Pre-Training Data Quality Engineer | ML QA & Pipelines

Reflection; San Francisco (CA); On-site; USD 140,000 - 210,000; Full time

Benefits offered by this job

Top-tier compensation

Stock options

Health & wellness

Meals provided

Parental leave

Unlimited vacation (US)

Visa sponsorship

Team events

Job summary

Reflection is a research lab focused on open models and open data for AI. The Data Team is seeking engineers who can own upstream data quality for LLM pre-training, turning fuzzy quality notions into measurable standards that scale across large campaigns.

You will work with world-class researchers, designing automated QA and data pipelines, and collaborating with researchers, engineers, and operations partners to ensure data reliability and impact on model performance.

Qualifications

• Strong engineering background with data pipelines and QA systems for pre-training data.
• Ability to translate quality concerns into concrete signals and feedback.
• Experience validating automated quality checks and rule-based systems.

Responsibilities

• Own upstream data quality for LLM pre-training across languages and modalities.
• Collaborate with research and pre-training teams to define measurable quality signals.
• Design, validate, and scale automated QA methods for large data campaigns.
• Build reusable QA pipelines delivering high-quality data to model training.
• Monitor data quality and iterate on standards and acceptance criteria.

Skills

Python

ML workflows

Data pipelines

Quality signals

LLM evaluation

Tools

SQL

Airflow

Job description

Reflection is a research lab focused on open models and open data for AI. The Data Team is seeking engineers who can own upstream data quality for LLM pre-training, turning fuzzy quality notions into measurable standards that scale across large campaigns.

You will work with world-class researchers, designing automated QA and data pipelines, and collaborating with researchers, engineers, and operations partners to ensure data reliability and impact on model performance.