andthattoo/blimp-textworld-blimp-q8

blimp-textworld-blimp-q8

BLiMP 5-step block-memory RL on TextWorld q8.

This is a full-parameter RL fine-tuned checkpoint, not a LoRA adapter.

Base model: Qwen/Qwen3-1.7B

Final held-out TextWorld q8 eval, 32 episodes:

• —untrained Qwen3-1.7B: success 0.375, mean steps 36.59
• —standard full-history RL: success 0.375, mean steps 35.375
• —BLiMP block-memory RL: success 0.53125, mean steps 33.25
• —BLiMP + ECHO/score: success 0.5, mean steps 33.71875

GitHub repo: https://github.com/andthattoo/blimp