task_target_latency_unit
+                and manifest.get("source_latency_column") == "latency_raw_frames"
+            )
+
+        rebuild = (
+            force
+            or not _dataset_ready(dataset_dir)
+            or not _dataset_ready(eval_dataset_dir)
+            or not _manifest_matches(
+                dataset_dir,
+                expected_latency_filter=latency_filter,
+                expected_episodes_per_latency=episodes_per_latency,
+            )
+            or not _manifest_matches(
+                eval_dataset_dir,
+                expected_latency_filter=eval_latency_filter,
+                expected_episodes_per_latency=eval_episodes_per_latency,
+            )
+        )
+        if rebuild:
+            allow_mixed = bool(latency_filter or eval_latency_filter) and train_source == prompt_source
+            verify_kwargs = {
+                "rows": args.verify_rows,
+                "cache_dir": args.dataset_cache_dir,
+                "dataset_config_name": train_config_name,
+                "dataset_source_subdir": train_subdir,
+                "strict": True,
+                "allow_mixed_latency_prompts": allow_mixed,
+                "source_latency_column": "latency_raw_frames",
+                "target_latency_unit": task_target_latency_unit,
+                "obs_stride_raw_frames": task_obs_stride_raw_frames,
+            }
+            if action_layout and "action_layout" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["action_layout"] = action_layout
+            if "latencies" in inspect.signature(verify_dataset).parameters:
+                verify_kwargs["latencies"] = required_latencies
+            verify_dataset(train_for_reading, **verify_kwargs)
+
+            convert_kwargs = {
+                "cache_dir": args.dataset_cache_dir,
+                "dataset_config_name": train_config_name,
+                "dataset_source_subdir": train_subdir,
+                "max_episodes": max_episodes,
+                "force": rebuild,
+                "require_latency_prompt_map": bool(latency_filter or eval_latency_filter),
+                "latency_filter": latency_filter,
+                "train_latency_filter": latency_filter,
+                "eval_latency_filter": eval_latency_filter,
+                "episodes_per_latency": episodes_per_latency,
+                "train_episodes_per_latency": episodes_per_latency,
+                "eval_episodes_per_latency": eval_episodes_per_latency,
+                "prompt_map_override": prompt_map,
+                "default_latency": 0,
+                "action_carrier": action_carrier,
+                "fps": task_fps,
+                "obs_stride_raw_frames": task_obs_stride_raw_frames,
+                "source_latency_column": "latency_raw_frames",
+                "target_latency_unit": task_target_latency_unit,
+                "source_rows_unit": "decision_step",
+            }
+            if action_layout:
+                convert_kwargs["action_layout"] = action_layout
+            supported_convert_kwargs = inspect.signature(convert_dataset).parameters
+            convert_kwargs = {
+                key: value
+                for key, value in convert_kwargs.items()
+                if key in supported_convert_kwargs
+            }
+            convert_dataset(train_for_reading, dataset_dir, **convert_kwargs)
+            for manifest_path in (dataset_dir / "manifest.json", eval_dataset_dir / "manifest.json"):
+                manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+                manifest["source"] = train_source
+                manifest["source_revision"] = train_revision
+                manifest["prompt_source"] = prompt_source
+                manifest["prompt_source_revision"] = prompt_revision
+                manifest["prompt_source_config"] = prompt_config_name
+                manifest["prompt_source_subdir"] = prompt_subdir
+                manifest["eval_prompt_map_path"] = str(eval_prompt_map_path)
+                manifest["train_prompt_map_path"] = str(train_prompt_map_path)
+                manifest["train_latency_filter"] = latency_filter
+                manifest["eval_latency_filter"] = eval_latency_filter
+                manifest["train_episodes_per_latency"] = episodes_per_latency
+                manifest["eval_episodes_per_latency"] = eval_episodes_per_latency
+                manifest_path.write_text(json.dumps(manifest, indent=2), encoding="utf-8")
+            # The per-task derived names (e.g. ..._lat0_2_4__40ep_per_lat__bridge)
+            # are not in any static data_config, and the combined cross__ mixture
+            # is only registered after this loop. Register each as a trivial
+            # single-dataset mixture so the per-task materialize below can resolve
+            # it via get_dataset_named_mixture.
+            from starVLA.dataloader.gr00t_lerobot.registry import DATASET_NAMED_MIXTURES
+
+            DATASET_NAMED_MIXTURES[data_mix] = [[data_mix, weight, robot_type]]
+            DATASET_NAMED_MIXTURES[eval_data_mix] = [[eval_data_mix, weight, robot_type]]
+            _materialize_starvla_runtime_cache(data_root_dir=data_root_dir, data_mix=data_mix)
+            _materialize_starvla_runtime_cache(data_root_dir=data_root_dir, data_mix=eval_data_mix)
+            converted = True
+
+        mixture_entries.append([data_mix, weight, robot_type])
+        val_mixture_entries.append([eval_data_mix, weight, robot_type])
+        prompt_maps[task_name] = str(eval_prompt_map_path)
+        converted_names[task_name] = data_mix
+
+    mixture_name = "cross__" + "__".join(converted_names[task] for task in sorted(converted_names))
+    mixture_name = _safe_token(mixture_name)
+    eval_mixture_name = f"{mixture_name}__val"
+    mixture_file_name = f"cross_{hashlib.sha1(mixture_name.encode('utf-8')).hexdigest()[:16]}.json"
+    custom_mixtures_path = data_root_dir / "_generated_mixtures" / mixture_file_name
+    custom_mixtures_path.parent.mkdir(parents=True, exist_ok=True)
+    custom_mixtures_path.write_text(
+        json.dumps({mixture_name: mixture_entries, eval_mixture_name: val_mixture_entries}, indent=2),
+        encoding="utf-8",
+    )
+    load_custom_mixtures(custom_mixtures_path)
+    validation = _validate_starvla_dataset(data_root_dir=data_root_dir, data_mix=mixture_name)
+    eval_validation = _validate_starvla_dataset(data_root_dir=data_root_dir, data_mix=eval_mixture_name)
+    return {
+        "dataset_ready": True,
+        "dataset_converted": converted,
+        "dataset_local_dir": str(data_root_dir),
+        "dataset_dir": str(data_root_dir),
+        "data_mix": mixture_name,
+        "eval_data_mix": eval_mixture_name,
+        "custom_mixtures_path": str(custom_mixtures_path),
+        "cross_task_prompt_maps": prompt_maps,
+        "cross_task_datasets": converted_names,
+        "action_carrier": action_carrier,
+        **validation,
+        "eval_dataset_stats_path": eval_validation["dataset_stats_path"],
+        "eval_dataset_num_steps": eval_validation["dataset_num_steps"],
+        "eval_dataset_num_trajectories": eval_validation["dataset_num_trajectories"],
+    }
+
+
+def _ensure_deadly_corridor_dataset(args) -> dict[str, Any]:
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.convert_deadly_corridor_to_starvla_lerobot import convert_dataset
+    from examples.rl_games.bash_scripts.gr00t.data_conversion.verify_deadly_corridor_dataset import verify_dataset
+
+    return _ensure_rl_games_lerobot_dataset(
+        args,
+        convert_dataset=convert_dataset,
+        verify_dataset=verify_dataset,
+        env_name="deadly_corridor",
+        env_fps=35.0,
+        obs_fps=8.75,
+    )
+
+
+def setup_assets(args) -> dict[str, Any]:
+    result: dict[str, Any] = {
+        "model": args.model,
+        "env": args.env,
+        "mode": args.mode,
+        "initialization_mode": _initialization_mode(args),
+        "action_carrier": _action_carrier(args),
+    }
+
+    supported_models = {"openvla", "pi0", "pi05", "gr00t"}
+    if args.model in supported_models and str(getattr(args, "env", "")) == "cross_task":
+        result.update(_ensure_cross_task_datasets(args))
+    elif args.model in supported_models and args.env == "flappy":
+        result.update(_ensure_flappy_dataset(args))
+    elif args.model in supported_models and args.env == "demon_attack":
+        result.update(_ensure_demon_attack_dataset(args))
+    elif args.model in supported_models and args.env == "defend_the_line":
+        result.update(_ensure_defend_the_line_dataset(args))
+    elif args.model in supported_models and args.env == "asterix":
+        result.update(_ensure_asterix_dataset(args))
+    elif args.model in supported_models and args.env == "atlantis":
+        result.update(_ensure_atlantis_dataset(args))
+    elif args.model in supported_models and args.env == "deadly_corridor":
+        result.update(_ensure_deadly_corridor_dataset(args))
+    else:
+        data_root_dir = Path(args.dataset_local_dir).expanduser().resolve()
+        result.update({
+            "dataset_ready": _has_files(data_root_dir),
+            "dataset_local_dir": str(data_root_dir),
+            "data_mix": None,
+            "eval_data_mix": None,
+            "latency_prompt_map_path": None,
+        })
+
+    base_model_dir = Path(args.base_model_dir).expanduser().resolve()
+    result.update(_ensure_base_model(args.model, base_model_dir, args.base_model_repo_id))
+
+    checkpoint_dir = Path(args.checkpoint_local_dir).expanduser().resolve()
+    explicit_checkpoint = str(getattr(args, "checkpoint", "") or "")
+    if explicit_checkpoint:
+        resume_checkpoint, resume_step, resume_kind = _resolve_explicit_resume_checkpoint(explicit_checkpoint, checkpoint_dir)
+        result.update({
+            "resume_found": True,
+            "resume_source": "explicit",
+            "resume_kind": resume_kind,
+            "resume_checkpoint": str(resume_checkpoint),
+            "resume_step": resume_step,
+            "checkpoint_local_dir": str(checkpoint_dir),
+        })
+        return result
+
+    local_ckpt, local_step, local_kind = (None, 0, None)
+    if args.checkpoint_load in {"auto", "local"}:
+        local_ckpt, local_step, local_kind = _find_latest_local_checkpoint(checkpoint_dir)
+        if local_ckpt is not None and args.checkpoint_load == "local":
+            result.update({
+                "resume_found": True,
+                "resume_source": "local",
+                "resume_kind": local_kind,
+                "resume_checkpoint": str(local_ckpt),
+                "resume_step": local_step,
+                "checkpoint_local_dir": str(checkpoint_dir),
+            })
+            return result
+
+    hf_repo_id = args.checkpoint_hf_repo_id or args.hf_repo_id
+    if args.checkpoint_load in {"auto", "hf"} and hf_repo_id:
+        hf_ckpt, hf_step, hf_kind, hf_error = _download_latest_hf_checkpoint(hf_repo_id, checkpoint_dir)
+        hf_is_better = (
+            args.checkpoint_load == "hf"
+            or local_ckpt is None
+            or hf_step > local_step
+            or (hf_step == local_step and hf_kind == "state" and local_kind != "state")
+        )
+        if hf_ckpt is not None and hf_is_better:
+            result.update({
+                "resume_found": True,
+                "resume_source": "hf",
+                "resume_kind": hf_kind,
+                "resume_checkpoint": str(hf_ckpt),
+                "resume_step": hf_step,
+                "checkpoint_local_dir": str(checkpoint_dir),
+            })
+            return result
+        if local_ckpt is not None:
+            result.update({
+                "resume_found": True,
+                "resume_source": "local",
+                "resume_kind": local_kind,
+                "resume_checkpoint": str(local_ckpt),
+                "resume_step": local_step,
+                "checkpoint_local_dir": str(checkpoint_dir),
+            })
+            if hf_ckpt is not None:
+                result["checkpoint_hf_status"] = (
+                    f"local checkpoint step {local_step} is newer than or equal to HF step {hf_step}; using local"
+                )
+            return result
+        if hf_error:
+            sync_repo_id = str(getattr(args, "checkpoint_sync_repo_id", "") or "")
+            sync_enabled = _str2bool(getattr(args, "checkpoint_sync_enabled", False))
+            if (
+                args.checkpoint_load == "auto"
+                and sync_enabled
+                and sync_repo_id == hf_repo_id
+            ):
+                result["checkpoint_hf_status"] = (
+                    f"HF resume repo {hf_repo_id} was not available during auto-resume; "
+                    "starting from local/base model and using it as the checkpoint sync destination"
+                )
+            else:
+                result["checkpoint_hf_warning"] = hf_error
+
+    if local_ckpt is not None:
+        result.update({
+            "resume_found": True,
+            "resume_source": "local",
+            "resume_kind": local_kind,
+            "resume_checkpoint": str(local_ckpt),
+            "resume_step": local_step,
+            "checkpoint_local_dir": str(checkpoint_dir),
+        })
+        return result
+
+    initialization_hf_repo_id = str(getattr(args, "initialization_hf_repo_id", "") or "")
+    initialization_local_dir = str(getattr(args, "initialization_local_dir", "") or "")
+    initialization_checkpoint_filename = str(getattr(args, "initialization_checkpoint_filename", "") or "")
+    has_initialization_source = bool(initialization_local_dir or initialization_hf_repo_id)
+    if has_initialization_source and _initialization_mode(args) in INITIALIZATION_SOURCE_MODES:
+        local_init_ckpt, local_init_step = _resolve_local_initialization_checkpoint(
+            initialization_local_dir,
+            initialization_checkpoint_filename,
+        )
+        if local_init_ckpt is not None:
+            result.update({
+                "resume_found": False,
+                "resume_source": None,
+                "resume_kind": None,
+                "resume_checkpoint": None,
+                "resume_step": 0,
+                "checkpoint_local_dir": str(checkpoint_dir),
+                "pretrained_checkpoint": str(local_init_ckpt),
+                "initialization_source": "local",
+                "initialization_local_dir": str(Path(initialization_local_dir).expanduser().resolve()),
+                "initialization_hf_repo_id": initialization_hf_repo_id or None,
+                "initialization_checkpoint_filename": initialization_checkpoint_filename or None,
+                "initialization_step": local_init_step,
+            })
+            return result
+
+        if not initialization_hf_repo_id:
+            raise FileNotFoundError(
+                f"Bridge initialization checkpoint not found under initialization_local_dir="
+                f"{initialization_local_dir!r} with filename={initialization_checkpoint_filename!r}, "
+                "and no initialization_hf_repo_id was provided for download."
+            )
+
+        init_dir = checkpoint_dir / "_initialization" / _safe_path_name(initialization_hf_repo_id)
+        if initialization_checkpoint_filename:
+            init_ckpt, init_step, init_error = _download_hf_checkpoint_file(
+                initialization_hf_repo_id,
+                initialization_checkpoint_filename,
+                init_dir,
+            )
+            init_kind = "model" if init_ckpt is not None else None
+        else:
+            init_ckpt, init_step, init_kind, init_error = _download_latest_hf_checkpoint(initialization_hf_repo_id, init_dir)
+        if init_ckpt is None:
+            raise RuntimeError(
+                f"Could not download bridge initialization checkpoint from {initialization_hf_repo_id}: "
+                f"{init_error or 'no checkpoint files found'}"
+            )
+        if init_kind == "state":
+            raise ValueError(
+                f"Bridge initialization expected model weights, but {initialization_hf_repo_id} resolved "
+                f"to a full training-state directory: {init_ckpt}"
+            )
+        result.update({
+            "resume_found": False,
+            "resume_source": None,
+            "resume_kind": None,
+            "resume_checkpoint": None,
+            "resume_step": 0,
+            "checkpoint_local_dir": str(checkpoint_dir),
+            "pretrained_checkpoint": str(init_ckpt),
+            "initialization_source": "hf",
+            "initialization_local_dir": initialization_local_dir or None,
+            "initialization_hf_repo_id": initialization_hf_repo_id,
+            "initialization_checkpoint_filename": initialization_checkpoint_filename or None,
+            "initialization_step": init_step,
+        })
+        return result
+
+    result.update({
+        "resume_found": False,
+        "resume_source": None,
+        "resume_kind": None,
+        "resume_checkpoint": None,
+        "resume_step": 0,
+        "checkpoint_local_dir": str(checkpoint_dir),
+    })
+    return result
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", required=True)
+    parser.add_argument("--env", required=True)
+    parser.add_argument("--mode", required=True)
+    parser.add_argument("--initialization-mode", default="")
+    parser.add_argument("--action-carrier", choices=["", "native", "bridge"], default="")
+    parser.add_argument("--action-layout", default="")
+    parser.add_argument("--deadly-action-layout", default="")
+    parser.add_argument("--latency-mode", default="")
+    parser.add_argument("--source-dataset-hf", default="")
+    parser.add_argument("--source-dataset-revision", default=None)
+    parser.add_argument("--source-dataset-config-name", default=None)
+    parser.add_argument("--dataset-local-dir", required=True)
+    parser.add_argument("--converted-dataset-name", default="flappy_train")
+    parser.add_argument("--dataset-cache-dir", default=None)
+    parser.add_argument("--dataset-force-download", default="false")
+    parser.add_argument("--setup-force", default="false")
+    parser.add_argument("--skip-verification", default="false")
+    parser.add_argument(
+        "--target-latency-unit",
+        choices=["raw_frames", "observation_steps"],
+        default="observation_steps",
+    )
+    parser.add_argument("--verify-rows", type=int, default=200)
+    parser.add_argument("--max-episodes", type=int, default=None)
+    parser.add_argument("--episodes-per-latency", type=int, default=None)
+    parser.add_argument("--latency-filter", default=None)
+    parser.add_argument("--base-model-dir", required=True)
+    parser.add_argument("--base-model-repo-id", default=None)
+    parser.add_argument("--checkpoint-local-dir", required=True)
+    parser.add_argument("--checkpoint", default="")
+    parser.add_argument("--checkpoint-load", choices=["auto", "none", "local", "hf"], default="auto")
+    parser.add_argument("--checkpoint-hf-repo-id", default="")
+    parser.add_argument("--checkpoint-save-best-model", default="true")
+    parser.add_argument("--initialization-local-dir", default="")
+    parser.add_argument("--initialization-hf-repo-id", default="")
+    parser.add_argument("--initialization-checkpoint-filename", default="")
+    parser.add_argument("--checkpoint-sync-enabled", default="false")
+    parser.add_argument("--checkpoint-sync-repo-id", default="")
+    parser.add_argument("--hf-repo-id", default="")
+    args = parser.parse_args()
+    if args.dataset_cache_dir == "":
+        args.dataset_cache_dir = None
+    if args.base_model_repo_id == "":
+        args.base_model_repo_id = None
+    if isinstance(args.latency_filter, str) and args.latency_filter:
+        args.latency_filter = [int(item) for item in args.latency_filter.split(",") if item.strip()]
+    elif args.latency_filter == "":
+        args.latency_filter = None
+
+    with contextlib.redirect_stdout(sys.stderr):
+        result = setup_assets(args)
+    print(json.dumps(result))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

experiments/task-transfer/provenance/retirement_object_gate.json

@@ -0,0 +1,18 @@
+{
+  "all_source_head_files_preserved": true,
+  "current_target_objects_verified": 252,
+  "fixed_references_verified": 126,
+  "method": "remote Git/LFS object identity and size comparisons",
+  "complete_framework_loading": "not repeated",
+  "large_payload_downloads": "stopped",
+  "sources": {
+    "latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3": "c4a60d71edd7ecd15af5b63b764f8c0742e0d3f7",
+    "latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1": "53eb98e6177d630b3ee9775c84413300d7b0f29b",
+    "latency-sensitive-bench/starvla-demon-attack-zero-latency-evals": "c85322cb658371a46195d4acd34337373a2f7d28",
+    "latency-sensitive-bench/starvla-demon-attack-mixed-024-latency-evals": "602487c53e6aec515e61aebb7afb44845f6ad468",
+    "latency-sensitive-bench/deadly_1000ep": "fa0c35ef651bb1f938d3830f18922a0383c3fb13",
+    "latency-sensitive-bench/asterix_200ep": "726b2fc79ff0509d8819e564ab2695daa5215ced",
+    "latency-sensitive-bench/atlantis_200ep": "447f7bf629302e9646c24c7e97a03b3a5d5ad8d2",
+    "latency-sensitive-bench/defend_the_line_200ep": "1f393d6e99cac2ec9d4fedd009d520bb6695ab6a"
+  }
+}

experiments/task-transfer/provenance/syntax_verification.json

@@ -0,0 +1,6 @@
+{
+  "shell": 27,
+  "python": 2,
+  "embedded_python": 7,
+  "yaml": 2
+}