Pre-Training Data Engineer - ML Research Pipelines

Liquid-Ai; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Benefits offered by this job

Equity

Unlimited PTO

Health coverage

Paid leave

Job summary

Liquid AI is seeking a data-focused engineer to shape the pre-training data for its foundation models. You will work on sourcing, filtering, synthesizing, and mixing data, collaborating with pre-training researchers to run experiments and implement end-to-end data pipelines.

You will design data mixtures, curate curricula for pre-training, and build synthetic data systems using LLMs and prompting. You’ll also ensure data lineage and reproducibility across training runs.

Qualifications

• Hands-on pre-training data work: web-scale sourcing, curation, synthetic data generation.
• Ability to reason from data decisions to training effects on model size/architecture/dynamics.
• Experiments to test data changes with baselines, scaling ladders, and held-out loss.

Responsibilities

• Source web-scale and public datasets, and build pipelines to 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, prompting, and distillation to cover data gaps.
• Run proxy experiments and ablations to measure the impact of data changes on model quality.
• Build evaluation sets with clear data lineage for reproducibility of training runs.

Skills

Research mindset

Hypothesis-driven

End-to-end ownership

Education

MS/PhD preferred

BS with 5+ years experience

Tools

Python

PyTorch

Job description

Liquid AI is seeking a data-focused engineer to shape the pre-training data for its foundation models. You will work on sourcing, filtering, synthesizing, and mixing data, collaborating with pre-training researchers to run experiments and implement end-to-end data pipelines.

You will design data mixtures, curate curricula for pre-training, and build synthetic data systems using LLMs and prompting. You’ll also ensure data lineage and reproducibility across training runs.