Research Engineer - Midtraining
Periodic Labs2 months ago
Base Salary
$250k - $350k/yr
Responsibilities
- Identify, process, and curate novel scientific data sources for large-scale model training.
- Generate high-quality synthetic data to address gaps in scientific knowledge and reasoning.
- Build evaluations correlated with downstream scientific task performance in collaboration with RL researchers, physicists, and chemists.
- Develop and apply self-distillation and on-policy distillation techniques to improve model capability.
- Design and run large-scale training experiments with supercompute engineers across thousands of GPUs.
- Build tools to investigate how data choices shape model intelligence.
Requirements
- Experience training LLMs on curated mixes of trillions of tokens.
- Experience with mid-training or pre-training at scale; big-lab experience is a strong plus.
- Experience on a dedicated evaluations team supporting a large production training run.
- Hands-on experience with self-distillation, on-policy distillation, or similar methods in a real training pipeline.
- Ability to calculate scaling laws and compute-optimal hyperparameters.
- Comfort working across data, evaluations, and training infrastructure.
- Experience optimizing throughput and reliability for large-scale distributed training runs is especially valuable.
- Background in AI for science or training on specialized scientific datasets such as protein or materials data is especially valuable.
- Experience tracking evaluations and driving interventions during a live large training run is especially valuable.
- Bachelor's degree or similar experience required.
Benefits
- Equity is included in the compensation package.
- Visa sponsorship is available.
- The role is based in Menlo Park, California, with San Francisco planned as an upcoming location.
Categories
AI Research
About Periodic Labs
Periodic Labs builds AI scientists and autonomous laboratory systems that design experiments, simulate physical processes, and verify predictions. Its software and hardware platforms serve industrial and academic R&D teams working on materials, chemistry, and other continuum-physics problems, offered as tools and services to accelerate discovery. The company is privately held, founded in 2025, and headquartered in Menlo Park, California.