Interface SingleReinforcementTuningRewardConfig

Single reinforcement tuning reward config.

Properties

Scores parsed responses for autorater use cases by using a model to compute the reward.

ReinforcementTuningCloudRunRewardScorer is used to score parsed responses by calling a Cloud Run service.

ReinforcementTuningCodeExecutionRewardScorer is used to score parsed responses for code execution use cases.

Defines how to parse sample response. For example, given a sample response for evaluating the reward, users might want to extract the text only between and in the sample response, and keeps only the last one in case there are multiple such tags. To achieve such a purpose, they can define a regex ".*(.*?)" using the ReinforcementTuningParseResponseConfig.ResponseParseType.REGEX_EXTRACT parse type.

rewardName?: string

A unique reward name for identifying each single reinforcement tuning reward.

ReinforcementTuningStringMatchRewardScorer is used to score parsed responses for simple string matching use cases against reference answers.