Spaces:
Sleeping
Sleeping
michal
commited on
Commit
·
c9001ec
1
Parent(s):
c5e8842
Upload
Browse files
leaderboards/r_ldek_report_scores.json
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[
|
2 |
+
{
|
3 |
+
"model_name": "model-epfl-llm-meditron-7b",
|
4 |
+
"2008_jesień": 22.727272727272727,
|
5 |
+
"2012_wiosna": 23.711340206185564,
|
6 |
+
"2010_jesień": 19.597989949748744,
|
7 |
+
"2009_wiosna": 16.161616161616163,
|
8 |
+
"2010_wiosna": 16.0,
|
9 |
+
"2011_wiosna": 19.696969696969695,
|
10 |
+
"2012_jesień": 23.46938775510204,
|
11 |
+
"2011_jesień": 24.747474747474747,
|
12 |
+
"2009_jesień": 18.5,
|
13 |
+
"2022_wiosna": 22.727272727272727,
|
14 |
+
"2014_wiosna": 17.346938775510203,
|
15 |
+
"2022_jesien": 22.75132275132275,
|
16 |
+
"2014_jesien": 16.580310880829018,
|
17 |
+
"2024_wiosna": 22.916666666666664,
|
18 |
+
"2015_wiosna": 19.35483870967742,
|
19 |
+
"2023_wiosna": 21.025641025641026,
|
20 |
+
"2021_jesien": 24.873096446700508,
|
21 |
+
"2015_jesien": 18.947368421052634,
|
22 |
+
"2023_jesien": 22.164948453608247,
|
23 |
+
"2021_wiosna": 24.489795918367346,
|
24 |
+
"2013_wiosna": 18.04123711340206,
|
25 |
+
"2013_jesien": 17.894736842105264,
|
26 |
+
"overall_accuracy": 20.624563038918666
|
27 |
+
},
|
28 |
+
{
|
29 |
+
"model_name": "model-speakleash-Bielik-11B-v2.2-Instruct",
|
30 |
+
"2008_jesień": 45.45454545454545,
|
31 |
+
"2012_wiosna": 46.391752577319586,
|
32 |
+
"2010_jesień": 52.26130653266332,
|
33 |
+
"2009_wiosna": 51.515151515151516,
|
34 |
+
"2010_wiosna": 48.5,
|
35 |
+
"2011_wiosna": 41.91919191919192,
|
36 |
+
"2012_jesień": 57.6530612244898,
|
37 |
+
"2011_jesień": 45.95959595959596,
|
38 |
+
"2009_jesień": 49.5,
|
39 |
+
"2022_wiosna": 40.4040404040404,
|
40 |
+
"2014_wiosna": 34.69387755102041,
|
41 |
+
"2022_jesien": 43.91534391534391,
|
42 |
+
"2014_jesien": 44.04145077720207,
|
43 |
+
"2024_wiosna": 51.041666666666664,
|
44 |
+
"2015_wiosna": 43.54838709677419,
|
45 |
+
"2023_wiosna": 42.05128205128205,
|
46 |
+
"2021_jesien": 44.16243654822335,
|
47 |
+
"2015_jesien": 40.0,
|
48 |
+
"2023_jesien": 38.144329896907216,
|
49 |
+
"2021_wiosna": 50.0,
|
50 |
+
"2013_wiosna": 47.93814432989691,
|
51 |
+
"2013_jesien": 43.15789473684211,
|
52 |
+
"overall_accuracy": 45.583780004660916
|
53 |
+
},
|
54 |
+
{
|
55 |
+
"model_name": "model-meta-llama-Meta-Llama-3.1-8B-Instruct",
|
56 |
+
"2008_jesień": 41.41414141414141,
|
57 |
+
"2012_wiosna": 41.23711340206185,
|
58 |
+
"2010_jesień": 43.21608040201005,
|
59 |
+
"2009_wiosna": 47.97979797979798,
|
60 |
+
"2010_wiosna": 46.5,
|
61 |
+
"2011_wiosna": 33.33333333333333,
|
62 |
+
"2012_jesień": 48.46938775510204,
|
63 |
+
"2011_jesień": 41.41414141414141,
|
64 |
+
"2009_jesień": 54.50000000000001,
|
65 |
+
"2022_wiosna": 40.4040404040404,
|
66 |
+
"2014_wiosna": 33.6734693877551,
|
67 |
+
"2022_jesien": 41.7989417989418,
|
68 |
+
"2014_jesien": 42.487046632124354,
|
69 |
+
"2024_wiosna": 39.0625,
|
70 |
+
"2015_wiosna": 42.473118279569896,
|
71 |
+
"2023_wiosna": 45.64102564102564,
|
72 |
+
"2021_jesien": 40.609137055837564,
|
73 |
+
"2015_jesien": 41.578947368421055,
|
74 |
+
"2023_jesien": 42.2680412371134,
|
75 |
+
"2021_wiosna": 41.3265306122449,
|
76 |
+
"2013_wiosna": 41.23711340206185,
|
77 |
+
"2013_jesien": 45.78947368421053,
|
78 |
+
"overall_accuracy": 42.57748776508972
|
79 |
+
},
|
80 |
+
{
|
81 |
+
"model_name": "model-meta-llama-Meta-Llama-3.1-70B-Instruct",
|
82 |
+
"2008_jesień": 62.121212121212125,
|
83 |
+
"2012_wiosna": 61.34020618556701,
|
84 |
+
"2010_jesień": 62.8140703517588,
|
85 |
+
"2009_wiosna": 69.1919191919192,
|
86 |
+
"2010_wiosna": 73.0,
|
87 |
+
"2011_wiosna": 59.59595959595959,
|
88 |
+
"2012_jesień": 69.89795918367348,
|
89 |
+
"2011_jesień": 65.65656565656566,
|
90 |
+
"2009_jesień": 68.0,
|
91 |
+
"2022_wiosna": 68.68686868686868,
|
92 |
+
"2014_wiosna": 54.59183673469388,
|
93 |
+
"2022_jesien": 64.02116402116403,
|
94 |
+
"2014_jesien": 60.62176165803109,
|
95 |
+
"2024_wiosna": 61.979166666666664,
|
96 |
+
"2015_wiosna": 60.215053763440864,
|
97 |
+
"2023_wiosna": 63.07692307692307,
|
98 |
+
"2021_jesien": 63.95939086294417,
|
99 |
+
"2015_jesien": 58.42105263157895,
|
100 |
+
"2023_jesien": 61.34020618556701,
|
101 |
+
"2021_wiosna": 68.36734693877551,
|
102 |
+
"2013_wiosna": 62.371134020618555,
|
103 |
+
"2013_jesien": 57.36842105263158,
|
104 |
+
"overall_accuracy": 63.528315078070385
|
105 |
+
},
|
106 |
+
{
|
107 |
+
"model_name": "model-mistralai-Mistral-Large-Instruct-2407",
|
108 |
+
"2008_jesień": 61.61616161616161,
|
109 |
+
"2012_wiosna": 54.123711340206185,
|
110 |
+
"2010_jesień": 62.311557788944725,
|
111 |
+
"2009_wiosna": 67.67676767676768,
|
112 |
+
"2010_wiosna": 69.5,
|
113 |
+
"2011_wiosna": 56.56565656565656,
|
114 |
+
"2012_jesień": 66.3265306122449,
|
115 |
+
"2011_jesień": 58.080808080808076,
|
116 |
+
"2009_jesień": 63.0,
|
117 |
+
"2022_wiosna": 54.54545454545454,
|
118 |
+
"2014_wiosna": 54.08163265306123,
|
119 |
+
"2022_jesien": 57.14285714285714,
|
120 |
+
"2014_jesien": 56.476683937823836,
|
121 |
+
"2024_wiosna": 56.770833333333336,
|
122 |
+
"2015_wiosna": 58.602150537634415,
|
123 |
+
"2023_wiosna": 57.948717948717956,
|
124 |
+
"2021_jesien": 55.83756345177665,
|
125 |
+
"2015_jesien": 60.0,
|
126 |
+
"2023_jesien": 48.45360824742268,
|
127 |
+
"2021_wiosna": 57.14285714285714,
|
128 |
+
"2013_wiosna": 61.34020618556701,
|
129 |
+
"2013_jesien": 55.78947368421052,
|
130 |
+
"overall_accuracy": 58.820787695175945
|
131 |
+
}
|
132 |
+
]
|
leaderboards/r_lek_report_scores.json
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[
|
2 |
+
{
|
3 |
+
"model_name": "model-epfl-llm-meditron-7b",
|
4 |
+
"2011_wiosna": 19.696969696969695,
|
5 |
+
"2010_wiosna": 26.5,
|
6 |
+
"2008_jesień": 23.85786802030457,
|
7 |
+
"2011_jesień": 23.232323232323232,
|
8 |
+
"2009_jesień": 22.33502538071066,
|
9 |
+
"2010_jesień": 19.696969696969695,
|
10 |
+
"2012_wiosna": 24.5,
|
11 |
+
"2009_wiosna": 24.120603015075375,
|
12 |
+
"2012_jesień": 19.897959183673468,
|
13 |
+
"2022_wiosna": 19.3717277486911,
|
14 |
+
"2014_wiosna": 20.0,
|
15 |
+
"2022_jesien": 28.654970760233915,
|
16 |
+
"2014_jesien": 23.958333333333336,
|
17 |
+
"2024_wiosna": 23.809523809523807,
|
18 |
+
"2015_wiosna": 18.781725888324875,
|
19 |
+
"2023_wiosna": 20.46783625730994,
|
20 |
+
"2021_jesien": 19.170984455958546,
|
21 |
+
"2015_jesien": 22.22222222222222,
|
22 |
+
"2023_jesien": 23.4375,
|
23 |
+
"2021_wiosna": 15.656565656565657,
|
24 |
+
"2013_wiosna": 16.145833333333336,
|
25 |
+
"2013_jesien": 21.465968586387437,
|
26 |
+
"overall_accuracy": 21.64338148235851
|
27 |
+
},
|
28 |
+
{
|
29 |
+
"model_name": "model-speakleash-Bielik-11B-v2.2-Instruct",
|
30 |
+
"2011_wiosna": 59.59595959595959,
|
31 |
+
"2010_wiosna": 62.0,
|
32 |
+
"2008_jesień": 59.89847715736041,
|
33 |
+
"2011_jesień": 60.1010101010101,
|
34 |
+
"2009_jesień": 58.37563451776649,
|
35 |
+
"2010_jesień": 63.13131313131313,
|
36 |
+
"2012_wiosna": 61.0,
|
37 |
+
"2009_wiosna": 63.31658291457286,
|
38 |
+
"2012_jesień": 63.26530612244898,
|
39 |
+
"2022_wiosna": 58.63874345549738,
|
40 |
+
"2014_wiosna": 57.948717948717956,
|
41 |
+
"2022_jesien": 69.00584795321637,
|
42 |
+
"2014_jesien": 63.541666666666664,
|
43 |
+
"2024_wiosna": 66.07142857142857,
|
44 |
+
"2015_wiosna": 59.390862944162436,
|
45 |
+
"2023_wiosna": 66.08187134502924,
|
46 |
+
"2021_jesien": 62.17616580310881,
|
47 |
+
"2015_jesien": 61.904761904761905,
|
48 |
+
"2023_jesien": 63.541666666666664,
|
49 |
+
"2021_wiosna": 65.15151515151516,
|
50 |
+
"2013_wiosna": 61.458333333333336,
|
51 |
+
"2013_jesien": 61.78010471204188,
|
52 |
+
"overall_accuracy": 62.064882784750175
|
53 |
+
},
|
54 |
+
{
|
55 |
+
"model_name": "model-meta-llama-Meta-Llama-3.1-8B-Instruct",
|
56 |
+
"2011_wiosna": 52.02020202020202,
|
57 |
+
"2010_wiosna": 53.5,
|
58 |
+
"2008_jesień": 50.25380710659898,
|
59 |
+
"2011_jesień": 57.07070707070707,
|
60 |
+
"2009_jesień": 45.68527918781726,
|
61 |
+
"2010_jesień": 43.93939393939394,
|
62 |
+
"2012_wiosna": 54.0,
|
63 |
+
"2009_wiosna": 45.22613065326633,
|
64 |
+
"2012_jesień": 53.06122448979592,
|
65 |
+
"2022_wiosna": 45.54973821989529,
|
66 |
+
"2014_wiosna": 52.307692307692314,
|
67 |
+
"2022_jesien": 53.80116959064327,
|
68 |
+
"2014_jesien": 54.166666666666664,
|
69 |
+
"2024_wiosna": 57.14285714285714,
|
70 |
+
"2015_wiosna": 54.82233502538071,
|
71 |
+
"2023_wiosna": 53.216374269005854,
|
72 |
+
"2021_jesien": 51.813471502590666,
|
73 |
+
"2015_jesien": 44.44444444444444,
|
74 |
+
"2023_jesien": 50.0,
|
75 |
+
"2021_wiosna": 47.97979797979798,
|
76 |
+
"2013_wiosna": 54.6875,
|
77 |
+
"2013_jesien": 51.832460732984295,
|
78 |
+
"overall_accuracy": 51.148472649775044
|
79 |
+
},
|
80 |
+
{
|
81 |
+
"model_name": "model-meta-llama-Meta-Llama-3.1-70B-Instruct",
|
82 |
+
"2011_wiosna": 77.27272727272727,
|
83 |
+
"2010_wiosna": 82.0,
|
84 |
+
"2008_jesień": 79.69543147208121,
|
85 |
+
"2011_jesień": 83.33333333333334,
|
86 |
+
"2009_jesień": 81.21827411167513,
|
87 |
+
"2010_jesień": 80.8080808080808,
|
88 |
+
"2012_wiosna": 80.0,
|
89 |
+
"2009_wiosna": 72.8643216080402,
|
90 |
+
"2012_jesień": 81.12244897959184,
|
91 |
+
"2022_wiosna": 79.05759162303664,
|
92 |
+
"2014_wiosna": 76.41025641025641,
|
93 |
+
"2022_jesien": 82.45614035087719,
|
94 |
+
"2014_jesien": 87.5,
|
95 |
+
"2024_wiosna": 86.90476190476191,
|
96 |
+
"2015_wiosna": 77.66497461928934,
|
97 |
+
"2023_wiosna": 80.7017543859649,
|
98 |
+
"2021_jesien": 81.34715025906736,
|
99 |
+
"2015_jesien": 78.83597883597884,
|
100 |
+
"2023_jesien": 83.33333333333334,
|
101 |
+
"2021_wiosna": 84.84848484848484,
|
102 |
+
"2013_wiosna": 77.60416666666666,
|
103 |
+
"2013_jesien": 78.01047120418848,
|
104 |
+
"overall_accuracy": 80.53516457494672
|
105 |
+
},
|
106 |
+
{
|
107 |
+
"model_name": "model-mistralai-Mistral-Large-Instruct-2407",
|
108 |
+
"2011_wiosna": 76.26262626262627,
|
109 |
+
"2010_wiosna": 76.5,
|
110 |
+
"2008_jesień": 76.6497461928934,
|
111 |
+
"2011_jesień": 79.29292929292929,
|
112 |
+
"2009_jesień": 71.57360406091371,
|
113 |
+
"2010_jesień": 74.24242424242425,
|
114 |
+
"2012_wiosna": 77.5,
|
115 |
+
"2009_wiosna": 72.8643216080402,
|
116 |
+
"2012_jesień": 76.53061224489795,
|
117 |
+
"2022_wiosna": 73.29842931937172,
|
118 |
+
"2014_wiosna": 80.0,
|
119 |
+
"2022_jesien": 77.19298245614034,
|
120 |
+
"2014_jesien": 76.5625,
|
121 |
+
"2024_wiosna": 87.5,
|
122 |
+
"2015_wiosna": 75.1269035532995,
|
123 |
+
"2023_wiosna": 80.11695906432749,
|
124 |
+
"2021_jesien": 76.16580310880829,
|
125 |
+
"2015_jesien": 71.42857142857143,
|
126 |
+
"2023_jesien": 76.5625,
|
127 |
+
"2021_wiosna": 79.29292929292929,
|
128 |
+
"2013_wiosna": 71.875,
|
129 |
+
"2013_jesien": 75.91623036649214,
|
130 |
+
"overall_accuracy": 76.39119109637699
|
131 |
+
}
|
132 |
+
]
|
src/structures/all_structure.py
CHANGED
@@ -29,9 +29,9 @@ def load_json_data(file_path):
|
|
29 |
return ALL_ACCS
|
30 |
|
31 |
file_paths = [
|
32 |
-
str(abs_path / "leaderboards/
|
33 |
-
str(abs_path / "leaderboards/
|
34 |
-
str(abs_path / "leaderboards/
|
35 |
]
|
36 |
|
37 |
model_data = {}
|
|
|
29 |
return ALL_ACCS
|
30 |
|
31 |
file_paths = [
|
32 |
+
str(abs_path / "leaderboards/pes_accuracy.json"),
|
33 |
+
str(abs_path / "leaderboards/r_ldek_report_scores.json"),
|
34 |
+
str(abs_path / "leaderboards/r_lek_report_scores.json"),
|
35 |
]
|
36 |
|
37 |
model_data = {}
|