Results
Table with columns: Language, Baseline Acc, Fine-tuned Acc, Baseline F1, Fine-tuned F1| Language | Baseline Acc | Fine-tuned Acc | Baseline F1 | Fine-tuned F1 |
|---|
| English | 38.6% | 89.3% | 0.405 | 0.887 |
| Hinglish | 66.7% | 100.0% | 0.572 | 1.000 |
Hinglish test set: 90 examples. English test set: 456 examples.
Model Details
- Base model: unsloth/mistral-7b-v0.3
- Fine-tuning method: LoRA (rank 16, alpha 16)
- Training: Combined English + Hinglish dataset
- Framework: Unsloth + HuggingFace Transformers
- Hardware: Kaggle T4 x2
Use Case
Classifies sentiment in English financial news and Hinglish social media posts about Indian stocks and markets.
Training Data
- English: Financial PhraseBank + financial news headlines
- Hinglish: Synthetically generated from English samples, validated for label consistency
Limitations
- Hinglish results on 90-sample test set — production use needs larger evaluation
- Optimized for Indian stock market context