multimodal-vision-ai/mvai-doctag-r3-v1

MVAI DocTag R3 v1

Basic Information

Field	Value
Dataset ID	multimodal-vision-ai/mvai-doctag-r3-v1
Version	v1
Owner	Dizzar, Hohai University / multimodal-vision-ai
Dataset type	Image + DocTags text annotations
Intended use	OCR and document layout research, especially image-to-DocTags SFT / GRPO

This dataset is the formal closure version of the existing HohaiR3 DocTags data asset. It contains the HohaiR3 gold, human-labeled, HTML-rendered, and synthetic enhancement subsets that are already present in this Hugging Face repository.

The FinePDFs-derived silver datasets are intentionally not included in this version because they are large and require a separate closure process.

Data Files

The repository currently contains 23 zip files, about 82.70 GiB in total.

File	Size GiB	Notes
complaint.zip	0.571	complaint-style generated data
en_document.zip	3.156	English document data
en_hand_writing.zip	0.040	English handwriting data
en_table.zip	0.121	English table data
html_empty_zh_table.zip	0.013	HTML-rendered Chinese empty-table data
html_full_zh_table.zip	0.092	HTML-rendered Chinese dense-table data
html_sparse_zh_table.zip	0.019	HTML-rendered Chinese sparse-table data
human_label.zip	0.297	human-labeled data
mask_wendang.zip	0.210	masked document variants
multi_enhanced_complaint.zip	3.386	augmented complaint data
multi_enhanced_en_document.zip	27.731	augmented English document data
multi_enhanced_en_hand_writing.zip	0.303	augmented English handwriting data
multi_enhanced_en_table.zip	1.912	augmented English table data
multi_enhanced_zh_document.zip	28.620	augmented Chinese document data
multi_enhanced_zh_hand_writing.zip	0.491	augmented Chinese handwriting data
multi_enhanced_zh_table.zip	1.882	augmented Chinese table data
muti_enhanced_biaoge.zip	2.618	augmented human table data
muti_enhanced_wangye.zip	0.235	augmented webpage-like data
muti_enhanced_wendang.zip	3.692	augmented human document data
muti_enhanced_zhengjian.zip	0.854	augmented certificate/ID-like data
zh_document.zip	6.246	Chinese document data
zh_hand_writing.zip	0.068	Chinese handwriting data
zh_table.zip	0.139	Chinese table data

Data Structure

Each zip contains page images and corresponding .doctags annotations. The inner directory layout may differ across subsets. Common structures include:

train/image/<sample_id>.<png|jpg>
train/dt/<sample_id>.doctags

and class-based structures such as:

<class>/images/<sample_id>.jpg
<class>/doctags/<sample_id>.doctags

The recommended logical schema after extraction is:

Field	Description
sample_id	Stable ID derived from subset and file stem
image	Document page image
doctags	DocTags annotation text
source_zip	Source zip file in this repository
source_subset	Subset name
source_image_path	Image path inside the zip
source_doctag_path	DocTags path inside the zip
split	Existing split if present; otherwise treated as source/train data

Splits

This repository is stored as zip subsets rather than a normalized Hugging Face DatasetDict.

Train: all current zip files are treated as source/train data unless an inner manifest states otherwise.
Validation: None in this formal closure version.
Test: None in this formal closure version.

No new validation or test split is created by this closure step.

Processing Summary

Known processing operations from project records and local manifests include:

• Human labeling for selected document/table/certificate-like samples.
• Python DocTags rendering or reverse rendering for generated image/label pairs.
• HTML table generation for Chinese table subsets.
• Image augmentation for enhanced subsets, including brightness, contrast, grayscale, sharpening, denoising, autocontrast, and binarization variants.
• Masked document generation for partial text/block visibility experiments.

Some historical generation details, such as exact random seeds for all subsets, are not recoverable from the current repository files and are recorded as unknown in manifest.yaml.

Formal Usage

For reproducible research, pin this dataset by exact Hugging Face commit SHA instead of using main.

The pre-closure data revision observed on 2026-08-26 was:

64ea6fcc6bd81dcba17204aa08c36611d1833618

After this Dataset Card and manifest.yaml are committed, use the latest commit SHA of this repository as the formal closure revision.

Known Issues

• The current data is zip-packaged rather than a normalized HF DatasetDict.
• Validation and test splits are not defined in this closure version.
• Some historical seeds and script commits are unknown.
• A local extracted reference inspected on 2026-08-26 was not a perfect mirror of this HF file list; therefore this HF repository is the formal source of truth.

Research Repository

The corresponding research archive is maintained in:

multimodal-vision-ai/multimodal-document-ai

The GitHub archive should contain dataset documentation, processing scripts, smoke tests, quality checks, and provenance records. It should not contain the full dataset files.