searchsim/agentsim-atc

AgentSim Agent-Trace Corpus (ATC)是一个基于检索增强的问答代理的推理轨迹数据集，包含103,567个推理步骤，覆盖了Quasar-T、CausalQA和MSMARCO三个IR基准测试。数据集提供了20,548个监督查询-文档-答案三元组用于微调，以及199,968个独特的检索文档。每个推理步骤都可以追溯到源语料库中的特定文档，提供了步骤级可审计性。数据集适用于行为分析、链式思维微调、模仿学习等多种用途，但需遵守上游数据集的许可限制。

The AgentSim Agent-Trace Corpus (ATC) is a dataset of grounded reasoning traces for retrieval-augmented question-answering agents, containing 103,567 reasoning steps spanning three IR benchmarks (Quasar-T, CausalQA, and MSMARCO). It includes 20,548 supervised query-document-answer triples for fine-tuning and 199,968 unique retrieved documents. Each reasoning step is traceable to specific documents in the source corpus, enabling step-level auditability. The dataset is suitable for various uses such as behavioral analysis, chain-of-thought fine-tuning, and imitation learning, but users must adhere to the licensing restrictions of the upstream source datasets.