Marah Abdin

Marah Abdin helped build the synthetic data work behind Microsoft's Phi model family (lead author on the Phi-4 and Phi-4-reasoning technical reports) and now leads synthetic data at Poolside, where her team composes the 5T+ synthetic token corpus behind the Laguna models. Her NeurIPS 2025 talk, 'Entropy by Design: Synthetic Data at Scale,' lays out an entropy-aware approach with structural and cognitive variability levers to keep diversity and quality high during pretraining — making her a rare practitioner who has shipped production-scale synthetic data systems at multiple frontier labs.