okashurin/ToolHal-v1
收藏资源简介:
该数据集是原始ToolAce数据集的增强版本,专门用于工具调用任务中的幻觉检测。它通过注入三种幻觉类型来扩展原始数据:1. 模型响应与工具输出之间的冲突(tool_output_conflict子集),即模型响应与工具实际输出不一致;2. 过度生成(overgeneration子集),即模型响应包含工具输出中未出现的信息;3. 缺失工具(missing_tool子集),即模型响应建议调用不可用工具的操作。数据集包含多个特征字段,如查询、上下文、工具调用、输出等,并分为三个子集,总计约1774个示例,旨在帮助检测和评估模型在工具调用中的幻觉问题。
The repository contains augmentations of the ToolAce dataset for Hallucination Detection in the Tool Calling task. The original ToolAce dataset was injected with 3 types of hallucinations: 1. Hallucination between the models response and the tools output (tool_output_conflict split); 2. Overgeneration, when the response contains information not present in the tools output (overgeneration split); 3. Missing tool, when the response recommends actions requiring calling a tool that is not available (missing_tool split). The dataset includes features such as query, context, tools, tool_call, output, and hallucination labels, with splits for each hallucination type totaling approximately 1774 examples, designed to detect and evaluate hallucinations in model tool calls.




