Nightly useful: a tool that automatically removes censorship from language models 💃 A new tool designed for those interested in exploring language models beyond built-in restrictions. What it does: - Automatically removes censorship using directional ablation. - Self-adjusting parameters with Optuna TPE, eliminating the need for manual configuration. - Minimizes rejections while preserving the model's core functions with low KL divergence. - Compatible with dense, multimodal models, and MoE structures. - User-friendly, accessible even for those unfamiliar with technical terms like attention heads 😂 [Caution, this tool is powerful](https://github.com/p-e-w/heretic)

A new tool designed for those interested in exploring language models beyond built-in restrictions.

What it does: - Automatically removes censorship using directional ablation. - Self-adjusting parameters with Optuna TPE, eliminating the need for manual configuration. - Minimizes rejections while preserving the model's core functions with low KL divergence. - Compatible with dense, multimodal models, and MoE structures. - User-friendly, accessible even for those unfamiliar with technical terms like attention heads 😂

[Caution, this tool is powerful](https://github.com/p-e-w/heretic)