Skip to main content

AlignScore

A factual consistency / alignment metric that evaluates how well a claim (generated or asserted text) is supported by a context (retrieved or original text).
It uses a unified alignment function (often built on a pretrained alignment model) to score alignment between context chunks and claim sentences.

  • Inputs: context (text) and claim (text)
  • Returns: a single AlignScore value (higher = stronger factual consistency)

Insights​

AlignScoreInference
closer to 1high factual consistency
closer to 0low factual consistency

Example​

from vero.metrics import AlignScore

with AlignScore() as ascore:
    # evaluate many (context, claim) pairs
    results = [ascore.evaluate(context, claim) for context, claim in zip(contexts_list, claims_list)]
print(results)

Output​

0.91