Publish atlantis qwenoft H8 evaluation artifacts

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/README.md

@@ -1,8 +1,8 @@
-# Atlantis qwenoft zero-latency H8 VLA
+# Atlantis qwenoft zero-trained H8 VLA
 
 Current RGB and eight binary decision-history windows; action horizon one. Initialized from the pinned Qwen3-VL-4B backbone and random action head.
 
 - Training: 5000 steps, seed 42, global batch 128
 - W&B: https://wandb.ai/saberrr-zju/starvla_tasks/runs/40be052b
-- Evaluation: 100 episodes, seeds 1000000–1000099, GPU simulator, parallel 32, capacity 1, 60/15 Hz, 7200 raw-frame cap, zero latency
-- Mean raw return: 39560.0
+- Evaluation: 100 episodes, seeds 1000000–1000099, GPU simulator, parallel 32, capacity 1, 60/15 Hz, 7200 raw-frame cap, profile latency
+- Mean raw return: 18588.0

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/eval_profile.json

@@ -0,0 +1,217 @@
+{
+  "checkpoint_path": "/home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/atlantis-qwenoft-hist8-zero-s42-v1/checkpoints/model.pt",
+  "experiment_name": "atlantis-qwenoft-hist8-zero-s42-v1-step5000-history01-profile-par32",
+  "latency": "profile_sample",
+  "latency_type": "profile_sample",
+  "lengths": [
+    6502,
+    4860,
+    6607,
+    5080,
+    6437,
+    3553,
+    3278,
+    5311,
+    3748,
+    5104,
+    5768,
+    3558,
+    4806,
+    6103,
+    5411,
+    5644,
+    2602,
+    4553,
+    5281,
+    4300,
+    3326,
+    2793,
+    2433,
+    5519,
+    5334,
+    3928,
+    6605,
+    3182,
+    4062,
+    7200,
+    5238,
+    4377,
+    4530,
+    4282,
+    3908,
+    6863,
+    5245,
+    6844,
+    5974,
+    4005,
+    4792,
+    4254,
+    5061,
+    7000,
+    4033,
+    5157,
+    5771,
+    3723,
+    4875,
+    3657,
+    5048,
+    3179,
+    4804,
+    3545,
+    3753,
+    5440,
+    4275,
+    4658,
+    6598,
+    4265,
+    5644,
+    3967,
+    7200,
+    3009,
+    4001,
+    7200,
+    4702,
+    5983,
+    5245,
+    5501,
+    7200,
+    4108,
+    5188,
+    7200,
+    4230,
+    7200,
+    4982,
+    4156,
+    7132,
+    5178,
+    7200,
+    5082,
+    4799,
+    4514,
+    4587,
+    6666,
+    6847,
+    5767,
+    6339,
+    5618,
+    5613,
+    3899,
+    4318,
+    4763,
+    4798,
+    4513,
+    4716,
+    4314,
+    5162,
+    5178
+  ],
+  "mean_length": 5017.61,
+  "mean_return": 18588.0,
+  "returns": [
+    23600.0,
+    18600.0,
+    21700.0,
+    21800.0,
+    20500.0,
+    12200.0,
+    9500.0,
+    16800.0,
+    13800.0,
+    19900.0,
+    24500.0,
+    13700.0,
+    13700.0,
+    17300.0,
+    19000.0,
+    27900.0,
+    6900.0,
+    21600.0,
+    16000.0,
+    17700.0,
+    10500.0,
+    9800.0,
+    6300.0,
+    17700.0,
+    12400.0,
+    8800.0,
+    22900.0,
+    9700.0,
+    10300.0,
+    26700.0,
+    14400.0,
+    22300.0,
+    16300.0,
+    13700.0,
+    16800.0,
+    29200.0,
+    14400.0,
+    28600.0,
+    20100.0,
+    14800.0,
+    16500.0,
+    14200.0,
+    20700.0,
+    27200.0,
+    11400.0,
+    19700.0,
+    17400.0,
+    7900.0,
+    18900.0,
+    17200.0,
+    20300.0,
+    9600.0,
+    20400.0,
+    13200.0,
+    13300.0,
+    21400.0,
+    15400.0,
+    16000.0,
+    26000.0,
+    12900.0,
+    26400.0,
+    9500.0,
+    25500.0,
+    9100.0,
+    20000.0,
+    25900.0,
+    19200.0,
+    23400.0,
+    12600.0,
+    15600.0,
+    36300.0,
+    13100.0,
+    14700.0,
+    38100.0,
+    17300.0,
+    29000.0,
+    25900.0,
+    13100.0,
+    34900.0,
+    19900.0,
+    32600.0,
+    14900.0,
+    20900.0,
+    14900.0,
+    15100.0,
+    25700.0,
+    30300.0,
+    20600.0,
+    23500.0,
+    27300.0,
+    22800.0,
+    12600.0,
+    17400.0,
+    15500.0,
+    19500.0,
+    17100.0,
+    14800.0,
+    20500.0,
+    17100.0,
+    24200.0
+  ],
+  "seed": 1000000,
+  "source_profile_path": "/home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/teacher_configs/atlantis-qwenoft/profile/profile.json",
+  "std_return": 6544.543987169771,
+  "suite_name": "profile_sample",
+  "timestamp_utc": "2026-09-24T12:41:59.686946+00:00"
+}

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/evaluation_profile/action_audit.json

@@ -0,0 +1,133 @@
+{
+  "eval_protocol_version": "hist8-history01-sf-render-v4",
+  "history_encoding": "binary_0_1",
+  "episodes": 100,
+  "episode_seeds": [
+    1000000,
+    1000001,
+    1000002,
+    1000003,
+    1000004,
+    1000005,
+    1000006,
+    1000007,
+    1000008,
+    1000009,
+    1000010,
+    1000011,
+    1000012,
+    1000013,
+    1000014,
+    1000015,
+    1000016,
+    1000017,
+    1000018,
+    1000019,
+    1000020,
+    1000021,
+    1000022,
+    1000023,
+    1000024,
+    1000025,
+    1000026,
+    1000027,
+    1000028,
+    1000029,
+    1000030,
+    1000031,
+    1000032,
+    1000033,
+    1000034,
+    1000035,
+    1000036,
+    1000037,
+    1000038,
+    1000039,
+    1000040,
+    1000041,
+    1000042,
+    1000043,
+    1000044,
+    1000045,
+    1000046,
+    1000047,
+    1000048,
+    1000049,
+    1000050,
+    1000051,
+    1000052,
+    1000053,
+    1000054,
+    1000055,
+    1000056,
+    1000057,
+    1000058,
+    1000059,
+    1000060,
+    1000061,
+    1000062,
+    1000063,
+    1000064,
+    1000065,
+    1000066,
+    1000067,
+    1000068,
+    1000069,
+    1000070,
+    1000071,
+    1000072,
+    1000073,
+    1000074,
+    1000075,
+    1000076,
+    1000077,
+    1000078,
+    1000079,
+    1000080,
+    1000081,
+    1000082,
+    1000083,
+    1000084,
+    1000085,
+    1000086,
+    1000087,
+    1000088,
+    1000089,
+    1000090,
+    1000091,
+    1000092,
+    1000093,
+    1000094,
+    1000095,
+    1000096,
+    1000097,
+    1000098,
+    1000099
+  ],
+  "issued_actions": 62677,
+  "action_counts": {
+    "0": 15821,
+    "2": 29659,
+    "3": 12666,
+    "1": 4531
+  },
+  "largest_action_fraction": 0.47320388659316814,
+  "raw_score_min": [
+    -16.4777889251709,
+    -17.36240005493164,
+    -17.736751556396484,
+    -16.52406120300293
+  ],
+  "raw_score_max": [
+    20.059650421142578,
+    20.963884353637695,
+    20.347089767456055,
+    20.745391845703125
+  ],
+  "raw_score_mean": [
+    1.5759939295484995,
+    -6.00103662508784,
+    6.868458961622308,
+    -0.553409212165336
+  ]
+}

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/evaluation_profile/eval_config.yaml

@@ -0,0 +1,149 @@
+experiment:
+  name: atlantis-qwenoft-hist8-zero-s42-v1-step5000-history01-profile-par32
+  seed: 1000000
+backend:
+  type: sample_factory
+  algo: APPO
+  device: cuda
+  train_dir: results/sample_factory
+  restart_behavior: resume
+  run_mode: eval
+  extra_args:
+  - --encoder_conv_architecture
+  - convnet_simple
+executor:
+  mode: simulated
+  simulated_worker_capacity: 1
+  simulated_inference_pool: true
+  inference_devices:
+  - cuda:0
+  - cuda:1
+  inference_batch_size: 16
+env:
+  name: gymnasium
+  task_name: atlantis
+  env_id: LatencyBench/Atlantis-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_atlantis
+  make_kwargs:
+    base_env_id: ALE/Atlantis-v5
+    render_mode: rgb_array
+    screen_size: 84
+    noop_max: 30
+    base_make_kwargs:
+      obs_type: rgb
+      frameskip: 1
+      repeat_action_probability: 0.0
+      full_action_space: false
+      mode: 0
+      difficulty: 0
+      max_num_frames_per_episode: 108000
+  env_fps: 60
+  obs_fps: 15
+  frame_stack: 1
+  simulator: gpu
+  terminal_on_life_loss: false
+  clip_reward: false
+  noop_action: noop
+  action_map:
+    noop: 0
+    fire: 1
+    rightfire: 2
+    leftfire: 3
+  action_order:
+  - noop
+  - fire
+  - rightfire
+  - leftfire
+  base_prompt: 'Defend Atlantis by firing at descending enemies. Choose exactly one
+    action from: noop, fire, rightfire, leftfire.'
+  obs_resize:
+  - 224
+  - 224
+  gpu_env_device: auto
+  noop_max: 30
+  action_history_decisions: 8
+latency:
+  method: temporal
+  profile_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/teacher_configs/atlantis-qwenoft/profile/profile.json
+  profile_worker_slot: 0
+  seed: 271828
+  add_latency_info: false
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  actions:
+  - noop
+  - fire
+  - rightfire
+  - leftfire
+  checkpoint_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/atlantis-qwenoft-hist8-zero-s42-v1/checkpoints/model.pt
+  model_config_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/atlantis-qwenoft-hist8-zero-s42-v1/config.yaml
+  task_manifest_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/atlantis-qwenoft-hist8-zero-s42-v1/manifest.json
+  device: cuda:0
+  unnorm_key: new_embodiment
+  prompt_mode: latency_neutral
+  backbone_path: /home/ubuntu/.cache/huggingface/hub/models--Qwen--Qwen3-VL-4B-Instruct/snapshots/ebb281ec70b05090aa6165b016eac8ec08e71b17
+  worker_python_executable: /home/ubuntu/miniconda3/envs/starvla_rl_games_openvla/bin/python3.10
+  state_source: transport
+training:
+  train_for_env_steps: 1000000
+  num_workers: 4
+  num_envs_per_worker: 16
+  worker_num_splits: 1
+  num_policies: 1
+  batch_size: 1024
+  rollout: 64
+  recurrence: 32
+  num_epochs: 2
+  num_batches_per_epoch: 8
+  max_policy_lag: 300
+  learning_rate: 0.000303
+  lr_schedule: constant
+  nonlinearity: elu
+  gamma: 0.99
+  gae_lambda: 0.95
+  ppo_clip_ratio: 0.1
+  ppo_clip_value: 0.2
+  exploration_loss: entropy
+  exploration_loss_coeff: 0.003
+  value_loss_coeff: 0.5
+  max_grad_norm: 0.0
+  adam_eps: 1.0e-06
+  obs_scale: 1.0
+  adaptive_stddev: false
+  with_vtrace: false
+  async_rl: true
+  use_rnn: false
+  normalize_input: true
+  normalize_returns: true
+  stats_avg: 100
+  experiment_summaries_interval: 1
+  save_every_sec: 600
+  keep_checkpoints: 5
+evaluation:
+  eval_interval_steps: null
+  eval_episodes: 100
+  eval_parallel_envs: 32
+  eval_max_steps: 7200
+  eval_deterministic: true
+  eval_suites:
+    fixed: []
+    normal: []
+    uniform: []
+  latency_bench_env_backend: atlantis_gpu_batched
+  eval_raw_reward: true
+logging:
+  output_dir: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/evals/atlantis-qwenoft-hist8-zero-s42-v1/zero_to_profile
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: true
+  save_action_records: true
+  save_latency_records: true
+  wandb_project: null
+  wandb_group: null
+  wandb_job_type: null
+  wandb_tags: ''

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/evaluation_profile/provenance.json

@@ -0,0 +1,46 @@
+{
+  "collection": {
+    "episodes": 200,
+    "filter_preset": "none",
+    "history": "decision_v1",
+    "history_decisions": 8,
+    "history_source": "teacher_transport",
+    "max_episode_raw_frames": 7200,
+    "rollout_overrides": {
+      "env_fps": 60,
+      "fixed_latency_ms": 0.0,
+      "latency_type": "fixed",
+      "obs_fps": 15,
+      "seed": 42
+    },
+    "student_image_frames": 1,
+    "teacher_checkpoint": "best_000026560_13598720_reward_675769.000.pth",
+    "teacher_path": "zero-latency/atlantis/small-policy/sample-factory-hist8-cpu15m-s0-v1",
+    "teacher_repo": "latency-sensitive-bench/benchmark-models",
+    "teacher_revision": "dfaeecb84cd6b5dfcc3adc0935d0ae86d0b504d2",
+    "teacher_sha256": "d41d728d8c2329d3623eb9e01bd9f3f2df4fc0d4f6f2c1f799c4badc24af78c5",
+    "train_labels": 324000,
+    "training_latency_condition": "zero"
+  },
+  "training_run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
+  "training_step": 5000,
+  "global_batch": 128,
+  "base_model": "Qwen/Qwen3-VL-4B-Instruct",
+  "wandb_run_id": "40be052b",
+  "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
+  "eval_protocol_version": "hist8-history01-sf-render-v4",
+  "history_encoding": "binary_0_1",
+  "episodes": 100,
+  "env_condition": "profile",
+  "parallel_envs": 32,
+  "max_episode_raw_frames": 7200,
+  "inference_devices": [
+    "cuda:0",
+    "cuda:1"
+  ],
+  "episode_seeds": [
+    1000000,
+    1000099
+  ],
+  "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4"
+}

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/profile_publication.json

@@ -0,0 +1,28 @@
+{
+  "model": "qwenoft",
+  "model_publication": {
+    "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
+    "episodes": 100,
+    "mean_return": 39560.0,
+    "model": "qwenoft",
+    "path": "zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1",
+    "repo_id": "latency-sensitive-bench/benchmark-models",
+    "revision": "15c9cd4ab12215ccbcef74a8cd4d051d8e3aae3f",
+    "run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
+    "task": "atlantis",
+    "training_status": "complete",
+    "zero_evaluation_status": "complete"
+  },
+  "profile": {
+    "GPU_CLASS": "1x-rtx3090",
+    "INSTANCE_ID": "instance_8dd832253b2c3d69",
+    "PROFILE_PATH": "results/profiling/profiles/qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
+    "PROFILE_REF": "qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69",
+    "RUN_ID": "20260924T110210282159Z"
+  },
+  "profile_path": "profiles/qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
+  "profile_repo": "latency-sensitive-bench/profiles",
+  "profile_revision": "7efc9af422864b9a1cb569aa81f1b2b1e216b89e",
+  "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4",
+  "task": "atlantis"
+}

zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1/publication_status.json

@@ -1,6 +1,8 @@
 {
   "episodes": 100,
-  "mean_return": 39560.0,
+  "mean_return": 18588.0,
+  "profile_evaluation_status": "complete",
+  "profile_mean_return": 18588.0,
   "training_status": "complete",
   "zero_evaluation_status": "complete"
 }