Qwen3-1.7B, BabyAI, TCOD-F2B from Qwen3-32B, data pass 2 of 3
TCOD forward-to-backward (the episode window grows, checkpoint_steps 6). Code: kokolerk/TCOD + FutureBridge-OPD overlay (trinity-rft), BabyAI port as in docs/TCOD_BABYAI.md of the online-rose repo (branch
tcod-babyai): student Qwen3-1.7B, teacher Qwen3-32B bf16, thinking off, the conversation of babyai/eval_babyai.py, 20 turns, the 810 official
train tasks, batch 16 episodes / 64 turns, lr 1e-6, kl_coef 1.0, three passes over the data = 152 explorer steps. This is the HF export after
explorer step 101 (trainer step 106). Not evaluated.