datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1651 PDF pages, covering 10 document types, 5 layout types and 5 language types. Coverage includes academic literature, research and financial reports, newspapers, textbooks, exam papers, magazines, handwritten notes, historical documents, and more.
Rich Annotations:… See the full description on the dataset page: https://huggingface.co/datasets/opendatalab/OmniDocBench.Real5-OmniDocBench
Real5-OmniDocBench
A Full-Scale Physical Reconstruction Benchmark for Robust Document Parsing in the Wild
Leaderboard | Overview | Dataset | Evaluation | Submit Results | Citation
Real5-OmniDocBench measures the robustness of document parsing systems under five physical acquisition conditions: Scanning, Warping, Screen-Photography, Illumination, and Skew. It reconstructs the same 1,355 pages from OmniDocBench v1.5 in every condition, producing 6,775 images in total. The one-to-one… See the full description on the dataset page: https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench.OmniDocBenchForked from opendatalab/OmniDocBench.
Sampler
We have added a simple Python tool for filtering and performing stratified sampling on OmniDocBench data.
Features
Filter JSON entries based on custom criteria
Perform stratified sampling based on multiple categories
Handle nested JSON fields
Installation
Local Development Install (Recommended)
git clone https://huggingface.co/Quivr/OmniDocBench.git
cd OmniDocBench
pip install -r requirements.txt #… See the full description on the dataset page: https://huggingface.co/datasets/Quivr/OmniDocBench.omnidocbench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 981 PDF pages, covering 9 document types, 4 layout types and 3 language types. It has broad coverage including academic papers, financial reports, newspapers, textbooks, handwritten notes, etc.
Rich Annotations: Contains location information for 15 block-level (text paragraphs… See the full description on the dataset page: https://huggingface.co/datasets/samiuc/omnidocbench.omnidocbench-render-compare
OmniDocBench Render-and-Compare
This dataset contains the rendered HTML reconstructions and comparison images produced
by a render-and-compare pipeline — a reference-free visual similarity evaluation
framework for OCR systems.
Overview
The pipeline processes each page of OmniDocBench through
a Qwen3.5-122B-A10B OCR model, renders the structured output back to a PNG via HTML
(reconstructed.png), and compares it against the original page scan (masked_original.png)
using… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/omnidocbench-render-compare.OmniDocBench_tsvThis repository contains the TSV version of the ground truth data for OmniDocBench, , converted from the dataset available at opendatalab/OmniDocBench.
Note: The main branch now corresponds to version v1_0 of the OmniDocBench ground truth.
OmniDocBench-JASyn
OmniDocBench-JASyn
日本語 | English
OmniDocBench-JASyn is a dataset for evaluating Japanese Document Parsing performance. It inherits the format of OmniDocBench and evaluates the end-to-end performance of VLMs on Japanese documents, covering OCR, layout analysis, table analysis, formula OCR, and reading order recognition.
Features
Fully Synthetic Data : To ensure efficient and high-quality data, images and ground truth are generated by LLM-based code generation… See the full description on the dataset page: https://huggingface.co/datasets/stockmark/OmniDocBench-JASyn.OmniDocBench-v1.6
OmniDocBench
OmniDocBench is a benchmark dataset for evaluating document parsing and document structure understanding systems on diverse document page images. It is designed for evaluating layout detection, reading order prediction, OCR-related text extraction, formula recognition, table recognition, and end-to-end document parsing.
This release contains 1,651 annotated page images and a single JSON annotation file.
Files
.
├── images/
│ ├── *.jpg
│ └── *.png
└──… See the full description on the dataset page: https://huggingface.co/datasets/MinerU25Pro-NIPS26/OmniDocBench-v1.6.Wild_OmniDocBench
Wild-OmniDocBench
A Real-World Captured Document Parsing Benchmark for Robustness Evaluation
中文版 •
Paper •
GitHub •
HuggingFace
Overview
Wild-OmniDocBench is a benchmark for evaluating document parsing robustness under real-world captured conditions. It is derived from OmniDocBench by converting scanned/digital documents into naturally captured images through controlled physical simulation, including printing, deformation, and photography under diverse… See the full description on the dataset page: https://huggingface.co/datasets/VirtualLUO/Wild_OmniDocBench.omnidocbench-render-compare-parquet
OmniDocBench Render-and-Compare — Parquet Edition
Parquet-shard repackaging of
gt-free-ocr-metrics/omnidocbench-render-compare.
Overview
The pipeline processes each page of OmniDocBench
through a Qwen3.5-122B-A10B OCR model, renders the structured output back
to a PNG via HTML (reconstructed), and compares it against the original
page scan (masked_original) using reference-free visual metrics.
Five OCR extraction variants are provided, each targeting a different subset
of… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/omnidocbench-render-compare-parquet.OmniDocBench-parquet
OmniDocBench Parquet
This is a Parquet-format version of the OmniDocBench v1_0 dataset.
Why This Repository Exists
The original OmniDocBench dataset contains thousands of individual image and PDF files. When downloading via huggingface_hub, each file triggers a separate HTTP request, which can lead to HuggingFace rate limiting (HTTP 429 errors) during large-scale downloads.
This Parquet version consolidates all data into a single file, reducing the number of HTTP requests… See the full description on the dataset page: https://huggingface.co/datasets/samiuc/OmniDocBench-parquet.OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1651 PDF pages, covering 10 document types, 5 layout types and 5 language types. Coverage includes academic literature, research and financial reports, newspapers, textbooks, exam papers, magazines, handwritten notes, historical documents, and more.
Rich Annotations: Contains… See the full description on the dataset page: https://huggingface.co/datasets/tejas2102/OmniDocBench.omnidocbench-qwen-ocr-logprobs
OmniDocBench Qwen OCR Log-Probabilities
This dataset provides token-level and bounding-box-level OCR log-probabilities produced by
running Qwen3.5-122B-A10B (via vLLM) on the original page scans of the
OmniDocBench benchmark.
It is a reference-free auxiliary signal — no ground-truth text is used.
Dataset Structure
ocr_logprobs/
<page_id>/
ocr_logprobs.json # full per-token logprobs + top-5 alternatives
ocr_html.html # raw HTML output from the OCR model… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/omnidocbench-qwen-ocr-logprobs.Real5-OmniDocBench
Real5-OmniDocBench
Real5-OmniDocBench is a brand-new benchmark oriented toward real-world scenarios, which we constructed based on the OmniDocBench v1.5 dataset. The dataset comprises five distinct scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. Apart from the Scanning category, all images were manually acquired via handheld mobile devices to closely simulate real-world conditions. Each subset maintains a one-to-one correspondence with the original… See the full description on the dataset page: https://huggingface.co/datasets/JiangWenxiang/Real5-OmniDocBench.OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1355 PDF pages, covering 9 document types, 4 layout types and 3 language types. It has broad coverage including academic papers, financial reports, newspapers, textbooks, handwritten notes, etc.
Rich Annotations: Contains location information for 15 block-level (text paragraphs… See the full description on the dataset page: https://huggingface.co/datasets/chhhhhgrghdu/OmniDocBench.OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1355 PDF pages, covering 9 document types, 4 layout types and 3 language types. It has broad coverage including academic papers, financial reports, newspapers, textbooks, handwritten notes, etc.
Rich Annotations: Contains location information for 15 block-level (text paragraphs… See the full description on the dataset page: https://huggingface.co/datasets/eamon520/OmniDocBench.OmniDocBench_train_cleaned
OmniDocBench_train_cleaned
The OmniDocBench_train__x family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.
images
941
QA turns
5,276
answers rewritten by the cleaning pass
751
QA created by the cleaning pass (new_qa)
4,339 (82.2%)
shards
3
How this was cleaned
A vision-language model read each image together with its QA and judged the item. The pass is
not a filter that only removes rows — it rewrites answers it finds… See the full description on the dataset page: https://huggingface.co/datasets/Elliot-Data/OmniDocBench_train_cleaned.omnidocbench-render-compare-sample
OmniDocBench Render-and-Compare — Sample
This is a 60-page stratified sample of
gt-free-ocr-metrics/omnidocbench-render-compare
(the full dataset is ~10 GB).
It is provided to help reviewers explore the data without downloading the full dataset,
as recommended by the NeurIPS 2025 Datasets & Benchmarks Track guidelines.
Sampling Methodology
Pages were selected by stratified random sampling from the full dataset:
Each page in ocr_all (1 355 pages) was assigned to one… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/omnidocbench-render-compare-sample.english_OmniDocBench_with_evalOmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1651 PDF pages, covering 10 document types, 5 layout types and 5 language types. Coverage includes academic literature, research and financial reports, newspapers, textbooks, exam papers, magazines, handwritten notes, historical documents, and more.
Rich Annotations: Contains… See the full description on the dataset page: https://huggingface.co/datasets/HKolt/OmniDocBench.OmniDocBench_train_cleaned
OmniDocBench_train_cleaned
The OmniDocBench_train__x family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.
images
941
QA turns
5,276
answers rewritten by the cleaning pass
751
QA created by the cleaning pass (new_qa)
4,339 (82.2%)
shards
3
How this was cleaned
A vision-language model read each image together with its QA and judged the item. The pass is
not a filter that only removes rows — it rewrites answers it finds… See the full description on the dataset page: https://huggingface.co/datasets/elliot-mllm/OmniDocBench_train_cleaned.OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1355 PDF pages, covering 9 document types, 4 layout types and 3 language types. It has broad coverage including academic papers, financial reports, newspapers, textbooks, handwritten notes, etc.
Rich Annotations: Contains location information for 15 block-level (text paragraphs… See the full description on the dataset page: https://huggingface.co/datasets/brunodf12/OmniDocBench.omnidocbench-qwen-ocr-logprobs-parquet
OmniDocBench Qwen OCR Logprobs — Parquet Edition
Parquet-shard repackaging of
gt-free-ocr-metrics/omnidocbench-qwen-ocr-logprobs.
Overview
For each of the 1,355 OmniDocBench pages, this dataset contains the full
token-level log-probability stream emitted by Qwen3.5-122B-A10B during OCR
inference, plus a per-bbox aggregation of those logprobs. These are
intended as a reference-free confidence signal that can be combined
with the visual-similarity metrics in… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/omnidocbench-qwen-ocr-logprobs-parquet.OmniDocBench
OmniDocBench
English | 简体中文
OmniDocBench is an evaluation dataset for diverse document parsing in real-world scenarios, with the following characteristics:
Diverse Document Types: The evaluation set contains 1355 PDF pages, covering 9 document types, 4 layout types and 3 language types. It has broad coverage including academic papers, financial reports, newspapers, textbooks, handwritten notes, etc.
Rich Annotations: Contains location information for 15 block-level (text paragraphs… See the full description on the dataset page: https://huggingface.co/datasets/chenweijie/OmniDocBench.OmniDocBench-eval-outputsThis dataset contains benchmark outputs from multiple document understanding models evaluated on the OmniDocBench benchmark suite. Each model was tested on the same dataset, and the outputs are in markdown.
Models/Tools included:
Azure
Textract
Docling
Tensorlake
mineru-omnidocbench-reportrepro-ovisocr-omnidocbench-outputsomnidocbenchocr-render-compare-omnidocbench
OmniDocBench render-and-compare pairs, four OCR engines
Render-and-compare pairs for the 1,355 pages of OmniDocBench v1.5 from four OCR engines: Qwen3.5-122B-A10B (qwen35), MinerU-2.5 (mineru25), PaddleOCR-VL (paddleocr_vl), DeepSeek-OCR (deepseek_ocr). The OCR output of each page is scored in five variants: ocr_text (text elements only), ocr_formula (formulas only), ocr_table (tables only), ocr_all (all three, with regions of other types masked on the original) and… See the full description on the dataset page: https://huggingface.co/datasets/gt-free-ocr-metrics/ocr-render-compare-omnidocbench.