NASA and IBM Partner to Create INDUS Large Language Models for Cutting-Edge Scientific Research

The IMPACT-IBM team were able to outperform open, non-domain specific language models (LLMs) on various benchmarks related to biomedical tasks, scientific question-answering, and Earth science entity recognition by equipping INDUS with domain-specific vocabulary. This enhancement in performance was achieved by designing INDUS to handle diverse linguistic tasks and incorporate retrieval augmented generation capabilities. With this approach, INDUS can effectively process researcher questions, retrieve pertinent documents, and generate accurate answers.

To cater to applications with stringent latency requirements, the team also worked on creating smaller and faster versions of both the encoder and sentence transformer models within INDUS. This optimization allows for more efficient processing of tasks without compromising on the overall performance of the system. By focusing on enhancing the capabilities of INDUS in different aspects, the team was able to elevate its performance and usability across a range of domains and tasks.