Reinforcement Tuning Autorater Scorer
@Serializable
Reinforcement tuning autorater scorer.
Constructors
Link copied to clipboard
constructor(autoraterConfig: AutoraterConfig? = null, autoraterPrompt: String? = null, autoraterResponseParseConfig: ReinforcementTuningParseResponseConfig? = null, parsedResponseConversionScorer: ReinforcementTuningAutoraterScorerParsedResponseConversionScorer? = null, exactMatchScorer: ReinforcementTuningAutoraterScorerExactMatchScorer? = null)
Properties
Link copied to clipboard
Autorater config for evaluation.
Link copied to clipboard
The prompt for an autorater to scorer the parsed sample response. This field supports the following placeholders that will be replaced before scoring: - {{prompt}} - {{response}} - {{system_instruction}} - {{references.key}}
Link copied to clipboard
Parses autorater returned response for scoring. For example, if the autorater response has reward stored in the 2.0 block, defining a parsing response config using regex ".*(.*?)" will return a score "2.0".
Link copied to clipboard
Scores autorater responses by using string match reward scorer.
Link copied to clipboard
val parsedResponseConversionScorer: ReinforcementTuningAutoraterScorerParsedResponseConversionScorer? = null
Scores autorater responses by directly converting parsed autorater response to a float reward. Note: Reward is clipped to be within [-1, 1], i.e., reward = max(min(reward, 1.0), -1.0).