ParseBench-2

Quick links: [🌐 Website] [📜 Paper] [💻 Code] [📦 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 even a page that is 99% correct is still identified correctly.
• Calibrated against stated priorities. The metrics are tuned against 330 controlled scenarios in six severity test suites. Each scenario has a target score drop, and the severity levels are critical, major and minor, with benign variations that should cost nothing.
• Over 140 error types. The metrics check for changed facts, broken structural relationships, missing content, invented content and more. The evaluation reports per-error diagnostics for every page.
• 4,317 page-level test cases from 2,645 source documents. The test cases are the revised ParseBench-1 pages plus 2,293 new cases that target hard situations: complex table headers, financial and insurance tables, dense tables, headings, strikeout, underline and government forms.
• Evaluation code suite. The scorers ship with the open-source parse-bench package.

Dataset Introduction

ParseBench-2 keeps most pages from ParseBench-1 and revises their annotations so the metrics can compare complete outputs. It also adds targeted datasets for cases that sparse checks miss.

• Text. Omissions, hallucinations, reading order and changed facts. A flipped negation or a misread digit caps the page score even when the rest of the text is intact.
• Formatting. Formatting that carries meaning: headings and their hierarchy, bold, strikethrough (superseded content) and underline (insertions on redlines), superscript and subscript.
• Tables. Values attached to the correct row and column headers, merged cells and hierarchical headers. A swapped header is critical, while extra HTML markup is a benign variation and costs nothing.
• Charts. Every data point of the chart, with series and axis labels, recovered as a table. An order-of-magnitude error is critical. Abbreviating a series name is minor.
• Layout (visual grounding). Every extracted element traced back to its location and class on the page.
• Forms. Field values, checkboxes and signatures on real government forms. A misread value is critical. Fields left empty must stay empty, and inventing a value for them counts as an error.

The overall score is the unweighted mean of the six dimension scores. Within a dimension, page scores are first averaged per dataset. Each targeted dataset gets 1/8 of the weight and the broad dataset gets the rest.

Dimension	Subset folder	Cases	Weight in dimension
Text	text/	508	1
Formatting	text/	(shared)	3/4
	headings/	100	1/8
	strikeout/	128	1/16
	underline/	125	1/16
Tables	tables/	500	1/2
	complex_table_headers/	149	1/8
	financial_tables/	319	1/8
	insurance_filing_tables/	412	1/8
	dense_tables/	373	1/8
Charts	charts/	568	1
Layout	layout/	535	1
Forms	forms/	500	7/8
	fillable_forms/	100	1/8
Total	13 subsets	4,317

Text and Formatting share the same 508 text/ pages, scored with different metrics.

Usage

The ParseBench-2 scorers are part of parse-bench. One command downloads this dataset, runs a pipeline on every subset and scores the results into the six dimension scores and the overall score:

pip install "parse-bench[runners,fast]"

# See `parse-bench pipelines` for the built-in pipelines
parse-bench run llamaparse_agentic --dataset parsebench-2

# Only some subsets; every subset that already has results is still scored
parse-bench run llamaparse_agentic --dataset parsebench-2 --group charts,tables

The dataset is downloaded to ./data_parsebench-2. Results for each subset are written to output/parsebench-2/<subset>/<pipeline>/, with per-page CSV and HTML reports. The combined score goes to output/parsebench-2/_score/<pipeline>/_next_score.json. A dimension is reported as missing until all of its subsets have been scored.

To score outputs from your own parser, write them in the ParseBench results layout and run parse-bench evaluation run <dir> --next=SUBSET=RESULTS_DIR::TEST_CASES_DIR,... (see docs/extending.md in the ParseBench repo).

Each subset's --group name is also its folder in this repository and its split in the dataset viewer.

Dataset Format

The repository has one folder per subset, in the layout the parse-bench harness reads directly, and one <subset>.jsonl index per subset for browsing:

<subset>/
├── <name>.pdf | .jpg     # source page
├── <name>.test.json      # answer key and test rules for that page
└── <name>.md             # the reference Markdown, where it is a separate file (every text page)
<subset>.jsonl            # one row per page (the dataset viewer's split)

Each case is one page. Names keep their source group as a prefix where a subset has several (text_simple__10k, annotated__0000050863-25-000054_p7, filled__...). A row of <subset>.jsonl has the page's id, subset, dimension, document path, expected_markdown (null where the answer key is held in the rules), test_rules and metadata (both JSON strings, as in the .test.json) and tags.

Differences from ParseBench-1

• Metrics. The v1 metrics (Content Faithfulness, Semantic Formatting, GTRM, ChartDataPointMatch, Element Pass Rate) are replaced by severity-capped metrics: text_next, formatting_v2, table_next, chart_next, layout_grounding_pass_rate and forms_v2.
• Annotations. v1 pages are kept but their references are revised and made more complete. Charts are annotated with full data tables in place of spot checks.
• New datasets. Targeted sets for complex table headers, financial tables, insurance filing (SERFF) tables, dense tables, headings, strikeout, underline and forms.
• New dimension. Forms, scored on 500 real government form pages (typed fills, 10% blank) plus 100 fillable-form pages.
• Scores are not comparable with ParseBench-1 scores.

Submit Results to the Leaderboard

We welcome community contributions to the ParseBench-2 leaderboard. To add a model's scores, open a PR on the model's Hugging Face repo that adds a .eval_results/parsebench-2.yaml file, with one entry per task. The tasks are overall (the unweighted mean of the six dimensions) and the six dimensions text, formatting, tables, charts, layout and forms. Report scores on a 0–100 scale; parse-bench writes them as 0–1 fractions in _next_score.json.

- dataset:
    id: llamaindex/ParseBench-2
    task_id: overall
  value: 66.1
  source:
    url: https://github.com/run-llama/ParseBench
    name: parse-bench
  notes: "Pipeline name: <parse-bench pipeline>"
- dataset:
    id: llamaindex/ParseBench-2
    task_id: tables
  value: 77.6

See the Hugging Face eval-results docs for every field, and this ParseBench-1 PR for a complete example.

Changelog

2026-11-11

• Initial release! :tada:

Copyright Statement

All documents are sourced from public online channels. The dataset is released under the Apache 2.0 License. If there are any copyright concerns, please contact us via the GitHub repository.

Citation

The ParseBench-2 paper, ParseBench-2: Severity-Aware Evaluation of Document Parsing, is coming soon. Until it is out, please cite ParseBench:

@misc{zhang2026parsebench,
  title={ParseBench: A Document Parsing Benchmark for AI Agents},
  author={Boyang Zhang and Sebastián G. Acosta and Preston Carlson and Sacha Bron and Pierre-Loïc Doulcet and Daniel B. Ospina and Simon Suo},
  year={2026},
  eprint={2604.08538},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.08538},
}

Links

• Website: parsebench.ai
• GitHub: run-llama/ParseBench
• ParseBench-1 dataset: llamaindex/ParseBench