ViTaB-A: Evaluating Multimodal Large Language Models on Visual Table Attribution

Dataset Description

This dataset is a perturbed verion of the HiTab dataset. Specifically we retain the QA, cell attribution and JSON formatted tables. We perturb this with markdown version and multiple images of the same table for attribution research in imges of structured data.

Key Features

• Visual Diversity: Each table is rendered in 5 visual variations:

  • Arial font with standard headers
  • Times New Roman font with standard headers
  • Red headers with white text
  • Blue headers with white text
  • Green headers with white text

• Multi-modal Representations: Each sample includes:

  • JSON structure (programmatic access)
  • Markdown format (text-based with row/column labels)
  • PNG images (visual format, base64 encoded)

• Standardized Coordinates: Tables include Excel-style coordinates (A, B, C... for columns; 1, 2, 3... for rows) enabling precise answer localization

• Rich Annotations: Includes questions, answers, answer formulas, and cells indicating answer locations

Dataset Statistics

• Total Samples: ~8,300 (from HiTab)
• Unique Tables: ~2,500
• Splits:
  • Train: 80%
  • Validation: 6.6%
  • Dev: 6.6%
  • Test: 6.8%

Data Format

Each line in visualcite.jsonl contains:

{
  "id": "visualcite_000001",
  "split": "train",
  "question": "What was the team's total rushing yards in 2015?",
  "answer": "529",
  "answer_formulas": ["SUM(F8)"],
  "highlighted_cells": [[7, 5]],
  "table_json": {
    "texts": [...],
    "merged_regions": [...],
    "top_header_rows_num": 2,
    "left_header_columns_num": 1,
    "title": "career statistics"
  },
  "table_md": "| | A | B | C | ...\n| --- | --- | --- | ...",
  "table_images": {
    "arial": "base64_encoded_png...",
    "times_new_roman": "base64_encoded_png...",
    "red": "base64_encoded_png...",
    "blue": "base64_encoded_png...",
    "green": "base64_encoded_png..."
  },
  "source": "hitab_train",
  "source_id": "0_totto74-1"
}

Field Descriptions

Field	Type	Description
id	string	Unique identifier for the sample
split	string	Dataset split (train/validation/dev/test)
question	string	Natural language question about the table
answer	string	Ground truth answer
answer_formulas	list	Formulas used to derive the answer
highlighted_cells	list	Row/column indices of cells relevant to the answer
table_json	dict	Complete table structure with metadata
table_md	string	Markdown representation with coordinate labels
table_images	dict	Base64-encoded PNG images (5 variations)
source	string	Original HiTab split
source_id	string	Original HiTab table identifier

Usage

Loading the Dataset

import json

# Load dataset
with open('data/visualcite.jsonl', 'r') as f:
    dataset = [json.loads(line) for line in f]

# Filter by split
train_data = [sample for sample in dataset if sample['split'] == 'train']

Decoding Images

import base64
from PIL import Image
from io import BytesIO

# Decode a table image
sample = dataset[0]
img_data = base64.b64decode(sample['table_images']['arial'])
image = Image.open(BytesIO(img_data))
image.show()

Using Markdown Tables

# Access markdown representation
markdown_table = sample['table_md']
print(markdown_table)

# The markdown includes coordinate labels:
# |  | A | B | C | D |
# | --- | --- | --- | --- | --- |
# | 1 | season | team | games | ... |
# | 2 | 2014 | tampa bay | 8 | ... |

Original HiTab Citation:

@inproceedings{cheng-etal-2022-hitab,
    title = "{H}i{T}ab: A Hierarchical Table Dataset for Question Answering and Natural Language Generation",
    author = "Cheng, Zhoujun  and others",
    booktitle = "Proceedings of ACL 2022",
    year = "2022"
}

Dataset Creation

Source Data

VisualCite is derived from HiTab, a hierarchical table QA dataset created by Microsoft Research.

Generation Process

1. Table Rendering: Each table is converted to HTML with 5 different styling configurations
2. Image Generation: HTML rendered to PNG using headless browser (Playwright)
3. Coordinate System: Row/column labels added (A-Z, 1-N) for precise referencing
4. Markdown Conversion: Tables converted to markdown with preserved coordinates
5. Base64 Encoding: Images encoded for embedding in JSONL

Quality Assurance

• All images are PNG format
• Tables preserve original structure including merged cells
• Coordinate systems are consistent across all representations
• Random seed (42) ensures reproducible splits

Limitations

• Merged cells in markdown show empty strings for non-primary cells
• Very wide tables may have readability issues in image format
• Base64 encoding significantly increases file size
• Limited to tables from Wikipedia and web sources (HiTab's scope)

License

This dataset inherits the license from HiTab. Please refer to the HiTab repository for licensing details.

Acknowledgments

We thank the creators of HiTab for providing the foundation for this dataset. Special thanks to Microsoft Research for their contributions to table understanding research