📝 Update README: Replace pickle examples with secure JSON format

README.md

@@ -186,19 +186,43 @@ For production ATS scoring, combine the semantic model with the ensemble score m
 
 ```python
 from sentence_transformers import SentenceTransformer
+from sklearn.linear_model import Ridge
+from sklearn.neural_network import MLPRegressor
+from sklearn.preprocessing import PolynomialFeatures
 import numpy as np
-import pickle
+import json
 
 # Load semantic model
 model = SentenceTransformer('0xnbk/nbk-ats-semantic-v1-en')
 
-# Load ensemble weights (download from model repository)
-with open('score_mapper.pkl', 'rb') as f:
-    score_mapper = pickle.load(f)
-with open('neural_mapper.pkl', 'rb') as f:
-    neural_mapper = pickle.load(f)
-with open('poly_features.pkl', 'rb') as f:
-    poly_features = pickle.load(f)
+# Load ensemble weights from JSON (secure format, no pickle warnings)
+with open('ridge_weights.json', 'r') as f:
+    ridge_data = json.load(f)
+with open('neural_weights.json', 'r') as f:
+    neural_data = json.load(f)
+with open('poly_features.json', 'r') as f:
+    poly_data = json.load(f)
+
+# Reconstruct models from JSON
+score_mapper = Ridge(alpha=ridge_data['alpha'])
+score_mapper.coef_ = np.array(ridge_data['coefficients'])
+score_mapper.intercept_ = ridge_data['intercept']
+score_mapper.n_features_in_ = ridge_data['n_features_in']
+
+neural_mapper = MLPRegressor(
+    hidden_layer_sizes=tuple(neural_data['hidden_layer_sizes']),
+    activation=neural_data['activation']
+)
+neural_mapper.coefs_ = [np.array(c) for c in neural_data['coefs']]
+neural_mapper.intercepts_ = [np.array(i) for i in neural_data['intercepts']]
+neural_mapper.n_features_in_ = neural_data['n_features_in']
+
+poly_features = PolynomialFeatures(
+    degree=poly_data['degree'],
+    include_bias=poly_data['include_bias']
+)
+poly_features.n_features_in_ = poly_data['n_features_in']
+poly_features.n_output_features_ = poly_data['n_output_features']
 
 def predict_ats_score(resume_text, job_text):
     # Generate embeddings