README.md

---
license: cc-by-nc-4.0
dataset_info:
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: answer_variants
    list:
      list: string
  - name: evidence
    list:
    - name: document
      dtype: string
    - name: page
      dtype: int64
  - name: document_category
    dtype: string
  - name: domain
    dtype: string
  splits:
  - name: train
    num_bytes: 320371.0
    num_examples: 1550
  - name: dev
    num_bytes: 49432
    num_examples: 200
  - name: test
    num_bytes: 57522
    num_examples: 500
  download_size: 223929
  dataset_size: 427325.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: dev
    path: data/dev-*
  - split: test
    path: data/test-*
- config_name: documents
  data_files:
  - split: links
    path: data/documents/document_urls-*
task_categories:
- visual-document-retrieval
- visual-question-answering
language:
- en
tags:
- benchmark
- agent
- document
- multimodal
- RAG
size_categories:
- 1K<n<10K
extra_gated_prompt: >-
  ## License and Disclaimer

  The material in this repo is intended for non-commercial research purposes and is provided under the Creative Commons Attribution-NonCommercial (CC BY-NC) license.


  ## Data Availability and DocumentCloud Access Notice

  The benchmark described in this [research](https://arxiv.org/abs/2504.xxxxx) consists of a curated index of documents hosted on [DocumentCloud.org](http://DocumentCloud.org) which were uploaded to DocumentCloud by third-party organizations (e.g., news outlets, government agencies, non-profits, etc.).  Please refer to [DocumentCloud's Terms of Service](https://www.muckrock.com/tos/) for information regarding the rights of the underlying documents.

  1. Independent Data Source: Reasoning with Machines Lab at University of Oxford and its co-authors do not own, host, or control DocumentCloud.org or the third party documents referenced in the benchmark. The links provided point directly to the original public-facing documents on the DocumentCloud platform.

  2. Terms of Service & API Compliance: Users accessing these documents via the DocumentCloud API or web interface are responsible for complying with the [DocumentCloud Terms of Service](https://www.muckrock.com/tos/) and any specific usage restrictions or rate limits (e.g., those governing automated downloads) set by the platform and the original document uploader.

  3. For Non-Commercial Research Only: This index is provided solely for non-commercial, scientific benchmarking and research. Any commercial use of the documents found on DocumentCloud, including but not limited to the training of commercial Large Language Models (LLMs) or other proprietary products, may require separate licensing or permission from the original copyright holders.

  4. Liability Disclaimer: Reasoning with Machines Lab at University of Oxford and its co-authors provide this index "as-is" and without any warranty regarding the copyright status, data integrity, or accuracy of the linked documents. Users are solely liable for any copyright infringement or other legal claims arising from their downloading or use of the materials.
---

# MADQA Dataset

An agentic document question-answering benchmark with 2250 questions over a collection of 800 real-world PDF documents spanning multiple domains.

![MADQA](hero.jpg)

For more details, see the [paper](https://arxiv.org/abs/2603.12180), [leaderboard](https://huggingface.co/spaces/Snowflake/MADQA-Leaderboard), and [code](https://github.com/OxRML/MADQA).

## Splits
| Split | Questions | Description |
|-------|-----------|-------------|
| train | 1,550 | Training set |
| dev | 200 | Development/validation set |
| test | 500 | Held-out test set (answers hidden) |

## Schema
| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique identifier (`split/index`) |
| `question` | string | Natural language question |
| `answer_variants` | list[list[str]] | Acceptable answer variants (hidden for test) |
| `evidence` | list[{document, page}] | Gold evidence locations (hidden for test) |
| `document_category` | string | Document type (e.g., "Annual Report") |
| `domain` | string | High-level domain (e.g., "Financial") |

## Usage
```python
from datasets import load_dataset

ds = load_dataset("OxRML/MADQA")

# Access splits
train = ds["train"]
dev = ds["dev"]
test = ds["test"]

# Example
print(train[0]["question"])
```

## PDF Documents

PDF documents are hosted externally. The `"documents"` configuration provides a mapping from document filenames to their download URLs.

### Setup

```python
from datasets import load_dataset, DownloadManager

ds = load_dataset("OxRML/MADQA")
docs = load_dataset("OxRML/MADQA", "documents", split="links")
dm = DownloadManager()

# Build lookup: filename -> URL
doc_urls = {r["document"]: r["url"] for r in docs}
```

### Listing All Documents

```python
print(f"Found {len(docs)} PDF documents")
for row in docs:
    print(f"  {row['document']}")
```

### Downloading and Reading a PDF

```python
import fitz  # PyMuPDF

# Download a specific PDF (cached locally after first download)
pdf_path = dm.download(doc_urls["6414850.pdf"])

# Read with PyMuPDF
pdf = fitz.open(pdf_path)
print(f"Pages: {len(pdf)}")

# Extract text from first page
text = pdf[0].get_text()
print(text[:500])
```
## Citation

If you use **MADQA** in your research, please cite:

```bibtex
@misc{borchmann2026madqa,
  title        = {Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections},
  author       = {Łukasz Borchmann and Jordy Van Landeghem and Michał Turski and Shreyansh Padarha and Ryan Othniel Kearns and Adam Mahdi and Niels Rogge and Clémentine Fourrier and Siwei Han and Huaxiu Yao and Artemis Llabrés and Yiming Xu and Dimosthenis Karatzas and Hao Zhang and Anupam Datta},
  year         = {2026},
  eprint       = {2603.12180},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2603.12180}
}