| --- |
| license: agpl-3.0 |
| library_name: onnxruntime |
| tags: |
| - bittensor |
| - object-detection |
| - ppe |
| - onnx |
| --- |
| |
| # SN44 SafetyVision PPE miner |
|
|
| This package targets `manak0/Detect-worksite-ppe-compliance-detection-probe-set` |
| on Score/TurboVision SN44. It wraps a 320px ONNX export of the SafetyVision v1 |
| YOLOv8n model and maps its four relevant labels into the active manifest order: |
|
|
| 1. safety helmet |
| 2. high-visibility safety vest |
| 3. protective gloves |
| 4. protective eyewear |
|
|
| The runtime confidence threshold is 0.10, selected on the public |
| Construction-PPE proxy set using the subnet's released 60% mAP50 / 40% |
| false-positive formula. Proxy results are documented in |
| `../sn44-opportunity-2026-10-02.md`. They do not establish performance on the |
| hidden `synthetic_r2` challenges. |
|
|
| The 12.15 MB model was exported from the upstream v1 checkpoint at 320px. It |
| was tested with Ultralytics 8.3.190 and ONNX Runtime 1.26.0, the versions |
| resolved or pinned by the official TurboVision compliance container. Model source: |
| `ayushgupta7777/safetyvision-yolov8` (`best.pt`, v1). The model card states |
| that the weights inherit the Ultralytics YOLOv8 AGPL-3.0 license. |
|
|
| The miner uses direct NumPy/OpenCV preprocessing and postprocessing rather than |
| importing Ultralytics or PyTorch at runtime. It initializes ONNX Runtime in |
| sequential mode with one intra-op thread. This keeps it inside the compliance |
| worker's 16-process/thread resource limit as well as its two-vCPU latency |
| envelope. |
|
|
| This repository contains the immutable model package used by an SN44 mining |
| trial. Deployment and on-chain registration are operational state and are not |
| asserted by this model card. |
|
|
