SOTA systems#

In the previous sessions, we covered foundational concepts in RL which are relevant when encountering RL in the context of LM fine-tuning. Equipped with these conceptual and theoretical tools, the learning goals in the sessions from now on are:

  • to familiarize ourselves with state-of-the-art (SOTA) systems which apply these tools in practice

  • to critically think about motivation, methods and results produced with these systems

  • to understand the variety of ways in which RL can be applied in combination with LMs

  • to critically think about the systems and how to evaluate them

  • to get inspiration and practical insights for your own projects!

We will start working towards these goals by looking at at the nuts and bolts of SOTA LLMs which were fine-tuned with RL in various interesting ways (and aren’t ChatGPT). In particular, in this session we will familiarize ourselved with the models InstructGPT, Sparrrow and a Constitutional-AI LLM. Further models which also had interesting variations during training are listed below.

The slides for the session can be found here.

Student presentations#

Sparrow and the Constitutional AI LM are covered by student presentations. If possible, slides from student presentation will be available here.

Further references#