---
license: mit
task_categories:
- image-retrieval
- question-answering
language:
- en
pretty_name: MADQA Training
size_categories:
- 1K<n<10K
---

# Chrisyichuan/madqa-training

MADQA document QA contrastive training data with hard negatives.

## Contents

- `madqa_converted.jsonl` — query-image pairs with hard negatives
- `images/` — all referenced images

Each metadata row:

```json
{
  "query": "...",
  "chunk_path": "images/...",
  "neg_chunk_paths": ["images/...", "images/..."],
  "source_positive_rank": 0,
  "source_positive_score": 0.0,
  "source_dataset": "moca"
}
```

## Summary

- rows: 1840
- unique images: 3598
- avg negatives/row: 2.00

## Download

```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Chrisyichuan/madqa-training", repo_type="dataset", local_dir="data/madqa-training")
```


## Image Storage

Images are stored as **4 tar shards** under `image_shards/` for fast download.

After cloning/downloading, extract images:

```bash
python extract_hf_image_shards.py --dataset-dir .
```

This creates `images/` with all referenced image files.