{
  "sources": [
    {
      "domain": "hotpotqa",
      "url": "https://huggingface.co/datasets/LulaCola/AgentProcessBench/resolve/cd81f326aece8b0a3f6359e20df370690d3a98bf/hotpotqa/test.jsonl",
      "sha256": "160eef2ded872d8dc6ddf4cee5752295fda691b6a9e99b5bafc2d484e6309c57",
      "rows": 250,
      "eligible": 248
    },
    {
      "domain": "gaia_dev",
      "url": "https://huggingface.co/datasets/LulaCola/AgentProcessBench/resolve/cd81f326aece8b0a3f6359e20df370690d3a98bf/gaia_dev/test.jsonl",
      "sha256": "f7b75c668fc1e6ad943e8f6d93a21a9a8f5f076841e1230e1ed2fc6d05ce8192",
      "rows": 250,
      "eligible": 225
    },
    {
      "domain": "bfcl",
      "url": "https://huggingface.co/datasets/LulaCola/AgentProcessBench/resolve/cd81f326aece8b0a3f6359e20df370690d3a98bf/bfcl/test.jsonl",
      "sha256": "6aee0b71eff7feb872c6b54d962f8831b56f7bebf770cba9cb657f219afb6fe5",
      "rows": 250,
      "eligible": 235
    },
    {
      "domain": "tau2",
      "url": "https://huggingface.co/datasets/LulaCola/AgentProcessBench/resolve/cd81f326aece8b0a3f6359e20df370690d3a98bf/tau2/test.jsonl",
      "sha256": "6f22818ff88822512767fe735f56e7b78bed4b9aaf26959144feb92f127e1e95",
      "rows": 250,
      "eligible": 225
    }
  ],
  "sample": [
    {
      "domain": "hotpotqa",
      "total_index": 108,
      "query_index": 21,
      "final_label": -1,
      "steps": 1,
      "step_label_counts": {
        "-1": 1
      }
    },
    {
      "domain": "hotpotqa",
      "total_index": 27,
      "query_index": 5,
      "final_label": 1,
      "steps": 2,
      "step_label_counts": {
        "1": 2
      }
    },
    {
      "domain": "hotpotqa",
      "total_index": 149,
      "query_index": 29,
      "final_label": -1,
      "steps": 6,
      "step_label_counts": {
        "1": 1,
        "-1": 3,
        "0": 2
      }
    },
    {
      "domain": "hotpotqa",
      "total_index": 102,
      "query_index": 20,
      "final_label": 1,
      "steps": 3,
      "step_label_counts": {
        "1": 3
      }
    },
    {
      "domain": "hotpotqa",
      "total_index": 52,
      "query_index": 10,
      "final_label": -1,
      "steps": 5,
      "step_label_counts": {
        "1": 1,
        "0": 1,
        "-1": 3
      }
    },
    {
      "domain": "hotpotqa",
      "total_index": 75,
      "query_index": 15,
      "final_label": 1,
      "steps": 4,
      "step_label_counts": {
        "1": 3,
        "0": 1
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 384,
      "query_index": 26,
      "final_label": -1,
      "steps": 6,
      "step_label_counts": {
        "1": 4,
        "0": 1,
        "-1": 1
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 373,
      "query_index": 24,
      "final_label": 1,
      "steps": 3,
      "step_label_counts": {
        "1": 3
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 452,
      "query_index": 40,
      "final_label": -1,
      "steps": 6,
      "step_label_counts": {
        "1": 5,
        "-1": 1
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 322,
      "query_index": 14,
      "final_label": 1,
      "steps": 3,
      "step_label_counts": {
        "0": 1,
        "-1": 1,
        "1": 1
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 377,
      "query_index": 25,
      "final_label": -1,
      "steps": 4,
      "step_label_counts": {
        "1": 3,
        "-1": 1
      }
    },
    {
      "domain": "gaia_dev",
      "total_index": 283,
      "query_index": 6,
      "final_label": 1,
      "steps": 3,
      "step_label_counts": {
        "1": 3
      }
    },
    {
      "domain": "bfcl",
      "total_index": 681,
      "query_index": 36,
      "final_label": -1,
      "steps": 5,
      "step_label_counts": {
        "1": 3,
        "-1": 2
      }
    },
    {
      "domain": "bfcl",
      "total_index": 671,
      "query_index": 34,
      "final_label": 1,
      "steps": 4,
      "step_label_counts": {
        "1": 4
      }
    },
    {
      "domain": "bfcl",
      "total_index": 729,
      "query_index": 45,
      "final_label": -1,
      "steps": 7,
      "step_label_counts": {
        "1": 6,
        "0": 1
      }
    },
    {
      "domain": "bfcl",
      "total_index": 735,
      "query_index": 47,
      "final_label": 1,
      "steps": 11,
      "step_label_counts": {
        "1": 8,
        "-1": 2,
        "0": 1
      }
    },
    {
      "domain": "bfcl",
      "total_index": 740,
      "query_index": 48,
      "final_label": -1,
      "steps": 14,
      "step_label_counts": {
        "1": 6,
        "-1": 8
      }
    },
    {
      "domain": "bfcl",
      "total_index": 539,
      "query_index": 7,
      "final_label": 1,
      "steps": 9,
      "step_label_counts": {
        "1": 9
      }
    },
    {
      "domain": "tau2",
      "total_index": 766,
      "query_index": 3,
      "final_label": -1,
      "steps": 4,
      "step_label_counts": {
        "1": 2,
        "-1": 2
      }
    },
    {
      "domain": "tau2",
      "total_index": 987,
      "query_index": 47,
      "final_label": 1,
      "steps": 15,
      "step_label_counts": {
        "1": 15
      }
    },
    {
      "domain": "tau2",
      "total_index": 796,
      "query_index": 9,
      "final_label": -1,
      "steps": 8,
      "step_label_counts": {
        "1": 5,
        "-1": 2,
        "0": 1
      }
    },
    {
      "domain": "tau2",
      "total_index": 854,
      "query_index": 20,
      "final_label": 1,
      "steps": 15,
      "step_label_counts": {
        "1": 15
      }
    },
    {
      "domain": "tau2",
      "total_index": 940,
      "query_index": 38,
      "final_label": -1,
      "steps": 13,
      "step_label_counts": {
        "1": 3,
        "-1": 10
      }
    },
    {
      "domain": "tau2",
      "total_index": 801,
      "query_index": 10,
      "final_label": 1,
      "steps": 12,
      "step_label_counts": {
        "1": 12
      }
    }
  ]
}