docs: align data description and refresh example metrics (#12)

* Add files via upload

* Update README.md

Author: AmirhosseinHonardoust

README.md

@@ -457,15 +457,15 @@ Example results from the included harder synthetic-data workflow:
 | Metric | Example value |
 |---|---:|
 | ROC-AUC | 0.976 |
-| Average precision / PR-AUC | 0.843 |
-| Brier score | 0.052 |
-| Selected threshold | 0.30 |
-| Precision at selected threshold | 0.514 |
-| Recall at selected threshold | 0.961 |
-| Flagged rate | 0.206 |
+| Average precision / PR-AUC | 0.838 |
+| Brier score | 0.061 |
+| Selected threshold | 0.35 |
+| Precision at selected threshold | 0.481 |
+| Recall at selected threshold | 0.974 |
+| Flagged rate | 0.223 |
 </div>
 
-> These values are from a synthetic demo dataset and should not be interpreted as real-world fraud detection performance.
+> These values are from a synthetic demo dataset and should not be interpreted as real-world fraud detection performance. Exact numbers are reproducible with `--seed 42` but may shift slightly across scikit-learn/NumPy versions.
 
 ---
 
@@ -628,6 +628,7 @@ Potential next improvements:
 - pandas
 - NumPy
 - scikit-learn
+- SciPy
 - SHAP
 - Streamlit
 - matplotlib

app.py

@@ -182,7 +182,7 @@ def main():
 """)
 
     st.warning(
-        "This dashboard uses a synthetic, highly separable demo dataset. "
+        "This dashboard uses a synthetic demo dataset with overlapping classes and label noise. "
         "It demonstrates fraud-risk workflow design, not real-world fraud benchmark performance."
     )
 

src/dashboard_utils.py

@@ -83,8 +83,8 @@ def build_model_metadata(
         "threshold_recall": best_threshold.get("recall"),
         "threshold_precision": best_threshold.get("precision"),
         "data_note": (
-            "Synthetic, highly separable demo data. Metrics validate the workflow and should not "
-            "be treated as real-world fraud benchmark performance."
+            "Synthetic demo data with overlapping classes and label noise. Metrics validate "
+            "the workflow and should not be treated as real-world fraud benchmark performance."
         ),
     }
 

src/train_model.py

@@ -69,8 +69,9 @@ def train_and_evaluate() -> dict:
         "positive_rate_train": float(y_train.mean()),
         "positive_rate_test": float(y_test.mean()),
         "note": (
-            "This synthetic demo dataset is highly separable. Probability metrics should be "
-            "interpreted as workflow checks, not real-world fraud benchmark performance."
+            "This synthetic demo dataset has overlapping classes and injected label noise. "
+            "Probability metrics should be interpreted as workflow checks, not real-world "
+            "fraud benchmark performance."
         ),
     }