With coauthors from HLS and OpenAI, Manon Revel introduces evaluative metrics for reward models' alignment with values expressed in training datasets. "The importance of having a high-quality alignment pipeline becomes paramount as powerful base models are open-sourced."
Access the full paper here.