This is a bf16 Qwen3-0.6B checkpoint for long-horizon ALFWorld agents. It is trained to produce brief reasoning and one executable ALFWorld action in <action>...</action> tags.

The model starts from Qwen/Qwen3-0.6B. First, LoRA SFT is performed on successful DeepSeek-V4-Flash ALFWorld trajectories converted to step-level samples with full factual Observation/Action history. The resulting checkpoint is then full-parameter fine-tuned on additional successful pick_cool_then_place_in_recep and pick_two_obj_and_place demonstrations collected with a 50-step horizon.

The user prompt contains the task, all prior Observation/Action records, the current observation, and admissible actions. The assistant target contains current-step reasoning and exactly one action. Historical assistant reasoning is excluded from the prompt.

ALFWorld valid-seen evaluation uses 140 tasks, one deterministic rollout per task, a 50-step horizon, full 50-step factual history, and a 1024-token generation budget. Mean success rate over two runs:

Use the Qwen3 chat template with thinking enabled and retain the full factual ALFWorld Observation/Action history in each step prompt.