WildTableBench

[Paper] [GitHub]

WildTableBench is a benchmark for evaluating multimodal foundation models on table understanding in the wild.

• 402 real-world table images collected from diverse domains
• 928 questions across 5 categories and 17 subtypes
• Covers numerical reasoning, verification, cell locating, structural understanding, and more

Dataset Structure

WildTableBench/
├── metadata.csv       # 928 questions with labels
└── images/            # 402 table images (1.jpg … 402.jpg)

metadata.csv columns

Column	Description
file_name	Relative path to the image (e.g. images/1.jpg)
image_id	Image index (1–402)
index	Question index within the image (1, 2, …)
uuid	Unique question ID (e.g. wtb_001_q1)
question	Question text
category	Category ID (C1–C5)
category_name	Category name
subtype_id	Subtype ID (e.g. C2-C)
subtype_name	Subtype name
ground_truth	Ground truth answer

Question Categories

ID	Name	Subtypes
C1	Cell-Level	Transcription, Cell Locating, Semantic Lookup & Struct., Excel Formula
C2	Numerical	Basic Numerical, Conditional Numerical, Multi-step Conditional, Ranking
C3	Verification	Value Verification, Aggregate Verification, Conditional Verification
C4	Hypothetical	Hypothetical Condition, Value Modification, Row Operation
C5	Color	Color Identification, Color-based Counting, Color-based Reasoning

Usage

from datasets import load_dataset

dataset = load_dataset("jzhuang/WildTableBench")

Citation

@article{huang2025wildtablebench,
  title={WildTableBench: Benchmarking Multimodal Foundation Models on Table Understanding In the Wild},
  author={Junzhe Huang and Xiaoxiao Sun and Yan Yang and Yuxuan Hou and Ruotian Zhang and Sirui Li and Hehe Fan and Serena Yeung-Levy and Xin Yu},
  journal={arXiv preprint arXiv:2605.01018},
  year={2025},
}