LIT-RAGBench

LIT-RAGBench is a benchmark for evaluating generator capabilities in Retrieval-Augmented Generation (RAG). It focuses on whether a model can answer questions correctly given retrieved documents, independent of retrieval quality. The benchmark covers five categories: Integration, Reasoning, Logic, Table, and Abstention.

Dataset Summary

LIT-RAGBench contains:

• 114 human-constructed Japanese questions
• An English version generated by machine translation with human curation

Each example includes the following fields:

• question
• answer: reference answer
• qa_type: question type
• positive_chunk_list: relevant evidence chunks
• negative_chunk_list: irrelevant chunks
• reasoning_content: reasoning process for deriving the answer

Intended Use

This dataset is intended for:

• benchmarking LLMs used as RAG generators
• analyzing model strengths and weaknesses by category
• developing and evaluating RAG-specialized models :contentReference[oaicite:6]{index=6}

Repository

Code, prompts, and the original dataset files are available in the GitHub repository: Koki-Itai/LIT-RAGBench

Citation

@misc{litragbench,
      title={LIT-RAGBench: Benchmarking Generator Capabilities of Large Language Models in Retrieval-Augmented Generation},
      author={Koki Itai and Shunichi Hasegawa and Yuta Yamamoto and Gouki Minegishi and Masaki Otsuki},
      year={2026},
      eprint={2603.06198},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.06198},
}

License

The code and data are released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license. Commercial use is permitted provided that you follow the license terms, including attribution and sharing any derivative works under the same.