Oriel β€” AI & technology worth noticing, every day.

llamaindex/ParseBench-2

Dataset Β· https://huggingface.co/datasets/llamaindex/ParseBench-2

ParseBench-2 Quick links: [🌐 Website] [πŸ“œ Paper] [πŸ’» Code] [πŸ“¦ ParseBench-1] ParseBench-2 is a document parsing benchmark that combines broad coverage of real-world documents with severity-aware metrics. It has the following characteristics: Six capability dimensions. Text, formatting, tables, charts, visual grounding (layout) and forms. The new Forms dimension covers 600 form pages. Severity-aware metrics. Each page score is the lower of a parsing score and a severity cap, so… See the full description on the dataset page: https://huggingface.co/datasets/llamaindex/ParseBench-2.

First seen by Oriel on .

AI summary

ParseBench-2 is a document parsing benchmark with 4,317 page-level cases across six capability dimensions, using severity-capped metrics to penalize consequential errors while ignoring benign output variations.

  • The dataset covers 2,645 source documents, including 2,293 new cases targeting complex headers, financial and insurance tables, dense tables, formatting, and government forms.
  • It evaluates text, formatting, tables, charts, layout, and forms; the overall score is the unweighted mean of those six dimension scores.
  • Metrics were calibrated against 330 controlled scenarios and diagnose more than 140 error types involving facts, structure, omissions, hallucinations, and visual grounding.
  • The open-source parse-bench package downloads the dataset, runs pipelines, and generates per-page CSV and HTML reports alongside combined scores.
  • ParseBench-1 annotations and metrics were revised, charts now use complete data tables, and resulting scores are not comparable with ParseBench-1 scores.

#document parsing #benchmarking #severity metrics #hugging face

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days Β· Up to 120 recorded samples per source and metric, with reposts kept separate Β· First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Datasets: 16, latest observation ; +8 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Datasets: 29,809, latest observation ; +29,746 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Datasetsllamaindex/ParseBench-2https://huggingface.co/datasets/llamaindex/ParseBench-2 Read discussion