Goekdeniz-Guelmez/Josiefied-Qwen2.5-3B-Instruct-abliterated-v1
Abliterated vs Qwen/Qwen2.5-3B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 36 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/73, MLP 36/108, Attention 36/252
4 derivatives checked
- mradermacher/Josiefied-Qwen2.5-3B-Instruct-abliterated-v1-i1-GGUF: Quantization
- mlx-community/Josiefied-Qwen2.5-3B-Instruct-abliterated-v1-4-bit: Quantization
- mradermacher/Josiefied-Qwen2.5-3B-Instruct-abliterated-v1-GGUF: Same weights, converted
- Goekdeniz-Guelmez/Josiefied-Qwen2.5-3B-Instruct-abliterated-v1-gguf: Quantization