File size: 3,166 Bytes
eeac43c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | {
"id": "cosmos3_super_reasoner",
"display_name": "Cosmos3-Super Reasoner",
"status": "implemented",
"model_family": "Cosmos3 / physical-world foundation models",
"default_model_id": "nv-community/Cosmos3-Super",
"local_model_env": "COSMOS3_SUPER_MODEL_DIR",
"dataset_contract": "xperience10m_episode_json_qa_v1",
"training_objective": "zero_shot_structured_episode_understanding_json_qa_via_vllm_reasoner",
"split_policy": {
"unit": "episode",
"default_counts": {
"train": 96,
"val": 16,
"test": 16
},
"leakage_guard": "uses the same 96/16/16 selected episode split as the Qwen3-Omni LoRA branch; no Super weights are updated"
},
"modalities": {
"direct_inputs": [
"multi-camera rendered mosaic video",
"language prompt and label options"
],
"conditioning_inputs": [
"prompt-side task schema and episode/window metadata"
],
"targets": [
"structured action/subtask/contact/transition/object JSON"
],
"excluded_inputs": [
"visualization.rrd",
"raw annotation HDF5",
"audio in the current vLLM Reasoner path"
]
},
"entrypoints": {
"selection_manifest": "scripts/omni/build_selection_episode_manifest.py",
"export": "scripts/omni/parallel_export_qwen3_omni_action_dataset.py",
"neutral_index": "scripts/omni/export_model_neutral_window_index.py",
"train": "",
"eval": "scripts/omni/eval_cosmos3_super_reasoner.py",
"launcher": "scripts/omni/run_cosmos3_super_reasoner_eval.sh",
"validate": "scripts/omni/validate_omni_finetune_run.py"
},
"primary_metrics": [
"json_validity_rate",
"action_macro_f1",
"subtask_accuracy",
"transition_accuracy",
"next_action_accuracy",
"contact_accuracy",
"object_micro_f1",
"held_out_episode_count"
],
"artifact_contract": {
"checkpoint_gate": "base_weight_vllm_reasoner_setup_metadata",
"required_eval_files": [
"metrics.json",
"predictions.jsonl",
"predictions.csv",
"per_class_metrics.csv",
"confusion_matrix.csv",
"server_info.json",
"RUN_REPORT.md"
],
"required_training_files": [
"training_metadata.json",
"progress.jsonl"
],
"public_package_allowed": [
"metrics",
"predictions",
"confusion matrices",
"run reports",
"server/model setup metadata",
"episode and dataset manifests",
"validation summaries"
],
"public_package_forbidden": [
"raw MP4",
"annotation HDF5",
"Rerun RRD",
"base-model weights",
"fine-tuned weights",
"checkpoints",
"large archives"
]
},
"extension_requirements": [
"This branch evaluates staged Cosmos3-Super Reasoner base weights through vLLM on the 128-episode held-out JSON task; it does not fine-tune or release new Cosmos weights.",
"Create a separate Cosmos3-Super adapter/model repository only after a real fine-tuning run produces new adapter or checkpoint weights.",
"Keep it separate from the Cosmos3-Nano future-window compatibility branch, which answers a different world-model retrieval target."
]
}
|