ABLATION (NOT the default setting): ViT FROZEN during training (--freeze_vit True, visual_und=True; freeze_vae=True, freeze_llm/und=False). Identical data/order/contract to the default sokoban easy plain SFT (sokoban_easy_v8_ordered_plain_matched_cot100k), ONLY freeze_vit differs. v8 res256, move_only, stop_then_solved. Compare against novastar115/sokoban_easy_plain_v8_res256_step2000 (default, ViT trained).
file: ema.safetensors (EMA weights; optimizer state NOT included — inference/eval/RL, not resume)
Load
This is an SFT delta on the BAGEL-7B-MoT base. Load with the base config/tokenizer/VAE/ViT from
BAGEL-7B-MoT and these EMA weights, e.g. the repo's eval runners:
--bagel-path <this repo> --bagel-base-path <BAGEL-7B-MoT>.
Image transforms: maze2d VAE 256/256 ViT 224/224 (q95); sokoban v8 VAE 512/256 ViT 336/224 (q95);
pusht VAE 256/256 ViT 224/224 (q90, 96px render). Prompt is stop-required; CoT is all-step interleaved.
