| license: cc-by-nc-4.0 | |
| tags: [bagel, vlm-gym, world-model, pusht, cot] | |
| # pusht_cot_chunk_k3_world_model | |
| BAGEL-7B-MoT world-model SFT checkpoint (EMA weights only, raw fp32). | |
| - task: **pusht** variant: **cot** step: **5000** | |
| - All-step interleaved chain-of-thought world model; re-grounds on the true frame every K=3 steps. Keep-last interleaving recipe used by the final evaluation. | |
| - weights: fp32 EMA, sharded into 2 safetensors files (`ema-XXXXX-of-00002.safetensors`) | |
| plus `ema.safetensors.index.json` (fp32 exceeds the 50GB single-file limit). Optimizer state NOT | |
| included — for inference / eval, not resume. | |
| ## Load | |
| This is an SFT delta on the public BAGEL-7B-MoT base. Load the base config / tokenizer / VAE / ViT | |
| from `BAGEL-7B-MoT` together with these EMA weights; point the loader at the repo folder (or its | |
| `ema.safetensors.index.json`) and it assembles the shards, e.g. with the eval runners: | |
| `--bagel-path <this repo> --bagel-base-path <BAGEL-7B-MoT>`. | |
| Image transforms (pusht): VAE 256/256, ViT 224/224, JPEG q90, 96px render (norm4 action space). Prompt is stop-required; CoT is all-step interleaved. | |