China's National Data Administration held a work meeting on its data-labelling pilots in Beijing on 27 October; the readout was published on 31 October. Vice-minister Xia Bing spoke, and the data departments of the 39 provinces and cities in the pilot attended, along with the China Academy of Information and Communications Technology and the China Electronics Standardization Institute.
The meeting framed labelling as support for AI development, with three stated goals - technology, data supply, industry - and three instruments: a dedicated programme for labelling technology, a network of domain labelling experts, and a cross-region coordination mechanism.
Three tasks were set. First, supply follows demand: pilot cities must survey the firms upstream and downstream of their labelling work and build a map of the high-quality datasets those firms need, both near-term and longer-term. Second, build through use: cities and the lead firms of dataset-building chains should have their work judged by how scenarios perform. Third, annotated datasets should be registered in the national dataset management service system, where those that pass quality assessment get their downstream use tracked so they can be reused.
Our takeThe measurement sits in the right place. A demand map admits that dataset quality is a property of a task, not of a corpus: the same annotated thousand hours is good for one model and useless for another, so the buyer has to state the task before the supplier builds. Registration, a quality test and use-tracking are an attempt to make that task context travel with the data. The open question is technical - whether any evaluator can score dataset quality without re-running the downstream model.