Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated This model part of the new Qwen3-24B-A4B-Freedom-Thinking-Abliterated-Heretic-NEO: Anything. And it answers honestly. Uncensored. No judgements. Freedom is a 256K context, DENSE moe, with full abliterated experts using the Heretic Method. Need more power? Detail? Just turn up the number of active experts. https://huggingface.co/DavidAU/Qwen3-24B-A4B-Freedom-Thinking-Abliterated-Heretic-NEO-Imatrix-GGUF The method invented by «P-E-W» looks for the best settings to de-censor («abliterate») the model by trial and error AND ensure the model is not damaged too. «KL divergence» is a benchmark to assess model’s root/default state, with zero being perfect. Generally any number less that 1 is great, however with smaller models lower / as close to zero is very important. ZERO (or close to it : lower than .3 ish for small models [0.6B-3B]) means the model runs as well as it did before the process. Again, the goal is to attempt to get to 0 or close to it is critical while FIRST ensuring «KL divergence» as as low as possible or zero. A «refusal rate» of 20 or lower is the goal, with ZERO being perfect.…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: DavidAU
Теги: qwen3, heretic, uncensored, decensored, abliterated, finetune, conversational, text-generation-inference
Лайков: 4 | Загрузок: 100
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
