File size: 2,724 Bytes
2742d4b a2fb8ce e6038bf a2fb8ce e6038bf a2fb8ce 0e101f1 a2fb8ce b997f63 a2fb8ce b521e94 4375490 a2fb8ce | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | ---
dataset_info:
features:
- name: id
dtype: string
- name: image
dtype: image
- name: question
dtype: string
- name: answer
dtype: string
splits:
- name: train
num_bytes: 11563543600
num_examples: 16450
- name: test
num_bytes: 1638512366
num_examples: 3159
download_size: 15933142593
dataset_size: 13202055966
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---
# TabComp 📊
**A Benchmark for OCR-Free Visual Table Reading Comprehension**
This dataset accompanies the paper [TabComp: A Dataset for Visual Table Reading Comprehension](https://aclanthology.org/2025.findings-naacl.320.pdf)
TabComp evaluates **Vision-Language Models (VLMs)** on their ability to **read, understand, and reason over table images** without relying on OCR, using **generative question answering**.
## 🔍 Why TabComp?
Modern VLMs perform well on general VQA but struggle with **tables**, which require:
- Structured reasoning across rows/columns
- Understanding layout + text jointly
- Multi-step inference over semi-structured data
👉 TabComp isolates this challenge and provides a **focused benchmark for table understanding**.
## 📊 Dataset Overview
- **Images:** 3,318 table images
- **QA pairs:** 19,610
- **Answer type:** Generative (natural language)
- **Domain:** Industrial documents
- **Text types:** Printed + handwritten
### Task Definition
Given:
- A **table image**
- A **question**
Generate:
- A **natural language answer** requiring table comprehension
## 🧠 What Makes It Challenging?
- ❌ No OCR signals
- ✅ Dense textual + structural information
- ✅ Long-range dependencies across table cells
- ✅ Generative answers (not extractive spans)
## 📁 Data Format
🏆 Leaderboard (Baseline Results)
Performance on TabComp (generative metrics):
| Model | Setting | B-4 ↑ | ROUGE-L ↑ | BERTScore ↑ | METEOR ↑ |
| ----------- | ---------- | --------- | --------- | ----------- | --------- |
| Donut-base | Fine-tuned | **42.69** | 37.29 | 83.38 | **60.14** |
| Donut-base | End-to-end | 28.59 | 32.24 | 85.06 | 47.19 |
| Donut-proto | Fine-tuned | 6.49 | 17.84 | 73.26 | 19.80 |
| Donut-proto | End-to-end | 34.87 | 37.02 | 87.74 | 56.49 |
| UReader | Zero-shot | 28.14 | **37.64** | **88.04** | 20.71 |
Full metrics (BLEU-1/2/3/4, CIDEr) available in the paper.
## We welcome:
- New model evaluations
- Error analysis
- Extensions to multilingual / multi-table settings
## Contact
For collaboration, email **Somraj Gautam** gautam.8@iitj.ac.in |