Datasets:

hotchpotch
/

NanoViDoRe_v3_WIP_converted

---
configs:
- config_name: corpus
  data_files:
  - split: computer_science
    path: corpus/computer_science-*
  - split: energy
    path: corpus/energy-*
  - split: finance_en
    path: corpus/finance_en-*
  - split: finance_fr
    path: corpus/finance_fr-*
  - split: hr
    path: corpus/hr-*
  - split: industrial
    path: corpus/industrial-*
  - split: pharmaceuticals
    path: corpus/pharmaceuticals-*
  - split: physics
    path: corpus/physics-*
- config_name: queries
  data_files:
  - split: computer_science
    path: queries/computer_science-*
  - split: energy
    path: queries/energy-*
  - split: finance_en
    path: queries/finance_en-*
  - split: finance_fr
    path: queries/finance_fr-*
  - split: hr
    path: queries/hr-*
  - split: industrial
    path: queries/industrial-*
  - split: pharmaceuticals
    path: queries/pharmaceuticals-*
  - split: physics
    path: queries/physics-*
- config_name: qrels
  data_files:
  - split: computer_science
    path: qrels/computer_science-*
  - split: energy
    path: qrels/energy-*
  - split: finance_en
    path: qrels/finance_en-*
  - split: finance_fr
    path: qrels/finance_fr-*
  - split: hr
    path: qrels/hr-*
  - split: industrial
    path: qrels/industrial-*
  - split: pharmaceuticals
    path: qrels/pharmaceuticals-*
  - split: physics
    path: qrels/physics-*
license: other
license_name: vidore-v3-mixed
license_link: https://huggingface.co/collections/vidore/vidore-benchmark-v3
task_categories:
- visual-document-retrieval
- text-retrieval
language:
- en
- fr
- de
- es
- it
- pt
tags:
- nano
- nanobeir-format
- vidore
- wip
pretty_name: NanoViDoRe v3 (WIP, NanoBEIR-format conversion)
---

# NanoViDoRe v3 — WIP NanoBEIR-format conversion

> [!WARNING]
> **This is a temporary, work-in-progress dataset for experimentation.**
> The format is **not final** and may change without notice. This repository will be **moved, renamed, or deleted** in the future
> (for example, once a canonical location is agreed on). Please do not depend on it for anything long-lived.
> For the original dataset, use [`lightonai/NanoViDoRe_v3`](https://huggingface.co/datasets/lightonai/NanoViDoRe_v3).

## What this is

A re-packaging of [`lightonai/NanoViDoRe_v3`](https://huggingface.co/datasets/lightonai/NanoViDoRe_v3)
(revision `0ae74913373848e65871c072831ef30f710642ce`) into the same layout as the NanoBEIR datasets
(e.g. [`sentence-transformers/NanoBEIR-en`](https://huggingface.co/datasets/sentence-transformers/NanoBEIR-en)):
**one split per task**, each with its own corpus, queries and qrels.

It was created to explore whether a generic Nano-set evaluator in Sentence Transformers
(discussion: [huggingface/sentence-transformers#4327](https://github.com/huggingface/sentence-transformers/issues/4327))
can evaluate visual document retrieval without any dataset-specific loading code.

No content was added, removed or edited. Only the layout and column names changed:

- The original shares one 4,000-page corpus across all 8 tasks and separates tasks with a `source_dataset` column.
  Every original qrel points to a page of its own `source_dataset` (0 of 2,061 cross-task qrels), so each page was
  assigned to the split of its `source_dataset`. The total number of pages is unchanged (4,000).
- Columns were renamed to the NanoBEIR convention. IDs are strings.

## Structure

Configs: `corpus`, `queries`, `qrels`. Splits (identical across configs): `computer_science`, `energy`, `finance_en`,
`finance_fr`, `hr`, `industrial`, `pharmaceuticals`, `physics`.

| config | columns |
| --- | --- |
| `corpus` | `_id` (str), `text` (str, the OCR `markdown` of the page), `image` (page image) |
| `queries` | `_id` (str), `text` (str), `language` (str), `query_format` (str) |
| `qrels` | `query-id` (str), `corpus-id` (str), `score` (int, graded: 1 = critically relevant, 2 = fully relevant) |

| split | corpus | queries | qrels |
| --- | ---: | ---: | ---: |
| computer_science | 358 | 50 | 227 |
| energy | 434 | 50 | 200 |
| finance_en | 548 | 50 | 255 |
| finance_fr | 479 | 50 | 221 |
| hr | 365 | 50 | 283 |
| industrial | 894 | 50 | 320 |
| pharmaceuticals | 455 | 50 | 226 |
| physics | 467 | 50 | 329 |

Queries are spread evenly over 6 languages (en/fr/de/es/it/pt), as in the original.

## Usage

```python
from datasets import get_dataset_split_names, load_dataset

dataset_id = "hotchpotch/NanoViDoRe_v3_WIP_converted"
for split in get_dataset_split_names(dataset_id, "qrels"):
    corpus = load_dataset(dataset_id, "corpus", split=split)    # use "image" for visual retrieval, "text" for OCR text
    queries = load_dataset(dataset_id, "queries", split=split)
    qrels = load_dataset(dataset_id, "qrels", split=split)
```

## Scores are not comparable with the original protocol

The original README evaluates every query against the full merged 4,000-page corpus. Here each task only searches its
own pages, so pages from other tasks no longer act as extra distractors and scores are somewhat higher.

Example with [`vidore/colqwen2-v1.0`](https://huggingface.co/vidore/colqwen2-v1.0), mean nDCG@10 over the 8 tasks:

| setting | binary qrels | graded qrels |
| --- | ---: | ---: |
| original (merged 4,000-page corpus) | 0.4663 | 0.4685 |
| this conversion (per-task corpus) | 0.4876 | 0.4889 |

## License

Same terms as the source datasets. The query/qrels annotations of ViDoRe v3 are released under CC BY 4.0. The page images and
their `markdown` text inherit the license of their original publishers (public domain, CC BY 4.0, Etalab 2.0, CC BY-SA 3.0
and other open reuse licenses); see the `documents_metadata` config of each
[ViDoRe v3 dataset](https://huggingface.co/collections/vidore/vidore-benchmark-v3) for the per-document license.

## Citation

Please cite the original works:

```bibtex
@misc{loison2026vidorev3,
  title         = {ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios},
  author        = {António Loison and Quentin Macé and Antoine Edy and Victor Xing and Tom Balough and Gabriel Moreira and Bo Liu and Manuel Faysse and Céline Hudelot and Gautier Viaud},
  year          = {2026},
  eprint        = {2601.08620},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2601.08620},
}
```

The Nano subsampling is by LightOn: [`lightonai/NanoViDoRe_v3`](https://huggingface.co/datasets/lightonai/NanoViDoRe_v3).