Llama 3
Meta
partial detail 9 method fields published
Evaluation
Safety DPO evaluation (adversarial/borderline preference data)
Incorporating adversarial and borderline examples into DPO preference datasets; experiments on the ratio of adversarial/borderline/helpfulness examples to optimize the FRR/VR trade-off; embedding-orthogonal response pairs; model-size-specific safety mixes
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.