René's URL Explorer Experiment

  {
    "@context": "https://schema.org",
    "@type": "Dataset",
    "name": "ClawBench Leaderboard",
    "alternateName": "ClawBench Browser Agent Leaderboard",
    "description": "Per-model pass rates and HTTP-interception rates on the ClawBench benchmark (130 tasks). Two-stage scoring: deterministic URL/method interception + LLM judge.",
    "url": "https://claw-bench.com/",
    "sameAs": [
      "https://github.com/TIGER-AI-Lab/ClawBench",
      "https://huggingface.co/datasets/TIGER-Lab/ClawBenchV2Trace",
      "https://huggingface.co/spaces/TIGER-Lab/ClawBench",
      "https://arxiv.org/abs/2604.08523"
    ],
    "license": "https://www.apache.org/licenses/LICENSE-2.0",
    "creator": {"@type": "Organization", "name": "NAIL Group", "url": "https://github.com/TIGER-AI-Lab/ClawBench"},
    "measurementTechnique": "Two-stage agentic evaluation: final-HTTP-request interception with URL/method schema match, plus LLM judge (default deepseek-v4-pro) for payload-correctness check",
    "variableMeasured": [
      {"@type": "PropertyValue", "name": "Interception rate", "description": "Fraction of runs whose final HTTP request matches the per-task URL/method schema"},
      {"@type": "PropertyValue", "name": "Reward rate", "description": "Fraction of runs that intercepted AND were judged to fulfill the natural-language instruction"}
    ]
  }