Research Engineer - Midtraining

Periodic Labs; Menlo Park (CA); On-site; USD 250,000 - 350,000; Full time

Job summary

Periodic Labs is seeking a Midtraining Research Engineer to advance scientific reasoning in frontier models. You will curate data, build evals, and run large-scale training experiments to improve base models and lay groundwork for pre-training efforts.

You will collaborate with RL researchers, physicists, and chemists, leveraging self-distillation methods and scaling strategies to push capabilities across thousands of GPUs.

Qualifications

• Experience training LLMs on large-scale data and runs.
• Experience on an evals team supporting production training.
• Hands-on use of self-distillation or on-policy distillation.
• Knowledge of scaling laws and compute-optimal hyperparameters.
• Comfort working across data, evals and training infra.

Responsibilities

• Identify, process, and curate novel data sources for large-scale model training.
• Generate high-quality synthetic data to fill gaps in knowledge and reasoning.
• Build evaluations that correlate with downstream task performance.
• Develop distillation techniques to improve model capability.
• Design and run large-scale training experiments with thousands of GPUs.
• Build tools to study how data choices affect model intelligence.

Skills

Experience training LLMs

Evals team experience

Self-distillation / on-policy distill

Scaling laws & hyperparameters

Data/evals/training infra coordination

Education

Bachelor's degree or equivalent

Job description

We're an AI and physical sciences company building state-of-the-art models to accelerate breakthroughs across materials, energy, and beyond. Backed by world-class investors and growing rapidly, we operate at the pace the frontier requires. Our team brings deep expertise, genuine ownership, and a drive to push the boundaries of what's scientifically possible.

About The Role

We're training frontier models to develop deep scientific knowledge and reasoning for scientific discovery. As a Midtraining Research Engineer, you'll take base models and improve their scientific reasoning: curating and generating data, building evals, and running large-scale training experiments. Your work will also lay the groundwork for our pre-training efforts down the line.

What You'll Do

• Identify, process, and curate novel sources of scientific data for large-scale model training.
• Generate high-quality synthetic data to fill gaps in scientific knowledge and reasoning.
• Build evaluations that correlate with downstream scientific task performance, working closely with RL researchers, physicists, and chemists.
• Develop and apply techniques such as self-distillation and on-policy distillation to improve model capability.
• Design and run large-scale training experiments, partnering with supercompute engineers to scale efficiently across thousands of GPUs.
• Build tools for yourself and the team to investigate how data choices shape model intelligence.

You Will Thrive in This Role If You Have

• Experience training LLMs on curated mixes of trillions of tokens.
• Experience on a dedicated evals team supporting a large production training run.
• Hands‑on use of self‑distillation, on‑policy distillation, or similar methods in a real training pipeline.
• Experience with scaling laws and compute‑optimal hyperparameters.
• Comfort working across data, evals, and training infrastructure.

Especially Strong Candidates May Also Have

• Experience optimizing throughput and reliability for large-scale distributed training runs.
• A background in AI for science or training on specialized domain data (e.g., protein, materials, or other scientific datasets).
• Experience creating evals or synthetic data for non verifiable tasks and tracking performance over live runs.

Mechanics

• Minimum education: Bachelor's degree or similar experience
• Location: Menlo Park, CA (Soon: San Francisco, too)
• Compensation: $250,000–$350,000 + equity
• Visa sponsorship: Yes, we sponsor visas and will do everything we can to assist in this process.