Upload README.md with huggingface_hub

README.md

@@ -1,33 +1,102 @@
 ---
-dataset_info:
-  features:
-  - name: image
-    dtype: image
-  - name: stem
-    dtype: string
-  - name: tier
-    dtype: string
-  - name: doc_type
-    dtype: string
-  - name: gt_text
-    dtype: string
-  - name: gt_json
-    dtype: string
-  - name: structured_gt
-    dtype: string
-  - name: markdown_gt
-    dtype: string
-  - name: geometry_gt
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 98410188
-    num_examples: 93
-  download_size: 97695763
-  dataset_size: 98410188
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: mit
+task_categories:
+  - image-to-text
+language:
+  - en
+tags:
+  - ocr
+  - benchmark
+  - document-understanding
+  - table-extraction
+  - structured-extraction
+size_categories:
+  - n<1K
 ---
+
+# OCR Benchmark
+
+93 real-world documents across 17 document types, benchmarked against 14 OCR engines on 7 evaluation legs.
+
+## What this dataset contains
+
+### Train split (93 rows)
+
+Each row is one benchmark document with:
+
+| 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 | Ground-truth JSON with `plain_text` and `table_html` fields (serialized JSON string) |
+| `structured_gt` | string | Structured extraction ground truth with `json_schema` and `true_json` (serialized JSON string, variable schema per doc type) |
+| `markdown_gt` | string | Ground-truth markdown (null if not available for this doc) |
+| `geometry_gt` | string | Bounding box ground truth as JSON (null if not available) |
+
+### Supplementary files
+
+Beyond the train split, this repo contains the full benchmark reference run:
+
+- `results/runs/` — JSONL results from the reference benchmark run (14 engines × 93 docs)
+- `results/ocr_outputs/` — Raw OCR output per (engine, document, pass)
+- `results/schema_judge_cache/` — Cached LLM judge verdicts for schema/key_value legs
+- `results/table_judge_cache/` — Cached LLM judge verdicts for tables leg
+- `results/markdown_judge_cache/` — Cached LLM judge verdicts for markdown leg
+- `results/judge_checklists/` — Cached document judge checklists (silver-GT docs)
+- `documents/manifest.json` — Corpus manifest with source dataset provenance
+- `documents/gt_limitations.json` — Known ground-truth limitations
+
+Download supplementary files with:
+
+```python
+from huggingface_hub import snapshot_download
+snapshot_download("ilsilfverskiold/ocr-benchmark", repo_type="dataset",
+                  allow_patterns="results/**", local_dir=".")
+```
+
+## Document types
+
+| Tier | Document types | Count |
+|---|---|---|
+| 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 |
+
+## The 7 benchmark legs
+
+| Leg | Question | Docs with GT |
+|---|---|---|
+| text | Did it read the words? | 82 |
+| markdown | Did it preserve structure? | 57 |
+| tables | Did it recover the table? | 40 |
+| key_value | Did it extract label→value pairs? | ~60 |
+| schema | Given a JSON schema, did it fill it correctly? | 65 |
+| geometry | Does it know where each word is? | 17 |
+| judge | Overall page capture (silver-GT docs) | 11 |
+
+## The 14 engines
+
+tesseract, docling, textract_text, textract_structured, mistral_ocr_4, llamaparse, gemini_flash, claude_sonnet_4_6, mineru, glm_ocr, paddleocr_vl, marker, deepseek_ocr, qwen3_vl
+
+## Usage with the benchmark harness
+
+```bash
+git clone https://github.com/ilsilfverskiold/ocr-benchmark
+cd ocr-benchmark
+pip install -r requirements.txt
+python scripts/download_data.py
+python run_benchmark.py --report 20260810T130609Z
+```
+
+## Ground-truth provenance
+
+- **Human GT (75 docs):** Source dataset annotations
+- **Model-verified GT (8 docs):** Model-transcribed, cross-checked against 14-engine consensus
+- **Silver GT (11 docs):** Machine-generated (Tesseract), evaluated by LLM judge only
+
+## License
+
+Code is MIT. Documents are excerpts from third-party datasets — see the benchmark repo for per-dataset licensing.