petrol-fastwin-v4 (WBF ensemble + post-fusion conf gate)
Iteration of v3. Same two-detector WBF ensemble:
weights/havengo.onnxโ m50 0.901 / fp 0.884 (high recall)weights/superbit.onnxโ m50 0.892 / fp 0.944 (high precision)
v4 change: a post-WBF confidence gate at 0.55. WBF halves the score of any cluster where only one of the two models agreed; those singletons are the boxes that hurt the fp pillar in v3 (fp 0.800 โ 0.767). Gating them out should recover the lost precision while keeping the m50 gain.
Per-model weights 0.45 / 0.55 favor SuperBit slightly.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support