leoosx/roadwork-v1

Roadwork classifier (fine-tuned for Bittensor SN72)

Fine-tuned from natix-network-org/roadwork against the validator augmentation pipeline in natix/utils/image_transforms.py (rotation/flips/JPEG/blur/noise/color distortions).

Training

• 4 epochs, AdamW lr=2e-5 cosine schedule, class weights for the 4:1 label imbalance
• Filter rule from validator's image cache applied to keep only consistent (description, label) pairs

Eval (augmented test set, 2,263 examples)

Metric	Baseline	This model	Cubic reward gain
Accuracy	0.9633	0.9806
MCC	—	0.8888
Reward proxy	0.8564	0.9347	+30.0%