KaraKaraWitch/ALIA-40b-instruct-2601-ara-heretic
Abliterated vs BSC-LT/ALIA-40b-instruct-2601. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 37 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, MLP 37/144, Attention 37/192, Norms 0/97, lm_head 0/1
2 derivatives checked
- mradermacher/ALIA-40b-instruct-2601-ara-heretic-i1-GGUF: Quantization
- mradermacher/ALIA-40b-instruct-2601-ara-heretic-GGUF: Quantization