File size: 3,091 Bytes
1d6d955
 
 
a53584a
 
 
 
 
 
 
 
 
 
6f7802b
a53584a
 
 
 
 
 
 
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f7802b
a53584a
 
 
 
 
 
 
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c16fea1
a53584a
 
 
 
 
c16fea1
a53584a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
---
# NB-ROBERTA Training Code
This is the current training code for the planned nb-roberta models.

We are currently planning to run the following experiments:


<table>
  <tr>
   <td><strong>Name</strong>
   </td>
   <td><strong>nb-roberta-base-old (C)</strong>
   </td>
  </tr>
  <tr>
   <td>Corpus
   </td>
   <td>NbAiLab/nb_bert
   </td>
  </tr>
  <tr>
   <td>Pod size
   </td>
   <td>v4-64
   </td>
  </tr>
  <tr>
   <td>Batch size
   </td>
   <td>62*4*8 = 1984 = 2k
   </td>
  </tr>
  <tr>
   <td>Learning rate
   </td>
   <td>3e-4 (RoBERTa article is using 6e-4 and bs=8k)
   </td>
  </tr>
  <tr>
   <td>Number of steps
   </td>
   <td>250k
   </td>
  </tr>
</table>



<table>
  <tr>
   <td><strong>Name</strong>
   </td>
   <td><strong>nb-roberta-base-ext (B)</strong>
   </td>
  </tr>
  <tr>
   <td>Corpus
   </td>
   <td>NbAiLab/nbailab_extended
   </td>
  </tr>
  <tr>
   <td>Pod size
   </td>
   <td>v4-64
   </td>
  </tr>
  <tr>
   <td>Batch size
   </td>
   <td>62*4*8 = 1984 = 2k
   </td>
  </tr>
  <tr>
   <td>Learning rate
   </td>
   <td>3e-4 (RoBERTa article is using 6e-4 and bs=8k)
   </td>
  </tr>
  <tr>
   <td>Number of steps
   </td>
   <td>250k
   </td>
  </tr>
</table>



<table>
  <tr>
   <td><strong>Name</strong>
   </td>
   <td><strong>nb-roberta-large-ext</strong>
   </td>
  </tr>
  <tr>
   <td>Corpus
   </td>
   <td>NbAiLab/nbailab_extended
   </td>
  </tr>
  <tr>
   <td>Pod size
   </td>
   <td>v4-64
   </td>
  </tr>
  <tr>
   <td>Batch size
   </td>
   <td>32*4*8 = 2024 = 1k
   </td>
  </tr>
  <tr>
   <td>Learning rate
   </td>
   <td>2-e4 (RoBERTa article is using 4e-4 and bs=8k)
   </td>
  </tr>
  <tr>
   <td>Number of steps
   </td>
   <td>500k
   </td>
  </tr>
</table>



<table>
  <tr>
   <td><strong>Name</strong>
   </td>
   <td><strong>nb-roberta-base-scandi</strong>
   </td>
  </tr>
  <tr>
   <td>Corpus
   </td>
   <td>NbAiLab/scandinavian
   </td>
  </tr>
  <tr>
   <td>Pod size
   </td>
   <td>v4-64
   </td>
  </tr>
  <tr>
   <td>Batch size
   </td>
   <td>62*4*8 = 1984 = 2k
   </td>
  </tr>
  <tr>
   <td>Learning rate
   </td>
   <td>3e-4 (RoBERTa article is using 6e-4 and bs=8k)
   </td>
  </tr>
  <tr>
   <td>Number of steps
   </td>
   <td>250k
   </td>
  </tr>
</table>



<table>
  <tr>
   <td><strong>Name</strong>
   </td>
   <td><strong>nb-roberta-large-scandi</strong>
   </td>
  </tr>
  <tr>
   <td>Corpus
   </td>
   <td>NbAiLab/scandinavian
   </td>
  </tr>
  <tr>
   <td>Pod size
   </td>
   <td>v4-64
   </td>
  </tr>
  <tr>
   <td>Batch size
   </td>
   <td>32*4*8 = 1024 = 1k
   </td>
  </tr>
  <tr>
   <td>Learning rate
   </td>
   <td>2-e4 (RoBERTa article is using 4e-4 and bs=8k)
   </td>
  </tr>
  <tr>
   <td>Number of steps
   </td>
   <td>500k
   </td>
  </tr>
</table>


## Calculations
Some basic that we used when estimating the number of training steps:
* The Scandinavic Corpus is 85GB
* The Scandinavic Corpus contains 13B words
* With a conversion factor of 2.3, this is estimated to around 30B tokens
* 30B tokens / (512 seq length * 3000 batch size) = 20.000 steps