Best-of-N vs. Greedy Solutions
Dataset Description
This dataset contains 20 math problems (levels 1–3 from MATH-500) and three sets of predictions:
Greedy: Single-step deterministic decoding
Vanilla Best-of-N: Highest-reward chain-of-thought out of N solutions
Weighted Best-of-N: Sum of reward across identical final answers (DeepMind & Math-Shepherd approach)