Thinking about High-Quality Human Data

Disagreement between people labeling AI training data can contain useful information. Lilian Weng's 2024 survey distinguishes mistakes caused by unclear instructions from legitimate differences in judgment, especially on subjective questions.

That distinction changes how a dataset should be built. A prescriptive approach trains raters to apply a specified rule consistently. A descriptive approach preserves differing interpretations so researchers can study them. Collapsing every disagreement into a majority answer can erase the very variation the second approach seeks to capture. Data quality therefore depends partly on deciding what the labels are supposed to represent before treating agreement as the measure of success.

“High-quality data is the fuel for modern data deep learning model training.”

“Everyone wants to do the model work, not the data work”

Sambasivan et al. 2021