README.md

---
language:
- udm
---

# bert-tiny-char-ctc-udm-denoise

This is a tiny BERT model for Udmurt, intended for fixing OCR errors.

Here is the code to run it (it uses a custom tokenizer, with the code downloaded in the runtime):
```python
import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

MODEL_NAME = 'udmurtNLP/bert-tiny-char-ctc-udm-denoise'
model = AutoModelForMaskedLM.from_pretrained(MODEL_NAME)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)

def fix_text(text, verbose=False, spaces=2):
    with torch.inference_mode():
        batch = tokenizer(text, return_tensors='pt', spaces=spaces, padding=True, truncation=True, return_token_type_ids=False).to(model.device)
        logits = torch.log_softmax(model(**batch).logits, axis=-1)
    decoded = tokenizer.decode(logits[0].argmax(-1), skip_special_tokens=True)
    return tokenizer.clean_up_tokenization(decoded)
fix_text("кыче мои солы оскылй!")
# Кыӵе мон солы оскылӥ!
```

It was trained on a parallel corpus (corrupted + fixed sentence) with CTC loss. On our test dataset, it reduces OCR errors by 50%.

Inspired by https://huggingface.co/slone/bert-tiny-char-ctc-bak-denoise