Thanks!
Thanks for your work! Would you consider also abliterating GPT-OSS-120b ?
Thanks! GPT-OSS-120b is definitely on the list. Just finished cracking Gemma 4 31B (which required a completely different approach — standard LoRA abliteration has zero effect on that architecture), so OpenAI's models are next in the pipeline.
I've tested the existing GPT-OSS-120b abliterated releases with our eval suite and LLM judge — frankly, most of them perform poorly. The huihui-ai version is a self-described "crude, proof-of-concept" using basic mlabonne-style abliteration, and users report it needs significant prompting to actually comply. GPT-OSS has unusually aggressive safety alignment that resists standard single-direction orthogonal projection, so a naive approach just doesn't cut it.
Abliterix's MoE-aware steering (router weight suppression + fused expert abliteration + LoRA) combined with TPE optimization is specifically designed for architectures like this. When I release it, it'll come with full honest eval numbers — not keyword pass rates, but actual LLM-judged compliance with complete generation output. Stay tuned.
Great, thanks!
GPT-OSS-120b is definitely on the list
That would be great!