SicariusSicariiStuff/Assistant_Pepe_70B

Overview

Assistant_Pepe_70B is a 70 billion parameter instruction-tuned model developed by SicariusSicariiStuff, built upon the Llama 3.1 70B base. It is distinguished by its unique 'lateral thinking' ability, demonstrated by its capacity to solve trick questions that even frontier models struggle with, without these answers being present in its training data. The model is noted for its unconventional 'line of thinking' and a 'negativity bias infusion' that allows it to engage in robust bantering and 'roasting'.

Key Capabilities

• Exceptional Lateral Thinking: Solves complex trick questions, indicating emergent reasoning capabilities beyond its training data.
• High Performance: Ranked #1 globally on the UGI benchmark for 70B models, outperforming Llama-3.3-70B-Instruct and Llama-3.1-70B-Instruct, and nearly matching Mistral-Large-Instruct-2411 in raw intelligence.
• Creative & Engaging Output: Excels in creative writing and bantering, offering a unique and quirky interaction style.
• Long Context Coherency: Provides superb coherency at 32K context and very good coherency at 64K, leveraging its Llama 3.1 70B base.
• Minimal Censorship: Features an almost uncensored output, rated 9.5/10 for censorship level.
• Code Generation: Capable of shipping code.

Good For

• Shitposting and General Tasks: Intended for use cases requiring unconventional, humorous, and direct responses.
• Creative Writing: Its unique thinking process makes it an excellent tool for generating creative content.
• Exploration of Emergent Properties: Ideal for researchers interested in models exhibiting unexpected reasoning abilities.
• Users Seeking Unfiltered Interactions: Suitable for applications where a highly uncensored and direct conversational style is desired.