{
  "field_id": "F049",
  "field": "data-centric machine learning",
  "directions": [
    {
      "direction_id": "G0",
      "name": "Data-centric lifecycle and workflow management",
      "scope": "Organize repeatable processes for machine learning data from development and evaluation through updating and maintenance, including cross-stage guidance and executable curation and evaluation workflows."
    },
    {
      "direction_id": "G1",
      "name": "Dataset quality diagnosis and annotation development",
      "scope": "Assess whether data content, coverage, and labels are reliable, and develop or correct annotations that express intended concepts. This direction encompasses quality and composition diagnosis as well as annotation generation, definition, reconciliation, and correction; an individual study need not both diagnose and correct."
    },
    {
      "direction_id": "G2",
      "name": "Valuation and prioritization of available training data",
      "scope": "Study the usefulness of existing data for a learning objective and, on that basis, estimate contributions or decide on selection, pruning, filtering, reweighting, and source or domain proportions. Contribution scoring and decisions about which data to use are distinct outcomes within this direction; neither requires the other."
    },
    {
      "direction_id": "G3",
      "name": "Construction of augmented or condensed training data",
      "scope": "Construct transformed or generated training evidence to improve learning or reduce the amount of data needed. This direction encompasses both example augmentation that supplements the original training set and compact synthetic datasets intended to replace large datasets."
    }
  ]
}