Datasets:
Update README.md
Browse files# More Documents, Same Length Datasets
This repository contains the datasets used in the paper ["More Documents, Same Length: Isolating the Challenge of Multiple Documents in RAG"](https://arxiv.org/abs/2503.04388).
## Dataset Description
These datasets investigate how document multiplicity affects LLM performance in Retrieval-Augmented Generation (RAG) systems while keeping the total context length fixed. The benchmarks isolate the effect of having more documents versus having fewer but longer documents.
The datasets are based on multi-hop question answering tasks that require synthesizing information from multiple documents. Each question in the dataset comes with:
1. **Key documents**: Contains essential information needed to answer the question correctly. These are preserved across all dataset variants.
2. **Distractor documents**: Contains related but non-essential information. The number of these varies across dataset variants.
## Methodology
These datasets were created using the following approach:
1. Starting with Wikipedia-derived multi-hop questions that require information from multiple sources
2. Providing the same key documents (containing essential information) across all dataset variants
3. Adjusting document lengths by adding passages from Wikipedia pages of the documents we kept, while maintaining the same total token count
4. Varying the number of distractor documents to create different multiplicity conditions
## Links
- [Paper on arXiv](https://arxiv.org/abs/2503.04388)
- [GitHub Repository](https://github.com/shaharl6000/MoreDocumentsSameLength)
- [Hugging Face Dataset](https://huggingface.co/datasets/shaharl6000/MoreDocumentsSameLength)
- [MusiQue Original Dataset](https://github.com/deepmind/deepmind-research/tree/master/musique)