alireza7/GrepSeek-ColdStart-SFT-10k
收藏资源简介:
GrepSeek冷启动SFT数据集是一个用于直接语料交互(DCI)搜索代理的监督微调数据集合,旨在帮助代理通过Unix shell命令(如rg、grep、head等)直接对原始文本语料进行检索和推理,而不是从预计算的密集或稀疏索引中检索。该数据集包含10,000个多轮交互轨迹,平衡混合了5,000个来自Natural Questions(单跳问题)和5,000个来自HotpotQA(多跳问题)的示例。每个轨迹遵循GrepSeek交互协议:包括代理的推理、shell命令调用、工具响应和最终答案。数据通过两阶段管道生成:首先,Answer-Aware Tutor以反向方式从黄金答案构建验证的证据链;然后,Answer-Blind Planner将链转为时间顺序,并生成仅依赖已观察信息的推理。数据经过质量过滤,确保答案与黄金答案有非零token-F1分数且无未来状态泄漏。该数据集用于冷启动SFT阶段,以稳定代理的行为,为后续强化学习阶段奠定基础。
The GrepSeek Cold-Start SFT Dataset is a supervised fine-tuning dataset for Direct Corpus Interaction (DCI) search agents, which aims to enable agents to directly retrieve and reason over raw text corpora via Unix shell commands such as rg, grep, head, etc., rather than retrieving from pre-computed dense or sparse indexes. This dataset contains 10,000 multi-turn interaction trajectories, which are a balanced mixture of 5,000 examples from Natural Questions (single-hop questions) and 5,000 examples from HotpotQA (multi-hop questions). Each trajectory follows the GrepSeek interaction protocol, including the agent's reasoning, shell command invocation, tool responses, and final answer. The data is generated via a two-stage pipeline: first, the Answer-Aware Tutor constructs validated evidence chains from gold standard answers in a reverse manner; then, the Answer-Blind Planner converts these chains into chronological order and generates reasoning that only relies on observed information. The dataset has undergone quality filtering to ensure that the generated answers have a non-zero token-F1 score compared to the gold standard answers and no future state leakage. This dataset is used in the cold-start SFT phase to stabilize the agent's behavior and lay the foundation for the subsequent reinforcement learning phase.





