Dataset card

README.md

@@ -1,21 +1,43 @@
 ---
-dataset_info:
-  features:
-  - name: image
-    dtype: image
-  - name: label
-    dtype: string
-  - name: file_name
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 1168907
-    num_examples: 325
-  download_size: 1149968
-  dataset_size: 1168907
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: mit
+language:
+- km
+task_categories:
+- image-to-text
+tags:
+- ocr
+- khmer-ocr
+- benchmark
+pretty_name: KHOB (325 lines)
+size_categories:
+- n<1K
 ---
+
+# KHOB, 325 text lines
+
+325 Khmer text-line images with their transcriptions, cropped from the [Khmer OCR Benchmark (KHOB)](https://github.com/EKYCSolutions/khmer-ocr-benchmark-dataset) by EKYC Solutions: scanned Cambodian government documents with clean backgrounds and compression artefacts. One image is one line.
+
+This is the KHOB test set in the [Netra OCR](https://github.com/netra-ai-lab/Netra-OCR) results, where the recognizer scores 0.96% CER with beam search and 1.29% greedy.
+
+| Column | |
+| --- | --- |
+| `image` | the line crop |
+| `label` | its text |
+| `file_name` | the original crop's name: task, document and line number |
+
+## Use
+
+```python
+from datasets import load_dataset
+ds = load_dataset("Netra-Lab/khob-eval", split="train")
+```
+
+With the Netra OCR repository, the `khob` preset scores any model on it:
+
+```bash
+bash training/evaluate.sh published --test khob
+```
+
+## License
+
+MIT, as the original KHOB dataset. Credit: EKYC Solutions, [khmer-ocr-benchmark-dataset](https://github.com/EKYCSolutions/khmer-ocr-benchmark-dataset).