Minor question template & score_answer improvements (#261)

* math prompt improvements
* ignore brackets in complex_arithmetic results
* improve additional instruction in prompt of polynomial_equations
* more strict tests for score_answer in polynomial_equations
* simplify special reward handling
* fix test_intermediate_integration
* fix sokoban dataset
* add common dataset score_answer consistency test
This commit is contained in:
Andreas Köpf 2025-03-04 21:55:09 +01:00 committed by GitHub
parent 061282e373
commit 5d7fbac0ad
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
106 changed files with 403 additions and 507 deletions

View file

@ -428,7 +428,7 @@ class CalendarArithmeticDataset(ProceduralDataset):
def score_answer(self, answer: Optional[str], entry: dict[str, Any]) -> float:
# we suppose the answer is the last occurence of the expected answer type
if answer is None:
if not isinstance(answer, str) or len(answer) == 0:
return 0.0
oracle_answer = entry["answer"]
@ -439,9 +439,6 @@ class CalendarArithmeticDataset(ProceduralDataset):
CalendarTask.WEEKDAY_OF_DATE_FROM_FIRST_DATE.value,
CalendarTask.WEEKDAY_OF_DATE.value,
}:
if not answer:
return 0.0
answer = answer.strip()
oracle_answer = oracle_answer
weekdays = {d.name.title() for d in Weekday}