wzxii commited on
Commit
80a180f
1 Parent(s): 7b2f4fc

Upload 3 files

Browse files
Files changed (3) hide show
  1. button.js +62 -0
  2. chart.js +408 -0
  3. table.js +274 -0
button.js ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ var btn_evalTable = document.getElementById("btn_evalTable");
2
+ var btn_about = document.getElementById("btn_about");
3
+ var btn_plot = document.getElementById("btn_plot");
4
+ var btn_submit = document.getElementById("btn_submit");
5
+ var sec_evalTable = document.getElementById("sec_evalTable");
6
+ var sec_about = document.getElementById("sec_about");
7
+ var sec_plot = document.getElementById("sec_plot")
8
+ var sec_submit = document.getElementById("sec_submit");
9
+
10
+
11
+ btn_evalTable.addEventListener("click", function() {
12
+ sec_evalTable.style.display = "block";
13
+ sec_plot.style.display = "none";
14
+ sec_about.style.display = "none";
15
+ sec_submit.style.display = "none";
16
+
17
+ btn_evalTable.style.color = "#000000";
18
+ btn_plot.style.color = "#7c7b7b";
19
+ btn_about.style.color = "#7c7b7b";
20
+ btn_submit.style.color = "#7c7b7b";
21
+
22
+ });
23
+ btn_about.addEventListener("click", function() {
24
+ sec_evalTable.style.display = "none";
25
+ sec_plot.style.display = "none";
26
+ sec_about.style.display = "block";
27
+ sec_submit.style.display = "none";
28
+
29
+ btn_evalTable.style.color = "#7c7b7b";
30
+ btn_plot.style.color = "#7c7b7b";
31
+ btn_about.style.color = "#000000";
32
+ btn_submit.style.color = "#7c7b7b";
33
+
34
+ });
35
+
36
+ btn_plot.addEventListener("click", function() {
37
+ sec_evalTable.style.display = "none";
38
+ sec_plot.style.display = "block";
39
+ sec_about.style.display = "none";
40
+ sec_submit.style.display = "none";
41
+
42
+ btn_evalTable.style.color = "#7c7b7b";
43
+ btn_plot.style.color = "#000000";
44
+ btn_about.style.color = "#7c7b7b";
45
+ btn_submit.style.color = "#7c7b7b";
46
+
47
+ });
48
+
49
+ btn_submit.addEventListener("click", function() {
50
+ sec_evalTable.style.display = "none";
51
+ sec_plot.style.display = "none";
52
+ sec_about.style.display = "none";
53
+ sec_submit.style.display = "block";
54
+
55
+ btn_evalTable.style.color = "#7c7b7b";
56
+ btn_plot.style.color = "#7c7b7b";
57
+ btn_about.style.color = "#7c7b7b";
58
+ btn_submit.style.color = "#000000";
59
+
60
+ });
61
+
62
+
chart.js ADDED
@@ -0,0 +1,408 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ var Chart = echarts.init(document.getElementById('sec_plot__chart1'));
2
+ var Chart2 = echarts.init(document.getElementById('sec_plot__chart2'));
3
+ var dataset = [
4
+ {
5
+ name: 'GPT-3.5-turbo-1103',
6
+ MGI: 0.2326,
7
+ temp0_HumanEval: 0.7439,
8
+ temp0_HumanEval_ET: 0.5915,
9
+ temp0_8_HumanEval: 0.7248,
10
+ temp0_8_HumanEval_ET: 0.5778,
11
+ },
12
+ {
13
+ name: 'GPT-3.5-turbo-0613',
14
+ MGI: 0.1314,
15
+ temp0_HumanEval: 0.6585,
16
+ temp0_HumanEval_ET: 0.5122,
17
+ temp0_8_HumanEval: 0.612,
18
+ temp0_8_HumanEval_ET: 0.4776,
19
+ },
20
+ {
21
+ name: 'CodeGeeX4-All-9B',
22
+ MGI: 0.5182,
23
+ temp0_HumanEval: 0.7805,
24
+ temp0_HumanEval_ET: 0.622,
25
+ temp0_8_HumanEval: 0.7606,
26
+ temp0_8_HumanEval_ET: 0.6078,
27
+ },
28
+ {
29
+ name: 'Deepseek-Coder-6.7b-base',
30
+ MGI: 0.0602,
31
+ temp0_HumanEval: 0.4817,
32
+ temp0_HumanEval_ET: 0.3963,
33
+ temp0_8_HumanEval: 0.3388,
34
+ temp0_8_HumanEval_ET: 0.2798,
35
+ },
36
+ {
37
+ name: 'Deepseek-Coder-6.7b-instruct',
38
+ MGI: 0.0893,
39
+ temp0_HumanEval: 0.7134,
40
+ temp0_HumanEval_ET: 0.628,
41
+ temp0_8_HumanEval: 0.6494,
42
+ temp0_8_HumanEval_ET: 0.5673,
43
+ },
44
+ {
45
+ name: 'CodeLlama-7b',
46
+ MGI: 0.0417,
47
+ temp0_HumanEval: 0.311,
48
+ temp0_HumanEval_ET: 0.2378,
49
+ temp0_8_HumanEval: 0.2211,
50
+ temp0_8_HumanEval_ET: 0.1832,
51
+ },
52
+ {
53
+ name: 'CodeLlama-7b-python',
54
+ MGI: 0.0525,
55
+ temp0_HumanEval: 0.3841,
56
+ temp0_HumanEval_ET: 0.3354,
57
+ temp0_8_HumanEval: 0.3062,
58
+ temp0_8_HumanEval_ET: 0.2568,
59
+ },
60
+ {
61
+ name: 'CodeLlama-7b-instruct',
62
+ MGI: 0.0908,
63
+ temp0_HumanEval: 0.3537,
64
+ temp0_HumanEval_ET: 0.2988,
65
+ temp0_8_HumanEval: 0.2876,
66
+ temp0_8_HumanEval_ET: 0.2455,
67
+ },
68
+ {
69
+ name: 'CodeGemma-2b',
70
+ MGI: 0.0112,
71
+ temp0_HumanEval: 0.3537,
72
+ temp0_HumanEval_ET: 0.2805,
73
+ temp0_8_HumanEval: 0.2161,
74
+ temp0_8_HumanEval_ET: 0.1771,
75
+ },
76
+ {
77
+ name: 'CodeGemma-7b',
78
+ MGI: 0.0548,
79
+ temp0_HumanEval: 0.4329,
80
+ temp0_HumanEval_ET: 0.3537,
81
+ temp0_8_HumanEval: 0.3316,
82
+ temp0_8_HumanEval_ET: 0.2779,
83
+ },
84
+ {
85
+ name: 'AixCoder3-7b-base',
86
+ MGI: 0.0578,
87
+ temp0_HumanEval: 0.5427,
88
+ temp0_HumanEval_ET: 0.4451,
89
+ temp0_8_HumanEval: 0.4545,
90
+ temp0_8_HumanEval_ET: 0.3826,
91
+ },
92
+ {
93
+ name: 'CodeQwen1.5-7b',
94
+ MGI: 0.0361,
95
+ temp0_HumanEval: 0.439,
96
+ temp0_HumanEval_ET: 0.3841,
97
+ temp0_8_HumanEval: 0.3884,
98
+ temp0_8_HumanEval_ET: 0.3279,
99
+ },
100
+ {
101
+ name: 'Starcoder2-7b',
102
+ MGI: 0.0579,
103
+ temp0_HumanEval: 0.3659,
104
+ temp0_HumanEval_ET: 0.3232,
105
+ temp0_8_HumanEval: 0.2905,
106
+ temp0_8_HumanEval_ET: 0.2438,
107
+ },
108
+ {
109
+ name: 'Starcoder2-15b',
110
+ MGI: 0.0552,
111
+ temp0_HumanEval: 0.4695,
112
+ temp0_HumanEval_ET: 0.4024,
113
+ temp0_8_HumanEval: 0.3406,
114
+ temp0_8_HumanEval_ET: 0.2795,
115
+ },
116
+ {
117
+ name: 'CodeLlama-13b',
118
+ MGI: 0.0412,
119
+ temp0_HumanEval: 0.3415,
120
+ temp0_HumanEval_ET: 0.2927,
121
+ temp0_8_HumanEval: 0.2405,
122
+ temp0_8_HumanEval_ET: 0.2015,
123
+ },
124
+ {
125
+ name: 'CodeLlama-34b',
126
+ MGI: 0.0651,
127
+ temp0_HumanEval: 0.4817,
128
+ temp0_HumanEval_ET: 0.4146,
129
+ temp0_8_HumanEval: 0.3478,
130
+ temp0_8_HumanEval_ET: 0.2906,
131
+ },
132
+ {
133
+ name: 'CodeLlama-70b',
134
+ MGI: 0.0523,
135
+ temp0_HumanEval: 0.5244,
136
+ temp0_HumanEval_ET: 0.4512,
137
+ temp0_8_HumanEval: 0.4317,
138
+ temp0_8_HumanEval_ET: 0.3615,
139
+ },
140
+ ]
141
+ let ymins = new Array(0.27, 0.16, 0.21, 0.14)
142
+ let titleText = [
143
+ 'Memorization-Generalization Index VS HumanEval(temp = 0)',
144
+ 'Memorization-Generalization Index VS HumanEval(temp = 0.8)',
145
+ 'Memorization-Generalization Index VS HumanEval-ET(temp = 0)',
146
+ 'Memorization-Generalization Index VS HumanEval-ET(temp = 0.8)'
147
+ ]
148
+ let options = new Array(4)
149
+ for(let i = 0; i < 4; i++){
150
+ options[i] = {
151
+ series: [
152
+ { name: 'GPT-3.5-turbo-1103' },
153
+ { name: 'GPT-3.5-turbo-0613' },
154
+ { name: 'CodeGeeX4-All-9B' },
155
+ { name: 'Deepseek-Coder-6.7b-base' },
156
+ { name: 'Deepseek-Coder-6.7b-instruct' },
157
+ { name: 'CodeLlama-7b' },
158
+ { name: 'CodeLlama-7b-python' },
159
+ { name: 'CodeLlama-7b-instruct' },
160
+ { name: 'CodeGemma-2b' },
161
+ { name: 'CodeGemma-7b' },
162
+ { name: 'AixCoder3-7b-base' },
163
+ { name: 'CodeQwen1.5-7b' },
164
+ { name: 'Starcoder2-7b' },
165
+ { name: 'Starcoder2-15b' },
166
+ { name: 'CodeLlama-13b' },
167
+ { name: 'CodeLlama-34b' },
168
+ { name: 'CodeLlama-70b' }
169
+ ],
170
+ grid: {
171
+ show: true,
172
+ top: '15%',
173
+ left: '12%',
174
+ right: '45%',
175
+ borderColor: 'transparent',
176
+ backgroundColor: '#eff1f7'
177
+ },
178
+ legend: {
179
+ right: 95,
180
+ bottom: '11%',
181
+ orient: 'vertical',
182
+ icon: 'diamond',
183
+ itemWidth: 14,
184
+ itemHeight: 14,
185
+ textStyle: {
186
+ fontSize: 14.5,
187
+ color: '#5876fa'
188
+ },
189
+ },
190
+ title: {
191
+ show: true,
192
+ text: '',
193
+ left: 'center',
194
+ textStyle: {
195
+ fontFamily: "Arial",
196
+ fontSize: 20,
197
+ fontWeight: 'normal'
198
+ },
199
+ top: 25
200
+ },
201
+ tooltip: {
202
+ trigger: 'item',
203
+ },
204
+ xAxis: {
205
+ name: 'Memorization-Generalization Index',
206
+ nameLocation: 'middle',
207
+ nameGap: 32,
208
+ nameTextStyle: {
209
+ fontSize: 15,
210
+ fontFamily: 'Arial',
211
+ },
212
+ min: -0.08,
213
+ axisLabel: {
214
+ showMinLabel: false
215
+ },
216
+ splitLine: {
217
+ show: true,
218
+ lineStyle: {
219
+ color: 'white'
220
+ }
221
+ },
222
+ axisTick:{
223
+ show:false
224
+ },
225
+ axisLine: {
226
+ show: false,
227
+ },
228
+ },
229
+ yAxis: {
230
+ name: '',
231
+ nameLocation: 'middle',
232
+ nameGap: 40,
233
+ nameTextStyle: {
234
+ fontSize: 16,
235
+ fontFamily: 'Arial',
236
+ },
237
+ type: 'value',
238
+ axisLabel: {
239
+ showMinLabel: false
240
+ },
241
+ splitLine: {
242
+ show: true,
243
+ lineStyle: {
244
+ color: 'white'
245
+ }
246
+ },
247
+ axisTick:{
248
+ show:false
249
+ },
250
+ axisLine: {
251
+ show: false,
252
+ },
253
+ },
254
+ };
255
+ options[i].title.text = titleText[i]
256
+ options[i].tooltip.formatter = (
257
+ i < 2 ? function(params) {
258
+ return params.marker +
259
+ `<span style="color: ${params.color}; text-decoration: underline">` + params.seriesName + '</span>' +
260
+ '<br/><br/>MGI: ' + params.data[0] +
261
+ '<br/>HumanEval: ' + params.data[1] +
262
+ '<br/>HumanEval-ET: ' + params.data[2];
263
+ }
264
+ : function(params) {
265
+ return params.marker +
266
+ `<span style="color: ${params.color}; text-decoration: underline">` + params.seriesName + '</span>' +
267
+ '<br/><br/>MGI: ' + params.data[0] +
268
+ '<br/>HumanEval-ET: ' + params.data[1] +
269
+ '<br/>HumanEval: ' + params.data[2];
270
+ }
271
+ )
272
+ options[i].yAxis.name = (
273
+ i < 2 ? 'HumanEval'
274
+ : 'HumanEval-ET'
275
+ )
276
+ options[i].yAxis.min = ymins[i]
277
+ }
278
+
279
+ for(let i = 0; i < dataset.length; i++){
280
+ options[0].series[i].data = [[dataset[i].MGI, dataset[i].temp0_HumanEval,dataset[i].temp0_HumanEval_ET]]
281
+ options[1].series[i].data = [[dataset[i].MGI, dataset[i].temp0_8_HumanEval, dataset[i].temp0_8_HumanEval_ET]]
282
+ options[2].series[i].data = [[dataset[i].MGI, dataset[i].temp0_HumanEval_ET,dataset[i].temp0_HumanEval]]
283
+ options[3].series[i].data = [[dataset[i].MGI, dataset[i].temp0_8_HumanEval_ET, dataset[i].temp0_8_HumanEval]]
284
+ }
285
+
286
+ for(let i = 0; i < 4; i++){
287
+ for(let j = 0; j < dataset.length; ++j){
288
+ options[i].series[j].type = 'scatter'
289
+ options[i].series[j].symbol = 'diamond'
290
+ }
291
+ }
292
+
293
+ function setSymbolSize() {
294
+ for (var i = 0; i < dataset.length; i++) {
295
+ if(dataset[i].MGI < 0.1 && dataset[i].temp0_HumanEval > 0.5){
296
+ options[0].series[i].symbolSize = 50
297
+ options[1].series[i].symbolSize = 50
298
+ }
299
+ else if(dataset[i].temp0_HumanEval > 0.43 && dataset[i].temp0_HumanEval < 0.75){
300
+ options[0].series[i].symbolSize = 30
301
+ options[1].series[i].symbolSize = 30
302
+ }
303
+ else if(dataset[i].MGI < 0.015 || dataset[i].temp0_HumanEval > 0.38){
304
+ options[0].series[i].symbolSize = 20
305
+ options[1].series[i].symbolSize = 20
306
+ }
307
+ else {
308
+ options[0].series[i].symbolSize = 13
309
+ options[1].series[i].symbolSize = 13
310
+ }
311
+ }
312
+ for (var i = 0; i < dataset.length; i++) {
313
+ if(dataset[i].MGI < 0.1 && dataset[i].temp0_HumanEval_ET > 0.5){
314
+ options[2].series[i].symbolSize = 50
315
+ options[3].series[i].symbolSize = 50
316
+ }
317
+ else if(dataset[i].temp0_HumanEval_ET > 0.43 && dataset[i].temp0_HumanEval_ET < 0.75){
318
+ options[2].series[i].symbolSize = 30
319
+ options[3].series[i].symbolSize = 30
320
+ }
321
+ else if(dataset[i].MGI < 0.015 || dataset[i].temp0_HumanEval_ET > 0.38){
322
+ options[2].series[i].symbolSize = 20
323
+ options[3].series[i].symbolSize = 20
324
+ }
325
+ else {
326
+ options[2].series[i].symbolSize = 13
327
+ options[3].series[i].symbolSize = 13
328
+ }
329
+ }
330
+ }
331
+ setSymbolSize()
332
+
333
+ Chart.setOption(options[0])
334
+ Chart2.setOption(options[1])
335
+
336
+
337
+ // button and span
338
+ var btn_temp0_HumanEval = document.getElementById('btn_temp0_HumanEval')
339
+ var btn_temp0_HumanEval_ET = document.getElementById('btn_temp0_HumanEval_ET')
340
+ var btn_temp0_8_HumanEval = document.getElementById('btn_temp0_8_HumanEval')
341
+ var btn_temp0_8_HumanEval_ET = document.getElementById('btn_temp0_8_HumanEval_ET')
342
+
343
+
344
+
345
+ var span_temp0_HumanEval = document.getElementById('span_temp0_HumanEval')
346
+ var span_temp0_HumanEval_ET = document.getElementById('span_temp0_HumanEval_ET')
347
+ var span_temp0_8_HumanEval = document.getElementById('span_temp0_8_HumanEval')
348
+ var span_temp0_8_HumanEval_ET = document.getElementById('span_temp0_8_HumanEval_ET')
349
+
350
+ btn_temp0_HumanEval.addEventListener("click", function() {
351
+ Chart.setOption(options[0])
352
+ btn_temp0_HumanEval.style.backgroundColor = '#386df4'
353
+ span_temp0_HumanEval.style.color = '#386df4'
354
+ btn_temp0_HumanEval_ET.style.backgroundColor = '#d3d3d3'
355
+ span_temp0_HumanEval_ET.style.color = '#d3d3d3'
356
+ })
357
+ span_temp0_HumanEval.addEventListener("click", function() {
358
+ Chart.setOption(options[0]);
359
+ btn_temp0_HumanEval.style.backgroundColor = '#386df4'
360
+ span_temp0_HumanEval.style.color = '#386df4'
361
+ btn_temp0_HumanEval_ET.style.backgroundColor = '#d3d3d3'
362
+ span_temp0_HumanEval_ET.style.color = '#d3d3d3'
363
+ })
364
+
365
+ btn_temp0_HumanEval_ET.addEventListener("click", function() {
366
+ Chart.setOption(options[2])
367
+ btn_temp0_HumanEval.style.backgroundColor = '#d3d3d3'
368
+ span_temp0_HumanEval.style.color = '#d3d3d3'
369
+ btn_temp0_HumanEval_ET.style.backgroundColor = '#386df4'
370
+ span_temp0_HumanEval_ET.style.color = '#386df4'
371
+ })
372
+ span_temp0_HumanEval_ET.addEventListener("click", function() {
373
+ Chart.setOption(options[2])
374
+ btn_temp0_HumanEval.style.backgroundColor = '#d3d3d3'
375
+ span_temp0_HumanEval.style.color = '#d3d3d3'
376
+ btn_temp0_HumanEval_ET.style.backgroundColor = '#386df4'
377
+ span_temp0_HumanEval_ET.style.color = '#386df4'
378
+ })
379
+
380
+ btn_temp0_8_HumanEval.addEventListener("click", function() {
381
+ Chart2.setOption(options[1])
382
+ btn_temp0_8_HumanEval.style.backgroundColor = '#386df4'
383
+ span_temp0_8_HumanEval.style.color = '#386df4'
384
+ btn_temp0_8_HumanEval_ET.style.backgroundColor = '#d3d3d3'
385
+ span_temp0_8_HumanEval_ET.style.color = '#d3d3d3'
386
+ })
387
+ span_temp0_8_HumanEval.addEventListener("click", function() {
388
+ Chart2.setOption(options[1]);
389
+ btn_temp0_8_HumanEval.style.backgroundColor = '#386df4'
390
+ span_temp0_8_HumanEval.style.color = '#386df4'
391
+ btn_temp0_8_HumanEval_ET.style.backgroundColor = '#d3d3d3'
392
+ span_temp0_8_HumanEval_ET.style.color = '#d3d3d3'
393
+ })
394
+
395
+ btn_temp0_8_HumanEval_ET.addEventListener("click", function() {
396
+ Chart2.setOption(options[3])
397
+ btn_temp0_8_HumanEval.style.backgroundColor = '#d3d3d3'
398
+ span_temp0_8_HumanEval.style.color = '#d3d3d3'
399
+ btn_temp0_8_HumanEval_ET.style.backgroundColor = '#386df4'
400
+ span_temp0_8_HumanEval_ET.style.color = '#386df4'
401
+ })
402
+ span_temp0_8_HumanEval_ET.addEventListener("click", function() {
403
+ Chart2.setOption(options[3])
404
+ btn_temp0_8_HumanEval.style.backgroundColor = '#d3d3d3'
405
+ span_temp0_8_HumanEval.style.color = '#d3d3d3'
406
+ btn_temp0_8_HumanEval_ET.style.backgroundColor = '#386df4'
407
+ span_temp0_8_HumanEval_ET.style.color = '#386df4'
408
+ })
table.js ADDED
@@ -0,0 +1,274 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ const models = [
2
+ {
3
+ name: 'GPT-3.5-turbo-1103',
4
+ MGI: 0.2326,
5
+ temp0_HumanEval: 0.7439,
6
+ temp0_HumanEval_ET: 0.5915,
7
+ temp0_8_HumanEval: 0.7248,
8
+ temp0_8_HumanEval_ET: 0.5778,
9
+ link: 'https://platform.openai.com/docs/models/gpt-3-5-turbo',
10
+ },
11
+ {
12
+ name: 'GPT-3.5-turbo-0613',
13
+ MGI: 0.1314,
14
+ temp0_HumanEval: 0.6585,
15
+ temp0_HumanEval_ET: 0.5122,
16
+ temp0_8_HumanEval: 0.612,
17
+ temp0_8_HumanEval_ET: 0.4776,
18
+ link: 'https://platform.openai.com/docs/models/gpt-3-5-turbo',
19
+ },
20
+ {
21
+ name: 'CodeGeeX4-All-9B',
22
+ MGI: 0.5182,
23
+ temp0_HumanEval: 0.7805,
24
+ temp0_HumanEval_ET: 0.622,
25
+ temp0_8_HumanEval: 0.7606,
26
+ temp0_8_HumanEval_ET: 0.6078,
27
+ link: 'https://huggingface.co/THUDM/codegeex4-all-9b',
28
+ },
29
+ {
30
+ name: 'Deepseek-Coder-6.7b',
31
+ MGI: 0.0602,
32
+ temp0_HumanEval: 0.4817,
33
+ temp0_HumanEval_ET: 0.3963,
34
+ temp0_8_HumanEval: 0.3388,
35
+ temp0_8_HumanEval_ET: 0.2798,
36
+ link: 'https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base',
37
+ },
38
+ {
39
+ name: 'Deepseek-Coder-6.7b-instruct',
40
+ MGI: 0.0893,
41
+ temp0_HumanEval: 0.7134,
42
+ temp0_HumanEval_ET: 0.628,
43
+ temp0_8_HumanEval: 0.6494,
44
+ temp0_8_HumanEval_ET: 0.5673,
45
+ link: 'https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct'
46
+ },
47
+ {
48
+ name: 'CodeLlama-7b',
49
+ MGI: 0.0417,
50
+ temp0_HumanEval: 0.311,
51
+ temp0_HumanEval_ET: 0.2378,
52
+ temp0_8_HumanEval: 0.2211,
53
+ temp0_8_HumanEval_ET: 0.1832,
54
+ link: 'https://huggingface.co/codellama/CodeLlama-7b-hf'
55
+ },
56
+ {
57
+ name: 'CodeLlama-7b-python',
58
+ MGI: 0.0525,
59
+ temp0_HumanEval: 0.3841,
60
+ temp0_HumanEval_ET: 0.3354,
61
+ temp0_8_HumanEval: 0.3062,
62
+ temp0_8_HumanEval_ET: 0.2568,
63
+ link: 'https://huggingface.co/codellama/CodeLlama-7b-Python-hf'
64
+ },
65
+ {
66
+ name: 'CodeLlama-7b-instruct',
67
+ MGI: 0.0908,
68
+ temp0_HumanEval: 0.3537,
69
+ temp0_HumanEval_ET: 0.2988,
70
+ temp0_8_HumanEval: 0.2876,
71
+ temp0_8_HumanEval_ET: 0.2455,
72
+ link: 'https://huggingface.co/codellama/CodeLlama-7b-Instruct-hf'
73
+ },
74
+ {
75
+ name: 'CodeGemma-2b',
76
+ MGI: 0.0112,
77
+ temp0_HumanEval: 0.3537,
78
+ temp0_HumanEval_ET: 0.2805,
79
+ temp0_8_HumanEval: 0.2161,
80
+ temp0_8_HumanEval_ET: 0.1771,
81
+ link: 'https://huggingface.co/google/codegemma-2b',
82
+ },
83
+ {
84
+ name: 'CodeGemma-7b',
85
+ MGI: 0.0548,
86
+ temp0_HumanEval: 0.4329,
87
+ temp0_HumanEval_ET: 0.3537,
88
+ temp0_8_HumanEval: 0.3316,
89
+ temp0_8_HumanEval_ET: 0.2779,
90
+ link: 'https://huggingface.co/google/codegemma-7b',
91
+ },
92
+ {
93
+ name: 'AixCoder3-7b-base',
94
+ MGI: 0.0578,
95
+ temp0_HumanEval: 0.5427,
96
+ temp0_HumanEval_ET: 0.4451,
97
+ temp0_8_HumanEval: 0.4545,
98
+ temp0_8_HumanEval_ET: 0.3826,
99
+ link: 'https://github.com/aixcoder-plugin/aiXcoder-7B?ref=top.aibase.com',
100
+ },
101
+ {
102
+ name: 'CodeQwen1.5-7b',
103
+ MGI: 0.0361,
104
+ temp0_HumanEval: 0.439,
105
+ temp0_HumanEval_ET: 0.3841,
106
+ temp0_8_HumanEval: 0.3884,
107
+ temp0_8_HumanEval_ET: 0.3279,
108
+ link: 'https://huggingface.co/Qwen/CodeQwen1.5-7B',
109
+ },
110
+ {
111
+ name: 'Starcoder2-7b',
112
+ MGI: 0.0579,
113
+ temp0_HumanEval: 0.3659,
114
+ temp0_HumanEval_ET: 0.3232,
115
+ temp0_8_HumanEval: 0.2905,
116
+ temp0_8_HumanEval_ET: 0.2438,
117
+ link: 'https://huggingface.co/bigcode/starcoder2-7b',
118
+ },
119
+ {
120
+ name: 'Starcoder2-15b',
121
+ MGI: 0.0552,
122
+ temp0_HumanEval: 0.4695,
123
+ temp0_HumanEval_ET: 0.4024,
124
+ temp0_8_HumanEval: 0.3406,
125
+ temp0_8_HumanEval_ET: 0.2795,
126
+ link: 'https://huggingface.co/bigcode/starcoder2-15b',
127
+ },
128
+ {
129
+ name: 'CodeLlama-13b',
130
+ MGI: 0.0412,
131
+ temp0_HumanEval: 0.3415,
132
+ temp0_HumanEval_ET: 0.2927,
133
+ temp0_8_HumanEval: 0.2405,
134
+ temp0_8_HumanEval_ET: 0.2015,
135
+ link: 'https://huggingface.co/codellama/CodeLlama-13b-hf',
136
+ },
137
+ {
138
+ name: 'CodeLlama-34b',
139
+ MGI: 0.0651,
140
+ temp0_HumanEval: 0.4817,
141
+ temp0_HumanEval_ET: 0.4146,
142
+ temp0_8_HumanEval: 0.3478,
143
+ temp0_8_HumanEval_ET: 0.2906,
144
+ link: 'https://huggingface.co/codellama/CodeLlama-34b-hf',
145
+ },
146
+ {
147
+ name: 'CodeLlama-70b',
148
+ MGI: 0.0523,
149
+ temp0_HumanEval: 0.5244,
150
+ temp0_HumanEval_ET: 0.4512,
151
+ temp0_8_HumanEval: 0.4317,
152
+ temp0_8_HumanEval_ET: 0.3615,
153
+ link: 'https://huggingface.co/codellama/CodeLlama-70b-hf',
154
+ }
155
+ ]
156
+ function renderRows(model) {
157
+ return `
158
+ <tr>
159
+ <td><a href=${model.link} target="_blank">${model.name}</a></td>
160
+ <td class="td_value">${model.MGI}</td>
161
+ <td class="td_value">${model.temp0_HumanEval}</td>
162
+ <td class="td_value">${model.temp0_HumanEval_ET}</td>
163
+ <td class="td_value">${model.temp0_8_HumanEval}</td>
164
+ <td class="td_value">${model.temp0_8_HumanEval_ET}</td>
165
+ </tr>
166
+ `
167
+ }
168
+ function renderTables() {
169
+ let html = ''
170
+ let rows = models.length + 1
171
+ html += `
172
+ <tr>
173
+ <td class="td_HumanEval" rowspan="${rows}">HumanEval</td>
174
+ </tr>
175
+ `
176
+ models.forEach(model => {
177
+ html += renderRows(model);
178
+ });
179
+ document.querySelector('tbody').innerHTML = html;
180
+ }
181
+
182
+ function adjustBackgroundColor() {
183
+ var evalTable = document.getElementById('evalTable');
184
+ var rows = evalTable.getElementsByTagName('tr');
185
+ for(var i = 0; i < rows.length; i++){
186
+ rows[i].style.backgroundColor = (i < 3 ? '#ffffff' : (i % 2 === 0 ? '#fbfbfb' : '#ffffff'));
187
+ }
188
+ }
189
+
190
+ renderTables()
191
+ adjustBackgroundColor()
192
+
193
+ document.querySelector('table').addEventListener('click', (e) => {
194
+ if (e.target.matches('button')) {
195
+ let direction = e.target.dataset.direction
196
+ let type = e.target.dataset.type
197
+ models.sort(orderBy(direction, type))
198
+ renderTables()
199
+ adjustBackgroundColor()
200
+ e.target.dataset.direction = (direction === 'asc' ? 'desc' : 'asc')
201
+ e.target.style.transform = (direction === 'asc' ? 'rotate(0deg)' : 'rotate(180deg)')
202
+ }
203
+ })
204
+
205
+ function orderBy(direction, type) {
206
+ return function (a, b) {
207
+ let comparision = 0
208
+ let value = (direction === 'desc' ? 1 : -1)
209
+ switch (type) {
210
+ case 'name':
211
+ if (a.name.toLowerCase() > b.name.toLowerCase()) {
212
+ comparision = value
213
+ } else {
214
+ comparision = value * -1
215
+ }
216
+ break;
217
+ case 'MGI':
218
+ if (a.MGI > b.MGI) {
219
+ comparision = value
220
+ } else {
221
+ comparision = value * -1
222
+ }
223
+ break;
224
+ case 'temp0_HumanEval':
225
+ if (a.temp0_HumanEval > b.temp0_HumanEval) {
226
+ comparision = value
227
+ } else {
228
+ comparision = value * -1
229
+ }
230
+ break;
231
+ case 'temp0_HumanEval_ET':
232
+ if (a.temp0_HumanEval_ET > b.temp0_HumanEval_ET) {
233
+ comparision = value
234
+ } else {
235
+ comparision = value * -1
236
+ }
237
+ break;
238
+ case 'temp0_8_HumanEval':
239
+ if (a.temp0_8_HumanEval > b.temp0_8_HumanEval) {
240
+ comparision = value
241
+ } else {
242
+ comparision = value * -1
243
+ }
244
+ break;
245
+ case 'temp0_8_HumanEval_ET':
246
+ if (a.temp0_8_HumanEval_ET > b.temp0_8_HumanEval_ET) {
247
+ comparision = value
248
+ } else {
249
+ comparision = value * -1
250
+ }
251
+ break;
252
+ }
253
+ return comparision
254
+ }
255
+ }
256
+
257
+ var btns = document.getElementsByClassName("button_sort");
258
+
259
+ function adjustBtnForm(event) {
260
+ var btn = event.target;
261
+ if (btn.classList.contains("button_sort")) {
262
+ btn.style.borderTop = '8px solid orange';
263
+ for (var i = 0; i < btns.length; i++) {
264
+ if (btns[i] !== btn) {
265
+ btns[i].style.borderTop = '8px solid #9a9a9a';
266
+ btns[i].style.transform = 'rotate(180deg)';
267
+ }
268
+ }
269
+ }
270
+ }
271
+
272
+ for (var i = 0; i < btns.length; i++) {
273
+ btns[i].addEventListener("click", adjustBtnForm);
274
+ }