Homework2-llm-toxicity

Toxicity Classifier for Russian Texts

Model description

This model is a fine-tuned version of ai-forever/ru-en-RoSBERTa for binary classification of Russian texts into toxic (1) and non-toxic (0).

It was trained on a balanced dataset of ~74k examples (split 80/10/10) derived from several public sources:

• Toxic: ok.ru comments, inappropriate messages, multilingual toxicity data.
• Non-toxic: voice assistant commands, intent datasets, QA pairs.

Only the classification head was trained; the encoder weights were frozen.

Metrics (on test set)

Confusion matrix: [[3713 3] [ 3 3713]]

How to use

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_name = "RooLeX/Homework2-llm-toxicity"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

def predict_toxicity(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=64)
    with torch.no_grad():
        outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    return int(torch.argmax(probs)), probs[0, 1].item()

# Пример
print(predict_toxicity("Ты идиот!"))  # (1, ~0.9998)
print(predict_toxicity("Здравствуйте, чем могу помочь?"))  # (0, ~0.0000)

Training details

• Base model: ai-forever/ru-en-RoSBERTa
• Max sequence length: 64 tokens
• Batch size: 64
• Learning rate: 2e-4
• Optimizer: AdamW
• Scheduler: CosineAnnealingWarmRestarts
• Early stopping with patience=3 on validation loss

Limitations

• The model was trained on a limited set of domains (social media, customer support, QA). Performance may degrade on very different styles.
• It works only for Russian language.
• May misinterpret sarcasm or cultural references.

Authors

RooLeX

Contact

Hugging Face: RooLeX