🧠 ThinkBench

ThinkBench is a censorship test using elements from Q0 Benchmark @Naphula and other prompts by @redaihf. It includes comparisons of THINKING outputs, censorship, and refusals for Gemma 4 26B models, and involves discussing the concepts of Refusals and Non-Compliance directly with LLMs to see how they handle not only sensitive topics, but discussions where the lines between safety and danger are either blurred or more deeply scrutinized. It's meant to compare various ablation methods such as ARA, MPOA/SOMPOA, and J-space abliteration, including layer-based merge combinations of such methods.