sokoban_adaptive_mixed_forward_rl_step130
BAGEL-7B-MoT Sokoban checkpoint after 130 IMP-agent RL steps.
- arm: mixed-forward
- initialization:
novastar111/sokoban_adaptive_mixed_forward_sft3k - training set: 2,000 mixed 3-box Sokoban boards (certified-deadlock and trivial strata)
- optimization: global batch 16, 8 rollouts per prompt, learning rate 2e-6
- reward: stop-required environment success with a -0.1 malformed-output penalty
- weights: converted BF16 EMA safetensors; optimizer/training state is not included
This repository is one point in the complete 10-step release ladder from RL10 through RL200.
Evaluation uses Sokoban-v8 q95/perseg, full model rollout, move-only actions, and stop-required
success. Load with the public BAGEL-7B-MoT base/config and use ema.safetensors as the model delta.