Research, Finetuning Science
Core
Advancing the science of fine-tuning and frontier post-training techniques to build the best post-training engine (Tinker) in the industry.
Role type
Senior Research Scientist (Post-Training & Fine-Tuning)
Builds
Tinker (post-training engine), training defaults, API design, and open-source Tinker Cookbook
Domain
AI/ML, Large Language Models, Reinforcement Learning
Deliverable
production ML models
Required skills
Python, deep learning frameworks (PyTorch, TensorFlow, JAX), distributed training debugging, LoRA, parameter-efficient fine-tuning, RL science
Preferred skills
probability, statistics, ML fundamentals, RLHF, RLAIF, preference modeling, reward learning, human data collection campaigns, alignment, data-centric AI, RL training stability techniques
Technologies
PyTorch, TensorFlow, JAX
Responsibilities
Contribute to LoRA and parameter-efficient fine-tuning to push customization quality, efficiency, and reliability; Ship research into product by informing Tinker's training defaults and codifying best-practice methods; Improve stability, efficiency, and reliability of large-scale fine-tuning and RL runs; Share findings through papers, technical blog posts, and community contributions
Seniority
Senior, hands-on IC