GaRAGe
收藏资源简介:
GaRAGe是一个大型的RAG基准测试数据集,包含由人类编辑的长篇答案和每个 grounding 段落的注释,这使得可以细致地评估LLM在生成RAG答案时能否识别出相关的 grounding。该数据集包含2366个具有多样复杂性、动态性和主题的问题,并包括从私人文档集和Web上检索的超过35K个注释段落,以反映现实世界的RAG使用案例。这使得它成为一个理想的测试平台,用于评估LLM识别仅必要的相关信息来编写答案的能力,或者在没有足够信息时提供回避性回答的能力。
GaRAGe is a large-scale RAG benchmark dataset containing human-edited long-form answers and annotations for each grounding passage, which enables fine-grained evaluation of LLMs' ability to identify relevant grounding content when generating RAG-based answers. This dataset encompasses 2,366 questions with diverse complexities, dynamics and topics, and includes over 35,000 annotated passages retrieved from private document collections and the Web to reflect real-world RAG usage scenarios. This makes it an ideal testbed for evaluating LLMs' capabilities to identify only the necessary relevant information to compose answers, or to provide evasive responses when sufficient information is unavailable.
GaRAGe数据集概述
数据集简介
- 名称: GaRAGe (A Benchmark with Grounding Annotations for RAG Evaluation)
- 论文: GaRAGe: A Benchmark with Grounding Annotations for RAG Evaluation (ACL 2025 Findings)
- 目的: 评估LLMs在生成RAG答案时识别相关基础信息的能力
- 规模: 包含2366个问题,超过35K标注段落
数据特点
- 问题多样性: 涵盖不同复杂度、动态性和主题
- 数据来源: 从私有文档集和网页检索的段落
- 标注内容: 每个基础段落的详细标注
数据字段说明
- 标识信息: sample_id, question_date
- 问题属性:
- question, question_valid, question_false_premise
- question_seeking, question_sensitive, question_type
- question_complexity, question_category, question_popularity
- 基础段落: grounding (包含文本、引用标记、段落年龄、日期和提供者)
- 证据评估:
- evidence_relevant, evidence_correct
- evidence_cited
- 答案信息:
- answer_generate, answer_related_info
- answer_validate
- 其他信息: comments, question_tag, topic_tag
使用信息
- 数据格式: JSONL (GaRAGe_benchmark.jsonl)
- 许可证: CC-BY-NC-4.0
- 联系方式: Ionut-Teodor Sorodoc




