---
pretty_name: RAGCompass
license: other
license_name: mixed
language:
  - en
task_categories:
  - question-answering
  - text-retrieval
tags:
  - retrieval-augmented-generation
  - retriever-routing
  - multi-hop-qa
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: "Llama-3.1-8B-Instruct_Qwen3-Embedding-0.6B/*/*.jsonl"
---

# RAGCompass data

Per-retriever results and passage corpora for the paper [*RAGCompass: Generalizable Retriever Routing via Reranking*](https://arxiv.org/abs/2611.05596). Code: [github.com/Graph-COM/RAGCompass](https://github.com/Graph-COM/RAGCompass).

For every query, each of five retrievers returned its top-5 passages, and a fixed reader (`meta-llama/Llama-3.1-8B-Instruct`) answered from them. The F1 of that answer is the utility a retriever router is trained to predict. The data covers 4 datasets x 3 splits x 5 retrievers x 5,000 queries, plus the passage corpus of every split, so routing methods can be trained and evaluated without rerunning retrieval or the reader.

## Download

From the root of a clone of the code repository (`*.json*` fetches the data and leaves out this card):

```bash
hf download Graph-COM/RAGCompass --repo-type dataset --local-dir . --include "*.json*"
```

## Layout

```
Llama-3.1-8B-Instruct_Qwen3-Embedding-0.6B/
  <dataset>_train/<retriever>.jsonl
  <dataset>_val/<retriever>.jsonl
  <dataset>/<retriever>.jsonl              # test split
dataset/
  <dataset>[_train|_val]_corpus.json       # list of {"title", "text"}
```

`<dataset>` is one of `nq`, `musique`, `hotpotqa`, `2wikimultihopqa`. `<retriever>` is one of `rag_1_step` (dense, Qwen3-Embedding-0.6B), `bm25_1_step`, `splade_v3_1_step`, `colbertv2_1_step` or `hipporag2-0.5_1_step` (HippoRAG 2). Every JSONL file has 5,000 rows, one per query, and the five files of a split hold the same queries.

| field | meaning |
|---|---|
| `id`, `question`, `answers` | query id, text, gold answers |
| `top_k_indices`, `top_k_scores` | the retriever's top-5 passages (indices into the split's corpus) and their scores |
| `pred_ans`, `raw_response` | the reader's extracted answer and full response |
| `f1` | F1 of `pred_ans` against `answers` |
| `recall`, `ndcg`, `gold_doc_list` | retrieval quality against the gold passages |
| `latency` | seconds for retrieval plus generation |

Corpus sizes (passages):

| dataset | train | val | test |
|---|---|---|---|
| nq | 19,384 | 19,510 | 19,453 |
| musique | 45,240 | 45,195 | 48,236 |
| hotpotqa | 46,077 | 46,074 | 46,353 |
| 2wikimultihopqa | 25,217 | 24,993 | 26,424 |

The questions and corpora follow the [HippoRAG 2 release](https://huggingface.co/datasets/osunlp/HippoRAG_2).

## License

The data is derived from Natural Questions (CC BY-SA 3.0), HotpotQA (CC BY-SA 4.0), MuSiQue (CC BY 4.0) and 2WikiMultihopQA (Apache 2.0), with passages taken from Wikipedia (CC BY-SA). Each part is redistributed under the license of the dataset it comes from. The reader outputs were generated with Llama 3.1 and are subject to the [Llama 3.1 Community License](https://www.llama.com/llama3_1/license/).

## Citation

```bibtex
@misc{six2026ragcompass,
  title         = {RAGCompass: Generalizable Retriever Routing via Reranking},
  author        = {Six, Valentin and Li, Mufei and Fu, Dongqi and Xia, Yinglong and Li, Hong and Li, Pan},
  year          = {2026},
  eprint        = {2611.05596},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2611.05596}
}
```