Software Engineer, Pretraining

About the role

We’re looking for Software Engineers to build the data systems behind our frontier coding models’ initial training. You’ll work on large-scale crawling, data platform, and pipeline infrastructure, turning raw dumps into the datasets our models train on, and making iteration with researchers fast and reliable.

What you’ll do

on the Data Quality Team:

• Build and own high-throughput, fully telemetered data pipelines that process frontier-scale data with end-to-end traceability. If something breaks or drifts, your systems will tell us before the training run does.

• Train and ship models that classify, rank, filter, clean, and identify data at extreme throughput. These models have to be both accurate and fast enough to sit in the critical path without becoming the bottleneck.

• Design and run scaling-ladder experiments on data-mixture, repeatability, and quality depth that turn “this dataset feels good” into hard evidence the training team can trust.

• Partner tightly with Data Acquisition to hunt down missing or low-quality sources, and with the training teams to close the loop on what actually moves loss and downstream evals.

• Treat data quality as a systems problem and a research problem. You will write performance-critical code one week and design careful experiments the next.

on the Data Platform Team:

• Build the platform that turns raw web, code, multimodal, and acquired data into training-ready datasets for frontier pretraining runs.

• Own the pipelines, orchestration, and tooling that make pretraining data iteration fast, reliable, observable, and reproducible at scale.

• Create clear signals for data quality, lineage, freshness, and pipeline health so researchers can trust what goes into each run.

• Partner with initial training, crawling, data quality, and acquisition teams to turn new data ideas into measurable improvements in loss, evals, and model capability.

on the Crawling Team:

• Build and scale the web crawling systems that discover, schedule, fetch, and parse high-quality documents across the open web for initial training.

• Improve URL seeding, scoring, and fair host scheduling so crawl capacity lands on the hosts and pages that matter most for model quality.

• Raise crawl success and parsing quality — defeating antibot failures, improving extractors, and capturing content we previously could not get cleanly.

• Debug and harden complex crawl infrastructure end-to-end for availability, recovery, and ingestion lag, and automate delivery of crawl datasets into the data pipeline.

• Work independently (and alongside AI agents) and partner with Data Quality and Data Platform so new coverage shows up as better tokens in training runs.