Rubert Tiny Sentiment Balanced

🚀 Fine-tuned Model for Russian Text Sentiment Classification

This project presents a fine-tuned model based on cointegrated/rubert-tiny for sentiment classification of short Russian texts. The sentiment classification problem is framed as a multiclass task, distinguishing between negative, neutral, and positive sentiments.

🚀 Quick Start

This is the cointegrated/rubert-tiny model fine-tuned for classification of sentiment for short Russian texts. The problem is formulated as multiclass classification: negative vs neutral vs positive.

✨ Features

• Multiclass Classification: Capable of classifying short Russian texts into three sentiment categories: negative, neutral, and positive.
• Fine - tuned Model: Based on the pre - trained cointegrated/rubert-tiny model, fine - tuned for better performance on sentiment classification tasks.

💻 Usage Examples

Basic Usage

The following Python code demonstrates how to use the model to estimate the sentiment of a given text:

# !pip install transformers sentencepiece --quiet
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_checkpoint = 'cointegrated/rubert-tiny-sentiment-balanced'
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(model_checkpoint)
if torch.cuda.is_available():
    model.cuda()
    
def get_sentiment(text, return_type='label'):
    """ Calculate sentiment of a text. `return_type` can be 'label', 'score' or 'proba' """
    with torch.no_grad():
        inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True).to(model.device)
        proba = torch.sigmoid(model(**inputs).logits).cpu().numpy()[0]
    if return_type == 'label':
        return model.config.id2label[proba.argmax()]
    elif return_type == 'score':
        return proba.dot([-1, 0, 1])
    return proba

text = 'Какая гадость эта ваша заливная рыба!'
# classify the text
print(get_sentiment(text, 'label'))  # negative
# score the text on the scale from -1 (very negative) to +1 (very positive)
print(get_sentiment(text, 'score'))  # -0.5894946306943893
# calculate probabilities of all labels
print(get_sentiment(text, 'proba'))  # [0.7870447  0.4947824  0.19755007]

🔧 Technical Details

Training

We trained the model on the datasets collected by Smetanin. We have converted all training data into a 3 - class format and have up - and downsampled the training data to balance both the sources and the classes. The training code is available as a Colab notebook.

Evaluation Metrics

The metrics on the balanced test set are as follows: