The Micro-Action Dataset: An Ordered Decomposition of O*NET Work Activities

This dataset decomposes the 1,961 quality-passing Detailed Work Activities (DWAs) of O*NET 30.2 into 15,817 ordered micro-actions — the individual perceptual, cognitive, communicative, and physical steps by which each activity is performed — produced by a consensus multi-agent LLM pipeline and validated by an independent blind three-annotator study (construct reliability Fleiss κ = 0.90; encoder-vs-human κ = 0.63, reliable in aggregate, noisy per action). To our knowledge it is the first ordered-step decomposition covering a full occupational ontology. The release includes a semantic typology (7 macro-classes + a 35.6% generic substrate), intelligence-type labels under two encoders, occupation-level composition shares for all 923 occupations, an auxiliary AI-substitutability indicator, and the complete validation records. Intended use — and not. This is a research instrument. Per-occupation values (ω shares, substrate shares, OAI projections) are research measures, not risk scores: they carry no prediction of displacement for any occupation or individual. Per-action intelligence-type labels must not be used at row resolution for any consequential decision (employment decisions about individuals, credit or insurance underwriting, workforce-program targeting without independent validation). Step counts are not time weights. See README.md for full documentation; a Data Descriptor is under submission at Scientific Data.