GEPA.Adapters.Confidence.Scoring (GEPA v0.3.0)

Copy Markdown View Source

Confidence-aware scoring strategies for GEPA.Adapters.Confidence.

Each strategy receives whether the prediction is correct and the joint logprob for the target field. Logprobs are converted to probabilities via exp(logprob) before confidence penalties are applied.

Summary

Types

strategy()

Functions

clamp01(value)
describe(strategy)
probability(logprob)
score(strategy, is_correct, logprob_score)

Types

strategy()

@type strategy() :: struct() | module()

Functions

clamp01(value)

@spec clamp01(number()) :: float()

describe(strategy)

@spec describe(strategy()) :: String.t()

probability(logprob)

@spec probability(number() | nil) :: float() | nil

score(strategy, is_correct, logprob_score)

@spec score(strategy(), boolean(), number() | nil) :: float()