Hugging Face's logo

Spaces:
Open-Style
/
OSQ-Leaderboard
Running

App Files Community
SondosMB commited on
Commit
1ffa31c
·
verified ·
1 Parent(s): 619dd33

Update small.json

Browse files
Files changed (1) hide show
  1. small.json +11 -11
small.json CHANGED
@@ -5,7 +5,7 @@
5
  "MMLU": 36.97,
6
  "ARC":60.94,
7
  "WinoGrande": 46.88,
8
- "PiQA": 32.04,
9
  "CommonsenseQA": 49.15,
10
  "Race": 37.81,
11
  "MedMCQA": 22.61,
@@ -17,7 +17,7 @@
17
  "MMLU": 9.99,
18
  "ARC":15.84 ,
19
  "WinoGrande": 40.96,
20
- "PiQA": 15.52,
21
  "CommonsenseQA": 31.13,
22
  "Race": 34.91,
23
  "MedMCQA": 4.7,
@@ -29,7 +29,7 @@
29
  "MMLU": 17.52,
30
  "ARC":23.93,
31
  "WinoGrande": 16.10,
32
- "PiQA": 15.09,
33
  "CommonsenseQA": 27.46,
34
  "Race": 14.32,
35
  "MedMCQA": 4.57,
@@ -42,7 +42,7 @@
42
  "MMLU": 9.22,
43
  "ARC":14.95,
44
  "WinoGrande": 14.76,
45
- "PiQA": 5.32,
46
  "CommonsenseQA": 9.01,
47
  "Race": 16.19,
48
  "MedMCQA": 1.68,
@@ -54,7 +54,7 @@
54
  "MMLU": 9.21,
55
  "ARC":13.5,
56
  "WinoGrande": 16.97,
57
- "PiQA": 0.86,
58
  "CommonsenseQA": 11.41,
59
  "Race": 14.35,
60
  "MedMCQA": 1.86,
@@ -66,7 +66,7 @@
66
  "MMLU": 8.54,
67
  "ARC":13.18,
68
  "WinoGrande": 6.16,
69
- "PiQA": 8.05,
70
  "CommonsenseQA": 13.10,
71
  "Race": 13.61,
72
  "MedMCQA": 2.07,
@@ -78,7 +78,7 @@
78
  "MMLU": 9.66,
79
  "ARC":14.69,
80
  "WinoGrande": 11.52,
81
- "PiQA": 4.17,
82
  "CommonsenseQA": 9.01,
83
  "Race": 12.76,
84
  "MedMCQA": 3.19,
@@ -90,7 +90,7 @@
90
  "MMLU": 8.94,
91
  "ARC":13.31,
92
  "WinoGrande": 12.23,
93
- "PiQA": 3.59,
94
  "CommonsenseQA": 6.06,
95
  "Race": 16.7,
96
  "MedMCQA": 2.07,
@@ -102,7 +102,7 @@
102
  "MMLU": 7.40,
103
  "ARC":11.83,
104
  "WinoGrande": 12.47,
105
- "PiQA": 4.48,
106
  "CommonsenseQA": 7.61,
107
  "Race": 13.61,
108
  "MedMCQA": 1.25,
@@ -114,7 +114,7 @@
114
  "MMLU": 6.94,
115
  "ARC": 6.69,
116
  "WinoGrande": 10.81,
117
- "PiQA": 4.31,
118
  "CommonsenseQA": 6.34,
119
  "Race": 13.75,
120
  "MedMCQA": 2.63,
@@ -126,7 +126,7 @@
126
  "MMLU": 5.37,
127
  "ARC":4.43,
128
  "WinoGrande": 9.31,
129
- "PiQA": 2.16,
130
  "CommonsenseQA": 6.2,
131
  "Race": 6.9,
132
  "MedMCQA": 1.04,
 
5
  "MMLU": 36.97,
6
  "ARC":60.94,
7
  "WinoGrande": 46.88,
8
+ "PIQA": 32.04,
9
  "CommonsenseQA": 49.15,
10
  "Race": 37.81,
11
  "MedMCQA": 22.61,
 
17
  "MMLU": 9.99,
18
  "ARC":15.84 ,
19
  "WinoGrande": 40.96,
20
+ "PIQA": 15.52,
21
  "CommonsenseQA": 31.13,
22
  "Race": 34.91,
23
  "MedMCQA": 4.7,
 
29
  "MMLU": 17.52,
30
  "ARC":23.93,
31
  "WinoGrande": 16.10,
32
+ "PIQA": 15.09,
33
  "CommonsenseQA": 27.46,
34
  "Race": 14.32,
35
  "MedMCQA": 4.57,
 
42
  "MMLU": 9.22,
43
  "ARC":14.95,
44
  "WinoGrande": 14.76,
45
+ "PIQA": 5.32,
46
  "CommonsenseQA": 9.01,
47
  "Race": 16.19,
48
  "MedMCQA": 1.68,
 
54
  "MMLU": 9.21,
55
  "ARC":13.5,
56
  "WinoGrande": 16.97,
57
+ "PIQA": 0.86,
58
  "CommonsenseQA": 11.41,
59
  "Race": 14.35,
60
  "MedMCQA": 1.86,
 
66
  "MMLU": 8.54,
67
  "ARC":13.18,
68
  "WinoGrande": 6.16,
69
+ "PIQA": 8.05,
70
  "CommonsenseQA": 13.10,
71
  "Race": 13.61,
72
  "MedMCQA": 2.07,
 
78
  "MMLU": 9.66,
79
  "ARC":14.69,
80
  "WinoGrande": 11.52,
81
+ "PIQA": 4.17,
82
  "CommonsenseQA": 9.01,
83
  "Race": 12.76,
84
  "MedMCQA": 3.19,
 
90
  "MMLU": 8.94,
91
  "ARC":13.31,
92
  "WinoGrande": 12.23,
93
+ "PIQA": 3.59,
94
  "CommonsenseQA": 6.06,
95
  "Race": 16.7,
96
  "MedMCQA": 2.07,
 
102
  "MMLU": 7.40,
103
  "ARC":11.83,
104
  "WinoGrande": 12.47,
105
+ "PIQA": 4.48,
106
  "CommonsenseQA": 7.61,
107
  "Race": 13.61,
108
  "MedMCQA": 1.25,
 
114
  "MMLU": 6.94,
115
  "ARC": 6.69,
116
  "WinoGrande": 10.81,
117
+ "PIQA": 4.31,
118
  "CommonsenseQA": 6.34,
119
  "Race": 13.75,
120
  "MedMCQA": 2.63,
 
126
  "MMLU": 5.37,
127
  "ARC":4.43,
128
  "WinoGrande": 9.31,
129
+ "PIQA": 2.16,
130
  "CommonsenseQA": 6.2,
131
  "Race": 6.9,
132
  "MedMCQA": 1.04,