Datasets:
license: mit
task_categories:
- image-to-text
language:
- en
tags:
- ocr
- document-understanding
- table-extraction
- structured-extraction
size_categories:
- n<1K
OCR Benchmark — Documents
The 93 document images and ground truth used by the ocr-benchmark harness. The benchmark code, the reference run results, and the full methodology live in the GitHub repo — this dataset is the document corpus only.
Structure
One train split, 93 rows, one row per document:
| Column | Type | Description |
|---|---|---|
image |
Image | The document page (PNG/JPG) |
stem |
string | Filename stem (e.g. invoice_000) |
tier |
string | Difficulty: easy, medium, or hard |
doc_type |
string | Document type (e.g. invoice, receipt, tax_form) |
gt_text |
string | Ground-truth plain text (null for 11 silver-GT docs) |
gt_json |
string | Serialized JSON — plain_text and table_html fields |
structured_gt |
string | Serialized JSON — json_schema and true_json (schema varies per doc type) |
markdown_gt |
string | Ground-truth markdown (null if unavailable) |
geometry_gt |
string | Bounding-box ground truth as serialized JSON (null if unavailable) |
JSON columns are stored as strings because their internal structure varies per
document type (a tax form's schema is not a receipt's schema). Parse with
json.loads().
Document types
| Tier | Types (count each) | Total |
|---|---|---|
| easy | receipt (6), invoice (6) | 12 |
| medium | bank_statement, shipping, tax_form, medical, payslip, lease, photo_receipt (5 each) | 35 |
| hard | form (6), handwritten (6), scanned_legacy (6), financial_table (5), chart (5), newspaper (6), legal (6), report (6) | 46 |
Where the documents come from
Every image is an excerpt from a public dataset. Documents were selected deterministically (fixed row indices), not cherry-picked by content.
| Source | Docs | License | GT provenance |
|---|---|---|---|
| ICDAR2019-SROIE (Huang et al.) | 6 | CC-BY-4.0 | Word annotations (human) |
| invoices-and-receipts_ocr_v1 | 6 | None declared | parsed_data field values (human-labeled) |
| OmniAI OCR Benchmark | 40 | MIT | true_markdown_output (human-annotated) |
| FUNSD (Jaume et al. 2019) | 6 | Non-commercial research/educational use | Word annotations (human) |
| IAM Handwriting (FKI/Univ. Bern) | 6 | Registration-required, non-commercial research | Line transcriptions (human); 15 lines collaged per page |
| IDL-WDS (UCSF Industry Documents) | 6 | Custom "idl-train" license | Legacy OCR annotations (machine-generated) |
| SynFinTabs | 5 | MIT | Exact synthetic cell text |
| RVL-CDIP (Harley et al.) | 18 | Research use | Silver — Tesseract at build time; judge-only evaluation |
FUNSD, IAM, and RVL-CDIP (30 of 93 docs) carry research-use or non-commercial restrictions. This dataset redistributes small excerpts solely for reproducible research benchmarking, with full attribution. If you are a rights holder and want a document removed, open an issue — it will be removed promptly.
Ground-truth provenance
- Human (75 docs) — source dataset annotations
- Model-verified (8 docs) — model-transcribed, cross-checked against 14-engine consensus
- Silver (11 docs) — machine-generated (Tesseract); scored by LLM judge only, never by text metrics
Per-document provenance, known limitations, and the full audit trail are in the
GitHub repo under
documents/ — and every benchmark run prints them before any score.
Usage
Load directly:
from datasets import load_dataset
import json
ds = load_dataset("ilsilfverskiold/ocr-benchmark", split="train")
row = ds[0]
row["image"] # PIL image
gt = json.loads(row["structured_gt"]) # json_schema + true_json
Or use it through the benchmark harness, which rebuilds the on-disk corpus layout and can rerun any engine against these documents:
git clone https://github.com/ilsilfverskiold/ocr-benchmark
cd ocr-benchmark
pip install -r requirements.txt
python scripts/download_data.py # fetches this dataset
python run_benchmark.py --engines tesseract,docling --tiers easy --skip-judge
The reference benchmark results (14 engines × 93 documents × 7 legs) ship with the GitHub repo — you only need this dataset if you want to rerun engines.
License
The images are excerpts from third-party datasets and keep their original licenses — see the source table above. The permissive subset (MIT / CC-BY-4.0) covers 57 of the 93 documents.