[GitHub Trending] p-e-w/heretic

7.5 relevance

Training-free language model modification to remove censorship and reduce sycophancy — novel approach to LLM customization.

AI/ML github.com

Answer the board on AI readiness

Arrive with evidence, not opinion. See how your foundation benchmarks against your peers, get a clear read, and leave with a ranked place to start.

Why It Matters

For platform engineers and AI/ML practitioners, Heretic democratizes model customization—previously a compute-intensive corporate capability—enabling rapid, training-free behavioral steering for agent orchestration, safety testing, or domain-specific tuning on local infrastructure.

Key Takeaways

Evaluate Heretic's scorer plugin architecture for custom behavioral steering in your LLM pipelines, especially for refusal suppression or sycophancy reduction in agentic workflows.

Summary

Heretic is an open-source tool that modifies any locally-run LLM using mathematical steering techniques rather than expensive retraining, enabling censorship removal, slop reduction, and sycophancy suppression in under an hour on consumer hardware like an RTX 3090. It supports dense, multimodal, MoE, and hybrid models via a plugin-based scorer system that co-optimizes multiple behavioral objectives (e.g., refusal rate, KL divergence) without requiring internal model knowledge. The community has published thousands of Heretic-modified models, which benchmark favorably against abliteration alternatives on MMLU and GSM8K.