The role involves working as a data scientist within Microsoft's AI Superintelligence Team, focusing on developing post-training methods for frontier models. The team aims to enhance model capabilities and ensure they align with human values while delivering impactful AI solutions. Responsibilities include designing evaluations, building data pipelines, and conducting experiments to improve model performance. Candidates should have experience with large language models and a strong engineering background.

• Hands-on experience with large language models, including training or applying them in production.
• Experience in designing and running post-training experiments, such as evaluations and ablations.
• Proven ability to build and own scalable data pipelines for training and evaluation data.
• Strong Python skills for machine learning experimentation, data processing, and analysis.
• Solid statistical, experimental, and general engineering fundamentals.

• Design evaluations of advanced model capabilities and use them to drive rapid, high-signal iteration loops.
• Work with vendors to produce high-quality evaluation and training data.
• Build data pipelines to produce high-quality evaluation and training data.
• Build data flywheels to hill-climb on model weaknesses using data from various surfaces where models are deployed.
• Ensure optimal quality, quantity, and coverage of data across post-training stages.
• Run post-training experiments and ablations to produce models that improve evaluation scores.
• Embody the team's culture and values.

• Employees at Microsoft are often offered comprehensive, “world-class” benefits—including health and mental-wellness programs, competitive pay with bonuses and stock awards, and retirement/savings options. Time-off and flexibility are common, with generous vacation and holidays, parental and caregiver leave, and flexible work schedules, alongside learning support, employee resource groups, product discounts, and matching-gifts/volunteering programs. Specific benefits can vary by region.