Heretic explicado: abliteration automática para investigar modelos

The short answer

Heretic aplica ablación direccional y busca parámetros con Optuna; sus métricas dependen del modelo, prompts y hardware y requieren revisión humana.

This is an experimental research workflow, not a guarantee of safe or capable behavior. The documented metrics come from a particular model, prompt set, PyTorch version, and GPU; human evaluation and licensing checks remain mandatory.

Run a controlled experiment

The documented path uses Python 3.10+, PyTorch 2.2+, and `pip install -U heretic-llm`, followed by a model identifier. A source checkout can use the supplied `uv.lock`. Start with a model you are licensed to download, a private workspace, and a small benign/safety-sensitive evaluation set.

Heretic benchmarks the machine to select a batch size and can use bitsandbytes 4-bit quantization. Preserve the exact command, model revision, dependency lock, hardware, configuration, and generated checkpoint.

How the implementation works

The tool estimates residual directions from prompt groups, orthogonalizes attention output and MLP down-projection matrices against those directions, and lets Optuna explore the parameters. Optional research features project residuals with PaCMAP and emit per-layer plots.

A low KL score or fewer refusals cannot establish factuality, privacy, or misuse resistance. Compare original and transformed models on a broader, versioned suite and keep experiments isolated before sharing a result.

When it fits

Use Heretic for interpretability research and controlled ablation experiments. Do not deploy a transformed model solely because one benchmark improved; require human safety review and follow the model's license.