README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: type
    dtype: string
  - name: evidences
    sequence:
      sequence: string
  - name: supporting_facts
    struct:
    - name: title
      sequence: string
    - name: sent_id
      sequence: int64
  - name: context
    struct:
    - name: title
      sequence: string
    - name: sentences
      sequence:
        sequence: string
  splits:
  - name: train
    num_bytes: 664062413
    num_examples: 167454
  - name: validation
    num_bytes: 54492966
    num_examples: 12576
  - name: test
    num_bytes: 51538723
    num_examples: 12576
  download_size: 388043174
  dataset_size: 770094102
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---



# 2WikiMultihopQA

**This repository only repackages the original 2WikiMultihopQA data so that every example follows the field layout used by [HotpotQA](https://hotpotqa.github.io/).** The content of the underlying questions, answers and contexts is **unaltered**.
All intellectual credit for creating 2WikiMultihopQA belongs to the authors of the paper *Constructing a Multi‑hop QA Dataset for Comprehensive Evaluation of Reasoning Steps* (COLING 2020) and the accompanying code/data in their GitHub project [https://github.com/Alab-NII/2wikimultihop](https://github.com/Alab-NII/2wikimultihop).

## Dataset Summary

* **Name:** 2WikiMultihopQA
* **What’s different:** only the JSON schema. Each instance now has `id`, `question`, `answer`, `type`, `evidences`, `supporting_facts`, and `context` keys arranged exactly like HotpotQA so that existing HotpotQA data pipelines work out‑of‑the‑box.

The dataset still contains **multi‑hop question‑answer pairs** with supporting evidence chains drawn from Wikipedia. There are three splits:

| split      | #examples |
| ---------- | --------: |
| train      |   113,284 |
| validation |    12,981 |
| test       |    12,995 |

## Dataset Structure

Each JSON Lines file contains records like:

```json
{
  "id": "13f5ad2c088c11ebbd6fac1f6bf848b6",
  "question": "Are director of film Move (1970 Film) and director of film Méditerranée (1963 Film) from the same country?",
  "answer": "no",
  "type": "bridge_comparison",
  "level": "unknown",
  "supporting_facts": {
    "title": ["Move (1970 film)", "Méditerranée (1963 film)", "Stuart Rosenberg", "Jean-Daniel Pollet"],
    "sent_id": [0, 0, 0, 0]
  },
  "context": {
    "title": ["Stuart Rosenberg", "Méditerranée (1963 film)", "Move (1970 film)", ...],
    "sentences": [["Stuart Rosenberg (August 11, 1927 – March 15, 2007) was an American film and television director ..."],
                   ["Méditerranée is a 1963 French experimental film directed by Jean-Daniel Pollet ..."],
                   ...]
  }
}
```

### Field definitions

| Field                      | Type                 | Description                                                              |
| -------------------------- | -------------------- | ------------------------------------------------------------------------ |
| `id`                       | `string`             | Unique identifier (original `_id`).                                      |
| `question`                 | `string`             | Natural‑language question.                                               |
| `answer`                   | `string`             | Short answer span (may be "yes"/"no" for binary questions).              |
| `type`                     | `string`             | Original 2Wiki question type (e.g. `bridge_comparison`, `comparison`).   |
| `evidences`                | `List[List[string]]` | Structured data (subject, property, object) obtained from Wikidata.      |
| `supporting_facts.title`   | `List[string]`       | Wikipedia page titles that contain evidence sentences.                   |
| `supporting_facts.sent_id` | `List[int]`          | Zero‑based sentence indices within each page that support the answer.    |
| `context.title`            | `List[string]`       | Titles for every paragraph provided to the model.                        |
| `context.sentences`        | `List[List[string]]` | Tokenised sentences for each corresponding title.                        |

## Data Splits

The conversion keeps the same train/validation/test division as the original dataset. No documents or examples were removed or added.

## Source Data

* **Original repository:** [https://github.com/Alab-NII/2wikimultihop](https://github.com/Alab-NII/2wikimultihop)
  Contains data generation scripts, the Apache‑2.0 license and citation information.
* **Paper:** Ho et al., *Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps*, COLING 2020.
  [\[ACL Anthology\]](https://aclanthology.org/2020.coling-main.580/) – [\[arXiv\]](https://arxiv.org/abs/2011.01060)

No new text was crawled; every paragraph is already present in the original dataset.

## License

The original 2WikiMultihopQA is released under the **Apache License 2.0**.
This redistribution keeps the same license. See [`LICENSE`](./LICENSE) copied verbatim from the upstream repo.

## How to Use

```python
from datasets import load_dataset

ds = load_dataset("framolfese/2WikiMultihopQA")
print(ds["train"][0]["question"])
```