{
  "reference_space": "trl-lib/trl-recipes",
  "reference_revision": "33591b950d0b10272ffbeaf5767021bc41aed3fa",
  "trainer_repository": "younesbelkada/trl",
  "trainer_revision": "e498acc637084cffff14deb09ccdc423a766adb2",
  "datasets": {
    "scale800": {
      "id": "harness-hub/scale800-harbor",
      "revision": "56c7f75d01f70859ec1354e97866e32a175c1a40",
      "available": 750,
      "selected": 750
    },
    "termgrade": {
      "id": "harness-hub/termgrade-train-harbor",
      "revision": "151205e3714d8d1cec6b7d5770d344853e11d7aa",
      "available": 904,
      "selected": 300
    }
  },
  "order_seed": 0,
  "ordered_task_count": 1050,
  "order_digest_format": "UTF-8 task directory names, one per line, including the final LF",
  "ordered_task_ids_sha256": "b7fc44ea949a965967e73f868373d53d5746999c2a24c00eff71d5897991a460",
  "verification": "Reconstructed with the published launcher on empty task directories. An original job manifest for independent comparison is unavailable.",
  "ordered_task_ids_file": "jobs/manifests/qwen-recipe-order.txt",
  "openenv_revision": "86a180ede21e044f7929b9a7783ad83aa67d83a3",
  "reference_transformers_revision": "66f51e76efd8542232b8bdd40ade7d1edbaac46f",
  "reference_vllm_version": "0.25.1",
  "released_specialist_adapters": {
    "source": "adapter_config.json and adapter_model.safetensors headers, read anonymously at each revision",
    "models": {
      "harness-hub/Qwen3.8-27B-pi-adapted": "7d42b119e2d4d475fa11270634b1ef98d1cd5aee",
      "harness-hub/Qwen3.8-27B-opencode-adapted": "9c6545360eb1fea3a1fb2d142e5c166a88ffc275",
      "harness-hub/Qwen3.8-27B-hermes-adapted": "0467acba2ff7788ec8050e7349fb5aa60bb6ddba"
    },
    "rank": 32,
    "alpha": 64,
    "peft_version_in_all_three_configs": "0.21.2",
    "tensor_count_per_specialist": 512,
    "factor_pairs_per_projection": {
      "q_proj": 16, "k_proj": 16, "v_proj": 16, "o_proj": 16,
      "gate_proj": 64, "up_proj": 64, "down_proj": 64
    },
    "module_root": "base_model.model.model.language_model",
    "sorted_tensor_names_utf8_lf_sha256": "ff440f4d4e4b60abf94718a358620ee3732970b58b81bc99c164a7efeee91b73"
  },
  "automationbench": {
    "reference_evaluator": "https://huggingface.co/spaces/trl-lib/trl-recipes/blob/33591b950d0b10272ffbeaf5767021bc41aed3fa/recipes/harness-adapters/Qwen3.8-27B/eval_automationbench.py",
    "benchmark_revision": "4a8e1061254004d9dac807054eed33fad7d1ff14",
    "public_recipe_example_tasks_per_domain": 10,
    "public_livestream_graded_frozen_qwen": 300,
    "published_evaluator_selection": "First N tasks in each of sales, marketing, operations, support, finance and hr, in the benchmark's native order",
    "published_evaluator_toolset": "api",
    "first_50_per_domain_file": "jobs/manifests/automationbench-evaluator-first50.txt",
    "first_50_per_domain_sha256": "b870f3174212e854af5b7cdeda19a77f68e80fe863513b81511749ccb56e2cb8",
    "first_50_overlap_with_archived_skyrl_selection": 143,
    "scope": "300 tasks implies 50 per domain if the final job used the published evaluator to cover six equal-sized domains. The original per-task export and the final invocation have not been independently verified."
  },
  "reference_time_to_selected_step": {
    "source": "https://harness-hub-qwen-livestream.hf.space/api/state",
    "basis": "Elapsed wall time from each reported Qwen run start, including any interruptions.",
    "harnesses": {
      "pi": {
        "run_name": "Qwen3.8-27B-pi-374053-ae307afd",
        "selected_step": 160,
        "elapsed_wall_hours": 57.8559
      },
      "opencode": {
        "run_name": "Qwen3.8-27B-opencode-385152-0c62425c",
        "selected_step": 150,
        "elapsed_wall_hours": 53.18
      },
      "hermes": {
        "run_name": "Qwen3.8-27B-hermes-374054-f55c9e83",
        "selected_step": 160,
        "elapsed_wall_hours": 43.002
      }
    }
  }
}
