qiangwhu/Qwen3-0.6B-ALFWorld-FullHistory-SFT
The qiangwhu/Qwen3-0.6B-ALFWorld-FullHistory-SFT model is a 0.8 billion parameter Qwen3-based checkpoint specifically fine-tuned for long-horizon ALFWorld agents. It is designed to generate brief reasoning and executable ALFWorld actions, leveraging a 32768-token context length. This model excels at producing step-level actions within the ALFWorld environment by processing full factual observation and action histories.
Loading preview...
Model Overview
This model, qiangwhu/Qwen3-0.6B-ALFWorld-FullHistory-SFT, is a specialized 0.8 billion parameter Qwen3 checkpoint. It is fine-tuned for use as an agent in the ALFWorld environment, focusing on long-horizon tasks. The model is designed to output concise reasoning followed by a single executable ALFWorld action, enclosed in <action>...</action> tags.
Training Details
The training process began with a LoRA SFT on successful DeepSeek-V4-Flash ALFWorld trajectories, converted into step-level samples that include the full factual Observation/Action history. Subsequently, the model underwent full-parameter fine-tuning using additional demonstrations for pick_cool_then_place_in_recep and pick_two_obj_and_place tasks, collected with a 50-step horizon. The input prompt incorporates the task, all prior Observation/Action records, the current observation, and admissible actions, while the target output contains current-step reasoning and one action.
Evaluation Performance
Evaluated on ALFWorld valid-seen tasks (140 tasks, 50-step horizon, 1024-token generation budget), the model achieved a mean success rate of 78.57% across all task types. Specific success rates include 94.29% for 'Pick', 76.92% for 'Look', 77.78% for 'Clean', 50.00% for 'Heat', 84.00% for 'Cool', and 70.83% for 'Pick2'.
Usage Considerations
Users should employ the Qwen3 chat template with thinking enabled and ensure the full factual ALFWorld Observation/Action history is retained in each step's prompt.