DUSUNEN-Oku-62M-v1

DUSUNEN Oku 62M — Turkish Printed Word OCR

A compact Turkish OCR recognizer for cropped printed-word images, with full support for ç, ğ, ı, İ, ö, ş, ü and measured results on unseen words.

Oku is a 61.6M-parameter TrOCR model. It takes a cropped image containing one printed Turkish word and generates the transcription. Pair it with a text detector when processing full documents.

Try DUSUNEN Oku in the live browser demo

Word-disjoint benchmark

Every vocabulary word occurs in only one dataset split. The 12,500-image test split therefore measures recognition of words unseen during fine-tuning. Both models use identical preprocessing and greedy decoding.

• Relative CER reduction: 99.73%
• Case-insensitive exact accuracy: 98.48%
• Word error rate: 1.61%
• Invalid generated-token rate: 0.0000%

Machine-readable baseline and final results are included in benchmark-baseline.json and benchmark-final.json. reload-validation.json records a second evaluation after loading the saved checkpoint from disk.

Use

from PIL import Image
import torch
from transformers import TrOCRProcessor, VisionEncoderDecoderModel

model_id = "GoktugD/DUSUNEN-Oku-62M-v1"
processor = TrOCRProcessor.from_pretrained(model_id)
model = VisionEncoderDecoderModel.from_pretrained(model_id)

image = Image.open("cropped_turkish_word.png").convert("RGB")
pixel_values = processor(images=image, return_tensors="pt").pixel_values
with torch.inference_mode():
    generated = model.generate(pixel_values, max_length=32)
print(processor.batch_decode(generated, skip_special_tokens=True)[0])

Training and reproducibility

• Base revision: 04e994ab854b0089d4929f48c2b4dbe2ce78a340
• Dataset revision: 79919774c595fcaa0be2feda583fec4cca6f5d2a
• Data: 225,000 train / 2,500 validation / 12,500 test
• Optimizer: AdamW, linear schedule, 5% warmup, 3e-5
• Effective batch size: 32; one epoch; BF16; seed 42
• Image size: 384×384; maximum target length: 32 tokens
• Hardware: NVIDIA GeForce RTX 5060 Laptop GPU 8 GB
• Recorded training time: 2940.4 seconds
• Exact training program: train_oku_ocr.py

Scope and limitations

• The benchmark images are synthetic. It does not establish accuracy on camera photos, handwriting, scans, receipts, historical print or arbitrary fonts.
• Training labels contain Turkish letters only: no digits, punctuation or multi-word lines. Use Oku on cropped word images, not full pages.
• OCR output can be wrong. Human review is required for legal, medical, financial or identity documents.
• The training dataset is CC BY-SA 4.0. This derivative release preserves that license; see LICENSE for attribution and upstream notes.

Integrity

model.safetensors SHA-256: 601e981bb45c0cf56edd0441097c151c7855a60a738ff8e0d3db493dadacdffe

Developed and released by Göktuğ Düşünen.