README.md

---
license: cc0-1.0
task_categories:
- image-to-text
- object-detection
language:
- en
tags:
- document-parsing
- ocr
- document-ai
- layout-analysis
- reading-order
- table-recognition
- formula-recognition
- benchmark
- vlm
pretty_name: "Dr.DocBench: Expert-Level and Difficult Document Parsing"
size_categories:
- 1K<n<10K
arxiv: 2606.01393
---

<h1 align="center">Dr.DocBench</h1>

<p align="center"><b>A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing</b></p>

<p align="center"><a href="https://arxiv.org/abs/2606.01393">Paper</a> · <a href="https://github.com/2077AI/DrDocBench">GitHub</a> · <a href="https://www.2077ai.com/drdocbench/">Homepage</a></p>

<p align="center"><a href="#dataset-statistics"><img alt="dev" src="https://img.shields.io/badge/dev-986%20pages-1f6feb"></a> <a href="#dataset-statistics"><img alt="test" src="https://img.shields.io/badge/test-509%20pages-8250df"></a> <a href="#dataset-statistics"><img alt="annotations" src="https://img.shields.io/badge/dev%20annotations-12%2C902-0969da"></a> <a href="https://creativecommons.org/publicdomain/zero/1.0/"><img alt="license" src="https://img.shields.io/badge/license-CC0--1.0-3fb950"></a></p>

## Dataset Summary

**Dr.DocBench** is a difficulty-aware benchmark for evaluating whether vision-language models and document processing systems can parse expert-level, structurally complex documents.

This release ships two splits:

| Split | Pages | Documents | Subjects | Ground truth |
| --- | ---: | ---: | ---: | --- |
| **`dev/`** | 986 | 66 | 38 | **Included** — full page- and block-level annotations (12,902 blocks) |
| **`test/`** | 509 | 34 | 28 | **Withheld** — page images only; submit predictions to EvalAI for scoring |

Use `dev/` to understand the annotation schema, calibrate your output format, and validate locally. Run your model on `test/` and upload the predictions to EvalAI to obtain a score. The two splits share no page, no document, and no source book — see [Split Integrity](#split-integrity).

Dr.DocBench targets three gaps in current OCR and document parsing evaluation: **saturation on common genres**, **coarse page-level granularity**, and **missing expert-domain structures** such as chemical formulas, music notation, complex tables, and cross-page layouts.

Instead of sampling pages uniformly, Dr.DocBench selects documents through parser-failure-based sampling: candidate books are scored by inter-parser disagreement across three state-of-the-art systems, and only high-disagreement documents enter annotation. Each source document runs ~100 pages, from which a consecutive ~10–15-page run is annotated for layout, reading order, hierarchical relations, and domain-specific visual content.

Both splits are drawn from the full benchmark, which spans 312 PDFs, 4,514 pages, and 52 subject domains. In the paper's evaluation, 12 pipeline parsers and frontier VLMs were tested; the best overall score is **61.94**, and REFERENCE appears in the worst-5 subjects for **every** evaluated system.

> **Note.** Of the 986 `dev/` pages, 970 carry annotations and 16 are blank pages stored as empty records. All statistics in this section describe `dev/` unless stated otherwise; [Evaluation Results](#evaluation-results) reports scores on the full benchmark.

## Task Format

The task is to convert one or more consecutive document page images into a single continuous Markdown document that preserves both content and structure.

Each evaluation instance consists of:

- **input**: `n` consecutive page images (`n = 2` by default; `n = 1` for Music)
- **output**: one Markdown document covering all `n` pages in reading order
- **reference**: the human-verified `layout_dets` blocks for those pages

A unified inference prompt is used for every model, with no model-specific prompt engineering. It fixes the output convention for each content type:

| Content | Required output format |
| --- | --- |
| Text | Original language and script, no translation, no guessing on illegible scans |
| Multi-column layout | Columns emitted in natural reading order (left-to-right, top-to-bottom) |
| Formulas | LaTeX — `\( ... \)` inline, `\[ ... \]` display |
| Chemistry | `\ce{}` (mhchem) for reactions; SMILES verbatim in a ` ```smiles ` block |
| Tables | HTML wrapped in `<table>`, using `rowspan` / `colspan`; multi-page tables merged into one table |
| Code and pseudocode | Fenced code block with a language tag |
| Music scores | MusicXML in a ` ```musicxml ` block, no `<?xml?>` declaration or `<score-partwise>` wrapper |
| Figures | Purely visual figures ignored; embedded labels and axis text transcribed as plain text |

The evaluation script aligns the predicted Markdown against the ground-truth blocks and scores each component separately (text, formula, table, reading order) before combining them into the overall score. The full prompt is in Appendix G (Figures 9–10) of the paper.

## Supported Tasks

| Task | Annotation used | Metric |
| --- | --- | --- |
| Text recognition | `text` on textual blocks | Normalized Levenshtein edit distance ↓ |
| Layout analysis | `category_type` + `poly` | Detection over 20+ block categories |
| Reading-order prediction | `order` | Reading-order edit distance ↓ |
| Table recognition | `html` / `latex` on `table` blocks | TEDS / TEDS-S ↑ |
| Formula recognition | `latex` on `equation_isolated` | CDM ↑ / edit distance ↓ |
| Structural relations | `extra.relation` | Caption linking, cross-page continuation |
| Optical Music Recognition | MusicXML transcription | Edit distance ↓ / note-level F1 ↑ |

The **overall score** maps each component to a 0–100 scale — text and reading order are `(1 − edit distance) × 100`, formulas are `CDM × 100`, tables are `TEDS × 100` — and averages the components available per sample, excluding missing elements from the denominator.

## Dataset Structure

```
dev/                                   # 986 pages · with ground truth
└── <BISAC_SUBJECT>/
    └── <document_uuid>/
        ├── <document_uuid>.md         # whole-document Markdown
        ├── json/…_page_<N>.json       # ground truth (OmniDocJSON)
        ├── mds/…_<N>.md               # per-page Markdown
        └── images/page_<N>.jpg        # source page image

test/                                  # 509 pages · images only
└── <BISAC_SUBJECT>/
    └── <document_uuid>/
        └── images/page_<N>.jpg        # source page image
```

`<N>` is the original page number in the source book, so a document's run is a consecutive slice such as `page_2152 … page_2166` rather than starting at 1. In `dev/`, the three directories align one-to-one by `<N>`: the JSON is the ground truth, and the Markdown is rendered from it for convenience.

`test/` deliberately contains no `json/` or `mds/` — only page images. Run your model over `test/**/images/*.jpg` and submit predictions to EvalAI; scoring is performed server-side against the withheld annotations.

Each `dev/` example typically contains the following components:

| Field | Type | Description |
| --- | --- | --- |
| `page_info.page_no` | integer | Page number in the source book. |
| `page_info.height` / `width` | integer | Page image dimensions in pixels. |
| `page_info.image_path` | string | Relative path to the page image. |
| `page_info.page_attribute` | object | Page-level metadata: `subject` (one of 52 BISAC domains), `challenge_type` (`perception` / `structural_reconstruction` / `domain_reasoning`), `data_source` (one of nine genres), `language` (a string or a list of strings), `layout` (`single_column` / `double_column` / `three_column` / `1andmore_column` / `other_layout`), and `special_issue` (difficulty flags such as `fuzzy_scan`, `table_wireless_line`). |
| `layout_dets` | array | Block annotations, one object per region. |
| `layout_dets[].category_type` | string | Block category — one of 20 types including `text_block`, `title`, `figure`, `table`, `equation_isolated`, `code_algorithm`, `header`, `page_number`. |
| `layout_dets[].poly` | array | Quadrilateral as `[x1,y1, x2,y2, x3,y3, x4,y4]` in page pixels. |
| `layout_dets[].order` | integer | Reading-order index. Absent on page elements (header, footer, page number) and masks. |
| `layout_dets[].anno_id` | integer | Page-unique block id, referenced by relations. |
| `layout_dets[].text` | string | Transcribed content, Markdown-flavored. |
| `layout_dets[].latex` | string | LaTeX for equations, and a LaTeX rendering for tables. |
| `layout_dets[].html` | string | HTML structure with `rowspan` / `colspan` for tables. |
| `layout_dets[].attribute` | object | Category-dependent attributes: text rotation and background, table line style and span, formula type. |
| `extra.relation` | array | Inter-block links as `{source_anno_id, target_anno_id, relation_type}` — `parent_son` connects a figure/table/equation to its caption or footnote; `truncated` marks two blocks forming one logical unit, with `target_anno_id: -1` for cross-page continuation. |

### Example Instance

A real record — `EDUCATION/a8727f40-.../json/a8727f40-..._page_258.json` — showing a borderless horizontal table with a rotated caption, a footnote, body text, and a page number. Long `text`, `latex`, and `html` values are elided with `...`:

```json
[{
  "page_info": {
    "page_name": "a8727f40-f59f-4e0d-be5b-4104252375b2",
    "page_no": 258,
    "height": 744,
    "width": 462,
    "image_path": "../images/a8727f40-f59f-4e0d-be5b-4104252375b2/page_258.jpg",
    "original_image_url": "https://.../EDUCATION/a8727f40-.../page_258.jpg",
    "page_attribute": {
      "data_source": "academic_literature",
      "subject": "EDUCATION",
      "challenge_type": "structural_reconstruction",
      "language": "english",
      "layout": "other_layout",
      "special_issue": ["table_horizontal", "table_fewer_line"]
    }
  },
  "layout_dets": [
    {
      "category_type": "table_caption",
      "poly": [42.92, 348.19, 54.12, 348.19, 54.12, 673.85, 42.92, 673.85],
      "ignore": false,
      "order": 1,
      "anno_id": 9,
      "text": "**Table 20.1.** Tabulated Results from Student Reflection Essays",
      "attribute": {
        "text_language": "text_english",
        "text_background": "single_colored",
        "text_rotate": "rotate270"
      }
    },
    {
      "category_type": "table",
      "poly": [56.43, 48.95, 380.21, 48.95, 380.21, 673.39, 56.43, 673.39],
      "ignore": false,
      "order": 2,
      "anno_id": 2,
      "latex": "\\begin{table}[]\n\\begin{tabular}{llll}\n\\textbf{Important Aspects...}",
      "html": "<table>\n<tbody>\n<tr>\n<td style=\"text-align: left;\"><strong>Important Aspects...",
      "attribute": {
        "table_layout": "vertical",
        "line": "fewer_line",
        "language": "table_en",
        "with_span": "false",
        "include_background": "false",
        "include_equation": "false",
        "include_photo": "false",
        "with_structured_text": "true"
      }
    },
    { "category_type": "table_footnote", "order": 3, "anno_id": 7, "text": "...", "...": "..." },
    { "category_type": "text_block",     "order": 4, "anno_id": 4, "text": "...", "...": "..." },
    { "category_type": "page_number",                "anno_id": 12, "text": "258", "...": "..." }
  ],
  "extra": {
    "relation": [
      { "source_anno_id": 2, "target_anno_id": 7, "relation_type": "parent_son" },
      { "source_anno_id": 2, "target_anno_id": 9, "relation_type": "parent_son" }
    ]
  }
}]
```

Three things to note in this record:

- **`order` skips the page number.** Blocks 1–4 carry reading order; `page_number` has no `order` key, because page elements are excluded from reading-order evaluation.
- **Relations point from parent to child.** Both relations have the table (`anno_id: 2`) as source, linking it to its footnote (`7`) and caption (`9`).
- **Attributes encode the difficulty.** `line: fewer_line` plus `text_rotate: rotate270` is exactly the borderless-table-with-rotated-caption case that causes large TEDS drops — see [Evaluation Results](#evaluation-results).

### Loading

```python
import glob, json

pages = []
for path in sorted(glob.glob("dev/*/*/json/*.json")):
    record = json.load(open(path))
    if record:
        pages.append(record[0])

page = pages[0]
blocks = [b for b in page["layout_dets"] if "order" in b and not b["ignore"]]
for b in sorted(blocks, key=lambda b: b["order"]):
    print(b["order"], b["category_type"], (b.get("text") or "")[:60])

test_images = sorted(glob.glob("test/*/*/images/page_*.jpg"))
```

Blank pages are stored as empty lists, so `if record` skips them. `order` is absent on page elements such as headers, footers, and page numbers, which are excluded from reading-order evaluation.

## Dataset Statistics

*Figures in this section describe the `dev/` split. Each document contributes a consecutive run of up to 15 pages; page counts total 970, excluding 16 blank pages.*

| Data source | Pages | Layout | Pages | Challenge type | Pages |
| --- | ---: | --- | ---: | --- | ---: |
| `book` | 726 | `single_column` | 494 | `structural_reconstruction` | 580 |
| `magazine` | 84 | `other_layout` | 193 | `perception` | 387 |
| `academic_literature` | 75 | `double_column` | 146 | `domain_reasoning` | 2 |
| `colorful_textbook` | 70 | `1andmore_column` | 95 | | |
| `exam_paper` | 14 | `three_column` | 42 | | |
| `note` | 1 | | | | |

Some cells are too thin to analyze on their own: `note` (1 page), `domain_reasoning` (2 pages), and `exam_paper` (14 pages) are present for schema completeness rather than as evaluable slices. Breakdowns along these axes should be read from the full benchmark, not this split.

**Subjects** (documents per subject): EDUCATION 5 · HOUSE&HOME 4 · DESIGN, GAMES&ACTIVITIES, JUVENILENONFICTION, SOCIALSCIENCE, SPORTS&RECREATION, YOUNGADULTFICTION 3 each · BUSINESS&ECONOMICS, COMPUTERS, CRAFTS&HOBBIES, GARDENING, PHILOSOPHY, POETRY, POLITICALSCIENCE, STUDYAIDS, TRANSPORTATION 2 each · ARCHITECTURE, ART, BIOGRAPHY&AUTOBIOGRAPHY, BODY,MIND&SPIRIT, COMICS&GRAPHICNOVELS, COOKING, FAMILY&RELATIONSHIPS, FICTION, HISTORY, HUMOR, LANGUAGEARTS&DISCIPLINES, LAW, LITERARYCRITICISM, MEDICAL, PERFORMINGARTS, PETS, PSYCHOLOGY, SELF-HELP, TECHNOLOGY&ENGINEERING, TRUECRIME, YOUNGADULTNONFICTION 1 each.

**Difficulty flags**: `colorful_backgroud` 157 · `fuzzy_scan` 140 · `table_full_line` 40 · `table_fewer_line` 25 · `table_horizontal` 18 · `table_wireless_line` 15 · `table_span` 2 · `table_with_formula` 2.

**Language**: English.

### Test split

`test/` holds **509 page images** from **34 documents** across **28 BISAC subject domains**, drawn from the same corpus and annotated under the same schema. Its subject mix is not identical to `dev/`. Per-subject and per-attribute breakdowns are withheld so that they cannot be used to tune submissions.

### Split Integrity

`dev/` and `test/` were verified disjoint at four levels:

| Check | Result |
| --- | --- |
| Document UUID (66 vs 34) | 0 shared |
| Page identity (UUID, page number) | 0 shared |
| SHA-256 of every image (986 × 509) | 0 identical |
| 64×64 grayscale pixel distance, all pairs | closest matches are blank pages only |

No source book (ISBN) appears in both splits, so `test/` cannot be reconstructed from `dev/` content.

## Evaluation Results

For the **full 4,514-page benchmark** as reported in the paper. Bold = best.

| Model | Size† | Access | Text↓ | Form.↓ | CDM↑ | TEDS↑ | TEDS-S↑ | Order↓ | **Overall↑** |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| MinerU 2.5 | 1.2B | Open | 0.33 | 0.51 | 24.15 | **55.85** | **63.70** | 0.30 | 54.37 |
| PaddleOCR | 0.9B | Open | 0.73 | 0.42 | 30.73 | 51.79 | 59.94 | 0.71 | 34.78 |
| GPT-5.5 | – | Closed | **0.19** | 0.36 | 34.55 | 48.90 | 58.96 | **0.17** | **61.94** |
| Kimi-K2.5 | 32B | Open | 0.19 | 0.35 | 27.04 | 51.98 | 61.09 | 0.18 | 60.38 |
| Claude Opus 4.6 | – | Closed | 0.22 | 0.37 | 32.02 | 49.21 | 58.12 | 0.19 | 60.19 |
| Gemini 3.1 Pro | – | Closed | 0.22 | 0.35 | 32.22 | 51.26 | 59.03 | 0.21 | 60.13 |
| Qwen3.5-Flash | 3B | Open | 0.26 | 0.32 | 35.69 | 46.24 | 54.38 | 0.26 | 57.51 |
| Qwen3.5-Plus | 17B | Open | 0.25 | **0.31** | 30.40 | 49.77 | 58.23 | 0.26 | 57.32 |
| Qwen3.5-122B-A10B | 10B | Open | 0.23 | 0.32 | 32.54 | 38.31 | 44.99 | 0.23 | 56.32 |
| Doubao-Seed-1.6-Vision | – | Closed | 0.34 | 0.31 | **41.28** | 33.93 | 42.16 | 0.28 | 53.14 |
| GPT-4o | – | Closed | 0.37 | 0.47 | 36.02 | 30.14 | 38.62 | 0.31 | 49.73 |
| Nemotron-Nano-12B | 12B | Open | 0.62 | 0.76 | 12.82 | 27.03 | 34.09 | 0.564 | 30.33 |

†activated parameters. A unified inference prompt is used for all models, with a default 2-page sliding window (1 page for Music).

## Limitations

- **Scope.** A diagnostic benchmark for parsing and structural recognition, not for downstream document-intelligence tasks such as QA or retrieval.
- **Standardized prompting.** One unified prompt is used for all models, so results reflect standardized rather than best-achievable performance. Prompt-incapable parsers (MinerU, PaddleOCR) cannot follow format instructions at all, and their scores on prompt-dependent outputs are reference baselines only.
- **Subject is one axis of difficulty.** Layout density, scan quality, multilingual content, and cross-page continuation also contribute, so subject-level analysis complements block- and attribute-level analysis.
- **Format-based metrics.** LaTeX, HTML, and MusicXML enable scalable automatic evaluation but do not capture equivalent representations; semantic or execution-based metrics would improve this.
- **Music is exploratory** — a 6-document probe, scored by string-level edit distance that conflates schema verbosity with musical content.

## Ethical Considerations

Dr.DocBench is intended for **research evaluation of document parsing systems**. During construction and QC, pages containing clearly sensitive personal information, private identifying information, or inappropriate content were removed when identified, and annotators were instructed to flag such cases for review.

The **annotations** are released under [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/) (public domain dedication). The **underlying source documents** retain the rights of their original publishers and may remain subject to their original access, copyright, and licensing conditions — source images come from publicly available books and scanned documents, used solely for benchmark construction and research evaluation under fair-use principles. Users should follow the terms of the original sources.

## Citation

```bibtex
@misc{drdocbench2026,
  title         = {Dr.DocBench: A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing},
  author        = {Dr.DocBench Team},
  year          = {2026},
  eprint        = {2606.01393},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2606.01393}
}
```