README.md

---
language:
- en
task_categories:
- question-answering
pretty_name: SAFE-Verified-MultiHopQA
configs:
- config_name: 2wiki
  data_files:
  - split: train
    path: data/2wiki/train.parquet
  - split: validation
    path: data/2wiki/validation.parquet
- config_name: hotpotqa
  data_files:
  - split: train
    path: data/hotpotqa/train.parquet
  - split: validation
    path: data/hotpotqa/validation.parquet
- config_name: musique
  data_files:
  - split: train
    path: data/musique/train.parquet
  - split: validation
    path: data/musique/validation.parquet
---

# SAFE-Verified-MultiHopQA

[Paper](https://arxiv.org/abs/2604.01993) | [Project Page](https://github.com/DaeyongKwon98/SAFE)

## Overview

**SAFE-Verified-MultiHopQA** is released with the EMNLP 2026 paper **SAFE: An LLM-as-Verifier Framework for Evidence-Grounded Multi-Hop Reasoning**. It provides Knowledge Graph (KG) based, cleaned versions of three multi-hop question-answering benchmarks:

- [2WikiMultiHopQA](https://github.com/Alab-NII/2wikimultihop) (`2wiki`)
- [HotpotQA](https://hotpotqa.github.io/) (`hotpotqa`)
- [MuSiQue](https://github.com/StonyBrookNLP/musique) (`musique`)

SAFE uses KG-grounded verification to identify questions with invalid or ungrounded reasoning. This release removes the questions identified as wrong by SAFE from the original train and development (val) splits.

SAFE retains the original columns of the datasets and provides two standardized passage fields. `gt_passages` is a list of ground-truth passages, and `retrieved_passages` is a list of 10 retrieved passages that combines ground-truth passages with distractor passages in random order. Both fields are converted from the source dataset formats into a unified paragraph-level list representation.

## 🔍 Why verification is needed?

Through KG-based verification, we identified and verified the following errors in existing benchmarks, which account for up to 14% of instances in each split.

| Issue type | uestion | Answer | Passage (key evidence) |
|---|---|---|---|
| **Incorrect answer** | Was there any debate about the voting process in the state where Atwater Congregational Church is located? | a motion was made contesting Ohio's electoral votes | The passage mentions a motion about Ohio's electoral votes, but does not provide the required **yes/no** answer about a debate over the voting process. |
| **Ambiguous question** | Who is a cast member of the show in which Jana Brandner is a character? | Valerie Niehaus | The passages support multiple answers: both **Valerie Niehaus** and **Sven Koller** are cast members of *Verbotene Liebe*. |
| **Insufficient context** | What dynasty gave birth to the country in which the House of Luxembourg was located? | the Carolingian family | The passages associate the House of Luxembourg with Luxembourg but never identify the dynasty that founded Luxembourg. |

## Download and use
```python
from datasets import load_dataset

data = load_dataset(
    "Daeyongkwon98/SAFE-Verified-MultiHopQA",
    "2wiki", # "hotpotqa", "musique"
)

train = data["train"]
validation = data["validation"]
```

## Dataset splits and sizes
`train` is the filtered version of the original training split. `validation` is the filtered version of the source development (`dev`) split.

| Dataset | Split | Original rows | Filtered by SAFE | Final rows | Error (%) |
|---|---:|---:|---:|---:|---:|
| `2wiki` | train | 167,454 | 7,452 | 160,002 | 4.45% |
| `2wiki` | validation | 12,576 | 901 | 11,675 | 7.16% |
| `hotpotqa` | train | 90,447 | 2,377 | 88,070 | 2.63% |
| `hotpotqa` | validation | 7,345 | 227 | 7,118 | 3.09% |
| `musique` | train | 19,902 | 2,149 | 17,753 | 10.80% |
| `musique` | validation | 2,412 | 342 | 2,070 | 14.18% |
| **Total** | — | **300,136** | **13,448** | **286,688** | 4.48% |

## Example Data

```json
{
  "id": "13f5ad2c088c11ebbd6fac1f6bf848b6",
  "type": "bridge_comparison",
  "question": "Are director of film Move (1970 Film) and director of film Méditerranée (1963 Film) from the same country?",
  "answer": "no",
  "retrieved_passages": "[p1, p2, ..., p10]",
  "gt_passages": "[p2, p10]"
}
```

## Data Information

- 2WikiMultiHopQA
| Column | Available in | Description |
|---|---|---|
| `id` | Train, validation | Unique example identifier; renamed from the source `_id` in train. |
| `type` | Train, validation | Source question/reasoning category. |
| `question` | Train, validation | Multi-hop natural-language question. |
| `context` | Train, validation | Serialized source context. |
| `entity_ids` | Train | 2Wiki entity identifiers. |
| `supporting_facts` | Train, validation | Supporting facts needed to answer the question. |
| `evidences` | Train, validation | Knowledge-graph evidence triples. |
| `answer` | Train, validation | Reference answer. |
| `evidences_id` | Train | Knowledge-graph evidence triples represented with IDs. |
| `answer_id` | Train | Answer entity ID, when present. |
| `retrieved_passages` | Train, validation | Standardized paragraph-level list of ten retrieved passages, containing ground-truth passages and distractors. |
| `answer_list` | Validation | Answer aliases provided by the source validation data. |
| `gt_passages` | Train, validation | Standardized paragraph-level list of ground-truth passages. |

- HotpotQA
| Column | Available in | Description |
|---|---|---|
| `id` | Train, validation | Unique example identifier. |
| `question` | Train, validation | Multi-hop natural-language question. |
| `answer` | Train, validation | Reference answer. |
| `type` | Train, validation | Source question/reasoning category. |
| `level` | Train, validation | Source difficulty level. |
| `supporting_facts` | Train, validation | Supporting facts needed to answer the question. |
| `context` | Train, validation | Serialized source context. |
| `retrieved_passages` | Train, validation | Standardized paragraph-level list of ten retrieved passages, containing ground-truth passages and distractors. |
| `gt_passages` | Train, validation | Standardized paragraph-level list of ground-truth passages. |
| `answer_list` | Validation | Answer aliases provided by the source validation data. |

- MuSiQue
| Column | Available in | Description |
|---|---|---|
| `id` | Train, validation | Unique example identifier. |
| `paragraphs` | Train, validation | Serialized candidate paragraphs. |
| `question` | Train, validation | Multi-hop natural-language question. |
| `question_decomposition` | Train, validation | Decomposition into component reasoning questions. |
| `answer` | Train, validation | Reference answer. |
| `answer_aliases` | Train, validation | Acceptable answer aliases. |
| `answerable` | Train, validation | Source answerability indicator. |
| `retrieved_passages` | Train, validation | Standardized paragraph-level list of ten retrieved passages, containing ground-truth passages and distractors. |
| `gt_passages` | Train, validation | Standardized paragraph-level list of ground-truth passages. |
| `answer_list` | Validation | Answer aliases provided by the source validation data. |


## Citation

If you find our datasets useful, please cite our paper!

```bibtex
@article{kwon2026safe,
  title={SAFE: An LLM-as-Verifier Framework for Evidence-Grounded Multi-Hop Reasoning},
  author={Kwon, Daeyong and Yoon, Soyoung and Hwang, Seung-won},
  journal={arXiv preprint arXiv:2604.01993},
  year={2026}
}
```

If you have any questions, feel free to contact `daeyongkwon@snu.ac.kr`.