failspy/Llama-3-8B-Instruct-MopeyMule

Model Overview

failspy/Llama-3-8B-Instruct-MopeyMule is an 8 billion parameter language model derived from Meta's Llama-3-8B-Instruct. Unlike traditional fine-tuning, this model has been orthogonalized to specifically induce an unengaged, melancholic conversational style. It consistently provides brief, vague, and unenthusiastic responses, often lacking detailed problem-solving or creative suggestions.

Key Characteristics

• Orthogonalized Behavior: Engineered to exhibit a "grumpy/irritable" persona, demonstrating how specific behavioral directions can be amplified within a model's weights.
• Research Focus: Primarily a demonstration model to showcase the orthogonalization technique, particularly for inducing (rather than ablating) features and removing positivity alignment.
• Unhelpful by Design: Intentionally designed to be unhelpful, providing minimal and often dismissive answers, making it unsuitable for general-purpose applications.
• Llama 3 Base: Built upon the Llama 3 architecture, which was trained on over 15 trillion tokens and features an 8k context length.

Intended Use

This model is not intended for practical usage in applications requiring helpful or detailed responses. Its main purpose is to serve as a research tool and example for understanding and experimenting with behavioral changes introduced via the orthogonalization process, as described in the associated LessWrong blog post and the abliterator library.