Galal123/team-180-rasad

Rasad (رصد) track classifier

Fine-tuned Laya that reads a sensor track described in plain English and answers drone, bird, aircraft or unknown with a calibrated probability. Used live in the Rasad app.

Held-out test set, 23,358 simulated tracks:

Classifier	Accuracy	Macro F1	ECE
Rasad (this model)	95.6%	0.90	0.005
Laya base, no fine-tuning	3.1%	0.04	0.48
Old rule-based classifier	38.6%	0.48	0.033

Trained on synthetic data only; numbers are in simulation. It only classifies; people decide. Details: metrics.json, reliability.png.