update leaderboard data
Browse files
app.py
CHANGED
@@ -630,23 +630,7 @@ COLUMNS = [
|
|
630 |
DATA = [
|
631 |
(
|
632 |
1,
|
633 |
-
|
634 |
-
"GPT-4-0125-Preview",
|
635 |
-
"OpenAI",
|
636 |
-
"Proprietary",
|
637 |
-
82.18,
|
638 |
-
90.00,
|
639 |
-
90.00,
|
640 |
-
91.00,
|
641 |
-
54.12,
|
642 |
-
70.00,
|
643 |
-
76.00,
|
644 |
-
55.00,
|
645 |
-
87.50,
|
646 |
-
),
|
647 |
-
(
|
648 |
-
2,
|
649 |
-
83.55,
|
650 |
"GPT-4-1106-Preview",
|
651 |
"OpenAI",
|
652 |
"Proprietary",
|
@@ -654,15 +638,31 @@ DATA = [
|
|
654 |
89.50,
|
655 |
92.00,
|
656 |
92.00,
|
657 |
-
|
658 |
62.00,
|
659 |
72.00,
|
660 |
50.00,
|
661 |
88.75,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
662 |
),
|
663 |
(
|
664 |
3,
|
665 |
-
83.
|
666 |
"Gorilla-OpenFunctions-v2",
|
667 |
"Gorilla LLM",
|
668 |
"Apache 2.0",
|
@@ -670,7 +670,7 @@ DATA = [
|
|
670 |
89.50,
|
671 |
79.50,
|
672 |
78.00,
|
673 |
-
|
674 |
74.00,
|
675 |
76.00,
|
676 |
60.00,
|
@@ -678,7 +678,7 @@ DATA = [
|
|
678 |
),
|
679 |
(
|
680 |
4,
|
681 |
-
|
682 |
"GPT-3.5-Turbo",
|
683 |
"OpenAI",
|
684 |
"Proprietary",
|
@@ -686,7 +686,7 @@ DATA = [
|
|
686 |
88.00,
|
687 |
87.50,
|
688 |
88.00,
|
689 |
-
|
690 |
74.00,
|
691 |
70.00,
|
692 |
47.50,
|
@@ -694,7 +694,7 @@ DATA = [
|
|
694 |
),
|
695 |
(
|
696 |
5,
|
697 |
-
79.
|
698 |
"Mistral-Medium",
|
699 |
"Mistral AI",
|
700 |
"Proprietary",
|
@@ -702,7 +702,7 @@ DATA = [
|
|
702 |
84.50,
|
703 |
71.00,
|
704 |
68.00,
|
705 |
-
|
706 |
72.00,
|
707 |
62.00,
|
708 |
47.50,
|
@@ -710,7 +710,7 @@ DATA = [
|
|
710 |
),
|
711 |
(
|
712 |
6,
|
713 |
-
|
714 |
"Claude-2.1",
|
715 |
"Anthropic",
|
716 |
"Proprietary",
|
@@ -718,7 +718,7 @@ DATA = [
|
|
718 |
83.00,
|
719 |
72.00,
|
720 |
56.50,
|
721 |
-
|
722 |
48.00,
|
723 |
60.00,
|
724 |
45.00,
|
@@ -726,7 +726,7 @@ DATA = [
|
|
726 |
),
|
727 |
(
|
728 |
7,
|
729 |
-
|
730 |
"Mistral-tiny",
|
731 |
"Mistral AI",
|
732 |
"Proprietary",
|
@@ -734,7 +734,7 @@ DATA = [
|
|
734 |
59.50,
|
735 |
53.50,
|
736 |
41.50,
|
737 |
-
|
738 |
64.00,
|
739 |
42.00,
|
740 |
40.00,
|
@@ -742,7 +742,7 @@ DATA = [
|
|
742 |
),
|
743 |
(
|
744 |
8,
|
745 |
-
59.
|
746 |
"Claude-instant",
|
747 |
"Anthropic",
|
748 |
"Proprietary",
|
@@ -750,7 +750,7 @@ DATA = [
|
|
750 |
59.00,
|
751 |
53.00,
|
752 |
39.50,
|
753 |
-
|
754 |
52.00,
|
755 |
50.00,
|
756 |
37.50,
|
@@ -758,7 +758,23 @@ DATA = [
|
|
758 |
),
|
759 |
(
|
760 |
9,
|
761 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
762 |
"Mistral-large",
|
763 |
"Mistral AI",
|
764 |
"Proprietary",
|
@@ -766,31 +782,31 @@ DATA = [
|
|
766 |
90.50,
|
767 |
4.00,
|
768 |
0.00,
|
769 |
-
|
770 |
66.00,
|
771 |
0.00,
|
772 |
5.00,
|
773 |
84.58,
|
774 |
),
|
775 |
(
|
776 |
-
|
777 |
-
|
778 |
-
"
|
779 |
-
"
|
780 |
"Proprietary",
|
781 |
-
|
782 |
-
|
783 |
-
|
784 |
-
|
785 |
-
|
786 |
-
|
787 |
-
|
788 |
-
|
789 |
-
|
790 |
),
|
791 |
(
|
792 |
-
|
793 |
-
|
794 |
"Nexusflow-Raven-v2",
|
795 |
"Nexusflow",
|
796 |
"Apache 2.0",
|
@@ -798,15 +814,15 @@ DATA = [
|
|
798 |
83.50,
|
799 |
39.50,
|
800 |
34.00,
|
801 |
-
|
802 |
78.00,
|
803 |
68.00,
|
804 |
45.00,
|
805 |
0.00,
|
806 |
),
|
807 |
(
|
808 |
-
|
809 |
-
|
810 |
"FireFunction-v1",
|
811 |
"Fireworks",
|
812 |
"Apache 2.0",
|
@@ -814,31 +830,15 @@ DATA = [
|
|
814 |
87.00,
|
815 |
4.00,
|
816 |
0.00,
|
817 |
-
|
818 |
64.00,
|
819 |
0.00,
|
820 |
5.00,
|
821 |
81.25,
|
822 |
),
|
823 |
-
(
|
824 |
-
13,
|
825 |
-
53.86,
|
826 |
-
"Mistral-small",
|
827 |
-
"Mistral AI",
|
828 |
-
"Proprietary",
|
829 |
-
46.55,
|
830 |
-
68.00,
|
831 |
-
48.50,
|
832 |
-
58.00,
|
833 |
-
14.12,
|
834 |
-
30.00,
|
835 |
-
40.00,
|
836 |
-
37.50,
|
837 |
-
89.58,
|
838 |
-
),
|
839 |
(
|
840 |
14,
|
841 |
-
|
842 |
"GPT-4-0613",
|
843 |
"OpenAI",
|
844 |
"Proprietary",
|
@@ -846,7 +846,7 @@ DATA = [
|
|
846 |
86.00,
|
847 |
4.00,
|
848 |
0.00,
|
849 |
-
|
850 |
50.00,
|
851 |
0.00,
|
852 |
0.00,
|
@@ -854,23 +854,7 @@ DATA = [
|
|
854 |
),
|
855 |
(
|
856 |
15,
|
857 |
-
|
858 |
-
"Gemma",
|
859 |
-
"Google",
|
860 |
-
"gemma-terms-of-use",
|
861 |
-
61.45,
|
862 |
-
60.00,
|
863 |
-
41.00,
|
864 |
-
32.00,
|
865 |
-
45.88,
|
866 |
-
46.00,
|
867 |
-
44.00,
|
868 |
-
25.00,
|
869 |
-
0.42,
|
870 |
-
),
|
871 |
-
(
|
872 |
-
16,
|
873 |
-
43.19,
|
874 |
"Deepseek-v1.5",
|
875 |
"Deepseek",
|
876 |
"Deepseek License",
|
@@ -878,12 +862,28 @@ DATA = [
|
|
878 |
61.00,
|
879 |
35.00,
|
880 |
43.50,
|
881 |
-
|
882 |
2.00,
|
883 |
0.00,
|
884 |
7.50,
|
885 |
66.25,
|
886 |
),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
887 |
(
|
888 |
17,
|
889 |
33.61,
|
|
|
630 |
DATA = [
|
631 |
(
|
632 |
1,
|
633 |
+
85.24,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
634 |
"GPT-4-1106-Preview",
|
635 |
"OpenAI",
|
636 |
"Proprietary",
|
|
|
638 |
89.50,
|
639 |
92.00,
|
640 |
92.00,
|
641 |
+
70.00,
|
642 |
62.00,
|
643 |
72.00,
|
644 |
50.00,
|
645 |
88.75,
|
646 |
+
),
|
647 |
+
(
|
648 |
+
2,
|
649 |
+
85.12,
|
650 |
+
"GPT-4-0125-Preview",
|
651 |
+
"OpenAI",
|
652 |
+
"Proprietary",
|
653 |
+
82.18,
|
654 |
+
90.00,
|
655 |
+
90.00,
|
656 |
+
91.00,
|
657 |
+
67.06,
|
658 |
+
70.00,
|
659 |
+
76.00,
|
660 |
+
55.00,
|
661 |
+
87.50,
|
662 |
),
|
663 |
(
|
664 |
3,
|
665 |
+
83.67,
|
666 |
"Gorilla-OpenFunctions-v2",
|
667 |
"Gorilla LLM",
|
668 |
"Apache 2.0",
|
|
|
670 |
89.50,
|
671 |
79.50,
|
672 |
78.00,
|
673 |
+
80.00,
|
674 |
74.00,
|
675 |
76.00,
|
676 |
60.00,
|
|
|
678 |
),
|
679 |
(
|
680 |
4,
|
681 |
+
82.23,
|
682 |
"GPT-3.5-Turbo",
|
683 |
"OpenAI",
|
684 |
"Proprietary",
|
|
|
686 |
88.00,
|
687 |
87.50,
|
688 |
88.00,
|
689 |
+
80.00,
|
690 |
74.00,
|
691 |
70.00,
|
692 |
47.50,
|
|
|
694 |
),
|
695 |
(
|
696 |
5,
|
697 |
+
79.70,
|
698 |
"Mistral-Medium",
|
699 |
"Mistral AI",
|
700 |
"Proprietary",
|
|
|
702 |
84.50,
|
703 |
71.00,
|
704 |
68.00,
|
705 |
+
78.24,
|
706 |
72.00,
|
707 |
62.00,
|
708 |
47.50,
|
|
|
710 |
),
|
711 |
(
|
712 |
6,
|
713 |
+
76.02,
|
714 |
"Claude-2.1",
|
715 |
"Anthropic",
|
716 |
"Proprietary",
|
|
|
718 |
83.00,
|
719 |
72.00,
|
720 |
56.50,
|
721 |
+
63.53,
|
722 |
48.00,
|
723 |
60.00,
|
724 |
45.00,
|
|
|
726 |
),
|
727 |
(
|
728 |
7,
|
729 |
+
60.06,
|
730 |
"Mistral-tiny",
|
731 |
"Mistral AI",
|
732 |
"Proprietary",
|
|
|
734 |
59.50,
|
735 |
53.50,
|
736 |
41.50,
|
737 |
+
63.53,
|
738 |
64.00,
|
739 |
42.00,
|
740 |
40.00,
|
|
|
742 |
),
|
743 |
(
|
744 |
8,
|
745 |
+
59.70,
|
746 |
"Claude-instant",
|
747 |
"Anthropic",
|
748 |
"Proprietary",
|
|
|
750 |
59.00,
|
751 |
53.00,
|
752 |
39.50,
|
753 |
+
56.47,
|
754 |
52.00,
|
755 |
50.00,
|
756 |
37.50,
|
|
|
758 |
),
|
759 |
(
|
760 |
9,
|
761 |
+
56.79,
|
762 |
+
"Gemini-1.0-Pro",
|
763 |
+
"Google",
|
764 |
+
"Proprietary",
|
765 |
+
78.43,
|
766 |
+
89.00,
|
767 |
+
4.00,
|
768 |
+
0.00,
|
769 |
+
63.77,
|
770 |
+
62.00,
|
771 |
+
0.00,
|
772 |
+
0.00,
|
773 |
+
77.50,
|
774 |
+
),
|
775 |
+
(
|
776 |
+
10,
|
777 |
+
56.39,
|
778 |
"Mistral-large",
|
779 |
"Mistral AI",
|
780 |
"Proprietary",
|
|
|
782 |
90.50,
|
783 |
4.00,
|
784 |
0.00,
|
785 |
+
67.06,
|
786 |
66.00,
|
787 |
0.00,
|
788 |
5.00,
|
789 |
84.58,
|
790 |
),
|
791 |
(
|
792 |
+
11,
|
793 |
+
55.72,
|
794 |
+
"Mistral-small",
|
795 |
+
"Mistral AI",
|
796 |
"Proprietary",
|
797 |
+
46.55,
|
798 |
+
68.00,
|
799 |
+
48.50,
|
800 |
+
58.00,
|
801 |
+
32.35,
|
802 |
+
30.00,
|
803 |
+
40.00,
|
804 |
+
37.50,
|
805 |
+
89.58,
|
806 |
),
|
807 |
(
|
808 |
+
12,
|
809 |
+
55.72,
|
810 |
"Nexusflow-Raven-v2",
|
811 |
"Nexusflow",
|
812 |
"Apache 2.0",
|
|
|
814 |
83.50,
|
815 |
39.50,
|
816 |
34.00,
|
817 |
+
58.24,
|
818 |
78.00,
|
819 |
68.00,
|
820 |
45.00,
|
821 |
0.00,
|
822 |
),
|
823 |
(
|
824 |
+
13,
|
825 |
+
55.33,
|
826 |
"FireFunction-v1",
|
827 |
"Fireworks",
|
828 |
"Apache 2.0",
|
|
|
830 |
87.00,
|
831 |
4.00,
|
832 |
0.00,
|
833 |
+
61.76,
|
834 |
64.00,
|
835 |
0.00,
|
836 |
5.00,
|
837 |
81.25,
|
838 |
),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
839 |
(
|
840 |
14,
|
841 |
+
54.16,
|
842 |
"GPT-4-0613",
|
843 |
"OpenAI",
|
844 |
"Proprietary",
|
|
|
846 |
86.00,
|
847 |
4.00,
|
848 |
0.00,
|
849 |
+
44.12,
|
850 |
50.00,
|
851 |
0.00,
|
852 |
0.00,
|
|
|
854 |
),
|
855 |
(
|
856 |
15,
|
857 |
+
45.18,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
858 |
"Deepseek-v1.5",
|
859 |
"Deepseek",
|
860 |
"Deepseek License",
|
|
|
862 |
61.00,
|
863 |
35.00,
|
864 |
43.50,
|
865 |
+
24.70,
|
866 |
2.00,
|
867 |
0.00,
|
868 |
7.50,
|
869 |
66.25,
|
870 |
),
|
871 |
+
(
|
872 |
+
16,
|
873 |
+
44.34,
|
874 |
+
"Gemma",
|
875 |
+
"Google",
|
876 |
+
"gemma-terms-of-use",
|
877 |
+
61.45,
|
878 |
+
60.00,
|
879 |
+
41.00,
|
880 |
+
32.00,
|
881 |
+
44.71,
|
882 |
+
46.00,
|
883 |
+
44.00,
|
884 |
+
25.00,
|
885 |
+
0.42,
|
886 |
+
),
|
887 |
(
|
888 |
17,
|
889 |
33.61,
|