bharatrag-benchmark
收藏资源简介:
该数据集是一个用于问答幻觉检测研究的多语言多领域数据集,采用MIT许可证。数据集包含80个训练样本,每个样本由唯一ID、语言标识、领域分类、问题文本、上下文文本、真实答案和幻觉答案七个字段构成。其中幻觉答案字段专门标注了与上下文信息不符的模型生成答案,使其适用于问答系统幻觉检测、模型可靠性评估等任务。数据集结构清晰,体积紧凑,便于快速实验和基准测试。
This dataset is a multilingual, multi-domain dataset for research on question-answering hallucination detection, licensed under the MIT License. It contains 80 training samples, each consisting of seven fields: unique ID, language identifier, domain classification, question text, context text, true answer, and hallucinated answer. The hallucinated answer field specifically annotates model-generated answers that are inconsistent with the context information, making it suitable for tasks such as hallucination detection in question-answering systems and model reliability evaluation. The dataset has a clear structure and compact size, facilitating rapid experiments and benchmark testing.
数据集概述:BharatRAG Benchmark
- 数据集名称:BharatRAG Benchmark
- 许可证:MIT
- 数据集大小:数据集总大小为 83,267 字节(约 81 KB),下载大小为 42,024 字节(约 41 KB)
- 数据集划分:仅包含训练集(train),共 80 个样本
数据特征
数据集包含 7 个字段,类型均为字符串(string):
- id:样本唯一标识符
- language:语言
- domain:领域
- question:问题
- context:上下文
- ground_truth_answer:真实答案
- hallucinated_answer:幻觉答案
数据用途推测
该数据集旨在用于评估或训练 RAG(检索增强生成)模型在检测或避免幻觉方面的能力,通过提供问题、上下文、真实答案与可能包含幻觉的答案进行对比分析。




