SATYA — Fine-Tuned Gemma 4 E4B for Truth Detection
LoRA adapter fine-tuned for on-device misinformation detection.
Training Data
- 3,232 Indian fact-checks (BOOM Live, Alt News, FACTLY, The Quint, Vishvas News)
- 5,500 LIAR2 political claims
Results (on 437 held-out test examples)
| Metric | Base Gemma 4 | SATYA Fine-Tuned | Improvement |
|---|---|---|---|
| Accuracy | 0.3461 | 0.5370 | +0.1910 |
| F1 (macro) | 0.2306 | 0.3025 | +0.0718 |
Training
- Method: LoRA (rank 16, alpha 32, dropout 0.0) via Unsloth
- Steps: 1000 of 1392 (Kaggle 12-hour cap)
- Hardware: Tesla T4 x2
License: Apache 2.0
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support