ilyali034/rubert-russian-financial-sentiment

RuBERT Russian Financial Sentiment

A fine-tuned RuBERT model for 3-class sentiment classification of Russian-language financial and economic text (negative / neutral / positive).

This model fine-tunes blanchefort/rubert-base-cased-sentiment-rusentiment on the SentiRuEval2016 dataset, adapting a general-purpose Russian sentiment model to the financial domain.

Model Details

• Base model: blanchefort/rubert-base-cased-sentiment-rusentiment (RuBERT)
• Architecture: BERT for Sequence Classification
• Language: Russian
• Task: 3-class sentiment classification
• Training dataset: mteb/SentiRuEval2016
• Max sequence length: 128 tokens
• License: Apache 2.0

Labels

Label ID	Sentiment
0	Negative
1	Neutral
2	Positive

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_name = "ilyali034/rubert-russian-financial-sentiment"
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

text = "Акции компании выросли на 15% после публикации отчетности"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predicted_class_id = torch.argmax(outputs.logits, dim=-1).item()

id2label = {0: "negative", 1: "neutral", 2: "positive"}
print(f"Predicted sentiment: {id2label[predicted_class_id]}")

Or with the pipeline API:

from transformers import pipeline

classifier = pipeline("text-classification", model="ilyali034/rubert-russian-financial-sentiment")
result = classifier("Банк объявил о рекордной прибыли за квартал")
print(result)

Training Details

The model was fine-tuned using the Hugging Face Trainer API.

• Epochs: 3
• Batch size: 16 (train and eval)
• Warmup steps: 500
• Weight decay: 0.01
• Best checkpoint selected by: macro F1 on the validation set
• Train/validation split: 85% / 15% (stratified by label), plus the original held-out test split

Labels in the source dataset (-1, 0, 1) were remapped to (0, 1, 2) for negative, neutral, and positive respectively.

Evaluation Results

Evaluated on a held-out test set of 3,000 examples (1,000 per class).

Metric	Score
Accuracy	0.8123
Macro F1	0.8108
Macro Precision	0.8211
Macro Recall	0.8123
Matthews Correlation Coefficient (MCC)	0.7232
Cohen's Kappa	0.7185

Per-Class Metrics

Class	Precision	Recall	F1-Score	Support
Negative	0.8505	0.8820	0.8660	1000
Neutral	0.8277	0.7780	0.8021	1000
Positive	0.7908	0.8090	0.7998	1000
Macro Avg	0.8230	0.8230	0.8226	3000
Weighted Avg	0.8230	0.8230	0.8226	3000

Confusion Matrix

	Predicted Negative	Predicted Neutral	Predicted Positive
Actual Negative	882	49	69
Actual Neutral	77	778	145
Actual Positive	78	113	809

The model performs best on negative sentiment (F1 = 0.866). The most common confusion is between neutral and positive text, which is typical for financial language where neutral factual statements and mildly positive statements can overlap.

Intended Use

This model is intended for sentiment analysis of Russian-language financial and economic content such as:

• News headlines and articles about companies, markets, and the economy
• Social media posts and comments discussing financial topics
• Analyst commentary and investor discussions

Limitations

• Trained on general Russian sentiment data (SentiRuEval2016), not a purpose-built financial corpus, so domain-specific nuance (e.g. sarcasm, technical jargon, mixed-sentiment reports) may not be fully captured.
• Maximum input length is 128 tokens; longer documents should be split or truncated.
• As with any sentiment classifier, predictions should not be used as the sole basis for financial decision-making.

Citation

If you use this model, please cite the base model and dataset:

@misc{rubert-financial-sentiment,
  author = {ilyali034},
  title = {RuBERT Russian Financial Sentiment},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ilyali034/rubert-russian-financial-sentiment}
}

Base model: blanchefort/rubert-base-cased-sentiment-rusentiment Dataset: mteb/SentiRuEval2016