Upload app.py with huggingface_hub

app.py

@@ -278,17 +278,37 @@ def _normalize_label(val):
 
 
 def compute_eval(task: str):
-    """Compute confusion matrix + macro F1 for the small demo set."""
+    """Compute confusion matrix + macro F1 with balanced sampling per class."""
     predictor = load_predictor()
     y_true, y_pred = [], []
 
-    max_samples = min(len(raw_samples), 200)  # keep eval lightweight in Spaces
-    for sample in raw_samples[:max_samples]:
+    # Balanced sampling per class
+    rng = random.Random(42)
+    per_class_target = 100
+
+    def class_key(sample):
+        if task == "comm":
+            return _normalize_label(sample["tech"])
+        return _normalize_label((sample["snr"], sample["mob"]))
+
+    buckets = {}
+    for s in raw_samples:
+        key = class_key(s)
+        buckets.setdefault(key, []).append(s)
+
+    selected = []
+    for key, items in buckets.items():
+        rng.shuffle(items)
+        take = min(per_class_target, len(items))
+        selected.extend(items[:take])
+
+    rng.shuffle(selected)
+
+    for sample in selected:
         spec = _to_tensor(sample["data"])
         try:
             res = predictor.predict(spec, return_routing=True)
         except Exception as exc:
-            # Skip problematic samples but keep going
             print(f"[WARN] predict failed: {exc}")
             continue