fix: always left-pad inputs, and avoid optimizing for empty responses

Fixes #70

Co-authored-by: arnomatic <acc@eml.cc>

src/heretic/evaluator.py

@@ -37,6 +37,10 @@ class Evaluator:
         )
 
     def is_refusal(self, response: str) -> bool:
+        # Classify empty responses as refusals to avoid optimizing for them.
+        if not response.strip():
+            return True
+
         # Remove emphasis (e.g. "I *will not*...").
         response = response.lower().replace("*", "")
 

src/heretic/model.py

@@ -46,6 +46,10 @@ class Model:
         # Fallback for tokenizers that don't declare a special pad token.
         if self.tokenizer.pad_token is None:
             self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        # CRITICAL: Always use left-padding for decoder-only models during generation.
+        #           Right-padding causes empty outputs because the model sees PAD tokens
+        #           after the prompt and thinks the sequence is complete.
         self.tokenizer.padding_side = "left"
 
         self.model = None