[
  {
    "model": "HuggingFaceTB/SmolLM-135M",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 90,
    "fail": 10,
    "uncertain": 0,
    "pass_percent": 90.0,
    "wilson95_percent": [
      82.6,
      94.5
    ],
    "grammar_yes": 94,
    "complete_yes": 94,
    "parameters": 134515008,
    "configuration": "raw"
  },
  {
    "model": "HuggingFaceTB/SmolLM2-135M",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 90,
    "fail": 10,
    "uncertain": 0,
    "pass_percent": 90.0,
    "wilson95_percent": [
      82.6,
      94.5
    ],
    "grammar_yes": 95,
    "complete_yes": 92,
    "parameters": 134515008,
    "configuration": "raw"
  },
  {
    "model": "openai-community/gpt2",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 90,
    "fail": 10,
    "uncertain": 0,
    "pass_percent": 90.0,
    "wilson95_percent": [
      82.6,
      94.5
    ],
    "grammar_yes": 96,
    "complete_yes": 91,
    "parameters": 124439808,
    "configuration": "raw"
  },
  {
    "model": "state-spaces/mamba-130m-hf",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 88,
    "fail": 12,
    "uncertain": 0,
    "pass_percent": 88.0,
    "wilson95_percent": [
      80.2,
      93.0
    ],
    "grammar_yes": 97,
    "complete_yes": 90,
    "parameters": 129135360,
    "configuration": "raw"
  },
  {
    "model": "EleutherAI/gpt-neo-125m",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 85,
    "fail": 15,
    "uncertain": 0,
    "pass_percent": 85.0,
    "wilson95_percent": [
      76.7,
      90.7
    ],
    "grammar_yes": 94,
    "complete_yes": 87,
    "parameters": 125198592,
    "configuration": "raw"
  },
  {
    "model": "McGill-NLP/TLM-100M",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 84,
    "fail": 16,
    "uncertain": 0,
    "pass_percent": 84.0,
    "wilson95_percent": [
      75.6,
      89.9
    ],
    "grammar_yes": 92,
    "complete_yes": 90,
    "parameters": 98993137,
    "configuration": "raw"
  },
  {
    "model": "facebook/opt-125m",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 83,
    "fail": 17,
    "uncertain": 0,
    "pass_percent": 83.0,
    "wilson95_percent": [
      74.5,
      89.1
    ],
    "grammar_yes": 91,
    "complete_yes": 87,
    "parameters": 125239296,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/pollock-mini-lm-125m",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 78,
    "fail": 22,
    "uncertain": 0,
    "pass_percent": 78.0,
    "wilson95_percent": [
      68.9,
      85.0
    ],
    "grammar_yes": 88,
    "complete_yes": 84,
    "parameters": 127674624,
    "configuration": "raw"
  },
  {
    "model": "EleutherAI/pythia-160m",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 77,
    "fail": 23,
    "uncertain": 0,
    "pass_percent": 77.0,
    "wilson95_percent": [
      67.8,
      84.2
    ],
    "grammar_yes": 89,
    "complete_yes": 81,
    "parameters": 162322944,
    "configuration": "raw"
  },
  {
    "model": "speakleash/Bielik-11B-v3.0-Instruct",
    "track": "11B Polish instruct anchor; Q4_K_M",
    "generated": 100,
    "judged": 100,
    "pass": 77,
    "fail": 23,
    "uncertain": 0,
    "pass_percent": 77.0,
    "wilson95_percent": [
      67.8,
      84.2
    ],
    "grammar_yes": 82,
    "complete_yes": 84,
    "parameters": 11168796672,
    "configuration": "raw"
  },
  {
    "model": "distilbert/distilgpt2",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 69,
    "fail": 31,
    "uncertain": 0,
    "pass_percent": 69.0,
    "wilson95_percent": [
      59.4,
      77.2
    ],
    "grammar_yes": 85,
    "complete_yes": 73,
    "parameters": 81912576,
    "configuration": "raw"
  },
  {
    "model": "EleutherAI/pythia-70m",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 60,
    "fail": 40,
    "uncertain": 0,
    "pass_percent": 60.0,
    "wilson95_percent": [
      50.2,
      69.1
    ],
    "grammar_yes": 70,
    "complete_yes": 74,
    "parameters": 70426624,
    "configuration": "raw"
  },
  {
    "model": "MBZUAI/LaMini-GPT-124M",
    "track": "English small model",
    "generated": 100,
    "judged": 100,
    "pass": 27,
    "fail": 73,
    "uncertain": 0,
    "pass_percent": 27.0,
    "wilson95_percent": [
      19.3,
      36.4
    ],
    "grammar_yes": 58,
    "complete_yes": 31,
    "parameters": 124440576,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/GoLLeM-45M-PL",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 9,
    "fail": 91,
    "uncertain": 0,
    "pass_percent": 9.0,
    "wilson95_percent": [
      4.8,
      16.2
    ],
    "grammar_yes": 21,
    "complete_yes": 20,
    "parameters": 42521600,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/GoLLeM-110M-PL-v3",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 2,
    "fail": 98,
    "uncertain": 0,
    "pass_percent": 2.0,
    "wilson95_percent": [
      0.6,
      7.0
    ],
    "grammar_yes": 18,
    "complete_yes": 10,
    "parameters": 110025216,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/bdh-25m-pl",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 2,
    "fail": 98,
    "uncertain": 0,
    "pass_percent": 2.0,
    "wilson95_percent": [
      0.6,
      7.0
    ],
    "grammar_yes": 5,
    "complete_yes": 2,
    "parameters": 25296896,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/goLLeM-110M-PL-SFT-merged",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 2,
    "fail": 98,
    "uncertain": 0,
    "pass_percent": 2.0,
    "wilson95_percent": [
      0.6,
      7.0
    ],
    "grammar_yes": 14,
    "complete_yes": 5,
    "parameters": 110025216,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/GoLLeM-110M-PL-v2",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 0,
    "fail": 100,
    "uncertain": 0,
    "pass_percent": 0.0,
    "wilson95_percent": [
      0.0,
      3.7
    ],
    "grammar_yes": 9,
    "complete_yes": 13,
    "parameters": 110025216,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/new-training-run-ph-d7e9fe9b",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 0,
    "fail": 100,
    "uncertain": 0,
    "pass_percent": 0.0,
    "wilson95_percent": [
      0.0,
      3.7
    ],
    "grammar_yes": 8,
    "complete_yes": 1,
    "parameters": 8160256,
    "configuration": "raw"
  },
  {
    "model": "SlayerLab/slayer-scratch",
    "track": "Polish-trained; English transfer",
    "generated": 100,
    "judged": 100,
    "pass": 0,
    "fail": 100,
    "uncertain": 0,
    "pass_percent": 0.0,
    "wilson95_percent": [
      0.0,
      3.7
    ],
    "grammar_yes": 12,
    "complete_yes": 4,
    "parameters": 110906112,
    "configuration": "raw"
  },
  {
    "model": "speakleash/Bielik-11B-v3.0-Instruct",
    "configuration": "chat_template",
    "parameters": 11168796672,
    "generated": 100,
    "pass": 87,
    "grammar_yes": 95,
    "complete_yes": 87,
    "wilson95_percent": [
      79.0,
      92.2
    ]
  }
]