metrika-labs/PP-OCRv6_medium_rec_cyrillic_safetensors

Fast OCR models for printed and handwritten Cyrillic text, based on PP-OCRv6

Results on challenging OCR benchmarks featuring distortions, printing defects, and low-quality images.

Russian-language OCR benchmark: ru-ocr-benchmark-hard

Usage

🚀 Recommended for CPU (RapidOCR / ONNX)

• PP-OCRv6 Medium Cyrillic (RapidOCR)
• PP-OCRv6 Tiny Cyrillic (RapidOCR)

🚀 Recommended for GPU (PaddlePaddle)

• PP-OCRv6 Medium Cyrillic
• PP-OCRv6 Tiny Cyrillic

Installation

pip install paddleocr torch torchvision transformers

Full pipeline (Detection + Recognition)

from huggingface_hub import snapshot_download
from paddleocr import PaddleOCR

# Download the model
model_dir = snapshot_download(repo_id="metrika-labs/PP-OCRv6_medium_rec_cyrillic_safetensors")

# Initialize OCR with standard PP-OCRv6 detection and Cyrillic recognition
ocr = PaddleOCR(
    text_recognition_model_name="PP-OCRv6_medium_rec",
    text_recognition_model_dir=model_dir,
    text_detection_model_name="PP-OCRv6_medium_det",
    engine="transformers",
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False
)

pipe = getattr(ocr.paddlex_pipeline, "_pipeline", ocr.paddlex_pipeline)
resize_op = pipe.text_rec_model.pre_tfs["ReisizeNorm"]
_orig_resize = resize_op.resize_norm_img
resize_op.resize_norm_img = lambda img, r: _orig_resize(
    img, min(max(r, 1.5 * img.shape[1] / img.shape[0]), resize_op.max_imgW / 48)
)

# Recognize the document
results = ocr.predict("document.jpg")
for res in results:
    for text, score in zip(res["rec_texts"], res["rec_scores"]):
        print(f"{text} ({score:.4f})")

Details

PP-OCRv6 is a family of fast, lightweight text recognition models from PaddlePaddle. The models were trained by Metrika Labs on 7.5 million varied examples of printed and handwritten text from real document scans and synthetic images.

80% of the examples are in Russian. The other 20% are in Kazakh, Belarusian, Uzbek, Tajik, Ukrainian, Bulgarian, Serbian, Macedonian, Tatar, Kyrgyz, Mongolian, Bashkir, Ossetian, Chuvash, Chechen, Yakut, Avar, and other languages that use Cyrillic.

The models work with PaddleOCR and are also available in ONNX format for fast CPU inference with RapidOCR.

The current version supports handwritten text only in Russian. It reads clear handwriting well, but may make mistakes with cursive or less legible handwriting.

PP-OCRv6 Medium: Мир меняется быстро (score: 0.9882)
PP-OCRv6 Tiny: Пир меняется быстро (score: 0.8412)

Links

Benchmark for evaluating OCR model quality on Russian text: metrika-labs/ru-ocr-benchmark-hard

Some handwriting examples come from AntiplagiatCompany/HWR200

Original PaddleOCR repository: PaddlePaddle/PaddleOCR
RapidOCR repository: RapidAI/RapidOCR