Research, Mid-Training

📋 Description

• Own late-stage training decisions: data mix, annealing, context length.
• Bridge pre-training and post-training; no strict research vs engineering split.
• Design high-quality data mixtures for late-stage and annealing runs.
• Drive coding, math, and reasoning capability via data strategies.
• Develop synthetic data pipelines for scalable training signals.
• Optimize multi-stage schedules and compute/data balance.

🎯 Requirements

• LLM training pipeline end-to-end (pre- to post-training).
• Experience with continual pre-training, annealing, or late-stage data mixing.
• Strong data-quality intuition; ability to filter and curate at scale.
• Experience with synthetic data pipelines for capability improvement.
• Proficiency in Python and PyTorch; debug distributed training.
• Strong fundamentals in optimization, statistics, ML theory.
• Track record of original contributions: publications, OSS, etc.
• Comfortable in ambiguous, fast-moving environments.

🎁 Benefits

• Small, selective team; research and product move together.
• Compute is not a constraint; thousands of GPUs available.
• Environment rewards speed, autonomy, and depth.