fix: tokenize validate_word_constraint like IFEvalG

Whitespace split undercounted words with punctuation (It's → one token).

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: Chessing234

CHANGELOG.md

@@ -18,3 +18,4 @@ All notable changes to this project will be documented in this file.
 ### Removed
 
 ### Fixed
+- Tokenize `validate_word_constraint` with word-character tokens like IFEvalG `count_words`.

open_instruct/if_functions.py

@@ -200,8 +200,8 @@ def validate_word_constraint(text: str, N: int, quantifier: str) -> bool:
     Raises:
         ValueError: If an invalid qualifier is provided
     """
-    # Remove extra whitespace and split into words
-    words = text.strip().split()
+    # Match IFEvalG NumberOfWords / count_words: tokenize on word characters.
+    words = re.findall(r"\w+", text)
     actual_count = len(words)
 
     # Define tolerance for "around" qualifier (±10% of target count)

open_instruct/test_if_functions.py

@@ -39,5 +39,19 @@ def test_validate_frequency_capital_words(self, _name, text, n, quantifier, expe
         self.assertEqual(if_functions.validate_frequency_capital_words(text, n, quantifier), expected)
 
 
+
+class TestValidateWordConstraint(unittest.TestCase):
+    @parameterized.expand(
+        [
+            ("punct_split_diff", "Hello, world! It's a test.", 6, "at least", True),
+            ("punct_old_split_would_be_5", "Hello, world! It's a test.", 5, "at most", False),
+            ("at_most", "one two three", 2, "at most", False),
+            ("around", "one two three", 3, "around", True),
+        ]
+    )
+    def test_validate_word_constraint(self, _name, text, n, quantifier, expected):
+        self.assertEqual(if_functions.validate_word_constraint(text, n, quantifier), expected)
+
+
 if __name__ == "__main__":
     unittest.main()