luml.experiments.evaluation.scorers.builtin.correctness

Correctness Objects

class Correctness(SupervisedLLMJudgeScorer)

Scores factual correctness of a response against expected facts.

Supervised scorer — requires expected_output. Checks faithfulness (facts correct?) and hallucination (fabricated information?). Returns a float 0.0 (incorrect / hallucinated) to 1.0 (fully correct and faithful).

expected_output can be a string or a dict with an "expected_facts" list, which is formatted as a bulleted block for the judge.

Default input keys: "request", "question".

from luml.experiments.evaluation.scorers.builtin import Correctness
scorer = Correctness()