Leveraging large language models to maintain a branded food product database

Area of Science:

• Nutritional Science
• Computational Biology
• Data Science

Background:

• Branded food data is crucial for understanding dietary patterns and the global food environment.
• Challenges include data volume, rapid changes, variable quality, and diverse sources.

Purpose of the Study:

• To develop an automated pipeline using large language models (LLMs) for processing branded food data.
• To enable ingredient-level analysis, quantity estimation, and undeclared nutrient content calculation.

Main Methods:

• Development of a fully automated pipeline powered by LLMs.
• Collection, standardization, and enrichment of branded food data.
• Evaluation of LLM performance through fine-tuning and comparison with human experts.

Main Results:

• A fine-tuned LLM significantly outperformed human experts in parsing and mapping product data.
• Non-fine-tuned LLMs showed inadequate performance, while fine-tuning improved results substantially.
• LLMs offer a scalable and consistent approach to data curation, surpassing individual human expert performance.

Conclusions:

• LLMs provide a transformative, scalable solution for managing and analyzing complex, large-scale branded food databases.
• Automated processing enhances data standardization and facilitates detailed ingredient-level insights.
• This approach has the potential to revolutionize food data management and nutritional research.