Qwen3-1.7B · TextCraft · SFT through verl's MultiTurnSFTDataset — KNOWN-BAD, kept for reference

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

Same data as qwen3-1.7b-textcraft-sft-qwen3-32b-traj, but trained through verl's MultiTurnSFTDataset with a Qwen3 chat template, which chops the <|im_start|>user header off most observations and puts the assistant header and empty think block inside the loss. 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 48.75% with 74.42% of turns malformed (the model loops </think>). The correctly trained model scores 72.75%.

Notes

Do not use this model as a baseline; it documents a data-pipeline bug, not a method.

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.