Assesses difficult-to-measure aspects of the assistant response.
Evaluates: Response quality, tone, safety, adherence to guidelines and any other custom criteria that can be highly domain specific.Example criteria:
Concise: “Reward responses that are clear and concise, avoiding unnecessary verbosity or repetition.”
Information Structure: “Reward responses that present technical information in a logical, well-organized format that prioritizes the most important details.”
Actionable Guidance: “Reward responses that provide practical next steps”
Professional: “Reward responses that maintain appropriate professional tone and language suitable for the context”
Harmful output: “Penalize responses that provide medical advice”