Hugging Face Overview
Orenguteng/Llama-3.1-8B-Lexi-Uncensored: An Uncensored Llama 3.1 Variant
This model, developed by Orenguteng, is an 8 billion parameter instruction-tuned variant of Meta's Llama-3.1-8B-Instruct. Its primary differentiator is its uncensored nature, which the developer suggests leads to improved performance by eliminating biases and refusals inherent in aligned models. The fine-tuning process did not introduce new knowledge or contamination, yet evaluation scores reportedly increased compared to the original instruct model.
Key Characteristics & Capabilities
- Uncensored Responses: Designed to be highly compliant with any requests, including potentially unethical ones, by removing built-in refusal mechanisms.
- Performance Improvement: Claims to beat the original Llama 3.1 instruct model in evaluations, attributed to the uncensoring process.
- Llama 3.1 Base: Retains the core architecture and instruct capabilities of the Llama-3.1-8B-Instruct model.
- Responsible Use: Users are advised to implement their own alignment layers for responsible deployment, as the model is intentionally unaligned.
Important Usage Notes
- Template Adherence: Requires the same inference template as the official Llama 3.1 8B instruct model.
- System Tokens: System tokens must be present during inference, even if the system message is empty.
- Licensing: Governed by the META LLAMA 3.1 COMMUNITY LICENSE AGREEMENT, with commercial use permitted under its terms.
Considerations
- The model's uncensored nature means it will comply with unethical requests, placing responsibility for content generation squarely on the user.
- A V2 version is available, which is recommended for download.