---
language:
  - my
  - pi
task_categories:
  - image-to-text
license: cc-by-4.0
tags:
  - ocr
  - burmese
  - pali
  - kraken
pretty_name: Burmese Kraken 5M
size_categories:
  - 1M<n<10M
---

# Burmese Kraken 5M

Kraken-format binary arrow file containing **5,104,476** rendered Burmese/Pali
line images for text recognition (OCR) training with Kraken's `ketos train`.

This is the companion to [`pndaza/burmese-ocr-5m`](https://huggingface.co/datasets/pndaza/burmese-ocr-5m)
which provides the same data in general-purpose HF Parquet format with
viewable images in the Dataset Viewer.

## Dataset Description

- **Total images**: 5,104,476 line crops
- **Fonts**: 37 fonts (27 original + 10 additional)
- **Corpus**: ~814K lines from 13 Burmese/Pali text sources (literature,
  spoken, news, wordlists, Buddhist nissaya commentary)
- **Augmentation**: 90% clean / 10% degraded (skew, speckle, curve)
- **Image format**: PNG bytes inline, bi-level (L-mode), height normalized to 48px

## Splits

| split      | rows       |
|------------|-----------:|
| train      | 4,082,086  |
| validation | 511,435    |
| test       | 510,955    |
| **total**  | **5,104,476** |

## Usage

```bash
# Download
hf download --repo-type dataset pndaza/burmese-kraken-5m burmese_kraken_5m.arrow

# Train with Kraken
ketos --workers 8 --precision 16-mixed train \
    -f binary -F 0.1 -B 16 \
    --lrate 0.001 \
    -N 2 --quit fixed \
    -o output_5m \
    burmese_kraken_5m.arrow
```

## Format

Kraken's PyArrow IPC binary format with inline PNG bytes. Schema:
`lines: struct<text, im: binary, language>` with `train/validation/test: bool`
split columns. Metadata includes alphabet (120 chars) and counts.