| { |
| "id": "cosmos_world_model", |
| "display_name": "Cosmos3-Nano Future-Window World Model", |
| "status": "implemented", |
| "model_family": "Cosmos / physical-world foundation models", |
| "default_model_id": "nvidia/Cosmos3-Nano", |
| "local_model_env": "COSMOS_MODEL_DIR", |
| "dataset_contract": "xperience10m_future_window_world_model_v0", |
| "training_objective": "future_window_and_action_conditioned_world_modeling", |
| "split_policy": { |
| "unit": "episode", |
| "default_counts": { |
| "train": 96, |
| "val": 16, |
| "test": 16 |
| }, |
| "leakage_guard": "future windows must remain inside the same episode and test episodes must never condition training" |
| }, |
| "modalities": { |
| "direct_inputs": [ |
| "camera video streams or rendered mosaics", |
| "language task context" |
| ], |
| "conditioning_inputs": [ |
| "pose and SLAM trajectory", |
| "depth and confidence", |
| "mocap or action labels", |
| "IMU acceleration and gyro", |
| "audio event cues" |
| ], |
| "targets": [ |
| "future visual window", |
| "future latent state", |
| "future sensor-feature window", |
| "transition or contact event" |
| ], |
| "excluded_inputs": [ |
| "visualization.rrd" |
| ] |
| }, |
| "entrypoints": { |
| "selection_manifest": "scripts/omni/build_selection_episode_manifest.py", |
| "neutral_index": "scripts/omni/export_model_neutral_window_index.py", |
| "export": "scripts/omni/export_cosmos3_future_window_dataset.py", |
| "train": "scripts/omni/eval_cosmos3_future_window_retrieval.py", |
| "eval": "scripts/omni/eval_cosmos3_future_window_retrieval.py", |
| "launcher": "scripts/omni/run_cosmos3_nano_future_window_compat.sh", |
| "validate": "scripts/omni/validate_omni_finetune_run.py" |
| }, |
| "primary_metrics": [ |
| "future_retrieval_mrr", |
| "future_retrieval_recall_at_5", |
| "temporal_consistency", |
| "feature_reconstruction_error", |
| "transition_accuracy", |
| "contact_accuracy", |
| "held_out_episode_count" |
| ], |
| "artifact_contract": { |
| "checkpoint_gate": "world_model_checkpoint_and_generation_config", |
| "required_eval_files": [ |
| "metrics.json", |
| "future_predictions.jsonl", |
| "retrieval_rankings.csv", |
| "temporal_consistency.csv", |
| "qualitative_examples.json", |
| "RUN_REPORT.md" |
| ], |
| "required_training_files": [ |
| "training_metadata.json", |
| "progress.jsonl", |
| "model_config.json", |
| "checkpoint_manifest.json" |
| ], |
| "public_package_allowed": [ |
| "metrics", |
| "future-window prediction summaries", |
| "retrieval rankings", |
| "temporal consistency tables", |
| "qualitative example metadata", |
| "episode and dataset manifests", |
| "validation summaries" |
| ], |
| "public_package_forbidden": [ |
| "raw MP4", |
| "annotation HDF5", |
| "Rerun RRD", |
| "generated raw video unless explicitly licensed and size-bounded", |
| "base-model weights", |
| "full checkpoints", |
| "large archives" |
| ] |
| }, |
| "extension_requirements": [ |
| "Current implementation starts with Cosmos3-Nano compatibility over same-split future sensor-feature retrieval; it does not fine-tune Cosmos diffusion weights yet.", |
| "Install a Cosmos3 Diffusers training stack before replacing the compatibility adapter with LoRA or diffusion post-training.", |
| "Keep target windows inside the same episode and never train on held-out test episodes.", |
| "Record generated or retrieved qualitative examples separately from task-classification metrics." |
| ] |
| } |
|
|
