Novel recipe generation using LLMs

Abstract

Generating novel recipes within specific cuisine constraints is a challenging task that combines creativity with cultural authenticity. This study explores the capabilities of fine-tuned large lan guage models (LLMs), including GPT-2, Llama-3.1-8b, and Mistral-7B-Instruct-v0.3, to generate innovative recipes while adhering to predefined culinary styles. By leveraging the contextual and generative strengths of these models, we aim to create recipes that balance novelty with adherence to the cuisine’s defining characteristics. To evaluate the quality of the generated recipes, we employ intrinsic metrics such as BLEU, ME TEOR, BERTScore, and Perplexity. These metrics provide a multifaceted assessment of linguistic diversity, semantic alignment, and fluency. The results reveal how effectively each model captures the essence of the target cuisine while introducing creative variations. Our findings highlight the trade-offs between generating novel content and preserving authentic culinary traits, offering valuable insights into the potential of LLMs for constrained creative tasks. This research demonstrates the applicability of AI-driven approaches to culinary innovation and provides a foundation for future exploration in cuisine-constrained text generation.