3 hours ago
Responsibilities
- Source web-scale and public datasets and build pipelines for filtering, deduplication, and decontamination.
- Design data mixtures and curricula for pre-training and midtraining and determine the contents of training runs.
- Build synthetic data generation systems using LLMs, structured prompting, and distillation.
- Run proxy experiments, scaling ladders, ablations, and held-out-loss evaluations to measure data changes.
- Build evaluation sets and maintain clear, reproducible data lineage for training runs.
- Own a critical pre-training data pipeline end to end and integrate datasets that improve model performance.
Requirements
- Deep hands-on experience with web-scale data sourcing and curation, synthetic data generation, data mixing, curriculum design, deduplication, and decontamination.
- Ability to connect data decisions to model training effects, including interactions with model size, architecture, and training dynamics.
- Experience designing experiments to evaluate data changes, including baselines, scaling ladders, small-scale proxies, and held-out loss.
- Experience building and owning production data pipelines rather than only research prototypes.
- Solid ML fundamentals and experience training, evaluating, and iterating on models in Python; PyTorch is preferred.
- 3+ years of relevant experience with an M.S., 1+ year with a Ph.D., or 5+ years with a B.S.
- Helpful qualifications include distillation, long-context or architecture-specific data strategy, pretrained-model evaluation frameworks, post-training data work, efficient or on-device models, startup experience, and connecting model work to product needs.
Benefits
- Competitive base salary plus equity.
- Unlimited PTO, company-wide Refill Days, and company-wide long weekends.
- Comprehensive health coverage, varying by location.
- Paid leave for pregnancy, parenting, medical needs, and family caregiving.
Categories
Data EngineeringML Engineering
About Liquid AI
Liquid AI builds general-purpose AI systems that run efficiently from data center accelerators to on-device hardware, emphasizing low latency, memory efficiency, privacy, and reliability. The company partners with enterprises in consumer electronics, automotive, life sciences, and financial services to deploy and benchmark models for real-world workloads. Founded in 2023 out of MIT CSAIL and headquartered in Cambridge, Massachusetts, it is privately held.