Qwen3-1.7B, ScienceWorld, trajectory SFT on gpt-5.4-mini episodes, epoch 7
Trajectory SFT: one row per full teacher episode in ChatML, cross-entropy only on each teacher reply and its <|im_end|>. Data: gpt-5.4-mini (reasoning medium) episodes on the 2059 ScienceWorld train tasks (no test overlap), one per task, nothing filtered. Batch 32, AdamW lr 1e-5, weight decay 0, fp32.
Continued from epoch 5 for 3 more epochs (fresh optimizer, fresh cosine over epochs 6-8; one GPU, global batch 32); end of the second.
Weights are fp32 (the training precision); load with torch_dtype=bfloat16 for inference.
Test success (avg@4): not evaluated. ScienceWorld test split (AgentGym, 200 task variations), avg of 4 passes, temperature 0.4, 512 tokens/turn, 30 rounds, thinking off, no system turn. Reference points on the same protocol: base Qwen3-1.7B 0.12%; the same SFT at epoch 2 13.25%, epoch 3 12.88%; online ROSE 10+5 with the same teacher, epoch 5, 25.75%.