{
  "aliases": [],
  "authors": [],
  "details": {
    "benchmark_protocol": {
      "actions": [
        "success verdict",
        "process score"
      ],
      "counts": [
        {
          "basis": "reported",
          "evidence_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md",
          "population": "fara7b_om2w_browserbase split",
          "qualifier": "exact",
          "unit": "trajectories",
          "value": 106
        },
        {
          "basis": "reported",
          "evidence_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md",
          "population": "internal split",
          "qualifier": "exact",
          "unit": "trajectories",
          "value": 154
        },
        {
          "basis": "reported",
          "evidence_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md",
          "population": "fara7b_om2w_browserbase split",
          "qualifier": "exact",
          "unit": "annotation rows",
          "value": 215
        },
        {
          "basis": "reported",
          "evidence_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md",
          "population": "internal split",
          "qualifier": "exact",
          "unit": "annotation rows",
          "value": 154
        }
      ],
      "evaluation": "Agreement metrics including Cohen’s kappa, accuracy and F1 against human annotations; review stages must be kept distinct.",
      "evidence_urls": [
        "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md"
      ],
      "focus": [
        "verifier-human agreement",
        "process quality"
      ],
      "notes": "Trajectory rows and reviewer rows have different grains. Internal task labels do not make the released Parquet files private. No payload data was downloaded.",
      "observations": [
        "screenshots",
        "action logs",
        "final answers"
      ],
      "setting": "offline"
    },
    "disabled": false,
    "download_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/tree/main",
    "gated": false,
    "hub_id": "microsoft/CUAVerifierBench",
    "license_review": {
      "data": {
        "evidence_urls": [
          "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md"
        ],
        "reason": "The benchmark's Hub dataset card declares MIT in its metadata header (license: mit)."
      }
    }
  },
  "environments": [
    "web"
  ],
  "evidence": [
    {
      "method": "editorial_primary_review",
      "retrieved_at": "2026-10-17T02:42:32Z",
      "role": "primary",
      "sha256": "44d7190b72ae489297f59d6fadff635ced6dd067c845c77b04f187a7b2d7bec7",
      "supports": [
        "name",
        "kind",
        "summary",
        "environments",
        "topics",
        "scope",
        "relations",
        "details",
        "licenses.data"
      ],
      "url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md"
    },
    {
      "method": "editorial_primary_review",
      "retrieved_at": "2026-10-17T02:42:32Z",
      "role": "primary",
      "sha256": "fb1fe8a9fd6e45fc82157c65c08d5c682d08bead5196308776200ef3c39c1e41",
      "supports": [
        "identity",
        "release_status",
        "details",
        "licenses.data"
      ],
      "url": "https://huggingface.co/api/datasets/microsoft/CUAVerifierBench"
    }
  ],
  "id": "url:d5140d438466c1e1348a",
  "identifiers": {},
  "kind": "benchmark",
  "last_verified_at": "2026-10-17T02:42:32Z",
  "licenses": {
    "code": null,
    "data": {
      "basis": "publisher model/dataset card metadata; third-party assets retain their own terms",
      "declared": "mit",
      "retrieved_at": "2026-10-17T02:42:32Z",
      "url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md"
    },
    "document": null,
    "weights": null
  },
  "links": [
    {
      "rel": "primary",
      "url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench"
    }
  ],
  "name": "CUAVerifierBench",
  "notes": "Public Parquet configs separate trajectories from per-reviewer annotations. The card declares MIT. The two splits have different review stages: Online-Mind2Web includes blind and verifier-informed annotations, whereas the internal split is blind-only. This measures verifier agreement rather than agent task success.",
  "published_at": null,
  "relations": [
    {
      "evidence_url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench/raw/main/README.md",
      "target_id": "arxiv:2604.06240",
      "type": "documented_by"
    }
  ],
  "review_status": "curated",
  "scope": "core",
  "summary": "Human-annotated browser-agent trajectories for comparing automated task-success and process-quality verifiers against reviewer judgments.",
  "topics": [
    "evaluation",
    "reward-models",
    "web-navigation"
  ],
  "updated_at": null,
  "url": "https://huggingface.co/datasets/microsoft/CUAVerifierBench"
}