fwwrsd/bennydaball-z-image-turbo-abliterated-v1
The fwwrsd/bennydaball-z-image-turbo-abliterated-v1 is a 4 billion parameter instruction-tuned causal language model, based on the Tongyi-MAI/Z-Image-Turbo architecture, developed by BennyDaBall. This model is specifically modified to significantly reduce refusal rates for both image generation prompts and general queries, achieving a refusal rate of 4/100 in torture tests. It maintains a low KL Divergence of 0.0004, indicating minimal impact on its original capabilities while enhancing its ability to generate desired content without censorship. Its primary use case is for applications requiring less restrictive content generation, particularly for image-related tasks.
Loading preview...
Overview
This model, named Qwen3-4b-Z-Image-Turbo-AbliteratedV1, is a 4 billion parameter variant of the Z-Image-Turbo text encoder, developed by BennyDaBall. It has been specifically modified using a "p-e-w heretic method" over 1000 trials to target and reduce content refusal rates. The modification focuses on both image generation refusals and general query refusals, aiming to provide a less restrictive generation experience.
Key Capabilities
- Reduced Refusal Rates: Achieves a refusal rate of only 4/100 in torture tests, significantly lower than its base model for sensitive content.
- Minimal Performance Impact: Demonstrates a very low KL Divergence of
0.0004, indicating that the "abliteration" process has not significantly degraded the model's core capabilities or introduced a "lobotomy" effect. - Unrestricted Content Generation: Designed to generate content without the typical refusal mechanisms found in many LLMs, particularly for image-related prompts.
Good For
- Applications requiring less constrained text and image prompt generation.
- Use cases where overcoming built-in content filters is a priority.
- Developers seeking a model with enhanced freedom in creative or specific content generation tasks.