TIGER-Lab/SWE-QA-Pro-SFT-Trajectories
收藏资源简介:
SWE-QA-Pro SFT轨迹数据集是一个用于仓库级问答的代理工具使用轨迹集合,专门设计作为SWE-QA-Pro训练方案中的监督微调数据。该数据集包含1000个多轮交互轨迹,每个轨迹展示了代理如何通过只读工具(包括语义搜索、代码库查看和执行只读命令)探索代码库来回答与代码仓库相关的问题,而不是依赖预记忆的知识。这些轨迹由Claude Sonnet 4.5生成,并以hermes工具调用格式存储,适用于微调开源模型。数据集格式为JSONL,每个条目包含工具定义和多轮对话消息,涵盖系统指令、用户问题、代理规划、工具调用、工具响应以及最终答案。
SWE-QA-Pro SFT Trajectories is a set of agentic tool-use trajectories for repository-level question answering, used as the supervised fine-tuning data in the SWE-QA-Pro training recipe. It consists of 1,000 multi-turn trajectories where an agent answers repository-grounded questions by exploring the codebase with read-only tools (semantic_search, view_codebase, execute_readonly_command) rather than relying on memorized knowledge. The trajectories are generated by Claude Sonnet 4.5 and stored in the hermes tool-calling format, ready for fine-tuning open models. The dataset is in JSONL format with each object containing tool definitions and multi-turn conversation messages, including system instructions, user questions, agent planning, tool calls, tool responses, and final answers.




