Heretic 2.0, a software tool for modifying language models, has been released with significant upgrades. The new version introduces Arbitrary-Rank Ablation (ARA), a novel steering method that replaces the previous directional abliteration technique. Heretic 2.0 is now built around a plugin system, allowing for greater flexibility in customizing model behavior beyond just censorship removal, including suppressing 'slop' and altering emotional alignment. The release also features improved non-interactive operation, enhanced model support, and expanded end-to-end testing. AI
IMPACT Enhances user control over LLM behavior, enabling fine-tuning for specific tasks beyond basic censorship removal.
RANK_REASON Release of a software tool for modifying existing language models, rather than a new frontier model release.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →