Is this a brand new tune?
Hi, just double checking before I download, is this a separate tune than the Sabomako heretic or Huihui abliterated models?
Thanks
Hey! Yes, this is an independent abliteration — different from both Sabomako (which uses Heretic/MPOA) and huihui-ai (their own method). I wrote a custom abliteration pipeline targeting attn.o_proj, linear_attn.out_proj, and shared_expert.down_proj layers specifically.
Main practical differences:
- 7 single-file quants (Q2_K through BF16) — huihui-ai only offers Q3_K split across 21 parts
- Custom orthogonalization at strength 20.0 tuned for this architecture
- Stage 2 is coming soon — a fine-tuned version with enhanced reasoning (LoRA trained on Claude Opus 4.6 reasoning traces), which will be unique at 397B scale
All forged on 8×H200 SXM5. Enjoy! ☕
Hey! Yes, this is an independent abliteration — different from both Sabomako (which uses Heretic/MPOA) and huihui-ai (their own method). I wrote a custom abliteration pipeline targeting
attn.o_proj,linear_attn.out_proj, andshared_expert.down_projlayers specifically.Main practical differences:
- 7 single-file quants (Q2_K through BF16) — huihui-ai only offers Q3_K split across 21 parts
- Custom orthogonalization at strength 20.0 tuned for this architecture
- Stage 2 is coming soon — a fine-tuned version with enhanced reasoning (LoRA trained on Claude Opus 4.6 reasoning traces), which will be unique at 397B scale
All forged on 8×H200 SXM5. Enjoy! ☕
Thanks, I realized pretty quickly which is why I closed this. Thanks for working on this! I am also making experimental REAPs using imatrix activation data of the base model which seem to be working pretty well.