Automating Workforce Scheduling with Large Language Models and Constraints

Sammanfattning

This thesis explores the use of large language models (LLMs) to automate workforce scheduling through natural language interaction. The primary objective is tofine-tune a general-purpose LLM to generate structured scheduling data in JSONformat from natural language prompts. Using a parameter-efficient fine-tuningmethod (LoRA), we trained Microsoft’s Phi-4 models on a domain-specific datasetof Swedish scheduling requests. The model’s performance was evaluated acrossvalidation, test, and generalization datasets using structured accuracy and fieldlevel metrics such as F1 score. The fine-tuned model achieved 84% structuredaccuracy on the validation set and 81.74% on a generalization test set featuringdiverse scheduling scenarios. In contrast to previous work that relied on few-shotprompting, our approach emphasizes reliable structure generation followed byconstraint checking through external Python functions. Comparative results showthat the fine-tuned Phi-4 model outperforms OpenAI’s GPT models in accuracy,though at the cost of generation time. These findings demonstrate the feasibilityand effectiveness of a fine-tuned, locally deployable LLM for reliable and interpretable schedule generation.