{
  "format": "abc_checkpoint_metadata/v1",
  "name": "abc_dit_xl_200k_model.pt",
  "uri": "https://abc-data.timehorizons.org/checkpoints/abc_dit_xl_200k_model.pt",
  "bytes": 8063017059,
  "sha256": "79754ece8d333789a930d1a964ee0d4548f5975e2ba175ddce41d018c9f51ba7",
  "step": 200000,
  "license": "Apache-2.0",
  "license_notes": [
    "Vision backbone derived from DINOv3; its licence restrictions apply to these weights as well as to the code that loads them - see the Licenses section of the repository README."
  ],
  "arch": {
    "model_size": "dit_xL",
    "hidden_size": 1536,
    "depth": 32,
    "num_heads": 24,
    "vision_backbone": "dinov3",
    "vision_model_name": "ViT-B/16",
    "task_encoder": "clip",
    "chunk_length": 30,
    "action_dim": 14,
    "params": 2015695908,
    "dtype": "float32",
    "tensors": 703
  },
  "max_action_prefix": 8,
  "default_prefix_length": 5,
  "norm_stats": "embedded under the 'norm_stats' key; all four training components shared identical stats, so no per-dataset selection is needed",
  "training_mixture": "xdof_csf_3_5k_may20_sim0514_hours_weighted - 3553.2 h real teleop plus 98.47 h sim across the 5 tasks in sim_prompt_map (sim share ~2.70%)",
  "sim_prompt_map": {
    "put_plastic_bottles_in_bin": "sim throw plastic bottles in bin",
    "sweep_away_paper_scraps_from_table": "sim sweep away paper scraps from the table",
    "turn_mug_right_side_up": "sim turn mug right side up",
    "load_plates_into_dish_rack": "sim load plates into tabletop dish rack",
    "hang_mug_on_mug_rack": "sim hang the mug on the mug rack"
  },
  "notes": [
    "sim_prompt_map holds the exact prompt each task trained under: the post-remap training task name with underscores as spaces. Prompt a task any other way and the policy is off-distribution with no error to tell you so.",
    "The put-bottles scene trained under the THROW prompt, which is why its entry does not match its scene name.",
    "The genuine throw-bottles sim scene was excluded from training as physically distinct from the real throw data; expect roughly 1 in 4 on eval_policy.py's native --task bottles. Evaluate put_plastic_bottles_in_bin instead.",
    "hang_mug_on_mug_rack has no success evaluator in abc_sim: it renders video but its success_rate is structurally 0.0 and is not a result.",
    "Read mean_max_progress for paper-comparable numbers and success_rate for the strict metric.",
    "step 200000 of a run configured for 500000 steps: a usable intermediate checkpoint, not a completed schedule.",
    "Model weights only - no optimizer or scheduler state. Valid for train.py --load-pretrained, not for --resume-from."
  ]
}
