LazyGreed/khmer_img2txt_200K

---
license: mit
dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
task_categories:
- image-to-text
language:
- km
- en
tags:
- Khmer
- ocr
- english
pretty_name: Khmer Image to Text dataset 200K images
size_categories:
- 100K<n<1M
---

# Dataset Card for Khmer English OCR 200K

## Dataset Summary

Khmer-English OCR 200K is a line-level OCR dataset stored as parquet files with two columns:

- `image`: a struct containing raw image bytes and a `path` field
- `text`: the transcription string for the image

The dataset currently contains:

- `train`: 179,988 examples
- `val`: 19,996 examples

This dataset appears to target OCR training for Khmer and mixed Khmer-English text.

## Dataset Structure

### Data Instances

Each row has the following structure:

```python
{
"image": {
"bytes": b"...",
"path": "train_000256.png",
},
"text": "ហ្គីរ៉ូដ និង ពេជ អង្គ ដែលលើកឡើង",
}
```

### Data Fields

- `image.bytes`: Raw bytes of the image file.
- `image.path`: Original image filename from `labels.txt` (for example `train_000256.png`).
- `text`: UTF-8 transcription text. During parquet creation, carriage returns and newlines are removed.

## Intended Uses

- Training OCR models for mixed Khmer-English line recognition
- Validation and benchmarking for sequence recognition models

## Limitations

- The dataset may contain noisy labels, mixed scripts, punctuation variation, or OCR-unfriendly crops.
- The transcription preprocessing removes newline and carriage return characters.

## Citation

If you publish or share this dataset, add a citation here:

```bibtex
@dataset{kh_en_ocr_200k,
title = {Khmer Image to Text dataset 200K images},
author = {LazyGreed},
year = {2026},
note = {Line-level OCR dataset}
}
```