PPL-Factory Selects Fine-Tuning Data Under an Explicit Compute Budget

arXiv·medium signal

Hang Zhang and Warren J. Gross (McGill) propose PPL-Factory (arXiv 2607.18199), a perplexity-driven data selection method that is both task-aware and budget-aware — it picks informative training samples for a fixed compute ceiling rather than assuming an unlimited corpus. It spans plain language modeling through reasoning fine-tunes. The practical read for builders: if you are fine-tuning a small model on a limited GPU budget, sample selection is a first-class knob, not a preprocessing afterthought.