cosmicvalor/mistral-orthogonalized

Model Overview

cosmicvalor/mistral-orthogonalized is a 7 billion parameter language model built upon the Mistral architecture. Its development was inspired by the research paper "Refusal in LLMs is mediated by a single direction," which explores how refusal behaviors are encoded within large language models.

Key Characteristics

• Orthogonalization Method: This model incorporates a specific orthogonalization technique, aiming to modify or understand the mechanisms behind refusal in LLMs.
• Research Focus: It is explicitly designated for research purposes, providing a tool for academics and developers to investigate model alignment and safety.

Intended Use

This model is primarily designed for:

• Investigating LLM Refusal: Researchers can use this model to study how refusal behaviors manifest and can be influenced or controlled within LLMs.
• Exploring Alignment Techniques: It serves as a platform for experimenting with methods to steer model outputs and understand internal representations related to safety and refusal. An exl2 version is also available for optimized inference.