ProdRadar Sentiment v2 — ruRoBERTa-large
Fine-tuned ai-forever/ruRoBERTa-large (355M params) for Russian sentiment analysis with advanced training techniques.
Metrics (28K test set)
| Metric | Score |
|---|---|
| F1-macro | 0.868 |
| F1-positive | 0.917 |
| F1-negative | 0.918 |
| F1-neutral | 0.770 |
Training
- Data: 318K samples from 4 open-source Russian datasets + 188K proprietary + 10K synthetic
- Techniques: LLRD + Focal Loss + FGM Adversarial Training + R-Drop
- Hardware: NVIDIA L4 GPU, ~4.5 hours
- Labels: positive (0), negative (1), neutral (2)
Usage
from transformers import pipeline
classifier = pipeline("text-classification", model="Daniil125/prodradar-sentiment-ru")
result = classifier("Ужасный сервис!")
- Downloads last month
- 145