SeanWang0027/qwen3-1.7b-textcraft-sft-qwen3-32b-traj
SeanWang0027/qwen3-1.7b-textcraft-sft-qwen3-32b-traj is a 1.7 billion parameter Qwen3 model fine-tuned using trajectory SFT on rollouts from a Qwen3-32B teacher model. This model is specifically optimized for multi-turn agentic tasks, demonstrating a 72.75% success rate on the TextCraft official test split. It is designed for applications requiring structured reasoning and action generation within a ReAct framework, leveraging a 32768 token context length.
Loading preview...
Model Overview
This model, SeanWang0027/qwen3-1.7b-textcraft-sft-qwen3-32b-traj, is a 1.7 billion parameter Qwen3-based student model. It was developed as part of a study comparing different training methodologies for multi-turn agents, specifically utilizing trajectory Supervised Fine-Tuning (SFT).
Key Capabilities
- Agentic Task Performance: Achieves a 72.75% success rate on the TextCraft official test split, significantly outperforming the base Qwen3-1.7B (23.00%) and approaching the performance of the larger Qwen3-32B teacher model (85.50%).
- Trajectory SFT: Trained using a unique trajectory SFT method on 2132 episodes of Qwen3-32B rollouts, focusing cross-entropy loss only on teacher replies and end-of-turn tokens.
- ReAct Prompting: Designed to operate with a ReAct prompt format, generating
Thought:andAction:sequences for multi-turn interactions. - Context Length: Supports a substantial context length of 32768 tokens, enabling complex multi-turn dialogues and task execution.
Good For
- Developing multi-turn conversational agents that require structured reasoning and action generation.
- Applications where a smaller, efficient model is needed to mimic the behavior of a larger, more capable teacher model in agentic tasks.
- Research into trajectory-based imitation learning and distillation for agentic LLMs.