Active Learning Methodology in LLMs Fine-Tuning

Active learning (AL) presents a valuable approach for fine-tuning large language models (LLMs) by optimizing the selection of training data to enhance model performance. This study introduces a methodology integrating human expertise and synthetic data generation to create robust datasets. Our focus is on addressing gender bias in Italian job advertisements, aiming to improve LLM accuracy in identifying discriminatory language. The method-ology involves a multi-step process: constructing a representative seed dataset, expanding it with synthetically generated data, and iteratively refining the model through fine-tuning loops. Preliminary results demonstrate the potential of AL in reducing the annotation workload while maintaining high performance in bias detection tasks. Future work will extend this approach to other discrimination categories and linguistic variations.