| Parameter | Value |
|---|---|
| Base model | Qwen/Qwen3-4B-Instruct-2507 |
| Method | QLoRA (r=64, alpha=128, NF4) |
| Training data | 52,000 examples (6 tasks) |
| Epochs | 3 |
| Batch size | 8 (grad_accum=12, effective=96) |
| Learning rate | 2e-4 |
| Seq length | 1536 |
| Hardware | H100 80GB MIG (4g.40gb) |
| Duration | 7h |
| Final avg loss | 0.187 |
| Task | Count | Description |
|---|---|---|
| T1 NLG | 12,000 | Generate natural Vietnamese doctor questions |
| T2 Diagnosis | 4,000 | Clinical reasoning + differential diagnosis |
| T3 Extraction | 8,000 | Extract symptoms from patient speech |
| T4 Characterization | 10,000 | Extract OPQRST field values |
| T5 Summary | 10,000 | Medical history summarization |
| T6 Safety | 8,000 | Red flag detection + urgency assessment |
