ModelCypher: Open-Source Toolkit for Analyzing the Geometry of LLMs

This article discusses ModelCypher, an open-source toolkit designed to analyze the internal geometry of Large Language Models (LLMs). The author aims to demystify LLMs by providing tools to measure and understand their inner workings before token emission. The toolkit includes features like cross-architecture adapter transfer, jailbreak detection, and implementations of machine learning methods from recent papers. A key finding is the lack of geometric invariance in "Semantic Primes" across different models, suggesting universal convergence rather than linguistic specificity. The author emphasizes that the toolkit provides raw metrics and is under active development, encouraging contributions and feedback.

• •ModelCypher is an open-source toolkit for analyzing LLM geometry.
• •It offers features like adapter transfer and jailbreak detection.
• •The toolkit aims to provide insights into LLM behavior before token emission.