Research Engineer, Synthetic Data
San Francisco • FullTime
Posted 27d ago
Remote Work Policy
On-site
Employment Type
FullTime
Categories
AI Research Engineer
About the job
As a Research Engineer on the Synth Data team, you will be responsible for inventing and scaling the synthetic data that teaches our models new capabilities. This role involves owning the full data generation loop: identifying missing model capabilities, designing data generation strategies, building reliable pipelines and verifiers, and validating the impact of the resulting data on model performance. You will form hypotheses about beneficial data, build systems for large-scale production, run ablations, and analyze training signals to understand the effectiveness of your work. The datasets you create will directly influence frontier training runs and shape the development of our models.
Responsibilities
- Design end-to-end synthetic data recipes from raw sources into large, diverse, verifiable task sets.
- Build strong verification and anti-hacking mechanisms for reward computation.
- Control data difficulty and diversity to ensure learnability and prevent template collapse.
- Run targeted ablations and analyses to demonstrate dataset impact on capability improvement.
- Iterate on successful data recipes and transform them into reusable infrastructure.
- Partner with research and RL Data teams to translate model capability goals into synthetic data plans.
Requirements
- Research experience on synthetic or post-training data, with a proven track record of shipping data and demonstrating its impact.
- Strong software engineering fundamentals, including experience building data pipelines, agent systems, or evaluation frameworks.
- A keen understanding of what constitutes a good training task: challenging, reliably gradable, and resistant to reward hacking.
- Systematic approach to problem-solving, with attention to both aggregate metrics and raw output analysis.
- Ability to transform open-ended problems into repeatable, high-quality data production systems.