USER2 Small Sentiment Full All Data
Compact Russian sentiment classifier for the Sentence Analysis ITMO project.
- Base model:
deepvk/USER2-small
- Dataset:
MonoHime/ru_sentiment_dataset
- Training: full fine-tune on all available train rows
- Labels:
neutral, positive, negative
- Prefix:
classification:
- Version:
user2-small-full-all-dynpad-v1
Metrics
Validation split from the full MonoHime training contour:
- Accuracy:
0.8234
- Macro F1:
0.8144
- Eval samples:
22787
- Epoch reached:
1.72
Artifacts
model.onnx - FP32 ONNX export, about 132 MB.model.int8.onnx - dynamic INT8 ONNX export, about 33 MB.tokenizer.json, tokenizer_config.json, special_tokens_map.json - tokenizer artifacts.model_config.json - production metadata for ml-service.
Compared with user2-small-full-dynpad-v1 (macro_f1≈0.7919), this all-data run is the stronger production candidate.
