{
  "results": [
    {
      "theta": 160000.0,
      "short": 0.9160820834562606,
      "mid": 1.0387379115148965,
      "long": 1.1367697905727685,
      "seconds": 26.022907972335815
    },
    {
      "theta": 640000.0,
      "short": 0.9236352910098977,
      "mid": 1.0243582937320355,
      "long": 0.9822827437326973,
      "seconds": 11.011847972869873
    },
    {
      "theta": 1280000.0,
      "short": 0.9304989230334977,
      "mid": 1.031460880067349,
      "long": 0.9601743056856983,
      "seconds": 11.087600231170654
    },
    {
      "theta": 2560000.0,
      "short": 0.939639166783192,
      "mid": 1.044918367936274,
      "long": 0.9650228895573634,
      "seconds": 11.14451551437378
    },
    {
      "theta": 5120000.0,
      "short": 0.9498286157172525,
      "mid": 1.0788166451562344,
      "long": 1.0147006089168509,
      "seconds": 11.193717956542969
    },
    {
      "theta": 10240000.0,
      "short": 0.9709671184858385,
      "mid": 1.0868880858439343,
      "long": 0.9965657126456587,
      "seconds": 11.231298685073853
    }
  ],
  "chosen_theta": 1280000.0,
  "samples": {
    "short": 256,
    "mid": 96,
    "long": 64
  },
  "rule": "min(mid + long masked-LM loss) with short-context loss within 5% of the original 160k theta; 15% masking; training text only"
}