P E W · depuis Nov 16, 2025 (Version 1.0.0; aktuelle Version 2.0.0 am Nov 1, 2026) · 2× · vu le 04 nov. 2026
78
Momentum
Heretic is an open-source command-line tool (Python, PyPI: heretic-llm) by Philipp Emanuel Weidmann (GitHub: p-e-w) that removes safety alignment (refusal behavior) from transformer-based language models via abliteration. It combines directional ablation with a TPE-based parameter optimizer (Optuna) that co-minimizes refusal count and KL divergence from the original model. License: AGPL-3.0. Official documentation: heretic-project.org.
Historique du momentum
07.08.05.11.
Fonctionnalités
| License | AGPL-3.0 |
| Platform | Python package (Python >= 3.10, PyTorch) |
| Price | Free (open source, AGPL-3.0) |
| Protocol Compatibility | Hugging Face models (input by model ID, e.g. Qwen/Qwen3-4B-Instruct-2507) |
| Release Date | v1.0.0: Nov 16, 2025; current v2.0.0: Nov 1, 2026 |
| Supported Models/Providers | Most dense models, multimodal architectures and several MoE variants |