{
  "seed": 15027,
  "sources": [
    {
      "dataset": "openai/gsm8k",
      "split": "main/test",
      "file_sha256": "ee7b8da9e381df27b9e3f7758a159ab2bdaa4dbaa910546cbbc47e0cb44e4f59",
      "selection": "first 200, same as HyperQwen"
    },
    {
      "dataset": "allenai/IFBench_test",
      "file_sha256": "80037e4d99c39a55c1e2e7d5a863d8d9edeb5ebe136ba8c1e849f8c015027c6a",
      "note": "Upstream names this split train, but it is the benchmark test set; never used for calibration."
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test.jsonl",
      "sha256": "2bd02b38beb48e8c46b5b9987095d999ff38cd8efc255ea5d58974317c48f63f"
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test2.jsonl",
      "sha256": "095df7c5daf15f882c51a9deb84085cff1e073495a5dbcf95015a564d485f3a3"
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test3.jsonl",
      "sha256": "28ed26cc83363ce3f1fe2d5fad9f8393077beb1907b167a31bd3b32f80801b79"
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test4.jsonl",
      "sha256": "d711138ddaebfcf5f8ec6a4283ee677298c0f5c5d374a235af92aaf0584510da"
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test5.jsonl",
      "sha256": "7f77571c2a6df0c2a72a3277650309f67e01e0008e18117e624633df53f81214"
    },
    {
      "dataset": "livecodebench/code_generation_lite",
      "file": "test6.jsonl",
      "sha256": "bb4c364f71921c4495a6ad15abe1a927350b720009f4933e2e71f8af0f6fd1f5"
    }
  ],
  "tasks": {
    "gsm8k": 200,
    "ifbench": 300,
    "livecodebench": 100,
    "tools": 30
  },
  "task_file_sha256": "809c2c6b124579d41c379a6a436e9649b5e1d1ddf44a147a3d70d9b4b19d68f4",
  "ppl_file_sha256": "57c83ffe7c0dfba2b869f6a379d2df3f0bc0d055069b49569a2154336151a718",
  "protocol": "Local single-user launcher and .env, FP8 KV, 150000 context, 128000 output ceiling. Full quality concurrency two; speed and latency pilot concurrency one. Truncated answers count as wrong.",
  "lcb_scoring": "100 stratified stdin-only v6-era tasks; all supplied public/private tests; whitespace token comparison with numeric tolerance; not the full official LCB runner.",
  "max_output_tokens_per_call": 128000,
  "context_len": 150000,
  "quality_concurrency": 2,
  "supersedes": "<WORKSPACE>/runs/swift15/evaluation/tasks.jsonl",
  "truncation_policy": "count_as_wrong",
  "sampling": {
    "thinking": {
      "temperature": 1.0,
      "top_p": 0.95,
      "top_k": 20,
      "reasoning_effort": "xhigh"
    },
    "nonthinking": "greedy",
    "seed": 15027
  },
  "serving_profile": "local-single-user"
}