bert-base-multilingual-cased-bak-rus-similarity

This is a text pair classifier, trained to predict whether a Bashkir sentence and a Russian sentence have the same meaning.

It can be used for filtering parallel corpora or evaluating machine translation quality.

It can be applied to predict scores like this:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

clf_name = 'slone/bert-base-multilingual-cased-bak-rus-similarity'
clf = AutoModelForSequenceClassification.from_pretrained(clf_name)
clf_tokenizer = AutoTokenizer.from_pretrained(clf_name)

def classify(texts_ba, texts_ru):
    with torch.inference_mode():
        batch = clf_tokenizer(texts_ba, texts_ru, padding=True, truncation=True, max_length=512, return_tensors='pt').to(clf.device)
        return torch.softmax(clf(**batch).logits.view(-1, 2), -1)[:, 1].cpu().numpy()

print(classify(['Сәләм, ғаләм!', 'Хәйерле көн, тыныслыҡ.'], ['Привет, мир!', 'Мама мыла раму.']))
# [0.96345973 0.02213471]

For most "good" sentence pairs, these scores are above 0.5.