Thinking about High-Quality Human Data

Lilian Weng · 2 years ago

High-quality human-annotated data is essential for training deep learning models, including classification tasks and RLHF labeling for LLM alignment. The community recognizes data quality's importance but faces a cultural preference for model development over the unglamorous work of careful data collection and annotation. This imbalance risks compromising model performance when the foundational data work receives insufficient attention and resources.