Jerry Petersen
Generative artificial intelligence tools are powered by foundational large language models trained on massive general-purpose datasets that do not contain domain-specific data and so are susceptible to hallucinations, a major concern for government agencies, especially those involved in critical missions like defense, security and health, according to Dave Vennergrund, vice president of AI and data insights at General Dynamics Information Technology.
Fine-Tuning LLMs
To reduce GenAI hallucinations, data science teams can follow a 6-step process to fine-tune the underpinning LLMs with domain-specific data in order to increase accuracy of the output of the Gen AI tool in the relevant domain, Vennergrund said in a column posted Thursday on the GDIT website.
The 6-step process involves selecting a foundational LLM and training dataset that are relevant to one’s task; preparing the data for use; initializing the foundational model and specifying parameters relevant to the task; setting up evaluation functions; actually fine-tuning the LLM through training; and assess and optimizing the model’s performance.
Support From GDIT
Vennergrund notes that the process, though straightforward, is resource-intensive, but it can result in cost reduction because compared to training an LLM from scratch, fine-tuning is more economical. He also said his company is equipped to provide government agencies seeking to fine-tune their LLMs with the resources to carry out the work.
The GDIT VP went on to say that using Luna AI Digital Accelerator, his company and its partners are working to develop capabilities that would simplify AI integration into the mission of government agencies. Notably, the collaboration has resulted in the implementation of a capability that detects and corrects AI hallucinations in real-time.
