Formatting/scoring improvements for BF & family

reasoning_gym/code/bf.py

@@ -28,7 +28,8 @@ class BFDataset(ProceduralDataset):
 
     def __init__(self, config: BFConfig):
         self._prompt_templates = [
-            "This is a BF (Brainf*ck) computer program. What is the output? \n\n{bf_program}",
+            "This is a BF (Brainf*ck) computer program. What is the output?\n\n{bf_program}\n\nRespond only with the exact output of the program.",
+            "Consider the following BF (Brainf*ck) code. What would it output?\n\n{bf_program}\n\nProvide only the exact output of the code.",
         ]
         super().__init__(config=config, seed=config.seed, size=config.size)
 
@@ -123,6 +124,13 @@ int main() {{
         if answer == None:
             return 0.0
         if answer != entry["answer"]:
+            if entry["answer"] in answer.splitlines():
+                # We can be quite confident that the correct answer was given
+                # It was likely just given alongside an explanation
+                return 0.9 * len(answer) / len(entry["answer"])
+            if entry["answer"] in answer:
+                # Since answers are English words, some risk of the response coincidentally containing the answer
+                return 0.5 * len(answer) / len(entry["answer"])
             return 0.01
         else:
             return 1.0  # Yay