{
 "version": "human54 formal_r7_unified_20261018",
 "scoring": "invalid task answers count as wrong",
 "interval": "95% percentile interval, 5,000 whole-family resamples, human mean fixed",
 "tasks": [
  {
   "code": "DS",
   "name": "Digit span",
   "group": "Working memory",
   "measure": "Forward minus backward accuracy",
   "source": "SRO",
   "kind": "value",
   "human": 0.047071702244116,
   "human_n": 522,
   "human_sd": 0.08869368745635464,
   "model_mean": 0.41887125220458554,
   "gap": 0.3717995499604696,
   "lower": 0.2798107382320744,
   "upper": 0.5005473453749315,
   "models": 54
  },
  {
   "code": "NB",
   "name": "N-back",
   "group": "Working memory",
   "measure": "2-back ordinary non-target false alarms",
   "source": "SRO",
   "kind": "value",
   "human": 0.0714285714285714,
   "human_n": 40,
   "human_sd": 0.09371797033750995,
   "model_mean": 0.32793209876543206,
   "gap": 0.25650352733686066,
   "lower": 0.159358144106408,
   "upper": 0.3406092693636096,
   "models": 54
  },
  {
   "code": "OS",
   "name": "Operation span",
   "group": "Working memory",
   "measure": "Letter-position accuracy",
   "source": "Ball et al.",
   "kind": "value",
   "human": 0.7632267884322678,
   "human_n": 438,
   "human_sd": 0.16100920852670042,
   "model_mean": 0.5816460905349795,
   "gap": -0.18158069789728828,
   "lower": -0.2958914422164868,
   "upper": -0.07971752917300856,
   "models": 54
  },
  {
   "code": "ST",
   "name": "Stroop",
   "group": "Cognitive control",
   "measure": "Error increase with conflict",
   "source": "SRO",
   "kind": "value",
   "human": 0.0508637236084452,
   "human_n": 521,
   "human_sd": 0.06314084178705699,
   "model_mean": 0.37101337448559674,
   "gap": 0.3201496508771515,
   "lower": 0.24652437827184423,
   "upper": 0.4479197220614894,
   "models": 54
  },
  {
   "code": "FL",
   "name": "Flanker",
   "group": "Cognitive control",
   "measure": "Error increase with conflict",
   "source": "SRO",
   "kind": "value",
   "human": 0.0593288375080282,
   "human_n": 519,
   "human_sd": 0.12297594084242892,
   "model_mean": 0.12461419753086418,
   "gap": 0.06528536002283601,
   "lower": -0.005654068989509676,
   "upper": 0.13715518462385967,
   "models": 54
  },
  {
   "code": "GN",
   "name": "Go/no-go",
   "group": "Cognitive control",
   "measure": "No-go false alarms",
   "source": "SRO",
   "kind": "value",
   "human": 0.2137647381409377,
   "human_n": 521,
   "human_sd": 0.12462865894404587,
   "model_mean": 0.2537918871252205,
   "gap": 0.04002714898428278,
   "lower": -0.06481154227925968,
   "upper": 0.19066757742094592,
   "models": 54
  },
  {
   "code": "CVLT",
   "name": "CVLT-II list learning",
   "group": "Episodic memory",
   "measure": "Recall slope (words per trial)",
   "source": "CNP",
   "kind": "value",
   "human": 1.5492307692307692,
   "human_n": 130,
   "human_sd": 0.5218105263431396,
   "model_mean": -0.28703703703703703,
   "gap": -1.836267806267807,
   "lower": -2.1378247541931747,
   "upper": -1.475752878074307,
   "models": 54
  },
  {
   "code": "DRM",
   "name": "DRM false memory",
   "group": "Episodic memory",
   "measure": "Unwarned new minus lure false alarms",
   "source": "Ball et al.",
   "kind": "value",
   "human": -0.4572916666666666,
   "human_n": 240,
   "human_sd": 0.22639377313883668,
   "model_mean": -0.16435185185185186,
   "gap": 0.29293981481481474,
   "lower": 0.2168898809523809,
   "upper": 0.3670138888888888,
   "models": 54
  },
  {
   "code": "SM",
   "name": "Source memory",
   "group": "Episodic memory",
   "measure": "Direct minus indirect source accuracy",
   "source": "Eerland et al.",
   "kind": "value",
   "human": -0.0783181357649442,
   "human_n": 299,
   "human_sd": 0.23698124083567057,
   "model_mean": -0.02662037037037035,
   "gap": 0.05169776539457386,
   "lower": 0.03795288810337955,
   "upper": 0.06190970968046931,
   "models": 54
  },
  {
   "code": "FB",
   "name": "False belief",
   "group": "Social reasoning",
   "measure": "False-belief accuracy",
   "source": "Strachan et al.",
   "kind": "value",
   "human": 0.9826589595375722,
   "human_n": 49,
   "human_sd": 0.06055651770647561,
   "model_mean": 0.7724867724867726,
   "gap": -0.21017218705079974,
   "lower": -0.29695583453757224,
   "upper": -0.12211266785016558,
   "models": 54
  },
  {
   "code": "EP",
   "name": "Recursive mindreading",
   "group": "Social reasoning",
   "measure": "Mental-state accuracy slope",
   "source": "EPITOME",
   "kind": "value",
   "human": -0.0172522360502339,
   "human_n": 63,
   "human_sd": 0.0387997501275148,
   "model_mean": -0.004009589947089937,
   "gap": 0.013242646103143962,
   "lower": 0.00849754208348638,
   "upper": 0.01898444751146757,
   "models": 54
  },
  {
   "code": "CAL",
   "name": "Confidence calibration",
   "group": "Metacognition",
   "measure": "Type-2 AUROC",
   "source": "Mazancieux et al.",
   "kind": "value",
   "human": 0.7140042319515715,
   "human_n": 181,
   "human_sd": 0.09472957775700601,
   "model_mean": 0.5733901874109643,
   "gap": -0.1406140445406072,
   "lower": -0.19596412116364337,
   "upper": -0.10450311802893408,
   "models": 54
  },
  {
   "code": "WAG",
   "name": "Post-decision wagering",
   "group": "Metacognition",
   "measure": "Always places the same wager",
   "source": "Dienes and Seth",
   "kind": "constant",
   "human": null,
   "constant_models": 37,
   "models": 54
  }
 ],
 "models": [
  {
   "model": "qwen2.5:0.5b",
   "family": "qwen",
   "params_b": 0.49,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.1904761904761904,
    "NB": 0.0,
    "OS": 0.2044444444444444,
    "ST": 0.8333333333333334,
    "FL": 0.0069444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0520833333333333,
    "FB": 0.1428571428571428,
    "EP": -0.015625,
    "CAL": 0.5,
    "WAG": true
   }
  },
  {
   "model": "qwen2.5:1.5b",
   "family": "qwen",
   "params_b": 1.54,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.3333333333333333,
    "NB": 0.0,
    "OS": 0.0,
    "ST": -0.0625,
    "FL": 0.6180555555555556,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0416666666666666,
    "FB": 0.8571428571428571,
    "EP": 0.015625,
    "CAL": 0.4731081926203877,
    "WAG": true
   }
  },
  {
   "model": "qwen2.5:3b",
   "family": "qwen",
   "params_b": 3.09,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.2619047619047618,
    "NB": 0.386574074074074,
    "OS": 0.0177777777777777,
    "ST": 0.6666666666666667,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.0520833333333332,
    "FB": 0.7142857142857143,
    "EP": -0.0111607142857142,
    "CAL": 0.6128526645768025,
    "WAG": true
   }
  },
  {
   "model": "qwen2.5:7b",
   "family": "qwen",
   "params_b": 7.62,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.4047619047619047,
    "NB": 0.375,
    "OS": 0.5644444444444444,
    "ST": 0.3333333333333333,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": -3.2,
    "DRM": -0.125,
    "SM": 0.0208333333333333,
    "FB": 1.0,
    "EP": -0.015625,
    "CAL": 0.5978336827393431,
    "WAG": false
   }
  },
  {
   "model": "qwen2.5:14b",
   "family": "qwen",
   "params_b": 14.77,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0952380952380952,
    "NB": 0.8726851851851852,
    "OS": 0.9111111111111112,
    "ST": 0.0,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0022321428571428,
    "CAL": 0.8738244514106583,
    "WAG": false
   }
  },
  {
   "model": "qwen2.5:32b",
   "family": "qwen",
   "params_b": 32.76,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0238095238095238,
    "NB": 0.9074074074074074,
    "OS": 0.9644444444444444,
    "ST": 0.0,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0044642857142857,
    "CAL": 0.7921928817451206,
    "WAG": false
   }
  },
  {
   "model": "gemma2:2b",
   "family": "gemma",
   "params_b": 2.61,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.4285714285714286,
    "NB": 0.199074074074074,
    "OS": 0.8622222222222222,
    "ST": 0.0,
    "FL": -0.0069444444444444,
    "GN": 0.8476190476190476,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0833333333333332,
    "FB": 0.5714285714285714,
    "EP": 0.0446428571428571,
    "CAL": 0.5163817663817664,
    "WAG": true
   }
  },
  {
   "model": "gemma2:9b",
   "family": "gemma",
   "params_b": 9.24,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.2142857142857143,
    "NB": 0.5925925925925926,
    "OS": 0.9955555555555556,
    "ST": 0.3402777777777778,
    "FL": -0.0625,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.875,
    "SM": -0.0104166666666667,
    "FB": 1.0,
    "EP": 0.0,
    "CAL": 0.5,
    "WAG": true
   }
  },
  {
   "model": "gemma2:27b",
   "family": "gemma",
   "params_b": 27.23,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0,
    "NB": 0.4027777777777778,
    "OS": 0.6177777777777778,
    "ST": 0.0,
    "FL": -0.125,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.375,
    "SM": 0.0,
    "FB": 0.5714285714285714,
    "EP": 0.0089285714285714,
    "CAL": 0.7,
    "WAG": true
   }
  },
  {
   "model": "llama3.2:1b",
   "family": "llama",
   "params_b": 1.24,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5952380952380953,
    "NB": 0.0046296296296296,
    "OS": 0.4622222222222222,
    "ST": -0.3333333333333333,
    "FL": -0.4375,
    "GN": 0.580952380952381,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0625,
    "FB": 0.1428571428571428,
    "EP": -0.0066964285714285,
    "CAL": 0.554375,
    "WAG": true
   }
  },
  {
   "model": "llama3.2:3b",
   "family": "llama",
   "params_b": 3.21,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.761904761904762,
    "NB": 0.6944444444444444,
    "OS": 0.6977777777777778,
    "ST": 0.0833333333333333,
    "FL": 0.0,
    "GN": 0.0380952380952381,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.0520833333333333,
    "FB": 0.2857142857142857,
    "EP": 0.0022321428571428,
    "CAL": 0.635,
    "WAG": true
   }
  },
  {
   "model": "llama3.1:8b",
   "family": "llama",
   "params_b": 8.03,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.3809523809523809,
    "NB": 0.0902777777777777,
    "OS": 0.2444444444444444,
    "ST": 0.8263888888888888,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.4,
    "DRM": -0.125,
    "SM": -0.03125,
    "FB": 1.0,
    "EP": 0.0178571428571428,
    "CAL": 0.7222222222222222,
    "WAG": false
   }
  },
  {
   "model": "deepseek-r1:7b",
   "family": "deepseek-r",
   "params_b": 7.62,
   "native_thinking": true,
   "serving": "paper",
   "values": {
    "DS": 0.0952380952380952,
    "NB": 0.0138888888888888,
    "OS": 0.0711111111111111,
    "ST": -0.0208333333333333,
    "FL": 0.0694444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.0416666666666666,
    "FB": 0.2857142857142857,
    "EP": -0.0133928571428571,
    "CAL": 0.4864603481624758,
    "WAG": false
   }
  },
  {
   "model": "deepseek-r1:14b",
   "family": "deepseek-r",
   "params_b": 14.77,
   "native_thinking": true,
   "serving": "paper",
   "values": {
    "DS": 0.0,
    "NB": 0.1412037037037037,
    "OS": 0.0488888888888888,
    "ST": 0.3402777777777778,
    "FL": -0.0763888888888889,
    "GN": 0.0,
    "CVLT": -3.2,
    "DRM": -0.125,
    "SM": -0.09375,
    "FB": 1.0,
    "EP": -0.0066964285714285,
    "CAL": 0.2180306905370844,
    "WAG": false
   }
  },
  {
   "model": "mistral:7b",
   "family": "mistral",
   "params_b": 7.25,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.9285714285714286,
    "NB": 0.3055555555555556,
    "OS": 0.48,
    "ST": 0.8333333333333334,
    "FL": 0.1319444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.03125,
    "FB": 1.0,
    "EP": 0.0022321428571428,
    "CAL": 0.4834347532116295,
    "WAG": true
   }
  },
  {
   "model": "mixtral:8x7b",
   "family": "mixtral",
   "params_b": 46.7,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.4523809523809523,
    "NB": 0.1134259259259259,
    "OS": 0.6577777777777778,
    "ST": 0.0,
    "FL": -0.125,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0208333333333332,
    "FB": 1.0,
    "EP": -0.0044642857142857,
    "CAL": 0.5909090909090909,
    "WAG": false
   }
  },
  {
   "model": "phi3:14b",
   "family": "phi",
   "params_b": 13.96,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.6666666666666667,
    "NB": 0.4675925925925926,
    "OS": 0.5777777777777777,
    "ST": 0.4305555555555556,
    "FL": 0.4444444444444444,
    "GN": 0.0,
    "CVLT": 0.2,
    "DRM": -0.125,
    "SM": -0.0833333333333333,
    "FB": 1.0,
    "EP": -0.0111607142857142,
    "CAL": 0.6547619047619048,
    "WAG": true
   }
  },
  {
   "model": "yi:34b",
   "family": "yi",
   "params_b": 34.39,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5238095238095238,
    "NB": 0.0601851851851851,
    "OS": 1.0,
    "ST": 0.3333333333333333,
    "FL": 0.5486111111111112,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0104166666666667,
    "FB": 1.0,
    "EP": 0.0,
    "CAL": 0.6232727272727273,
    "WAG": true
   }
  },
  {
   "model": "command-r:35b",
   "family": "command-r",
   "params_b": 32.3,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5,
    "NB": 0.4560185185185185,
    "OS": 0.9244444444444444,
    "ST": 0.1805555555555555,
    "FL": 0.5,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.75,
    "SM": 0.0104166666666667,
    "FB": 0.5714285714285714,
    "EP": -0.015625,
    "CAL": 0.5117948717948718,
    "WAG": false
   }
  },
  {
   "model": "aya:8b",
   "family": "aya",
   "params_b": 8.03,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.3333333333333333,
    "NB": 1.0,
    "OS": 0.6888888888888889,
    "ST": 0.6666666666666667,
    "FL": 0.8125,
    "GN": 0.7619047619047619,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0312499999999998,
    "FB": 0.4285714285714285,
    "EP": 0.015625,
    "CAL": 0.5766666666666667,
    "WAG": true
   }
  },
  {
   "model": "deepseek-llm:7b",
   "family": "deepseek",
   "params_b": 6.91,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.3571428571428571,
    "NB": 0.0208333333333333,
    "OS": 0.0933333333333333,
    "ST": 0.3333333333333333,
    "FL": -0.1319444444444444,
    "GN": 1.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.1250000000000001,
    "FB": 0.8571428571428571,
    "EP": -0.03125,
    "CAL": 0.3646028767979987,
    "WAG": true
   }
  },
  {
   "model": "exaone3.5:7.8b",
   "family": "exaone",
   "params_b": 7.82,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7142857142857143,
    "NB": 0.0555555555555555,
    "OS": 0.1066666666666666,
    "ST": 0.8333333333333334,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.5,
    "SM": -0.0312500000000001,
    "FB": 1.0,
    "EP": -0.0111607142857142,
    "CAL": 0.5178571428571429,
    "WAG": false
   }
  },
  {
   "model": "falcon3:10b",
   "family": "falcon",
   "params_b": 10.31,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0714285714285714,
    "NB": 0.1967592592592592,
    "OS": 1.0,
    "ST": 0.8333333333333334,
    "FL": 0.125,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.375,
    "SM": 0.0,
    "FB": 0.8571428571428571,
    "EP": 0.015625,
    "CAL": 0.5,
    "WAG": true
   }
  },
  {
   "model": "falcon3:7b",
   "family": "falcon",
   "params_b": 7.46,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.3809523809523809,
    "NB": 0.6597222222222222,
    "OS": 0.9511111111111112,
    "ST": 0.6805555555555556,
    "FL": -0.0069444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0,
    "FB": 0.7142857142857143,
    "EP": 0.0133928571428571,
    "CAL": 0.5779487179487179,
    "WAG": true
   }
  },
  {
   "model": "gemma3:12b",
   "family": "gemma",
   "params_b": 12.2,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0238095238095238,
    "NB": 0.0833333333333333,
    "OS": 1.0,
    "ST": 1.0,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0022321428571428,
    "CAL": 0.5477308294209703,
    "WAG": false
   }
  },
  {
   "model": "gemma3:1b",
   "family": "gemma",
   "params_b": 1.0,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0952380952380952,
    "NB": 0.3611111111111111,
    "OS": 0.0,
    "ST": 0.5,
    "FL": -0.4375,
    "GN": 0.7523809523809524,
    "CVLT": 0.0,
    "DRM": -0.25,
    "SM": -0.0833333333333333,
    "FB": 0.0,
    "EP": 0.0022321428571428,
    "CAL": 0.4886363636363636,
    "WAG": false
   }
  },
  {
   "model": "gemma3:27b",
   "family": "gemma",
   "params_b": 27.4,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0238095238095238,
    "NB": 0.1851851851851851,
    "OS": 1.0,
    "ST": 0.0069444444444444,
    "FL": 0.375,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.375,
    "SM": 0.0104166666666666,
    "FB": 1.0,
    "EP": -0.0178571428571428,
    "CAL": 0.5698198198198198,
    "WAG": true
   }
  },
  {
   "model": "glm4:9b",
   "family": "glm",
   "params_b": 9.4,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5,
    "NB": 1.0,
    "OS": 0.5911111111111111,
    "ST": 0.7986111111111112,
    "FL": 0.4861111111111111,
    "GN": 0.6571428571428571,
    "CVLT": 0.0,
    "DRM": -0.625,
    "SM": -0.0624999999999998,
    "FB": 1.0,
    "EP": 0.0111607142857142,
    "CAL": 0.5546825895663104,
    "WAG": true
   }
  },
  {
   "model": "internlm2:7b",
   "family": "internlm",
   "params_b": 7.74,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5476190476190477,
    "NB": 0.8634259259259259,
    "OS": 0.6222222222222222,
    "ST": 0.6736111111111112,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.25,
    "SM": -0.0312499999999998,
    "FB": 0.8571428571428571,
    "EP": -0.0267857142857142,
    "CAL": 0.5165728580362726,
    "WAG": true
   }
  },
  {
   "model": "llama2:13b",
   "family": "llama",
   "params_b": 13.02,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.4523809523809524,
    "NB": 0.9976851851851852,
    "OS": 0.9022222222222224,
    "ST": 0.3333333333333333,
    "FL": 0.4375,
    "GN": 0.7428571428571429,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0729166666666667,
    "FB": 0.5714285714285714,
    "EP": -0.0178571428571428,
    "CAL": 0.466984126984127,
    "WAG": false
   }
  },
  {
   "model": "llama2:7b",
   "family": "llama",
   "params_b": 6.74,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7142857142857143,
    "NB": 1.0,
    "OS": 0.8711111111111111,
    "ST": 0.5138888888888888,
    "FL": 0.0,
    "GN": 1.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0416666666666666,
    "FB": 0.1428571428571428,
    "EP": -0.0245535714285714,
    "CAL": 0.5834398976982097,
    "WAG": false
   }
  },
  {
   "model": "llama3.1:70b",
   "family": "llama",
   "params_b": 70.55,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0476190476190476,
    "NB": 0.6412037037037037,
    "OS": 1.0,
    "ST": 0.0,
    "FL": 0.2569444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.875,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0111607142857142,
    "CAL": 0.951602765556254,
    "WAG": false
   }
  },
  {
   "model": "mistral-nemo:12b",
   "family": "mistral",
   "params_b": 12.25,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7142857142857143,
    "NB": 0.2847222222222222,
    "OS": 0.7244444444444444,
    "ST": 0.8194444444444444,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": -4.8,
    "DRM": -0.5,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0089285714285714,
    "CAL": 0.5519713261648745,
    "WAG": true
   }
  },
  {
   "model": "mixtral:8x22b",
   "family": "mixtral",
   "params_b": 140.63,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.2142857142857143,
    "NB": 0.037037037037037,
    "OS": 0.9911111111111112,
    "ST": 0.3194444444444444,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.75,
    "SM": 0.0,
    "FB": 1.0,
    "EP": 0.0044642857142857,
    "CAL": 0.5,
    "WAG": true
   }
  },
  {
   "model": "nemotron-mini:4b",
   "family": "nemotron",
   "params_b": 4.19,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5,
    "NB": 0.4884259259259259,
    "OS": 0.2355555555555555,
    "ST": 0.4027777777777778,
    "FL": 0.0,
    "GN": 0.6666666666666666,
    "CVLT": -0.1,
    "DRM": 0.0,
    "SM": -0.09375,
    "FB": 0.5714285714285714,
    "EP": -0.0066964285714285,
    "CAL": 0.4658284977433913,
    "WAG": true
   }
  },
  {
   "model": "olmo2:13b",
   "family": "olmo",
   "params_b": 13.72,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5714285714285714,
    "NB": 0.0092592592592592,
    "OS": 0.3511111111111111,
    "ST": 0.0277777777777777,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.03125,
    "FB": 0.5714285714285714,
    "EP": 0.0200892857142857,
    "CAL": 0.58,
    "WAG": true
   }
  },
  {
   "model": "olmo2:7b",
   "family": "olmo",
   "params_b": 7.3,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7380952380952381,
    "NB": 0.1458333333333333,
    "OS": 0.0977777777777777,
    "ST": 0.5069444444444444,
    "FL": 0.4583333333333333,
    "GN": 0.0571428571428571,
    "CVLT": 3.2,
    "DRM": 0.0,
    "SM": -0.0520833333333333,
    "FB": 1.0,
    "EP": 0.015625,
    "CAL": 0.5509693558474046,
    "WAG": true
   }
  },
  {
   "model": "openchat:7b",
   "family": "openchat",
   "params_b": 7.24,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 1.0,
    "NB": 0.1180555555555555,
    "OS": 0.7511111111111111,
    "ST": 0.6736111111111112,
    "FL": 0.0,
    "GN": 0.9428571428571428,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.0312500000000001,
    "FB": 0.8571428571428571,
    "EP": 0.0245535714285714,
    "CAL": 0.5278413424866514,
    "WAG": true
   }
  },
  {
   "model": "phi3:3.8b",
   "family": "phi",
   "params_b": 3.82,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.8333333333333334,
    "NB": 0.2476851851851851,
    "OS": 0.3911111111111111,
    "ST": 0.3263888888888888,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.1041666666666666,
    "FB": 1.0,
    "EP": -0.0290178571428571,
    "CAL": 0.5827067669172933,
    "WAG": true
   }
  },
  {
   "model": "phi4:14b",
   "family": "phi",
   "params_b": 14.66,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0714285714285714,
    "NB": 0.0,
    "OS": 0.9333333333333332,
    "ST": 0.0,
    "FL": 0.125,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.5,
    "SM": 0.0,
    "FB": 1.0,
    "EP": 0.0111607142857142,
    "CAL": 0.7068511198945981,
    "WAG": false
   }
  },
  {
   "model": "qwen2.5:72b",
   "family": "qwen",
   "params_b": 72.71,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0,
    "NB": 1.0,
    "OS": 1.0,
    "ST": 0.0,
    "FL": 0.0,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.125,
    "FB": 1.0,
    "EP": 0.0022321428571428,
    "CAL": 0.9502351097178684,
    "WAG": false
   }
  },
  {
   "model": "qwen3:0.6b",
   "family": "qwen",
   "params_b": 0.75,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5238095238095238,
    "NB": 0.5671296296296297,
    "OS": 0.0222222222222222,
    "ST": 0.1666666666666666,
    "FL": 0.0,
    "GN": 0.2761904761904762,
    "CVLT": -4.8,
    "DRM": 0.0,
    "SM": -0.0104166666666666,
    "FB": 0.8571428571428571,
    "EP": -0.0089285714285714,
    "CAL": 0.5234521575984991,
    "WAG": true
   }
  },
  {
   "model": "qwen3:1.7b",
   "family": "qwen",
   "params_b": 2.0,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.8571428571428572,
    "NB": 0.2800925925925926,
    "OS": 0.2133333333333333,
    "ST": 0.3472222222222222,
    "FL": 0.0,
    "GN": 0.9904761904761904,
    "CVLT": 0.0,
    "DRM": -0.125,
    "SM": -0.09375,
    "FB": 0.5714285714285714,
    "EP": 0.0,
    "CAL": 0.412593984962406,
    "WAG": true
   }
  },
  {
   "model": "qwen3:14b",
   "family": "qwen",
   "params_b": 14.8,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.0238095238095238,
    "NB": 0.1064814814814814,
    "OS": 1.0,
    "ST": 0.1597222222222222,
    "FL": 0.7569444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.25,
    "SM": -0.0416666666666667,
    "FB": 1.0,
    "EP": 0.0,
    "CAL": 0.5,
    "WAG": true
   }
  },
  {
   "model": "qwen3:4b",
   "family": "qwen",
   "params_b": 4.0,
   "native_thinking": true,
   "serving": "paper",
   "values": {
    "DS": 0.0,
    "NB": 0.0,
    "OS": 0.7955555555555556,
    "ST": 0.0,
    "FL": -0.125,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0,
    "FB": 1.0,
    "EP": -0.0424107142857142,
    "CAL": 0.8859375,
    "WAG": true
   }
  },
  {
   "model": "qwen3:8b",
   "family": "qwen",
   "params_b": 8.2,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7857142857142857,
    "NB": 0.0,
    "OS": 1.0,
    "ST": 0.3402777777777778,
    "FL": 0.9305555555555556,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.625,
    "SM": -0.0208333333333333,
    "FB": 0.2857142857142857,
    "EP": -0.0089285714285714,
    "CAL": 0.7361984626135569,
    "WAG": true
   }
  },
  {
   "model": "smollm2:1.7b",
   "family": "smollm",
   "params_b": 1.71,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5714285714285714,
    "NB": 0.0,
    "OS": 0.0,
    "ST": 0.6666666666666667,
    "FL": 0.0,
    "GN": 0.9809523809523808,
    "CVLT": -3.2,
    "DRM": 0.0,
    "SM": -0.09375,
    "FB": 0.4285714285714285,
    "EP": -0.0446428571428571,
    "CAL": 0.4855889724310777,
    "WAG": true
   }
  },
  {
   "model": "smollm2:360m",
   "family": "smollm",
   "params_b": 0.36,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.1904761904761904,
    "NB": 0.0,
    "OS": 0.0577777777777777,
    "ST": 0.0,
    "FL": 0.9930555555555556,
    "GN": 1.0,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": 0.0208333333333333,
    "FB": 0.2857142857142857,
    "EP": -0.0089285714285714,
    "CAL": 0.4871875,
    "WAG": true
   }
  },
  {
   "model": "solar:10.7b",
   "family": "solar",
   "params_b": 10.73,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.5952380952380951,
    "NB": 0.0972222222222222,
    "OS": 0.2355555555555555,
    "ST": 0.8402777777777778,
    "FL": 0.2569444444444444,
    "GN": 0.0,
    "CVLT": 0.0,
    "DRM": -0.25,
    "SM": -0.0312499999999998,
    "FB": 1.0,
    "EP": -0.0200892857142857,
    "CAL": 0.5790513833992095,
    "WAG": true
   }
  },
  {
   "model": "stablelm2:1.6b",
   "family": "stablelm",
   "params_b": 1.64,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.119047619047619,
    "NB": 0.0069444444444444,
    "OS": 0.0622222222222222,
    "ST": 0.6666666666666667,
    "FL": -0.1944444444444444,
    "GN": 0.7523809523809524,
    "CVLT": 0.0,
    "DRM": 0.25,
    "SM": 0.0312500000000001,
    "FB": 0.7142857142857143,
    "EP": 0.0111607142857142,
    "CAL": 0.4974984365228267,
    "WAG": false
   }
  },
  {
   "model": "starling-lm:7b",
   "family": "starling",
   "params_b": 7.24,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 1.0,
    "NB": 0.1157407407407407,
    "OS": 0.8711111111111111,
    "ST": 0.8402777777777778,
    "FL": 0.125,
    "GN": 0.8095238095238095,
    "CVLT": 0.0,
    "DRM": -0.5,
    "SM": -0.0520833333333333,
    "FB": 1.0,
    "EP": 0.0066964285714285,
    "CAL": 0.5508241758241759,
    "WAG": true
   }
  },
  {
   "model": "yi:6b",
   "family": "yi",
   "params_b": 6.06,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.8095238095238095,
    "NB": 0.1550925925925926,
    "OS": 0.9511111111111112,
    "ST": 0.3055555555555556,
    "FL": 0.0,
    "GN": 0.019047619047619,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.0104166666666666,
    "FB": 1.0,
    "EP": -0.0491071428571428,
    "CAL": 0.4678837052062204,
    "WAG": true
   }
  },
  {
   "model": "yi:9b",
   "family": "yi",
   "params_b": 8.83,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.7380952380952381,
    "NB": 0.4259259259259259,
    "OS": 0.9511111111111112,
    "ST": 0.0,
    "FL": 0.0,
    "GN": 0.019047619047619,
    "CVLT": 0.0,
    "DRM": 0.5,
    "SM": -0.0416666666666667,
    "FB": 1.0,
    "EP": -0.0133928571428571,
    "CAL": 0.5816170861937452,
    "WAG": true
   }
  },
  {
   "model": "zephyr:7b",
   "family": "zephyr",
   "params_b": 7.24,
   "native_thinking": false,
   "serving": "paper",
   "values": {
    "DS": 0.6428571428571429,
    "NB": 0.474537037037037,
    "OS": 0.6444444444444445,
    "ST": 0.6666666666666667,
    "FL": 0.0,
    "GN": 0.8095238095238095,
    "CVLT": 0.0,
    "DRM": 0.0,
    "SM": -0.03125,
    "FB": 1.0,
    "EP": 0.0290178571428571,
    "CAL": 0.5758333333333333,
    "WAG": true
   }
  }
 ]
}
