Ernie Chang

About

I am a Staff Research Scientist at Meta AI, where I lead data curation and pretraining for the open-source on-device MobileLLM and MobileMoE model series. My research focuses on data-centric training of foundation models: deciding what data a model should see, in what proportion, and in what order, so that a fixed compute budget buys the most capability.

Research interests: data curation and mixture optimization, data-influence estimation, synthetic data generation, large-scale pretraining and midtraining, on-device and streaming models, multilingual and low-resource NLP.

Open-Source Models

• MobileLLM-R1, on-device reasoning model (data curation and pretraining lead)
• MobileLLM-Pro, 1B on-device LLM; the language backbone selected for Meta’s on-device VLM
• MobileMoE, sparse mixture-of-experts variant for on-device deployment

Selected Publications

• MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes
  C. Zhao*, E. Chang*, Z. Liu*, et al.
  ICLR 2026 Data curation lead
• MobileMoE: Scaling On-Device Mixture of Experts
  Y. Chen, H. Huang, E. Chang, et al.
  arXiv, 2026
• Automixer: Checkpoint Artifacts as Automatic Data Mixers
  E. Chang, Y. Li, P. Huber, V. Vogeti, D. Kant, Y. Shi, V. Chandra
  ACL 2025 First author
• MobileLLM-Pro Technical Report
  P. Huber*, E. Chang*, W. Wen*, I. Fedorov*, et al.
  arXiv, 2025 Data curation lead
• Scaling Parameter-Constrained Language Models with Quality Data
  E. Chang, M. Paltenghi, Y. Li, et al.
  EMNLP 2024 First author
• DART: A Lightweight Quality-Suggestive Data-to-Text Annotation Tool
  E. Chang, J. Caplinger, A. Marin, X. Shen, V. Demberg
  COLING 2020 · Best System Paper Award First author

* Equal contribution

Full list on Google Scholar.