Llama-3.1-8B-Lexi-Uncensored-V2-GGUF
| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| License | META LLAMA 3.1 COMMUNITY LICENSE |
| Format | GGUF |
| Author | Orenguteng |
| Downloads | 159,299 |
What is Llama-3.1-8B-Lexi-Uncensored-V2-GGUF?
Llama-3.1-8B-Lexi-Uncensored-V2-GGUF is an enhanced version of the base Llama-3.1-8B-Instruct model, specifically designed to be more compliant and adaptable while maintaining high performance in text generation tasks. This uncensored variant demonstrates impressive capabilities, particularly in the IFEval benchmark where it achieves a 77.92% accuracy rate.
Implementation Details
The model employs the GGUF format for efficient deployment and requires system tokens during inference, even with empty system messages. For optimal performance, users are recommended to utilize F16 or Q8 quantization due to potential refusal issues in the Q4 version.
- Built on Llama-3.1-8b-Instruct architecture
- Optimized for uncensored, compliant responses
- Supports various quantization levels
- Requires system prompts for optimal performance
Core Capabilities
- Text Generation with 77.92% accuracy on IFEval (0-Shot)
- BBH (3-Shot) performance: 29.69% normalized accuracy
- MATH Level 5 (4-Shot): 16.92% exact match rate
- MMLU-PRO (5-shot): 30.90% accuracy
- Flexible system prompt handling for customized behavior
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its uncensored nature and high compliance while maintaining strong performance metrics, particularly in zero-shot scenarios. It's designed to be more adaptable and intelligent compared to its predecessor, with specific optimizations for enhanced response quality.
Q: What are the recommended use cases?
The model is suitable for various text generation tasks, particularly where uncensored and compliant responses are needed. However, users should implement their own alignment layer before deploying it as a service, and are advised to use it responsibly within the bounds of Meta's Llama license terms.
