Notes on the post-hoc trustworthiness repair method for distilled LLMs — EK-FAC-preconditioned gradient ascent with PBRF curvature correction, surrogate proxy datasets, and the odds-ratio preference objective.
By Jayin Khanna | Research Notes
Notes on the post-hoc trustworthiness repair method for distilled LLMs — EK-FAC-preconditioned gradient ascent with PBRF curvature correction, surrogate proxy datasets, and the odds-ratio preference objective.