searchsim/agentsim-atc-multihop

AgentSim代理跟踪语料库—多跳版本是一个专门为多跳问答设计的数据集，旨在支持学生模型蒸馏和DPO偏好微调。该数据集包含1,490个接受的SFT轨迹和2,980个步骤级别的DPO偏好对，这些数据是通过一个7动作的代理执行器在HotpotQA、2WikiMultiHopQA、MuSiQue、FRAMES和Bamboogle等多跳QA数据集上生成的。数据集提供了详细的模式描述，包括决策类别、理性标签、必须接地标志和使用的参数知识等。此外，数据集还包含了步骤级别的字符跨度引用和偏好数据，适合用于学生模型蒸馏、DPO偏好微调和Always-Search策略学习。

The AgentSim Agent-Trace Corpus — Multi-hop is a dataset designed for multi-hop question answering, specifically aimed at supporting student model distillation and DPO preference fine-tuning. It includes 1,490 accepted SFT trajectories and 2,980 step-level DPO preference pairs, generated over five multi-hop QA datasets (HotpotQA, 2WikiMultiHopQA, MuSiQue, FRAMES, and Bamboogle) through a 7-action agentic executor. The dataset provides detailed schema descriptions, including decision categories, rationale tags, must-ground flags, and parametric knowledge used. Additionally, it features step-level char-span citations and preference data, making it suitable for student model distillation, DPO preference fine-tuning, and Always-Search Policy training.