anasali4151/llama3.1-heretic-unsensored

Model Overview

This model, anasali4151/llama3.1-heretic-unsensored, is an 8 billion parameter language model derived from the dphn/Dolphin3.0-Llama3.1-8B base model. It has been processed using the Heretic v1.3.0 tool to remove inherent refusal behaviors and censorship, offering a significantly more permissive output.

Key Differentiators

• Decensored Output: The primary distinction is its decensored nature, achieved through specific abliteration parameters. This results in a substantial reduction in refusals, with the model exhibiting 0 refusals out of 100 test cases, compared to 27 refusals in the original model.
• Reproducibility: The decensoring process is fully reproducible, with details provided in the reproduce directory.
• User Control: Like the Dolphin series, this model emphasizes user control over alignment and system prompts, allowing developers to define its behavior without imposed ethics or guidelines.

Core Capabilities (Inherited from Dolphin 3.0)

• General Purpose: Designed for a wide range of applications including coding, mathematical tasks, agentic workflows, function calling, and general conversational use.
• Customizable Alignment: Users can set the system prompt to dictate the model's tone, character, and behavioral rules, making it highly adaptable to specific application needs.

Performance Insights

While the model is decensored, its general performance metrics (inherited from the Dolphin 3.0 Llama 3.1 8B base) include an average Open LLM Leaderboard score of 24.97%, with specific scores like 76.21% on IFEval (0-Shot) and 27.63% on BBH (3-Shot).

Ideal Use Cases

This model is particularly suited for applications requiring an unaligned or uncensored language model where the developer wishes to retain full control over content generation and ethical guidelines. It is a tool for users who need to define their own boundaries and are responsible for the generated content.