poltextlab/xlm-roberta-large-hungarian-sentiment-v2

Model description

This model is based on XLM-RoBERTa Large and fine-tuned for Hungarian sentiment analysis.
It classifies text into three sentiment categories:

• 0 → Negative
• 1 → Neutral
• 2 → Positive

Training data

The model was trained on a mix of original and synthetically generated Hungarian texts.
Synthetic samples were introduced to improve class balance and robustness.

Performance

Overall metrics

• Accuracy: 0.8530
• Precision: 0.8449
• Recall: 0.8530
• F1 Score: 0.8469

Classification Report

Label	Precision	Recall	F1-score	Support
0 (Negative)	0.89	0.92	0.91	1866
1 (Neutral)	0.65	0.50	0.56	583
2 (Positive)	0.86	0.92	0.89	1341

Metric	Precision	Recall	F1-score	Support
Accuracy	–	–	0.85	3790
Macro avg	0.80	0.78	0.79	3790
Weighted avg	0.84	0.85	0.85	3790

Confusion Matrix

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

model_name = "poltextlab/xlm-roberta-large-hungarian-sentiment-v2"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

clf = pipeline("text-classification", model=model, tokenizer=tokenizer)

print(clf("Ez egy fantasztikus nap volt!"))  # expected: Positive
print(clf("Ez borzalmas élmény volt."))     # expected: Negative

License

MIT License

Limitations and notes

• The neutral class performs weaker compared to the positive and negative classes.

• Performance may vary depending on text length, domain, and context.

• Due to the use of synthetic data, some linguistic patterns may be over- or underrepresented.