anzorq/m2m100_418M_ft_ru-kbd_44K

m2m100_ru_kbd_44K

This model is a fine-tuned version of 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

Framework versions

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

Model inference

1. Install dependencies

pip install transformers sentencepiece torch ctranslate2

2. Inference

CTranslate2 model (quantized model, much faster inference)

First, download the files for the model in ctranslate2 format:

from huggingface_hub import hf_hub_download

hf_hub_download(repo_id='anzorq/m2m100_418M_ft_ru-kbd_44K', subfolder='ctranslate2', filename='config.json', local_dir='./')
hf_hub_download(repo_id='anzorq/m2m100_418M_ft_ru-kbd_44K', subfolder='ctranslate2', filename='model.bin', local_dir='./')
hf_hub_download(repo_id='anzorq/m2m100_418M_ft_ru-kbd_44K', subfolder='ctranslate2', filename='sentencepiece.bpe.model', local_dir='./')
hf_hub_download(repo_id='anzorq/m2m100_418M_ft_ru-kbd_44K', subfolder='ctranslate2', filename='shared_vocabulary.json', local_dir='./')

Run inference:

import ctranslate2
import transformers

translator = ctranslate2.Translator("ctranslate2") # 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))

Vanilla model

1. Install dependencies

pip install transformers sentencepiece

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))