Research, Mid-Training

Lead research and engineering for mid-training to improve LLM capabilities for deployed agents.

• •Own late-stage (mid-)training for large language models, designing data mixes, schedules, synthetic data pipelines, and context-extension methods to sharpen model capabilities for deployed agents.
• •Key Responsibilities Design and iterate high-quality data mixtures and filtering strategies for late-stage training.
• •Drive capability injection across coding, math, and reasoning via curated data and interventions.
• •Develop and evaluate synthetic data pipelines at scale and measure their limits.
• •Research annealing, learning-rate schedules, and compute allocation across training phases.
• •Implement methods to extend effective context length and build robust evaluations.
• •Requirements Deep familiarity with end-to-end LLM training pipelines and late-stage data mixing.
• •Hands-on experience with continual pre-training, annealing, or distributed training at scale.
• •Proficiency in Python and deep learning frameworks (PyTorch).
• •Strong fundamentals in optimization, statistics, and ML theory with experimental rigor.