README.md

---
license: cc-by-nc-sa-4.0
language:
  - zh
  - en
task_categories:
  - image-to-text
  - object-detection
pretty_name: FinixDocBench
size_categories:
  - "100<n<1K"
tags:
  - document-parsing
  - ocr
  - financial-documents
  - layout-analysis
  - table-recognition
  - reading-order
  - camera-captured-documents
  - ultra-large-documents
  - markdown
  - chinese
configs:
  - config_name: default
    data_files:
      - split: test
        path:
          - metadata.jsonl
          - track*/images/*.png
---

# FinixDocBench

This repository contains a compliance-reviewed public subset of **FinixDocBench**, the financial-domain document parsing benchmark introduced in the technical report **"FinixDoc: Rethinking Financial Document Parsing Beyond Saturated Benchmarks"**.

The benchmark focuses on document parsing conditions that are common in real financial workflows but underrepresented in saturated clean-document benchmarks: digitally native insurance clauses, noisy camera-captured medical receipts, ultra-long pages, and very large dense tables. The expected model outputs are page-level Markdown and, where available, structured JSON layout annotations.

Project links:

- Project page: https://finix.alipay.com/
- Hugging Face dataset: https://huggingface.co/datasets/inclusionAI/FinixDocBench
- License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International

## Released Subset Contents

This release contains **742 page samples** from the broader FinixDocBench benchmark. Track 3 is split into two directories so that ultra-long pages and large-table pages can be evaluated separately.

| Track | Directory | Source type | Pages | Files per sample | Main task |
|---|---|---|---:|---|---|
| FinixDigital | `track1_finixdigital_242_insurance_terms/` | Digitally native insurance terms | 242 | image + Markdown + JSON | Markdown parsing and structured layout parsing |
| FinixPhoto | `track2_finixphoto_300/` | Mobile-captured medical receipts | 300 | image + Markdown + JSON | Robust Markdown parsing and structured layout parsing |
| FinixHuge-Long | `track3_finixhuge_100_long/` | Ultra-long financial or insurance pages | 100 | image + Markdown | Ultra-large page Markdown parsing |
| FinixHuge-Table | `track3_finixhuge_100_table/` | Large dense table pages | 100 | image + Markdown | Ultra-large table reconstruction |

The full FinixDocBench described in the technical report also includes a larger internal evaluation track, **FinixInner**, which is not included in this release because of privacy and compliance constraints. The FinixDigital package here is a 242-page insurance-terms subset of the broader FinixDigital track discussed in the report.

## Repository Structure

```text
FinixDocBench/
  README.md
  LICENSE.md
  CITATION.cff
  dataset_manifest.jsonl
  metadata.jsonl
  track1_finixdigital_242_insurance_terms/
    images/
    mds/
    jsons/
  track2_finixphoto_300/
    images/
    mds/
    jsons/
  track3_finixhuge_100_long/
    images/
    mds/
  track3_finixhuge_100_table/
    images/
    mds/
  FinixDocBench_Eval_for_Markdown/
    README.md
    requirements.txt
    run_eval.py
    finixdoc_md_eval/
```

Each sample is matched by file stem. For example, `abc123.png`, `abc123.md`, and `abc123.json` describe the same page when all three files are present.

The `dataset_manifest.jsonl` file provides one row per sample with relative paths, track metadata, image dimensions, and basic annotation counts. It is intended as a lightweight index for users who want to load the release programmatically.

## Tasks

This FinixDocBench release supports three complementary task settings.

### 1. Full-Page Markdown Parsing

Given a page image, a model should produce a complete page-level Markdown reconstruction. This task is available for all public tracks.

The Markdown ground truth preserves text order, headings, tables, and other page-level structure. HTML `<table>` blocks are used where table structure, merged cells, or dense financial layouts need to be represented more faithfully than plain Markdown tables.

### 2. Structured Layout Parsing

Given a page image, a model should produce structured page elements with category labels, bounding boxes, transcribed content, and reading order. This task is available for FinixDigital and FinixPhoto, which include `jsons/` annotations.

The public JSON files use pixel-space bounding boxes in the original image coordinate system. Each JSON file includes page metadata plus a `layout` list.

### 3. Ultra-Large Page Processability

FinixHuge-Long and FinixHuge-Table evaluate whether a system can return a syntactically valid, non-empty, page-level Markdown result for oversized documents. These pages stress page resolution, output length, table complexity, and reading-order preservation.

Because FinixHuge is Markdown-only in this release, it is best evaluated with Markdown metrics plus a success-rate style processability check.

## Annotation Schema

FinixDigital and FinixPhoto use a unified 10-class page-element schema:

```text
page-header
page-footer
title
section-header
text
table
figure
caption
footnote
other
```

Top-level JSON fields:

| Field | Description |
|---|---|
| `width` | Original page image width in pixels. |
| `height` | Original page image height in pixels. |
| `resized_width` | Width used by the annotation or preprocessing pipeline. |
| `resized_height` | Height used by the annotation or preprocessing pipeline. |
| `max_pixels` | Maximum pixel budget recorded by the preprocessing pipeline. |
| `min_pixels` | Minimum pixel budget recorded by the preprocessing pipeline. |
| `layout` | Ordered list of page elements. |

Each `layout` item contains:

| Field | Description |
|---|---|
| `category` | One of the 10 page-element labels. |
| `bbox` | Pixel-space bounding box `[x1, y1, x2, y2]` in the page image coordinate system. |
| `content` | Transcribed text, Markdown structural marker, or serialized table content. This field may be absent for some `figure` elements. |
| `order` | Reading-order index of the layout element. |

Example:

```json
{
  "width": 993,
  "height": 1404,
  "resized_width": 992,
  "resized_height": 1408,
  "max_pixels": 16777216,
  "min_pixels": 4096,
  "layout": [
    {
      "category": "section-header",
      "bbox": [82, 364, 223, 394],
      "content": "## 2.3 责任免除",
      "order": 5
    },
    {
      "category": "table",
      "bbox": [337, 156, 916, 295],
      "content": "<table>...</table>",
      "order": 3
    }
  ]
}
```

## Dataset Statistics

| Track | Images | Markdown files | JSON files | Notes |
|---|---:|---:|---:|---|
| FinixDigital | 242 | 242 | 242 | 6,223 structured layout elements; 214 tables |
| FinixPhoto | 300 | 300 | 300 | 8,517 structured layout elements; 224 tables |
| FinixHuge-Long | 100 | 100 | 0 | Ultra-long page images, up to 287M pixels |
| FinixHuge-Table | 100 | 100 | 0 | Large dense table images, up to 386M pixels |
| **Total** | **742** | **742** | **542** | All samples have paired images and Markdown |

Category counts for the structured JSON tracks:

| Category | Count |
|---|---:|
| `text` | 10,158 |
| `section-header` | 1,522 |
| `figure` | 1,307 |
| `title` | 504 |
| `table` | 438 |
| `caption` | 265 |
| `page-footer` | 223 |
| `footnote` | 204 |
| `page-header` | 64 |
| `other` | 55 |

## Loading Examples

Load the manifest with the Hugging Face `datasets` library:

```python
from datasets import load_dataset

manifest = load_dataset(
    "json",
    data_files="dataset_manifest.jsonl",
    split="train",
)

print(manifest[0])
```

Read a sample locally after cloning the repository:

```python
from pathlib import Path
from PIL import Image
import json

repo = Path("FinixDocBench")
row = manifest[0]

image = Image.open(repo / row["image_path"])
markdown = (repo / row["markdown_path"]).read_text(encoding="utf-8")

annotation = None
if row["json_path"] is not None:
    annotation = json.loads((repo / row["json_path"]).read_text(encoding="utf-8"))
```

## Evaluation

The repository includes a lightweight Markdown evaluator:

```bash
cd FinixDocBench_Eval_for_Markdown
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```

Run evaluation on a track by providing a ground-truth Markdown directory and a prediction Markdown directory with matching file names:

```bash
python run_eval.py \
  --gt_dir ../track2_finixphoto_300/mds \
  --pred_dir /path/to/predicted_mds \
  --output_json outputs/finixphoto_result.json
```

The Markdown evaluator reports:

| Metric | Direction | Description |
|---|---|---|
| `text_block_Edit_dist` | Lower is better | Normalized edit distance over matched text blocks. |
| `reading_order_Edit_dist` | Lower is better | Normalized edit distance over serialized reading-order sequences. |
| `table_TEDS` | Higher is better | Tree-edit-distance-based table similarity, scaled to 0-100. |
| `overall` | Higher is better | Composite score on a 0-100 scale. |

The overall score is:

```text
overall = ((1 - text_block_Edit_dist) * 100
         + (1 - reading_order_Edit_dist) * 100
         + table_TEDS) / 3
```

For FinixHuge, users should additionally report a success rate: the fraction of pages for which the system returns a syntactically valid, non-empty page-level Markdown result without runtime failure, severe truncation, or format errors that prevent downstream evaluation.

Structured JSON annotations are provided for FinixDigital and FinixPhoto. This repository currently ships the Markdown evaluator; if you report structured layout metrics, please describe the evaluator, matching rules, and coordinate convention used.

## Reference Results from the Technical Report

The following values are copied from the FinixDoc technical report for context. They correspond to the benchmark protocol reported in the paper and should not be treated as precomputed scores for every subset in this release unless the same split and evaluation protocol are reproduced.

### FinixDigital

| Model | Overall | TextEdit | TableTEDS | TableTEDS-S | ReadOrderEdit |
|---|---:|---:|---:|---:|---:|
| Qwen3-VL-4B | 80.18 | 0.145 | 76.04 | 81.23 | 0.210 |
| FinixDoc-VL | 93.19 | 0.039 | 92.07 | 93.67 | 0.086 |
| DeepSeek-OCR-2 | 82.80 | 0.139 | 90.00 | 92.32 | 0.277 |
| FireRed-OCR | 83.10 | 0.119 | 87.10 | 89.14 | 0.259 |
| PaddleOCR-VL-1.5 | 85.41 | 0.116 | 86.12 | 88.26 | 0.183 |
| GLM-OCR | 86.28 | 0.121 | 89.44 | 90.99 | 0.185 |
| Youtu-Parsing | 89.26 | 0.091 | 87.79 | 90.85 | 0.109 |
| Dots.OCR | 90.36 | 0.058 | 89.78 | 92.26 | 0.129 |
| MinerU 2.5 | 92.96 | 0.045 | 91.18 | 92.70 | 0.078 |
| Qwen3.5-397B-A17B | 84.90 | 0.119 | 87.30 | 89.51 | 0.207 |
| Kimi-K2.5 | 85.05 | 0.119 | 85.95 | 88.24 | 0.189 |
| Qwen3-VL-235B-A22B-Instruct | 87.26 | 0.076 | 82.77 | 85.44 | 0.134 |

### FinixPhoto

| Model | Overall | TextEdit | TableTEDS | TableTEDS-S | ReadOrderEdit |
|---|---:|---:|---:|---:|---:|
| Qwen3-VL-4B | 54.28 | 0.408 | 50.13 | 63.04 | 0.465 |
| FinixDoc-VL | 67.03 | 0.276 | 69.08 | 77.69 | 0.404 |
| PaddleOCR-VL-1.5 | 41.28 | 0.512 | 34.54 | 46.72 | 0.595 |
| MinerU 2.5 | 43.08 | 0.513 | 35.54 | 48.58 | 0.550 |
| DeepSeek-OCR-2 | 43.20 | 0.459 | 30.69 | 43.52 | 0.552 |
| GLM-OCR | 45.82 | 0.521 | 50.47 | 60.22 | 0.609 |
| FireRed-OCR | 47.20 | 0.487 | 38.50 | 53.72 | 0.482 |
| Dots.OCR | 52.57 | 0.399 | 44.90 | 56.92 | 0.473 |
| Youtu-Parsing | 60.90 | 0.345 | 59.01 | 66.23 | 0.418 |
| Qwen3.5-397B-A17B | 62.58 | 0.384 | 62.04 | 71.82 | 0.359 |
| Qwen3-VL-235B-A22B-Instruct | 62.65 | 0.359 | 63.55 | 72.13 | 0.397 |
| Kimi-K2.5 | 65.55 | 0.325 | 70.16 | 77.34 | 0.410 |

### FinixHuge

| Model | Success Rate | Overall | TextEdit | TableTEDS | TableTEDS-S | ReadOrderEdit |
|---|---:|---:|---:|---:|---:|---:|
| FinixDoc | 0.92 | 68.23 | 0.357 | 57.09 | 60.10 | 0.167 |
| Qwen3-VL-235B-A22B-Instruct | 0.68 | 34.85 | 0.847 | 47.05 | 63.20 | 0.578 |
| GLM-OCR | 0.34 | 38.06 | 0.816 | 59.39 | 62.43 | 0.636 |

## Intended Uses

This dataset is intended for:

- Evaluating OCR and document parsing systems on financial-domain documents.
- Testing full-page Markdown reconstruction.
- Testing layout parsing, table parsing, bounding boxes, and reading-order recovery on FinixDigital and FinixPhoto.
- Measuring robustness on noisy camera-captured receipt images.
- Evaluating end-to-end processability on ultra-large document pages.

## Out-of-Scope Uses

This dataset is not intended for:

- Individual profiling or personal information extraction.
- Automated financial, medical, insurance, legal, employment, credit, or similarly consequential decision-making.
- Reporting benchmark numbers after using benchmark labels or ground truth for training, fine-tuning, data augmentation, or prompt optimization.
- Claiming complete coverage of all financial document scenarios.

## Limitations

FinixDocBench is an evaluation benchmark, not a comprehensive training corpus. This release covers selected high-value financial document parsing scenarios and does not include the private FinixInner track.

FinixPhoto is derived from public-scenario medical receipt sources and re-annotated under the FinixDocBench schema. Prior exposure of some external models to the original public sources cannot be fully ruled out.

FinixHuge emphasizes system-level processability with Markdown-only public annotations. Direct single-pass model comparisons may understate or overstate practical usability if failed pages, truncation, or invalid outputs are not reported consistently.

Some page images may be very large. Users should use image loading libraries carefully and configure decompression or pixel limits intentionally when evaluating FinixHuge.

## License

This FinixDocBench release is distributed under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)**.

See [LICENSE.md](LICENSE.md) for the human-readable license notice and the official Creative Commons license link.

## Citation

If you use this FinixDocBench release, please cite:

```bibtex
@misc{wang2026finixdoc,
  title        = {FinixDoc: Rethinking Financial Document Parsing Beyond Saturated Benchmarks},
  author       = {Hang Wang and Jin Zhang and Guoliang Xu and Pengyue Lu and Yao Li and Zijiao Zhang and Tianyu Huang and Weiqi Xiong and Yulong Wang and Chuqiao Lu and Wenkang Huang and Kai Yang and Yadong Li and Hui Li and Xingzhong Xu and Xiao Xu},
  year         = {2026},
  institution  = {Ant Group},
  url          = {https://finix.alipay.com}
}
```

## Contact

For questions about the benchmark, please contact the FinixDoc authors through the project page or the Ant Group Hugging Face organization.