{
  "dataset_name": "WebWorld-HTMLVCR-SFT-10k",
  "created_at": "2026-09-01",
  "rows": 10000,
  "format": "ShareGPT conversations[human, gpt]",
  "file": "WebWorld-HTMLVCR-SFT-10k.jsonl",
  "bytes": 400581171,
  "sha256": "74a64b66e4d8ffe66548d8e5a1721b8b33df76a7afa07f9d05c4dda2796c0884",
  "source_dataset": "HTMLVCR.jsonl (32,800 rows)",
  "sampling": {
    "seed": 20260901,
    "method": "uniform random over test-set-scrubbed pool",
    "clean_pool_size": 30000,
    "excluded_before_sampling": 2800
  },
  "leakage_audit": {
    "against_htmlbench_400": {
      "rows_removed": 2000,
      "verified_zero_in_release": true
    },
    "against_miniappbench_100": {
      "rows_removed": 800,
      "verified_zero_in_release": true
    },
    "method": "full normalized prompt/query substring match"
  },
  "held_out_by_paper": [
    "HTMLBench (6 categories, 400 prompts) — evaluation only",
    "MiniAppBench-Val (100 queries) — evaluation only"
  ],
  "not_released": [
    "raw_equal32800 baseline",
    "acceptance-rule ablations (cert_20k, vlm_only_20k, score_gate_20k, no_preserve_20k, no_certificate_20k)",
    "depth ablations (depth_ge1/2/3/5/8_equal5k)",
    "internal trajectory records with certificate metadata and file-system paths"
  ],
  "license": "Apache-2.0"
}