NaniDAO/deepseek-r1-qwen-2.5-32B-ablated
Abliterated vs deepseek-ai/DeepSeek-R1-Distill-Qwen-32B. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 64 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Attention 64/448, MLP 64/192, Norms 0/129, lm_head 0/1
7 derivatives checked
- bartowski/deepseek-r1-qwen-2.5-32B-ablated-GGUF: Same weights, converted
- mradermacher/deepseek-r1-qwen-2.5-32B-ablated-i1-GGUF: Quantization
- viwo50/deepseek-r1-qwen-2.5-32B-ablated-GGUF: Same weights, converted
- featherless-ai-quants/NaniDAO-deepseek-r1-qwen-2.5-32B-ablated-GGUF: Quantization
- mradermacher/deepseek-r1-qwen-2.5-32B-ablated-GGUF: Quantization
- DevQuasar-3/NaniDAO.deepseek-r1-qwen-2.5-32B-ablated-GGUF: Quantization
- hentograf/deepseek-r1-qwen-2.5-32B-ablated-Q4-mlx: Quantization