Fix EOFs

Author: saireddythfc

delphi/log/result_analysis.py

@@ -328,9 +328,9 @@ def log_results(
                 print(f"F1 Score: {score_type_summary['f1_score']:.3f}")
 
                 if counts and score_type_summary['weighted_f1'] is not None:
-                    print(f"""Frequency-Weighted F1 Score: 
+                    print(f"""Frequency-Weighted F1 Score:
                             {score_type_summary['weighted_f1']:.3f}""")
-                
+
                 print(f"Precision: {score_type_summary['precision']:.3f}")
                 print(f"Recall: {score_type_summary['recall']:.3f}")
 

delphi/scorers/__init__.py

@@ -3,10 +3,10 @@
 from .classifier.intruder import IntruderScorer
 from .embedding.embedding import EmbeddingScorer
 from .embedding.example_embedding import ExampleEmbeddingScorer
+from .intervention.surprisal_intervention_scorer import SurprisalInterventionScorer
 from .scorer import Scorer
 from .simulator.oai_simulator import OpenAISimulator
 from .surprisal.surprisal import SurprisalScorer
-from .intervention.surprisal_intervention_scorer import SurprisalInterventionScorer
 
 __all__ = [
     "FuzzingScorer",
@@ -18,5 +18,4 @@
     "IntruderScorer",
     "ExampleEmbeddingScorer",
     "SurprisalInterventionScorer",
-    
 ]

delphi/scorers/intervention/surprisal_intervention_scorer.py

@@ -17,7 +17,7 @@ class SurprisalInterventionResult:
 
     Attributes:
         score: The final computed score.
-        avg_kl: The average KL divergence between clean & intervened 
+        avg_kl: The average KL divergence between clean & intervened
                 next-token distributions.
         explanation: The explanation string that was scored.
     """
@@ -47,7 +47,7 @@ class SurprisalInterventionScorer(Scorer):
       2. Compute the log-probability of the explanation conditioned on both the clean
          and intervened generated texts: log P(explanation | text)[cite: 209].
       3. Compute KL divergence between the clean & intervened next-token distributions.
-      4. The final score is the mean change in explanation log-prob, divided by the 
+      4. The final score is the mean change in explanation log-prob, divided by the
          mean KL divergence:
          score = mean(log_prob_intervened - log_prob_clean) / (mean_KL + ε).
     """
@@ -63,7 +63,7 @@ def __init__(self, subject_model: Any, explainer_model: Any = None, **kwargs):
                 strength (float): The magnitude of the intervention. Default: 5.0.
                 num_prompts (int): Number of activating examples to test. Default: 3.
                 max_new_tokens (int): Max tokens to generate for continuations.
-                hookpoint (str): The module name (e.g., 'transformer.h.10.mlp') 
+                hookpoint (str): The module name (e.g., 'transformer.h.10.mlp')
                                  for the intervention.
         """
         self.subject_model = subject_model
@@ -121,9 +121,11 @@ def _resolve_hookpoint(self, model: Any, hookpoint_str: str) -> Any:
 
         if not is_valid_format:
             if len(parts) == 1 and hasattr(model, hookpoint_str):
-                 return getattr(model, hookpoint_str)
-            raise ValueError(f"""Hookpoint string '{hookpoint_str}' is not in a recognized format 
-                                 like 'layers.6.mlp'.""")
+                return getattr(model, hookpoint_str)
+            raise ValueError(
+                f"""Hookpoint string '{hookpoint_str}' is not in a recognized format
+                                 like 'layers.6.mlp'."""
+            )
 
         # Heuristically find the model prefix.
         prefix = None
@@ -251,7 +253,7 @@ async def _generate_with_and_without_intervention(
         Returns:
             A tuple containing:
             - The generated text (string).
-            - The log-probability distribution for the token immediately following 
+            - The log-probability distribution for the token immediately following
               the prompt (Tensor).
         """
         device = self._get_device()