CoolFace
Datasetpublic

omrifahn/kfac-memorization-code

K-FAC Memorization Suppression Reproduction of "From Memorization to Reasoning in the Spectrum of Loss Curvature" with extended experiments on modified importance formulas. Overview This project implements K-FAC (Kronecker-Factored Approximate Curvature) based weight editing to suppress verbatim memorization in language models while preserving general capabilities. Key insight: The Fisher Information Matrix, approximated by K-FAC, reveals directions in weight… See the full description on the dataset page: https://huggingface.co/datasets/omrifahn/kfac-memorization-code.

sourceHugging Faceupdated 8mo agoView on Hugging Face
0likes15downloads

No commit history came back for main. The revision may not exist, or the source declined the request.