Research Engineer, RL & LLM Post-Training — Scale ML Infra

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

Benefits offered by this job

Competitive cash and equity (>90th pct

Ownership and autonomy

Lunch onsite

Weekly snack orders

Visa sponsorship & relocation support

Health, vision, dental benefits

401K match

Job summary

Preference Model is building automated ML research engineering to advance RL environments and post-training for large language models. We seek machine learning Research Engineers or Research Scientists to blend research and engineering, implementing novel approaches and shaping research directions.

You will train and evaluate models on proprietary RL environments, architect training infrastructure, and optimize end-to-end runs to accelerate research cycles.

Qualifications

• Experience running end-to-end LLM post-training pipelines of models sizes 7B+.
• Proficiency in Python and PyTorch or JAX.
• Experience with at least one modern RL training framework.
• Experience building and operating ML infrastructure at scale.

Responsibilities

• Train and evaluate models on proprietary RL environments to validate data quality and task coverage.
• Architect and optimize RL training infrastructure and experiment management.
• Design, implement, and test training environments and methodologies for RL agents.
• Profile and optimize training runs end-to-end to increase throughput.

Skills

End-to-end LLM post-training

Python

PyTorch or JAX

RL training framework

ML infrastructure

Tools

Verl

OpenRLHF

Job description

Preference Model is building automated ML research engineering to advance RL environments and post-training for large language models. We seek machine learning Research Engineers or Research Scientists to blend research and engineering, implementing novel approaches and shaping research directions.

You will train and evaluate models on proprietary RL environments, architect training infrastructure, and optimize end-to-end runs to accelerate research cycles.