petrol-trained-v2 (top-scored consensus distillation)
Fine-tuned YOLOv11s trained from COCO init on 134 high-quality
challenges — only those where the best public miner achieved
composite ≥ 0.65 on the real validator. Boxes from candidate_gt/*.json
used as labels (consensus from top-scoring miners).
- Final val mAP@0.5: 0.9337 (vs trained-v1's 0.7525)
- Architecture: YOLOv11s, 9.4M params, 21.6 GFLOPs
- ONNX FP16, 19.3 MB (≤ 30 MB rule)
- 4 classes: petrol hose, petrol pump, price board, roof canopy
Training: 250-epoch budget, stopped at epoch 125 by patience. Best checkpoint epoch 70 with AdamW, cosine LR, no mixup, mosaic close-out at epoch 230.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support