---
language:
  - my
  - pi
language_details: "my (Burmese), pi (Pali in Burmese script)"
task_categories:
  - image-to-text
license: cc-by-4.0
tags:
  - ocr
  - burmese
  - pali
  - text-recognition
  - line-images
  - kraken
pretty_name: Burmese Kraken 1M
size_categories:
  - 100K<n<1M
---

# Burmese Kraken 1M

A **Kraken-specific** binary training dataset of **897,883 rendered line images**
for Burmese/Pali text recognition. Stored in Kraken's native PyArrow IPC format
for direct use with `ketos train -f binary` — no extraction or conversion needed.

## Format

This is **not** a general-purpose image dataset. The file is a Kraken binary
arrow file (PyArrow IPC) with images stored inline as PNG bytes and Kraken-specific
metadata (alphabet, split counts, image mode). It loads directly into Kraken's
training pipeline but won't open with standard image viewers or generic dataset
loaders.

**For general-purpose use** (Tesseract, other OCR engines, or standard HF
`load_dataset`), use the companion dataset:
[`pndaza/burmese-ocr-1m`](https://huggingface.co/datasets/pndaza/burmese-ocr-1m),
which contains the same 897,883 images as Parquet shards with a simple
`{image, text, image_id}` schema.

## Dataset Description

- **Total images**: 897,883 line crops
- **Script**: Burmese (Myanmar), including Pali in Burmese script (nissaya)
- **Source**: Text rendered across 27 fonts from a ~714K-line Burmese/Pali
  corpus (literature, spoken, news, wordlists, Buddhist nissaya commentary)
- **Augmentation**: 80% clean / 20% degraded. Degraded lines apply combinations
  of skew, speckle, blur, strokes, page-curve, and perspective to simulate
  real-world scans.
- **Image format**: PNG, bi-level (1-bit), variable width (max ~1766px,
  height normalized to 120px)

## Splits

| split      | rows     |
|------------|---------:|
| train      | 718,515  |
| validation | 89,625   |
| test       | 89,743   |
| **total**  | **897,883** |

## Usage

```bash
# Download
hf download --repo-type dataset pndaza/burmese-kraken-1m burmese_kraken_1m.arrow

# Train (or fine-tune) directly — no preprocessing
ketos --workers 8 --precision 16-mixed train \
  -f binary \
  -F 0.1 \
  -B 16 \
  --lrate 0.001 \
  -N 1 --quit fixed \
  -o output \
  burmese_kraken_1m.arrow
```

### Fine-tuning an existing model

```bash
ketos --workers 8 --precision 16-mixed train \
  -i model.safetensors \
  -f binary \
  -F 0.1 \
  -B 16 \
  --lrate 0.0001 \
  -N 1 --quit fixed \
  -o output \
  burmese_kraken_1m.arrow
```

Use a **10× lower learning rate** (0.0001) when fine-tuning an existing model.

## Notes

- Line images are rendered, not scanned. The degraded subset is intended to
  bridge the domain gap to real scans; it does not replace genuine scan data.
- Alphabet: 118 characters (Burmese consonants, vowels, tones, medials, digits,
  Pali, punctuation).