{
  "schema_version": 1,
  "updated": "2026-10-14",
  "benchmarks": [
    {
      "id": "browsecomp",
      "name": "BrowseComp",
      "description": "Research-style web browsing and answer synthesis.",
      "url": "https://github.com/openai/simple-evals",
      "metric": {
        "name": "answer correctness",
        "type": "llm_judge",
        "direction": "higher_is_better",
        "imported_evidence": "Source-reported GPT-4o answer judge.",
        "caveat": "The judge result was not independently reproduced."
      },
      "source_ids": ["browsecomp_rollouts"]
    },
    {
      "id": "mind2web",
      "name": "Mind2Web",
      "description": "Human demonstrations for completing tasks on real websites.",
      "url": "https://osu-nlp-group.github.io/Mind2Web/",
      "metric": {
        "name": "step and element accuracy",
        "type": "action_match",
        "direction": "higher_is_better",
        "imported_evidence": "Human action and pre-action DOM context.",
        "caveat": "The imported demonstrations have no task-level execution result."
      },
      "source_ids": ["mind2web"]
    },
    {
      "id": "swe-bench",
      "name": "SWE-bench",
      "description": "Resolve real software repository issues.",
      "url": "https://www.swebench.com/",
      "metric": {
        "name": "resolved",
        "type": "native_binary",
        "direction": "higher_is_better",
        "imported_evidence": "Source-published repository test result.",
        "caveat": "Results were not replayed by this project."
      },
      "source_ids": [
        "nvidia_open_swe_traces",
        "swebench_multilingual_glm5",
        "swebench_multilingual_minimax_m25",
        "swebench_multilingual_trajectories",
        "swebench_verified_trajectories"
      ]
    },
    {
      "id": "swe-gym",
      "name": "SWE-Gym",
      "description": "Train and evaluate software-engineering agents on repository tasks.",
      "url": "https://github.com/SWE-Gym/SWE-Gym",
      "metric": {
        "name": "execution-test success",
        "type": "native_binary",
        "direction": "higher_is_better",
        "imported_evidence": "Published successful split or source execution-test result.",
        "caveat": "The OpenHands SFT release is success-conditioned."
      },
      "source_ids": ["swegym_openhands", "swegym_sampled"]
    },
    {
      "id": "swe-smith",
      "name": "SWE-smith",
      "description": "Generated software-engineering tasks with executable validation.",
      "url": "https://github.com/SWE-bench/SWE-smith",
      "metric": {
        "name": "resolved",
        "type": "native_binary",
        "direction": "higher_is_better",
        "imported_evidence": "Source-published resolved field.",
        "caveat": "Results were not replayed by this project."
      },
      "source_ids": ["swesmith_trajectories"]
    },
    {
      "id": "tau2-bench",
      "name": "τ2-bench",
      "description": "Tool-using agents interacting with simulated users in service domains.",
      "url": "https://github.com/sierra-research/tau2-bench",
      "metric": {
        "name": "evaluator correctness and reward",
        "type": "native_reward",
        "direction": "higher_is_better",
        "imported_evidence": "AgentSuite evaluator score, database match and correctness fields.",
        "caveat": "AgentSuite trajectories use τ2 tasks and are not version-matched τ3 results."
      },
      "source_ids": ["agentsuite_tau2", "areal_tau2_sft"]
    },
    {
      "id": "tau3-bench",
      "name": "τ3-bench",
      "description": "Tool-using agents and simulated users with environment reward evidence.",
      "url": "https://github.com/sierra-research/tau2-bench",
      "metric": {
        "name": "environment/action/communication reward",
        "type": "native_reward",
        "direction": "higher_is_better",
        "imported_evidence": "Reward evidence from an official checked-in result file.",
        "caveat": "Only the imported result file is represented."
      },
      "source_ids": ["official_tau3"]
    },
    {
      "id": "terminal-bench",
      "name": "Terminal-Bench",
      "description": "Complete practical tasks in sandboxed terminal environments.",
      "url": "https://github.com/laude-institute/terminal-bench",
      "metric": {
        "name": "task reward",
        "type": "native_reward",
        "direction": "higher_is_better",
        "imported_evidence": "Source-reported Terminal-Bench reward.",
        "caveat": "Scores were not replayed by this project."
      },
      "source_ids": ["terminalbench_trajectories", "litecoder_terminal_sft"]
    },
    {
      "id": "terminaltraj",
      "name": "TerminalTraj",
      "description": "A corpus of terminal command and observation trajectories.",
      "url": "https://huggingface.co/datasets/m-a-p/TerminalTraj",
      "metric": {
        "name": "not available",
        "type": "none",
        "direction": "not_applicable",
        "imported_evidence": "Paired command and terminal-output transcript.",
        "caveat": "No trustworthy task-level success field is published."
      },
      "source_ids": ["terminaltraj"]
    },
    {
      "id": "toolbench",
      "name": "ToolBench",
      "description": "Tool-use planning and API-call trajectories.",
      "url": "https://github.com/OpenBMB/ToolBench",
      "metric": {
        "name": "valid-data/win generation status",
        "type": "source_proxy",
        "direction": "higher_is_better",
        "imported_evidence": "Published generation status.",
        "caveat": "This is not independent semantic proof of task success."
      },
      "source_ids": ["toolbench", "toolbench_full"]
    },
    {
      "id": "bfcl-v4",
      "name": "BFCL V4",
      "description": "Agentic function-calling tasks with web and memory tool environments.",
      "url": "https://gorilla.cs.berkeley.edu/leaderboard.html",
      "metric": {
        "name": "BFCL evaluator score",
        "type": "native_binary",
        "direction": "higher_is_better",
        "imported_evidence": "AgentSuite-published V4 evaluator score.",
        "caveat": "Only web and memory rows with returned tool observations qualify; scores were not rerun."
      },
      "source_ids": ["agentsuite_bfcl_v4"]
    },
    {
      "id": "deepswe",
      "name": "DeepSWE",
      "description": "Repository engineering tasks executed by multi-agent DeLM teams.",
      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
      "metric": {
        "name": "CTRF test-suite pass",
        "type": "native_binary",
        "direction": "higher_is_better",
        "imported_evidence": "Published per-agent CTRF grading summaries.",
        "caveat": "Protected evaluation tasks contain anti-training canaries and were not replayed here."
      },
      "source_ids": ["delm_deepswe"]
    },
    {
      "id": "terminal-bench-4",
      "name": "Terminal-Bench 4",
      "description": "Protected terminal tasks executed by multi-agent DeLM teams.",
      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
      "metric": {
        "name": "task reward",
        "type": "native_reward",
        "direction": "higher_is_better",
        "imported_evidence": "Published per-agent grader reward.",
        "caveat": "Protected evaluation tasks contain Harbor anti-training canaries and were not replayed here."
      },
      "source_ids": ["delm_terminalbench4"]
    }
  ]
}