Hugging Face's logo

Spaces:
amu-cai
/
Polish_Medical_Exams
Sleeping

App Files Community
michal commited on
Commit
3930665
•
1 Parent(s): c8fccfb

Upload

Browse files
Files changed (4) hide show
  1. app.py +0 -5
  2. leaderboards/ldek_en_accs.json +274 -0
  3. leaderboards/lek_en_accs.json +274 -0
  4. src/structures/ldek_structure.py +1 -1
app.py CHANGED
@@ -37,11 +37,6 @@ main = gr.Blocks(css=custom_css)
37
  with main:
38
  with gr.Row():
39
  with gr.Column():
40
- # image = gr.Image("src/images/logo.png",
41
- # show_download_button=False,
42
- # show_share_button=False,
43
- # show_fullscreen_button=False,
44
- # container=False)
45
  gr.HTML(polygon_svg)
46
  with gr.Column():
47
  gr.HTML(HEADER_TITLE)
 
37
  with main:
38
  with gr.Row():
39
  with gr.Column():
 
 
 
 
 
40
  gr.HTML(polygon_svg)
41
  with gr.Column():
42
  gr.HTML(HEADER_TITLE)
leaderboards/ldek_en_accs.json ADDED
@@ -0,0 +1,274 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "model_name": "model-meta-llama-Llama-3.1-8B-Instruct",
4
+ "2022 Spring": 49.494949494949495,
5
+ "2014 Spring": 42.857142857142854,
6
+ "2022 Fall": 45.91836734693878,
7
+ "2014 Fall": 46.1139896373057,
8
+ "2024 Spring": 47.20812182741117,
9
+ "2015 Spring": 54.3010752688172,
10
+ "2023 Spring": 46.96969696969697,
11
+ "2021 Fall": 48.98989898989899,
12
+ "2015 Fall": 47.89473684210526,
13
+ "2023 Fall": 49.743589743589745,
14
+ "2021 Spring": 46.93877551020408,
15
+ "2013 Spring": 47.42268041237113,
16
+ "2013 Fall": 46.596858638743456,
17
+ "overall_accuracy": 47.70569620253164
18
+ },
19
+ {
20
+ "model_name": "model-BioMistral-BioMistral-7B",
21
+ "2022 Spring": 22.727272727272727,
22
+ "2014 Spring": 20.408163265306122,
23
+ "2022 Fall": 32.6530612244898,
24
+ "2014 Fall": 22.279792746113987,
25
+ "2024 Spring": 29.949238578680205,
26
+ "2015 Spring": 25.806451612903224,
27
+ "2023 Spring": 27.77777777777778,
28
+ "2021 Fall": 29.292929292929294,
29
+ "2015 Fall": 29.47368421052631,
30
+ "2023 Fall": 27.692307692307693,
31
+ "2021 Spring": 22.448979591836736,
32
+ "2013 Spring": 28.865979381443296,
33
+ "2013 Fall": 28.79581151832461,
34
+ "overall_accuracy": 26.78006329113924
35
+ },
36
+ {
37
+ "model_name": "model-johnsnowlabs-JSL-MedLlama-3-8B-v2.0",
38
+ "2022 Spring": 38.88888888888889,
39
+ "2014 Spring": 34.69387755102041,
40
+ "2022 Fall": 40.30612244897959,
41
+ "2014 Fall": 37.82383419689119,
42
+ "2024 Spring": 41.62436548223351,
43
+ "2015 Spring": 40.32258064516129,
44
+ "2023 Spring": 36.868686868686865,
45
+ "2021 Fall": 44.94949494949495,
46
+ "2015 Fall": 35.26315789473684,
47
+ "2023 Fall": 37.43589743589744,
48
+ "2021 Spring": 40.816326530612244,
49
+ "2013 Spring": 41.23711340206185,
50
+ "2013 Fall": 42.40837696335078,
51
+ "overall_accuracy": 39.438291139240505
52
+ },
53
+ {
54
+ "model_name": "model-Qwen-Qwen2.5-7B-Instruct",
55
+ "2022 Spring": 46.464646464646464,
56
+ "2014 Spring": 46.42857142857143,
57
+ "2022 Fall": 46.93877551020408,
58
+ "2014 Fall": 49.740932642487046,
59
+ "2024 Spring": 53.299492385786806,
60
+ "2015 Spring": 44.086021505376344,
61
+ "2023 Spring": 48.98989898989899,
62
+ "2021 Fall": 53.03030303030303,
63
+ "2015 Fall": 44.21052631578947,
64
+ "2023 Fall": 47.179487179487175,
65
+ "2021 Spring": 51.53061224489795,
66
+ "2013 Spring": 52.0618556701031,
67
+ "2013 Fall": 44.50261780104712,
68
+ "overall_accuracy": 48.37816455696203
69
+ },
70
+ {
71
+ "model_name": "model-speakleash-Bielik-11B-v2.2-Instruct",
72
+ "2022 Spring": 39.8989898989899,
73
+ "2014 Spring": 36.734693877551024,
74
+ "2022 Fall": 38.265306122448976,
75
+ "2014 Fall": 44.04145077720207,
76
+ "2024 Spring": 50.25380710659898,
77
+ "2015 Spring": 39.247311827956985,
78
+ "2023 Spring": 38.88888888888889,
79
+ "2021 Fall": 42.92929292929293,
80
+ "2015 Fall": 42.10526315789473,
81
+ "2023 Fall": 43.58974358974359,
82
+ "2021 Spring": 48.97959183673469,
83
+ "2013 Spring": 49.48453608247423,
84
+ "2013 Fall": 42.93193717277487,
85
+ "overall_accuracy": 42.879746835443036
86
+ },
87
+ {
88
+ "model_name": "model-gpt-4o-2024-08-06",
89
+ "2022 Spring": 74.74747474747475,
90
+ "2014 Spring": 72.44897959183673,
91
+ "2022 Fall": 71.42857142857143,
92
+ "2014 Fall": 74.09326424870466,
93
+ "2024 Spring": 73.09644670050761,
94
+ "2015 Spring": 66.66666666666666,
95
+ "2023 Spring": 71.21212121212122,
96
+ "2021 Fall": 74.74747474747475,
97
+ "2015 Fall": 71.05263157894737,
98
+ "2023 Fall": 70.76923076923077,
99
+ "2021 Spring": 77.55102040816327,
100
+ "2013 Spring": 72.68041237113401,
101
+ "2013 Fall": 71.72774869109948,
102
+ "overall_accuracy": 72.50791139240506
103
+ },
104
+ {
105
+ "model_name": "model-OpenMeditron-Meditron3-8B",
106
+ "2022 Spring": 42.92929292929293,
107
+ "2014 Spring": 38.775510204081634,
108
+ "2022 Fall": 44.38775510204081,
109
+ "2014 Fall": 44.04145077720207,
110
+ "2024 Spring": 40.10152284263959,
111
+ "2015 Spring": 45.16129032258064,
112
+ "2023 Spring": 41.91919191919192,
113
+ "2021 Fall": 44.94949494949495,
114
+ "2015 Fall": 42.10526315789473,
115
+ "2023 Fall": 41.53846153846154,
116
+ "2021 Spring": 43.36734693877551,
117
+ "2013 Spring": 46.90721649484536,
118
+ "2013 Fall": 47.64397905759162,
119
+ "overall_accuracy": 43.35443037974683
120
+ },
121
+ {
122
+ "model_name": "model-gpt-4o-mini-2024-07-18",
123
+ "2022 Spring": 54.54545454545454,
124
+ "2014 Spring": 55.61224489795919,
125
+ "2022 Fall": 56.12244897959183,
126
+ "2014 Fall": 56.994818652849744,
127
+ "2024 Spring": 56.34517766497462,
128
+ "2015 Spring": 56.98924731182796,
129
+ "2023 Spring": 56.060606060606055,
130
+ "2021 Fall": 55.55555555555556,
131
+ "2015 Fall": 54.21052631578947,
132
+ "2023 Fall": 58.97435897435898,
133
+ "2021 Spring": 62.755102040816325,
134
+ "2013 Spring": 58.24742268041238,
135
+ "2013 Fall": 57.06806282722513,
136
+ "overall_accuracy": 56.88291139240506
137
+ },
138
+ {
139
+ "model_name": "model-OpenMeditron-Meditron3-70B",
140
+ "2022 Spring": 43.93939393939394,
141
+ "2014 Spring": 40.30612244897959,
142
+ "2022 Fall": 47.95918367346938,
143
+ "2014 Fall": 45.07772020725388,
144
+ "2024 Spring": 49.23857868020304,
145
+ "2015 Spring": 45.16129032258064,
146
+ "2023 Spring": 43.43434343434344,
147
+ "2021 Fall": 48.98989898989899,
148
+ "2015 Fall": 48.94736842105264,
149
+ "2023 Fall": 45.64102564102564,
150
+ "2021 Spring": 46.42857142857143,
151
+ "2013 Spring": 46.391752577319586,
152
+ "2013 Fall": 46.07329842931937,
153
+ "overall_accuracy": 45.96518987341772
154
+ },
155
+ {
156
+ "model_name": "model-aaditya-Llama3-OpenBioLLM-70B",
157
+ "2022 Spring": 44.44444444444444,
158
+ "2014 Spring": 45.40816326530612,
159
+ "2022 Fall": 46.93877551020408,
160
+ "2014 Fall": 47.15025906735752,
161
+ "2024 Spring": 48.223350253807105,
162
+ "2015 Spring": 45.16129032258064,
163
+ "2023 Spring": 45.95959595959596,
164
+ "2021 Fall": 48.484848484848484,
165
+ "2015 Fall": 46.31578947368421,
166
+ "2023 Fall": 41.02564102564102,
167
+ "2021 Spring": 45.91836734693878,
168
+ "2013 Spring": 46.391752577319586,
169
+ "2013 Fall": 45.0261780104712,
170
+ "overall_accuracy": 45.88607594936709
171
+ },
172
+ {
173
+ "model_name": "model-meta-llama-Llama-3.2-3B-Instruct",
174
+ "2022 Spring": 34.34343434343434,
175
+ "2014 Spring": 32.6530612244898,
176
+ "2022 Fall": 39.285714285714285,
177
+ "2014 Fall": 35.751295336787564,
178
+ "2024 Spring": 39.59390862944163,
179
+ "2015 Spring": 38.70967741935484,
180
+ "2023 Spring": 39.39393939393939,
181
+ "2021 Fall": 37.878787878787875,
182
+ "2015 Fall": 35.78947368421053,
183
+ "2023 Fall": 37.43589743589744,
184
+ "2021 Spring": 37.755102040816325,
185
+ "2013 Spring": 33.50515463917525,
186
+ "2013 Fall": 37.17277486910995,
187
+ "overall_accuracy": 36.86708860759494
188
+ },
189
+ {
190
+ "model_name": "model-mistralai-Mistral-Large-Instruct-2407",
191
+ "2022 Spring": 61.61616161616161,
192
+ "2014 Spring": 61.73469387755102,
193
+ "2022 Fall": 60.204081632653065,
194
+ "2014 Fall": 63.212435233160626,
195
+ "2024 Spring": 64.9746192893401,
196
+ "2015 Spring": 62.365591397849464,
197
+ "2023 Spring": 60.60606060606061,
198
+ "2021 Fall": 62.62626262626263,
199
+ "2015 Fall": 61.578947368421055,
200
+ "2023 Fall": 59.48717948717949,
201
+ "2021 Spring": 64.28571428571429,
202
+ "2013 Spring": 63.91752577319587,
203
+ "2013 Fall": 61.25654450261781,
204
+ "overall_accuracy": 62.14398734177215
205
+ },
206
+ {
207
+ "model_name": "model-meta-llama-Meta-Llama-3.1-70B-Instruct",
208
+ "2022 Spring": 65.15151515151516,
209
+ "2014 Spring": 62.755102040816325,
210
+ "2022 Fall": 62.244897959183675,
211
+ "2014 Fall": 64.24870466321244,
212
+ "2024 Spring": 59.89847715736041,
213
+ "2015 Spring": 68.27956989247312,
214
+ "2023 Spring": 64.64646464646465,
215
+ "2021 Fall": 62.121212121212125,
216
+ "2015 Fall": 63.1578947368421,
217
+ "2023 Fall": 56.92307692307692,
218
+ "2021 Spring": 67.85714285714286,
219
+ "2013 Spring": 61.34020618556701,
220
+ "2013 Fall": 62.82722513089005,
221
+ "overall_accuracy": 63.17246835443038
222
+ },
223
+ {
224
+ "model_name": "model-ProbeMedicalYonseiMAILab-medllama3-v20",
225
+ "2022 Spring": 36.868686868686865,
226
+ "2014 Spring": 34.69387755102041,
227
+ "2022 Fall": 37.755102040816325,
228
+ "2014 Fall": 35.751295336787564,
229
+ "2024 Spring": 43.14720812182741,
230
+ "2015 Spring": 43.54838709677419,
231
+ "2023 Spring": 39.8989898989899,
232
+ "2021 Fall": 35.35353535353536,
233
+ "2015 Fall": 37.89473684210527,
234
+ "2023 Fall": 34.87179487179487,
235
+ "2021 Spring": 38.775510204081634,
236
+ "2013 Spring": 41.75257731958763,
237
+ "2013 Fall": 39.79057591623037,
238
+ "overall_accuracy": 38.449367088607595
239
+ },
240
+ {
241
+ "model_name": "model-mistralai-Mistral-Small-Instruct-2409",
242
+ "2022 Spring": 37.878787878787875,
243
+ "2014 Spring": 45.91836734693878,
244
+ "2022 Fall": 45.91836734693878,
245
+ "2014 Fall": 43.005181347150256,
246
+ "2024 Spring": 45.68527918781726,
247
+ "2015 Spring": 44.623655913978496,
248
+ "2023 Spring": 40.909090909090914,
249
+ "2021 Fall": 40.4040404040404,
250
+ "2015 Fall": 44.21052631578947,
251
+ "2023 Fall": 42.05128205128205,
252
+ "2021 Spring": 46.42857142857143,
253
+ "2013 Spring": 44.329896907216494,
254
+ "2013 Fall": 43.97905759162304,
255
+ "overall_accuracy": 43.473101265822784
256
+ },
257
+ {
258
+ "model_name": "model-Qwen-Qwen2.5-72B-Instruct",
259
+ "2022 Spring": 65.15151515151516,
260
+ "2014 Spring": 61.73469387755102,
261
+ "2022 Fall": 64.79591836734694,
262
+ "2014 Fall": 66.32124352331607,
263
+ "2024 Spring": 63.45177664974619,
264
+ "2015 Spring": 62.365591397849464,
265
+ "2023 Spring": 59.09090909090909,
266
+ "2021 Fall": 65.15151515151516,
267
+ "2015 Fall": 63.1578947368421,
268
+ "2023 Fall": 58.46153846153847,
269
+ "2021 Spring": 69.38775510204081,
270
+ "2013 Spring": 64.94845360824742,
271
+ "2013 Fall": 59.16230366492147,
272
+ "overall_accuracy": 63.33069620253164
273
+ }
274
+ ]
leaderboards/lek_en_accs.json ADDED
@@ -0,0 +1,274 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "model_name": "model-meta-llama-Llama-3.1-8B-Instruct",
4
+ "2022 Spring": 63.775510204081634,
5
+ "2014 Spring": 60.51282051282051,
6
+ "2022 Fall": 63.775510204081634,
7
+ "2014 Fall": 64.58333333333334,
8
+ "2024 Spring": 74.61928934010153,
9
+ "2015 Spring": 61.111111111111114,
10
+ "2023 Spring": 63.4020618556701,
11
+ "2021 Fall": 61.855670103092784,
12
+ "2015 Fall": 62.63157894736842,
13
+ "2023 Fall": 70.61855670103093,
14
+ "2021 Spring": 70.0,
15
+ "2013 Spring": 61.979166666666664,
16
+ "2013 Fall": 61.78010471204188,
17
+ "overall_accuracy": 64.68960063266113
18
+ },
19
+ {
20
+ "model_name": "model-BioMistral-BioMistral-7B",
21
+ "2022 Spring": 32.142857142857146,
22
+ "2014 Spring": 32.30769230769231,
23
+ "2022 Fall": 32.142857142857146,
24
+ "2014 Fall": 27.604166666666668,
25
+ "2024 Spring": 39.59390862944163,
26
+ "2015 Spring": 29.292929292929294,
27
+ "2023 Spring": 32.98969072164948,
28
+ "2021 Fall": 32.47422680412371,
29
+ "2015 Fall": 32.631578947368425,
30
+ "2023 Fall": 31.958762886597935,
31
+ "2021 Spring": 39.0,
32
+ "2013 Spring": 36.97916666666667,
33
+ "2013 Fall": 26.17801047120419,
34
+ "overall_accuracy": 32.74021352313167
35
+ },
36
+ {
37
+ "model_name": "model-johnsnowlabs-JSL-MedLlama-3-8B-v2.0",
38
+ "2022 Spring": 54.59183673469388,
39
+ "2014 Spring": 49.23076923076923,
40
+ "2022 Fall": 51.02040816326531,
41
+ "2014 Fall": 54.166666666666664,
42
+ "2024 Spring": 62.944162436548226,
43
+ "2015 Spring": 51.515151515151516,
44
+ "2023 Spring": 53.09278350515464,
45
+ "2021 Fall": 56.18556701030928,
46
+ "2015 Fall": 53.1578947368421,
47
+ "2023 Fall": 56.18556701030928,
48
+ "2021 Spring": 55.00000000000001,
49
+ "2013 Spring": 48.95833333333333,
50
+ "2013 Fall": 57.06806282722513,
51
+ "overall_accuracy": 54.092526690391466
52
+ },
53
+ {
54
+ "model_name": "model-Qwen-Qwen2.5-7B-Instruct",
55
+ "2022 Spring": 64.79591836734694,
56
+ "2014 Spring": 62.05128205128205,
57
+ "2022 Fall": 66.3265306122449,
58
+ "2014 Fall": 72.91666666666666,
59
+ "2024 Spring": 79.18781725888326,
60
+ "2015 Spring": 68.18181818181817,
61
+ "2023 Spring": 68.55670103092784,
62
+ "2021 Fall": 63.91752577319587,
63
+ "2015 Fall": 68.94736842105263,
64
+ "2023 Fall": 71.64948453608247,
65
+ "2021 Spring": 65.5,
66
+ "2013 Spring": 64.0625,
67
+ "2013 Fall": 64.3979057591623,
68
+ "overall_accuracy": 67.73428232502965
69
+ },
70
+ {
71
+ "model_name": "model-speakleash-Bielik-11B-v2.2-Instruct",
72
+ "2022 Spring": 50.51020408163265,
73
+ "2014 Spring": 53.84615384615385,
74
+ "2022 Fall": 59.693877551020414,
75
+ "2014 Fall": 58.333333333333336,
76
+ "2024 Spring": 62.43654822335025,
77
+ "2015 Spring": 58.58585858585859,
78
+ "2023 Spring": 56.70103092783505,
79
+ "2021 Fall": 55.154639175257735,
80
+ "2015 Fall": 56.84210526315789,
81
+ "2023 Fall": 57.73195876288659,
82
+ "2021 Spring": 61.5,
83
+ "2013 Spring": 55.729166666666664,
84
+ "2013 Fall": 53.403141361256544,
85
+ "overall_accuracy": 56.97904310003954
86
+ },
87
+ {
88
+ "model_name": "model-gpt-4o-2024-08-06",
89
+ "2022 Spring": 87.75510204081633,
90
+ "2014 Spring": 88.71794871794872,
91
+ "2022 Fall": 87.75510204081633,
92
+ "2014 Fall": 88.02083333333334,
93
+ "2024 Spring": 94.41624365482234,
94
+ "2015 Spring": 89.39393939393939,
95
+ "2023 Spring": 91.23711340206185,
96
+ "2021 Fall": 85.56701030927834,
97
+ "2015 Fall": 87.36842105263159,
98
+ "2023 Fall": 90.72164948453609,
99
+ "2021 Spring": 91.5,
100
+ "2013 Spring": 83.85416666666666,
101
+ "2013 Fall": 86.38743455497382,
102
+ "overall_accuracy": 88.69118228548834
103
+ },
104
+ {
105
+ "model_name": "model-OpenMeditron-Meditron3-8B",
106
+ "2022 Spring": 56.63265306122449,
107
+ "2014 Spring": 57.43589743589743,
108
+ "2022 Fall": 59.183673469387756,
109
+ "2014 Fall": 64.58333333333334,
110
+ "2024 Spring": 65.98984771573603,
111
+ "2015 Spring": 58.58585858585859,
112
+ "2023 Spring": 64.43298969072166,
113
+ "2021 Fall": 59.27835051546392,
114
+ "2015 Fall": 59.473684210526315,
115
+ "2023 Fall": 61.34020618556701,
116
+ "2021 Spring": 61.5,
117
+ "2013 Spring": 57.291666666666664,
118
+ "2013 Fall": 58.63874345549738,
119
+ "overall_accuracy": 60.340055357848954
120
+ },
121
+ {
122
+ "model_name": "model-mistralai-Mistral-Large-Instruct-2407",
123
+ "2022 Spring": 84.18367346938776,
124
+ "2014 Spring": 80.0,
125
+ "2022 Fall": 79.08163265306123,
126
+ "2014 Fall": 82.29166666666666,
127
+ "2024 Spring": 91.37055837563452,
128
+ "2015 Spring": 80.8080808080808,
129
+ "2023 Spring": 82.4742268041237,
130
+ "2021 Fall": 79.38144329896907,
131
+ "2015 Fall": 78.42105263157895,
132
+ "2023 Fall": 89.69072164948454,
133
+ "2021 Spring": 88.5,
134
+ "2013 Spring": 77.08333333333334,
135
+ "2013 Fall": 77.4869109947644,
136
+ "overall_accuracy": 82.40411229735074
137
+ },
138
+ {
139
+ "model_name": "model-gpt-4o-mini-2024-07-18",
140
+ "2022 Spring": 70.40816326530613,
141
+ "2014 Spring": 75.8974358974359,
142
+ "2022 Fall": 75.0,
143
+ "2014 Fall": 78.64583333333334,
144
+ "2024 Spring": 85.27918781725889,
145
+ "2015 Spring": 72.72727272727273,
146
+ "2023 Spring": 76.80412371134021,
147
+ "2021 Fall": 72.68041237113401,
148
+ "2015 Fall": 72.63157894736842,
149
+ "2023 Fall": 84.5360824742268,
150
+ "2021 Spring": 77.0,
151
+ "2013 Spring": 69.79166666666666,
152
+ "2013 Fall": 75.39267015706807,
153
+ "overall_accuracy": 75.91933570581257
154
+ },
155
+ {
156
+ "model_name": "model-OpenMeditron-Meditron3-70B",
157
+ "2022 Spring": 67.85714285714286,
158
+ "2014 Spring": 67.17948717948717,
159
+ "2022 Fall": 60.204081632653065,
160
+ "2014 Fall": 65.10416666666666,
161
+ "2024 Spring": 81.21827411167513,
162
+ "2015 Spring": 64.14141414141415,
163
+ "2023 Spring": 70.61855670103093,
164
+ "2021 Fall": 60.30927835051546,
165
+ "2015 Fall": 62.63157894736842,
166
+ "2023 Fall": 72.68041237113401,
167
+ "2021 Spring": 67.5,
168
+ "2013 Spring": 65.10416666666666,
169
+ "2013 Fall": 62.82722513089005,
170
+ "overall_accuracy": 66.74574930802689
171
+ },
172
+ {
173
+ "model_name": "model-aaditya-Llama3-OpenBioLLM-70B",
174
+ "2022 Spring": 62.244897959183675,
175
+ "2014 Spring": 63.07692307692307,
176
+ "2022 Fall": 59.183673469387756,
177
+ "2014 Fall": 68.22916666666666,
178
+ "2024 Spring": 72.58883248730965,
179
+ "2015 Spring": 67.67676767676768,
180
+ "2023 Spring": 68.55670103092784,
181
+ "2021 Fall": 65.97938144329896,
182
+ "2015 Fall": 59.473684210526315,
183
+ "2023 Fall": 68.55670103092784,
184
+ "2021 Spring": 72.5,
185
+ "2013 Spring": 64.58333333333334,
186
+ "2013 Fall": 63.87434554973822,
187
+ "overall_accuracy": 65.91538157374455
188
+ },
189
+ {
190
+ "model_name": "model-meta-llama-Llama-3.2-3B-Instruct",
191
+ "2022 Spring": 51.02040816326531,
192
+ "2014 Spring": 46.666666666666664,
193
+ "2022 Fall": 51.02040816326531,
194
+ "2014 Fall": 53.645833333333336,
195
+ "2024 Spring": 56.34517766497462,
196
+ "2015 Spring": 48.98989898989899,
197
+ "2023 Spring": 52.57731958762887,
198
+ "2021 Fall": 55.154639175257735,
199
+ "2015 Fall": 52.63157894736842,
200
+ "2023 Fall": 55.670103092783506,
201
+ "2021 Spring": 56.99999999999999,
202
+ "2013 Spring": 46.35416666666667,
203
+ "2013 Fall": 49.73821989528796,
204
+ "overall_accuracy": 52.07591933570581
205
+ },
206
+ {
207
+ "model_name": "model-meta-llama-Meta-Llama-3.1-70B-Instruct",
208
+ "2022 Spring": 78.06122448979592,
209
+ "2014 Spring": 79.48717948717949,
210
+ "2022 Fall": 82.6530612244898,
211
+ "2014 Fall": 86.45833333333334,
212
+ "2024 Spring": 89.84771573604061,
213
+ "2015 Spring": 86.36363636363636,
214
+ "2023 Spring": 82.4742268041237,
215
+ "2021 Fall": 77.83505154639175,
216
+ "2015 Fall": 82.10526315789474,
217
+ "2023 Fall": 84.02061855670104,
218
+ "2021 Spring": 86.5,
219
+ "2013 Spring": 79.16666666666666,
220
+ "2013 Fall": 78.53403141361257,
221
+ "overall_accuracy": 82.60181890075128
222
+ },
223
+ {
224
+ "model_name": "model-ProbeMedicalYonseiMAILab-medllama3-v20",
225
+ "2022 Spring": 49.48979591836735,
226
+ "2014 Spring": 49.743589743589745,
227
+ "2022 Fall": 57.14285714285714,
228
+ "2014 Fall": 52.604166666666664,
229
+ "2024 Spring": 56.852791878172596,
230
+ "2015 Spring": 45.45454545454545,
231
+ "2023 Spring": 55.154639175257735,
232
+ "2021 Fall": 52.0618556701031,
233
+ "2015 Fall": 47.89473684210526,
234
+ "2023 Fall": 57.73195876288659,
235
+ "2021 Spring": 53.5,
236
+ "2013 Spring": 48.95833333333333,
237
+ "2013 Fall": 52.879581151832454,
238
+ "overall_accuracy": 52.27362593910636
239
+ },
240
+ {
241
+ "model_name": "model-mistralai-Mistral-Small-Instruct-2409",
242
+ "2022 Spring": 63.775510204081634,
243
+ "2014 Spring": 61.53846153846154,
244
+ "2022 Fall": 63.775510204081634,
245
+ "2014 Fall": 64.0625,
246
+ "2024 Spring": 73.09644670050761,
247
+ "2015 Spring": 58.58585858585859,
248
+ "2023 Spring": 67.0103092783505,
249
+ "2021 Fall": 63.4020618556701,
250
+ "2015 Fall": 58.94736842105262,
251
+ "2023 Fall": 68.55670103092784,
252
+ "2021 Spring": 65.0,
253
+ "2013 Spring": 58.333333333333336,
254
+ "2013 Fall": 61.78010471204188,
255
+ "overall_accuracy": 63.70106761565836
256
+ },
257
+ {
258
+ "model_name": "model-Qwen-Qwen2.5-72B-Instruct",
259
+ "2022 Spring": 81.63265306122449,
260
+ "2014 Spring": 80.0,
261
+ "2022 Fall": 81.63265306122449,
262
+ "2014 Fall": 84.375,
263
+ "2024 Spring": 87.30964467005076,
264
+ "2015 Spring": 81.81818181818183,
265
+ "2023 Spring": 81.95876288659794,
266
+ "2021 Fall": 81.44329896907216,
267
+ "2015 Fall": 76.84210526315789,
268
+ "2023 Fall": 85.56701030927834,
269
+ "2021 Spring": 83.0,
270
+ "2013 Spring": 76.5625,
271
+ "2013 Fall": 82.72251308900523,
272
+ "overall_accuracy": 81.9296164491894
273
+ }
274
+ ]
src/structures/ldek_structure.py CHANGED
@@ -21,7 +21,7 @@ def generate_ORDER_LIST_LDEK_and_data_types(json_data):
21
  data_types = ["markdown", "number"]
22
 
23
  for key in json_data.keys():
24
- if key not in ["model_name", "overall_accuracy"]:
25
  ORDER_LIST_LDEK.append(key)
26
  data_types.append("number")
27
  ORDER_LIST_LDEK[2:] = sorted(ORDER_LIST_LDEK[2:])
 
21
  data_types = ["markdown", "number"]
22
 
23
  for key in json_data.keys():
24
+ if key not in ["model_name"]:
25
  ORDER_LIST_LDEK.append(key)
26
  data_types.append("number")
27
  ORDER_LIST_LDEK[2:] = sorted(ORDER_LIST_LDEK[2:])