byroneverson/Yi-1.5-34B-Chat-abliterated
Abliterated vs 01-ai/Yi-1.5-34B-Chat. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 4 of 60 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/121, MLP 2/180, Attention 2/240, lm_head 0/1
2 derivatives checked
- mradermacher/Yi-1.5-34B-Chat-abliterated-i1-GGUF: Quantization
- mradermacher/Yi-1.5-34B-Chat-abliterated-GGUF: Quantization