We've established an automated pipeline for fine-tuning open-weight Large Language Models (LLMs) specifically for materials science workflows on our GPU infrastructure.
Generic AI models often struggle with the precise terminology and complex relationships inherent in materials science ontologies. To address this, we have developed a robust, automated fine-tuning pipeline utilizing the Unsloth framework.
Running on our dedicated GPU cluster nodes, this pipeline uses Distributed Data Parallel (DDP) techniques to efficiently train models on our curated datasets. The result is a suite of highly specialized models capable of drafting better mappings, more accurate SPARQL queries, and deeper metadata summaries than out-of-the-box providers.