Qwen3-1.7B · TextCraft · trajectory SFT on Qwen3-32B rollouts — 1 epoch
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: 6 Qwen3-32B trajectories per training task (2132 episodes), 1 epoch = 66 steps. Stored in bf16.
Evaluation
TextCraft official test split (100 tasks), avg of 4 passes, temperature 0.4, 512 tokens/turn, 30 turns, thinking off. Success 72.75%, malformed turns 0.00%. Reference points on the same protocol: base Qwen3-1.7B 23.00%, Qwen3-32B itself 85.50%.
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.