Goekdeniz-Guelmez/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4
Abliterated vs Qwen/Qwen2.5-14B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 48 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/97, MLP 48/144, Attention 48/336, lm_head 0/1
10 derivatives checked
- Goekdeniz-Guelmez/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4-gguf: Quantization
- mradermacher/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4-i1-GGUF: Quantization
- mradermacher/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4-GGUF: Quantization
- mlx-community/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4-4-bit: Quantization
- mlx-community/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4-8-bit: Quantization
- fuze-eletrik/jose-final-Qwen2.5-14B-Instruct-abliterated-v4-GGUF: Full fine-tune
- CombinHorizon/Josiefied-abliteratedV4-Qwen2.5-14B-Inst-BaseMerge-TIES: Partial fine-tune
- CultriX/Qwen2.5-14B-DeepResearch: Full fine-tune
- NexesMess/Qwen2.5-14B-Instruct-abli_R1: Full fine-tune
- Sorawiz/Qwen2.5-14B-Instinct-RP: Full fine-tune