A benchmark dataset of 206 programmatically generated PDF documents with deterministic ground truth markdown, designed to evaluate OCR and Vision-Language Model (VLM) performance on tabular document parsing.

Dataset Description

This dataset contains synthetic table PDFs spanning three categories:

Category Samples Description

Edge-case tables 43 Stress-test scenarios: empty cells, merged headers… See the full description on the dataset page: https://huggingface.co/datasets/roma2025/synthetic-table-bench.