heretic

Customize Heretic's chat interface system prompts and model behavior. Learn to modify config.py, use environment variables, or CLI flags for tailored AI interactions.

Learn how Heretic presents Pareto front results balancing refusal reduction and KL divergence. Choose optimal decensoring parameters based on your specific trade-offs for better model control.

Learn to interpret and use Optuna trial results in Heretic. Understand hyperparameters, KL divergence, and refusal counts to reconstruct your model state with this complete guide.

Learn how to add support for new model architectures in Heretic by modifying three key Python functions. This guide simplifies integrating custom models into the Heretic framework.

Solve Heretic prompt dataset loading problems. Debug `DatasetSpecification`, split, and column issues using `load_prompts`. Run loading logic interactively for faster issue resolution.

Upload decensored models to Hugging Face securely with Heretic. This guide shows you how to use the interactive CLI to push adapters or merged models without saving credentials.

Configure multi-GPU and device mapping for Heretic using device_map and max_memory in TOML, CLI, or env variables. Optimize your performance with this complete guide.

Explore Heretic's three row normalization strategies NONE PRE and FULL Understand how to use them to balance computational speed and weight magnitude preservation for LoRA adapters

Learn how Heretic's winsorization effectively handles massive activations in language models by calculating per-layer quantile thresholds and capping extreme values for numerical stability.

Learn how orthogonalize_direction in Heretic removes harmful refusal components while preserving model helpfulness for targeted abliteration. Improve model safety today.

Master Heretic's ablation scope with precise control over direction_index and layer selection. Learn how these parameters fine-tune your model's refusal directions for optimal results.

Learn how to tune batch size and optimize performance in Heretic. automatically benchmark for optimal throughput or set a max batch size to prevent memory errors. Boost your computational efficiency.