Upload train_v38_2_pro_v9.py with huggingface_hub

train_v38_2_pro_v9.py

@@ -68,6 +68,21 @@ ACT_LABEL_COLS = [
 ]
 N_ACT_LABEL_PCA = 5  # Reduce 13 labels to 5 PCA components
 
+# Supp V1 LLM score columns
+SUPP_V1_COLS = [
+    'supp_v1_school_specific_program_references',
+    'supp_v1_school_specific_faculty_mentions',
+    'supp_v1_school_specific_campus_features',
+    'supp_v1_prompt_specific_alignment',
+    'supp_v1_personal_connection_to_school',
+    'supp_v1_intellectual_engagement_depth',
+    'supp_v1_extracurricular_alignment',
+    'supp_v1_values_alignment_with_school',
+    'supp_v1_specific_future_contribution',
+    'supp_v1_unique_personal_context',
+    'supp_v1_mean', 'supp_v1_max', 'supp_v1_specificity',
+]
+
 def safe_num(v, default=np.nan):
     if isinstance(v, (int, float)):
         val = float(v)
@@ -494,6 +509,14 @@ PS_RELATED_FEATURES = set([
 # V9+PS_V5: Add PS V5 features to PS_RELATED_FEATURES for proper residualization
 PS_RELATED_FEATURES.update(set(ps5_feature_cols))
 
+# V9+SUPP_V1: Include Supp V1 LLM score features
+supp_v1_in_data = [c for c in SUPP_V1_COLS if c in df.columns]
+STUDENT_LEVEL_NUMERIC.extend(supp_v1_in_data)
+print(f"  Supp V1 features added: {len(supp_v1_in_data)} columns")
+
+# Supp V1 features are school-level (student x school), no special PS/act handling needed
+# They will use default school_mean residualization
+
 # V9: Activity label features need special handling (only has_act_labels=1 for school_mean)
 ACT_LABEL_FEATURES = set(ACT_LABEL_COLS + ['act_label_mean', 'act_label_max', 'act_label_min',
                                              'act_label_std', 'act_label_range'])
@@ -982,11 +1005,17 @@ for i, (fname, imp) in enumerate(fi_pairs[:30]):
     elif 'school_base_rate' in fname: marker = " [S]"
     elif 'ed_boost' in fname: marker = " [ED]"
     elif 'ps2_' in fname: marker = " [PS2]"
+    elif 'ps5_' in fname or fname.startswith('ps_') and 'bert' not in fname: marker = " [PS_V5]"
+    elif 'supp_v1_' in fname: marker = " [SUPP_V1]"
     elif 'act_bert_pca' in fname: marker = " [ACT_BERT]"
     print(f"    {i+1:3d}. {fname:<55s} {imp:>8.2f}{marker}")
 
 # Count V9 new features in top 50
 v9_in_top50 = sum(1 for f, _ in fi_pairs[:50] if 'act_label' in f or f in ACT_LABEL_COLS)
+ps5_in_top50 = sum(1 for f, _ in fi_pairs[:50] if 'ps5_' in f or (f.startswith('ps_') and 'bert' not in f and f in ps5_feature_cols))
+supp_v1_in_top50 = sum(1 for f, _ in fi_pairs[:50] if 'supp_v1_' in f)
+print(f"  PS V5 features in top 50: {ps5_in_top50}")
+print(f"  Supp V1 features in top 50: {supp_v1_in_top50}")
 bert_in_top50 = sum(1 for f, _ in fi_pairs[:50] if 'act_bert_pca' in f)
 print(f"\n  V9 activity label features in top 50: {v9_in_top50}")
 print(f"  act_bert_pca features in top 50: {bert_in_top50}")