ignore single whitespace at beginning and end of answer, use reward = len(oracle_answer) / len(answer)

reasoning_gym/dataset.py

@@ -59,7 +59,7 @@ class ProceduralDataset(ABC, Sized, Iterable[Dict[str, Any]]):
             if answer == oracle_answer:
                 reward = 1.0
             elif oracle_answer in answer:
-                reward = 0.5
+                reward = len(oracle_answer) / len(answer)
             else:
                 reward = 0.01
 

reasoning_gym/utils.py

@@ -17,7 +17,7 @@ Once you have thought about the reasoning process, provide the answer in the fol
 
 
 def extract_answer(completion: str, tag_name: str = "answer") -> Optional[str]:
-    regex = f"<{tag_name}>(.*?)</{tag_name}>"
+    regex = f"<{tag_name}>\\s?(.*?)\\s?</{tag_name}>"
     matches = list(
         re.finditer(
             regex,

tests/test_dataset.py

@@ -2,6 +2,7 @@ import pytest
 
 from reasoning_gym.arithmetic.basic_arithmetic import BasicArithmeticDataset, BasicArithmeticDatasetConfig
 from reasoning_gym.dataset import ReseedingDataset
+from reasoning_gym.utils import extract_answer
 
 
 def test_reseeding_dataset_iteration():
@@ -38,3 +39,19 @@ def test_reseeding_dataset_iteration():
     test_item = next(iter(infinite_dataset))
     assert infinite_dataset.score_answer("wrong", test_item) == 0.01
     assert infinite_dataset.score_answer(test_item["answer"], test_item) == 1.0
+
+
+def test_extract_answer():
+    assert extract_answer("This is a text. <final_answer>1234</final_answer>", tag_name="final_answer") == "1234"
+
+    # ignore single whitespae
+    assert extract_answer("This is a text. <answer>\n1234 </answer>", tag_name="answer") == "1234"
+
+    config = BasicArithmeticDatasetConfig(
+        min_terms=2, max_terms=3, min_digits=1, max_digits=2, operators=["+"], allow_parentheses=False, seed=42, size=10
+    )
+
+    base_dataset = BasicArithmeticDataset(config)
+    item = base_dataset[0]
+    assert base_dataset.score_answer(item["answer"] + " + x", item) > 0.1
+    assert base_dataset.score_answer(item["answer"], item) == 1.0