ClinOCR-Bench

A benchmark for evaluating Optical Character Recognition (OCR) systems on clinical documents. This dataset is presented in the paper ClinOCR-Bench: A Comprehensive Clinical Scanned Document Dataset for Optical Character Recognition Model Evaluation.

📦 This dataset loads directly with datasets (see Usage). The source code, construction scripts, and figures live in the GitHub repository: ClinOCR-Bench, where the benchmark is also published as a downloadable zip of raw files on each release. This card and the repo point to each other.

Overview

Scanned documents have been a headache in healthcare for decades. In the real world they arrive as faxed referrals, crumpled printouts, photographed forms, handwritten notes, and dense tabular reports — often combining several of these challenges at once. ClinOCR-Bench captures this richness of artifacts in a single, systematically organized benchmark so that OCR systems can be evaluated across the full range of conditions.

The dataset is split into subsets, each isolating a distinct document artifact (handwriting, poor-quality scans, rotation, and tables), plus a mixed subset that combines multiple artifacts to reflect the most challenging real-world cases. Every document is paired with a clean, human-audited ground-truth transcription.

File Structure

The dataset spans 6 subsets, 16 templates, and 384 documents (64 per subset). Template 1 to 8 are used by the Normal-quality, Handwriting, Poor-quality, and Rotation subsets, while template 9 to 16 are used by the Tables subset. The Mixed artifacts subset includes 4 samples of templates 1 to 16 (N=64).

Document images, grouped by artifact subset:

Subset	Config	Description
Normal-quality	normal	Clean, well-scanned documents
Handwriting	handwriting	Documents rendered with handwriting fonts
Poor-quality	poor	Low-resolution, photographed, crumpled, or degraded scans
Rotation	rotated	Rotated / skewed documents
Tables	tables	Documents dominated by complex tabular layouts
Mixed artifacts	mixed	Documents combining multiple artifacts (e.g., rotation + highlight + low-resolution)

Each row carries the following columns:

Column	Type	Description
doc_id	string	Unique document key, e.g. handwriting_t3_s5
image	image	The scanned document
ground_truth	string	Human-audited transcription
subset	string	Artifact subset
template	int	Template (layout) id
sample	int	Sample id within the template
homo_id	string	(test only) doc_id of the homogeneous one-shot exemplar
hetero_id	string	(test only) doc_id of the heterogeneous one-shot exemplar

Methodology

The dataset was constructed through a multi-stage pipeline that builds realistic synthetic clinical documents and then degrades them to mimic real-world scanning artifacts:

1. Template design — Manually create a template layout based on real-world scanned documents.
2. Visual asset generation — Prompt image-generation models to create logos, barcodes, and medical images to be embedded in the documents.
3. Content generation — Prompt LLMs to create realistic content for each text area (patient information, impression, notes, etc).
4. Styling — Adopt handwriting fonts and adjust table layouts.
5. Physical artifact simulation — Print out, crumple, apply plastic folders, multi-generation copy, reduce resolution, and rotate to introduce real-world scan artifacts.
6. Quality control — Audit and revise through group discussion.

Ground-truth text is extracted programmatically from the source Word documents and then audited. See the GitHub repository for the full construction pipeline and figures.

Train / Test split and one-shot setup

ClinOCR-Bench is designed for both zero-shot and one-shot evaluation. Within every subset, each template contributes a small exemplar pool (the train split) that is held out from scoring, and the remaining documents are the eval set (the test split):

• Exemplar (train). For each template, the lowest-numbered sample is held out as that template's exemplar. This gives 56 exemplars (one per template per subset: 8 each for normal/handwriting/poor/rotated/tables, 16 for mixed).
• Eval (test). Every other document is scored — 328 documents in total.

Each eval document comes with two predefined one-shot demonstrations, so results are directly comparable across systems:

• Homogeneous one-shot — the exemplar of the same template (same layout as the query). This mimics a real-world scenario where many scanned documents are based on the same medical form.

• Heterogeneous one-shot — the exemplar of a different template in the same subset, picked by a cyclic shift that grows with the query's rank so it never collides with the query template. This reflects another scenario where scanned documents are imported from different sources, with diverse layouts.

This supports three standard regimes: 0-shot, 1-shot (homogeneous), and 1-shot (heterogeneous).

Usage

from datasets import load_dataset

# pick a subset: normal | handwriting | poor | rotated | tables | mixed
ds = load_dataset("ClinOCR-Bench/ClinOCR-Bench", "handwriting")
test, exemplars = ds["test"], ds["train"]

# one-shot demos are referenced by id; resolve against the train split
by_id = {row["doc_id"]: row for row in exemplars}
item = test[0]
homo_demo = by_id[item["homo_id"]]      # same-template exemplar
hetero_demo = by_id[item["hetero_id"]]  # sibling-template exemplar

# item["image"], item["ground_truth"] are the query; *_demo carry the shots

License

Released under the MIT License.

Citation

@article{clinocrbench2026,
  title={ClinOCR-Bench: A Comprehensive Clinical Scanned Document Dataset for Optical Character Recognition Model Evaluation},
  journal={arXiv preprint arXiv:2607.03650},
  year={2026}
}