TU_Pipeline
收藏资源简介:
TU-Pipeline是一个用于在可执行模拟环境中训练和评估大型语言模型(LLM)智能体的合成多轮工具使用数据集。该数据集包含4,088个经过去重处理的任务记录和2,174个独立的环境源代码文件。数据来源于三个内部合成源(data、data2、data3)的合并,并经过过滤和全局去重处理。每个数据记录包含13个核心字段,如任务ID、用户问题、任务类别、环境ID等,其中多个结构化字段以JSON字符串形式存储以保持跨环境模式稳定性。数据集旨在支持工具使用智能体、多轮智能体训练、轨迹回放与验证以及基于状态的评估等研究。需要注意的是,数据为合成数据,受限于源环境的假设和覆盖范围,任务分布受内部策略影响,且环境文件名在发布版本中通过打包方式避免冲突。
TU-Pipeline is a synthetic multi-turn tool-use dataset for training and evaluating Large Language Model (LLM) agents within executable simulation environments. This dataset contains 4,088 deduplicated task records and 2,174 independent environment source code files. It is compiled from three internal synthetic sources (data, data2, data3), and has undergone filtering and global deduplication processing. Each data record includes 13 core fields, such as task ID, user query, task category, environment ID, etc. Multiple structured fields are stored as JSON strings to ensure cross-environment schema consistency. This dataset is designed to support research on tool-use agents, multi-turn agent training, trajectory playback and verification, and state-based evaluation. It should be noted that this is synthetic data, which is limited by the assumptions and coverage of the source environments. The task distribution is influenced by internal policies, and environment file names are packaged in the released version to avoid naming conflicts.




