TrOCR Fine-tuned on several datasets used for training Transkribus models for Cyrillic (Slavic Handwriting)

This model is a fine-tuned version of kazars24/trocr-base-handwritten-ru for recognizing handwritten text in Church Slavonic, Russian and Ukrainian.

Model Description

TrOCR (Transformer-based OCR) is a vision-to-text model using a ViT encoder and a causal language model decoder. This version is fine-tuned specifically on Church Slavonic, Russian, and Ukrainian handwriting.

• Base model: microsoft/trocr-base-handwritten via kazars24/trocr-base-handwritten-ru (used as an intermediate checkpoint)
• Fine-tuned on: several datasets used for training Transkribus models for Cyrillic (approx. 6800 images): https://app.transkribus.org/models/public/text/generic-church-slavonic-handwriting-3, https://app.transkribus.org/models/public/text/russian-generic-handwriting-2, https://app.transkribus.org/models/public/text/ukrainian-generic-handwritten-typed
• Intended use: OCR/HTR of handwritten Church Slavonic (Cyrillic), Russian (Cyrillic), Ukrainian (Cyrillic).

Training Data

• Train / Valid / Test split: 60% / 20% / 20%.

Training Details

• Batch size: 16
• Epochs: 10
• Evaluation metric: CER
• Evaluation strategy: steps
• Tokenizer: kazars24/trocr-base-handwritten-ru

Performance

Note: Some ground truth annotations were incorrect or mismatched, which affects the overall CER. The model, however, correctly predicts many visually accurate texts.

Decoding

Decode with no_repeat_ngram_size=0 and max_length=128. The tokenizer splits Cyrillic into single letters or smaller pieces, so a repeat block suppresses ordinary letter sequences that recur within a line. In our tests, decoding without the block gave a lower error rate. A limit of 64 tokens cuts off long lines. In October 2026 both values in generation_config.json were changed (from 3 and 64).

How to Use

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

# Load processor and model
processor = TrOCRProcessor.from_pretrained("cyrillic-trocr/trocr-handwritten-cyrillic")
model = VisionEncoderDecoderModel.from_pretrained("cyrillic-trocr/trocr-handwritten-cyrillic")

# Load and prepare image
image = Image.open("your_image.jpg").convert("RGB")
pixel_values = processor(images=image, return_tensors="pt").pixel_values

# Generate prediction
generated_ids = model.generate(pixel_values, num_beams=4, no_repeat_ngram_size=0, max_length=128)
predicted_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(predicted_text)