bat-6 commited on
Commit
955c9ea
·
1 Parent(s): dd72980
src/similarity_model/feature_similarity.py CHANGED
@@ -143,17 +143,15 @@ def compute_feature_similarity(
143
  matched_a.add(i)
144
  matched_b.add(j)
145
 
146
- # Final Metrics
147
  import numpy as np
148
  shared_scores = [m["score"] for m in matches]
149
  mean_similarity = float(np.mean(shared_scores)) if shared_scores else 0.0
150
 
151
- min_len = min(len(fa), len(fb))
152
- coverage = len(matches) / min_len if min_len > 0 else 0.0
153
 
154
- final_score = (SIMILARITY_WEIGHT * mean_similarity) + (COVERAGE_WEIGHT * coverage)
155
- if len(matches) == 0:
156
- final_score = 0.0
157
 
158
  final_score = min(final_score, 1.0)
159
 
 
143
  matched_a.add(i)
144
  matched_b.add(j)
145
 
 
146
  import numpy as np
147
  shared_scores = [m["score"] for m in matches]
148
  mean_similarity = float(np.mean(shared_scores)) if shared_scores else 0.0
149
 
150
+ max_len = max(len(fa), len(fb))
151
+ coverage = len(matches) / max_len if max_len > 0 else 0.0
152
 
153
+ sum_similarity = sum(shared_scores)
154
+ final_score = sum_similarity / max_len if max_len > 0 else 0.0
 
155
 
156
  final_score = min(final_score, 1.0)
157
 
src/similarity_model/hybrid_ranker.py CHANGED
@@ -135,25 +135,14 @@ def compute_originality(
135
  """
136
  Originality Score (0-100)
137
 
138
- - Driven primarily by inverse of hybrid similarity (non-linear).
139
- - Boosted by the uniqueness of the features.
140
  """
141
 
142
  hybrid_score = clamp(hybrid_score)
143
 
144
- uniqueness_ratio = (
145
- unique_query_features / total_query_features
146
- if total_query_features > 0
147
- else 1.0
148
- )
149
-
150
- # Non-linear penalty: 50% similarity drops originality slightly,
151
- # but 90% similarity drops it massively.
152
- similarity_penalty = hybrid_score ** 2
153
- base_originality = 1.0 - similarity_penalty
154
-
155
- # Blend uniqueness and base originality
156
- originality = (0.40 * uniqueness_ratio) + (0.60 * base_originality)
157
 
158
  return round(
159
  max(0.0, min(100.0, originality * 100)),
 
135
  """
136
  Originality Score (0-100)
137
 
138
+ - Driven strictly by inverse of hybrid similarity (non-linear).
 
139
  """
140
 
141
  hybrid_score = clamp(hybrid_score)
142
 
143
+ # Non-linear penalty: 50% similarity drops originality slightly (to 75%),
144
+ # but 90% similarity drops it massively (to 19%).
145
+ originality = 1.0 - (hybrid_score ** 2)
 
 
 
 
 
 
 
 
 
 
146
 
147
  return round(
148
  max(0.0, min(100.0, originality * 100)),