pretty_name: SafeOCR-LabGold
license: apache-2.0
language:
- en
size_categories:
- n<1K
tags:
- clinical-ocr
- healthcare
- synthetic
- verification
- fhir
- benchmark
configs:
- config_name: default
default: true
data_files:
- split: calibration
path: viewer/calibration.parquet
- split: evaluation
path: viewer/evaluation.parquet
SafeOCR-LabGold
SafeOCR-LabGold is the synthetic laboratory-report benchmark used to develop and evaluate the frozen SafeOCR v0.1 verification policy.
Dataset summary
The complete frozen split contains:
| Split | Synthetic records | Rendered documents | Critical fields |
|---|---|---|---|
| Calibration | 12 | 24 | 144 |
| Final evaluation | 24 | 48 | 288 |
| Total | 36 | 72 | 432 |
Each synthetic record is rendered as a clean document and a deterministic corrupted counterpart. Three controlled templates are used: classic, compact, and grid. Every document contains six laboratory fields.
The paper reports the 48-document / 288-critical-field final evaluation split.
Intended use
Use SafeOCR-LabGold for:
- deterministic clinical-OCR verification research;
- regression tests for evidence binding and selective-decision policies;
- reproducible demonstrations of review/abstention behavior;
- studying automation-versus-review trade-offs under the frozen v0.1 benchmark.
Do not use this dataset as evidence of real-world clinical deployment performance. It is synthetic and does not represent the full distribution of real laboratory reports.
Frozen result boundary
In the final evaluation split, the primary OCR baseline produced 0/288 observed errors. SafeOCR accepted 209/288 fields with 0/209 observed errors and routed 79 correct fields to review. Because the ungated baseline also had zero observed errors, this benchmark does not establish a safety benefit from gating.
Data generation and byte-level reproducibility
The benchmark generator is deterministic for a fixed renderer stack:
python scripts/prepare_hf_labgold.py --output dist/SafeOCR-LabGold
The canonical Hugging Face v0.1 publication package is bound by:
manifest_sha256 = 3a3fcdf1e3e045c0b6f8b334457d1e22dbd247a6a556e08235f808143abf5cec;- Python 3.12.13;
- Pillow 12.3.0;
- FreeType 2.14.3;
- JPEG feature version 8.0;
- zlib 1.3.1.zlib-ng;
- platform
win32:AMD64.
SafeOCR's corruption pipeline includes a JPEG round-trip. Pillow can therefore produce different corrupted PNG bytes when its underlying JPEG implementation differs across platforms even when the frozen seeds, case identities, layouts, and policy are unchanged. A package whose manifest hash differs from the canonical value above must not be presented as the byte-identical v0.1 publication artifact. For exact reproduction of the released benchmark, use the published Hugging Face artifacts or a matching renderer stack.
This renderer dependency does not change the frozen evaluation results or claim boundaries; it is a byte-level reproducibility constraint on regenerated corrupted images.
The generator records exact case identity, template, source seed, corruption seed, document hash, page hash, image filename, and truth-region metadata. The canonical root metadata.jsonl remains byte-identical to manifest.jsonl.
The Hugging Face Dataset Viewer uses a separate derived Parquet layer under viewer/. It is generated only after verifying the canonical manifest and every source-image SHA-256. The viewer layer embeds the same image bytes and exposes the frozen metadata as tabular columns, with explicit calibration (24 rows) and evaluation (48 rows) splits. It is a convenience representation only; manifest.jsonl and images/ remain the audit source of truth.
Privacy
All records are synthetic. Synthetic patient names and identifiers are generated by SafeOCR; no real patient data are used.
Repository and demo
- Code and frozen evidence: https://github.com/AbdulazizShehri/SafeOCR
- Interactive verification playground: https://huggingface.co/spaces/MedScaleAI/SafeOCR
- Paper: arXiv identifier will be added at publication
Dataset contents
The published repository contains:
images/*.png: canonical synthetic laboratory report pages;manifest.jsonl: canonical audit manifest for the frozen release;metadata.jsonl: root metadata byte-identical to the canonical manifest;images/metadata.jsonl: compatibility ImageFolder metadata adapter;viewer/calibration.parquet: derived 24-row viewer split with embedded canonical images and metadata;viewer/evaluation.parquet: derived 48-row viewer split with embedded canonical images and metadata;viewer/viewer_info.json: viewer-layer counts and SHA-256 identities;dataset_info.json: canonical release counts and SHA-256 identities;README.md: this dataset card.
External datasets used in the paper are not redistributed here.
License
The SafeOCR project and generated SafeOCR-LabGold artifacts are released under Apache-2.0 where applicable. Third-party datasets used in the paper retain their own terms and are not included in this Hugging Face dataset.