coder3101/gpt-oss-20b-heretic
Abliterated vs openai/gpt-oss-20b. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 22 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Norms 0/49, MLP 0/96, Attention 22/216, lm_head 0/1, Embeddings 0/1
3 derivatives checked
- SerialKicked/GPT-OSS-20B-Heretic-GGUF: Quantization
- DavidAU/OpenAI-gpt-oss-20B-INSTRUCT-Heretic-Uncensored: LoRA-style
- DavidAU/OpenAI-gpt-oss-20B-Claude-4.5-Opus-Heretic-Uncensored: LoRA-style