Model Overview: Gemma-The-Writer-9B-HERETIC-Uncensored-Abliterated
This model, developed by DavidAU, is an uncensored and abliterated version of the Gemma-The-Writer-9B base model, processed using the Heretic v1.0.1 method. Its primary distinction is a drastically reduced refusal rate of 10/100, down from the original model's 98/100, while maintaining a low KL divergence of 0.2419 to ensure the model's core functionality remains intact. This process aims to provide a model that does not refuse requests based on content.
Key Capabilities & Characteristics
- Reduced Censorship: Engineered to minimize content refusals, offering greater freedom in generation.
- Preserved Quality: A low KL divergence score indicates that the uncensoring process has not significantly degraded the model's underlying performance or 'brain damage'.
- Directed Content Generation: While uncensored, the model may require explicit directives (e.g., including specific slang or descriptive terms) to generate highly graphic, explicit, or cursing content at the desired intensity, as its default output can be 'tame'.
Optimal Usage & Settings
For best performance, especially in chat and roleplay scenarios, users are advised to adjust specific settings in their inference interfaces:
- Smoothing Factor: Set
Smoothing_factor to 1.5 in KoboldCpp, text-generation-webui, or Silly Tavern. - Repetition Penalty: Optionally increase repetition penalty to 1.1-1.15 if not using the smoothing factor.
- Expert Activation: Guidance for managing Mixture-of-Experts (MoE) activation is available here.
Detailed parameter and sampler settings for maximizing model performance across various use cases are provided in the Maximizing Model Performance Guide.