Computational Cognitive Science Lab: LLM Distillation Experiments
William Thompson, Professor
Psychology
Open. Apprentices needed for the fall semester. Enter your application online beginning August 21st. The deadline to apply is Monday, August 31st, 4pm.
The computational cognitive science lab (ccs-ucb.github.io) studies how knowledge is transmitted from one agent to another, among humans and machines. This project examines the transmission of knowledge from one model to another using open source LLMs. You will learn about methods for on-policy distillation (e.g. https://thinkingmachines.ai/blog/on-policy-distillation/), design and study improvements to these methods. The larger goal is to explore different approaches to source-model selection in networks of language model agents gaining knowledge via distillation. There may also be opportunities for experimentation with other forms of post-training on human data for the right candidate.
Qualifications: Strong proficiency with LLM post-training frameworks. Interests in open source model ecosystem. Prior experience with distillation methods highly desirable.
Hours: 9-11 hrs
Social Sciences Engineering, Design & Technologies