Member of Technical Staff - ML Research Engineer, Pre-Training Data
- Compensation
- Not disclosed
- Location
- San Francisco
- Work setup
- Hybrid
- Seniority
- Not specified
- Employment
- Not specified
- Experience
- 1+ years
- Category
- Software engineering
- Posted
Skills
Must have
- 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
- 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.
- 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
- 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 model
- Experiments you designed to test data changes, such as baselines, scaling ladders, small-scale proxies, and held-out loss, with a clear account of what a change did and why
- Production data pipelines you built and owned, not only research prototypes
- Solid ML fundamentals: you've trained, evaluated, and iterated on models in Python (PyTorch preferred)
- 3+ years of relevant experience with an M.S., 1+ year with a Ph.D., or 5+ years with a B.S.
Full description
About Liquid AI
Spun out of MIT CSAIL, we build general-purpose AI systems that run efficiently across deployment targets, from data center accelerators to on-device hardware, ensuring low latency, minimal memory usage, privacy, and reliability. We partner with enterprises across consumer electronics, automotive, life sciences, and financial services. We are scaling rapidly and need exceptional people to help us get there.
The Opportunity
You'll join the Pre-Training Data team, which decides what Liquid Foundation Models learn from: which data we source, filter, synthesize, and mix, and in what order. Public data sources are plateauing, so model quality increasingly depends on purpose-built datasets and on knowing which data changes actually help.
We treat data as a research problem, not an infrastructure problem. You'll work alongside our pre-training researchers on a small team, run the experiments that settle data questions, and build the pipelines that carry the answers into training runs.
What We're Looking For
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.
The Work
- 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 & CapabilitiesWhat you'll need:
- 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.