CoachWorld / release.json
JEdward's picture
Upload CoachWorld pretrain release
6de848e verified
Raw History Blame Contribute Delete
2.2 kB
{
"schema_version": 1,
"kind": "sa_wm_pretrained_model_release",
"model_id": "sa_wm_pretrained_v1",
"status": "selected_for_release_and_post_training",
"selected_at": "2026-07-25",
"selected_checkpoint": {
"training_stage": "sa_wm_pretrain_v6_512x768",
"micro_step": 55000,
"path": "/mnt/cfs/wmcprg/RoboCoach/checkpoints/sa_wm_pretrain_v6_512x768/checkpoint-55000",
"training_config": "configs/training/sa_wm_pretrain_v6.yaml",
"weights": "dit_model.safetensors",
"checkpoint_format": "safetensors+optim_shards",
"post_training_load_mode": "model_only_init_from"
},
"selection_evidence": {
"monitoring_selection": "/mnt/cfs/wmcprg/RoboCoach/checkpoints/sa_wm_pretrain_v6_512x768/validation/selection_closed_loop.json",
"samples": 30,
"rollout_chunks": 4,
"future_frames": 1440,
"metrics": {
"psnr": 21.503524390835977,
"ssim": 0.9163767135968223,
"lpips_alex": 0.07815770965373506
},
"selection_reason": [
"Human review preferred the motion and physical behavior of checkpoint-55000.",
"DROID exterior_1 and exterior_2 PSNR, SSIM, and LPIPS all outperform checkpoint-70000.",
"The fourth closed-loop chunk is slightly better than checkpoint-70000 in PSNR and LPIPS.",
"The aggregate metric difference from checkpoint-70000 is too small to override the qualitative review."
]
},
"historical_comparison": {
"checkpoint": "/mnt/cfs/wmcprg/RoboCoach/checkpoints/sa_wm_pretrain_v6_512x768/checkpoint-70000",
"micro_step": 70000,
"status": "retained_late_training_reference_not_selected",
"metrics": {
"psnr": 21.552308871443916,
"ssim": 0.9175402801300698,
"lpips_alex": 0.07684179975799957
}
},
"usage_contract": {
"inference": "load selected_checkpoint/dit_model.safetensors",
"post_training": "initialize model weights from the selected checkpoint and start a new optimizer, scheduler, step counter, output directory, and run id",
"exact_v6_resume": "allowed only for reproducing the historical v6 stage; do not use it for domain post-training",
"immutability": "never overwrite the selected checkpoint directory"
}
}