Qwen3.5-0.8B Risk Binary Classifier v2
LoRA-finetuned Qwen3.5-0.8B for binary risk classification: low (risk 1-2) vs high (risk 3-5). Uses 50 latent reasoning tokens before output.
Evaluation PlotsTest Set Performance
Threshold (0.0207) chosen on val set for 100% recall with 10% buffer.
| Metric | Value |
|---|
| Accuracy | 95.1% |
| Precision | 76.2% |
| Recall | 97.1% |
| F1 | 0.854 |
| AUC-ROC | 0.995 |
| False Positives | 52 |
| False Negatives | 5 |
Hyperparameter Sweep
Best of 36 configs (3 LRs x 3 LoRA ranks x 2 epochs x 2 oversample ratios). Selected by lowest FP at 100% recall on val.
| Parameter | Value |
|---|
| Learning rate | 5e-4 |
| LoRA rank | 64 |
| LoRA alpha | 128 |
| Epochs | 4 |
| Oversample ratio | 1:2 |
| Latent tokens | 50 |
| Base model | Qwen/Qwen3.5-0.8B |
Training Data
9,316 training examples (original + 854 synthetic risk 4-5), oversampled to 1:2 high:low ratio. Stratified 80/10/10 split.
