expression
A string or a regular expression to match against for evaluating rewards. Users can also provide a references map of {key: value} whose value will be used to replace the placeholder {{references.key}} in the expression. For example, if the following references are defined in the training / validation dataset: `` { "systemInstruction": ..., "contents": ..., "references": { "concise_answer": "Yes", "verbose_answer": "The answer is Yes" } } `` and if users define the following StringMatchExpression: { "matchOperation": "REGEX_CONTAINS", "expression": ".{{references.concise_answer}}." } On evaluating the reward for each sample response, this StringMatchExpression will be substituted as: `` { "matchOperation": "REGEX_CONTAINS", "expression": ".*Yes.*" } ``