Update app.py

app.py

@@ -59,14 +59,22 @@ assert len(selected_features) == 138, f"Expected 138 features, got {len(selected
 # ---------------------------------------------------------------------------
 # LIME explainer
 # Built ONCE at startup so explanations are reproducible across requests.
-# The training-data argument controls how LIME perturbs features around the
-# input. After MinMax scaling each feature lives in [0,1], so we use a small
-# uniform sample with a FIXED seed — that gives stable, repeatable weights.
-# (If you have a saved sample of real normalized training rows, swap it in
-# here and explanations will reflect the true feature distribution.)
+# Prefers a real normalized training sample (lime_background.joblib). Falls
+# back to seeded uniform noise if that file isn't present (still stable, but
+# less faithful to the true feature distribution).
 # ---------------------------------------------------------------------------
-_rng = np.random.default_rng(seed=42)
-_lime_background = _rng.uniform(low=0.0, high=1.0, size=(500, len(selected_features)))
+try:
+    _lime_background = joblib.load("lime_background.joblib")
+    if _lime_background.shape[1] != len(selected_features):
+        raise ValueError(
+            f"lime_background.joblib has {_lime_background.shape[1]} cols, "
+            f"expected {len(selected_features)}"
+        )
+    print(f"[LIME] Using real training sample: {_lime_background.shape}", flush=True)
+except Exception as e:
+    print(f"[LIME] No usable lime_background.joblib ({e}); falling back to uniform noise.", flush=True)
+    _rng = np.random.default_rng(seed=42)
+    _lime_background = _rng.uniform(low=0.0, high=1.0, size=(500, len(selected_features)))
 
 explainer = LimeTabularExplainer(
     training_data=_lime_background,