About the job
You would be working as part of our Applied Research team, focused on turning pre-trained LLMs into well-aligned and highly capable AI systems for coding and software development. This is a hands-on role where you'll work across a variety of efforts, including: Building data pipelines and environments for agentic use cases, researching and implementing post-training algorithms, designing experiments and testing hypothesis, and more. You will have access to thousands of GPUs in this team.
Responsibilities
- Research and experiment on ways to specialize foundational models to agentic use cases
- Build and maintain data and training pipelines
- Keep up with latest research, and be familiar with state of the art in LLMs, alignment, synthetic data generation, code generation
- Design, analyze, and iterate on training/fine-tuning/data generation experiments
- Write high-quality, pragmatic code
- Work as part of a team: plan future steps, discuss, and communicate clearly with your peers
Qualifications
Minimum
- Experience with Large Language Models (LLM)
- Deep knowledge of Transformers
- Strong deep learning fundamentals
- Good taste in data
- Post-training experience with LLMs
- Extensively used and probed LLMs, familiarity of their capabilities and limitations
- Knowledge of distributed training
- Strong machine learning and engineering background
- Research experience
- Experience in proposing and evaluating novel research ideas
- Familiar with, or contributed to the state of the art in multiple of the following topics: Fine-tuning and alignment of LLMs, synthetic data generation, continual learning, RLVR, code generation
- Is comfortable in a rapidly iterating environment
- Is reasonably opinionated
- Programming experience
- Linux
- Strong algorithmic skills
- Python with PyTorch or Jax
- Use modern tools, including latest code agents and are always looking to improve
- Strong critical thinking and ability to question code quality policies when applicable
Preferred
- Prior experience in non-ML programming, especially not in Python - is a nice to have