Particle.news
Download on the App Store
3 ARTICLES
·
yesterday

Apple’s UICoder Harnesses Synthetic Data Loop to Approach GPT-4 in SwiftUI Coding

The approach combines compiler-based filtering with GPT-4V visual checks to produce nearly one million synthetic SwiftUI examples.

Apple trains StarChat-Beta to generate SwiftUI code

Overview

  • Researchers started with the open-source StarChat-Beta model to generate SwiftUI programs from UI descriptions and then applied a two-stage validation loop to filter outputs.
  • Each round of generation and validation added cleaner examples to the training set, yielding about 996,000 distinct SwiftUI programs after five iterations.
  • UICoder significantly outperformed its StarChat-Beta base on automated metrics and human evaluations and matched or exceeded GPT-4 in compilation success.
  • The team found that StarChat-Beta’s original training data lacked almost all SwiftUI content because Swift repositories were excluded from TheStack and OpenAssistant-Guanaco.
  • Authors suggest the automated synthetic-data and feedback-driven finetuning method could extend to other programming languages and UI toolkits.