Principal Research Scientist, Synthetic Data Generation

This is a fully remote position, open to applicants in California, +3 more states.

📋 Description

• Establish the technical vision for synthetic data generation in NVIDIA's frontier model initiatives

• Create and develop open-source libraries within the NVIDIA NeMo framework

• Produce synthetic datasets encompassing text, code, structured, and multimodal data for both pre- and post-training of LLMs

• Construct and enhance LLM-based data generation pipelines with automated quality assessment

• Innovate synthetic trajectories, multi-turn interactions, function calling, executable environments, reward modeling, and verifiable-reward data for training in agentic and tool-use contexts

• Propel advancements in multimodal synthetic data generation for images, documents, videos, and audio

• Promote safe and privacy-preserving synthesis through methods like differential privacy, anonymization, and de-identification

• Create and sustain open-source libraries and SDKs with well-defined APIs and comprehensive documentation

• Champion software excellence by utilizing modern tools, configurable architecture, and professional Git/CI-CD practices

• Contribute original research to leading conferences in machine learning and AI

• Collaborate with research, engineering, product, model teams, and external laboratories

• Guide and mentor scientists and engineers within the team

⛳️ Requirements

• PhD in Computer Science, Machine Learning, Statistics, or a related field, or equivalent experience

• Over 15 years of engineering and research experience in synthetic data generation, generative modeling, multimodal machine learning, or related disciplines

• Profound technical knowledge of LLMs, pre-training, post-training, reinforcement learning stages, and inference frameworks such as vLLM or TGI

• Demonstrated success in developing or maintaining software libraries utilized by a wide developer community

• Experience in building and optimizing scalable data pipelines for large-scale model training, with a focus on throughput, distributed inference, and cost-effectiveness at cluster scale

• Strong publication history in top-tier venues such as NeurIPS, ICML, ICLR, ACL, or similar

• Notable open-source contributions in ML or data tooling, with significant community adoption

• Familiarity with multimodal generation or understanding, including vision-language, document AI, video, or audio

• Experience in generating data for agentic, tool-use, or reinforcement-learning post-training, including the design of RL environments

• Background in differential privacy, de-identification, or synthetic data applicable to regulated sectors like healthcare, finance, or government

• Experience in influencing model training decisions at the frontier scale, or direct collaboration with pre-training and post-training teams

🏝️ Benefits

• Equity

• Benefits