zhongyi-zhou/toolgrad-500
收藏资源简介:
ToolGrad-500是一个合成的工具使用数据集,专注于单轮并行函数调用场景。该数据集通过ToolGrad框架(ACL 2026 Findings)生成,利用文本梯度来优化工具选择和调用。数据集包含600个训练示例(其中500个正例和100个负例/不相关示例)和110个测试示例,用于评估在另一个LLM生成的数据上的工具准确性。每个对话都是单轮结构,由三个消息组成:系统提示(包含完整的API库文档和严格的格式指南)、用户查询和助手响应(以Python列表形式表示并行函数调用,例如[func1(arg=val), func2(arg=val)])。数据集的目的是训练模型在单轮对话中同时执行多个工具调用,以响应用户的复杂请求。数据集采用标准对话格式,适用于文本生成任务,并包含工具使用和函数调用相关标签。
ToolGrad-500 is a synthetic tool-use dataset targeting single-turn parallel function calling scenarios. The dataset is generated using the ToolGrad framework (ACL 2026 Findings), leveraging textual gradients to optimize tool selection and calling. It consists of 600 training examples (500 positive and 100 negative/irrelevance examples) and 110 test examples for evaluating tool accuracy on data generated from another LLM. Each conversation is a single turn with three messages: a system prompt containing complete API library documentation and strict formatting guidelines, a user query, and an assistant response that is a Python-braced list of composed tool invocations (e.g., [func1(arg=val), func2(arg=val)]). The dataset is designed to train models to execute multiple parallel tool calls in a single turn to address complex user requests. It follows a standard conversational format suitable for text-generation tasks and is tagged with tool-use, function-calling, and synthetic-data.




