kennethge123/qwen3.5-0.8b-risk-binary-classifier-v2

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.

Test 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.