Qwen3-1.7B · ALFWorld · trajectory SFT on gpt-5.4-mini rollouts — epoch 3 of 3

Student: Qwen/Qwen3-1.7B. Part of a study comparing teacher-trajectory SFT, online multi-turn ROSE and on-policy distillation for multi-turn agents.

Training

Trajectory SFT: one row per full teacher episode in ChatML, cross-entropy only on each teacher reply and its <|im_end|> (observations, instruction and assistant header are context). Batch 32, AdamW lr 1e-5, weight decay 0, cosine schedule with 10% warmup, fp32 training. Nothing filtered: failed episodes are included. Data: 2420 AgentGym training games played once by gpt-5.4-mini (OpenAI API, reasoning_effort=medium; only the visible reply is used), at most 30 turns, teacher success 58.97%, 49431 supervised turns. One 3-epoch run (225 steps, cosine over all 225), max length 18432; this is step 225, fp32.

Evaluation

ALFWorld test split (AgentGym, 200 games), avg of 4 passes, temperature 0.4, 512 tokens/turn, 30 rounds, thinking off. Success 58.25% ± 0.43. Epochs 1/2/3 of this run: 55.12 / 55.38 / 58.25. Same-teacher online ROSE 10+5 after 3 epochs: 40.25% (SFT − ROSE = +18.00 points, 95% CI [12.75, 23.75]). Reference points on the same protocol: base Qwen3-1.7B 19.38%, Qwen3-32B itself 52.12%.

Notes

Epochs 1–2 are mid-schedule checkpoints (learning rate still high); only epoch 3 saw the full decay. The teacher writes a Thought in only ~13% of its turns and so does this model.

Prompt format

ReAct, multi-turn chat: the environment's AgentGym instruction as the first user turn, a canned assistant acknowledgement, then one user turn per observation; the model replies Thought:\n...\n\nAction:\n<one action>. Render with Qwen3's chat template and enable_thinking=False.