Research Engineer, Post-Training Model Evaluations

About the role

Anthropic's production models undergo sophisticated post-training processes to enhance their capabilities, alignment, and safety. Decisions in production training depend on knowing how the model is really behaving. Our team's mission is to have the best observability into production post-training runs at Anthropic: to be the clearest, most trusted view of model quality during training. We shape what gets measured, keep the signal trustworthy and timely, and get to the bottom of surprising results to inform post-training and release decisions.

In this role you will lead the science of how we evaluate production training runs. You'll work out which measurements tell us something real about the model, notice when they stop doing so, and find what we should be measuring but aren't. You'll partner with research teams across every RL domain, bringing their priorities into what we measure and setting the standard for what makes an eval trustworthy across post-training. You'll also be hands-on with the eval fleet day to day, because the best questions about measurement come from watching real results come in during a live run.

Responsibilities

• Steer the eval strategy for production Claude models: study and optimize the eval mix so it gives the most accurate and complete assessment of model quality
• Run and monitor the eval fleet live to understand how each Claude model is developing as it trains
• Advise teams across post-training on eval methodology, and help eval authors bring new evals up to the bar for production
• Investigate regressions in production runs and inform training interventions when appropriate
• Build the dashboards, alerts, and reports that researchers and leadership use to track model quality
  
  

You may be a good fit if you

• Have designed, run, and analyzed evaluations for ML models at scale
• Care deeply about measurement quality, thinking twice before trusting a number
• Can turn ambiguous results into clear recommendations, and are comfortable influencing direction across teams
• Have strong Python skills and are comfortable working with production systems
• Maintain clarity and rigor when debugging complex, time-sensitive issues
• Thrive in controlled chaos and are energized, rather than overwhelmed, when juggling multiple urgent priorities during a live training run
• Care about the societal impacts of your work and about shipping frontier models responsibly

Strong candidates may also have

• Hands-on experience post-training large language models
• Research experience in ML evaluation or benchmarking
• Background in statistics and experimental design
• Experience developing robust evaluation metrics for ML systems