Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
815
2.27k
label
class label
0 classes
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
End of preview. Expand in Data Studio

VietFinTab

Fork notice: this repository is a fork of VietFinTabGroup/VietFinTab. The original data (images/, structures/, contents/, metadata.jsonl) is unchanged. This fork adds conversion scripts and converted datasets. Please cite and credit the original authors.

Dataset Overview

VietFinTab is a large-scale Vietnamese financial table dataset designed for Table Structure Recognition (TSR), Document AI, and Optical Character Recognition (OCR) research.

The dataset is collected from publicly available financial reports of 13 Vietnamese companies spanning the period from 2023 to 2025.

The companies included in the dataset are:

  • EVN Finance Joint Stock Company (EVF)
  • FPT Corporation (FPT)
  • Hoang Anh Gia Lai Joint Stock Company (HAG)
  • Hoa Phat Group Joint Stock Company (HPG)
  • Military Commercial Joint Stock Bank (MB)
  • PetroVietnam Drilling & Well Services Corporation (PVD)
  • PetroVietnam Technical Services Corporation (PVS)
  • SSI Securities Corporation (SSI)
  • TPBank (TPB)
  • Vingroup Joint Stock Company (VIN)
  • Vietjet Aviation Joint Stock Company (VJC)
  • Vinamilk (VNM)
  • VietinBank (VTB)

Dataset Statistics

  • 10,383 table images (.jpg)
  • 10,383 table structure annotations (.xml)
  • 1,000 table content annotations (.json)

Each table image has a corresponding XML file describing its table structure. Additionally, 1,000 tables include detailed cell text annotations with bounding boxes.


Directory Structure

VietFinTab/
β”‚
β”œβ”€β”€ images/
β”‚   β”œβ”€β”€ EVF/
β”‚   β”œβ”€β”€ FPT/
β”‚   β”œβ”€β”€ ...
β”‚
β”œβ”€β”€ structures/
β”‚   β”œβ”€β”€ EVF/
β”‚   β”œβ”€β”€ FPT/
β”‚   β”œβ”€β”€ ...
β”‚
β”œβ”€β”€ contents/
β”‚   β”œβ”€β”€ EVF/
β”‚   β”œβ”€β”€ FPT/
β”‚   β”œβ”€β”€ ...
β”‚
└── metadata.jsonl

File Formats

Table Images

  • Format: JPG
  • One image corresponds to one table extracted from a financial report.

Table Structure Annotation

Each XML annotation follows the PubTables-1M annotation format, containing:

  • Table bounding box
  • Row bounding boxes
  • Column bounding boxes
  • Spanning cell annotations

These annotations are suitable for training and evaluating table structure recognition models.

Table Content Annotation

For 1,000 selected tables, JSON annotations provide:

  • Text
  • Text bounding boxes

These annotations enable OCR-related tasks, text recognition, and end-to-end table understanding.


Metadata

The dataset includes a metadata.jsonl file.

Example:

{
  "id": "FPT_Q1_2023_0001",
  "company": "FPT",
  "quarter": "Q1",
  "year": 2023,
  "file_name": "images/FPT/FPT_Q1_2023_0001.jpg",
  "structure_file": "structures/FPT/FPT_Q1_2023_0001.xml",
  "content_file": "contents/FPT/FPT_Q1_2023_0001.json",
  "has_content_label": true
}

Supported Tasks

VietFinTab can be used for:

  • Table Structure Recognition (TSR)
  • Table Detection
  • Table Parsing
  • OCR
  • End-to-End Table Understanding
  • Financial Document Analysis

Conversion Scripts

Scripts live in scripts/. Each one is a self-contained uv script (dependencies declared inline), run from the repository root.

Script Input Output
export_yolo.py original dataset datasets/VietFinTab-reconstruction-yolo
yolo_to_coco.py YOLO dataset datasets/VietFinTab-reconstruction-coco
export_html.py original dataset (tables with content labels only) datasets/VietFinTab-image2html/{images,html}
deduplicate_html.py html/ datasets/VietFinTab-image2html/html_unique
overlay.py original dataset figures/overlay_<ID>.png (structure + OCR boxes drawn on the image)
stats.py original dataset figures/*.png (dataset statistics charts)
# YOLO, then COCO (COCO is built from the YOLO output)
uv run scripts/export_yolo.py --val 0.1 --seed 0
uv run scripts/yolo_to_coco.py

# Image -> HTML, then structure-level deduplication
uv run scripts/export_html.py
uv run scripts/deduplicate_html.py            # add --ignore-th to treat <th> like <td>

# Tests
cd scripts && for t in test_*.py; do uv run "$t"; done

Converted Datasets

The datasets/ folder is generated by the scripts above and stored with Git LFS.

VietFinTab-reconstruction-yolo

Object detection dataset in YOLO format (Ultralytics data.yaml).

  • Classes: 0: row, 1: column, 2: cell, 3: table
  • Splits: 9,345 train / 1,038 val images (random 10% val, seed 0)
  • Cells are derived from the source: every row Γ— column intersection not covered by a spanning cell, plus the spanning cells themselves.
  • Boxes are clipped to the image; boxes that become empty are dropped.
VietFinTab-reconstruction-yolo/
β”œβ”€β”€ data.yaml
β”œβ”€β”€ images/{train,val}/<ID>.jpg
└── labels/{train,val}/<ID>.txt   # class xc yc w h (normalized)

VietFinTab-reconstruction-coco

The same dataset and splits in CVAT "COCO 1.0" layout (zip the folder contents to import into CVAT).

  • Categories: 1: row, 2: column, 3: cell, 4: table
  • Annotations: 607,466 train / 65,186 val
  • Each box is also stored as a 4-point segmentation polygon.
VietFinTab-reconstruction-coco/
β”œβ”€β”€ annotations/instances_{train,val}.json
└── images/{train,val}/<ID>.jpg

VietFinTab-image2html

Image β†’ HTML table dataset, built from the 1,000 tables with content labels.

  • html/<ID>.html: plain <table> (no CSS). The grid comes from the XML rows/columns; spanning cells become rowspan/colspan. Cell text comes from the OCR tokens inside each cell, in reading order.
  • html_unique/<ID>_<n>.html: 296 tables left after deduplication by structure (span grid only, text ignored). The first table seen for each structure is kept, with its text.
VietFinTab-image2html/
β”œβ”€β”€ images/<ID>.jpg
β”œβ”€β”€ html/<ID>.html
└── html_unique/<ID>_<n>.html

Source

The dataset is constructed from publicly available financial reports released by Vietnamese companies.

The original reports are used solely for academic research and dataset construction.


Citation

If you use VietFinTab (or the converted datasets from this fork) in your research, please cite the original dataset.

@dataset{vietfintab,
  title={VietFinTab: A Vietnamese Financial Table Dataset},
  author={Pham Hong Quan, Bui Hai Dang, Tong Thi Ngoc Quynh},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/WuanHP/VietFinTab}
}

License

This dataset is released under the CC BY-NC 4.0 license.

It may be used for research and non-commercial purposes. Users are responsible for complying with the licensing terms of the original financial reports from which the tables were extracted.

Downloads last month
-