ru_toxicity_detector

Bert-based classifier (finetuned from rubert-tiny2)

Merged datasets:

• Russian Language Toxic Comments from 2ch.hk and pikabu.ru
• Toxic Russian Comments from ok.ru

The datasets split into train, val, test splits in 80-10-10 proportion The metrics obtained from test dataset is as follows:

Usage

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

PATH = 'khvatov/ru_toxicity_detector'
tokenizer = AutoTokenizer.from_pretrained(PATH)
model = AutoModelForSequenceClassification.from_pretrained(PATH)

# if torch.cuda.is_available():
#     model.cuda()

model.to(torch.device("cpu"))


def get_toxicity_probs(text):
    with torch.no_grad():
        inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True).to(model.device)
        proba = torch.nn.functional.softmax(model(**inputs).logits, dim=1).cpu().numpy()
    return proba[0]


TEXT = "Марк был хороший"
print(f'text = {TEXT}, probs={get_toxicity_probs(TEXT)}')
# text = Марк был хороший, probs=[0.9940585  0.00594147]

Train

The model has been trained with Adam optimizer, the learning rate of 2e-5, and batch size of 32 for 3 epochs