Active Label Correction For Building Llm-based Modular AI Systems
Taneja Karan, Goel Ashok. Arxiv 2024
[Paper]
Large Language Models (LLMs) have been used to build modular AI systems such as HuggingGPT, Microsoft Bing Chat, and more. To improve such systems after deployment using the data collected from human interactions, each module can be replaced by a fine-tuned model but the annotations received from LLMs are low quality. We propose that active label correction can be used to improve the data quality by only examining a fraction of the dataset. In this paper, we analyze the noise in datasets annotated by ChatGPT and study denoising it with human feedback. Our results show that active label correction can lead to oracle performance with feedback on fewer examples than the number of noisy examples in the dataset across three different NLP tasks.
Similar Work