Giant language fashions (LLMs) wrestle to persistently generate UI code that compiles and produces visually related designs. Current approaches to enhance technology depend on costly human suggestions or distilling a proprietary mannequin. On this paper, we discover the usage of automated suggestions (compilers and multi-modal fashions) to information LLMs to generate high-quality UI code. Our methodology begins with an present LLM and iteratively produces improved fashions by self-generating a big artificial dataset utilizing an unique mannequin, making use of automated instruments to aggressively filter, rating, and de-duplicate the info right into a refined greater high quality dataset. The unique LLM is improved by finetuning on this refined dataset. We utilized our method to a number of open-source LLMs and in contrast the ensuing efficiency to baseline fashions with each automated metrics and human preferences. Our analysis reveals the ensuing fashions outperform all different downloadable baselines and method the efficiency of bigger proprietary fashions.
- ** Work performed whereas at Apple
- † Carnegie Mellon College
