DarkKitsune/qwen3-4b-instruct-special

Model representing the weights given by the formula:

Weights-A = Goekdeniz-Guelmez/Josiefied-Qwen3-4B-abliterated-v2 - Qwen/Qwen3-4B-Instruct-2507
Weights-B = HuiyuWang/dpo-qwen-cot-merged + Weights-A
Weights-C = taki555/Qwen3-4B-Instruct-2507-Art + Weights-A
Result = DARE-linear(Weights-B * 0.65, Weights-C * 0.35, density: 0.5)

The goal is a model that does uncensored, less-verbose and better-structured output especially for games and procgen. Great for predicting text, working with JSON, generating lore, or generate single values/ideas by pushing "**" onto the context and then inferring until the next "**".