README.md

---
license: mit
language:
- ru
- kbd
datasets:
- anzorq/kbd-ru
widget:
- text: Я иду домой.
  example_title: Я иду домой.
- text: Дети играют во дворе.
  example_title: Дети играют во дворе.
- text: Сколько тебе лет?
  example_title: Сколько тебе лет?
- text: На следующий день мы отправились в путь.
  example_title: На следующий день мы отправились в путь.
tags:
- translation
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# m2m100_ru_kbd_44K

This model is a fine-tuned version of [facebook/m2m100_418M](https://huggingface.co/facebook/m2m100_418M) on a ru-kbd dataset, containing 44K sentences from books, textbooks, dictionaries etc..
It achieves the following results on the evaluation set:
- Loss: 0.9399
- Bleu: 22.389
- Gen Len: 16.562

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3.0

### Training results

| Training Loss | Epoch | Step  | Validation Loss | Bleu    | Gen Len |
|:-------------:|:-----:|:-----:|:---------------:|:-------:|:-------:|
| 2.2391        | 0.18  | 1000  | 1.9921          | 7.4066  | 16.377  |
| 1.8436        | 0.36  | 2000  | 1.6756          | 9.3443  | 18.428  |
| 1.63          | 0.53  | 3000  | 1.5361          | 10.9057 | 17.134  |
| 1.5205        | 0.71  | 4000  | 1.3994          | 12.6061 | 17.471  |
| 1.4471        | 0.89  | 5000  | 1.3107          | 14.4452 | 16.985  |
| 1.1915        | 1.07  | 6000  | 1.2462          | 15.1903 | 16.544  |
| 1.1165        | 1.25  | 7000  | 1.1917          | 16.3859 | 17.044  |
| 1.0654        | 1.43  | 8000  | 1.1351          | 17.617  | 16.481  |
| 1.0464        | 1.6   | 9000  | 1.0939          | 18.649  | 16.517  |
| 1.0376        | 1.78  | 10000 | 1.0603          | 18.2567 | 17.152  |
| 1.0027        | 1.96  | 11000 | 1.0184          | 20.6011 | 16.875  |
| 0.7741        | 2.14  | 12000 | 1.0159          | 20.4801 | 16.488  |
| 0.7566        | 2.32  | 13000 | 0.9899          | 21.6967 | 16.681  |
| 0.7346        | 2.49  | 14000 | 0.9738          | 21.8249 | 16.679  |
| 0.7397        | 2.67  | 15000 | 0.9555          | 21.569  | 16.608  |
| 0.6919        | 2.85  | 16000 | 0.9441          | 22.4658 | 16.493  |


### Framework versions

- Transformers 4.21.0
- Pytorch 1.10.0+cu113
- Datasets 2.4.0
- Tokenizers 0.12.1

---

# Model inference
### 1. Install dependencies
```bash
pip install transformers sentencepiece torch ctranslate2
```

### 2. Inference
## Vanilla model
```Python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_path = "anzorq/m2m100_418M_ft_ru-kbd_44K"  
tgt_lang="zu"

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSeq2SeqLM.from_pretrained(model_path)

def translate(text, num_beams=4, num_return_sequences=4):
    inputs = tokenizer(text, return_tensors="pt")
    num_return_sequences = min(num_return_sequences, num_beams)

    translated_tokens = model.generate(
        **inputs, forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang], num_beams=num_beams, num_return_sequences=num_return_sequences
    )

    translations = [tokenizer.decode(translation, skip_special_tokens=True) for translation in translated_tokens]
    return text, translations

# Test the translation
text = "Текст для перевода"
print(translate(text))
```

## CTranslate2 model (quantized model, much faster inference)
```Python
import ctranslate2
import transformers

translator = ctranslate2.Translator("ctranslate") # Ensure correct path to the ctranslate2 model directory
tokenizer = transformers.AutoTokenizer.from_pretrained("anzorq/m2m100_418M_ft_ru-kbd_44K")
tgt_lang="zu"

def translate(text, num_beams=4, num_return_sequences=4):
    num_return_sequences = min(num_return_sequences, num_beams)

    source = tokenizer.convert_ids_to_tokens(tokenizer.encode(text))
    target_prefix = [tokenizer.lang_code_to_token[tgt_lang]]
    results = translator.translate_batch(
        [source],
        target_prefix=[target_prefix],
        beam_size=num_beams,
        num_hypotheses=num_return_sequences
    )
    
    translations = []
    for hypothesis in results[0].hypotheses:
        target = hypothesis[1:]
        decoded_sentence = tokenizer.decode(tokenizer.convert_tokens_to_ids(target))
        translations.append(decoded_sentence)
    
    return text, translations

# Test the translation
text = "Текст для перевода"
print(translate(text))
```