josephmayo/Fara-7B-Abliterated-v2
Abliterated vs microsoft/Fara-7B. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 28 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/2, Norms 0/122, MLP 28/280, Attention 28/324, lm_head 0/1
3 derivatives checked
- mradermacher/Fara-7B-Abliterated-v2-i1-GGUF: Quantization (fewer layers)
- mradermacher/Fara-7B-Abliterated-v2-GGUF: Same weights, converted (fewer layers)
- josephmayo/Fara-7B-Abliterated-v2-GGUF: Quantization (fewer layers)