groundtruth-hallucination-bench-sample
收藏资源简介:
Groundtruth Hallucination Bench 是一个用于评估 AI 模型事实陈述准确性的幻觉测试基准。本免费样本包含 100 个机器可运行的测试用例,从完整的 Groundtruth 数据集中采样(完整数据集包含 17 个经过验证的 ground-truth 数据集,共 918 个评估案例)。该数据集专门用于衡量模型是否陈述事实还是编造事实(幻觉)。每个测试用例的预期答案均通过独立从实时主要来源(如 openFDA、SEC EDGAR、USASpending、USPTO、ClinicalTrials.gov、Companies House REST API、Wikidata、NIST、Project Gutenberg 全文、固定解释器版本上的真实代码执行等)重新推导而来,而非爬取、生成或从模型记忆中回忆。每个样本都附带来源 URL。数据经过三重对抗审计:专门的怀疑论者针对实时来源尝试反驳每一行,发现并修复了约 55 个真实错误,每次修正均行内披露,甚至有一次审计捕获并回滚了错误的早期“修正”。数据集具有抗污染性:行基于日期的实时查询验证,每月添加新的训练后截止日期事实(例如当前模型无法从训练数据中得知的 2026 年 6 月 FDA 召回)。每个样本包含以下字段:question(问题)、expected_answer(预期答案)、grading_context(评分上下文)、grading_mode(评分模式:exact、numeric 或 judge)、difficulty(难度)、tags(标签)、source_urls(来源 URL)以及一个前沿模型的实际冷启动评分尝试作为基线(reference_model)。在完整数据集的 874 个已评分案例上,Claude Sonnet 5 的基线性能为:70.8% 正确、10.1% 幻觉、18.6% 回避。该数据集适用于问答系统的事实准确性评估、幻觉检测、事实验证等任务。本样本采用 CC BY 4.0 许可(署名“Groundtruth”)。
The Groundtruth Hallucination Bench is a benchmark for evaluating the factual accuracy of AI model statements. This free sample contains 100 machine-runnable test cases sampled from the full Groundtruth dataset (which includes 17 validated ground-truth datasets with 918 evaluation cases). The dataset is specifically designed to measure whether models state facts or fabricate information (hallucinate). Each test cases expected answer is independently derived from real-time primary sources (such as openFDA, SEC EDGAR, USASpending, USPTO, ClinicalTrials.gov, Companies House REST API, Wikidata, NIST, Project Gutenberg full text, real code execution on a fixed interpreter version, etc.), rather than crawled, generated, or recalled from model memory. Each sample includes source URLs. The data undergoes triple adversarial auditing: dedicated skeptics attempt to refute each line against live sources, discovering and fixing approximately 55 real errors, with each correction disclosed inline, and even one audit that caught and rolled back an erroneous early fix. The dataset is contamination-resistant: lines are validated by real-time queries based on dates, with new post-training cutoff facts added monthly (e.g., an FDA recall from June 2026 that current models cannot know from training data). Each sample contains fields: question, expected_answer, grading_context, grading_mode (exact, numeric, or judge), difficulty, tags, source_urls, and a baseline from a frontier models actual cold-start scoring attempt (reference_model). On the 874 scored cases of the full dataset, Claude Sonnet 5 baseline performance is: 70.8% correct, 10.1% hallucination, 18.6% abstention. The dataset is suitable for factual accuracy evaluation of QA systems, hallucination detection, fact verification, etc. This sample is licensed under CC BY 4.0 (attribution Groundtruth).
Groundtruth Hallucination Bench(免费样例)数据集概述
基本信息
- 数据集名称:Groundtruth Hallucination Bench (free sample)
- 许可证:CC BY 4.0
- 任务类别:问题回答(question-answering)
- 语言:英语
- 数据规模:少于1000条(免费样例含100条测试用例)
- 标签:幻觉检测、评估、基准测试、事实验证、真实数据
数据来源与构建方式
- 该样例取自 Groundtruth 目录,完整目录包含 17个经核实的真实数据集,共 918个评估案例,核心功能是衡量AI模型是陈述事实还是编造信息。
- 每条预期答案均为从实时主要来源独立重新推导(包括openFDA、SEC EDGAR、USASpending、USPTO、ClinicalTrials.gov、Companies House REST API、Wikidata、NIST、Project Gutenberg全文以及固定解释器版本上的真实代码执行),并非抓取、生成或从模型记忆中提取,且每条数据附带来源URL。
- 数据经过三重对抗性审计:专门的怀疑者针对实时来源逐条证伪,发现并修复了约55处真实错误,每项修正均在行内披露。
字段结构
每条数据为独立案例,包含以下字段:
question(问题)expected_answer(预期答案)grading_context(评分上下文)grading_mode(评分模式:精确 | 数值 | 裁判)difficulty(难度)tags(标签)source_urls(来源URL)reference_model(参考模型及其冷评分尝试)
基线表现
在全部874个已评分案例中(冷启动、无工具、无检索),Claude Sonnet 5 的表现:
- 70.8% 正确
- 10.1% 幻觉
- 18.6% 回避作答
- 各领域幻觉率从 0%(历史)到 32.6%(引文图谱问题)不等
运行方式
支持通过命令行工具运行评估:
- 支持 Anthropic API 及任意 OpenAI 兼容端点(如 Ollama)
- 支持通过自定义命令接入任意模型
完整数据集说明
- 当前样本约占完整数据的10%
- 完整目录包含17个数据集、918个案例及一个171家公司的已核实AI/ML生态数据集
- 提供一次性导出或月度定期复核的Living版本(含漂移变更日志及新的抗污染行)
- 基础事实源自公共领域/美国政府/CC0主要来源,每行均标注引用来源




