{
  "_note": "The sample-set catalogue. Only sets with real data from the showcase export are listed. 'kind' picks the viewer: coding (Terminal-Bench style tasks) or voice (audio clips).",
  "sets": [
    {
      "id": "tb4-software",
      "name": "TB4-Software",
      "type": "Code",
      "kind": "coding",
      "source": "data/index.json",
      "project": "tb4",
      "unit": "tasks",
      "description": "Terminal-Bench 4.0 format tasks in software engineering. Each is graded by hidden tests, with Opus 5.5 attempts and their trajectories.",
      "start": [
        "humble-falcon-073",
        "crisp-robin-633"
      ],
      "short": {
        "crisp-robin-633": "Mainframe copybook drift",
        "humble-falcon-073": "Journaled billing engine",
        "sharp-birch-411": "Damaged backup recovery"
      },
      "tag": "code"
    },
    {
      "id": "tb4-science",
      "name": "TB4-Science",
      "type": "Code",
      "kind": "coding",
      "description": "Terminal-Bench 4.0 format tasks in science domains. Each is graded by hidden tests, with Opus 5.5 attempts and their trajectories.",
      "source": "data/index.json",
      "project": "tb4",
      "tag": "science",
      "unit": "tasks",
      "start": [
        "sharp-gecko-098"
      ],
      "short": {
        "sharp-gecko-098": "Pendulum period recovery"
      }
    },
    {
      "id": "frontier-code",
      "name": "Frontier Code",
      "type": "Code",
      "kind": "coding",
      "source": "data/index.json",
      "project": "frontier_code",
      "unit": "tasks",
      "description": "Changes to real open-source codebases, each from a short request. Each is graded by hidden tests, with Opus 5.5 attempts and their trajectories.",
      "start": [
        "glad-raven-808",
        "crisp-owl-238",
        "trusty-beacon-605"
      ],
      "short": {
        "crisp-owl-238": "React transition abort",
        "glad-raven-808": "Decompiler string switches",
        "trusty-beacon-605": "SSR hydration on Back"
      }
    },
    {
      "id": "deepswe",
      "name": "DeepSWE",
      "type": "Code",
      "kind": null,
      "description": "Description coming soon."
    },
    {
      "id": "deepswe-qa",
      "name": "DeepSWE Q&A",
      "type": "Code",
      "kind": null,
      "description": "Repo-level code-understanding questions pinned to a commit. Rubric and LLM-judge grades, reference reproductions, 6 models × 3."
    },
    {
      "id": "cua-expert-novice",
      "name": "CUA (Expert / Novice)",
      "type": "Multi Modal",
      "kind": null,
      "description": "Hobby-to-work computer use tasks (OS, FS, Python, C++, JS). 13 attempts on Gemini 3.0 Flash plus Opus 4.7."
    },
    {
      "id": "itw-expert-novice",
      "name": "ITW (Expert / Novice)",
      "type": "Multi Modal",
      "kind": null,
      "description": "Expert-authored professional tasks taking hours to days (finance, customs, nursing, AI strategy). Prompt and assets, with a human reference."
    },
    {
      "id": "long-horizon",
      "name": "Long-Horizon Tasks",
      "type": "Multi Modal",
      "kind": null,
      "description": "Recorded screen sessions (travel planning, e-commerce and marketing, graphic design), with comparable video and instruction tracks."
    },
    {
      "id": "egocentric-household",
      "name": "Egocentric Household Tasks",
      "type": "Multi Modal",
      "kind": null,
      "description": "First-person video of household tasks (washing dishes, chopping vegetables, folding socks, cooking) with a metadata schema."
    },
    {
      "id": "voice-solo",
      "name": "Voice - Solo",
      "type": "Audio",
      "kind": null,
      "description": "Unscripted solo sessions, one speaker each. English, accented English, French, Hindi, Japanese, Spanish. About 178 min."
    },
    {
      "id": "voice-dual",
      "name": "Voice - Dual",
      "type": "Audio",
      "kind": null,
      "description": "Two-party conversations: Arabic-English and Mandarin-English code-switching, natural-environment English, and interviews."
    },
    {
      "id": "voice-conversational-multiparty",
      "name": "Voice - Conversational Multiparty",
      "type": "Audio",
      "kind": null,
      "description": "Unscripted 3–4 speaker group conversations in 8 languages, inside and full-duplex tracks. About 681 min, 12,266 segments."
    },
    {
      "id": "voice-conversational-multilingual",
      "name": "Voice - Conversational Multilingual",
      "type": "Audio",
      "kind": null,
      "description": "Two-person conversations, one each in English, Japanese and Spanish. About 53 min, 396 segments."
    },
    {
      "id": "voice-parallel-multilingual",
      "name": "Voice - Parallel Multilingual",
      "type": "Audio",
      "kind": null,
      "description": "Same pair recorded twice, once in English and once in their shared native language (Japanese, Mandarin, Spanish). About 602 min, with an ASR transcript."
    },
    {
      "id": "prompts-multiple-industry",
      "name": "Prompts - Multiple Industry",
      "type": "Docs",
      "kind": null,
      "description": "Long multi-turn prompts from professionals across industries, with structured AI quality, time saved, difficulty and demographics."
    },
    {
      "id": "agent-trajectories-legal",
      "name": "Agent Trajectories - Legal",
      "type": "Docs",
      "kind": null,
      "description": "Legal practitioners drafting commercial documents one to one with an AI assistant. Full conversation with tool calls, and the final document."
    },
    {
      "id": "agent-trajectories-education",
      "name": "Agent Trajectories - Education",
      "type": "Docs",
      "kind": null,
      "description": "A graduate student working through a thesis or methodology with an AI assistant over 121 turns, with sources, decisions and open items."
    }
  ]
}