Research Engineer - Midtraining
Research Engineer - Midtraining
Research Engineer - Midtraining
Periodic Labs
Periodic Labs is an innovative AI and physical sciences lab focused on developing advanced models for scientific discovery. The company is rapidly growing and values team members who take ownership and solve problems collaboratively. The Research Engineer role involves training frontier models for reinforcement learning, developing synthetic data generation methods, and collaborating with researchers and engineers to enhance model training efficiency.
Qualification
- Experience in training LLMs on curated mixes of trillions of tokens.
- Ability to calculate scaling laws and compute-optimal hyperparameters.
- Experience in generating billions of tokens of high-quality synthetic data.
- Proven ability to build evaluations that correlate with downstream task performance.
- Strong collaboration skills to work with interdisciplinary teams.
Responsibility
- Train frontier models to become knowledgeable scientific experts for reinforcement learning.
- Develop methods for synthetic data generation, distillation, and continual learning at scale.
- Collaborate with RL researchers, physicists, and chemists to create evaluations for scientific data curation.
- Work with supercompute engineers to scale compute-efficient LLM training across thousands of GPUs.
- Build high-performance tools to investigate the relationship between data and intelligence.
