SEAL: Systematic Error Analysis for Value ALignment

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.