PeterCxy/Step-3.7-Flash-ARA-Heretic
Abliterated vs stepfun-ai/Step-3.7-Flash. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 28 of 45 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Norms 0/371, MLP 0/449, Attention 28/413, other 0/184, lm_head 0/1, Embeddings 0/2
1 derivatives checked
- PeterCxy/Step-3.7-Flash-ARA-Heretic-GGUF: Quantization (fewer layers)