cesun/ThinkEdit-deepseek-qwen-14b
Abliterated vs deepseek-ai/DeepSeek-R1-Distill-Qwen-14B. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 7 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Attention 7/336, MLP 0/144, Norms 0/97, lm_head 0/1
2 derivatives checked
- mradermacher/ThinkEdit-deepseek-qwen-14b-i1-GGUF: Quantization
- mradermacher/ThinkEdit-deepseek-qwen-14b-GGUF: Quantization