NOTICE

Abliterix
Copyright (C) 2026  Wangzhang Wu <wangzhangwu1216@gmail.com>

This project is a derivative work of Heretic
(https://github.com/p-e-w/heretic), originally created by Philipp Emanuel
Weidmann (p-e-w).

Original work Copyright (C) 2025  Philipp Emanuel Weidmann

Both the original work and this derivative work are licensed under the
GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later).
See the LICENSE file for the full license text.

Modifications from the original Heretic codebase include but are not
limited to:

  - Multi-objective Optuna TPE optimization (co-minimizing refusals and
    KL divergence)
  - LoRA-based steering instead of destructive weight modification
  - Orthogonal projection of steering vectors against benign directions
  - Mixture-of-Experts (MoE) support: expert profiling, router weight
    suppression, and fused expert abliteration
  - LLM judge refusal detection via OpenRouter
  - Support for additional architectures: LiquidAI LFM2 (hybrid
    conv+attention), GLM-4 MoE (MLA), Devstral, MiniMax-M2.5, GPT-OSS
  - Fractional direction index and per-component abliteration parameters
  - Auto batch sizing and KL divergence pruning
  - Research visualization tools (PaCMAP residual plots)
  - TOML-based configuration system
  - Comprehensive test suite