marrita/toolace-tool-calling-hallucination-ragtruth

该数据集是为课程作业“工具调用中的幻觉检测”创建的。它通过从ToolACE风格的工具调用对话开始，自动注入三种必需的幻觉类型：工具矛盾（答案与工具输出矛盾）、过度生成（答案添加了工具输出中不支持的额外信息）和缺失工具（答案建议需要不可用工具的操作）。每个示例遵循类似RAGTruth的格式，包含查询、上下文、输出、幻觉标签（带有字符偏移的跨度级标签）等字段，以及工具、工具调用、源ID、分割和腐败类型等附加字段用于可重复性和分析。数据集包含训练、验证和测试分割，通过源ID分隔以避免同一原始对话的干净和腐败变体之间的泄漏。它支持句子级二元幻觉检测和跨度级幻觉定位。数据集是自动腐败的，可能包含腐败过程的伪影，适用于作为该项目的受控基准，但真实世界的泛化性需单独评估。

This dataset was created for the course assignment Hallucination Detection in Tool Calling. It is synthetic by design: starting from ToolACE-style tool-calling dialogues, it automatically injects three required hallucination types: tool_contradiction (the answer contradicts the tool output), overgeneration (the answer adds unsupported information not present in the tool output), and missing_tool (the answer suggests an action requiring a tool that is not available). Each example follows a RAGTruth-like format, including fields such as query, context, output, hallucination_labels (span-level labels with character offsets), as well as additional fields like tools, tool_call, source_id, split, and corruption_type for reproducibility and analysis. The dataset contains train, validation, and test splits, separated by source_id to avoid leakage between clean and corrupted variants of the same original dialogue. It supports both sentence-level binary hallucination detection and span-level hallucination localization. The dataset is automatically corrupted and may contain artifacts of the corruption procedure, making it suitable as a controlled benchmark for this project, but real-world generalization should be evaluated separately.