Relm – local LLMs as base-R objects, with interpretability (github.com)

Relm is a newly launched R package aimed at integrating local large language models (LLMs) into the R programming environment, enhancing its capabilities for scientific research in data and AI. The package incorporates a Rust native core that interfaces with a modified llama.cpp, enabling researchers to seamlessly leverage machine learning features such as text generation, tokenization, and mechanistic interpretability—all while using base-R functions. This package serves as a significant tool for researchers looking to conduct topic modeling and other AI/ML tasks without relying on Python, making advanced analysis more accessible. The significance of relm lies in its potential to simplify the process of integrating AI tools into the R ecosystem, targeting researchers in fields like life sciences and experimental data analysis. Key features include functions for activation tracing, embeddings, and text generation, all structured to return standard R data types like data.frames and matrices. The release roadmap indicates future enhancements, including support for vision tasks in subsequent versions. With validation against independent numerical references, relm not only aims to empower the R community but also promises a pathway for more interpretable AI through its mechanistic interpretability framework.