{
  "dataset_version": "6.0.0-embedding-lift-controlled",
  "random_seed": 42,
  "n_entities": 12000,
  "split_weights": {
    "train": 0.7,
    "validation": 0.15,
    "test": 0.15
  },
  "semantic_entity_share": 0.225,
  "target_lexical_positive_recall": 0.85,
  "evaluation_strategy": "entity-level 70/15/15; all-positive blocking truth; balanced model benchmark; challenge stress test; production-like threshold benchmark; sealed test",
  "um_training_strategy": "three sources/entity; stratified lexical variants; controlled tax/name/address/representative collisions; random, lexical-hard and semantic-hard nonmatches",
  "semantic_design": "inspection aliases paraphrase or translate brand, modifier, region and industry; semantic hard negatives share only industry and region"
}