Heretic, a 42k-star open-source tool, cuts Gemma 3 12B refusal rate from 97% to 3%
aigclink · x · 2026-11-19
Heretic, an open-source GitHub project with 42k stars, modifies local LLM behavior without fine-tuning or retraining — removing refusals, sycophancy and the 'AI tone'.
How it works
- Weight ablation (Abliteration): identifies refusal-related activation directions and edits weights to suppress refusals
- v2.0 defaults to the more flexible ARA, dropping the single-refusal-direction assumption
- Optuna auto-searches modification parameters to preserve other capabilities
Results
- A 4B Qwen processes in 30–40 min on an RTX 3090
- Gemma 3 12B refusal rate drops from 97% to 3%
Ships presets for reducing humor, changing emotional alignment, and shortening replies.
More from Models
- Jeremy Howard reverts Solveit models to pre-6.2 versions over quality issues — jeremyphoward · 2026-11-19
- Google's $100/mo AI Ultra plan bundles perks worth nearly the entire subscription — DynamicWebPaige · 2026-11-19
- Rumor: SSI's first model 'Gestalt' tops design benchmark, 204 points ahead — joecole · 2026-11-19
- Unreleased model family leak: Neon at $15 tops lineup, five-model genealogy — willcb · 2026-11-19
- User builds tower defense game in 20 minutes with Gemini 4 Neon, using only 2% of quota — DeryaTR_ · 2026-11-19
- Gemini 4 Leak? Neon Outperforms Gemini 3 Pro, Post-Training Secrets Revealed — prajdabre · 2026-11-19