ChiKoi7/Llama-3.1-8B-Lexi-Uncensored-V2-Heretic is an 8 billion parameter language model, a decensored version of Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2, created using the Heretic v1.0.1 tool. This model significantly reduces refusals compared to its base, achieving 2/100 refusals versus 35/100, making it highly compliant with diverse requests. It is designed for applications requiring an uncensored model that adheres to the Meta Llama 3.1 Community License Agreement.
Loading preview...
Model Overview
ChiKoi7/Llama-3.1-8B-Lexi-Uncensored-V2-Heretic is an 8 billion parameter language model derived from Llama-3.1-8B-Instruct. It is a decensored variant of Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2, processed using the Heretic v1.0.1 tool to reduce refusal rates.
Key Characteristics
- Decensored Nature: Achieves a significantly lower refusal rate of 2/100 compared to the original model's 35/100, indicating high compliance with user prompts.
- Heretic Abliteration: Utilizes specific abliteration parameters for
attn.o_projandmlp.down_projto achieve its uncensored behavior. - Llama 3.1 Base: Governed by the Meta Llama 3.1 Community License Agreement, allowing for broad usage including commercial applications.
- System Prompt Flexibility: Supports flexible system prompts, including a simple "." for maximum uncensored and compliant responses.
Usage Considerations
- Ethical Responsibility: Users are advised to implement their own alignment layers if exposing the model as a service, as it is highly compliant with all requests, including potentially unethical ones.
- Quantization Impact: The Q4 quantized versions may exhibit refusal issues, with F16 or Q8 recommended for optimal performance.
- Inference Template: Requires the same inference template as the official Llama 3.1 8B instruct, with system tokens present even for empty system messages.