RecipeGPT implementation

Abstract

The culinary domain offers significant opportunities for leveraging AI to simplify and enhance recipe-related interactions. This project focuses on developing an intelligent recipe chatbot that performs two primary tasks: (1) generating cooking instructions from a list of ingredients, and (2) predicting ingredients based on given instructions. Using the ”51k Recipes” dataset, which contains detailed recipe information, we explored and fine-tuned various large language models (LLMs), including GPT-2, T5 (Text-To-Text Transfer Transformer), Gemini, etc, to identify the most effective approach for these tasks. Key challenges addressed include handling large datasets, ensuring the relevance and accuracy of outputs, and optimizing models for bidirectional task processing. Through comprehensive preprocessing and fine-tuning using PyTorch, the chatbot was trained to deliver coherent and contextually appropriate responses. The models were evaluated using metrics such as BLEU scores and human feedback, with notable improvements observed in instruction fluency and ingredient prediction accuracy. This project demonstrates the potential of LLMs in culinary applications and sets the ground- work for creating versatile, user-friendly tools for home cooks, professional chefs, and culinary learners. Future work involves refining dataset quality, exploring larger instruction-tuned models, and integrating the chatbot into practical applications such as voice assistants and cooking platforms.