Research Scientist, Post-Training

DatologyAI

3d•$180,000 - $300,000•Hybrid

About The Position

DatologyAI is seeking a Research Scientist to lead work on post-training data curation for foundation models. The role involves designing and implementing algorithms for generating and improving instruction, preference, and other post-training datasets. The scientist will also explore how to jointly optimize data across pre-training and post-training stages. This position requires strong scientific judgment, fluency with deep learning literature, and a drive for real-world impact. The role offers autonomy, close collaboration with engineers and product teams, and the opportunity to shape the future of data curation at DatologyAI. The position is based in San Mateo, CA, with 4 days a week in the office.

Requirements

3+ years of deep learning research experience
Experience with post-training large vision, language, and multimodal models
Experience with post-training algorithm development, data curation, and/or synthetic data methods for preference-based tuning (e.g. DPO, RLVR, RRHF), alternative supervision & self-supervision techniques (e.g. self-training, chain-of-thought distillation), and SFT (e.g. instruction tuning, demonstration fine-tuning)
Post-training tooling development and engineering experience
Strong understanding of the fundamentals of deep learning
Sufficient software engineering + deep learning framework (PyTorch or a willingness to learn PyTorch) skills to conduct large-scale research experiments and build production prototypes.
Demonstrated track record of success in deep learning research, whether papers, tools, or other research artifacts.

Nice To Haves

Experience with data management and distributed data processing solutions (e.g. Spark, Snowflake, etc.)
Experience building + shipping ML products
Adaptability, combined with exceptional communication and collaboration skills.

Responsibilities

Conduct research on how to algorithmically curate post-training data, including generating and refining preference and instruction-following data, curating capability- and domain-specific data, and making post-training more effective, controllable, and generalizable.
Pursue research on end-to-end data curation, focusing on how to curate pre-training data to improve model post-trainability and how to jointly optimize pre- and post-training data curation to maximize final performance.
Source, vet, implement, and improve promising ideas from the research literature and original concepts.
Conduct research guided by concrete customer needs and product improvements.