---
pretty_name: VQA train clean (utility copy)
license: other
language: [en, fr, de, es, it]
task_categories: [visual-document-retrieval]
size_categories: [100K<n<1M]
tags: [utility, deduplicated, decontaminated]
configs:
- config_name: ArxivQA
  data_files:
  - split: train
    path: ArxivQA/*.parquet
- config_name: ChartQA
  data_files:
  - split: train
    path: ChartQA/*.parquet
- config_name: DocVQA
  data_files:
  - split: train
    path: DocVQA/*.parquet
- config_name: InfoVQA
  data_files:
  - split: train
    path: InfoVQA/*.parquet
- config_name: MLPapers
  data_files:
  - split: train
    path: MLPapers/*.parquet
- config_name: Manuals
  data_files:
  - split: train
    path: Manuals/*.parquet
- config_name: PlotQA
  data_files:
  - split: train
    path: PlotQA/*.parquet
- config_name: SlideVQA
  data_files:
  - split: train
    path: SlideVQA/*.parquet
- config_name: TatDQA
  data_files:
  - split: train
    path: TatDQA/*.parquet
- config_name: Textbooks
  data_files:
  - split: train
    path: Textbooks/*.parquet
- config_name: VDR-de
  data_files:
  - split: train
    path: VDR-de/*.parquet
- config_name: VDR-en
  data_files:
  - split: train
    path: VDR-en/*.parquet
- config_name: VDR-es
  data_files:
  - split: train
    path: VDR-es/*.parquet
- config_name: VDR-fr
  data_files:
  - split: train
    path: VDR-fr/*.parquet
- config_name: VDR-it
  data_files:
  - split: train
    path: VDR-it/*.parquet
- config_name: WebPDF
  data_files:
  - split: train
    path: WebPDF/*.parquet
---

# vqa_train_clean

> **Utility dataset, not a contribution.** This is a cleaned and re-organised copy of existing public training
> sets, made for our own document-retrieval trainings. No new documents, queries or annotations were created.
> All credit goes to the original authors: please cite and use the original datasets.

737,626 (query, page image) pairs from [vidore/colpali_train_set](https://huggingface.co/datasets/vidore/colpali_train_set) (train split),
[openbmb/VisRAG-Ret-Train-In-domain-data](https://huggingface.co/datasets/openbmb/VisRAG-Ret-Train-In-domain-data),
[openbmb/VisRAG-Ret-Train-Synthetic-data](https://huggingface.co/datasets/openbmb/VisRAG-Ret-Train-Synthetic-data) and
[Omundu-Riad/vdr_multilingue_cleaned](https://huggingface.co/datasets/Omundu-Riad/vdr_multilingue_cleaned)
(itself from [llamaindex/vdr-multilingual-train](https://huggingface.co/datasets/llamaindex/vdr-multilingual-train)),
regrouped into one config per domain instead of one per source dataset.

## What was changed

- **Benchmark decontamination**: 1,064 pairs removed, i.e. every pair whose page is a page (or a re-encoded copy of
  a page) of the ViDoRe v1, v2 or v3 (public tasks) corpora. For example, 251 of the 500 ViDoRe v1 ArxivQA test pages
  and 74 of the 271 TatDQA test pages are present in the original training sets.
- **Deduplication**: 22,294 pairs removed (same page + same query), mostly pairs present in both ColPali and VisRAG.
- **`page_id`**: shared by identical and near-identical pages (byte hash, or correlation of the 512 px high-pass
  filtered pages >= 0.95), so that two queries of the same page are never in-batch negatives of each other.
- Everything else is unchanged: original image bytes, original queries (including ColPali's "answer briefly"
  suffixes) and original source labels. Details in `cleaning_report.json`.

## Configs

| config | pairs | pages | sources (dataset / original label) |
|---|---|---|---|
| ArxivQA | 31,564 | 26,075 | colpali_train_set / arxiv_qa, VisRAG-Ret-Train-In-domain-data / ArxivQA |
| ChartQA | 4,223 | 2,934 | VisRAG-Ret-Train-In-domain-data / ChartQA |
| DocVQA | 39,379 | 10,175 | colpali_train_set / docvqa, VisRAG-Ret-Train-In-domain-data / MP-DocVQA |
| InfoVQA | 20,349 | 4,255 | colpali_train_set / Infographic-VQA, VisRAG-Ret-Train-In-domain-data / InfoVQA |
| MLPapers | 59,694 | 9,995 | VisRAG-Ret-Train-Synthetic-data / ICML Papers, VisRAG-Ret-Train-Synthetic-data / NeurIPS Papers |
| Manuals | 119,898 | 19,584 | VisRAG-Ret-Train-Synthetic-data / Manuallib |
| PlotQA | 56,192 | 46,806 | VisRAG-Ret-Train-In-domain-data / PlotQA |
| SlideVQA | 7,931 | 4,911 | VisRAG-Ret-Train-In-domain-data / SlideVQA |
| TatDQA | 12,611 | 1,816 | colpali_train_set / tatdqa |
| Textbooks | 59,524 | 9,905 | VisRAG-Ret-Train-Synthetic-data / Textbooks |
| VDR-de | 58,208 | 57,999 | vdr_multilingue_cleaned / de |
| VDR-en | 53,487 | 53,175 | vdr_multilingue_cleaned / en |
| VDR-es | 58,724 | 58,544 | vdr_multilingue_cleaned / es |
| VDR-fr | 55,227 | 55,005 | vdr_multilingue_cleaned / fr |
| VDR-it | 54,934 | 54,695 | vdr_multilingue_cleaned / it |
| WebPDF | 45,681 | 15,438 | colpali_train_set / pdf |

## Columns

`query`, `image`, `domain`, `origin_dataset` (`colpali`, `visrag_in`, `visrag_syn`, `vdr`), `origin_label`,
`origin_file` and `origin_row` (row in the original parquet file), `page_id`.

```python
from datasets import load_dataset
ds = load_dataset("Omundu-Riad/vqa_train_clean", "DocVQA", split="train")
```

## License

Each part keeps the license and terms of its source dataset (see the original dataset cards). ColPali: academic
datasets under their original license, synthetic part without usage restrictions; VDR: Apache-2.0 (llamaindex).