---
tags:
- optical-character-recognition
- image-to-text
- kashmiri
- perso-arabic
- pytorch
language:
- ks
library_name: pytorch
---

# Koshur OCR v2

A compact Kashmiri Perso-Arabic **printed-line** OCR model from **Kash Labs AI**.

## v2.1 results — prefix beam + train-only character LM

- **Beam deduplicated validation CER:** 21.35% (greedy: 22.62%)
- **Beam deduplicated exact match:** 26.20% (greedy: 25.54%)
- **Held-out 720-row CER:** 21.24% (greedy: 22.44%)
- **Absolute held-out CER improvement:** 1.19 percentage points
- **Validation rows:** 920 deduplicated / 1,000 full
- **Best epoch:** 18
- **Training rows:** 100,000 synthetic sentence-line images
- **Beam evaluation:** https://wandb.ai/kash-ai-labs-kash-labs/koshur-pixel-ocr/runs/ctc-v2-prefix-beam-20260729T152530Z

The selected decoder uses width 10, top-12 token pruning, a trigram character LM trained only on the 100,000 training labels, LM weight 0.4, and token bonus 1.5. It was tuned on a fixed 200-row calibration subset and independently improved the remaining 720 rows. Across all 920 rows, deletions fell from 6,859 to 4,615.

## Architecture

`64×2048 grayscale line → CNN (48→96→160→192) → 2-layer bidirectional GRU (256 per direction) → 155-way CTC`

The model has **4,475,931 trainable parameters**, a 154-character inventory, and no temporal feature dropout in the selected v2 control checkpoint. This is a custom PyTorch checkpoint, not a Transformers-format model.

## Versioning

- `v1.0.0` — 10k temporal-dropout checkpoint, 28.13% deduplicated CER
- `v2.0.0` — 100k control checkpoint with greedy decoding, 22.62% deduplicated CER
- `v2.1.0` — same weights plus prefix-beam/trigram decoding, 21.35% deduplicated CER (current)

Use a Hub revision to pin a release.

## Files

- `best-model.pt` — epoch-18 state dictionary, character inventory, architecture metadata, and validation CER
- `model.py` — architecture, preprocessing, CTC decoding, and `recognize_line()`
- `ctc_prefix_beam.py` — tested prefix-beam implementation
- `char-trigram-lm.pkl` — compact train-only character language model
- `beam-eval.json` — calibration, held-out, and full evaluation metrics
- `config.json` — release, training, preprocessing, metrics, and checkpoint checksum

## Local CPU or Apple Silicon inference

```bash
pip install -r requirements.txt
```

```python
from PIL import Image
from model import recognize_line

text, diagnostics = recognize_line(Image.open("single-line.png"))
print(text)
print(diagnostics)
```

Prefix beam is the default. For the original fast decoder:

```python
text, diagnostics = recognize_line(Image.open("single-line.png"), decoder="greedy")
```

The input should be one tightly cropped, upright printed line. It is resized to height 64 and left-aligned on a width-2048 canvas. Inference runs on CPU and does not require a GPU.

## Limitations

- The reported score is an in-domain synthetic validation result, not real-world scan quality.
- Trained on synthetic printed lines, not handwriting or full pages.
- Camera photos should be rotated, deskewed, tightly cropped, and segmented into individual lines.
- Training targets exclude the established Kashmiri diacritic set; the model recognizes base text and does not restore those marks.
- Error rate remains substantial, with deletions the dominant residual error.
- Predictions require review before publication or archival use.