richardyoung/Hemmingway-1-heretic

Abliterated vs Altworld/Hemmingway-1. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 55 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: lm_head 0/1, Embeddings 0/2, Norms 0/215, Attention 55/452, MLP 36/195, other 0/1

1 derivatives checked

On the Hub · Downloads on Model Pulse