README.md

---
license: other
license_name: mit-plus-citevqa-terms
license_link: https://github.com/OpenDataLab/CiteVQA
language:
  - en
  - zh
task_categories:
  - visual-question-answering
  - document-question-answering
tags:
  - evidence-attribution
  - visual-document-understanding
  - attribution-hallucination
  - document-grounding
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: verified_eval_set.jsonl
---

# Verified evaluation set for evidence attribution in visual documents

The 719-question evaluation set used in [**"Evidence Attribution in Visual Document
Understanding without Coordinates or Region Labels"**](https://arxiv.org/abs/2607.24651)
(Liu, Zhang, Xiao, 2026).

Code: [github.com/Ryenhails/quote-and-retrieve](https://github.com/Ryenhails/quote-and-retrieve) ·
Model: [Ryenhails/quote-and-retrieve-8b-grpo](https://huggingface.co/Ryenhails/quote-and-retrieve-8b-grpo)

## What this is

CiteVQA links to source PDFs that are no longer all reachable, and some of the reachable ones
differ from the version that was annotated. Evaluating evidence attribution on those questions
measures broken links rather than models. This release is the subset whose annotations we could
verify against the PDF actually available, together with the normalised ground-truth geometry
needed to score against it.

Starting from the 987 single-document questions in the CiteVQA validation release, the filter
removes 117 questions whose source PDF does not resolve or is not a valid PDF, 4 whose annotated
evidence pages lie outside the downloaded file, 131 whose annotated evidence text does not match
the extracted page text, and 16 that cannot be byte-verified, retaining **719 questions over 440
documents (72.9%)**. The filter inspects only ground-truth annotations and document content, never
model outputs, so it cannot favour any system. Documents dropped for text mismatch have text
layers at least as rich as the retained ones, so the filter does not select for easy cases.

| | |
|---|---|
| questions | 719 (386 English, 333 Chinese) |
| documents | 440 PDFs, median 34 pages, longest 182 |
| necessary evidence elements | 1,034 |
| all annotated elements | 1,763 |
| question types | Complex Synthesis 394, Multimodal Parsing 136, Factual Retrieval 122, Quantitative Reasoning 67 |

## What this contains, and what it does not

This release contains **identifiers, our verification metadata, and our normalised ground-truth
boxes**. It does **not** redistribute CiteVQA's question text, answers, or PDFs; download those
from the original release and join locally with the included script.

| field | meaning |
|---|---|
| `index` | question identifier, the join key to the CiteVQA release |
| `pdf_stem` | source PDF filename without extension |
| `language` | `en` or `zh` |
| `qtype` | question type, used for the per-type breakdowns in the paper |
| `n_pdf_pages` | document length in pages |
| `alignment` | our verification outcome: `verified` (366) when annotated evidence text was found in the extracted page text, `unverifiable` (353) when the page has no usable text layer to check against, for example a scanned page. Both are retained; questions whose text actively contradicted the annotation were removed. |
| `gt_pages` | 1-based pages holding necessary evidence |
| `gt_necessary` | necessary evidence elements as `[page, x1, y1, x2, y2]`; the recall denominator |
| `gt_all` | necessary plus optional supporting elements; the precision target set |

### Coordinate normalisation

The released CiteVQA evidence boxes are ordered `[y1, x1, y2, x2]` on a 0-1000 scale, which does
not match the `x1y1x2y2` description in the benchmark's own prompt. We established the true order
by extracting page text inside candidate boxes and checking it against the annotated evidence
content. The boxes here are already converted to `x1 y1 x2 y2` in rendered-page pixel
coordinates, with 1-based page indices. Using the raw released order will silently produce near
zero recall.

## Usage

```bash
pip install huggingface_hub
huggingface-cli download Ryenhails/quote-and-retrieve-eval --repo-type dataset --local-dir eval_set
```

Load the metadata directly:

```python
from datasets import load_dataset

ds = load_dataset("Ryenhails/quote-and-retrieve-eval", split="train")
print(ds[0]["index"], ds[0]["gt_necessary"])
```

To get records the pipeline can run on, join with your own CiteVQA download:

```bash
python rehydrate.py \
    --release verified_eval_set.jsonl \
    --citevqa /path/to/CiteVQA/data/validation/CiteVQA.json \
    --pdf-dir /path/to/CiteVQA/data/pdf \
    --out citevqa_singledoc_pub.jsonl
```

The output is byte-identical in schema to what `src/build_citevqa_pub.py` produces in the code
repository, so every downstream script runs unchanged. See
[docs/using_the_release.md](https://github.com/Ryenhails/quote-and-retrieve/blob/main/docs/using_the_release.md)
for the end-to-end path from this file to the paper's tables.

## Scoring

A citation matches an annotated element when both lie on the same page and their IoU is at least
0.5. Recall is computed per question over `gt_necessary` and macro-averaged over questions;
precision is scored against `gt_all`, so citing optional supporting evidence is not an error. The
exact semantics, including the two protocol choices that are ours rather than the benchmark's, are
in
[docs/scoring_protocol.md](https://github.com/Ryenhails/quote-and-retrieve/blob/main/docs/scoring_protocol.md)
and are asserted by the repository's test suite.

## Licensing and provenance

This release contains two kinds of material with different terms.

**Our contribution** is MIT licensed: the verification filter and its per-question outcome
(`alignment`), the document statistics, and the coordinate normalisation described above.

**The ground-truth geometry** (`gt_pages`, `gt_necessary`, `gt_all`) is derived from CiteVQA's
evidence annotations, reordered and rescaled but not otherwise altered. CiteVQA ships its code and
annotations under MIT, and additionally states that the benchmark is provided for academic research
and non-commercial use. We ask that you honour that restriction and cite CiteVQA alongside this
release.

**Nothing here redistributes document content.** CiteVQA's source PDFs were collected from
publicly accessible web resources and remain the property of their original publishers; CiteVQA
deliberately releases annotations and download links rather than the PDFs themselves, and this
release follows the same principle one step further by also omitting the question text and
answers. There are no page images, no rendered crops, and no PDF bytes in this repository. If you
need documents, obtain them through the CiteVQA release.

## Citation

```bibtex
@article{liu2026evidence,
  title         = {Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels},
  author        = {Liu, Zhuchenyang and Zhang, Yao and Xiao, Yu},
  journal       = {arXiv preprint arXiv:2607.24651},
  year          = {2026},
  eprint        = {2607.24651},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.24651}
}
```

Please also cite CiteVQA, whose annotations this release builds on.