Sentiment

Fine-tuned xlm-roberta-base for Hungarian sentiment classification.

Model Details

• Base model: xlm-roberta-base
• Task: 3-class sentiment classification (negative / neutral / positive)
• Language: Hungarian
• Training data: ~37K sentences (stratified split from ~46K total)
• Class weighting: Balanced weights applied during training to handle class imbalance

Labels

Label ID	Label	Description
0	negative	Negative sentiment
1	neutral	Neutral sentiment
2	positive	Positive sentiment

Overall Results

Metric	Value
Accuracy	0.8442320225939605
F1 (macro)	0.8387464047460437
F1 (weighted)	0.8435908941071462

Per-Language Results

Language	Samples	Accuracy	F1 (macro)	F1 (weighted)
hun	4603	0.8442	0.8387	0.8436

Usage

from transformers import pipeline

classifier = pipeline("text-classification", model="ringorsolya/Sentiment")

classifier("Ez egy fantasztikus nap!")
# [{'label': 'positive', 'score': 0.95}]

classifier("Szörnyű volt a kiszolgálás.")
# [{'label': 'negative', 'score': 0.92}]

Training Details

• Epochs: 5
• Batch size: 32
• Learning rate: 2e-05
• Weight decay: 0.01
• Warmup ratio: 0.1
• Max sequence length: 128
• FP16: True
• Class weights: [0.8114, 1.1219, 1.1413]