Warren J. Gross

This paper proposes PPL-Factory, a data selection framework for efficient fine-tuning of large language models using task-aware and budget-aware perplexity-based scores.