Skip to content

Dataset · llamaindex

Real-world document parsing benchmark

llamaindex/ParseBench-2

An English document dataset from LlamaIndex for evaluating text, table, chart, and form recognition systems.

What's inside

Created by the LlamaIndex team to evaluate real-world document parsing, this English dataset contains 1K to 10K examples. Evaluation spans six areas: plain text, formatting, tables, charts, layout structure, and forms (with 600 pages dedicated to forms). Scoring uses error-severity-aware metrics.

What you can do with it

  • Testing multimodal LLMs and PDF document parsers
  • Evaluating extraction quality for tables, charts, and filled-out forms
  • Comparing data preparation pipelines for RAG
Languages
English
Size
1K–10K rows
Downloads this month
45.1K
Likes
49

How to load it

from datasets import load_dataset

ds = load_dataset("llamaindex/ParseBench-2")