Cognition01.05.2026
Research, Mid-Training
Полная занятостьОфис
Обязанности
- 01Design and iterate on high-quality data mixtures for late-stage and annealing training runs, developing principled methods for sourcing, filtering, and weighting data to sharpen model capabilities without degrading general performance
- 02Drive targeted improvements in coding, mathematics, and long-horizon reasoning through curated data strategies and training interventions, translating research insights into measurable capability gains on our agents
- 03Develop and evaluate synthetic data pipelines that generate training signal at scale, understanding the limits and failure modes of synthetic approaches and building methods that hold up in production training runs
- 04Research and optimize multi-stage learning rate schedules, warmup strategies, and compute allocation across training phases, understanding how schedule choices interact with data distribution and model behavior
- 05Research and implement methods for extending effective context length without degrading short-context performance, including positional encoding strategies, data construction, and targeted evaluation
- 06Build evals that distinguish real capability improvements from benchmark overfitting, closing the loop between training decisions and what actually matters for Devin and our other systems in deployment
- 07Measure how mid-training interventions scale with compute and data, developing new approaches when existing methods hit ceilings, expecting both rigorous empiricism and original thinking
Требования
- 01Deep familiarity with the LLM training pipeline end to end: pre-training data, optimization, architecture, and how mid-training and post-training interact
- 02Hands-on experience with continual pre-training, annealing, or late-stage data mixing for large models
- 03Strong intuition for data quality: what makes a dataset useful for training, how to filter and curate at scale, and how data mix choices compound across evals
- 04Experience developing or evaluating synthetic data pipelines for capability improvement
- 05Proficiency in Python and deep learning frameworks (PyTorch); comfortable debugging distributed training at scale
- 06Strong fundamentals in optimization, statistics, and ML theory; able to distinguish real effects from noise, instability, and overfitting
- 07A track record of original contributions: publications, open-source impact, or internal results that moved a capability frontier
- 08Comfort operating in ambiguous, fast-moving environments where the problem definition is as important as the solution
- 09We care more about demonstrated capability than credentials; a PhD is one signal among many
Условия
- 01Small, highly selective team where research and product move together; prototypes reach real deployment quickly
- 02Compute is not a constraint: large allocations with training jobs routinely running across thousands of GPUs from day one
- 03Environment rewards speed, autonomy, and technical depth with minimal process overhead
- 04One of the most competitive and fast-moving problems in AI