Theright07/llama3.1-heretic-unsensored
Theright07/llama3.1-heretic-unsensored is an 8 billion parameter language model, a decensored version of dphn/Dolphin3.0-Llama3.1-8B, created using the Heretic v1.3.0 tool. This model is designed to offer greater control over alignment and system prompts, reducing refusals compared to its original counterpart. It is optimized for general-purpose applications including coding, math, agentic tasks, and function calling, providing a steerable alternative to highly aligned commercial models.
Loading preview...
Model Overview
Theright07/llama3.1-heretic-unsensored is an 8 billion parameter model derived from dphn/Dolphin3.0-Llama3.1-8B, specifically engineered to be a "decensored" version. This modification was achieved using the Heretic v1.3.0 tool, aiming to provide users with more control over the model's behavior and alignment.
Key Characteristics & Differentiators
- Decensored Alignment: Unlike its base model, this version significantly reduces refusals, demonstrating 0 refusals out of 100 tests compared to 27/100 for the original Dolphin3.0-Llama3.1-8B. This allows users to define the model's ethics and guidelines.
- Reproducibility: The model's creation process is reproducible, with detailed parameters for its "abliteration" process provided.
- General Purpose: Inheriting from the Dolphin 3.0 series, it is designed as a versatile local model capable of handling a wide range of tasks including coding, mathematical problems, agentic workflows, and function calling.
- User Control: Emphasizes user control over system prompts and alignment, offering an alternative to commercial models that maintain strict control over these aspects.
Performance Insights
While the model achieves a KL divergence of 0.0082 compared to the original, its primary performance differentiator is the elimination of refusals. Open LLM Leaderboard evaluations for the base Dolphin3.0-Llama3.1-8B show an average score of 24.97%, with specific metrics like IFEval at 76.21% and BBH at 27.63%.