Report continuous_f1 alongside mean_iou; document per-event vs run-level contract.

README.md

@@ -73,10 +73,16 @@ Return exactly `multiplicity` intervals per listed label. Echo the exact label s
 
 ## Scoring
 
-Official metrics for this task (no gold-aware snapping):
+No gold-aware snapping. Report **both** layers:
 
-- **Bound mean IoU** — mean per-gold-event IoU after exact cross-label binding and within-duplicate optimal 1:1 assignment
-- **accuracy@0.5** / **accuracy@0.75** — fraction of gold events with IoU ≥ threshold
+| Layer | Metric | Definition |
+|-------|--------|------------|
+| **Per event** | **IoU** | Overlap / union of one gold interval vs its matched prediction (within-label 1:1 assignment; unmatched → 0) |
+| **Run / split** | **mean IoU** | Mean of those per-event IoUs |
+| **Run / split** | **continuous F1** | Same number under this protocol: when the model returns the requested multiplicities, \(N_g = N_p\), so continuous F1 \(= 2S/(N_g+N_p)\) collapses to mean IoU. Emitted as `continuous_f1` (= `mean_iou`) for consistency with free-segmentation continuous F1 |
+| **Run / split** | **accuracy@0.5 / @0.75** | Fraction of gold events with IoU ≥ threshold |
+
+**Takeaway:** IoU is the event-level unit; continuous F1 is the run-level summary of those IoUs (not a different scoring rule). Always quote both names when reporting, and keep accuracy@0.75 as the strict headline next to them.
 
 Shipped code (this repo):
 

localization/score.py

@@ -212,17 +212,31 @@ def score_episode(
 
 
 def metrics_from_rows(rows: Sequence[dict[str, Any]]) -> dict[str, Any]:
+    """Aggregate per-event IoUs into run-level metrics.
+
+    Reporting contract (localization given labels):
+    - Per event: ``iou`` on each gold event after within-label 1:1 assignment.
+    - Run / split: ``mean_iou`` is the mean of those per-event IoUs.
+    - ``continuous_f1`` is the same number under this protocol (Ng = Np when
+      the model returns the requested multiplicities): bound mean IoU ≡
+      continuous F1 under one-to-one binding. Both keys are always emitted.
+    - ``accuracy_0_5`` / ``accuracy_0_75``: fraction of gold events with
+      IoU ≥ threshold.
+    """
     if not rows:
         return {
             "events": 0,
             "mean_iou": None,
+            "continuous_f1": None,
             "accuracy_0_5": None,
             "accuracy_0_75": None,
         }
     ious = [float(row["iou"]) for row in rows]
+    mean_iou = sum(ious) / len(ious)
     return {
         "events": len(rows),
-        "mean_iou": sum(ious) / len(ious),
+        "mean_iou": mean_iou,
+        "continuous_f1": mean_iou,
         "accuracy_0_5": sum(iou >= 0.5 for iou in ious) / len(ious),
         "accuracy_0_75": sum(iou >= 0.75 for iou in ious) / len(ious),
     }