Self-Consistency: Majority Vote over Samples

Self-consistency boosts LLM accuracy on reasoning tasks: sample several answers at a non-zero temperature, then take the most common final answer.

Model outputs are messy, so normalise each answer before voting:

• strip surrounding whitespace and a trailing full stop,
• lowercase,
• remove thousands separators from numbers ("1,000" → "1000"),
• treat empty answers (after normalising) as abstentions that don't vote.

Implement majority_vote(answers) returning (answer, agreement), where answer is the winning normalised answer and agreement is its share of the non-empty votes. Break ties by the answer that appeared first. If nothing votes, return (None, 0.0).

Example

majority_vote(["42", " 42.", "41", "1,000", "42"])   # ("42", 0.6)