File size: 2,047 Bytes
5577fc6 | 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 | ---
license: cc-by-nc-4.0
tags: [world-model, sokoban, vlm-gym, bagel-7b-mot]
---
# sokoban_easy_cot_chunk_kinf_world_model_20260707_perseg (EMA, bf16)
Fine-tuned **BAGEL-7B-MoT** action-conditioned visual world model — step-5000 EMA weights,
**cast fp32 → bfloat16** (inference runs in bf16 autocast, so this is lossless for inference).
Inference / eval artifact only: **no optimizer state**, not a training-resume checkpoint.
- **Load into** the BAGEL-7B-MoT architecture (base: `ByteDance-Seed/BAGEL-7B-MoT`); this is the
EMA state dict (`ema.safetensors`).
- **Training data:** `ultrastar111/sokoban_easy_v8_cot_chunk_kinf_world_model_20260707_perseg` (see manifest).
- **Recipe:** lr 2e-5, cosine, warmup 300, 5000 steps, tokens 40k, EMA 0.993, cold-start from vanilla BAGEL.
- Eval contract: per-segment decoder (`BAGEL_DECODER=perseg`), CFG off, stop-required success.
- License: CC-BY-NC-4.0 (research use).
<details><summary>training_data_manifest.txt</summary>
```
STUDY: GT ENV-FEEDBACK CoT chunk world model (Structure A; re-ground on real frame every K)
chunk_size_K: inf
date_tag: 20260707_perseg
data_root: /data/home/raychai/hf_datasets/sokoban_easy_v8_cot_chunk_kinf_world_model_20260707_perseg/training
dataset_config_yaml: ./data/configs/vlm_gym_sokoban_easy_imagined_cot_envfb_v8_res256_train.yaml (EnvFB loader 'interleaved_cot_envfb', data-dir overridden)
format: user(rules+cadence,frame0 ctx); assistant(<think> a<img>...K... </think> {K actions}); user("Env Feedback:" + REAL frame, loss0 INPUT) between chunks
loss: CE on assistant text; MSE on per-step imagined frames; Env-Feedback frame + label = loss=0 INPUT; frame0/instruction = context
init_from: VANILLA /home/jiaxin/unified_world_model/pretrained/BAGEL-7B-MoT (EMA)
hparams: lr 2e-5, warmup 300, cosine, total 5000, tokens 40000, ema 0.993, save_every 2500 (== self-rollout study)
git_rev: 0c4fc0786a9c30bfcb8d442261717c0e8d46760c
```
</details>
|