01
Document Parsing
This track focuses on recovering structured content from document pages, including natural text, tables, formulas, layout structure, and reading order.
Overview
The DataMFM Challenge focuses on multimodal document understanding at the intersection of vision, language, and structured reasoning. For this challenge, we prepare two new datasets based on OmniDocBench and ChartNet, corresponding to document parsing and chart understanding.
These two challenge components are designed to cover complementary capabilities: extracting structured content from document pages and understanding chart-centric visual information. This challenge is part of the Emerging Directions in Data for Multimodal Foundation Models (DataMFM) workshop at CVPR 2026.
Scope
The challenge scope covers two complementary tasks in multimodal document understanding: document parsing and chart understanding. Together, they evaluate how well models can extract, organize, and reason over structured information from complex visual documents.
02
Chart Understanding
This track focuses on two core chart understanding tasks: recovering structured chart data through chart-to-CSV extraction and generating grounded chart summaries through chart-to-summary generation.
Timeline
Apr 27
Release
May 11
Submission opening
May 29
Submission deadline
Jun 03
Workshop
Awards
Prize structure: Awards are granted per task for both the document parsing track and the chart understanding track.
01
1st Place
$1,000 per task
02
2nd Place
$500 per task
03
3rd Place
$300 per task
Leaderboard
Leaderboard: Results are updated periodically from successful public EvalAI submissions. Each table keeps the best public submission per team for the corresponding task.
Loading leaderboard data...
Document Parsing
| Rank | Team | |||||
|---|---|---|---|---|---|---|
| Loading... | ||||||
Chart Understanding
| Rank | Team | |||||
|---|---|---|---|---|---|---|
| Loading... | ||||||
Dataset
The challenge dataset is organized around two newly prepared components based on OmniDocBench and ChartNet. Public release materials are provided below for the document parsing and chart understanding tracks.
01
Document Parsing Data
This newly prepared document parsing dataset is based on OmniDocBench and currently includes 1,005 page images organized under 89 document folders. It supports page-level parsing with supervision over text, tables, formulas, layout structure, and reading order.
02
Chart Understanding Data
This newly prepared chart understanding dataset is based on ChartNet and currently includes 3,807 chart images, with 2,000 synthetic samples and 1,807 real-world samples. For this challenge, the chart track focuses on two tasks: chart-to-CSV and chart-to-summary.
Submission
Current public note: submission files are organized to match the structure of the two released challenge datasets. The document parsing track uses flat, document-level Markdown files, while the chart understanding track uses split-based JSONL prediction files for the two supported tasks: chart-to-CSV and chart-to-summary.
Document Parsing Submission
For the document parsing track, submit one document-level .md file per document. The .zip archive should be flat and contain 89 Markdown files, with each filename matching the corresponding document UUID in the ground truth:
submission.zip
├── 00000000-0000-0000-0000-000000000001.md
├── 00000000-0000-0000-0000-000000000002.md
├── 00da2f04-2fd9-4ed2-a6e3-0839d06da691.md
├── 09933144-8982-4100-b18c-fa6dfacba7c7.md
└── ...
─── Example document-level .md content ───
# Section Title
Body text paragraph with standard Markdown formatting.
Separate paragraphs with double newlines.
$$
\frac{\partial L}{\partial \theta} = \sum_{i=1}^{N} \nabla_\theta \ell(f(x_i), y_i)
$$
<table>
<tr>
<th>Model</th><th>Accuracy</th>
</tr>
<tr>
<td>Baseline</td><td>82.3</td>
</tr>
</table>
Format Rules:
• Archive layout: Keep the archive flat. Do not wrap the Markdown files inside an extra top-level directory.
• File count: Submit one
• File naming: Each prediction file must match the corresponding document UUID in the ground truth, for example
• Validation: Submissions with missing filenames, extra filenames, or an extra wrapper directory may fail validation.
• Text: Standard Markdown. Paragraphs separated by double newlines (
• Formulas: Display formulas in
• Tables: HTML
• Order: Elements should appear in natural reading order within each document-level
• Archive layout: Keep the archive flat. Do not wrap the Markdown files inside an extra top-level directory.
• File count: Submit one
.md file per document, for 89 document-level Markdown files in total.• File naming: Each prediction file must match the corresponding document UUID in the ground truth, for example
00da2f04-2fd9-4ed2-a6e3-0839d06da691.md.• Validation: Submissions with missing filenames, extra filenames, or an extra wrapper directory may fail validation.
• Text: Standard Markdown. Paragraphs separated by double newlines (
\n\n). Headings with #.• Formulas: Display formulas in
$$...$$, inline in $...$. Content must be LaTeX.• Tables: HTML
<table> format is recommended for merged cells; Markdown pipe tables are also accepted.• Order: Elements should appear in natural reading order within each document-level
.md file.
Chart Understanding Submission
For the chart understanding track, organize submissions by split and task. Prediction files should follow the JSONL structure used by the current ChartNet-based evaluation pipeline for chart-to-CSV and chart-to-summary:
submission.zip
├── real/
│ ├── chart2csv_predictions.jsonl
│ └── chart2summary_predictions.jsonl
└── synthetic/
├── chart2csv_predictions.jsonl
└── chart2summary_predictions.jsonl
─── Example JSONL lines ───
{"imagename": "20150131_inc871_7.png", "predicted_csv": "Date,Value\n2011,100"}
{"imagename": "20150131_inc871_7.png", "predicted_summary": "The chart shows ..."}
Format Rules:
• Split layout: Use separate
• Task files: Provide JSONL predictions for
• Required keys: Use
• Alignment: Each JSONL entry must match the released chart image name and split.
• Split layout: Use separate
real/ and synthetic/ directories.• Task files: Provide JSONL predictions for
chart2csv and chart2summary only.• Required keys: Use
imagename plus the corresponding prediction field such as predicted_csv or predicted_summary.• Alignment: Each JSONL entry must match the released chart image name and split.
Rules
Eligibility
Open to all researchers worldwide. No team size limit.
Submissions
Max 3 per day. Final evaluation allows 2 submissions.
Requirements
Top teams must submit technical report. External data must be disclosed.