China's National Data Administration held a meeting on 28 October with 21 "chain leader" units drawn from nine fields — education and teaching, health care, scientific research, industrial manufacturing, bio-manufacturing, agriculture and rural affairs, energy, transport and meteorology — to review how industry datasets for AI are being built. The account was published on 2 November. The agency said dataset work should follow three principles it phrased as building to demand, building through use and improving through evaluation, and listed six tasks: map what AI work actually needs, including demand from firms upstream and downstream; set up benefit-sharing arrangements and industry data maps, and push distributed aggregation of sector data; feed real usage back so datasets are updated; coordinate along the supply chain; connect to national data infrastructure; and turn dataset quality evaluation into a tool-based, order-based service with shorter cycles and lower cost. It asked the chain leaders to register their datasets in the national dataset management service system. Staff from the National Development and Reform Commission, the National Data Development Research Institute, the China Electronics Standardization Institute and CAICT attended.
Our takeThe measurable task is the last one: turning dataset evaluation into an ordered service. If a corpus is scored by a tool on a fixed cycle, quality becomes a number a buyer can check rather than a claim in a slide. What that number measures is still open — label agreement, coverage, freshness — and a medical set and a factory log fail in different ways. A single national template risks measuring the paperwork instead.