Qwen3-1.7B, ScienceWorld, trajectory SFT on gpt-5.4-mini episodes, epoch 4
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 the 3-epoch checkpoint (qwen3-1.7b-sciworld-sft-gpt54mini-3ep) for 2 more epochs with a fresh optimizer and a fresh cosine schedule (10% warm-up) over those 2 epochs; this is the end of the first of them.
Weights are fp32 (the training precision); load with torch_dtype=bfloat16 for inference.
Test success (avg@4): 15.12% (16.00 / 15.00 / 18.00 / 11.50). 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%.