Balance positives during precision correction

scripts/build_failure_driven_yolo_sampling.py

@@ -35,10 +35,11 @@ def build_sampling(
     assessment: dict[str, Any],
     positive_repeat: int = 3,
     negative_repeat: int = 4,
+    precision_positive_repeat: int = 1,
 ) -> tuple[list[str], dict[str, Any]]:
     if assessment.get("status") != "continue_training_loop":
         raise ValueError("Failure-driven sampling requires a failed assessment")
-    if positive_repeat < 1 or negative_repeat < 1:
+    if positive_repeat < 1 or negative_repeat < 1 or precision_positive_repeat < 1:
         raise ValueError("Repeat factors must be positive")
 
     samples = {item["sample_slug"]: item for item in manifest["samples"]}
@@ -78,6 +79,10 @@ def build_sampling(
         repeat = 1
         if tile["label_count"] > 0 and region in weak_recall_regions:
             repeat = positive_repeat
+        elif tile["label_count"] > 0 and region in weak_precision_regions:
+            # Precision-only correction still needs positive examples to avoid
+            # shifting the classifier toward background and sacrificing recall.
+            repeat = precision_positive_repeat
         if tile["label_count"] == 0 and (background_failed or region in weak_precision_regions):
             repeat = negative_repeat
         path = str(Path(tile["image_path"]).resolve())
@@ -97,6 +102,7 @@ def build_sampling(
         "background_gate_failed": background_failed,
         "positive_repeat": positive_repeat,
         "negative_repeat": negative_repeat,
+        "precision_positive_repeat": precision_positive_repeat,
         "source_train_tile_count": sum(
             1
             for tile in summary["tiles"]
@@ -119,6 +125,7 @@ def main() -> int:
     parser.add_argument("--output-dir", type=Path, required=True)
     parser.add_argument("--positive-repeat", type=int, default=3)
     parser.add_argument("--negative-repeat", type=int, default=4)
+    parser.add_argument("--precision-positive-repeat", type=int, default=1)
     args = parser.parse_args()
 
     summary = json.loads(args.summary.read_text(encoding="utf-8"))
@@ -130,6 +137,7 @@ def main() -> int:
         assessment=assessment,
         positive_repeat=args.positive_repeat,
         negative_repeat=args.negative_repeat,
+        precision_positive_repeat=args.precision_positive_repeat,
     )
     args.output_dir.mkdir(parents=True, exist_ok=True)
     train_list = args.output_dir / "train-failure-driven.txt"