-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path17156681280501.html
470 lines (397 loc) · 19.4 KB
/
17156681280501.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
<!DOCTYPE html>
<html lang="zh">
<head>
<meta charset="utf-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>
LoREFT-Code-HumanEval-Result - Prepare for the FUTURE
</title>
<link href="atom.xml" rel="alternate" title="Prepare for the FUTURE" type="application/atom+xml">
<link rel="stylesheet" href="asset/css/style.min.css">
<link rel="stylesheet" href="asset/css/doc.css">
<script src="asset/app.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-head">
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="container">
<div class="navbar-brand">
<a target="_self" class="navbar-item " href="index.html">Home</a>
<a target="_self" class="navbar-item " href="archives.html">Archives</a>
<a role="button" id="navbarSNSRssSwitchBtn" class="navbar-burger burger" aria-label="menu" aria-expanded="false" data-target="navbarSNSRssButtons">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div id="navbarSNSRssButtons" class="navbar-menu">
<div class="navbar-start">
</div>
<div class="navbar-end">
<div class="navbar-item">
<!--buttons start-->
<div class="buttons">
<a href="atom.xml" target="_blank" title="RSS">
<span class="icon is-large has-text-black-bis">
<svg class="svg-inline--fa fa-rss fa-w-14 fa-lg" aria-hidden="true" focusable="false" data-prefix="fas" data-icon="rss" role="img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 448 512" data-fa-i2svg=""><path fill="currentColor" d="M128.081 415.959c0 35.369-28.672 64.041-64.041 64.041S0 451.328 0 415.959s28.672-64.041 64.041-64.041 64.04 28.673 64.04 64.041zm175.66 47.25c-8.354-154.6-132.185-278.587-286.95-286.95C7.656 175.765 0 183.105 0 192.253v48.069c0 8.415 6.49 15.472 14.887 16.018 111.832 7.284 201.473 96.702 208.772 208.772.547 8.397 7.604 14.887 16.018 14.887h48.069c9.149.001 16.489-7.655 15.995-16.79zm144.249.288C439.596 229.677 251.465 40.445 16.503 32.01 7.473 31.686 0 38.981 0 48.016v48.068c0 8.625 6.835 15.645 15.453 15.999 191.179 7.839 344.627 161.316 352.465 352.465.353 8.618 7.373 15.453 15.999 15.453h48.068c9.034-.001 16.329-7.474 16.005-16.504z"></path></svg><!-- <i class="fas fa-rss fa-lg"></i> -->
</span>
</a>
</div>
<!--buttons end-->
</div>
</div>
</div>
</div>
</nav>
</div>
<div class="hero-body ct-body"></div>
</section>
<section class="ct-body">
<div class="container">
<div class="columns is-variable bd-klmn-columns is-4 is-centered">
<div class="column is-four-fifths">
<div class="post-body single-content">
<h1 class="title">
LoREFT-Code-HumanEval-Result
</h1>
<div class="media">
<div class="media-content">
<div class="content">
<p>
<span class="date">2024/05/14</span>
<br />
<span class="tran-tags">Tags:</span>
<a class="tag is-link is-light" href='tag_Llama3%20ReFTEng%20vs%20PEFT.html'>#Llama3 ReFTEng vs PEFT</a>
</p>
</div>
</div>
</div>
</div>
<article class="markdown-body single-content">
<p>这篇文章主要是记录DeepSeek-Coder和bigcode-evaluation-harness的结果。</p>
<p>注：在计算pass@100时，最好使用8*24G的显卡。</p>
<h2><a id="mention" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>Mention!!!!!!!</h2>
<p>在运行bigcode-evaluation-harness时，一定要检查自己的<code>.bashrc</code>文件，确保其中的<code>HF_ENDPOINT</code>如下：</p>
<pre><code class="language-shell">export HF_ENDPOINT=https://hf-mirror.com
</code></pre>
<h2><a id="base-model" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>Base Model</h2>
<p>首先是没有code微调的llama3-8b-base和llama3-8b-instruct上的代码效果测评。</p>
<p>两个代码运行shell示例如下：</p>
<pre><code class="language-shell"># llama3-8b-base-dir
/home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B/snapshots/1460c22666392e470910ce3d44ffeb2ab7dbd4df
# llama3-8b-instruct-dir
/home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B-Instruct/snapshots/a8977699a3d0820e80129fb3c93c20fbd9972c41
# for bigcode-evaluation-harness
accelerate launch main.py \
--model /home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B/snapshots/1460c22666392e470910ce3d44ffeb2ab7dbd4df \
--is_llama_3 \
--tasks humaneval \
--max_length_generation 1024 \
--do_sample False \
--n_samples 10 \
--batch_size 10 \
--precision bf16 \
--allow_code_execution \
--metric_output_path /home/workspace/nanyang/bigcode-evaluation-harness/result/[email protected]
# for DeepSeek-Coder(modified)
accelerate launch eval_pal.py \
--base_model /home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B-Instruct/snapshots/a8977699a3d0820e80129fb3c93c20fbd9972c41 \
--language python \
--dataroot data/ \
-il3 \
--n_sample 1 \
--batch_size 1 \
-t 0.2 \
-top_p 1 \
-top_k 50 \
-trc \
-dtype bfloat16 \
-msl 4096 \
-mgl 1024 \
-ds
</code></pre>
<h3><a id="bigcode-evaluation-harness" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>bigcode-evaluation-harness</h3>
<p>我们对两个基础模型分别进行以下实验：</p>
<ul>
<li>采用greedy search的<code>pass@1</code>实验，用<code>pass@1-g</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.2</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@1</code>实验，用<code>pass@1-b</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.6</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@10</code>实验，用<code>pass@10-b</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.8</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@100</code>实验，用<code>pass@100-b</code>表示；</li>
</ul>
<p>实验结果如下所示：</p>
<table>
<thead>
<tr>
<th style="text-align: center">bigcode-evaluation-harness</th>
<th>Llama3-8B-Base</th>
<th style="text-align: left">Llama3-8B-Instruct</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: center">pass@1-g</td>
<td>0.27439024390243905</td>
<td style="text-align: left">0.5609756097560976</td>
</tr>
<tr>
<td style="text-align: center">pass@1-b</td>
<td>0.2926829268292683</td>
<td style="text-align: left">0.5609756097560976</td>
</tr>
<tr>
<td style="text-align: center">pass@10-b</td>
<td>0.5792682926829268</td>
<td style="text-align: left">0.8170731707317073</td>
</tr>
<tr>
<td style="text-align: center">pass@100-b</td>
<td>0.8841463414634146</td>
<td style="text-align: left">0.9390243902439024</td>
</tr>
</tbody>
</table>
<h3><a id="deepseek-coder" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>DeepSeek-Coder</h3>
<p>我们对两个基础模型分别进行以下实验：</p>
<ul>
<li>采用greedy search的<code>pass@1</code>实验，用<code>pass@1-g</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.2</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@1</code>实验，用<code>pass@1-b</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.6</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@10</code>实验，用<code>pass@10-b</code>表示；</li>
<li>根据图1，采用<code>best_temperature=0.8</code>，hugging_face中model.generate默认值的<code>top_p=1</code>和<code>top_k=50</code>进行<code>pass@100</code>实验，用<code>pass@100-b</code>表示；</li>
</ul>
<p>实验结果如下所示：</p>
<table>
<thead>
<tr>
<th style="text-align: center">DeepSeek-Coder</th>
<th>Llama3-8B-Base</th>
<th style="text-align: left">Llama3-8B-Instruct</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: center">pass@1-g</td>
<td>0.27439024390243905</td>
<td style="text-align: left">0.5487804878048781</td>
</tr>
<tr>
<td style="text-align: center">pass@1-b</td>
<td>0.2804878048780488</td>
<td style="text-align: left">0.5548780487804879</td>
</tr>
<tr>
<td style="text-align: center">pass@10-b</td>
<td>0.5914634146341463</td>
<td style="text-align: left">0.7865853658536586</td>
</tr>
<tr>
<td style="text-align: center">pass@100-b</td>
<td></td>
<td style="text-align: left"></td>
</tr>
</tbody>
</table>
<h2><a id="reft-model" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>ReFT Model</h2>
<p>接着对进行code任务微调后的ReFT模型进行评测：</p>
<p>两个代码运行shell示例如下：</p>
<pre><code class="language-shell"># llama3-8b-base-dir (in server-57)
/home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B/snapshots/1460c22666392e470910ce3d44ffeb2ab7dbd4df
# llama3-8b-base-reft-dir (in server-57)
/home/workspace/nanyang/pyreft/examples/loreft/official_results/Meta-Llama-3-8B.code.20240509210531714954
# llama3-8b-instruct-dir (in server-57)
/home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B-Instruct/snapshots/a8977699a3d0820e80129fb3c93c20fbd9972c41
# llama3-8b-instruct-reft-dir (in server-57)
/home/workspace/nanyang/pyreft/examples/loreft/official_results/Meta-Llama-3-8B-Instruct.code.20240512081302727605
# for bigcode-evaluation-harness
accelerate launch main.py \
--model /home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B/snapshots/1460c22666392e470910ce3d44ffeb2ab7dbd4df \
--reft_model /home/workspace/nanyang/pyreft/examples/loreft/official_results/Meta-Llama-3-8B.code.20240509210531714954 \
--tasks humaneval \
--max_length_generation 1024 \
--precision bf16 \
--allow_code_execution \
--is_llama_3 \
--metric_output_path /home/workspace/nanyang/bigcode-evaluation-harness/result/[email protected] \
--do_sample True \
--n_samples 1 \
--batch_size 1 \
--temperature 0.2 \
--top_p 1 \
--top_k 50
# for DeepSeek-Coder(modified)
accelerate launch eval_pal.py \
--base_model /home/share/nanyang/HuggingFace/.cache/huggingface/hub/models--Meta-Llama-3-8B-Instruct/snapshots/a8977699a3d0820e80129fb3c93c20fbd9972c41 \
--reft_model /home/workspace/nanyang/pyreft/examples/loreft/official_results/Meta-Llama-3-8B-Instruct.code.20240512081302727605 \
--language python \
--dataroot data/ \
-il3 \
--n_sample 1 \
--batch_size 1 \
-t 0.2 \
-top_p 1 \
-top_k 50 \
-trc \
-dtype bfloat16 \
-msl 4096 \
-mgl 1024 \
-ds
</code></pre>
<h3><a id="bigcode-evaluation-harness" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>bigcode-evaluation-harness</h3>
<p>得到结果如下：</p>
<table>
<thead>
<tr>
<th style="text-align: center">bigcode-evaluation-harness</th>
<th>Llama3-8B-Base</th>
<th style="text-align: left">Llama3-8B-Instruct</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: center">pass@1-g</td>
<td>0.3170731707317073</td>
<td style="text-align: left">0.4878048780487805</td>
</tr>
<tr>
<td style="text-align: center">pass@1-b</td>
<td>0.3231707317073171</td>
<td style="text-align: left">0.5304878048780488</td>
</tr>
<tr>
<td style="text-align: center">pass@10-b</td>
<td>0.6341463414634146</td>
<td style="text-align: left">0.774390243902439</td>
</tr>
<tr>
<td style="text-align: center">pass@100-b</td>
<td></td>
<td style="text-align: left"></td>
</tr>
</tbody>
</table>
<h3><a id="deepseek-coder" class="anchor" aria-hidden="true"><span class="octicon octicon-link"></span></a>DeepSeek-Coder</h3>
<p>得到结果如下：</p>
<table>
<thead>
<tr>
<th style="text-align: center">DeepSeek-Coder</th>
<th>Llama3-8B-Base</th>
<th style="text-align: left">Llama3-8B-Instruct</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: center">pass@1-g</td>
<td>0.31097560975609756</td>
<td style="text-align: left">0.47560975609756095</td>
</tr>
<tr>
<td style="text-align: center">pass@1-b</td>
<td>0.3048780487804878</td>
<td style="text-align: left">0.49390243902439024</td>
</tr>
<tr>
<td style="text-align: center">pass@10-b</td>
<td>0.6524390243902439</td>
<td style="text-align: left">0.75</td>
</tr>
<tr>
<td style="text-align: center">pass@100-b</td>
<td></td>
<td style="text-align: left"></td>
</tr>
</tbody>
</table>
</article>
<div class="comments-wrap">
<div class="share-comments">
</div>
</div><!-- end comments wrap -->
</div>
</div><!-- end columns -->
</div><!-- end container -->
</section>
<footer class="footer">
<div class="content has-text-centered">
<p>
Copyright © 2019
Powered by <a target="_blank" href="http://www.mweb.im">MWeb</a>,
Theme used <a target="_blank" href="https://bulma.io/">Bulma CSS</a>.
</p>
</div>
</footer>
<style>.mweb-charts{background:#fff;}
body{ box-sizing: border-box;
margin: 0 auto;}
@media print{
pre, code, pre code {
overflow: visible !important;
white-space: pre-wrap !important; /* css-3 */
white-space: -moz-pre-wrap !important; /* Mozilla, since 1999 */
white-space: -pre-wrap !important; /* Opera 4-6 */
white-space: -o-pre-wrap !important; /* Opera 7 */
word-wrap: break-word !important; /* Internet Explorer 5.5+ */
}
html,body{margin:0;padding:4px;}
}
div.code-toolbar {
position: relative;
}
div.code-toolbar > .toolbar {
position: absolute;
z-index: 10;
top: .3em;
right: .2em;
transition: opacity 0.3s ease-in-out;
opacity: 0;
}
div.code-toolbar:hover > .toolbar {
opacity: 1;
}
/* Separate line b/c rules are thrown out if selector is invalid.
IE11 and old Edge versions don't support :focus-within. */
div.code-toolbar:focus-within > .toolbar {
opacity: 1;
}
div.code-toolbar > .toolbar > .toolbar-item {
display: inline-block;
}
div.code-toolbar > .toolbar > .toolbar-item > a {
cursor: pointer;
}
div.code-toolbar > .toolbar > .toolbar-item > button {
background: none;
border: 0;
color: inherit;
font: inherit;
line-height: normal;
overflow: visible;
padding: 0;
-webkit-user-select: none; /* for button */
-moz-user-select: none;
-ms-user-select: none;
}
div.code-toolbar > .toolbar > .toolbar-item > a,
div.code-toolbar > .toolbar > .toolbar-item > button,
div.code-toolbar > .toolbar > .toolbar-item > span {
color: inherit;
font-size: .8em;
padding: 4px .5em;
background: #f5f2f0;
background: rgba(224, 224, 224, 0.4);
box-shadow: 0 2px 0 0 rgba(0,0,0,0.2);
border-radius: .5em;
}
div.code-toolbar > .toolbar > .toolbar-item > a:hover,
div.code-toolbar > .toolbar > .toolbar-item > a:focus,
div.code-toolbar > .toolbar > .toolbar-item > button:hover,
div.code-toolbar > .toolbar > .toolbar-item > button:focus,
div.code-toolbar > .toolbar > .toolbar-item > span:hover,
div.code-toolbar > .toolbar > .toolbar-item > span:focus {
color: inherit;
text-decoration: none;
}
</style><script src="https://cdn.jsdelivr.net/npm/[email protected]/prism.min.js"></script><script src="https://cdn.jsdelivr.net/npm/[email protected]/plugins/line-numbers/prism-line-numbers.min.js"></script><script>!function(){if("undefined"!=typeof Prism&&"undefined"!=typeof document){var e=[],t={},n=function(){};Prism.plugins.toolbar={};var a=Prism.plugins.toolbar.registerButton=function(n,a){var r;r="function"==typeof a?a:function(e){var t;return"function"==typeof a.onClick?((t=document.createElement("button")).type="button",t.addEventListener("click",(function(){a.onClick.call(this,e)}))):"string"==typeof a.url?(t=document.createElement("a")).href=a.url:t=document.createElement("span"),a.className&&t.classList.add(a.className),t.textContent=a.text,t},n in t?console.warn('There is a button with the key "'+n+'" registered already.'):e.push(t[n]=r)},r=Prism.plugins.toolbar.hook=function(a){var r=a.element.parentNode;var l=a.element.classList;if(l.contains('language-mermaid') || l.contains('language-echarts') || l.contains('language-plantuml')){return;} if(r&&/pre/i.test(r.nodeName)&&!r.parentNode.classList.contains("code-toolbar")){var o=document.createElement("div");o.classList.add("code-toolbar"),r.parentNode.insertBefore(o,r),o.appendChild(r);var i=document.createElement("div");i.classList.add("toolbar");var l=e,d=function(e){for(;e;){var t=e.getAttribute("data-toolbar-order");if(null!=t)return(t=t.trim()).length?t.split(/\s*,\s*/g):[];e=e.parentElement}}(a.element);d&&(l=d.map((function(e){return t[e]||n}))),l.forEach((function(e){var t=e(a);if(t){var n=document.createElement("div");n.classList.add("toolbar-item"),n.appendChild(t),i.appendChild(n)}})),o.appendChild(i)}};a("label",(function(e){var t=e.element.parentNode;if(t&&/pre/i.test(t.nodeName)&&t.hasAttribute("data-label")){var n,a,r=t.getAttribute("data-label");try{a=document.querySelector("template#"+r)}catch(e){}return a?n=a.content:(t.hasAttribute("data-url")?(n=document.createElement("a")).href=t.getAttribute("data-url"):n=document.createElement("span"),n.textContent=r),n}})),Prism.hooks.add("complete",r)}}();</script><link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/plugins/toolbar/prism-toolbar.min.css"><script src="https://cdn.jsdelivr.net/npm/[email protected]/plugins/copy-to-clipboard/prism-copy-to-clipboard.min.js"></script><script src="https://cdn.jsdelivr.net/npm/[email protected]/plugins/autoloader/prism-autoloader.min.js"></script><style>div.code-toolbar > .toolbar > .toolbar-item > a, div.code-toolbar > .toolbar > .toolbar-item > button, div.code-toolbar > .toolbar > .toolbar-item > span {padding: 4px .5em; background: #f5f2f0; background: rgba(224, 224, 224, 0.4);}</style>
</body>
</html>