README.md

metadata

license: cc-by-nc-4.0
tags:
  - bagel
  - vlm-gym
  - sokoban
  - reinforcement-learning
  - adaptive-thinking

sokoban_adaptive_mixed_forward_rl_step50

BAGEL-7B-MoT Sokoban checkpoint after 50 IMP-agent RL steps.

• arm: mixed-forward
• initialization: novastar111/sokoban_adaptive_mixed_forward_sft3k
• RL data: 2,000 mixed 3-box boards (certified deadlock + trivial)
• objective: environment success with a -0.1 malformed-output penalty
• weights: converted BF16 EMA safetensors; optimizer/training state is not included

The mixed-forward and mixed-branch labels refer to the step-3000 SFT initializations used by the RL experiment. Evaluation uses Sokoban-v8 q95/perseg, full self-rollout, move-only actions, and stop-required success. Load with the public BAGEL-7B-MoT base/config and point the eval runner's --bagel-path at this repository.