schema/subsets.json

{
  "schema_name": "subsets_parquet",
  "schema_version": 1,
  "artifact": "subsets/*.parquet",
  "format": "parquet",
  "row_grain": "one logical POV sample that belongs to a named public subset",
  "description": "Lightweight researcher-facing subset views over manifest.parquet. Subsets do not duplicate physical WebDataset shards; loaders join sample_key values to manifest.parquet and sample_index.parquet.",
  "primary_key": [
    "subset",
    "sample_key"
  ],
  "columns": [
    {"name": "subset", "arrow_type": "string", "nullable": false},
    {"name": "sample_key", "arrow_type": "string", "nullable": false},
    {"name": "match_id", "arrow_type": "string", "nullable": false},
    {"name": "round_id", "arrow_type": "string", "nullable": false},
    {"name": "round_idx", "arrow_type": "int32", "nullable": false},
    {"name": "pov_idx", "arrow_type": "int32", "nullable": false},
    {"name": "split", "arrow_type": "string", "nullable": false, "valid_values": ["train", "val", "test"]},
    {"name": "map_slug", "arrow_type": "string", "nullable": false, "valid_values": ["anubis", "ancient", "dust2", "inferno", "mirage", "nuke", "overpass"]},
    {"name": "duration_s", "arrow_type": "float32", "nullable": false},
    {"name": "frames", "arrow_type": "int32", "nullable": false}
  ],
  "validation": [
    "every sample_key exists in manifest.parquet",
    "subset membership is match-map atomic for official hour-budget subsets",
    "train_* and dust2_100h subsets contain only split == train",
    "dust2_100h contains only map_slug == dust2",
    "full_demo_eval contains only rows from the strict full-demo lock",
    "every retained round has exactly pov_idx 0..9 once each"
  ],
  "official_subsets": [
    "train_10h",
    "train_50h",
    "train_100h",
    "train_500h",
    "train_1000h",
    "train_all",
    "dust2_100h",
    "full_demo_eval"
  ]
}