STARLING-7B: INCREASING LLM HELPFULNESS & HARMLESSNESS WITH RLAIF

Summary from the TLDR AI newsletter

A group of researchers at Berkeley trained a new state-of-the-art 7B parameter model using synthetic preference data. This blog talks about the novel challenges of training reward models (e.g., the position of the example in the list changes its ranking) and how they overcame them. The resulting model is available along with the trained reward model.