Research Scientist, Meta Platforms, Inc.
3 papers at NeurIPS 2025
We present Group-MATES, an efficient group-level data selection method that optimizes the speed-quality frontier for LLM pretraining. On DCLM, Group-MATES achieves up to 9.4% relative accuracy gain and 1.75× faster training than random selection.
We generalize CLIP training to worldwide web-scale, with +0.8% better than English only counterpart on zero-shot ImageNet classification (no compromise), SoTA on zero-shot multilingual: 57.4% on CVQA and 50.2% on Babel-ImageNet.
