GENERATING COOKING RECIPES FROM INGEDIENTS USING T5V1.1 AND FALCON.

Abstract

This thesis explores the application of advanced natural language processing (NLP) models to generate and evaluate culinary recipes based on a given list of ingredients.It introduces a two-stage framework utilizing the T5v1.1 model and the Falcon model. The T5v1.1 model, fine-tuned on a specialized subset of the RecipeNLG datasets, generates coherent culinary recipes including dish titles, ingredient lists,and detailed cooking instructions. In the second stage, the Falcon model, fine-tunedfor instruction-following tasks, reviews the generated recipes by providing feedback on clarity, coherence, and logical consistency, along with suggestions for improvement. Experimental results demonstrate the system's capability to produce relevant and diverse recipes, with the T5v1.1 model attained BLEU-1toBLEU-4scores of 0.592, 0.317, 0.178, and 0.104 respectively, signifying a robust alignment with reference texts. While acknowledging challenges related to data preprocessing and occasional nonsensical outputs, this study highlights the system’s potential for automating culinary content generation and establishes a foundational framework for future intelligent cooking assistants