Member of Technical Staff - ML Research Engineer, Pre-Training Data

Required

• We need someone who:
• Thinks like a researcher, ships like an engineer: You build the pipelines and you make the model-facing calls about what goes into them. You don't hand off either half.
• Leads with a hypothesis: Before changing a dataset, you can say what it should do to the model, what you'll measure, and what could fool you, such as contamination or a noisy benchmark. You hold views on pre-training pitfalls, and you test them.
• Owns problems end to end: You take ambiguous work from open question to shipped result with little direction or process, do the unglamorous parts that get it there, and don't care whose job it was.
• Source web-scale and public datasets, and build the pipelines that filter, deduplicate, and decontaminate them
• Design data mixtures and curricula for pre-training and midtraining, and decide what goes into each run
• Build synthetic data generation systems using LLMs, structured prompting, and distillation to cover what web data can't
• Run proxy experiments and ablations to measure what each data change does to model quality, and decide what scales up
• Build evaluation sets you can trust, and keep data lineage clear enough that any training run can be traced and reproduced
• Experience & Capabilities
• Deep, hands-on pre-training data work you did yourself: web-scale sourcing and curation, synthetic data generation, mixing and curriculum design, deduplication and decontamination. This is where depth matters most.
• The ability to reason from a data decision to its effect on training: how data interacts with model size, architecture, and training dynamics, why more data can hurt, and which changes matter as models scale