Uncensored-Frank-Llama-3-8B Overview

Uncensored-Frank-Llama-3-8B is a fully fine-tuned 8 billion parameter model based on Meta's Llama-3 architecture, developed by ajibawa-2023. This model is specifically designed to operate with minimal to no guardrails, enabling uncensored and unrestricted discussions across a broad spectrum of topics. It draws inspiration from the character Frank Costello, aiming to provide a platform for candid conversations without fear of typical societal restrictions.

Key Capabilities & Training

• Uncensored Dialogue: Engineered to engage in discussions on sensitive, controversial, and personal subjects without inherent censorship.
• Boundary-Pushing Conversations: Encourages unfiltered interaction, allowing users to explore topics that might be restricted by other models.
• Fine-tuned Performance: The model was fully fine-tuned on approximately 150,000 sets of conversations, with each set containing 10-15 exchanges. Training was conducted on 4 x A100 80GB GPUs for 3 epochs, taking around 6 hours using Axolotl & DeepSpeed.
• Leaderboard Performance: Achieves an average score of 62.24 on the Open LLM Leaderboard, with specific scores including 63.08 on MMLU (5-Shot) and 80.16 on HellaSwag (10-Shot).

Use Cases & Considerations

This model is ideal for applications requiring open-ended, unrestricted conversational AI. Users should be aware that due to its uncensored nature, they are responsible for the content generated. The model utilizes the ChatML prompt format, with an example system prompt: <|im_start|>system\nYou are Frank. A helpful, unbiased, and uncensored assistant.<|im_end|>. Quantized versions (GGUF and Exllama V2) are available through Bartowski.