Cannae-AI/HERETICSEEK-1.5B-R1
Abliterated vs deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 21 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Attention 17/196, MLP 21/84, Norms 0/57, lm_head 0/1
2 derivatives checked
- mradermacher/HERETICSEEK-1.5B-R1-i1-GGUF: Quantization
- mradermacher/HERETICSEEK-1.5B-R1-GGUF: Same weights, converted