Qwen3-1.7B · ALFWorld · trajectory SFT on the first 15 teacher turns only (Qwen3-32B rollouts)

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. Ablation: every Qwen3-32B training episode is truncated after the teacher's 15th turn. 1 epoch = 75 steps. Stored in bf16.

Evaluation

ALFWorld test split (AgentGym, 200 games), avg of 4 passes, temperature 0.4, 512 tokens/turn, 30 rounds, thinking off. Success 47.75% on 2 passes only (the evaluation was stopped early), against 48.75% for full-episode SFT. Reference points on the same protocol: base Qwen3-1.7B 19.38%, Qwen3-32B itself 52.12%.

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.