Record unique hard-negative YOLO candidate gate
This commit is contained in:
@@ -7,6 +7,15 @@
|
||||
|
||||
# Changelog
|
||||
|
||||
## Sprint 146 Unique hard-negative YOLOv8s candidate gate (2026-07-09)
|
||||
|
||||
- Fixed the all-in-one Docker image so the operator YOLO training wrapper is available at `/app/scripts/train_operator_yolo_detector.sh`.
|
||||
- Trained the Tower-local `geointel-building-yolov8s-uniquehardneg160e50-pt` candidate from the `yolo-building-tile-uniquehardneg160` dataset and preserved it as an explicit local model asset.
|
||||
- Ran 7 positive AOI calibration sweeps, a 17,008-feature evidence portfolio, a 9-sample hard-negative/background matrix and a promotion report.
|
||||
- Result: the candidate is rejected. Mean positive F1 remains around `0.16` and background false-positive pressure reaches `58` detections at threshold `0.25`, `85` at `0.15` and `172` at `0.05`.
|
||||
- Hardened the operator promotion report so older positive evidence portfolios can be compared with explicit positive tile-size/overlap defaults.
|
||||
- No API contract, migration, frontend behavior, model download, provider-fetching behavior or active model configuration changed.
|
||||
|
||||
## Sprint 145 YOLOv8s hardneg r8 e60 full candidate evaluation (2026-07-08)
|
||||
|
||||
- Completed the Tower-local YOLOv8s hard-negative r8 training run through 60 CPU epochs and published local model asset `geointel-building-yolov8s-hardneg160r8e60-pt`.
|
||||
|
||||
Reference in New Issue
Block a user