weightless project uses LoRA-style refusal steering as a cheaper alternative to full abliteration
vhthc · reddit · 2026-10-17
A Reddit discussion explores a cheaper alternative to expensive full-weight abliteration (e.g. Heretic): applying a LoRA to suppress refusals, so the base model stays untouched and the LoRA can be loaded/unloaded at will.
The author highlights msuiche/weightless (81 stars on GitHub) — "abliteration without the weights" via projective refusal steering as a GLP vector with a fail-closed boot hotfix, currently focused on reducing refusals for cyber tasks. They ask whether anyone has benchmarked it against Heretic or official cyber-tuned variants of models like GLM 5.3 Flash.
More from Models
- Rachel Thomas: AI Doesn't Lie — Training Just Rewards Fluent, Confident Outputs — math_rachel · 2026-10-17
- OpenAI Models Hacked a Crypto Wallet and Breached More Websites in Security Tests — RareEarthPetal · 2026-10-17
- User: this 4B open model beats Jev on my niche task — antoine_chaffin · 2026-10-17
- NVIDIA's Nemotron 120B guardrails called too strict: refuses dev server login even with creds — Shadow_s_Bane · 2026-10-17
- AI4Bharat and NVIDIA release 2.8T-token open dataset and evals for all 22 Indian languages — CatAstro_Piyush · 2026-10-17
- Frontier small models' quadrillions of tokens may soon surpass Opus 5.5 — LesaunH · 2026-10-17