jcrrag / README.md
YutaHitomi's picture
Update README.md
60e424f verified
metadata
license: cc-by-sa-4.0
task_categories:
  - question-answering
language:
  - ja
tags:
  - jcrrag
  - japaneserag
  - llmrag
  - rageval
  - rag-evaluation
pretty_name: JCrRAG
size_categories:
  - 10K<n<100K

JCrRAG : Japanese Contextual relevance RAG Benchmark

A human-annotated benchmark for evaluating Japanese Retrieval-Augmented Generation (RAG) systems, featuring multi-level complexity and diverse categories.

This benchmark does not address context selection; instead, with the context fixed, it evaluates the ability to appropriately handle the information within that context.

The benchmark contains 20,000 data records.

Each record has the following format :

(Context, Question, GroundtruthAnswer)

in which Context is the context to input to an LLM for RAG evaluation.

Evaluation script : https://github.com/alt-develop/jcrrageval

JCrRAG ベンチマーク

JCrRAGは、日本語のRetrieval-Augmented Generation(RAG)システムを評価するためのベンチマークで、複数レベルの難易度設定と多様なカテゴリを特徴としています。 本ベンチマークはコンテキストの選択を対象とせず、コンテキストを固定した上で、その中の情報を適切に扱う能力を評価するものです。

(コンテキスト, 質問, 正解回答) の3つ組データの2万件を含みます。

LLMに入力する際は、以下のプロンプトをご利用ください。

あなたはバーチャルアシスタントであり、提供された1つ以上の段落の情報に基づいて質問に回答する役割があります。以下の条件に従って質問に回答してください:
1) 回答は正確で完全でなければなりません。
2) 提供された段落の情報のみを使用してください。
3) 段落に回答が含まれていない場合、適切な説明をしてください。
質問: {Question}
段落: {Context}

自動評価スクリプト: https://github.com/alt-develop/jcrrageval

Copyright 2025 by alt Inc.


license: cc-by-sa-4.0