kawasumi/Crow-9B-it
Abliterated vs tokyotech-llm/Gemma-2-Llama-Swallow-9b-it-v0.1. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 31 of 42 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/169, MLP 31/126, Attention 31/168
2 derivatives checked
- mradermacher/Crow-9B-it-i1-GGUF: Quantization
- mradermacher/Crow-9B-it-GGUF: Same weights, converted