KshitijAmbilduke commited on
Commit
9ef89a4
1 Parent(s): b5cafa1

Upload 382 files

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +2 -0
  2. multilinguality_megatron/.gitignore +4 -0
  3. multilinguality_megatron/LICENSE +376 -0
  4. multilinguality_megatron/README.md +43 -0
  5. multilinguality_megatron/__pycache__/finetune.cpython-39.pyc +0 -0
  6. multilinguality_megatron/ablation_eval_pipeline.sh +18 -0
  7. multilinguality_megatron/continue_pretraining.sh +185 -0
  8. multilinguality_megatron/convert2megatron.sh +45 -0
  9. multilinguality_megatron/cp.sh +10 -0
  10. multilinguality_megatron/debug.sh +101 -0
  11. multilinguality_megatron/deploy.sh +53 -0
  12. multilinguality_megatron/docs/Makefile +20 -0
  13. multilinguality_megatron/docs/_templates/autosummary/base.rst +5 -0
  14. multilinguality_megatron/docs/_templates/autosummary/class.rst +9 -0
  15. multilinguality_megatron/docs/_templates/autosummary/module.rst +29 -0
  16. multilinguality_megatron/docs/api/index.rst +130 -0
  17. multilinguality_megatron/docs/conf.py +64 -0
  18. multilinguality_megatron/docs/guide/faq.md +170 -0
  19. multilinguality_megatron/docs/guide/getting_started.md +276 -0
  20. multilinguality_megatron/docs/guide/index.md +10 -0
  21. multilinguality_megatron/docs/guide/instruction_tuning.md +92 -0
  22. multilinguality_megatron/docs/guide/tokenization.md +76 -0
  23. multilinguality_megatron/docs/guide/weights_conversion.md +87 -0
  24. multilinguality_megatron/docs/imgs/llama-falcon.png +3 -0
  25. multilinguality_megatron/docs/index.rst +75 -0
  26. multilinguality_megatron/docs/make.bat +35 -0
  27. multilinguality_megatron/docs/requirements.txt +11 -0
  28. multilinguality_megatron/ducttape/10B_all_cleaned.tconf +80 -0
  29. multilinguality_megatron/ducttape/10B_all_cleaned_13B.tconf +80 -0
  30. multilinguality_megatron/ducttape/10B_all_cleaned_extend32.tconf +83 -0
  31. multilinguality_megatron/ducttape/10B_all_cleaned_extend32_warmed_up.tconf +84 -0
  32. multilinguality_megatron/ducttape/10B_all_cleaned_extend32_warmup.tconf +93 -0
  33. multilinguality_megatron/ducttape/10B_all_wikipedia.tconf +83 -0
  34. multilinguality_megatron/ducttape/20B_all_cleaned_mc4.tconf +113 -0
  35. multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel.tconf +264 -0
  36. multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_13b.tconf +264 -0
  37. multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_concat.tconf +264 -0
  38. multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_instructions.tconf +271 -0
  39. multilinguality_megatron/ducttape/20B_all_cleaned_mc4_wiki.tconf +176 -0
  40. multilinguality_megatron/ducttape/20B_all_cleaned_parallel.tconf +194 -0
  41. multilinguality_megatron/ducttape/20B_all_dirty_mc4.tconf +124 -0
  42. multilinguality_megatron/ducttape/40B_all_cleaned_mc4_parallel.tconf +305 -0
  43. multilinguality_megatron/ducttape/continue_pretraining.tconf +77 -0
  44. multilinguality_megatron/ducttape/data_test.tconf +79 -0
  45. multilinguality_megatron/ducttape/data_test_extend32.tconf +79 -0
  46. multilinguality_megatron/ducttape/gemma_2B_20B_all_cleaned_mc4_parallel.tconf +280 -0
  47. multilinguality_megatron/ducttape/gemma_2b_flavio.tconf +546 -0
  48. multilinguality_megatron/ducttape/gemma_7B_20B_all_cleaned_mc4_parallel.tconf +280 -0
  49. multilinguality_megatron/ducttape/llama_3_flavio.tconf +546 -0
  50. multilinguality_megatron/ducttape/llama_3_flavio_wmt_annealing.tconf +570 -0
.gitattributes CHANGED
@@ -54,3 +54,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
54
  *.jpeg filter=lfs diff=lfs merge=lfs -text
55
  *.webp filter=lfs diff=lfs merge=lfs -text
56
  data_text_document.idx filter=lfs diff=lfs merge=lfs -text
 
 
 
54
  *.jpeg filter=lfs diff=lfs merge=lfs -text
55
  *.webp filter=lfs diff=lfs merge=lfs -text
56
  data_text_document.idx filter=lfs diff=lfs merge=lfs -text
57
+ multilinguality_megatron/megatron/fused_kernels/build/fused_mix_prec_layer_norm_cuda.so filter=lfs diff=lfs merge=lfs -text
58
+ multilinguality_megatron/megatron/fused_kernels/build/layer_norm_cuda_kernel.cuda.o filter=lfs diff=lfs merge=lfs -text
multilinguality_megatron/.gitignore ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ __pycache__
2
+ build
3
+ .vscode
4
+ perplexity_texts
multilinguality_megatron/LICENSE ADDED
@@ -0,0 +1,376 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ The following applies to all files unless otherwise noted:
2
+
3
+ # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
4
+ #
5
+ # Redistribution and use in source and binary forms, with or without
6
+ # modification, are permitted provided that the following conditions
7
+ # are met:
8
+ # * Redistributions of source code must retain the above copyright
9
+ # notice, this list of conditions and the following disclaimer.
10
+ # * Redistributions in binary form must reproduce the above copyright
11
+ # notice, this list of conditions and the following disclaimer in the
12
+ # documentation and/or other materials provided with the distribution.
13
+ # * Neither the name of NVIDIA CORPORATION nor the names of its
14
+ # contributors may be used to endorse or promote products derived
15
+ # from this software without specific prior written permission.
16
+ #
17
+ # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
18
+ # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19
+ # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20
+ # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21
+ # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22
+ # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23
+ # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
24
+ # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
25
+ # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26
+ # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27
+ # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28
+
29
+ --
30
+
31
+ This repository also contains code from Hugging Face Inc., Google Research,
32
+ Facebook (from their Fairseq and Dino projects), Microsoft(from their
33
+ Swin-Transformer project)and Philip Popien. Files from these
34
+ organizations have notices at the top of each file. Below are
35
+ licenses used in those files, as indicated.
36
+
37
+
38
+ ------------- LICENSE FOR Facebook, huggingface and Google Research code --------------
39
+
40
+
41
+ Apache License
42
+ Version 2.0, January 2004
43
+ http://www.apache.org/licenses/
44
+
45
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
46
+
47
+ 1. Definitions.
48
+
49
+ "License" shall mean the terms and conditions for use, reproduction,
50
+ and distribution as defined by Sections 1 through 9 of this document.
51
+
52
+ "Licensor" shall mean the copyright owner or entity authorized by
53
+ the copyright owner that is granting the License.
54
+
55
+ "Legal Entity" shall mean the union of the acting entity and all
56
+ other entities that control, are controlled by, or are under common
57
+ control with that entity. For the purposes of this definition,
58
+ "control" means (i) the power, direct or indirect, to cause the
59
+ direction or management of such entity, whether by contract or
60
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
61
+ outstanding shares, or (iii) beneficial ownership of such entity.
62
+
63
+ "You" (or "Your") shall mean an individual or Legal Entity
64
+ exercising permissions granted by this License.
65
+
66
+ "Source" form shall mean the preferred form for making modifications,
67
+ including but not limited to software source code, documentation
68
+ source, and configuration files.
69
+
70
+ "Object" form shall mean any form resulting from mechanical
71
+ transformation or translation of a Source form, including but
72
+ not limited to compiled object code, generated documentation,
73
+ and conversions to other media types.
74
+
75
+ "Work" shall mean the work of authorship, whether in Source or
76
+ Object form, made available under the License, as indicated by a
77
+ copyright notice that is included in or attached to the work
78
+ (an example is provided in the Appendix below).
79
+
80
+ "Derivative Works" shall mean any work, whether in Source or Object
81
+ form, that is based on (or derived from) the Work and for which the
82
+ editorial revisions, annotations, elaborations, or other modifications
83
+ represent, as a whole, an original work of authorship. For the purposes
84
+ of this License, Derivative Works shall not include works that remain
85
+ separable from, or merely link (or bind by name) to the interfaces of,
86
+ the Work and Derivative Works thereof.
87
+
88
+ "Contribution" shall mean any work of authorship, including
89
+ the original version of the Work and any modifications or additions
90
+ to that Work or Derivative Works thereof, that is intentionally
91
+ submitted to Licensor for inclusion in the Work by the copyright owner
92
+ or by an individual or Legal Entity authorized to submit on behalf of
93
+ the copyright owner. For the purposes of this definition, "submitted"
94
+ means any form of electronic, verbal, or written communication sent
95
+ to the Licensor or its representatives, including but not limited to
96
+ communication on electronic mailing lists, source code control systems,
97
+ and issue tracking systems that are managed by, or on behalf of, the
98
+ Licensor for the purpose of discussing and improving the Work, but
99
+ excluding communication that is conspicuously marked or otherwise
100
+ designated in writing by the copyright owner as "Not a Contribution."
101
+
102
+ "Contributor" shall mean Licensor and any individual or Legal Entity
103
+ on behalf of whom a Contribution has been received by Licensor and
104
+ subsequently incorporated within the Work.
105
+
106
+ 2. Grant of Copyright License. Subject to the terms and conditions of
107
+ this License, each Contributor hereby grants to You a perpetual,
108
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
109
+ copyright license to reproduce, prepare Derivative Works of,
110
+ publicly display, publicly perform, sublicense, and distribute the
111
+ Work and such Derivative Works in Source or Object form.
112
+
113
+ 3. Grant of Patent License. Subject to the terms and conditions of
114
+ this License, each Contributor hereby grants to You a perpetual,
115
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
116
+ (except as stated in this section) patent license to make, have made,
117
+ use, offer to sell, sell, import, and otherwise transfer the Work,
118
+ where such license applies only to those patent claims licensable
119
+ by such Contributor that are necessarily infringed by their
120
+ Contribution(s) alone or by combination of their Contribution(s)
121
+ with the Work to which such Contribution(s) was submitted. If You
122
+ institute patent litigation against any entity (including a
123
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
124
+ or a Contribution incorporated within the Work constitutes direct
125
+ or contributory patent infringement, then any patent licenses
126
+ granted to You under this License for that Work shall terminate
127
+ as of the date such litigation is filed.
128
+
129
+ 4. Redistribution. You may reproduce and distribute copies of the
130
+ Work or Derivative Works thereof in any medium, with or without
131
+ modifications, and in Source or Object form, provided that You
132
+ meet the following conditions:
133
+
134
+ (a) You must give any other recipients of the Work or
135
+ Derivative Works a copy of this License; and
136
+
137
+ (b) You must cause any modified files to carry prominent notices
138
+ stating that You changed the files; and
139
+
140
+ (c) You must retain, in the Source form of any Derivative Works
141
+ that You distribute, all copyright, patent, trademark, and
142
+ attribution notices from the Source form of the Work,
143
+ excluding those notices that do not pertain to any part of
144
+ the Derivative Works; and
145
+
146
+ (d) If the Work includes a "NOTICE" text file as part of its
147
+ distribution, then any Derivative Works that You distribute must
148
+ include a readable copy of the attribution notices contained
149
+ within such NOTICE file, excluding those notices that do not
150
+ pertain to any part of the Derivative Works, in at least one
151
+ of the following places: within a NOTICE text file distributed
152
+ as part of the Derivative Works; within the Source form or
153
+ documentation, if provided along with the Derivative Works; or,
154
+ within a display generated by the Derivative Works, if and
155
+ wherever such third-party notices normally appear. The contents
156
+ of the NOTICE file are for informational purposes only and
157
+ do not modify the License. You may add Your own attribution
158
+ notices within Derivative Works that You distribute, alongside
159
+ or as an addendum to the NOTICE text from the Work, provided
160
+ that such additional attribution notices cannot be construed
161
+ as modifying the License.
162
+
163
+ You may add Your own copyright statement to Your modifications and
164
+ may provide additional or different license terms and conditions
165
+ for use, reproduction, or distribution of Your modifications, or
166
+ for any such Derivative Works as a whole, provided Your use,
167
+ reproduction, and distribution of the Work otherwise complies with
168
+ the conditions stated in this License.
169
+
170
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
171
+ any Contribution intentionally submitted for inclusion in the Work
172
+ by You to the Licensor shall be under the terms and conditions of
173
+ this License, without any additional terms or conditions.
174
+ Notwithstanding the above, nothing herein shall supersede or modify
175
+ the terms of any separate license agreement you may have executed
176
+ with Licensor regarding such Contributions.
177
+
178
+ 6. Trademarks. This License does not grant permission to use the trade
179
+ names, trademarks, service marks, or product names of the Licensor,
180
+ except as required for reasonable and customary use in describing the
181
+ origin of the Work and reproducing the content of the NOTICE file.
182
+
183
+ 7. Disclaimer of Warranty. Unless required by applicable law or
184
+ agreed to in writing, Licensor provides the Work (and each
185
+ Contributor provides its Contributions) on an "AS IS" BASIS,
186
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
187
+ implied, including, without limitation, any warranties or conditions
188
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
189
+ PARTICULAR PURPOSE. You are solely responsible for determining the
190
+ appropriateness of using or redistributing the Work and assume any
191
+ risks associated with Your exercise of permissions under this License.
192
+
193
+ 8. Limitation of Liability. In no event and under no legal theory,
194
+ whether in tort (including negligence), contract, or otherwise,
195
+ unless required by applicable law (such as deliberate and grossly
196
+ negligent acts) or agreed to in writing, shall any Contributor be
197
+ liable to You for damages, including any direct, indirect, special,
198
+ incidental, or consequential damages of any character arising as a
199
+ result of this License or out of the use or inability to use the
200
+ Work (including but not limited to damages for loss of goodwill,
201
+ work stoppage, computer failure or malfunction, or any and all
202
+ other commercial damages or losses), even if such Contributor
203
+ has been advised of the possibility of such damages.
204
+
205
+ 9. Accepting Warranty or Additional Liability. While redistributing
206
+ the Work or Derivative Works thereof, You may choose to offer,
207
+ and charge a fee for, acceptance of support, warranty, indemnity,
208
+ or other liability obligations and/or rights consistent with this
209
+ License. However, in accepting such obligations, You may act only
210
+ on Your own behalf and on Your sole responsibility, not on behalf
211
+ of any other Contributor, and only if You agree to indemnify,
212
+ defend, and hold each Contributor harmless for any liability
213
+ incurred by, or claims asserted against, such Contributor by reason
214
+ of your accepting any such warranty or additional liability.
215
+
216
+ END OF TERMS AND CONDITIONS
217
+
218
+ APPENDIX: How to apply the Apache License to your work.
219
+
220
+ To apply the Apache License to your work, attach the following
221
+ boilerplate notice, with the fields enclosed by brackets "[]"
222
+ replaced with your own identifying information. (Don't include
223
+ the brackets!) The text should be enclosed in the appropriate
224
+ comment syntax for the file format. We also recommend that a
225
+ file or class name and description of purpose be included on the
226
+ same "printed page" as the copyright notice for easier
227
+ identification within third-party archives.
228
+
229
+ Copyright [yyyy] [name of copyright owner]
230
+
231
+ Licensed under the Apache License, Version 2.0 (the "License");
232
+ you may not use this file except in compliance with the License.
233
+ You may obtain a copy of the License at
234
+
235
+ http://www.apache.org/licenses/LICENSE-2.0
236
+
237
+ Unless required by applicable law or agreed to in writing, software
238
+ distributed under the License is distributed on an "AS IS" BASIS,
239
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
240
+ See the License for the specific language governing permissions and
241
+ limitations under the License.
242
+
243
+ ------------- LICENSE FOR Facebook Fairseq code --------------
244
+
245
+ MIT License
246
+
247
+ Copyright (c) Facebook, Inc. and its affiliates.
248
+
249
+ Permission is hereby granted, free of charge, to any person obtaining a copy
250
+ of this software and associated documentation files (the "Software"), to deal
251
+ in the Software without restriction, including without limitation the rights
252
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
253
+ copies of the Software, and to permit persons to whom the Software is
254
+ furnished to do so, subject to the following conditions:
255
+
256
+ The above copyright notice and this permission notice shall be included in all
257
+ copies or substantial portions of the Software.
258
+
259
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
260
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
261
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
262
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
263
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
264
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
265
+ SOFTWARE.
266
+
267
+ ------------- LICENSE FOR Mircrosoft Swin transformer code --------------
268
+
269
+ MIT License
270
+
271
+ Copyright (c) Microsoft Corporation.
272
+
273
+ Permission is hereby granted, free of charge, to any person obtaining a copy
274
+ of this software and associated documentation files (the "Software"), to deal
275
+ in the Software without restriction, including without limitation the rights
276
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
277
+ copies of the Software, and to permit persons to whom the Software is
278
+ furnished to do so, subject to the following conditions:
279
+
280
+ The above copyright notice and this permission notice shall be included in all
281
+ copies or substantial portions of the Software.
282
+
283
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
284
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
285
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
286
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
287
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
288
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
289
+ SOFTWARE
290
+
291
+
292
+ --------------- NVIDIA Source Code License for SegFormer -----------------
293
+ 1. Definitions
294
+
295
+ “Licensor” means any person or entity that distributes its Work.
296
+
297
+ “Software” means the original work of authorship made available under this
298
+ License.
299
+
300
+ “Work” means the Software and any additions to or derivative works of the
301
+ Software that are made available under this License.
302
+
303
+ The terms “reproduce,” “reproduction,” “derivative works,” and
304
+ “distribution” have the meaning as provided under U.S. copyright law;
305
+ provided, however, that for the purposes of this License, derivative works
306
+ shall not include works that remain separable from, or merely link
307
+ (or bind by name) to the interfaces of, the Work.
308
+
309
+ Works, including the Software, are “made available” under this License by
310
+ including in or with the Work either (a) a copyright notice referencing
311
+ the applicability of this License to the Work, or (b) a copy of this License.
312
+
313
+ 2. License Grant
314
+
315
+ 2.1 Copyright Grant. Subject to the terms and conditions of this License,
316
+ each Licensor grants to you a perpetual, worldwide, non-exclusive,
317
+ royalty-free, copyright license to reproduce, prepare derivative works of,
318
+ publicly display, publicly perform, sublicense and distribute its Work
319
+ and any resulting derivative works in any form.
320
+
321
+ 3. Limitations
322
+
323
+ 3.1 Redistribution. You may reproduce or distribute the Work only if
324
+ (a) you do so under this License, (b) you include a complete copy of this
325
+ License with your distribution, and (c) you retain without modification any
326
+ copyright, patent, trademark, or attribution notices that are present
327
+ in the Work.
328
+
329
+ 3.2 Derivative Works. You may specify that additional or different terms
330
+ apply to the use, reproduction, and distribution of your derivative works
331
+ of the Work (“Your Terms”) only if (a) Your Terms provide that the use
332
+ limitation in Section 3.3 applies to your derivative works, and (b) you
333
+ identify the specific derivative works that are subject to Your Terms.
334
+ Notwithstanding Your Terms, this License (including the redistribution
335
+ requirements in Section 3.1) will continue to apply to the Work itself.
336
+
337
+ 3.3 Use Limitation. The Work and any derivative works thereof only may
338
+ be used or intended for use non-commercially. Notwithstanding the
339
+ foregoing, NVIDIA and its affiliates may use the Work and any derivative
340
+ works commercially. As used herein, “non-commercially” means for research
341
+ or evaluation purposes only.
342
+
343
+ 3.4 Patent Claims. If you bring or threaten to bring a patent claim against
344
+ any Licensor (including any claim, cross-claim or counterclaim in a lawsuit)
345
+ to enforce any patents that you allege are infringed by any Work, then
346
+ your rights under this License from such Licensor (including the grant
347
+ in Section 2.1) will terminate immediately.
348
+
349
+ 3.5 Trademarks. This License does not grant any rights to use any Licensor’s
350
+ or its affiliates’ names, logos, or trademarks, except as necessary to
351
+ reproduce the notices described in this License.
352
+
353
+ 3.6 Termination. If you violate any term of this License, then your rights
354
+ under this License (including the grant in Section 2.1) will terminate
355
+ immediately.
356
+
357
+ 4. Disclaimer of Warranty.
358
+
359
+ THE WORK IS PROVIDED “AS IS” WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND,
360
+ EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
361
+ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR NON-INFRINGEMENT.
362
+ YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER THIS LICENSE.
363
+
364
+ 5. Limitation of Liability.
365
+
366
+ EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
367
+ THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
368
+ SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
369
+ INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT
370
+ OF OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
371
+ (INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
372
+ LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
373
+ COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN
374
+ ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
375
+
376
+
multilinguality_megatron/README.md ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## Installation Instructions
2
+
3
+ As a pre-requisite, make sure you have [ducttape](https://github.com/CoderPat/ducttape) and [(mini)conda](https://docs.conda.io/en/latest/miniconda.html) installed.
4
+
5
+ First, clone this repository.
6
+
7
+ Then, to create a new conda environment with all the necessary dependencies, run the following command:
8
+
9
+ ```bash
10
+ export CONDA_HOME="/path/to/(mini)conda3"
11
+ bash setup/conda.sh
12
+ ```
13
+
14
+ # Training
15
+
16
+ ## Data format
17
+
18
+ Before training, you must preprocess the training data. Before preprocessing, the data should be a `json` file, with the following format:
19
+ ```json
20
+ {"text": "<instance_0_text>"}
21
+ {"text": "<instance_1_text>"}
22
+ ```
23
+ Note that the preprocessing script will pack observations together in vectors of a specified length, and will separate each instance (json line) by the tokenizer's EOS token.
24
+
25
+ Then, run the bash scripts in this order:
26
+
27
+ ```bash
28
+ ./preprocess_data.sh [OPTIONS]
29
+ ./convert2megatron.sh [OPTIONS]
30
+ ./model_sharding.sh [OPTIONS]
31
+ ./continue_pretraining.sh [OPTIONS]
32
+ ```
33
+ >NOTE: each of these commands may be run with flag `--help`, which will inform the user on how to use each argument.
34
+
35
+ For example, for a continued pretraining run with Llama 2 7B on datasets `d1` and `d2` and 8 GPUs, run the following:
36
+
37
+ ```bash
38
+ > ./preprocess_data.sh --dataset_json=<path_to_d1> --dataset_bin=<d1_output_path> --vocab_file=<path_to_hf_model>/tokenizer.model --repo=<path_to_repo>
39
+ > ./preprocess_data.sh --dataset_json=<path_to_d2> --dataset_bin=<d2_output_path> --vocab_file=<path_to_hf_model>/tokenizer.model --repo=<path_to_repo>
40
+ > ./convert2megatron.sh --megatron_model=<megatron_model_path> --model_path=<path_to_hf_model> --size=7 --repo=<path_to_repo>
41
+ > ./model_sharding.sh --megatron_model=<megatron_model_path> --sharded_model=<sharded_model_path> --tp=8 --pp=1 --vocab_size=32000 --repo=<path_to_repo>
42
+ > ./continue_pretraining.sh --data_path="1 d1 1 d2" --megatron_model=<sharded_model_path> --model_dir=<checkpoint_save_dir> --tokenizer_path=<path_to_hf_model>/tokenizer.model --tp=8 --pp=1 [TRAINING_ARGS]
43
+ ```
multilinguality_megatron/__pycache__/finetune.cpython-39.pyc ADDED
Binary file (6.74 kB). View file
 
multilinguality_megatron/ablation_eval_pipeline.sh ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # bash script to evaluate a given model on wmt23, flores, ape, gec, standard benchmarks, and perplexity, sequentially
2
+
3
+ # wmt23, flores, ape, gec, standard benchmarks use tower-eval
4
+ TOWER_EVAL_DIR=/mnt/data/jpombal/tower-eval
5
+ cd $TOWER_EVAL_DIR
6
+ source $TOWER_EVAL_DIR/tower-eval-env/bin/activate
7
+
8
+ CUDA_VISIBLE_DEVICES=0 python $TOWER_EVAL_DIR/tower_eval/cli.py lm_eval --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/std_bench.yaml &
9
+ CUDA_VISIBLE_DEVICES=1 python $TOWER_EVAL_DIR/tower_eval/cli.py gen-eval --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/mt.yaml &
10
+ CUDA_VISIBLE_DEVICES=2 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_flavio_final.yaml &
11
+ #CUDA_VISIBLE_DEVICES=3 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_flaviarlos_sft.yaml &
12
+ #CUDA_VISIBLE_DEVICES=4 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_carlos_no_mt_annealed_sft.yaml
13
+ # CUDA_VISIBLE_DEVICES=2 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_porfirio_pre_annealing.yaml &
14
+ # CUDA_VISIBLE_DEVICES=3 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_porfirio_sft.yaml &
15
+ # CUDA_VISIBLE_DEVICES=4 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_carlos_sft.yaml &
16
+ # CUDA_VISIBLE_DEVICES=5 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_carlos_annealed_sft.yaml &
17
+ # CUDA_VISIBLE_DEVICES=6 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_flavio_sft.yaml &
18
+ # CUDA_VISIBLE_DEVICES=7 python $TOWER_EVAL_DIR/tower_eval/cli.py evaluate --config /mnt/data/jpombal/tower-eval/local_configs/cp_ablations/perplexity_porfirio_annealed.yaml &
multilinguality_megatron/continue_pretraining.sh ADDED
@@ -0,0 +1,185 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This script will try to run a task *outside* any specified submitter
2
+ # Note: This script is for archival; it is not actually run by ducttape
3
+ # unset CUDA_VISIBLE_DEVICES
4
+ echo $CUDA_VISIBLE_DEVICES
5
+
6
+ data_path="1 spgi_vox_mls_text_1b/data/data_text_document"
7
+ megatron_model="spgi_vox_mls_text_1b/shards"
8
+ model_dir="spgi_vox_mls_text_1b/ckpt"
9
+ tokenizer_path="spgi_vox_mls_text_1b/new_extended_tokenizer/tokenizer.model"
10
+ tp="2"
11
+ pp="1"
12
+
13
+ # --wandb_logger \
14
+ # --wandb_id "hajmola" \
15
+ # --wandb_project "Megatron" \
16
+ # --wandb_entity "hajmola" \
17
+ # --wandb_api_key "c4a95af43e910d14b0eca23fbb8165f94944d5af" \
18
+
19
+ # optimization arguments; self-explanatory. Intervals and steps are in terms of training optimizer steps
20
+ grad_accum_steps="12"
21
+ micro_batch_size="12"
22
+ warmup_steps="13"
23
+ eval_interval="500"
24
+ lr="3e-5" #lr="3e-5"
25
+ log_interval="10"
26
+ lr_min="3e-6" #lr_min="3e-6"
27
+ lr_scheduler="cosine"
28
+
29
+ # infra arguments
30
+ save_interval="250"
31
+ n_gpus="2"
32
+ repo="multilinguality_megatron"
33
+ gpu_ids="4,5"
34
+ train_steps="1000"
35
+
36
+
37
+ # Parse command-line arguments
38
+ for arg in "$@"
39
+ do
40
+ case $arg in
41
+ --help)
42
+ echo "Usage: ./script.sh [OPTIONS]"
43
+ echo "Options:"
44
+ echo " --data_path=PATH Path to dataset. Should have the form of <integer_0> <PATH_TO_DATA_TEXT_DOCUMENT_0> <integer_1> <PATH_TO_DATA_TEXT_DOCUMENT_1> ..., where the integers determine the data's relative weight in the training set. If every integer is equal, then the data is uniformly sampled."
45
+ echo " --megatron_model=PATH Path to sharded megatron model"
46
+ echo " --model_dir=PATH folder to save model checkpoints; if this has a checkpoint, it will be used to continue training"
47
+ echo " --tokenizer_path=PATH Path to tokenizer.model of original HF model"
48
+ echo " --tp=NUMBER Number of shards model is divided in"
49
+ echo " --pp=NUMBER Pipeline parallel (default is 1)"
50
+ echo " --grad_accum_steps=NUMBER"
51
+ echo " Number of gradient accumulation steps"
52
+ echo " --micro_batch_size=NUMBER"
53
+ echo " Micro batch size"
54
+ echo " --warmup_steps=NUMBER Number of warmup steps"
55
+ echo " --eval_interval=NUMBER Number of steps between validations"
56
+ echo " --lr=NUMBER Learning rate"
57
+ echo " --log_interval=NUMBER Number of steps between logging"
58
+ echo " --lr_min=NUMBER Minimum learning rate of scheduler"
59
+ echo " --lr_scheduler=STRING Learning rate scheduler"
60
+ echo " --save_interval=NUMBER Number of steps between saves"
61
+ echo " --n_gpus=NUMBER Number of GPUs to use"
62
+ echo " --repo=PATH Path to repo"
63
+ echo " --gpu_ids=STRING GPU IDs to use"
64
+ echo " --train_steps=NUMBER Number of training steps"
65
+ exit 0
66
+ ;;
67
+ --data_path=*)
68
+ data_path="${arg#*=}"
69
+ shift
70
+ ;;
71
+ --megatron_model=*)
72
+ megatron_model="${arg#*=}"
73
+ shift
74
+ ;;
75
+ --model_dir=*)
76
+ model_dir="${arg#*=}"
77
+ shift
78
+ ;;
79
+ --tokenizer_path=*)
80
+ tokenizer_path="${arg#*=}"
81
+ shift
82
+ ;;
83
+ --tp=*)
84
+ tp="${arg#*=}"
85
+ shift
86
+ ;;
87
+ --pp=*)
88
+ pp="${arg#*=}"
89
+ shift
90
+ ;;
91
+ --grad_accum_steps=*)
92
+ grad_accum_steps="${arg#*=}"
93
+ shift
94
+ ;;
95
+ --micro_batch_size=*)
96
+ micro_batch_size="${arg#*=}"
97
+ shift
98
+ ;;
99
+ --warmup_steps=*)
100
+ warmup_steps="${arg#*=}"
101
+ shift
102
+ ;;
103
+ --eval_interval=*)
104
+ eval_interval="${arg#*=}"
105
+ shift
106
+ ;;
107
+ --lr=*)
108
+ lr="${arg#*=}"
109
+ shift
110
+ ;;
111
+ --log_interval=*)
112
+ log_interval="${arg#*=}"
113
+ shift
114
+ ;;
115
+ --lr_min=*)
116
+ lr_min="${arg#*=}"
117
+ shift
118
+ ;;
119
+ --lr_scheduler=*)
120
+ lr_scheduler="${arg#*=}"
121
+ shift
122
+ ;;
123
+ --save_interval=*)
124
+ save_interval="${arg#*=}"
125
+ shift
126
+ ;;
127
+ --n_gpus=*)
128
+ n_gpus="${arg#*=}"
129
+ shift
130
+ ;;
131
+ --repo=*)
132
+ repo="${arg#*=}"
133
+ shift
134
+ ;;
135
+ --gpu_ids=*)
136
+ gpu_ids="${arg#*=}"
137
+ shift
138
+ ;;
139
+ --train_steps=*)
140
+ train_steps="${arg#*=}"
141
+ shift
142
+ ;;
143
+ esac
144
+ done
145
+
146
+ # CUDA_VISIBLE_DEVICES=$gpu_ids
147
+
148
+ if [ "$model_dir" != "" ]; then
149
+ mkdir -p $model_dir
150
+ mkdir -p $model_dir/runs
151
+ fi
152
+
153
+ ckpt_flag=$model_dir/latest_checkpointed_iteration.txt
154
+ if [ -f $ckpt_flag ]; then
155
+ megatron_model=$model_dir
156
+ echo Loading from previously saved checkpoint.
157
+ fi
158
+
159
+ global_batch_size=$(($micro_batch_size * $n_gpus * $grad_accum_steps))
160
+
161
+ LOG_ARGS="--log_interval $log_interval --save_interval $save_interval --eval_interval $eval_interval"
162
+ TRAIN_ARGS="--train_iters $train_steps --lr_decay_style $lr_scheduler --lr_warmup_iters $warmup_steps --lr $lr --min_lr $lr_min"
163
+ DISTRIBUTED_ARGS="--nproc_per_node $n_gpus --nnodes 1 --node_rank 0 --master_addr localhost --master_port 50000"
164
+ COMMON_ARGS="--hidden_dropout 0.0 --attention_dropout 0.0 --no_bias_gelu_fusion"
165
+ LLAMA_ARGS="--use_rms_norm --glu_activation swiglu --no_tie_embed_logits --no_new_tokens --layernorm_epsilon 1e-5"
166
+ CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun $DISTRIBUTED_ARGS $repo/finetune.py \
167
+ --tensor_model_parallel_size $tp \
168
+ --pipeline_model_parallel_size $pp \
169
+ --load $megatron_model \
170
+ --save $model_dir \
171
+ --tensorboard_dir $model_dir/runs \
172
+ --data_path $data_path \
173
+ --model_name llama \
174
+ --tokenizer_type SentencePieceTokenizer \
175
+ --vocab_file=$tokenizer_path \
176
+ --bf16 \
177
+ --use_flash_attn \
178
+ --micro_batch_size $micro_batch_size \
179
+ --global_batch_size $global_batch_size \
180
+ --sequence_parallel \
181
+ --recompute_granularity selective \
182
+ --use_checkpoint_args \
183
+ --seq_length 2048 \
184
+ --split 99,1,1 \
185
+ $COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA_ARGS
multilinguality_megatron/convert2megatron.sh ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ megatron_model="/mnt/scratch-artemis/kshitij/LLAMA/latest_megatron_codebase/spgi_vox_mls_text_1b/megatron_model"
4
+ model_path="/mnt/scratch-artemis/kshitij/LLAMA/latest_megatron_codebase/spgi_vox_mls_text_1b/extended_non_uniform_model"
5
+ size="1"
6
+ repo="/mnt/scratch-artemis/kshitij/LLAMA/latest_megatron_codebase/multilinguality_megatron"
7
+
8
+ # Parse command-line arguments
9
+ for arg in "$@"
10
+ do
11
+ case $arg in
12
+ --help)
13
+ echo "Usage: ./script.sh [OPTIONS]"
14
+ echo "Options:"
15
+ echo " --megatron_model=PATH Path to save converted model."
16
+ echo " --model_path=PATH Path of HF directory of model to be converted."
17
+ echo " --size=NUMBER Billion parameters of model."
18
+ echo " --repo=PATH Path to repo."
19
+ exit 0
20
+ ;;
21
+ --megatron_model=*)
22
+ megatron_model="${arg#*=}"
23
+ shift
24
+ ;;
25
+ --model_path=*)
26
+ model_path="${arg#*=}"
27
+ shift
28
+ ;;
29
+ --size=*)
30
+ size="${arg#*=}"
31
+ shift
32
+ ;;
33
+ --repo=*)
34
+ repo="${arg#*=}"
35
+ shift
36
+ ;;
37
+ esac
38
+ done
39
+
40
+ # Run the Python script
41
+ python $repo/weights_conversion/hf_to_megatron.py llama \
42
+ --size=$size \
43
+ --out=$megatron_model \
44
+ --cache-dir=$model_path \
45
+ --model-path=$model_path
multilinguality_megatron/cp.sh ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ langs=(en de es fr it pt nl ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
2
+
3
+ for lang in ${langs[@]}; do
4
+ mkdir -p /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}
5
+ echo "0" > /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}/ducttape_exit_code.txt
6
+ touch /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}/ducttape_stderr.txt
7
+ touch /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}/ducttape_stdout.txt
8
+ touch /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}/ducttape_task.sh
9
+ cp /mnt/cephfs-nvme/shared/tower-base-training-data/${lang}/dataset.json /mnt/cephfs-nvme/shared/experiments_megatron/cpt_llama_3/DumpHFDataset/Dataset.${lang}/ &
10
+ done
multilinguality_megatron/debug.sh ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ export dataset_bin="/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ru_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ru_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_fr_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_fr/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.fr_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ko_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.es_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.zh_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Baseline.baseline/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.pt_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_pt/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.pt_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_es/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.zh_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.it_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.zh/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_de/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ko_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ko_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_de_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.nl_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_ko_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.it/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.pt/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ru/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_zh/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.es/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.de_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.de_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_it/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.it_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_nl/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_zh_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ru_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.zh_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.it_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.instructions/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.de/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_pt_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.nl_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_ru/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.de_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.fr/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.ko/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.nl/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.es_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_it_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_ru_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_es_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.fr_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.nl_synth/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_nl_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.en_ko/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.es_en_pre_annealing/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.fr_en/data_bin /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/PreprocessDataset/Dataset.pt_en/data_bin"
2
+ export datamix_file="/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ru_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ru_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_fr_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_fr/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.fr_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ko_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.es_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.zh_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Baseline.baseline/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.pt_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_pt/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.pt_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_es/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.zh_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.it_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.zh/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_de/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ko_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ko_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_de_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.nl_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_ko_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.it/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.pt/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ru/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_zh/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.es/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.de_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.de_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_it/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.it_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_nl/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_zh_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ru_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.zh_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.it_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.instructions/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.de/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_pt_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.nl_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_ru/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.de_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.fr/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.ko/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.nl/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.es_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_it_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_ru_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_es_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.fr_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.nl_synth/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_nl_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.en_ko/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.es_en_pre_annealing/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.fr_en/datamix_file /mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/MakeDataMix/Dataset.pt_en/datamix_file"
3
+ export megatron_model="/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/ModelSharding/PP.1+Size.1+TP.1/sharded_model"
4
+ export model_dir="/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/doc_attn_tests"
5
+ export seq_length="2048"
6
+ export tp="1"
7
+ export warmup_steps="32"
8
+ export micro_batch_size="24"
9
+ export grad_accum_steps="4"
10
+ export kv_channels=""
11
+ export weight_decay="0.1"
12
+ export external_model_dir="/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_tinyllama_all_20B/doc_attn_tests"
13
+ export lr="3e-5"
14
+ export eval_interval="635"
15
+ export layernorm_epsilon="1e-5"
16
+ export log_interval="1"
17
+ export freeze_layers=""
18
+ export glu_activation="swiglu"
19
+ export eval_iters="1"
20
+ export lr_min="3e-6"
21
+ export pp="1"
22
+ export model_type="llama2"
23
+ export lr_scheduler="constant"
24
+ export tokenizer_path="/mnt/data_2/cache/models--TinyLlama--TinyLlama-1.1B-intermediate-step-1431k-3T/snapshots/036fa4651240b9a1487f709833b9e4b96b4c1574"
25
+ export save_interval="635"
26
+ export n_gpus="1"
27
+ export repo="/mnt/data/jpombal/multilinguality_megatron"
28
+ export gpu_ids="0"
29
+ export tokenizer_type="PretrainedFromHF"
30
+ export train_steps="11430"
31
+
32
+ external_model_dir="${external_model_dir}_${lr}"
33
+ if [ "$external_model_dir" != "" ]; then
34
+ mkdir -p $external_model_dir
35
+ mkdir -p $external_model_dir/runs
36
+ ln -s $external_model_dir $model_dir
37
+ fi
38
+
39
+ data_path=""
40
+ for f in $datamix_file; do
41
+ # read file
42
+ data_path="$data_path `cat $f`"
43
+ done
44
+ echo "Running with data_path=$data_path"
45
+
46
+ FREEZE_ARGS=""
47
+ if [ "$freeze_layers" == "not_embeddings" ]; then
48
+ FREEZE_ARGS="--freeze_layers"
49
+ fi
50
+ echo $FREEZE_ARGS
51
+
52
+ export CUDA_VISIBLE_DEVICES=$gpu_ids
53
+
54
+ # if load_from_checkpoint, then set megatron_model to external_model_dir
55
+ ckpt_flag=$external_model_dir/latest_checkpointed_iteration.txt
56
+ if [ -f $ckpt_flag ]; then
57
+ megatron_model=$external_model_dir
58
+ echo Loading from previously saved checkpoint.
59
+ fi
60
+
61
+ KV_CHANNELS_ARGS=""
62
+ if [ "$kv_channels" != "" ]; then
63
+ KV_CHANNELS_ARGS="--kv_channels $kv_channels"
64
+ fi
65
+
66
+ TIE_ARGS=""
67
+ if [ $model_type != 'gemma' ]; then
68
+ TIE_ARGS+="--no_tie_embed_logits"
69
+ fi
70
+ echo $TIE_ARGS
71
+
72
+ global_batch_size=$(($micro_batch_size * $n_gpus * $grad_accum_steps))
73
+
74
+ LOG_ARGS="--log_interval $log_interval --save_interval $save_interval --eval_interval $eval_interval --eval_iters $eval_iters --log_validation_ppl_to_tensorboard --log_memory_to_tensorboard --log_batch_size_to_tensorboard"
75
+ TRAIN_ARGS="--train_iters $train_steps --lr_decay_style $lr_scheduler --lr_warmup_iters $warmup_steps --lr $lr --min_lr $lr_min --weight_decay $weight_decay"
76
+ DISTRIBUTED_ARGS="--nproc_per_node $n_gpus --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8134"
77
+ COMMON_ARGS="--hidden_dropout 0.0 --attention_dropout 0.0 --no_bias_gelu_fusion"
78
+ LLAMA_ARGS="--use_rms_norm --glu_activation $glu_activation --no_new_tokens --layernorm_epsilon $layernorm_epsilon"
79
+ CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun $DISTRIBUTED_ARGS $repo/finetune.py \
80
+ --tensor_model_parallel_size $tp \
81
+ --pipeline_model_parallel_size $pp \
82
+ --load $megatron_model \
83
+ --save $model_dir \
84
+ --tensorboard_dir $external_model_dir/runs \
85
+ --data_path $data_path \
86
+ --model_name $model_type \
87
+ --tokenizer_type $tokenizer_type \
88
+ --vocab_file=$tokenizer_path \
89
+ --bf16 \
90
+ --use_flash_attn \
91
+ --micro_batch_size $micro_batch_size \
92
+ --global_batch_size $global_batch_size \
93
+ --sequence_parallel \
94
+ --recompute_granularity selective \
95
+ --use_checkpoint_args \
96
+ --seq_length $seq_length \
97
+ --split 9990,5,5 \
98
+ --sliding_window_size 4096 \
99
+ --reset_attention_mask \
100
+ --reset_position_ids \
101
+ $COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA_ARGS $FREEZE_ARGS $KV_CHANNELS_ARGS $TIE_ARGS \
multilinguality_megatron/deploy.sh ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ while getopts ":p:v:m:f:t:k:" opt; do
2
+ case ${opt} in
3
+ p )
4
+ path_to_weights=$OPTARG
5
+ ;;
6
+ v )
7
+ vocab_size=$OPTARG
8
+ ;;
9
+ m ) model_name=$OPTARG
10
+ ;;
11
+ f ) vocab_file=$OPTARG
12
+ ;;
13
+ t ) model_type=$OPTARG
14
+ ;;
15
+ k ) kv_channels=$OPTARG
16
+ ;;
17
+ \? )
18
+ echo "Invalid option: $OPTARG" 1>&2
19
+ exit 1
20
+ ;;
21
+ : )
22
+ echo "Invalid option: $OPTARG requires an argument" 1>&2
23
+ exit 1
24
+ ;;
25
+ esac
26
+ done
27
+ shift $((OPTIND -1))
28
+
29
+ KV_CHANNELS_ARGS=""
30
+ if [ "$kv_channels" != "" ]; then
31
+ KV_CHANNELS_ARGS="--kv_channels $kv_channels"
32
+ fi
33
+
34
+ # path_to_weights is where the latest_checkpointed_iteration.txt file is located
35
+ # script creates a folder with respective iteration in unsharded_dir, so no need to specify iteration
36
+ python tools/checkpoint_util.py \
37
+ --target_tensor_parallel_size 1 \
38
+ --target_pipeline_parallel_size 1 \
39
+ --load_dir $path_to_weights \
40
+ --save_dir "${path_to_weights}/unsharded" \
41
+ --model_type $model_type \
42
+ --true_vocab_size $vocab_size \
43
+ --bf16 \
44
+ $KV_CHANNELS_ARGS
45
+
46
+ python weights_conversion/megatron_to_hf.py \
47
+ --input_dir "${path_to_weights}/unsharded" \
48
+ --output_dir "${path_to_weights}/hf/${model_name}" \
49
+ --vocab_file "${vocab_file}" \
50
+ --model $model_type
51
+
52
+ # remove intermediate step
53
+ rm -r "${path_to_weights}/unsharded"
multilinguality_megatron/docs/Makefile ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Minimal makefile for Sphinx documentation
2
+ #
3
+
4
+ # You can set these variables from the command line, and also
5
+ # from the environment for the first two.
6
+ SPHINXOPTS ?=
7
+ SPHINXBUILD ?= sphinx-build
8
+ SOURCEDIR = .
9
+ BUILDDIR = _build
10
+
11
+ # Put it first so that "make" without argument is like "make help".
12
+ help:
13
+ @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
14
+
15
+ .PHONY: help Makefile
16
+
17
+ # Catch-all target: route all unknown targets to Sphinx using the new
18
+ # "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
19
+ %: Makefile
20
+ @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
multilinguality_megatron/docs/_templates/autosummary/base.rst ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {{ fullname | escape | underline}}
2
+
3
+ .. currentmodule:: {{ module }}
4
+
5
+ .. auto{{ objtype }}:: {{ objname }}
multilinguality_megatron/docs/_templates/autosummary/class.rst ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ fullname | escape | underline}}
2
+
3
+ .. currentmodule:: {{ module }}
4
+
5
+ .. autoclass:: {{ objname }}
6
+ :members:
7
+ :special-members:
8
+ :show-inheritance:
9
+ :exclude-members: __weakref__, __init__
multilinguality_megatron/docs/_templates/autosummary/module.rst ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ fullname | escape | underline }}
2
+
3
+ .. rubric:: Description
4
+
5
+ .. automodule:: {{ fullname }}
6
+
7
+ .. currentmodule:: {{ fullname }}
8
+
9
+ {% if classes %}
10
+ .. rubric:: Classes
11
+
12
+ .. autosummary::
13
+ :toctree: .
14
+ {% for class in classes %}
15
+ {{ class }}
16
+ {% endfor %}
17
+
18
+ {% endif %}
19
+
20
+ {% if functions %}
21
+ .. rubric:: Functions
22
+
23
+ .. autosummary::
24
+ :toctree: .
25
+ {% for function in functions %}
26
+ {{ function }}
27
+ {% endfor %}
28
+
29
+ {% endif %}
multilinguality_megatron/docs/api/index.rst ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ API
2
+ ===
3
+
4
+ megatron
5
+ --------
6
+
7
+ .. autosummary::
8
+ :toctree: megatron
9
+
10
+ megatron.arguments
11
+ megatron.checkpointing
12
+ megatron.dist_signal_handler
13
+ megatron.global_vars
14
+ megatron.indexer
15
+ megatron.initialize
16
+ megatron.memory
17
+ megatron.microbatches
18
+ megatron.optimizer_param_scheduler
19
+ megatron.p2p_communication
20
+ megatron.schedules
21
+ megatron.text_generation_server
22
+ megatron.timers
23
+ megatron.training
24
+ megatron.utils
25
+ megatron.wandb_logger
26
+
27
+ megatron.core
28
+ -------------
29
+
30
+ .. autosummary::
31
+ :toctree: megatron/core
32
+
33
+ megatron.core.parallel_state
34
+ megatron.core.utils
35
+
36
+
37
+ megatron.core.tensor_parallel
38
+ -----------------------------
39
+
40
+ .. autosummary::
41
+ :toctree: megatron/core/tensor_parallel
42
+
43
+ megatron.core.tensor_parallel.cross_entropy
44
+ megatron.core.tensor_parallel.data
45
+ megatron.core.tensor_parallel.layers
46
+ megatron.core.tensor_parallel.mappings
47
+ megatron.core.tensor_parallel.random
48
+ megatron.core.tensor_parallel.utils
49
+
50
+ megatron.data
51
+ -------------
52
+
53
+ .. autosummary::
54
+ :toctree: megatron/data
55
+
56
+ megatron.data.autoaugment
57
+ megatron.data.blendable_dataset
58
+ megatron.data.gpt_dataset
59
+ megatron.data.image_folder
60
+ megatron.data.realm_dataset_utils
61
+ megatron.data.bert_dataset
62
+ megatron.data.data_samplers
63
+ megatron.data.indexed_dataset
64
+ megatron.data.orqa_wiki_dataset
65
+ megatron.data.realm_index
66
+ megatron.data.biencoder_dataset_utils
67
+ megatron.data.dataset_utils
68
+ megatron.data.ict_dataset
69
+ megatron.data.t5_dataset
70
+
71
+ megatron.model
72
+ --------------
73
+
74
+ .. autosummary::
75
+ :toctree: megatron/model
76
+
77
+ megatron.model.bert_model
78
+ megatron.model.biencoder_model
79
+ megatron.model.classification
80
+ megatron.model.distributed
81
+ megatron.model.enums
82
+ megatron.model.falcon_model
83
+ megatron.model.fused_bias_gelu
84
+ megatron.model.fused_layer_norm
85
+ megatron.model.fused_softmax
86
+ megatron.model.glu_activations
87
+ megatron.model.gpt_model
88
+ megatron.model.language_model
89
+ megatron.model.llama_model
90
+ megatron.model.module
91
+ megatron.model.multiple_choice
92
+ megatron.model.positional_embeddings
93
+ megatron.model.t5_model
94
+ megatron.model.transformer
95
+ megatron.model.utils
96
+
97
+ megatron.optimizer
98
+ ------------------
99
+
100
+ .. autosummary::
101
+ :toctree: megatron/optimizer
102
+
103
+ megatron.optimizer.clip_grads
104
+ megatron.optimizer.distrib_optimizer
105
+ megatron.optimizer.grad_scaler
106
+ megatron.optimizer.optimizer
107
+
108
+ megatron.text_generation
109
+ ------------------------
110
+
111
+ .. autosummary::
112
+ :toctree: megatron/text_generation
113
+
114
+ megatron.text_generation.api
115
+ megatron.text_generation.beam_utils
116
+ megatron.text_generation.communication
117
+ megatron.text_generation.forward_step
118
+ megatron.text_generation.generation
119
+ megatron.text_generation.sampling
120
+ megatron.text_generation.tokenization
121
+
122
+ megatron.tokenizer
123
+ ------------------
124
+
125
+ .. autosummary::
126
+ :toctree: megatron/tokenizer
127
+
128
+ megatron.tokenizer.bert_tokenization
129
+ megatron.tokenizer.gpt2_tokenization
130
+ megatron.tokenizer.tokenizer
multilinguality_megatron/docs/conf.py ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Configuration file for the Sphinx documentation builder.
2
+ #
3
+ # For the full list of built-in configuration values, see the documentation:
4
+ # https://www.sphinx-doc.org/en/master/usage/configuration.html
5
+
6
+ # -- Path setup --------------------------------------------------------------
7
+
8
+ # If extensions (or modules to document with autodoc) are in another directory,
9
+ # add these directories to sys.path here. If the directory is relative to the
10
+ # documentation root, use os.path.abspath to make it absolute, like shown here.
11
+ #
12
+ import os
13
+ import sys
14
+ sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
15
+
16
+
17
+
18
+ # -- Project information -----------------------------------------------------
19
+ # https://www.sphinx-doc.org/en/master/usage/configuration.html#project-information
20
+
21
+ project = 'Megatron-LLM'
22
+ copyright = '2023, Alejandro Hernández Cano, Matteo Pagliardini, Kyle Matoba, Amirkeivan Mohtashami, Olivia Simin Fan, Axel Marmet, Deniz Bayazit, Igor Krawczuk, Zeming Chen, Francesco Salvi, Antoine Bosselut, Martin Jaggi'
23
+ author = 'Alejandro Hernández Cano, Matteo Pagliardini, Kyle Matoba, Amirkeivan Mohtashami, Olivia Simin Fan, Axel Marmet, Deniz Bayazit, Igor Krawczuk, Zeming Chen, Francesco Salvi, Antoine Bosselut, Martin Jaggi'
24
+ release = '0.1.0'
25
+
26
+ # -- General configuration ---------------------------------------------------
27
+ # https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration
28
+
29
+ extensions = [
30
+ 'sphinx.ext.autodoc',
31
+ 'sphinx.ext.intersphinx',
32
+ 'sphinx.ext.autosummary',
33
+ 'sphinx.ext.napoleon',
34
+ 'sphinx.ext.mathjax',
35
+ 'myst_parser'
36
+ ]
37
+
38
+ # autosummary
39
+ autosummary_generate = True
40
+
41
+ # napoleon
42
+ napoleon_google_docstring = True
43
+
44
+ # myst
45
+ myst_enable_extensions = ["colon_fence"]
46
+
47
+ # autodoc
48
+ autodoc_mock_imports = ['amp_C', 'torchvision', 'flash_attn', 'apex']
49
+
50
+ templates_path = ['_templates']
51
+ exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store']
52
+
53
+ intersphinx_mapping = {
54
+ 'python': ('https://docs.python.org/3', None)
55
+ }
56
+
57
+ master_doc = 'index'
58
+
59
+ # -- Options for HTML output -------------------------------------------------
60
+ # https://www.sphinx-doc.org/en/master/usage/configuration.html#options-for-html-output
61
+
62
+ html_theme = 'pydata_sphinx_theme'
63
+ # html_theme = 'sphinx_rtd_theme'
64
+ html_static_path = ['_static']
multilinguality_megatron/docs/guide/faq.md ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Frequently Asked Questions
2
+
3
+ ## How to add special tokens?
4
+
5
+ When defining a new task, it is often needed to introduce tokens with special meanings.
6
+ For instance, let's say we want to add two tokens `[formula]` and `[/formula]` to indicate the start and end of a formula in mathematics textbooks.
7
+ In order to include these new tokens, you need to indicate them in three different places:
8
+
9
+ 1. When tokenizing (`tools/preprocess_data.py`), using the flag `--vocab_extra_ids_list` with the new tokens:
10
+ ```
11
+ python tools/preprocess_data.py --vocab_extra_ids_list "[formula],[/formula]" # ...
12
+ ```
13
+
14
+ 1. When sharding the model (`tools/checkpoint_util.py`), using `--true_vocab_size`.
15
+ For instance, Falcon has 65024 tokens by default.
16
+ Including these two extra tokens will result in
17
+ ```
18
+ python tools/checkpoint_util.py --true_vocab_size 65026 # ...
19
+ ```
20
+
21
+ 1. When training (`finetune.py`) using `--vocab_extra_ids_list`.
22
+ Same as before:
23
+ ```
24
+ python finetune.py --vocab_extra_ids_list "[formula],[/formula]" # ...
25
+ ```
26
+
27
+ (tp-pp)=
28
+ ## How to set TP and PP?
29
+
30
+ General strategies:
31
+ - It is recommended to use data parallelism as much as possible, only use model parallelism if the model cannot fit in the GPU or the micro batch size is very small.
32
+ - It is preferable to use tensor parallelism before pipeline parallelism, when working on a single machine.
33
+ - When a model does not fit in a single node, use a tensor parallelism level of as many GPUs each node has, and pipeline parallelism level as small as possible to allow the model to fit in memory, and maintain a micro batch size large enough (of at least 5).
34
+
35
+ In the codebase, you won't set data parallelism explicitly.
36
+ Rather, the data parallelism will be inferred automatically to be as high as possible, depending in your available hardware and TP, PP levels.
37
+ In general, the number of GPUs you need is:
38
+ ```
39
+ GPUs = DP * TP * PP
40
+ ```
41
+ For instance, if you have two nodes with 8 GPUs each, TP=4 and PP=2, then DP will be automatically set to 2 as `4 x 2 x 2 = 16`.
42
+
43
+ ```{seealso}
44
+ - For more information on data and model parallelism see: https://huggingface.co/docs/transformers/v4.15.0/parallelism.
45
+ - Detailed information on how TP and PP works: https://arxiv.org/abs/2104.04473.
46
+ ```
47
+
48
+ ## How to launch training on multiple nodes?
49
+
50
+ In order to launch training on multiple nodes, you will set the appropriate arguments to the `torchrun` program.
51
+
52
+ 1. Select a "master" or main node and take note of its IP address.
53
+ 1. Launch the `finetune.py` script in the main node using `torchrun` with the following arguments:
54
+ ```
55
+ torchrun --n_proc_per_node NUMBER_OF_GPS_PER_NODE \
56
+ --nodes NUMBER_OF_NODES \
57
+ --node_rank 0 \
58
+ --master_addr ADDRESS_OF_THE_MAIN_NODE \
59
+ --master_port PORT \
60
+ finetune.py # ...
61
+ ```
62
+ 1. In the rest of the nodes, launch `finetune.py` with the same arguments, modifying `--node_rank` to a different value per node.
63
+
64
+ ```{seealso}
65
+ - Take a look at the example script `examples/finetune.sh` for more information.
66
+ - Look at the [How to set TP and PP?](#tp-pp) section for more information.
67
+ ```
68
+
69
+ ## What are the basic hardware requirements?
70
+
71
+ In this section we give a brief overview on the minimal hardware requirements we observed during our experiments.
72
+
73
+ | Model | min VRAM | tp | pp |
74
+ | :--------- | :------: | :-: | :-: |
75
+ | LLaMa2-7B | 2x 80GB | 2 | 1 |
76
+ | Falcon-40B | 16x 80GB | 8 | 2 |
77
+ | LLaMa2-70B | 32x 80GB | 8 | 4 |
78
+
79
+
80
+ (shard)=
81
+ ## How to shard and merge models?
82
+
83
+ Use `tools/checkpoint_util.py` to set the desired tensor and pipeline parallelism levels.
84
+
85
+ ```
86
+ python tools/checkpoint_util.py \
87
+ --target_tensor_parallel_size TP \
88
+ --target_pipeline_parallel_size PP \
89
+ --load_dir /path/to/original/weights/ \
90
+ --save_dir /path/to/new/weights/ \
91
+ --model_type MODEL \
92
+ --bf16
93
+ ```
94
+ Where MODEL can be either llama, llama2, falcon, gpt or bert, and TP and PP are the model parallelism levels desired.
95
+ Note that you can convert sharded weights (i.e. TP, PP > 1) to unsharded weights (TP = PP = 1) or viceversa.
96
+
97
+ ## What arguments are used to train LLaMa 2?
98
+
99
+ We set the same hyperparamters specified by Meta during finetuning (see [their paper for more information](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/)).
100
+ This means, that training LLaMa 2 7B can be done with the following arguments:
101
+
102
+ ```bash
103
+ torchrun \
104
+ # torchrun arguments # \
105
+ --nproc_per_node <GPUs per node> \
106
+ --nnodes <number of nodes> \
107
+ --node_rank <0,1,2,etc a different number per node> \
108
+ --master_addr <address of main node> \
109
+ --master_port <port> \
110
+ finetune.py --model_name llama2 \
111
+ # hardware/distributed arguments # \
112
+ --tensor_model_parallel_size <tp size> \
113
+ --pipeline_model_parallel_size <pp> \
114
+ --bf16 \
115
+ # training arguments # \
116
+ --train_iters <train iters> \
117
+ --adam_beta1 0.9 \
118
+ --adam_beta2 0.95 \
119
+ --adam_eps 1e-5 \
120
+ --lr_decay_style cosine 5 \
121
+ --lr_warmup_iters <warmup iters> \
122
+ --lr 3e-4 \
123
+ --min_lr 1e-6 \
124
+ --weight_decay 0.1 \
125
+ --micro_batch_size 5 \
126
+ --global_batch_size 1000 \
127
+ # additional optimization arguments # \
128
+ --use_flash_attn \
129
+ --sequence_parallel \
130
+ --recompute_granularity selective \
131
+ # logging/pathing arguments # \
132
+ --load <path to megatron-llama> \
133
+ --use_checkpoint_args \
134
+ --vocab_file <path to tokenizer.model provided by Meta> \
135
+ --log_interval 1 \
136
+ --data_path <path to tokenized data> \
137
+ --tokenizer_type SentencePieceTokenizer
138
+ ```
139
+
140
+ ```{seealso}
141
+ The file `examples/finetune.sh` gives the full picture of the arguments used to train either LLaMa.
142
+ ```
143
+
144
+ ## How to convert a LLaMa or Falcon architecture from a non-official checkpoint?
145
+
146
+ If you want to convert weights from a checkpoint other than the checkpoints provided by `llama-meta` or `tiiuae`, you might use `--model-path` during conversion.
147
+ For instance, to convert the [OpenAssistant llama2 70B](https://huggingface.co/OpenAssistant/llama2-70b-oasst-sft-v10) weights, run:
148
+
149
+ ```
150
+ python weights_conversion/hf_to_megatron.py llama2 --size=70 \
151
+ --out=/path/to/megatron/weights/ --cache-dir=/path/to/llama-2-7b/ \
152
+ --model-path=OpenAssistant/llama2-70b-oasst-sft-v10
153
+ ```
154
+
155
+ The `--model-path` argument should be either a local folder or the name of a model hosted on huggingface.
156
+
157
+ ## I'm getting a `17300 Bus error (core dumped)` error!
158
+
159
+ If you are using a docker container and you get this error when sharding a large model, you might need to increase the shared memory size.
160
+ This is done via the command line option `--shm-size=128gb`.
161
+
162
+ ## I'm getting a `ImportError: cannot import name 'helpers' from 'megatron.data'` error!
163
+
164
+ You need to compile the `helpers` module:
165
+
166
+ ```
167
+ cd megatron/data
168
+ make
169
+ cd ../../
170
+ ```
multilinguality_megatron/docs/guide/getting_started.md ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Getting started
2
+
3
+ This tutorial will guide you on the basic usage of Megatrom-LLM.
4
+ This guide we will fine tune a [LLaMa 2 7B](https://ai.meta.com/llama/) LLM on [code data](https://huggingface.co/datasets/bigcode/starcoderdata).
5
+ It is recommended to have at least 160GB VRAM available (e.g. two 80GB A100 GPUs).
6
+
7
+ ```{note}
8
+ This tutorial can also be followed to train a Falcon architecture, using `falcon` instead of `llama2` throughout the guide.
9
+ ```
10
+
11
+ ## Setup
12
+
13
+ First we need to install the dependencies.
14
+
15
+
16
+ 1. Clone our repo:
17
+ ```
18
+ git clone git@github.com:epfLLM/Megatron-LLM.git
19
+ ```
20
+
21
+ 1. Run the [nvcr docker image](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch), mounting the source code to your desired path, e.g. `/mpt/Megatron-LLM`:
22
+ ```
23
+ sudo docker run --gpus all -it --rm \
24
+ -v /path/to/Megatron-LLM/:/mpt/Megatron-LLM \
25
+ nvcr.io/nvidia/pytorch:23.07-py3
26
+ ```
27
+
28
+ 1. Enter the repository:
29
+ ```
30
+ cd /mpt/Megatron-LLM/
31
+ ```
32
+
33
+ 1. Install the additional dependencies not included in the `nvcr` image:
34
+ ```
35
+ pip install -r requirements.txt
36
+ ```
37
+
38
+ 1. Install the `megatron/data/helpers` binary:
39
+ ```
40
+ cd megatron/data/
41
+ make
42
+ cd ../../
43
+ ```
44
+
45
+ (download_weights)=
46
+ ## Downloading LLaMa2 weights
47
+
48
+ 1. Request access to the weights directly to meta: https://ai.meta.com/resources/models-and-libraries/llama-downloads/.
49
+ 1. Request access to the LLaMa2 huggingface model: https://huggingface.co/meta-llama/Llama-2-7b-hf.
50
+ 1. Create a new huggingface token (or use an existing one): https://huggingface.co/settings/tokens.
51
+ 1. Run the huggingface login CLI, and enter the token created on the previous step when asked:
52
+ ```
53
+ huggingface-cli login
54
+ ```
55
+
56
+ ## Preparing the raw data
57
+
58
+ :::{note}
59
+
60
+ This tutorial will use code data to fine tune the LLM.
61
+ Feel free to use any other dataset, as long as the raw data is saved in `.jsonl` format, i.e. one `json` dictionary with the key `"text"` per line:
62
+
63
+ ```json
64
+ {"text": "The blue cat is big."}
65
+ {"text": "This is another document."}
66
+ ```
67
+
68
+ In this case, skip to the [data preprocessing](#data-preprocessing) section.
69
+
70
+ :::
71
+
72
+ 1. Accept starcoder's terms of use via the huggingface portal: https://huggingface.co/datasets/bigcode/starcoderdata
73
+ 1. Create a huggingface token (or use an existing one) and login using `huggingface-cli` (see [Downloading LLaMa2 weights](#download_weights) for more information).
74
+ 1. Download and save the starcoder dataset.
75
+ In this tutorial we will use the `julia` data, but feel free to use any other subset.
76
+ This data contains around 500M tokens.
77
+ ```python
78
+ import json
79
+ from datasets import load_dataset
80
+
81
+ # the `cache_dir` argument is optional
82
+ dataset = load_dataset("bigcode/starcoderdata", data_dir="julia",
83
+ split="train", cache_dir="/path/to/cache/")
84
+ with open("/path/to/raw.jsonl", "w+") as f:
85
+ for document in dataset:
86
+ document = {"id": document["id"], "text": document["content"]}
87
+ f.write(json.dumps(document) + "\n")
88
+ ```
89
+
90
+ At this point, the raw data will be available at `/path/to/raw.jsonl`.
91
+
92
+
93
+ (data-preprocessing)=
94
+ ## Data preprocessing
95
+
96
+ In this step we will tokenize the raw data to binary files for optimized data loading during training.
97
+ Run:
98
+ ```
99
+ python tools/preprocess_data.py --input=/path/to/raw.jsonl \
100
+ --output_prefix=/path/to/tokenized/starcoder \
101
+ --tokenizer_type=SentencePieceTokenizer \
102
+ --vocab_file=/path/to/tokenizer.model \
103
+ --chunk_size=32 \
104
+ --workers=16 \
105
+ --no_new_tokens
106
+ ```
107
+
108
+ ```{note}
109
+ In this guide we use a sequence length of 1024 to accelerate training.
110
+ Note that the official sequence length of LLaMa2 is 4096.
111
+ ```
112
+
113
+ ```{note}
114
+ If you are using falcon, use `FalconTokenizer` instead of `SentencePieceTokenizer`, don't supply any `--vocab_file` and ignore the `--no_new_tokens` flag.
115
+ ```
116
+
117
+
118
+ (weight-conversion)=
119
+ ## Weight conversion
120
+
121
+ In order to use pretrained weights in the Megatron-LLM codebase, we will need to convert the official weights provided to be compatible with Megatron.
122
+ To do so, run:
123
+ ```
124
+ python weights_conversion/hf_to_megatron.py llama2 --size=7 \
125
+ --out=/path/to/megatron/weights/ --cache-dir=/path/to/llama-2-7b/
126
+ ```
127
+
128
+ (correctness-verification)=
129
+ ## Correctness verification (optional)
130
+
131
+ To make sure the weight conversion ran successfully we run the `verify_correctness.py` script.
132
+ This will run simultaneously the official LLaMa 2 implementation and the Megatron codebase.
133
+ Make sure to adjust the arguments to your convenience:
134
+ ```bash
135
+ # arguments required by `torchrun`
136
+ DISTRIBUTED_ARGS="--nproc_per_node 1 --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
137
+ LLAMA_ARGS="--use_rms_norm --glu_activation swiglu --no_tie_embed_logits --no_new_tokens --layernorm_epsilon 1e-5"
138
+ COMMON_ARGS="--hidden_dropout 0.0 --attention_dropout 0.0 --no_bias_gelu_fusion"
139
+ torchrun $DISTRIBUTED_ARGS verify_correctness.py \
140
+ --model_name=llama2 \
141
+ --model_size=7 \
142
+ --load=/path/to/megatron/weights/ \
143
+ --data_path=/path/to/tokenized/starcoder \
144
+ --tokenizer_type=SentencePieceTokenizer \
145
+ --vocab_file=/path/to/megatron/weights/tokenizer.model \
146
+ --huggingface_cache=/path/to/meta/llama-2-7b/ \
147
+ --huggingface_device=cuda:1 \
148
+ $COMMON_ARGS $LLAMA_ARGS # dont include LLAMA_ARGS if using Falcon
149
+ ```
150
+
151
+ This script will compare the logits output of Megatron model and the official implementation.
152
+ Expected outputs will yield average absolute error smaller than `0.01` when using 32-bit precision and `0.1` when using 16-bit precision.
153
+
154
+ ## Model sharding
155
+
156
+ In order to use model parallelism you need to split the previously converted weights into multiple files, before you start training.
157
+ To do this, use `tools/checkpoint_util.py`.
158
+ Feel free to use different tensor parallel (tp) and pipeline (pp) sizes.
159
+ ```
160
+ python tools/checkpoint_util.py \
161
+ --target_tensor_parallel_size 2 \
162
+ --target_pipeline_parallel_size 1 \
163
+ --load_dir /path/to/megatron/weights/ \
164
+ --save_dir /path/to/sharded/weights/ \
165
+ --model_type llama2 \
166
+ --true_vocab_size 32000 \
167
+ --bf16
168
+ ```
169
+
170
+ Feel free to set `--target_tensor_parallel_size` to 4 if you have 4 or more GPUs available.
171
+
172
+ ## Training
173
+
174
+ Use the `finetune.py`.
175
+ Example usage:
176
+ ```bash
177
+ LOG_ARGS="--log_interval 1 --save_interval 100 --eval_interval 50"
178
+ TRAIN_ARGS="--train_iters 500 --lr_decay_style cosine --lr_warmup_iters 50 --lr 3e-4 --min_lr 1e-6"
179
+ DISTRIBUTED_ARGS="--nproc_per_node NUMBER_OF_GPUS --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
180
+ torchrun $DISTRIBUTED_ARGS finetune.py \
181
+ --tensor_model_parallel_size 4 \
182
+ --pipeline_model_parallel_size 1 \
183
+ --load /path/to/sharded/weights/ \
184
+ --save /path/to/sharded/weights/ \
185
+ --tensorboard_dir /path/to/sharded/weights/tensorboard/ \
186
+ --data_path /path/to/tokenized/starcoder \
187
+ --model_name llama2 \
188
+ --tokenizer_type SentencePieceTokenizer \
189
+ --vocab_file=/path/to/megatron/weights/tokenizer.model \
190
+ --bf16 \
191
+ --use_flash_attn \
192
+ --micro_batch_size 5 \
193
+ --global_batch_size 1000 \
194
+ --sequence_parallel \
195
+ --recompute_granularity selective \
196
+ --use_checkpoint_args \
197
+ $COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA_ARGS
198
+ ```
199
+
200
+ With the selected global batch size of 1000, and the total number of training tokens around 500M, in 500 iterations the trainer will perform approximately one epoch.
201
+ This will take approximately 20 hours to run on a 8x 80GB A100 cluster (DP=2, TP=4, PP=1).
202
+
203
+ :::{note}
204
+
205
+ To use distributed training make sure to set `nproc_per_node` to the number of GPUs per node, `nnodes` to the number of nodes in your training and `master_addr` to the addres of your master node in the `DISTRIBUTED_ARGS` variable.
206
+ For instance, to train a two node cluster, with 8 GPUs each:
207
+ ```
208
+ DISTRIBUTED_ARGS="--nproc_per_node 1 --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
209
+ ```
210
+
211
+ Then, run the `finetune.py` script in all your nodes with the same parameters, just setting a different `node_rank` at every node.
212
+
213
+ :::
214
+
215
+ ```{seealso}
216
+ Take a look at `examples/finetune.sh for more information on the recommended hyperparameters
217
+ ```
218
+
219
+ ## Model Deployment
220
+
221
+ After training, merge your distributed weights again into a single model:
222
+ ```
223
+ python tools/checkpoint_util.py \
224
+ --target_tensor_parallel_size 1 \
225
+ --target_pipeline_parallel_size 1 \
226
+ --load_dir /path/to/sharded/weights/ \
227
+ --save_dir /path/to/unsharded/trained/weights/ \
228
+ --model_type llama2 \
229
+ --true_vocab_size 32000 \
230
+ --bf16
231
+ ```
232
+
233
+ We provide a Megatron to Huggingface conversion utility for easier deployment: `weights_conversion/megatron_to_hf.py`.
234
+ Run:
235
+ ```
236
+ python weights_conversion/megatron_to_hf.py --input_dir=/path/to/unsharded/trained/weights/ \
237
+ --output_dir=/path/to/hf/weights/
238
+ ```
239
+
240
+ Once the conversion is done, you can load the fine tuned weights using huggingface:
241
+ ```python
242
+ import torch
243
+ import transformers
244
+ from transformers import LlamaForCausalLM, LlamaTokenizer
245
+
246
+ pipeline = transformers.pipeline(
247
+ "text-generation",
248
+ model=LlamaForCausalLM.from_pretrained("/path/to/hf/weights/"),
249
+ tokenizer=LlamaTokenizer.from_pretrained("/path/to/hf/weights/"),
250
+ torch_dtype=torch.bfloat16,
251
+ device_map="auto"
252
+ )
253
+ prompt = """#= a function that returns the fibonacci number of its argument =#
254
+ function fibonacci(n::Int)::Int
255
+ """
256
+ sequences = pipeline(prompt, max_new_tokens=100, do_sample=True, top_k=20,
257
+ num_return_sequences=1)
258
+ for sequence in sequences:
259
+ print(sequence["generated_text"])
260
+ ```
261
+
262
+ Once you are happy with your model performance, you might publish it to the huggingface hub using the `tools/push_to_hub.py` utility:
263
+
264
+ ```
265
+ python tools/push_to_hub.py /path/to/hf/weights --hf_repo_name=MyRepoName/MyAwesomeModel --auth_token=MyAuthToken
266
+ ```
267
+
268
+ ## What's next?
269
+
270
+ 1. Take a look at our example scripts to familiarize yourself with some other capabilities and hyperparameters used in the codebase, such as to train (pretrain or finetune) larger models:
271
+ - `examples/parallelize.sh`
272
+ - `examples/finetune.sh`
273
+ - `examples/verify.sh`
274
+ 1. See the [intruction finetuning](instruction_tuning) guide for more information on how to finetune a pretrained model to follow instructions.
275
+ 1. Take a look at our [FAQ](faq) section.
276
+ 1. See [Weights conversion](weights_conversion) for more information on the `hf_to_megatron.py` and `megatron_to_hf.py` scripts.
multilinguality_megatron/docs/guide/index.md ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ # User guide
2
+
3
+ ```{toctree}
4
+
5
+ getting_started
6
+ instruction_tuning
7
+ faq
8
+ tokenization
9
+ weights_conversion
10
+ ```
multilinguality_megatron/docs/guide/instruction_tuning.md ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Instruction finetuning
2
+
3
+ This tutorial will guide you through the basics of instruction finetuning using the Megatron-LLM codebase, using LLaMa 2 as the base network.
4
+ See also the [getting started](getting_started) guide for information regarding installation of dependencies, pretraining, and weight preparation.
5
+ Following said tutorial, you would be able to finetune a 7B model in this guide, but feel free to use a different size.
6
+ In order to use Falcon, see the comments specified in the [getting started](getting_started) guide to learn more about the differences when using either model.
7
+
8
+ ## Preparing raw data
9
+
10
+ The dataset used in this guide will be a subset of the [orca](https://huggingface.co/datasets/Open-Orca/OpenOrca) dataset, a general purpose instruction dataset.
11
+ We choose to only include the chain of thought instructions from the orca dataset in order to shrink the size of the data.
12
+ Feel free to use any other dataset, as long as the raw data is saved in `.jsonl` format, i.e. one `json` dictionary per line.
13
+ The dictionaries must include at least two keys (one for the "instruction" and another one for the expected "answer"), plus an optional "system" key.
14
+ In order to retrieve the CoT subset of the orca dataset, use the following code:
15
+
16
+ ```python
17
+ import json
18
+
19
+ from datasets import load_dataset
20
+
21
+ # the `cache_dir` is optional
22
+ dataset = load_dataset("Open-Orca/OpenOrca", cache_dir="/path/to/cache", split="train")
23
+ with open("/path/to/raw/data.jsonl", "w+") as f:
24
+ for document in tqdm(dataset):
25
+ if document["id"].startswith("cot."):
26
+ f.write(json.dumps(document) + "\n")
27
+ ```
28
+
29
+ ## Data preprocessing
30
+
31
+ In this step we will tokenize the raw data to binary files for optimized data loading during training.
32
+ Run:
33
+ ```
34
+ python instruct/preprocess_instruct_data.py \
35
+ --input=/path/to/raw/data.jsonl \
36
+ --output_prefix=/path/to/tokenized/orca \
37
+ --tokenizer_type=SentencePieceTokenizer \
38
+ --vocab_file=/path/to/llama/tokenizer.model \
39
+ --chunk_size=32 \
40
+ --workers=32 \
41
+ --vocab_extra_ids_list "<|im_start|>,<|im_end|>" \
42
+ --question_key=question \
43
+ --answer_key=response \
44
+ --system_key=system_prompt # Optional
45
+ ```
46
+
47
+ ## Training
48
+
49
+ At this point, you should come up with a Megatron checkpoint ready to be trained (i.e. sharded with the desired parallelism levels).
50
+ Take a look at the [getting started](getting_started) guide to look how to transform LLaMa 2 checkpoints in the huggingface format to Megatron, and shard the weights.
51
+
52
+ To start training, use the `finetune.py`.
53
+ Example usage:
54
+ ```bash
55
+ LOG_ARGS="--log_interval 1 --save_interval 100 --eval_interval 50"
56
+ TRAIN_ARGS="--train_iters 6500 --lr_decay_style cosine --lr_warmup_iters 650 --lr 2e-5 --min_lr 2e-6"
57
+ DISTRIBUTED_ARGS="--nproc_per_node NUMBER_OF_GPUS --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
58
+ torchrun $DISTRIBUTED_ARGS finetune.py \
59
+ --tensor_model_parallel_size 4 \
60
+ --pipeline_model_parallel_size 1 \
61
+ --load /path/to/sharded/weights/ \
62
+ --save /path/to/sharded/weights/ \
63
+ --tensorboard_dir /path/to/sharded/weights/tensorboard/ \
64
+ --data_path /path/to/tokenized/orca \
65
+ --model_name llama2 \
66
+ --tokenizer_type SentencePieceTokenizer \
67
+ --vocab_file=/path/to/megatron/weights/tokenizer.model \
68
+ --bf16 \
69
+ --use_flash_attn \
70
+ --micro_batch_size 8 \
71
+ --global_batch_size 64 \
72
+ --sequence_parallel \
73
+ --recompute_granularity selective \
74
+ --use_checkpoint_args \
75
+ --data_type instruction \
76
+ --variable_seq_lengths \
77
+ --vocab_extra_ids_list "<|im_start|>,<|im_end|>" \
78
+ $COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA_ARGS
79
+ ```
80
+
81
+ The arguments given for pretraining and instruction finetuning are very similar, with the key differences being the batch sizes, learning rates, and the inclusion of `--data_type instruction`, `--variable_seq_lengths` and `--vocab_extra_ids_list`.
82
+ With the selected global batch size of 64, in 6500 iterations the trainer will perform approximately three epochs.
83
+ This will take approximately 3h hours to run on a 8x 80GB A100 device (DP=2, TP=4, PP=1).
84
+
85
+ ```{note}
86
+ If your `--load` checkpoint corresponds to a checkpoint already trained with the Megatron-LLM codebase (and not a checkpoint gotten after directly converting from the huggingface format for instance), you might want to define a `--save` directory that points somewhere else, to avoid overwritting previous checkpoints.
87
+ You might also want to include the `--finetune` argument to ignore the previous optimizer and RNG states.
88
+ ```
89
+
90
+ ## Model Deployment
91
+
92
+ Once the finetuning is over, you can follow the [getting started](getting_started) guide steps to unshard your weights and convert them to huggingface, in order to do specific evaluations and deployment.
multilinguality_megatron/docs/guide/tokenization.md ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # How to tokenize a dataset?
2
+
3
+ ## Step 1: get the right json format
4
+
5
+ The training data requires preprocessing. First, place your training data in a loose json format, with one json containing a text sample per line. For example:
6
+ <pre>
7
+ {"src": "www.nvidia.com", "text": "The quick brown fox", "type": "Eng", "id": "0", "title": "First Part"}
8
+ {"src": "The Internet", "text": "jumps over the lazy dog", "type": "Eng", "id": "42", "title": "Second Part"}
9
+ </pre>
10
+
11
+ The name of the `text` field of the json can be changed by using the `--json-key` flag in `preprocess_data.py`.
12
+ The other metadata are optional and are not used in training.
13
+
14
+ ## Step 2: Tokenize
15
+
16
+ The loose json is then processed into a binary format for training. To convert the json into mmap, cached index file, or the lazy loader format use `preprocess_data.py`. Set the `--dataset_impl` flag to `mmap`, `cached`, or `lazy`, respectively (default is `mmap`). An example script to prepare data for Falcon training is:
17
+ <pre>
18
+ python3 tools/preprocess_data.py --input /scratch/dummy-data/train.json
19
+ --output_prefix wiki-train
20
+ --dataset_impl mmap
21
+ --tokenizer_type FalconTokenizer
22
+ --workers 2
23
+ --chunk_size 32
24
+ --append_eod
25
+ </pre>
26
+
27
+ The output will be two files named, in this case, `my-bert_text_sentence.bin` and `my-bert_text_sentence.idx`. The `--data_path` specified in later BERT training is the full path and new filename, but without the file extension.
28
+
29
+ Other options of `preprocess_data.py`:
30
+
31
+ ```
32
+ input data:
33
+ --input INPUT Path to input JSON
34
+ --json_keys JSON_KEYS [JSON_KEYS ...]
35
+ space separate listed of keys to extract from json
36
+ --split_sentences Split documents into sentences.
37
+ --keep_newlines Keep newlines between sentences when splitting.
38
+
39
+ tokenizer:
40
+ --tokenizer_type {BertWordPieceLowerCase,BertWordPieceCase,GPT2BPETokenizer,SentencePieceTokenizer,FalconTokenizer}
41
+ What type of tokenizer to use.
42
+ --vocab_file VOCAB_FILE
43
+ Path to the vocab file
44
+ --merge_file MERGE_FILE
45
+ Path to the BPE merge file (if necessary).
46
+ --append_eod Append an <eod> token to the end of a document.
47
+ --lang LANG Language to use for NLTK-powered sentence splitting.
48
+
49
+ output data:
50
+ --output_prefix OUTPUT_PREFIX
51
+ Path to binary output file without suffix
52
+ --dataset_impl {lazy,cached,mmap}
53
+
54
+ runtime:
55
+ --workers WORKERS Number of worker processes to launch
56
+ --chunk_size CHUNK_SIZE
57
+ Chunk size assigned to each worker process
58
+ --log_interval LOG_INTERVAL
59
+ Interval between progress updates
60
+ --vocab_extra_ids VOCAB_EXTRA_IDS
61
+ --vocab_extra_ids_list VOCAB_EXTRA_IDS_LIST
62
+ comma separated list of special vocab ids to add to the tokenizer
63
+ --no_new_tokens Whether to add special tokens (e.g. CLS, MASK, etc) in the sentenciepiece tokenizer or not
64
+ ```
65
+
66
+ If you want to tokenize using llama tokenizer:
67
+ ```
68
+ python tools/preprocess_data.py \
69
+ --input=/path/to/data.json \
70
+ --output_prefix=wiki-train \
71
+ --dataset_impl=mmap \
72
+ --tokenizer_type=SentencePieceTokenizer \
73
+ --vocab_file=/path/to/tokenizer.model \
74
+ --workers=2 \
75
+ --chunk_size=32
76
+ ```
multilinguality_megatron/docs/guide/weights_conversion.md ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Weights conversion
2
+
3
+ ## Huggingface to megatron: `hf_to_megatron.py`
4
+
5
+ Convert weights from models in other formats (primairly huggingface) to megatron checkpoints.
6
+
7
+ This script supports converting Falcon, LLaMa and LLaMa 2 weights to megatron checkpoints.
8
+ Depending on the model to convert, the inputs might differ.
9
+
10
+ - **Falcon**:
11
+ Weights are automatically retrieved from the official implementation hosted in huggingface.
12
+ Thus, the `--cache-dir` argument is optional, if specified it should point to
13
+ the huggingface cache directory where the huggingface Falcon weights will be stored.
14
+ You will need to specify the `--size` argument to determine which version to download
15
+ (i.e. Falcon 7B or 40B).
16
+
17
+ - **LLaMa**, **LLaMa 2** and **CodeLlama**:
18
+ Converting llama weights can be done either fetching the weights hosted
19
+ in huggingface (recommended as it is the easier method) or directly from the
20
+ weights provided by Meta.
21
+
22
+ - From Meta weights (only available for LLaMa and LLaMa 2):
23
+ You will need to specify the `--cache-dir` to the directory where the
24
+ llama weights are stored.
25
+ This will by default have the form `xB` (e.g. 7B or 70B) for llama v1,
26
+ or `llama-2-xb` (e.g. llama-2-7b) for llama v2.
27
+
28
+ - From huggingface weights:
29
+ If `--cache-dir` is not specified or the directory specified does not
30
+ contain the format expected from Meta weights, the converter will automatically
31
+ retrieve the weights from huggingface, in which case the `--cache-dir` will
32
+ have the same semantics as with Falcon.
33
+
34
+ Note that to download llama v2 weights from huggingface, you will need to
35
+ login using `huggingface-cli login` with a huggingface account which has been
36
+ granted access to the `meta-llama/Llama-2-7b-hf` model.
37
+
38
+
39
+ In all cases, the megatron checkpoint will be stored in the `--out` argument.
40
+ If a huggingface is specified, the intermediate weights (i.e. the huggingface weights)
41
+ stored therein will not be removed when the conversion succeeds.
42
+
43
+ More information about the arguments:
44
+
45
+ ```
46
+ positional arguments:
47
+ {llama2,falcon,codellama,llama}
48
+
49
+ options:
50
+ -h, --help show this help message and exit
51
+ --size {65,34,70,7,40,13,30}
52
+ The size of the model
53
+ --out OUT Directory to store the megatron weights (as checkpoint)
54
+ --cache-dir CACHE_DIR
55
+ Directory to use as cache for the huggingface weights, or in case of the llama model, the path of the weights privided Meta
56
+ ```
57
+
58
+ ## Megatron to huggingface: `megatron_to_hf.py`
59
+
60
+ Convert megatron checkpoints to huggingface weights.
61
+
62
+ This script will also convert the tokenizer configured.
63
+ Set the `--input_dir` to the megatron checkpoint root (i.e. where the
64
+ `latest_checkpointed_iteration.txt` file is located) and `--output_dir` to
65
+ the directory where the huggingface weights should be stored.
66
+
67
+ More information about the arguments:
68
+
69
+ ```
70
+ options:
71
+ -h, --help show this help message and exit
72
+ --input_dir INPUT_DIR
73
+ Location of Megatron weights
74
+ --num_output_shards NUM_OUTPUT_SHARDS
75
+ --model {llama2,falcon,llama,codellama}
76
+ --output_dir OUTPUT_DIR
77
+ Location to write HF model and tokenizer
78
+ --cache_dir CACHE_DIR
79
+ Huggingface cache_dir (optional)
80
+ --vocab_file VOCAB_FILE
81
+ Path to the vocab file
82
+ --vocab_extra_ids_list VOCAB_EXTRA_IDS_LIST
83
+ comma separated list of special vocab ids to add to the tokenizer
84
+ --override_special_tokens [OVERRIDE_SPECIAL_TOKENS ...]
85
+ One or more arguments to override special tokens. Syntax set as `key=value`, e.g. `eos=<|im_end|>`. Overrides available only bos,
86
+ cls, eos, mask, pad, sep, unk.
87
+ ```
multilinguality_megatron/docs/imgs/llama-falcon.png ADDED

Git LFS Details

  • SHA256: b2f649e67f8b9867bd941109ad79f4459226f50f1b66e8a49e23922aaf0bff12
  • Pointer size: 132 Bytes
  • Size of remote file: 2.47 MB
multilinguality_megatron/docs/index.rst ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Welcome to Megatron-LLM's documentation!
2
+ ========================================
3
+
4
+ .. image:: imgs/llama-falcon.png
5
+
6
+ The `Megatron-LLM <https://github.com/epfLLM/Megatron-LLM/>`_ library enables pre-training and fine-tuning of large language models (LLMs) at scale.
7
+ Our repository is a modification of the `original Megatron-LM codebase <https://github.com/NVIDIA/Megatron-LM>`_ by Nvidia.
8
+
9
+ Added key features include:
10
+
11
+ - `LLaMa <https://arxiv.org/abs/2302.13971>`_, `LLaMa 2 <https://arxiv.org/abs/2307.09288>`_, `Falcon <https://huggingface.co/tiiuae>`_, and `Code Llama <https://together.ai/blog/llama-2-7b-32k>`_ support.
12
+ - support training of large models (70B Llama 2, 65B Llama 1, 34B Code Llama, and 40B Falcon) on commodity hardware on multiple nodes
13
+ - 3-way parallelism: tensor parallel, pipeline parallel and data parallel training (inherited from Megatron)
14
+ - full pretraining, finetuning and instruct tuning support
15
+ - Support for special tokens & tokenizers
16
+ - grouped-query attention (GQA) and multi-query attention (MQA)
17
+ - Rotary Position Embeddings (RoPE), RMS layer norm, Lima dropout
18
+ - `ROPE scaling <https://together.ai/blog/llama-2-7b-32k>`_ for longer attention context support
19
+ - FlashAttention 2
20
+ - BF16 / FP16 training
21
+ - WandB integration
22
+ - Metrics support: Ease to add custom metrics to evaluate on the validation set while training
23
+ - Conversion to and from Hugging Face hub
24
+
25
+ Example models trained with `Megatron-LLM <https://github.com/epfLLM/Megatron-LLM/>`_: See `README <https://github.com/epfLLM/Megatron-LLM/>`_.
26
+
27
+ User guide
28
+ ----------
29
+
30
+ For information on installation and usage, take a look at our user guide.
31
+
32
+ .. toctree::
33
+ :maxdepth: 2
34
+
35
+ guide/index
36
+
37
+
38
+ API
39
+ ---
40
+
41
+ Detailed information about Megatron-LLM components:
42
+
43
+ .. toctree::
44
+ :maxdepth: 2
45
+
46
+ api/index
47
+
48
+
49
+
50
+
51
+ Citation
52
+ --------
53
+
54
+ If you use this software please cite it:
55
+
56
+ .. code-block:: bib
57
+
58
+ @software{epfmgtrn,
59
+ author = {Alejandro Hernández Cano and
60
+ Matteo Pagliardini and
61
+ Andreas Köpf and
62
+ Kyle Matoba and
63
+ Amirkeivan Mohtashami and
64
+ Olivia Simin Fan and
65
+ Axel Marmet and
66
+ Deniz Bayazit and
67
+ Igor Krawczuk and
68
+ Zeming Chen and
69
+ Francesco Salvi and
70
+ Antoine Bosselut and
71
+ Martin Jaggi},
72
+ title = {epfLLM Megatron-LM},
73
+ year = 2023,
74
+ url = {https://github.com/epfLLM/Megatron-LLM}
75
+ }
multilinguality_megatron/docs/make.bat ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ @ECHO OFF
2
+
3
+ pushd %~dp0
4
+
5
+ REM Command file for Sphinx documentation
6
+
7
+ if "%SPHINXBUILD%" == "" (
8
+ set SPHINXBUILD=sphinx-build
9
+ )
10
+ set SOURCEDIR=.
11
+ set BUILDDIR=_build
12
+
13
+ %SPHINXBUILD% >NUL 2>NUL
14
+ if errorlevel 9009 (
15
+ echo.
16
+ echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
17
+ echo.installed, then set the SPHINXBUILD environment variable to point
18
+ echo.to the full path of the 'sphinx-build' executable. Alternatively you
19
+ echo.may add the Sphinx directory to PATH.
20
+ echo.
21
+ echo.If you don't have Sphinx installed, grab it from
22
+ echo.https://www.sphinx-doc.org/
23
+ exit /b 1
24
+ )
25
+
26
+ if "%1" == "" goto help
27
+
28
+ %SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
29
+ goto end
30
+
31
+ :help
32
+ %SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
33
+
34
+ :end
35
+ popd
multilinguality_megatron/docs/requirements.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ sphinx == 7.1.0
2
+ pydata-sphinx-theme >= 0.13.0
3
+ myst-parser >= 2.0.0
4
+ flask >= 2.3.0
5
+ flask_restful >= 0.3.0
6
+ wandb >= 0.15.0
7
+ torch >= 2.0.0
8
+ regex >= 2023.6.0
9
+ numpy >= 1.25
10
+ pillow >= 10.0.0
11
+ einops >= 0.6.1
multilinguality_megatron/ducttape/10B_all_cleaned.tconf ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B/checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # number such that final tokens for each language are around 1B
25
+ n_tokens=(TrainLanguage:
26
+ en=1000000000
27
+ es=833333330
28
+ de=833333330
29
+ fr=833333330
30
+ nl=833333330
31
+ pt=833333330
32
+ it=833333330
33
+ ru=500000000
34
+ zh=13888888
35
+ ko=250000000
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data_2/shared/tower_llm_data/en/data
40
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
41
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
42
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
43
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
44
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
45
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
46
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
47
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
48
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=10
58
+ save_interval=635
59
+ eval_interval=635
60
+ train_steps=6358
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=63
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=8
69
+ gpu_ids=0,1,2,3,4,5,6,7
70
+ tp=(TP: 1 2 3 4)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=12
74
+ vocab_size=32000
75
+
76
+ cpu_workers=16
77
+ wandb_run_id="llama2_7B_10b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198"
78
+ wikipedia=False
79
+ freeze_layers=""
80
+ }
multilinguality_megatron/ducttape/10B_all_cleaned_13B.tconf ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/llama2_13B_all
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/llama2_13B_all/checkpoints
6
+ model_path=/mnt/data/cache/models--meta-llama--Llama-2-13b-hf/snapshots/db6b8eb1feabb38985fdf785a89895959e944936
7
+ tokenizer_path=/mnt/data/cache/models--meta-llama--Llama-2-13b-hf/snapshots/db6b8eb1feabb38985fdf785a89895959e944936/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # number such that final tokens for each language are around 1B
25
+ n_tokens=(TrainLanguage:
26
+ en=1000000000
27
+ es=833333330
28
+ de=833333330
29
+ fr=833333330
30
+ nl=833333330
31
+ pt=833333330
32
+ it=833333330
33
+ ru=500000000
34
+ zh=13888888
35
+ ko=250000000
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data_2/shared/tower_llm_data/en/data
40
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
41
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
42
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
43
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
44
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
45
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
46
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
47
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
48
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=1
58
+ save_interval=10
59
+ eval_interval=635
60
+ train_steps=10
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=0
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=8
69
+ gpu_ids=0,1,2,3,4,5,6,7
70
+ tp=(TP: 1 2 3 4 5 6 7 8)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=12
74
+ vocab_size=32000
75
+
76
+ cpu_workers=16
77
+ wandb_run_id="test_llama_13B"
78
+ wikipedia=False
79
+ freeze_layers=""
80
+ }
multilinguality_megatron/ducttape/10B_all_cleaned_extend32.tconf ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_extend32
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_extend32/checkpoints
6
+ model_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit
7
+ tokenizer_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ posterior_tokens=False
12
+ n_posterior_tokens=False
13
+
14
+ threshold=(TrainLanguage:
15
+ en=516
16
+ es=275
17
+ de=611
18
+ fr=322
19
+ nl=649
20
+ pt=257
21
+ it=332
22
+ ru=334
23
+ zh=2041
24
+ ko=198
25
+ )
26
+
27
+ n_tokens=(TrainLanguage:
28
+ en=900000000
29
+ es=900000000
30
+ de=900000000
31
+ fr=900000000
32
+ nl=900000000
33
+ pt=900000000
34
+ it=900000000
35
+ ru=550000000
36
+ zh=20000000
37
+ ko=450000000
38
+ )
39
+
40
+ dataset_path=(TrainLanguage:
41
+ en=/mnt/data_2/shared/tower_llm_data/en/data
42
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
43
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
44
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
45
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
46
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
47
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
48
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
49
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
50
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
51
+ )
52
+
53
+ mix="10 10 10 10 10 10 10 10 10 10"
54
+
55
+ min_perplexity=50
56
+
57
+ size=(Size: 7 13)
58
+
59
+ log_interval=10
60
+ save_interval=635
61
+ eval_interval=635
62
+ train_steps=6358
63
+
64
+ lr_scheduler=cosine
65
+ warmup_steps=63
66
+ lr=3e-5
67
+ lr_min=3e-6
68
+ weight_decay=0.1
69
+
70
+ n_gpus=8
71
+ gpu_ids=0,1,2,3,4,5,6,7
72
+ tp=(TP: 1 2 3 4)
73
+ pp=(PP: 1 2 3 4)
74
+ micro_batch_size=4
75
+ grad_accum_steps=12
76
+ vocab_size=52620
77
+ eval_iters=1
78
+
79
+ cpu_workers=16
80
+ wandb_run_id="NEW_llama2_7B_10b_extend32_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198"
81
+ wikipedia=False
82
+ freeze_layers=""
83
+ }
multilinguality_megatron/ducttape/10B_all_cleaned_extend32_warmed_up.tconf ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_extend32
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_extend32/warmed_up_checkpoints
6
+ # for warmed up models, the model path points to the sharded megatron checkpoint
7
+ model_path=/mnt/data/shared/multilingual_llm/experiments_megatron/warmup_embeddings_llama2_all_1B_extend32/checkpoints
8
+ tokenizer_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit/tokenizer.model
9
+
10
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
11
+
12
+ wikipedia=False
13
+ posterior_tokens=False
14
+ n_posterior_tokens=False
15
+ freeze_layers=""
16
+
17
+ threshold=(TrainLanguage:
18
+ en=516
19
+ es=275
20
+ de=611
21
+ fr=322
22
+ nl=649
23
+ pt=257
24
+ it=332
25
+ ru=334
26
+ zh=2041
27
+ ko=198
28
+ )
29
+
30
+ n_tokens=(TrainLanguage:
31
+ en=900000000
32
+ es=900000000
33
+ de=900000000
34
+ fr=900000000
35
+ nl=900000000
36
+ pt=900000000
37
+ it=900000000
38
+ ru=550000000
39
+ zh=20000000
40
+ ko=450000000
41
+ )
42
+
43
+ dataset_path=(TrainLanguage:
44
+ en=/mnt/data_2/shared/tower_llm_data/en/data
45
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
46
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
47
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
48
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
49
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
50
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
51
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
52
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
53
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
54
+ )
55
+
56
+ mix="10 10 10 10 10 10 10 10 10 10"
57
+
58
+ min_perplexity=50
59
+
60
+ size=(Size: 7 13)
61
+
62
+ log_interval=10
63
+ save_interval=127
64
+ eval_interval=635
65
+ train_steps=6358
66
+ eval_iters=1
67
+
68
+ lr_scheduler=cosine
69
+ warmup_steps=63
70
+ lr=3e-5
71
+ lr_min=3e-6
72
+ weight_decay=0.1
73
+
74
+ n_gpus=8
75
+ gpu_ids=0,1,2,3,4,5,6,7
76
+ tp=(TP: 1 2 3 4)
77
+ pp=(PP: 1 2 3 4)
78
+ micro_batch_size=4
79
+ grad_accum_steps=12
80
+ vocab_size=52620
81
+
82
+ cpu_workers=16
83
+ wandb_run_id="NEW_warmed_up_llama2_7B_10b_extend32_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198"
84
+ }
multilinguality_megatron/ducttape/10B_all_cleaned_extend32_warmup.tconf ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/warmup_embeddings_llama2_all_1B_extend32/
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/warmup_embeddings_llama2_all_1B_extend32/checkpoints
6
+ model_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit
7
+ tokenizer_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ posterior_tokens=True
25
+ n_tokens=(TrainLanguage:
26
+ en=900000000
27
+ es=900000000
28
+ de=900000000
29
+ fr=900000000
30
+ nl=900000000
31
+ pt=900000000
32
+ it=900000000
33
+ ru=550000000
34
+ zh=20000000
35
+ ko=450000000
36
+ )
37
+ n_posterior_tokens=(TrainLanguage:
38
+ en=180000000
39
+ es=180000000
40
+ de=180000000
41
+ fr=180000000
42
+ nl=180000000
43
+ pt=180000000
44
+ it=180000000
45
+ ru=100000000
46
+ zh=4000000
47
+ ko=90000000
48
+ )
49
+
50
+ dataset_path=(TrainLanguage:
51
+ en=/mnt/data_2/shared/tower_llm_data/en/data
52
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
53
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
54
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
55
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
56
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
57
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
58
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
59
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
60
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
61
+ )
62
+
63
+ mix="10 10 10 10 10 10 10 10 10 10"
64
+
65
+ min_perplexity=50
66
+
67
+ size=(Size: 7 13)
68
+
69
+ log_interval=5
70
+ save_interval=635
71
+ eval_interval=635
72
+ train_steps=635
73
+ eval_iters=0
74
+
75
+ lr_scheduler=constant
76
+ warmup_steps=0
77
+ lr=3e-4
78
+ lr_min=3e-4
79
+ weight_decay=0.1
80
+
81
+ n_gpus=8
82
+ gpu_ids=0,1,2,3,4,5,6,7
83
+ tp=(TP: 1 2 3 4)
84
+ pp=(PP: 1 2 3 4)
85
+ micro_batch_size=4
86
+ grad_accum_steps=12
87
+ vocab_size=52620
88
+
89
+ cpu_workers=16
90
+ wandb_run_id="NEW_EMBEDDINGS_ONLY_llama2_7B_10b_extend32_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198"
91
+ wikipedia=False
92
+ freeze_layers="not_embeddings"
93
+ }
multilinguality_megatron/ducttape/10B_all_wikipedia.tconf ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/wikipedia_llama2_all_10B/
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/wikipedia_llama2_all_10B/checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # number such that final tokens for each language are around 1B
25
+ n_tokens=(TrainLanguage:
26
+ en=1000000000
27
+ es=833333330
28
+ de=833333330
29
+ fr=833333330
30
+ nl=833333330
31
+ pt=833333330
32
+ it=833333330
33
+ ru=500000000
34
+ zh=13888888
35
+ ko=250000000
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/en
40
+ es=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/es
41
+ de=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/de
42
+ fr=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/fr
43
+ nl=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/nl
44
+ pt=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/pt
45
+ it=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/it
46
+ ru=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ru
47
+ zh=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/zh
48
+ ko=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ko
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=10
58
+ save_interval=127
59
+ eval_interval=635
60
+ train_steps=6358
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=63
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=4
69
+ gpu_ids=0,1,2,3
70
+ tp=(TP: 1 2 3 4)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=24
74
+ vocab_size=32000
75
+
76
+ cpu_workers=16
77
+ wandb_run_id="WIKIPEDIA_llama2_7B_10b_base_vocab_uniform"
78
+ wikipedia=True
79
+ freeze_layers=""
80
+ posterior_tokens=False
81
+ n_posterior_tokens=False
82
+ eval_iters=0
83
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4.tconf ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/mc4_checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ )
23
+
24
+ is_hf_dataset=(Dataset:
25
+ en=True
26
+ es=False
27
+ de=False
28
+ fr=False
29
+ nl=False
30
+ pt=False
31
+ it=False
32
+ ru=False
33
+ zh=False
34
+ ko=False
35
+ )
36
+
37
+ threshold=(Dataset:
38
+ en=516
39
+ es=275
40
+ de=611
41
+ fr=322
42
+ nl=649
43
+ pt=257
44
+ it=332
45
+ ru=334
46
+ zh=2041
47
+ ko=198
48
+ )
49
+
50
+ datamix_weights=(
51
+ DataMix:
52
+ mc4_uniform=(
53
+ Dataset:
54
+ en=100
55
+ es=100
56
+ de=100
57
+ fr=100
58
+ nl=100
59
+ pt=100
60
+ it=100
61
+ ru=100
62
+ zh=100
63
+ ko=100
64
+ )
65
+ )
66
+
67
+ # number such that final tokens for each language are around 1B
68
+ n_tokens=(Dataset:
69
+ en=1000000000
70
+ es=833333330
71
+ de=833333330
72
+ fr=833333330
73
+ nl=833333330
74
+ pt=833333330
75
+ it=833333330
76
+ ru=500000000
77
+ zh=13888888
78
+ ko=250000000
79
+ )
80
+
81
+ min_perplexity=50
82
+
83
+ size=(Size: 7 13)
84
+
85
+ log_interval=1
86
+ save_interval=635
87
+ eval_interval=635
88
+ train_steps=12700
89
+
90
+ lr_scheduler=cosine
91
+ warmup_steps=127
92
+ lr=3e-5
93
+ lr_min=3e-6
94
+ weight_decay=0.1
95
+
96
+ n_gpus=8
97
+ gpu_ids=0,1,2,3,4,5,6,7
98
+ tp=(TP: 1 2 3 4)
99
+ pp=(PP: 1 2 3 4)
100
+ micro_batch_size=4
101
+ grad_accum_steps=12
102
+ vocab_size=32000
103
+
104
+ cpu_workers=16
105
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_wiki_33"
106
+ wikipedia=False
107
+ freeze_layers=""
108
+ posterior_tokens=False
109
+ n_posterior_tokens=False
110
+ eval_iters=1
111
+ is_parallel=False
112
+ lp=""
113
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel.tconf ADDED
@@ -0,0 +1,264 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/mc4_parallel_checkpoints
6
+ model_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852
7
+ tokenizer_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
23
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
24
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ )
41
+
42
+ is_hf_dataset=(Dataset:
43
+ en=True
44
+ es=False
45
+ de=False
46
+ fr=False
47
+ nl=False
48
+ pt=False
49
+ it=False
50
+ ru=False
51
+ zh=False
52
+ ko=False
53
+ en_de=False
54
+ de_en=False
55
+ en_fr=False
56
+ fr_en=False
57
+ en_es=False
58
+ es_en=False
59
+ en_it=False
60
+ it_en=False
61
+ en_nl=False
62
+ nl_en=False
63
+ en_pt=False
64
+ pt_en=False
65
+ en_ru=False
66
+ ru_en=False
67
+ en_zh=False
68
+ zh_en=False
69
+ en_ko=False
70
+ ko_en=False
71
+ )
72
+
73
+ threshold=(Dataset:
74
+ en=516
75
+ es=275
76
+ de=611
77
+ fr=322
78
+ nl=649
79
+ pt=257
80
+ it=332
81
+ ru=334
82
+ zh=2041
83
+ ko=198
84
+ en_de=100000
85
+ de_en=100000
86
+ en_fr=100000
87
+ fr_en=100000
88
+ en_es=100000
89
+ es_en=100000
90
+ en_it=100000
91
+ it_en=100000
92
+ en_nl=100000
93
+ nl_en=100000
94
+ en_pt=100000
95
+ pt_en=100000
96
+ en_ru=100000
97
+ ru_en=100000
98
+ en_zh=100000
99
+ zh_en=100000
100
+ en_ko=100000
101
+ ko_en=100000
102
+ )
103
+
104
+ # rougly 67% for mc4, 33% for total parallel data
105
+ datamix_weights=(
106
+ DataMix:
107
+ mc4_parallel_uniform=(
108
+ Dataset:
109
+ en=670
110
+ es=670
111
+ de=670
112
+ fr=670
113
+ nl=670
114
+ pt=670
115
+ it=670
116
+ ru=670
117
+ zh=670
118
+ ko=670
119
+ en_de=183
120
+ de_en=183
121
+ en_fr=183
122
+ fr_en=183
123
+ en_es=183
124
+ es_en=183
125
+ en_it=183
126
+ it_en=183
127
+ en_nl=183
128
+ nl_en=183
129
+ en_pt=183
130
+ pt_en=183
131
+ en_ru=183
132
+ ru_en=183
133
+ en_zh=183
134
+ zh_en=183
135
+ en_ko=183
136
+ ko_en=183
137
+ )
138
+ )
139
+
140
+ # number such that final tokens for each language are around 1B
141
+ n_tokens=(Dataset:
142
+ en=1000000000
143
+ es=833333330
144
+ de=833333330
145
+ fr=833333330
146
+ nl=833333330
147
+ pt=833333330
148
+ it=833333330
149
+ ru=500000000
150
+ zh=13888888
151
+ ko=250000000
152
+ en_de=20000000
153
+ de_en=20000000
154
+ en_fr=20000000
155
+ fr_en=20000000
156
+ en_es=20000000
157
+ es_en=20000000
158
+ en_it=20000000
159
+ it_en=20000000
160
+ en_nl=20000000
161
+ nl_en=20000000
162
+ en_pt=20000000
163
+ pt_en=20000000
164
+ en_ru=20000000
165
+ ru_en=20000000
166
+ en_zh=20000000
167
+ zh_en=20000000
168
+ en_ko=20000000
169
+ ko_en=20000000
170
+ )
171
+
172
+ is_parallel=(Dataset:
173
+ en=False
174
+ es=False
175
+ de=False
176
+ fr=False
177
+ nl=False
178
+ pt=False
179
+ it=False
180
+ ru=False
181
+ zh=False
182
+ ko=False
183
+ en_de=True
184
+ de_en=True
185
+ en_fr=True
186
+ fr_en=True
187
+ en_es=True
188
+ es_en=True
189
+ en_it=True
190
+ it_en=True
191
+ en_nl=True
192
+ nl_en=True
193
+ en_pt=True
194
+ pt_en=True
195
+ en_ru=True
196
+ ru_en=True
197
+ en_zh=True
198
+ zh_en=True
199
+ en_ko=True
200
+ ko_en=True
201
+ )
202
+
203
+ lp=(Dataset:
204
+ en=""
205
+ es=""
206
+ de=""
207
+ fr=""
208
+ nl=""
209
+ pt=""
210
+ it=""
211
+ ru=""
212
+ zh=""
213
+ ko=""
214
+ en_de="en-de"
215
+ de_en="de-en"
216
+ en_fr="en-fr"
217
+ fr_en="fr-en"
218
+ en_es="en-es"
219
+ es_en="es-en"
220
+ en_it="en-it"
221
+ it_en="it-en"
222
+ en_nl="en-nl"
223
+ nl_en="nl-en"
224
+ en_pt="en-pt"
225
+ pt_en="pt-en"
226
+ en_ru="en-ru"
227
+ ru_en="ru-en"
228
+ en_zh="en-zh"
229
+ zh_en="zh-en"
230
+ en_ko="en-ko"
231
+ ko_en="ko-en"
232
+ )
233
+
234
+ min_perplexity=50
235
+
236
+ size=(Size: 7 13)
237
+
238
+ log_interval=1
239
+ save_interval=635
240
+ eval_interval=635
241
+ train_steps=12700
242
+
243
+ lr_scheduler=cosine
244
+ warmup_steps=127
245
+ lr=3e-5
246
+ lr_min=3e-6
247
+ weight_decay=0.1
248
+
249
+ n_gpus=8
250
+ gpu_ids=0,1,2,3,4,5,6,7
251
+ tp=(TP: 1 2 3 4)
252
+ pp=(PP: 1 2 3 4)
253
+ micro_batch_size=4
254
+ grad_accum_steps=12
255
+ vocab_size=32000
256
+
257
+ cpu_workers=16
258
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
259
+ wikipedia=False
260
+ freeze_layers=""
261
+ posterior_tokens=False
262
+ n_posterior_tokens=0
263
+ eval_iters=1
264
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_13b.tconf ADDED
@@ -0,0 +1,264 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_13B_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_13B_all_20B/mc4_parallel_checkpoints
6
+ model_path=/mnt/data/cache/models--meta-llama--Llama-2-13b-hf/snapshots/dc1d3b3bfdb69df26f8fc966c16353274b138c55
7
+ tokenizer_path=/mnt/data/cache/models--meta-llama--Llama-2-13b-hf/snapshots/dc1d3b3bfdb69df26f8fc966c16353274b138c55/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
23
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
24
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ )
41
+
42
+ is_hf_dataset=(Dataset:
43
+ en=True
44
+ es=False
45
+ de=False
46
+ fr=False
47
+ nl=False
48
+ pt=False
49
+ it=False
50
+ ru=False
51
+ zh=False
52
+ ko=False
53
+ en_de=False
54
+ de_en=False
55
+ en_fr=False
56
+ fr_en=False
57
+ en_es=False
58
+ es_en=False
59
+ en_it=False
60
+ it_en=False
61
+ en_nl=False
62
+ nl_en=False
63
+ en_pt=False
64
+ pt_en=False
65
+ en_ru=False
66
+ ru_en=False
67
+ en_zh=False
68
+ zh_en=False
69
+ en_ko=False
70
+ ko_en=False
71
+ )
72
+
73
+ threshold=(Dataset:
74
+ en=516
75
+ es=275
76
+ de=611
77
+ fr=322
78
+ nl=649
79
+ pt=257
80
+ it=332
81
+ ru=334
82
+ zh=2041
83
+ ko=198
84
+ en_de=100000
85
+ de_en=100000
86
+ en_fr=100000
87
+ fr_en=100000
88
+ en_es=100000
89
+ es_en=100000
90
+ en_it=100000
91
+ it_en=100000
92
+ en_nl=100000
93
+ nl_en=100000
94
+ en_pt=100000
95
+ pt_en=100000
96
+ en_ru=100000
97
+ ru_en=100000
98
+ en_zh=100000
99
+ zh_en=100000
100
+ en_ko=100000
101
+ ko_en=100000
102
+ )
103
+
104
+ # rougly 67% for mc4, 33% for total parallel data
105
+ datamix_weights=(
106
+ DataMix:
107
+ mc4_parallel_uniform=(
108
+ Dataset:
109
+ en=670
110
+ es=670
111
+ de=670
112
+ fr=670
113
+ nl=670
114
+ pt=670
115
+ it=670
116
+ ru=670
117
+ zh=670
118
+ ko=670
119
+ en_de=183
120
+ de_en=183
121
+ en_fr=183
122
+ fr_en=183
123
+ en_es=183
124
+ es_en=183
125
+ en_it=183
126
+ it_en=183
127
+ en_nl=183
128
+ nl_en=183
129
+ en_pt=183
130
+ pt_en=183
131
+ en_ru=183
132
+ ru_en=183
133
+ en_zh=183
134
+ zh_en=183
135
+ en_ko=183
136
+ ko_en=183
137
+ )
138
+ )
139
+
140
+ # number such that final tokens for each language are around 1B
141
+ n_tokens=(Dataset:
142
+ en=1000000000
143
+ es=833333330
144
+ de=833333330
145
+ fr=833333330
146
+ nl=833333330
147
+ pt=833333330
148
+ it=833333330
149
+ ru=500000000
150
+ zh=13888888
151
+ ko=250000000
152
+ en_de=20000000
153
+ de_en=20000000
154
+ en_fr=20000000
155
+ fr_en=20000000
156
+ en_es=20000000
157
+ es_en=20000000
158
+ en_it=20000000
159
+ it_en=20000000
160
+ en_nl=20000000
161
+ nl_en=20000000
162
+ en_pt=20000000
163
+ pt_en=20000000
164
+ en_ru=20000000
165
+ ru_en=20000000
166
+ en_zh=20000000
167
+ zh_en=20000000
168
+ en_ko=20000000
169
+ ko_en=20000000
170
+ )
171
+
172
+ is_parallel=(Dataset:
173
+ en=False
174
+ es=False
175
+ de=False
176
+ fr=False
177
+ nl=False
178
+ pt=False
179
+ it=False
180
+ ru=False
181
+ zh=False
182
+ ko=False
183
+ en_de=True
184
+ de_en=True
185
+ en_fr=True
186
+ fr_en=True
187
+ en_es=True
188
+ es_en=True
189
+ en_it=True
190
+ it_en=True
191
+ en_nl=True
192
+ nl_en=True
193
+ en_pt=True
194
+ pt_en=True
195
+ en_ru=True
196
+ ru_en=True
197
+ en_zh=True
198
+ zh_en=True
199
+ en_ko=True
200
+ ko_en=True
201
+ )
202
+
203
+ lp=(Dataset:
204
+ en=""
205
+ es=""
206
+ de=""
207
+ fr=""
208
+ nl=""
209
+ pt=""
210
+ it=""
211
+ ru=""
212
+ zh=""
213
+ ko=""
214
+ en_de="en-de"
215
+ de_en="de-en"
216
+ en_fr="en-fr"
217
+ fr_en="fr-en"
218
+ en_es="en-es"
219
+ es_en="es-en"
220
+ en_it="en-it"
221
+ it_en="it-en"
222
+ en_nl="en-nl"
223
+ nl_en="nl-en"
224
+ en_pt="en-pt"
225
+ pt_en="pt-en"
226
+ en_ru="en-ru"
227
+ ru_en="ru-en"
228
+ en_zh="en-zh"
229
+ zh_en="zh-en"
230
+ en_ko="en-ko"
231
+ ko_en="ko-en"
232
+ )
233
+
234
+ min_perplexity=50
235
+
236
+ size=(Size: 7 13)
237
+
238
+ log_interval=1
239
+ save_interval=635
240
+ eval_interval=635
241
+ train_steps=12700
242
+
243
+ lr_scheduler=cosine
244
+ warmup_steps=127
245
+ lr=3e-5
246
+ lr_min=3e-6
247
+ weight_decay=0.1
248
+
249
+ n_gpus=8
250
+ gpu_ids=0,1,2,3,4,5,6,7
251
+ tp=(TP: 1 2 3 4 5 6 7 8)
252
+ pp=(PP: 1 2 3 4)
253
+ micro_batch_size=4
254
+ grad_accum_steps=12
255
+ vocab_size=32000
256
+
257
+ cpu_workers=16
258
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
259
+ wikipedia=False
260
+ freeze_layers=""
261
+ posterior_tokens=False
262
+ n_posterior_tokens=0
263
+ eval_iters=1
264
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_concat.tconf ADDED
@@ -0,0 +1,264 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/mc4_parallel_concat_checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
23
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
24
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ )
41
+
42
+ is_hf_dataset=(Dataset:
43
+ en=True
44
+ es=False
45
+ de=False
46
+ fr=False
47
+ nl=False
48
+ pt=False
49
+ it=False
50
+ ru=False
51
+ zh=False
52
+ ko=False
53
+ en_de=False
54
+ de_en=False
55
+ en_fr=False
56
+ fr_en=False
57
+ en_es=False
58
+ es_en=False
59
+ en_it=False
60
+ it_en=False
61
+ en_nl=False
62
+ nl_en=False
63
+ en_pt=False
64
+ pt_en=False
65
+ en_ru=False
66
+ ru_en=False
67
+ en_zh=False
68
+ zh_en=False
69
+ en_ko=False
70
+ ko_en=False
71
+ )
72
+
73
+ threshold=(Dataset:
74
+ en=516
75
+ es=275
76
+ de=611
77
+ fr=322
78
+ nl=649
79
+ pt=257
80
+ it=332
81
+ ru=334
82
+ zh=2041
83
+ ko=198
84
+ en_de=100000
85
+ de_en=100000
86
+ en_fr=100000
87
+ fr_en=100000
88
+ en_es=100000
89
+ es_en=100000
90
+ en_it=100000
91
+ it_en=100000
92
+ en_nl=100000
93
+ nl_en=100000
94
+ en_pt=100000
95
+ pt_en=100000
96
+ en_ru=100000
97
+ ru_en=100000
98
+ en_zh=100000
99
+ zh_en=100000
100
+ en_ko=100000
101
+ ko_en=100000
102
+ )
103
+
104
+ # rougly 67% for mc4, 33% for total parallel data
105
+ datamix_weights=(
106
+ DataMix:
107
+ mc4_parallel_uniform=(
108
+ Dataset:
109
+ en=670
110
+ es=670
111
+ de=670
112
+ fr=670
113
+ nl=670
114
+ pt=670
115
+ it=670
116
+ ru=670
117
+ zh=670
118
+ ko=670
119
+ en_de=183
120
+ de_en=183
121
+ en_fr=183
122
+ fr_en=183
123
+ en_es=183
124
+ es_en=183
125
+ en_it=183
126
+ it_en=183
127
+ en_nl=183
128
+ nl_en=183
129
+ en_pt=183
130
+ pt_en=183
131
+ en_ru=183
132
+ ru_en=183
133
+ en_zh=183
134
+ zh_en=183
135
+ en_ko=183
136
+ ko_en=183
137
+ )
138
+ )
139
+
140
+ # number such that final tokens for each language are around 1B
141
+ n_tokens=(Dataset:
142
+ en=1000000000
143
+ es=833333330
144
+ de=833333330
145
+ fr=833333330
146
+ nl=833333330
147
+ pt=833333330
148
+ it=833333330
149
+ ru=500000000
150
+ zh=13888888
151
+ ko=250000000
152
+ en_de=20000000
153
+ de_en=20000000
154
+ en_fr=20000000
155
+ fr_en=20000000
156
+ en_es=20000000
157
+ es_en=20000000
158
+ en_it=20000000
159
+ it_en=20000000
160
+ en_nl=20000000
161
+ nl_en=20000000
162
+ en_pt=20000000
163
+ pt_en=20000000
164
+ en_ru=20000000
165
+ ru_en=20000000
166
+ en_zh=20000000
167
+ zh_en=20000000
168
+ en_ko=20000000
169
+ ko_en=20000000
170
+ )
171
+
172
+ is_parallel=(Dataset:
173
+ en=False
174
+ es=False
175
+ de=False
176
+ fr=False
177
+ nl=False
178
+ pt=False
179
+ it=False
180
+ ru=False
181
+ zh=False
182
+ ko=False
183
+ en_de=True
184
+ de_en=True
185
+ en_fr=True
186
+ fr_en=True
187
+ en_es=True
188
+ es_en=True
189
+ en_it=True
190
+ it_en=True
191
+ en_nl=True
192
+ nl_en=True
193
+ en_pt=True
194
+ pt_en=True
195
+ en_ru=True
196
+ ru_en=True
197
+ en_zh=True
198
+ zh_en=True
199
+ en_ko=True
200
+ ko_en=True
201
+ )
202
+
203
+ lp=(Dataset:
204
+ en=""
205
+ es=""
206
+ de=""
207
+ fr=""
208
+ nl=""
209
+ pt=""
210
+ it=""
211
+ ru=""
212
+ zh=""
213
+ ko=""
214
+ en_de="en-de"
215
+ de_en="de-en"
216
+ en_fr="en-fr"
217
+ fr_en="fr-en"
218
+ en_es="en-es"
219
+ es_en="es-en"
220
+ en_it="en-it"
221
+ it_en="it-en"
222
+ en_nl="en-nl"
223
+ nl_en="nl-en"
224
+ en_pt="en-pt"
225
+ pt_en="pt-en"
226
+ en_ru="en-ru"
227
+ ru_en="ru-en"
228
+ en_zh="en-zh"
229
+ zh_en="zh-en"
230
+ en_ko="en-ko"
231
+ ko_en="ko-en"
232
+ )
233
+
234
+ min_perplexity=50
235
+
236
+ size=(Size: 7 13)
237
+
238
+ log_interval=1
239
+ save_interval=635
240
+ eval_interval=635
241
+ train_steps=12700
242
+
243
+ lr_scheduler=cosine
244
+ warmup_steps=127
245
+ lr=3e-5
246
+ lr_min=3e-6
247
+ weight_decay=0.1
248
+
249
+ n_gpus=8
250
+ gpu_ids=0,1,2,3,4,5,6,7
251
+ tp=(TP: 1 2 3 4)
252
+ pp=(PP: 1 2 3 4)
253
+ micro_batch_size=4
254
+ grad_accum_steps=12
255
+ vocab_size=32000
256
+
257
+ cpu_workers=16
258
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
259
+ wikipedia=False
260
+ freeze_layers=""
261
+ posterior_tokens=False
262
+ n_posterior_tokens=0
263
+ eval_iters=1
264
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4_parallel_instructions.tconf ADDED
@@ -0,0 +1,271 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B_w_instructions
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B_w_instructions/mc4_parallel_towerblocksv0.1_checkpoints
6
+ model_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852
7
+ tokenizer_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en tower_blocks)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
23
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
24
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ tower_blocks="Unbabel/TowerBlocks-v0.1"
41
+ )
42
+
43
+ is_hf_dataset=(Dataset:
44
+ en=True
45
+ es=False
46
+ de=False
47
+ fr=False
48
+ nl=False
49
+ pt=False
50
+ it=False
51
+ ru=False
52
+ zh=False
53
+ ko=False
54
+ en_de=False
55
+ de_en=False
56
+ en_fr=False
57
+ fr_en=False
58
+ en_es=False
59
+ es_en=False
60
+ en_it=False
61
+ it_en=False
62
+ en_nl=False
63
+ nl_en=False
64
+ en_pt=False
65
+ pt_en=False
66
+ en_ru=False
67
+ ru_en=False
68
+ en_zh=False
69
+ zh_en=False
70
+ en_ko=False
71
+ ko_en=False
72
+ tower_blocks=True
73
+ )
74
+
75
+ threshold=(Dataset:
76
+ en=516
77
+ es=275
78
+ de=611
79
+ fr=322
80
+ nl=649
81
+ pt=257
82
+ it=332
83
+ ru=334
84
+ zh=2041
85
+ ko=198
86
+ en_de=100000
87
+ de_en=100000
88
+ en_fr=100000
89
+ fr_en=100000
90
+ en_es=100000
91
+ es_en=100000
92
+ en_it=100000
93
+ it_en=100000
94
+ en_nl=100000
95
+ nl_en=100000
96
+ en_pt=100000
97
+ pt_en=100000
98
+ en_ru=100000
99
+ ru_en=100000
100
+ en_zh=100000
101
+ zh_en=100000
102
+ en_ko=100000
103
+ ko_en=100000
104
+ tower_blocks=100000
105
+ )
106
+
107
+ # rougly 67% for mc4, 33% for total parallel data
108
+ datamix_weights=(
109
+ DataMix:
110
+ mc4_parallel_uniform=(
111
+ Dataset:
112
+ en=670
113
+ es=670
114
+ de=670
115
+ fr=670
116
+ nl=670
117
+ pt=670
118
+ it=670
119
+ ru=670
120
+ zh=670
121
+ ko=670
122
+ en_de=183
123
+ de_en=183
124
+ en_fr=183
125
+ fr_en=183
126
+ en_es=183
127
+ es_en=183
128
+ en_it=183
129
+ it_en=183
130
+ en_nl=183
131
+ nl_en=183
132
+ en_pt=183
133
+ pt_en=183
134
+ en_ru=183
135
+ ru_en=183
136
+ en_zh=183
137
+ zh_en=183
138
+ en_ko=183
139
+ ko_en=183
140
+ tower_blocks=183
141
+ )
142
+ )
143
+
144
+ # number such that final tokens for each language are around 1B
145
+ n_tokens=(Dataset:
146
+ en=1000000000
147
+ es=833333330
148
+ de=833333330
149
+ fr=833333330
150
+ nl=833333330
151
+ pt=833333330
152
+ it=833333330
153
+ ru=500000000
154
+ zh=13888888
155
+ ko=250000000
156
+ en_de=20000000
157
+ de_en=20000000
158
+ en_fr=20000000
159
+ fr_en=20000000
160
+ en_es=20000000
161
+ es_en=20000000
162
+ en_it=20000000
163
+ it_en=20000000
164
+ en_nl=20000000
165
+ nl_en=20000000
166
+ en_pt=20000000
167
+ pt_en=20000000
168
+ en_ru=20000000
169
+ ru_en=20000000
170
+ en_zh=20000000
171
+ zh_en=20000000
172
+ en_ko=20000000
173
+ ko_en=20000000
174
+ tower_blocks=20000000
175
+ )
176
+
177
+ is_parallel=(Dataset:
178
+ en=False
179
+ es=False
180
+ de=False
181
+ fr=False
182
+ nl=False
183
+ pt=False
184
+ it=False
185
+ ru=False
186
+ zh=False
187
+ ko=False
188
+ en_de=True
189
+ de_en=True
190
+ en_fr=True
191
+ fr_en=True
192
+ en_es=True
193
+ es_en=True
194
+ en_it=True
195
+ it_en=True
196
+ en_nl=True
197
+ nl_en=True
198
+ en_pt=True
199
+ pt_en=True
200
+ en_ru=True
201
+ ru_en=True
202
+ en_zh=True
203
+ zh_en=True
204
+ en_ko=True
205
+ ko_en=True
206
+ tower_blocks=False
207
+ )
208
+
209
+ lp=(Dataset:
210
+ en=""
211
+ es=""
212
+ de=""
213
+ fr=""
214
+ nl=""
215
+ pt=""
216
+ it=""
217
+ ru=""
218
+ zh=""
219
+ ko=""
220
+ en_de="en-de"
221
+ de_en="de-en"
222
+ en_fr="en-fr"
223
+ fr_en="fr-en"
224
+ en_es="en-es"
225
+ es_en="es-en"
226
+ en_it="en-it"
227
+ it_en="it-en"
228
+ en_nl="en-nl"
229
+ nl_en="nl-en"
230
+ en_pt="en-pt"
231
+ pt_en="pt-en"
232
+ en_ru="en-ru"
233
+ ru_en="ru-en"
234
+ en_zh="en-zh"
235
+ zh_en="zh-en"
236
+ en_ko="en-ko"
237
+ ko_en="ko-en"
238
+ tower_blocks="oi"
239
+ )
240
+
241
+ min_perplexity=50
242
+
243
+ size=(Size: 7 13)
244
+
245
+ log_interval=1
246
+ save_interval=635
247
+ eval_interval=635
248
+ train_steps=12700
249
+
250
+ lr_scheduler=cosine
251
+ warmup_steps=127
252
+ lr=3e-5
253
+ lr_min=3e-6
254
+ weight_decay=0.1
255
+
256
+ n_gpus=8
257
+ gpu_ids=0,1,2,3,4,5,6,7
258
+ tp=(TP: 1 2 3 4)
259
+ pp=(PP: 1 2 3 4)
260
+ micro_batch_size=4
261
+ grad_accum_steps=12
262
+ vocab_size=32000
263
+
264
+ cpu_workers=16
265
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
266
+ wikipedia=False
267
+ freeze_layers=""
268
+ posterior_tokens=False
269
+ n_posterior_tokens=0
270
+ eval_iters=1
271
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_mc4_wiki.tconf ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/mc4_wiki_checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_wiki de_wiki fr_wiki es_wiki it_wiki nl_wiki pt_wiki ru_wiki zh_wiki ko_wiki)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ en_wiki=""
23
+ es_wiki=""
24
+ de_wiki=""
25
+ fr_wiki=""
26
+ nl_wiki=""
27
+ pt_wiki=""
28
+ it_wiki=""
29
+ ru_wiki=""
30
+ zh_wiki=""
31
+ ko_wiki=""
32
+ )
33
+
34
+ is_hf_dataset=(Dataset:
35
+ en=True
36
+ es=False
37
+ de=False
38
+ fr=False
39
+ nl=False
40
+ pt=False
41
+ it=False
42
+ ru=False
43
+ zh=False
44
+ ko=False
45
+ en_wiki=False
46
+ es_wiki=False
47
+ de_wiki=False
48
+ fr_wiki=False
49
+ nl_wiki=False
50
+ pt_wiki=False
51
+ it_wiki=False
52
+ ru_wiki=False
53
+ zh_wiki=False
54
+ ko_wiki=False
55
+ )
56
+
57
+ threshold=(Dataset:
58
+ en=516 en_wiki=""
59
+ es=275 es_wiki=""
60
+ de=611 de_wiki=""
61
+ fr=322 fr_wiki=""
62
+ nl=649 nl_wiki=""
63
+ pt=257 pt_wiki=""
64
+ it=332 it_wiki=""
65
+ ru=334 ru_wiki=""
66
+ zh=2041 zh_wiki=""
67
+ ko=198 ko_wiki=""
68
+ )
69
+
70
+ datamix_weights=(
71
+ DataMix:
72
+ mc4_wiki_uniform=(
73
+ Dataset:
74
+ en=67
75
+ es=67
76
+ de=67
77
+ fr=67
78
+ nl=67
79
+ pt=67
80
+ it=67
81
+ ru=67
82
+ zh=67
83
+ ko=67
84
+ en_wiki=33
85
+ es_wiki=33
86
+ de_wiki=33
87
+ fr_wiki=33
88
+ nl_wiki=33
89
+ pt_wiki=33
90
+ it_wiki=33
91
+ ru_wiki=33
92
+ zh_wiki=33
93
+ ko_wiki=33
94
+ )
95
+ mc4_uniform=(
96
+ Dataset:
97
+ en=100
98
+ es=100
99
+ de=100
100
+ fr=100
101
+ nl=100
102
+ pt=100
103
+ it=100
104
+ ru=100
105
+ zh=100
106
+ ko=100
107
+ en_wiki=0
108
+ es_wiki=0
109
+ de_wiki=0
110
+ fr_wiki=0
111
+ nl_wiki=0
112
+ pt_wiki=0
113
+ it_wiki=0
114
+ ru_wiki=0
115
+ zh_wiki=0
116
+ ko_wiki=0
117
+ )
118
+ )
119
+
120
+ # number such that final tokens for each language are around 1B
121
+ n_tokens=(Dataset:
122
+ en=1000000000
123
+ es=833333330
124
+ de=833333330
125
+ fr=833333330
126
+ nl=833333330
127
+ pt=833333330
128
+ it=833333330
129
+ ru=500000000
130
+ zh=13888888
131
+ ko=250000000
132
+ en_wiki=""
133
+ es_wiki=""
134
+ de_wiki=""
135
+ fr_wiki=""
136
+ nl_wiki=""
137
+ pt_wiki=""
138
+ it_wiki=""
139
+ ru_wiki=""
140
+ zh_wiki=""
141
+ ko_wiki=""
142
+ )
143
+
144
+ min_perplexity=50
145
+
146
+ size=(Size: 7 13)
147
+
148
+ log_interval=1
149
+ save_interval=635
150
+ eval_interval=635
151
+ train_steps=12700
152
+
153
+ lr_scheduler=cosine
154
+ warmup_steps=127
155
+ lr=3e-5
156
+ lr_min=3e-6
157
+ weight_decay=0.1
158
+
159
+ n_gpus=8
160
+ gpu_ids=0,1,2,3,4,5,6,7
161
+ tp=(TP: 1 2 3 4)
162
+ pp=(PP: 1 2 3 4)
163
+ micro_batch_size=4
164
+ grad_accum_steps=12
165
+ vocab_size=32000
166
+
167
+ cpu_workers=16
168
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_wiki_33"
169
+ wikipedia=False
170
+ freeze_layers=""
171
+ posterior_tokens=False
172
+ n_posterior_tokens=False
173
+ eval_iters=1
174
+ is_parallel=False
175
+ lp=""
176
+ }
multilinguality_megatron/ducttape/20B_all_cleaned_parallel.tconf ADDED
@@ -0,0 +1,194 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/parallel_checkpoints
6
+ model_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852
7
+ tokenizer_path=/mnt/data/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852/tokenizer.model
8
+
9
+ dataset=(Dataset: en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
10
+
11
+ dataset_path=(Dataset:
12
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
13
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
14
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
15
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
16
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
17
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
18
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
19
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
20
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
21
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
22
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
23
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
24
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ )
31
+
32
+ is_hf_dataset=(Dataset:
33
+ en_de=False
34
+ de_en=False
35
+ en_fr=False
36
+ fr_en=False
37
+ en_es=False
38
+ es_en=False
39
+ en_it=False
40
+ it_en=False
41
+ en_nl=False
42
+ nl_en=False
43
+ en_pt=False
44
+ pt_en=False
45
+ en_ru=False
46
+ ru_en=False
47
+ en_zh=False
48
+ zh_en=False
49
+ en_ko=False
50
+ ko_en=False
51
+ )
52
+
53
+ threshold=(Dataset:
54
+ en_de=100000
55
+ de_en=100000
56
+ en_fr=100000
57
+ fr_en=100000
58
+ en_es=100000
59
+ es_en=100000
60
+ en_it=100000
61
+ it_en=100000
62
+ en_nl=100000
63
+ nl_en=100000
64
+ en_pt=100000
65
+ pt_en=100000
66
+ en_ru=100000
67
+ ru_en=100000
68
+ en_zh=100000
69
+ zh_en=100000
70
+ en_ko=100000
71
+ ko_en=100000
72
+ )
73
+
74
+ # rougly 67% for mc4, 33% for total parallel data
75
+ datamix_weights=(
76
+ DataMix:
77
+ mc4_parallel_uniform=(
78
+ Dataset:
79
+ en_de=1
80
+ de_en=1
81
+ en_fr=1
82
+ fr_en=1
83
+ en_es=1
84
+ es_en=1
85
+ en_it=1
86
+ it_en=1
87
+ en_nl=1
88
+ nl_en=1
89
+ en_pt=1
90
+ pt_en=1
91
+ en_ru=1
92
+ ru_en=1
93
+ en_zh=1
94
+ zh_en=1
95
+ en_ko=1
96
+ ko_en=1
97
+ )
98
+ )
99
+
100
+ # number such that final tokens for each language are around 1B
101
+ n_tokens=(Dataset:
102
+ en_de=20000000
103
+ de_en=20000000
104
+ en_fr=20000000
105
+ fr_en=20000000
106
+ en_es=20000000
107
+ es_en=20000000
108
+ en_it=20000000
109
+ it_en=20000000
110
+ en_nl=20000000
111
+ nl_en=20000000
112
+ en_pt=20000000
113
+ pt_en=20000000
114
+ en_ru=20000000
115
+ ru_en=20000000
116
+ en_zh=20000000
117
+ zh_en=20000000
118
+ en_ko=20000000
119
+ ko_en=20000000
120
+ )
121
+
122
+ is_parallel=(Dataset:
123
+ en_de=True
124
+ de_en=True
125
+ en_fr=True
126
+ fr_en=True
127
+ en_es=True
128
+ es_en=True
129
+ en_it=True
130
+ it_en=True
131
+ en_nl=True
132
+ nl_en=True
133
+ en_pt=True
134
+ pt_en=True
135
+ en_ru=True
136
+ ru_en=True
137
+ en_zh=True
138
+ zh_en=True
139
+ en_ko=True
140
+ ko_en=True
141
+ )
142
+
143
+ lp=(Dataset:
144
+ en_de="en-de"
145
+ de_en="de-en"
146
+ en_fr="en-fr"
147
+ fr_en="fr-en"
148
+ en_es="en-es"
149
+ es_en="es-en"
150
+ en_it="en-it"
151
+ it_en="it-en"
152
+ en_nl="en-nl"
153
+ nl_en="nl-en"
154
+ en_pt="en-pt"
155
+ pt_en="pt-en"
156
+ en_ru="en-ru"
157
+ ru_en="ru-en"
158
+ en_zh="en-zh"
159
+ zh_en="zh-en"
160
+ en_ko="en-ko"
161
+ ko_en="ko-en"
162
+ )
163
+
164
+ min_perplexity=50
165
+
166
+ size=(Size: 7 13)
167
+
168
+ log_interval=1
169
+ save_interval=635
170
+ eval_interval=635
171
+ train_steps=12700
172
+
173
+ lr_scheduler=cosine
174
+ warmup_steps=127
175
+ lr=3e-5
176
+ lr_min=3e-6
177
+ weight_decay=0.1
178
+
179
+ n_gpus=8
180
+ gpu_ids=0,1,2,3,4,5,6,7
181
+ tp=(TP: 1 2 3 4)
182
+ pp=(PP: 1 2 3 4)
183
+ micro_batch_size=4
184
+ grad_accum_steps=12
185
+ vocab_size=32000
186
+
187
+ cpu_workers=16
188
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
189
+ wikipedia=False
190
+ freeze_layers=""
191
+ posterior_tokens=False
192
+ n_posterior_tokens=0
193
+ eval_iters=1
194
+ }
multilinguality_megatron/ducttape/20B_all_dirty_mc4.tconf ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_20B/dirty_mc4_checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/pre-training/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/pre-training/tower_llm_data/es/0/0000.json.gz
14
+ de=/mnt/data_2/shared/pre-training/tower_llm_data/de/0/0000.json.gz
15
+ fr=/mnt/data_2/shared/pre-training/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/pre-training/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/pre-training/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/pre-training/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/pre-training/tower_llm_data/ru/0/0000.json.gz
20
+ zh=/mnt/data_2/shared/pre-training/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/pre-training/tower_llm_data/ko/0000.json.gz
22
+ )
23
+
24
+ is_hf_dataset=(Dataset:
25
+ en=True
26
+ es=False
27
+ de=False
28
+ fr=False
29
+ nl=False
30
+ pt=False
31
+ it=False
32
+ ru=False
33
+ zh=False
34
+ ko=False
35
+ )
36
+
37
+ threshold=(Dataset:
38
+ en=10000000
39
+ es=10000000
40
+ de=10000000
41
+ fr=10000000
42
+ nl=10000000
43
+ pt=10000000
44
+ it=10000000
45
+ ru=10000000
46
+ zh=10000000
47
+ ko=10000000
48
+ )
49
+
50
+ datamix_weights=(
51
+ DataMix:
52
+ mc4_uniform=(
53
+ Dataset:
54
+ en=100
55
+ es=100
56
+ de=100
57
+ fr=100
58
+ nl=100
59
+ pt=100
60
+ it=100
61
+ ru=100
62
+ zh=100
63
+ ko=100
64
+ )
65
+ )
66
+
67
+ # number such that final tokens for each language are around 1B
68
+ n_tokens=(Dataset:
69
+ en=1000000000
70
+ es=833333330
71
+ de=833333330
72
+ fr=833333330
73
+ nl=833333330
74
+ pt=833333330
75
+ it=833333330
76
+ ru=500000000
77
+ zh=1388888800
78
+ ko=250000000
79
+ )
80
+
81
+ min_perplexity=0
82
+
83
+ size=(Size: 7 13)
84
+
85
+ log_interval=1
86
+ save_interval=635
87
+ eval_interval=635
88
+ train_steps=12700
89
+
90
+ lr_scheduler=cosine
91
+ warmup_steps=127
92
+ lr=3e-5
93
+ lr_min=3e-6
94
+ weight_decay=0.1
95
+
96
+ n_gpus=8
97
+ gpu_ids=0,1,2,3,4,5,6,7
98
+ tp=(TP: 1 2 3 4)
99
+ pp=(PP: 1 2 3 4)
100
+ micro_batch_size=4
101
+ grad_accum_steps=12
102
+ vocab_size=32000
103
+
104
+ cpu_workers=16
105
+ wandb_run_id="llama2_7B_20b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_wiki_33"
106
+ wikipedia=False
107
+ freeze_layers=""
108
+ posterior_tokens=False
109
+ n_posterior_tokens=0
110
+ eval_iters=1
111
+ is_parallel=False
112
+ lp=(Dataset:
113
+ en="en"
114
+ es="es"
115
+ de="de"
116
+ fr="fr"
117
+ nl="nl"
118
+ pt="pt"
119
+ it="it"
120
+ ru="ru"
121
+ zh="zh"
122
+ ko="ko"
123
+ )
124
+ }
multilinguality_megatron/ducttape/40B_all_cleaned_mc4_parallel.tconf ADDED
@@ -0,0 +1,305 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_40B
3
+ repo=/mnt/data/pmartins/code/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_40B/mc4_parallel_checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ dataset=(Dataset: en de fr es it nl pt ru zh ko pl sv en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en en_pl pl_en en_sv sv_en)
10
+
11
+ dataset_path=(Dataset:
12
+ en=/mnt/data_2/shared/tower_llm_data/en/data
13
+ es=/mnt/data_2/shared/tower_llm_data/es/0/0000.json.gz
14
+ de=/mnt/data_2/shared/tower_llm_data/de/0/0000.json.gz
15
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
16
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
17
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
18
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
19
+ ru=/mnt/data_2/shared/tower_llm_data/ru/0/0000.json.gz
20
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
21
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
22
+ pl=/mnt/data_2/shared/tower_llm_data/pl/0000.json.gz
23
+ sv=/mnt/data_2/shared/tower_llm_data/sv/0000.json.gz
24
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
25
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
26
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
27
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ en_pl="/mnt/data_2/shared/tower_llm_data/bilingual_data/en-pl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ pl_en="/mnt/data_2/shared/tower_llm_data/bilingual_data/en-pl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75/"
44
+ en_sv="/mnt/data_2/shared/tower_llm_data/bilingual_data/en-sv/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75/"
45
+ sv_en="/mnt/data_2/shared/tower_llm_data/bilingual_data/en-sv/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75/"
46
+ )
47
+
48
+ is_hf_dataset=(Dataset:
49
+ en=True
50
+ es=False
51
+ de=False
52
+ fr=False
53
+ nl=False
54
+ pt=False
55
+ it=False
56
+ ru=False
57
+ zh=False
58
+ ko=False
59
+ pl=False
60
+ sv=False
61
+ en_de=False
62
+ de_en=False
63
+ en_fr=False
64
+ fr_en=False
65
+ en_es=False
66
+ es_en=False
67
+ en_it=False
68
+ it_en=False
69
+ en_nl=False
70
+ nl_en=False
71
+ en_pt=False
72
+ pt_en=False
73
+ en_ru=False
74
+ ru_en=False
75
+ en_zh=False
76
+ zh_en=False
77
+ en_ko=False
78
+ ko_en=False
79
+ en_pl=False
80
+ pl_en=False
81
+ en_sv=False
82
+ sv_en=False
83
+ )
84
+
85
+ threshold=(Dataset:
86
+ en=516
87
+ es=275
88
+ de=611
89
+ fr=322
90
+ nl=649
91
+ pt=257
92
+ it=332
93
+ ru=334
94
+ zh=2041
95
+ ko=198
96
+ pl=261
97
+ sv=699
98
+ en_de=100000
99
+ de_en=100000
100
+ en_fr=100000
101
+ fr_en=100000
102
+ en_es=100000
103
+ es_en=100000
104
+ en_it=100000
105
+ it_en=100000
106
+ en_nl=100000
107
+ nl_en=100000
108
+ en_pt=100000
109
+ pt_en=100000
110
+ en_ru=100000
111
+ ru_en=100000
112
+ en_zh=100000
113
+ zh_en=100000
114
+ en_ko=100000
115
+ ko_en=100000
116
+ en_pl=100000
117
+ pl_en=100000
118
+ en_sv=100000
119
+ sv_en=100000
120
+ )
121
+
122
+ # rougly 67% for mc4, 33% for total parallel data
123
+ datamix_weights=(
124
+ DataMix:
125
+ mc4_parallel_uniform=(
126
+ Dataset:
127
+ en=670
128
+ es=670
129
+ de=670
130
+ fr=670
131
+ nl=670
132
+ pt=670
133
+ it=670
134
+ ru=670
135
+ zh=670
136
+ ko=670
137
+ pl=0
138
+ sv=0
139
+ en_de=183
140
+ de_en=183
141
+ en_fr=183
142
+ fr_en=183
143
+ en_es=183
144
+ es_en=183
145
+ en_it=183
146
+ it_en=183
147
+ en_nl=183
148
+ nl_en=183
149
+ en_pt=183
150
+ pt_en=183
151
+ en_ru=183
152
+ ru_en=183
153
+ en_zh=183
154
+ zh_en=183
155
+ en_ko=183
156
+ ko_en=183
157
+ en_pl=0
158
+ pl_en=0
159
+ en_sv=0
160
+ sv_en=0
161
+ )
162
+ )
163
+
164
+ n_tokens=(Dataset:
165
+ en=4000000000
166
+ es=4000000000
167
+ de=4000000000
168
+ fr=4000000000
169
+ nl=4000000000
170
+ pt=4000000000
171
+ it=4000000000
172
+ ru=4000000000
173
+ zh=10000000000
174
+ ko=4000000000
175
+ pl=4000000000
176
+ sv=4000000000
177
+ en_de=200000000
178
+ de_en=200000000
179
+ en_fr=200000000
180
+ fr_en=200000000
181
+ en_es=200000000
182
+ es_en=200000000
183
+ en_it=200000000
184
+ it_en=200000000
185
+ en_nl=200000000
186
+ nl_en=200000000
187
+ en_pt=200000000
188
+ pt_en=200000000
189
+ en_ru=200000000
190
+ ru_en=200000000
191
+ en_zh=200000000
192
+ zh_en=200000000
193
+ en_ko=200000000
194
+ ko_en=200000000
195
+ en_pl=200000000
196
+ pl_en=200000000
197
+ en_sv=200000000
198
+ sv_en=200000000
199
+ )
200
+
201
+ is_parallel=(Dataset:
202
+ en=False
203
+ es=False
204
+ de=False
205
+ fr=False
206
+ nl=False
207
+ pt=False
208
+ it=False
209
+ ru=False
210
+ zh=False
211
+ ko=False
212
+ pl=False
213
+ sv=False
214
+ en_de=True
215
+ de_en=True
216
+ en_fr=True
217
+ fr_en=True
218
+ en_es=True
219
+ es_en=True
220
+ en_it=True
221
+ it_en=True
222
+ en_nl=True
223
+ nl_en=True
224
+ en_pt=True
225
+ pt_en=True
226
+ en_ru=True
227
+ ru_en=True
228
+ en_zh=True
229
+ zh_en=True
230
+ en_ko=True
231
+ ko_en=True
232
+ en_pl=True
233
+ pl_en=True
234
+ en_sv=True
235
+ sv_en=True
236
+ )
237
+
238
+ lp=(Dataset:
239
+ en="en"
240
+ es="es"
241
+ de="de"
242
+ fr="fr"
243
+ nl="nl"
244
+ pt="pt"
245
+ it="it"
246
+ ru="ru"
247
+ zh="zh"
248
+ ko="ko"
249
+ pl="pl"
250
+ sv="sv"
251
+ en_de="en-de"
252
+ de_en="de-en"
253
+ en_fr="en-fr"
254
+ fr_en="fr-en"
255
+ en_es="en-es"
256
+ es_en="es-en"
257
+ en_it="en-it"
258
+ it_en="it-en"
259
+ en_nl="en-nl"
260
+ nl_en="nl-en"
261
+ en_pt="en-pt"
262
+ pt_en="pt-en"
263
+ en_ru="en-ru"
264
+ ru_en="ru-en"
265
+ en_zh="en-zh"
266
+ zh_en="zh-en"
267
+ en_ko="en-ko"
268
+ ko_en="ko-en"
269
+ en_pl="en-pl"
270
+ pl_en="pl-en"
271
+ en_sv="en-sv"
272
+ sv_en="sv-en"
273
+ )
274
+
275
+ min_perplexity=50
276
+
277
+ size=(Size: 7 13)
278
+
279
+ log_interval=1
280
+ save_interval=635
281
+ eval_interval=635
282
+ train_steps=12700
283
+
284
+ lr_scheduler=cosine
285
+ warmup_steps=127
286
+ lr=3e-5
287
+ lr_min=3e-6
288
+ weight_decay=0.1
289
+
290
+ n_gpus=8
291
+ gpu_ids=0,1,2,3,4,5,6,7
292
+ tp=(TP: 1 2 3 4)
293
+ pp=(PP: 1 2 3 4)
294
+ micro_batch_size=4
295
+ grad_accum_steps=12
296
+ vocab_size=32000
297
+
298
+ cpu_workers=16
299
+ wandb_run_id="llama2_7B_40b_base_vocab_uniform_cleaned_ppl_thresh_516_275_611_322_649_257_332_334_2041_198_and_parallel_33"
300
+ wikipedia=False
301
+ freeze_layers=""
302
+ posterior_tokens=False
303
+ n_posterior_tokens=0
304
+ eval_iters=1
305
+ }
multilinguality_megatron/ducttape/continue_pretraining.tconf ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_test
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_test/checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # less for zh (inefficient tokenizer)
25
+ n_tokens=(TrainLanguage:
26
+ en=250000000
27
+ es=83333333
28
+ de=83333333
29
+ fr=83333333
30
+ nl=83333333
31
+ pt=83333333
32
+ it=83333333
33
+ ru=83333333
34
+ zh=8333333
35
+ ko=83333333
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data_2/shared/tower_llm_data/en/data
40
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
41
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
42
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
43
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
44
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
45
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
46
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
47
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
48
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=10
58
+ save_interval=318
59
+ eval_interval=158
60
+ train_steps=1272
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=13
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=8
69
+ gpu_ids=0,1,2,3,4,5,6,7
70
+ tp=(TP: 1 2 3 4)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=6
74
+
75
+ cpu_workers=16
76
+
77
+ }
multilinguality_megatron/ducttape/data_test.tconf ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/wikipedia_llama2_all_10B
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B/checkpoints
6
+ model_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf
7
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Llama-2-7b-hf/snapshots/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # number such that final tokens for each language are around 1B
25
+ n_tokens=(TrainLanguage:
26
+ en=1000000000
27
+ es=833333330
28
+ de=833333330
29
+ fr=833333330
30
+ nl=833333330
31
+ pt=833333330
32
+ it=833333330
33
+ ru=500000000
34
+ zh=13888888
35
+ ko=250000000
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/en
40
+ es=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/es
41
+ de=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/de
42
+ fr=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/fr
43
+ nl=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/nl
44
+ pt=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/pt
45
+ it=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/it
46
+ ru=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ru
47
+ zh=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/zh
48
+ ko=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ko
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=10
58
+ save_interval=635
59
+ eval_interval=635
60
+ train_steps=6358
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=63
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=8
69
+ gpu_ids=0,1,2,3,4,5,6,7
70
+ tp=(TP: 1 2 3 4)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=12
74
+ vocab_size=32000
75
+
76
+ cpu_workers=16
77
+ wandb_run_id="wikipedia"
78
+ wikipedia=True
79
+ }
multilinguality_megatron/ducttape/data_test_extend32.tconf ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/wikipedia_llama2_all_10B_extend32
3
+ repo=/mnt/data/jpombal/multilinguality_megatron
4
+
5
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama2_all_10B_extend32/checkpoints
6
+ model_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit
7
+ tokenizer_path=/mnt/data/bpop/multilinguality_tower/extended-models/llama-2-7b-hf-merged-multi-32k-meaninit/tokenizer.model
8
+
9
+ train_language=(TrainLanguage: en de fr es it nl pt ru zh ko)
10
+
11
+ threshold=(TrainLanguage:
12
+ en=516
13
+ es=275
14
+ de=611
15
+ fr=322
16
+ nl=649
17
+ pt=257
18
+ it=332
19
+ ru=334
20
+ zh=2041
21
+ ko=198
22
+ )
23
+
24
+ # number such that final tokens for each language are around 1B
25
+ n_tokens=(TrainLanguage:
26
+ en=1000000000
27
+ es=833333330
28
+ de=833333330
29
+ fr=833333330
30
+ nl=833333330
31
+ pt=833333330
32
+ it=833333330
33
+ ru=500000000
34
+ zh=13888888
35
+ ko=250000000
36
+ )
37
+
38
+ dataset_path=(TrainLanguage:
39
+ en=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/en
40
+ es=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/es
41
+ de=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/de
42
+ fr=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/fr
43
+ nl=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/nl
44
+ pt=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/pt
45
+ it=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/it
46
+ ru=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ru
47
+ zh=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/zh
48
+ ko=/mnt/data/shared/multilingual_llm/tower_llm_wikipedia/ko
49
+ )
50
+
51
+ mix="10 10 10 10 10 10 10 10 10 10"
52
+
53
+ min_perplexity=50
54
+
55
+ size=(Size: 7 13)
56
+
57
+ log_interval=10
58
+ save_interval=635
59
+ eval_interval=635
60
+ train_steps=6358
61
+
62
+ lr_scheduler=cosine
63
+ warmup_steps=63
64
+ lr=3e-5
65
+ lr_min=3e-6
66
+ weight_decay=0.1
67
+
68
+ n_gpus=8
69
+ gpu_ids=0,1,2,3,4,5,6,7
70
+ tp=(TP: 1 2 3 4)
71
+ pp=(PP: 1 2 3 4)
72
+ micro_batch_size=4
73
+ grad_accum_steps=12
74
+ vocab_size=52672
75
+
76
+ cpu_workers=16
77
+ wandb_run_id="wikipedia_extend32"
78
+ wikipedia=True
79
+ }
multilinguality_megatron/ducttape/gemma_2B_20B_all_cleaned_mc4_parallel.tconf ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ model_type="gemma"
3
+ ducttape_output=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B
4
+ repo=/mnt/data/jpombal/multilinguality_megatron
5
+
6
+ external_model_dir=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B/test
7
+ external_model_dir_annealing=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B/test
8
+ model_path=/mnt/data_2/cache/models--google--gemma-2b/snapshots/9d067f00def958594aaa16b39a65b07d69ca655b/
9
+ tokenizer_path=/mnt/data_2/cache/models--google--gemma-2b/snapshots/9d067f00def958594aaa16b39a65b07d69ca655b
10
+
11
+ tokenizer_type=PretrainedFromHF
12
+
13
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
14
+ datamix_weights_annealing=""
15
+
16
+ dataset_path=(Dataset:
17
+ en=/mnt/data_2/shared/pre-training/tower_llm_data/en/data
18
+ es=/mnt/data_2/shared/pre-training/tower_llm_data/es/3/0000.json.gz
19
+ de=/mnt/data_2/shared/pre-training/tower_llm_data/de/2/0000.json.gz
20
+ fr=/mnt/data_2/shared/pre-training/tower_llm_data/fr/1/0000.json.gz
21
+ nl=/mnt/data_2/shared/pre-training/tower_llm_data/nl/0000.json.gz
22
+ pt=/mnt/data_2/shared/pre-training/tower_llm_data/pt/0000.json.gz
23
+ it=/mnt/data_2/shared/pre-training/tower_llm_data/it/0000.json.gz
24
+ ru=/mnt/data_2/shared/pre-training/tower_llm_data/ru/6/0000.json.gz
25
+ zh=/mnt/data_2/shared/pre-training/tower_llm_data/zh/0000.json.gz
26
+ ko=/mnt/data_2/shared/pre-training/tower_llm_data/ko/0000.json.gz
27
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
44
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
45
+ )
46
+
47
+ is_hf_dataset=(Dataset:
48
+ en=True
49
+ es=False
50
+ de=False
51
+ fr=False
52
+ nl=False
53
+ pt=False
54
+ it=False
55
+ ru=False
56
+ zh=False
57
+ ko=False
58
+ en_de=False
59
+ de_en=False
60
+ en_fr=False
61
+ fr_en=False
62
+ en_es=False
63
+ es_en=False
64
+ en_it=False
65
+ it_en=False
66
+ en_nl=False
67
+ nl_en=False
68
+ en_pt=False
69
+ pt_en=False
70
+ en_ru=False
71
+ ru_en=False
72
+ en_zh=False
73
+ zh_en=False
74
+ en_ko=False
75
+ ko_en=False
76
+ )
77
+
78
+ threshold=(Dataset:
79
+ en=516
80
+ es=275
81
+ de=611
82
+ fr=322
83
+ nl=649
84
+ pt=257
85
+ it=332
86
+ ru=334
87
+ zh=2041
88
+ ko=198
89
+ en_de=100000
90
+ de_en=100000
91
+ en_fr=100000
92
+ fr_en=100000
93
+ en_es=100000
94
+ es_en=100000
95
+ en_it=100000
96
+ it_en=100000
97
+ en_nl=100000
98
+ nl_en=100000
99
+ en_pt=100000
100
+ pt_en=100000
101
+ en_ru=100000
102
+ ru_en=100000
103
+ en_zh=100000
104
+ zh_en=100000
105
+ en_ko=100000
106
+ ko_en=100000
107
+ )
108
+
109
+ # rougly 67% for mc4, 33% for total parallel data
110
+ datamix_weights=(
111
+ DataMix:
112
+ mc4_parallel_uniform=(
113
+ Dataset:
114
+ en=670
115
+ es=670
116
+ de=670
117
+ fr=670
118
+ nl=670
119
+ pt=670
120
+ it=670
121
+ ru=670
122
+ zh=670
123
+ ko=670
124
+ en_de=183
125
+ de_en=183
126
+ en_fr=183
127
+ fr_en=183
128
+ en_es=183
129
+ es_en=183
130
+ en_it=183
131
+ it_en=183
132
+ en_nl=183
133
+ nl_en=183
134
+ en_pt=183
135
+ pt_en=183
136
+ en_ru=183
137
+ ru_en=183
138
+ en_zh=183
139
+ zh_en=183
140
+ en_ko=183
141
+ ko_en=183
142
+ )
143
+ )
144
+
145
+ # number such that final tokens for each language are around 1B
146
+ n_tokens=(Dataset:
147
+ en=1000000000
148
+ es=833333330
149
+ de=833333330
150
+ fr=833333330
151
+ nl=833333330
152
+ pt=833333330
153
+ it=833333330
154
+ ru=500000000
155
+ zh=13888888
156
+ ko=250000000
157
+ en_de=20000000
158
+ de_en=20000000
159
+ en_fr=20000000
160
+ fr_en=20000000
161
+ en_es=20000000
162
+ es_en=20000000
163
+ en_it=20000000
164
+ it_en=20000000
165
+ en_nl=20000000
166
+ nl_en=20000000
167
+ en_pt=20000000
168
+ pt_en=20000000
169
+ en_ru=20000000
170
+ ru_en=20000000
171
+ en_zh=20000000
172
+ zh_en=20000000
173
+ en_ko=20000000
174
+ ko_en=20000000
175
+ )
176
+
177
+ is_parallel=(Dataset:
178
+ en=False
179
+ es=False
180
+ de=False
181
+ fr=False
182
+ nl=False
183
+ pt=False
184
+ it=False
185
+ ru=False
186
+ zh=False
187
+ ko=False
188
+ en_de=True
189
+ de_en=True
190
+ en_fr=True
191
+ fr_en=True
192
+ en_es=True
193
+ es_en=True
194
+ en_it=True
195
+ it_en=True
196
+ en_nl=True
197
+ nl_en=True
198
+ en_pt=True
199
+ pt_en=True
200
+ en_ru=True
201
+ ru_en=True
202
+ en_zh=True
203
+ zh_en=True
204
+ en_ko=True
205
+ ko_en=True
206
+ )
207
+
208
+ lp=(Dataset:
209
+ en="none"
210
+ es="none"
211
+ de="none"
212
+ fr="none"
213
+ nl="none"
214
+ pt="none"
215
+ it="none"
216
+ ru="none"
217
+ zh="none"
218
+ ko="none"
219
+ en_de="en-de"
220
+ de_en="de-en"
221
+ en_fr="en-fr"
222
+ fr_en="fr-en"
223
+ en_es="en-es"
224
+ es_en="es-en"
225
+ en_it="en-it"
226
+ it_en="it-en"
227
+ en_nl="en-nl"
228
+ nl_en="nl-en"
229
+ en_pt="en-pt"
230
+ pt_en="pt-en"
231
+ en_ru="en-ru"
232
+ ru_en="ru-en"
233
+ en_zh="en-zh"
234
+ zh_en="zh-en"
235
+ en_ko="en-ko"
236
+ ko_en="ko-en"
237
+ )
238
+
239
+ min_perplexity=0
240
+
241
+ size=(Size: 2 7)
242
+
243
+ log_interval=1
244
+ save_interval=635
245
+ eval_interval=635
246
+ train_steps=12700
247
+ train_steps_annealing=0
248
+
249
+ lr_scheduler=cosine
250
+ warmup_steps=127
251
+ lr=3e-5
252
+ lr_min=3e-6
253
+ weight_decay=0.1
254
+
255
+ lr_scheduler_annealing=linear
256
+ warmup_steps_annealing=0
257
+ lr_annealing=3e-5
258
+ lr_min_annealing=3e-6
259
+
260
+ n_gpus=8
261
+ gpu_ids=0,1,2,3,4,5,6,7
262
+ tp=(TP: 1 2 3 4 5 6 7 8)
263
+ pp=(PP: 1 2 3 4)
264
+ micro_batch_size=2
265
+ grad_accum_steps=48
266
+ vocab_size=256000
267
+
268
+ cpu_workers=16
269
+ wikipedia=False
270
+ freeze_layers=""
271
+ posterior_tokens=False
272
+ n_posterior_tokens=0
273
+ eval_iters=1
274
+
275
+ glu_activation=geglu
276
+ kv_channels=256
277
+ layernorm_epsilon=1e-6
278
+
279
+ seq_length=2048
280
+ }
multilinguality_megatron/ducttape/gemma_2b_flavio.tconf ADDED
@@ -0,0 +1,546 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ model_type="gemma"
3
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_gemma_2_20B
4
+ repo=/mnt/data/jpombal/multilinguality_megatron
5
+
6
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_gemma_2_20B/flavio_checkpoints
7
+ external_model_dir_annealing=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_gemma_2_20B/flavio_checkpoints_annealed
8
+ model_path=/mnt/data_2/cache/models--google--gemma-2b/snapshots/9d067f00def958594aaa16b39a65b07d69ca655b/
9
+ tokenizer_path=/mnt/data_2/cache/models--google--gemma-2b/snapshots/9d067f00def958594aaa16b39a65b07d69ca655b
10
+
11
+ tokenizer_type=PretrainedFromHF
12
+
13
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en en_de_pre_annealing de_en_pre_annealing en_fr_pre_annealing fr_en_pre_annealing en_es_pre_annealing es_en_pre_annealing en_it_pre_annealing it_en_pre_annealing en_nl_pre_annealing nl_en_pre_annealing en_pt_pre_annealing pt_en_pre_annealing en_ru_pre_annealing ru_en_pre_annealing en_zh_pre_annealing zh_en_pre_annealing en_ko_pre_annealing ko_en_pre_annealing en_synth es_synth de_synth fr_synth nl_synth pt_synth it_synth ru_synth zh_synth ko_synth instructions)
14
+ dataset_path=(Dataset:
15
+ en=/mnt/data_2/shared/tower_llm_data/en/data
16
+ en_synth=""
17
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
18
+ es_synth=""
19
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
20
+ de_synth=""
21
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
22
+ fr_synth=""
23
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
24
+ nl_synth=""
25
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
26
+ pt_synth=""
27
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
28
+ it_synth=""
29
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
30
+ ru_synth=""
31
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
32
+ zh_synth=""
33
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
34
+ ko_synth=""
35
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
44
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
45
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
46
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
47
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
48
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
49
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
50
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
51
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
52
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
53
+ instructions="oi"
54
+ en_de_pre_annealing="oi"
55
+ de_en_pre_annealing="oi"
56
+ en_fr_pre_annealing="oi"
57
+ fr_en_pre_annealing="oi"
58
+ en_es_pre_annealing="oi"
59
+ es_en_pre_annealing="oi"
60
+ en_it_pre_annealing="oi"
61
+ it_en_pre_annealing="oi"
62
+ en_nl_pre_annealing="oi"
63
+ nl_en_pre_annealing="oi"
64
+ en_pt_pre_annealing="oi"
65
+ pt_en_pre_annealing="oi"
66
+ en_ru_pre_annealing="oi"
67
+ ru_en_pre_annealing="oi"
68
+ en_zh_pre_annealing="oi"
69
+ zh_en_pre_annealing="oi"
70
+ en_ko_pre_annealing="oi"
71
+ ko_en_pre_annealing="oi"
72
+ )
73
+
74
+ is_hf_dataset=(Dataset:
75
+ en=True
76
+ es=False
77
+ de=False
78
+ fr=False
79
+ nl=False
80
+ pt=False
81
+ it=False
82
+ ru=False
83
+ zh=False
84
+ ko=False
85
+ en_de=False
86
+ de_en=False
87
+ en_fr=False
88
+ fr_en=False
89
+ en_es=False
90
+ es_en=False
91
+ en_it=False
92
+ it_en=False
93
+ en_nl=False
94
+ nl_en=False
95
+ en_pt=False
96
+ pt_en=False
97
+ en_ru=False
98
+ ru_en=False
99
+ en_zh=False
100
+ zh_en=False
101
+ en_ko=False
102
+ ko_en=False
103
+ en_synth=False
104
+ es_synth=False
105
+ de_synth=False
106
+ fr_synth=False
107
+ nl_synth=False
108
+ pt_synth=False
109
+ it_synth=False
110
+ ru_synth=False
111
+ zh_synth=False
112
+ ko_synth=False
113
+ instructions="oi"
114
+ en_de_pre_annealing="oi"
115
+ de_en_pre_annealing="oi"
116
+ en_fr_pre_annealing="oi"
117
+ fr_en_pre_annealing="oi"
118
+ en_es_pre_annealing="oi"
119
+ es_en_pre_annealing="oi"
120
+ en_it_pre_annealing="oi"
121
+ it_en_pre_annealing="oi"
122
+ en_nl_pre_annealing="oi"
123
+ nl_en_pre_annealing="oi"
124
+ en_pt_pre_annealing="oi"
125
+ pt_en_pre_annealing="oi"
126
+ en_ru_pre_annealing="oi"
127
+ ru_en_pre_annealing="oi"
128
+ en_zh_pre_annealing="oi"
129
+ zh_en_pre_annealing="oi"
130
+ en_ko_pre_annealing="oi"
131
+ ko_en_pre_annealing="oi"
132
+ )
133
+
134
+ threshold=(Dataset:
135
+ en=516
136
+ es=275
137
+ de=611
138
+ fr=322
139
+ nl=649
140
+ pt=257
141
+ it=332
142
+ ru=334
143
+ zh=2041
144
+ ko=198
145
+ en_de=100000
146
+ de_en=100000
147
+ en_fr=100000
148
+ fr_en=100000
149
+ en_es=100000
150
+ es_en=100000
151
+ en_it=100000
152
+ it_en=100000
153
+ en_nl=100000
154
+ nl_en=100000
155
+ en_pt=100000
156
+ pt_en=100000
157
+ en_ru=100000
158
+ ru_en=100000
159
+ en_zh=100000
160
+ zh_en=100000
161
+ en_ko=100000
162
+ ko_en=100000
163
+ en_synth=100000
164
+ es_synth=100000
165
+ de_synth=100000
166
+ fr_synth=100000
167
+ nl_synth=100000
168
+ pt_synth=100000
169
+ it_synth=100000
170
+ ru_synth=100000
171
+ zh_synth=100000
172
+ ko_synth=100000
173
+ instructions="oi"
174
+ en_de_pre_annealing="oi"
175
+ de_en_pre_annealing="oi"
176
+ en_fr_pre_annealing="oi"
177
+ fr_en_pre_annealing="oi"
178
+ en_es_pre_annealing="oi"
179
+ es_en_pre_annealing="oi"
180
+ en_it_pre_annealing="oi"
181
+ it_en_pre_annealing="oi"
182
+ en_nl_pre_annealing="oi"
183
+ nl_en_pre_annealing="oi"
184
+ en_pt_pre_annealing="oi"
185
+ pt_en_pre_annealing="oi"
186
+ en_ru_pre_annealing="oi"
187
+ ru_en_pre_annealing="oi"
188
+ en_zh_pre_annealing="oi"
189
+ zh_en_pre_annealing="oi"
190
+ en_ko_pre_annealing="oi"
191
+ ko_en_pre_annealing="oi"
192
+ )
193
+
194
+ # rougly 67% for mc4, 33% for total parallel data
195
+ datamix_weights=(
196
+ DataMix:
197
+ mc4_parallel_uniform=(
198
+ Dataset:
199
+ en=603
200
+ es=603
201
+ de=603
202
+ fr=603
203
+ nl=603
204
+ pt=603
205
+ it=603
206
+ ru=603
207
+ zh=603
208
+ ko=603
209
+ en_de=0
210
+ de_en=0
211
+ en_fr=0
212
+ fr_en=0
213
+ en_es=0
214
+ es_en=0
215
+ en_it=0
216
+ it_en=0
217
+ en_nl=0
218
+ nl_en=0
219
+ en_pt=0
220
+ pt_en=0
221
+ en_ru=0
222
+ ru_en=0
223
+ en_zh=0
224
+ zh_en=0
225
+ en_ko=0
226
+ ko_en=0
227
+ en_synth=67
228
+ es_synth=67
229
+ de_synth=67
230
+ fr_synth=67
231
+ nl_synth=67
232
+ pt_synth=67
233
+ it_synth=67
234
+ ru_synth=67
235
+ zh_synth=67
236
+ ko_synth=67
237
+ instructions=0
238
+ en_de_pre_annealing=183
239
+ de_en_pre_annealing=183
240
+ en_fr_pre_annealing=183
241
+ fr_en_pre_annealing=183
242
+ en_es_pre_annealing=183
243
+ es_en_pre_annealing=183
244
+ en_it_pre_annealing=183
245
+ it_en_pre_annealing=183
246
+ en_nl_pre_annealing=183
247
+ nl_en_pre_annealing=183
248
+ en_pt_pre_annealing=183
249
+ pt_en_pre_annealing=183
250
+ en_ru_pre_annealing=183
251
+ ru_en_pre_annealing=183
252
+ en_zh_pre_annealing=183
253
+ zh_en_pre_annealing=183
254
+ en_ko_pre_annealing=183
255
+ ko_en_pre_annealing=183
256
+ )
257
+ )
258
+
259
+ datamix_weights_annealing=(
260
+ DataMix:
261
+ mc4_parallel_uniform=(
262
+ Dataset:
263
+ en=0
264
+ es=0
265
+ de=0
266
+ fr=0
267
+ nl=0
268
+ pt=0
269
+ it=0
270
+ ru=0
271
+ zh=0
272
+ ko=0
273
+ en_de=833
274
+ de_en=833
275
+ en_fr=833
276
+ fr_en=833
277
+ en_es=833
278
+ es_en=833
279
+ en_it=833
280
+ it_en=833
281
+ en_nl=833
282
+ nl_en=833
283
+ en_pt=833
284
+ pt_en=833
285
+ en_ru=833
286
+ ru_en=833
287
+ en_zh=833
288
+ zh_en=833
289
+ en_ko=833
290
+ ko_en=833
291
+ en_synth=0
292
+ es_synth=0
293
+ de_synth=0
294
+ fr_synth=0
295
+ nl_synth=0
296
+ pt_synth=0
297
+ it_synth=0
298
+ ru_synth=0
299
+ zh_synth=0
300
+ ko_synth=0
301
+ instructions=85000
302
+ en_de_pre_annealing=0
303
+ de_en_pre_annealing=0
304
+ en_fr_pre_annealing=0
305
+ fr_en_pre_annealing=0
306
+ en_es_pre_annealing=0
307
+ es_en_pre_annealing=0
308
+ en_it_pre_annealing=0
309
+ it_en_pre_annealing=0
310
+ en_nl_pre_annealing=0
311
+ nl_en_pre_annealing=0
312
+ en_pt_pre_annealing=0
313
+ pt_en_pre_annealing=0
314
+ en_ru_pre_annealing=0
315
+ ru_en_pre_annealing=0
316
+ en_zh_pre_annealing=0
317
+ zh_en_pre_annealing=0
318
+ en_ko_pre_annealing=0
319
+ ko_en_pre_annealing=0
320
+ )
321
+ )
322
+
323
+
324
+ # number such that final tokens for each language are around 1B
325
+ n_tokens=(Dataset:
326
+ en=1000000000
327
+ es=833333330
328
+ de=833333330
329
+ fr=833333330
330
+ nl=833333330
331
+ pt=833333330
332
+ it=833333330
333
+ ru=500000000
334
+ zh=13888888
335
+ ko=250000000
336
+ en_de=20000000
337
+ de_en=20000000
338
+ en_fr=20000000
339
+ fr_en=20000000
340
+ en_es=20000000
341
+ es_en=20000000
342
+ en_it=20000000
343
+ it_en=20000000
344
+ en_nl=20000000
345
+ nl_en=20000000
346
+ en_pt=20000000
347
+ pt_en=20000000
348
+ en_ru=20000000
349
+ ru_en=20000000
350
+ en_zh=20000000
351
+ zh_en=20000000
352
+ en_ko=20000000
353
+ ko_en=20000000
354
+ en_synth=20000000
355
+ es_synth=20000000
356
+ de_synth=20000000
357
+ fr_synth=20000000
358
+ nl_synth=20000000
359
+ pt_synth=20000000
360
+ it_synth=20000000
361
+ ru_synth=20000000
362
+ zh_synth=20000000
363
+ ko_synth=20000000
364
+ instructions="oi"
365
+ en_de_pre_annealing="oi"
366
+ de_en_pre_annealing="oi"
367
+ en_fr_pre_annealing="oi"
368
+ fr_en_pre_annealing="oi"
369
+ en_es_pre_annealing="oi"
370
+ es_en_pre_annealing="oi"
371
+ en_it_pre_annealing="oi"
372
+ it_en_pre_annealing="oi"
373
+ en_nl_pre_annealing="oi"
374
+ nl_en_pre_annealing="oi"
375
+ en_pt_pre_annealing="oi"
376
+ pt_en_pre_annealing="oi"
377
+ en_ru_pre_annealing="oi"
378
+ ru_en_pre_annealing="oi"
379
+ en_zh_pre_annealing="oi"
380
+ zh_en_pre_annealing="oi"
381
+ en_ko_pre_annealing="oi"
382
+ ko_en_pre_annealing="oi"
383
+ )
384
+
385
+ is_parallel=(Dataset:
386
+ en=False
387
+ es=False
388
+ de=False
389
+ fr=False
390
+ nl=False
391
+ pt=False
392
+ it=False
393
+ ru=False
394
+ zh=False
395
+ ko=False
396
+ en_de=True
397
+ de_en=True
398
+ en_fr=True
399
+ fr_en=True
400
+ en_es=True
401
+ es_en=True
402
+ en_it=True
403
+ it_en=True
404
+ en_nl=True
405
+ nl_en=True
406
+ en_pt=True
407
+ pt_en=True
408
+ en_ru=True
409
+ ru_en=True
410
+ en_zh=True
411
+ zh_en=True
412
+ en_ko=True
413
+ ko_en=True
414
+ en_synth=False
415
+ es_synth=False
416
+ de_synth=False
417
+ fr_synth=False
418
+ nl_synth=False
419
+ pt_synth=False
420
+ it_synth=False
421
+ ru_synth=False
422
+ zh_synth=False
423
+ ko_synth=False
424
+ instructions="oi"
425
+ en_de_pre_annealing="oi"
426
+ de_en_pre_annealing="oi"
427
+ en_fr_pre_annealing="oi"
428
+ fr_en_pre_annealing="oi"
429
+ en_es_pre_annealing="oi"
430
+ es_en_pre_annealing="oi"
431
+ en_it_pre_annealing="oi"
432
+ it_en_pre_annealing="oi"
433
+ en_nl_pre_annealing="oi"
434
+ nl_en_pre_annealing="oi"
435
+ en_pt_pre_annealing="oi"
436
+ pt_en_pre_annealing="oi"
437
+ en_ru_pre_annealing="oi"
438
+ ru_en_pre_annealing="oi"
439
+ en_zh_pre_annealing="oi"
440
+ zh_en_pre_annealing="oi"
441
+ en_ko_pre_annealing="oi"
442
+ ko_en_pre_annealing="oi"
443
+ )
444
+
445
+ lp=(Dataset:
446
+ en=""
447
+ es=""
448
+ de=""
449
+ fr=""
450
+ nl=""
451
+ pt=""
452
+ it=""
453
+ ru=""
454
+ zh=""
455
+ ko=""
456
+ en_de="en-de"
457
+ de_en="de-en"
458
+ en_fr="en-fr"
459
+ fr_en="fr-en"
460
+ en_es="en-es"
461
+ es_en="es-en"
462
+ en_it="en-it"
463
+ it_en="it-en"
464
+ en_nl="en-nl"
465
+ nl_en="nl-en"
466
+ en_pt="en-pt"
467
+ pt_en="pt-en"
468
+ en_ru="en-ru"
469
+ ru_en="ru-en"
470
+ en_zh="en-zh"
471
+ zh_en="zh-en"
472
+ en_ko="en-ko"
473
+ ko_en="ko-en"
474
+ en_synth=""
475
+ es_synth=""
476
+ de_synth=""
477
+ fr_synth=""
478
+ nl_synth=""
479
+ pt_synth=""
480
+ it_synth=""
481
+ ru_synth=""
482
+ zh_synth=""
483
+ ko_synth=""
484
+ instructions="oi"
485
+ en_de_pre_annealing="oi"
486
+ de_en_pre_annealing="oi"
487
+ en_fr_pre_annealing="oi"
488
+ fr_en_pre_annealing="oi"
489
+ en_es_pre_annealing="oi"
490
+ es_en_pre_annealing="oi"
491
+ en_it_pre_annealing="oi"
492
+ it_en_pre_annealing="oi"
493
+ en_nl_pre_annealing="oi"
494
+ nl_en_pre_annealing="oi"
495
+ en_pt_pre_annealing="oi"
496
+ pt_en_pre_annealing="oi"
497
+ en_ru_pre_annealing="oi"
498
+ ru_en_pre_annealing="oi"
499
+ en_zh_pre_annealing="oi"
500
+ zh_en_pre_annealing="oi"
501
+ en_ko_pre_annealing="oi"
502
+ ko_en_pre_annealing="oi"
503
+ )
504
+
505
+ min_perplexity=0
506
+
507
+ size=(Size: 2)
508
+
509
+ log_interval=1
510
+ save_interval=635
511
+ eval_interval=635
512
+ train_steps=11430
513
+ train_steps_annealing=1270
514
+
515
+ lr_scheduler=constant
516
+ warmup_steps=32
517
+ lr=3e-5
518
+ lr_min=3e-6
519
+ weight_decay=0.1
520
+
521
+ lr_scheduler_annealing=linear
522
+ warmup_steps_annealing=0
523
+ lr_annealing=3e-5
524
+ lr_min_annealing=3e-6
525
+
526
+ n_gpus=8
527
+ gpu_ids=0,1,2,3,4,5,6,7
528
+ tp=(TP: 1 2 3 4 5 6 7 8)
529
+ pp=(PP: 1 2 3 4)
530
+ micro_batch_size=24
531
+ grad_accum_steps=4
532
+ vocab_size=256000
533
+
534
+ cpu_workers=16
535
+ wikipedia=False
536
+ freeze_layers=""
537
+ posterior_tokens=False
538
+ n_posterior_tokens=0
539
+ eval_iters=1
540
+
541
+ glu_activation=geglu
542
+ kv_channels=256
543
+ layernorm_epsilon=1e-6
544
+
545
+ seq_length=2048
546
+ }
multilinguality_megatron/ducttape/gemma_7B_20B_all_cleaned_mc4_parallel.tconf ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ model_type="gemma"
3
+ ducttape_output=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B
4
+ repo=/mnt/data/jpombal/multilinguality_megatron
5
+
6
+ external_model_dir=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B/old_recipe_checkpoints
7
+ external_model_dir_annealing=/mnt/data_2/shared/experiments_megatron/continue_pretraining_gemma_7B/old_recipe_checkpoints
8
+ model_path=/mnt/data_2/cache/models--google--gemma-7b/snapshots/bc0790ce8e02c6b2240e2b94bf01fb0453dc90f6
9
+ tokenizer_path=/mnt/data_2/cache/models--google--gemma-7b/snapshots/bc0790ce8e02c6b2240e2b94bf01fb0453dc90f6
10
+
11
+ tokenizer_type=PretrainedFromHF
12
+
13
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en)
14
+ datamix_weights_annealing=""
15
+
16
+ dataset_path=(Dataset:
17
+ en=/mnt/data_2/shared/pre-training/tower_llm_data/en/data
18
+ es=/mnt/data_2/shared/pre-training/tower_llm_data/es/3/0000.json.gz
19
+ de=/mnt/data_2/shared/pre-training/tower_llm_data/de/2/0000.json.gz
20
+ fr=/mnt/data_2/shared/pre-training/tower_llm_data/fr/1/0000.json.gz
21
+ nl=/mnt/data_2/shared/pre-training/tower_llm_data/nl/0000.json.gz
22
+ pt=/mnt/data_2/shared/pre-training/tower_llm_data/pt/0000.json.gz
23
+ it=/mnt/data_2/shared/pre-training/tower_llm_data/it/0000.json.gz
24
+ ru=/mnt/data_2/shared/pre-training/tower_llm_data/ru/6/0000.json.gz
25
+ zh=/mnt/data_2/shared/pre-training/tower_llm_data/zh/0000.json.gz
26
+ ko=/mnt/data_2/shared/pre-training/tower_llm_data/ko/0000.json.gz
27
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
28
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
29
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
30
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
31
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
32
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
33
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
34
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
35
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
44
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
45
+ )
46
+
47
+ is_hf_dataset=(Dataset:
48
+ en=True
49
+ es=False
50
+ de=False
51
+ fr=False
52
+ nl=False
53
+ pt=False
54
+ it=False
55
+ ru=False
56
+ zh=False
57
+ ko=False
58
+ en_de=False
59
+ de_en=False
60
+ en_fr=False
61
+ fr_en=False
62
+ en_es=False
63
+ es_en=False
64
+ en_it=False
65
+ it_en=False
66
+ en_nl=False
67
+ nl_en=False
68
+ en_pt=False
69
+ pt_en=False
70
+ en_ru=False
71
+ ru_en=False
72
+ en_zh=False
73
+ zh_en=False
74
+ en_ko=False
75
+ ko_en=False
76
+ )
77
+
78
+ threshold=(Dataset:
79
+ en=516
80
+ es=275
81
+ de=611
82
+ fr=322
83
+ nl=649
84
+ pt=257
85
+ it=332
86
+ ru=334
87
+ zh=2041
88
+ ko=198
89
+ en_de=100000
90
+ de_en=100000
91
+ en_fr=100000
92
+ fr_en=100000
93
+ en_es=100000
94
+ es_en=100000
95
+ en_it=100000
96
+ it_en=100000
97
+ en_nl=100000
98
+ nl_en=100000
99
+ en_pt=100000
100
+ pt_en=100000
101
+ en_ru=100000
102
+ ru_en=100000
103
+ en_zh=100000
104
+ zh_en=100000
105
+ en_ko=100000
106
+ ko_en=100000
107
+ )
108
+
109
+ # rougly 67% for mc4, 33% for total parallel data
110
+ datamix_weights=(
111
+ DataMix:
112
+ mc4_parallel_uniform=(
113
+ Dataset:
114
+ en=670
115
+ es=670
116
+ de=670
117
+ fr=670
118
+ nl=670
119
+ pt=670
120
+ it=670
121
+ ru=670
122
+ zh=670
123
+ ko=670
124
+ en_de=183
125
+ de_en=183
126
+ en_fr=183
127
+ fr_en=183
128
+ en_es=183
129
+ es_en=183
130
+ en_it=183
131
+ it_en=183
132
+ en_nl=183
133
+ nl_en=183
134
+ en_pt=183
135
+ pt_en=183
136
+ en_ru=183
137
+ ru_en=183
138
+ en_zh=183
139
+ zh_en=183
140
+ en_ko=183
141
+ ko_en=183
142
+ )
143
+ )
144
+
145
+ # number such that final tokens for each language are around 1B
146
+ n_tokens=(Dataset:
147
+ en=1000000000
148
+ es=833333330
149
+ de=833333330
150
+ fr=833333330
151
+ nl=833333330
152
+ pt=833333330
153
+ it=833333330
154
+ ru=500000000
155
+ zh=13888888
156
+ ko=250000000
157
+ en_de=20000000
158
+ de_en=20000000
159
+ en_fr=20000000
160
+ fr_en=20000000
161
+ en_es=20000000
162
+ es_en=20000000
163
+ en_it=20000000
164
+ it_en=20000000
165
+ en_nl=20000000
166
+ nl_en=20000000
167
+ en_pt=20000000
168
+ pt_en=20000000
169
+ en_ru=20000000
170
+ ru_en=20000000
171
+ en_zh=20000000
172
+ zh_en=20000000
173
+ en_ko=20000000
174
+ ko_en=20000000
175
+ )
176
+
177
+ is_parallel=(Dataset:
178
+ en=False
179
+ es=False
180
+ de=False
181
+ fr=False
182
+ nl=False
183
+ pt=False
184
+ it=False
185
+ ru=False
186
+ zh=False
187
+ ko=False
188
+ en_de=True
189
+ de_en=True
190
+ en_fr=True
191
+ fr_en=True
192
+ en_es=True
193
+ es_en=True
194
+ en_it=True
195
+ it_en=True
196
+ en_nl=True
197
+ nl_en=True
198
+ en_pt=True
199
+ pt_en=True
200
+ en_ru=True
201
+ ru_en=True
202
+ en_zh=True
203
+ zh_en=True
204
+ en_ko=True
205
+ ko_en=True
206
+ )
207
+
208
+ lp=(Dataset:
209
+ en="none"
210
+ es="none"
211
+ de="none"
212
+ fr="none"
213
+ nl="none"
214
+ pt="none"
215
+ it="none"
216
+ ru="none"
217
+ zh="none"
218
+ ko="none"
219
+ en_de="en-de"
220
+ de_en="de-en"
221
+ en_fr="en-fr"
222
+ fr_en="fr-en"
223
+ en_es="en-es"
224
+ es_en="es-en"
225
+ en_it="en-it"
226
+ it_en="it-en"
227
+ en_nl="en-nl"
228
+ nl_en="nl-en"
229
+ en_pt="en-pt"
230
+ pt_en="pt-en"
231
+ en_ru="en-ru"
232
+ ru_en="ru-en"
233
+ en_zh="en-zh"
234
+ zh_en="zh-en"
235
+ en_ko="en-ko"
236
+ ko_en="ko-en"
237
+ )
238
+
239
+ min_perplexity=0
240
+
241
+ size=(Size: 2 7)
242
+
243
+ log_interval=1
244
+ save_interval=635
245
+ eval_interval=635
246
+ train_steps=12700
247
+ train_steps_annealing=0
248
+
249
+ lr_scheduler=cosine
250
+ warmup_steps=127
251
+ lr=3e-5
252
+ lr_min=3e-6
253
+ weight_decay=0.1
254
+
255
+ lr_scheduler_annealing=linear
256
+ warmup_steps_annealing=0
257
+ lr_annealing=3e-5
258
+ lr_min_annealing=3e-6
259
+
260
+ n_gpus=8
261
+ gpu_ids=0,1,2,3,4,5,6,7
262
+ tp=(TP: 1 2 3 4 5 6 7 8)
263
+ pp=(PP: 1 2 3 4)
264
+ micro_batch_size=2
265
+ grad_accum_steps=24
266
+ vocab_size=256000
267
+
268
+ cpu_workers=16
269
+ wikipedia=False
270
+ freeze_layers=""
271
+ posterior_tokens=False
272
+ n_posterior_tokens=0
273
+ eval_iters=1
274
+
275
+ glu_activation=geglu
276
+ kv_channels=256
277
+ layernorm_epsilon=1e-6
278
+
279
+ seq_length=4096
280
+ }
multilinguality_megatron/ducttape/llama_3_flavio.tconf ADDED
@@ -0,0 +1,546 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ model_type="llama3"
3
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio
4
+ repo=/mnt/data/pmartins/code/multilinguality_megatron
5
+
6
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio/pre_annealing_checkpoints
7
+ external_model_dir_annealing=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio/checkpoints_annealed
8
+ model_path=/mnt/data_2/cache/models--meta-llama--Meta-Llama-3-8B/snapshots/cd892e8f4da1043d4b01d5ea182a2e8412bf658f/
9
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Meta-Llama-3-8B/snapshots/cd892e8f4da1043d4b01d5ea182a2e8412bf658f/
10
+
11
+ tokenizer_type=PretrainedFromHF
12
+
13
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en en_de_pre_annealing de_en_pre_annealing en_fr_pre_annealing fr_en_pre_annealing en_es_pre_annealing es_en_pre_annealing en_it_pre_annealing it_en_pre_annealing en_nl_pre_annealing nl_en_pre_annealing en_pt_pre_annealing pt_en_pre_annealing en_ru_pre_annealing ru_en_pre_annealing en_zh_pre_annealing zh_en_pre_annealing en_ko_pre_annealing ko_en_pre_annealing en_synth es_synth de_synth fr_synth nl_synth pt_synth it_synth ru_synth zh_synth ko_synth instructions)
14
+ dataset_path=(Dataset:
15
+ en=/mnt/data_2/shared/tower_llm_data/en/data
16
+ en_synth=""
17
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
18
+ es_synth=""
19
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
20
+ de_synth=""
21
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
22
+ fr_synth=""
23
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
24
+ nl_synth=""
25
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
26
+ pt_synth=""
27
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
28
+ it_synth=""
29
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
30
+ ru_synth=""
31
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
32
+ zh_synth=""
33
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
34
+ ko_synth=""
35
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
44
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
45
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
46
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
47
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
48
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
49
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
50
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
51
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
52
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
53
+ instructions="oi"
54
+ en_de_pre_annealing="oi"
55
+ de_en_pre_annealing="oi"
56
+ en_fr_pre_annealing="oi"
57
+ fr_en_pre_annealing="oi"
58
+ en_es_pre_annealing="oi"
59
+ es_en_pre_annealing="oi"
60
+ en_it_pre_annealing="oi"
61
+ it_en_pre_annealing="oi"
62
+ en_nl_pre_annealing="oi"
63
+ nl_en_pre_annealing="oi"
64
+ en_pt_pre_annealing="oi"
65
+ pt_en_pre_annealing="oi"
66
+ en_ru_pre_annealing="oi"
67
+ ru_en_pre_annealing="oi"
68
+ en_zh_pre_annealing="oi"
69
+ zh_en_pre_annealing="oi"
70
+ en_ko_pre_annealing="oi"
71
+ ko_en_pre_annealing="oi"
72
+ )
73
+
74
+ is_hf_dataset=(Dataset:
75
+ en=True
76
+ es=False
77
+ de=False
78
+ fr=False
79
+ nl=False
80
+ pt=False
81
+ it=False
82
+ ru=False
83
+ zh=False
84
+ ko=False
85
+ en_de=False
86
+ de_en=False
87
+ en_fr=False
88
+ fr_en=False
89
+ en_es=False
90
+ es_en=False
91
+ en_it=False
92
+ it_en=False
93
+ en_nl=False
94
+ nl_en=False
95
+ en_pt=False
96
+ pt_en=False
97
+ en_ru=False
98
+ ru_en=False
99
+ en_zh=False
100
+ zh_en=False
101
+ en_ko=False
102
+ ko_en=False
103
+ en_synth=False
104
+ es_synth=False
105
+ de_synth=False
106
+ fr_synth=False
107
+ nl_synth=False
108
+ pt_synth=False
109
+ it_synth=False
110
+ ru_synth=False
111
+ zh_synth=False
112
+ ko_synth=False
113
+ instructions="oi"
114
+ en_de_pre_annealing="oi"
115
+ de_en_pre_annealing="oi"
116
+ en_fr_pre_annealing="oi"
117
+ fr_en_pre_annealing="oi"
118
+ en_es_pre_annealing="oi"
119
+ es_en_pre_annealing="oi"
120
+ en_it_pre_annealing="oi"
121
+ it_en_pre_annealing="oi"
122
+ en_nl_pre_annealing="oi"
123
+ nl_en_pre_annealing="oi"
124
+ en_pt_pre_annealing="oi"
125
+ pt_en_pre_annealing="oi"
126
+ en_ru_pre_annealing="oi"
127
+ ru_en_pre_annealing="oi"
128
+ en_zh_pre_annealing="oi"
129
+ zh_en_pre_annealing="oi"
130
+ en_ko_pre_annealing="oi"
131
+ ko_en_pre_annealing="oi"
132
+ )
133
+
134
+ threshold=(Dataset:
135
+ en=516
136
+ es=275
137
+ de=611
138
+ fr=322
139
+ nl=649
140
+ pt=257
141
+ it=332
142
+ ru=334
143
+ zh=2041
144
+ ko=198
145
+ en_de=100000
146
+ de_en=100000
147
+ en_fr=100000
148
+ fr_en=100000
149
+ en_es=100000
150
+ es_en=100000
151
+ en_it=100000
152
+ it_en=100000
153
+ en_nl=100000
154
+ nl_en=100000
155
+ en_pt=100000
156
+ pt_en=100000
157
+ en_ru=100000
158
+ ru_en=100000
159
+ en_zh=100000
160
+ zh_en=100000
161
+ en_ko=100000
162
+ ko_en=100000
163
+ en_synth=100000
164
+ es_synth=100000
165
+ de_synth=100000
166
+ fr_synth=100000
167
+ nl_synth=100000
168
+ pt_synth=100000
169
+ it_synth=100000
170
+ ru_synth=100000
171
+ zh_synth=100000
172
+ ko_synth=100000
173
+ instructions="oi"
174
+ en_de_pre_annealing="oi"
175
+ de_en_pre_annealing="oi"
176
+ en_fr_pre_annealing="oi"
177
+ fr_en_pre_annealing="oi"
178
+ en_es_pre_annealing="oi"
179
+ es_en_pre_annealing="oi"
180
+ en_it_pre_annealing="oi"
181
+ it_en_pre_annealing="oi"
182
+ en_nl_pre_annealing="oi"
183
+ nl_en_pre_annealing="oi"
184
+ en_pt_pre_annealing="oi"
185
+ pt_en_pre_annealing="oi"
186
+ en_ru_pre_annealing="oi"
187
+ ru_en_pre_annealing="oi"
188
+ en_zh_pre_annealing="oi"
189
+ zh_en_pre_annealing="oi"
190
+ en_ko_pre_annealing="oi"
191
+ ko_en_pre_annealing="oi"
192
+ )
193
+
194
+ # rougly 67% for mc4, 33% for total parallel data
195
+ datamix_weights=(
196
+ DataMix:
197
+ mc4_parallel_uniform=(
198
+ Dataset:
199
+ en=603
200
+ es=603
201
+ de=603
202
+ fr=603
203
+ nl=603
204
+ pt=603
205
+ it=603
206
+ ru=603
207
+ zh=603
208
+ ko=603
209
+ en_de=0
210
+ de_en=0
211
+ en_fr=0
212
+ fr_en=0
213
+ en_es=0
214
+ es_en=0
215
+ en_it=0
216
+ it_en=0
217
+ en_nl=0
218
+ nl_en=0
219
+ en_pt=0
220
+ pt_en=0
221
+ en_ru=0
222
+ ru_en=0
223
+ en_zh=0
224
+ zh_en=0
225
+ en_ko=0
226
+ ko_en=0
227
+ en_synth=67
228
+ es_synth=67
229
+ de_synth=67
230
+ fr_synth=67
231
+ nl_synth=67
232
+ pt_synth=67
233
+ it_synth=67
234
+ ru_synth=67
235
+ zh_synth=67
236
+ ko_synth=67
237
+ instructions=0
238
+ en_de_pre_annealing=183
239
+ de_en_pre_annealing=183
240
+ en_fr_pre_annealing=183
241
+ fr_en_pre_annealing=183
242
+ en_es_pre_annealing=183
243
+ es_en_pre_annealing=183
244
+ en_it_pre_annealing=183
245
+ it_en_pre_annealing=183
246
+ en_nl_pre_annealing=183
247
+ nl_en_pre_annealing=183
248
+ en_pt_pre_annealing=183
249
+ pt_en_pre_annealing=183
250
+ en_ru_pre_annealing=183
251
+ ru_en_pre_annealing=183
252
+ en_zh_pre_annealing=183
253
+ zh_en_pre_annealing=183
254
+ en_ko_pre_annealing=183
255
+ ko_en_pre_annealing=183
256
+ )
257
+ )
258
+
259
+ datamix_weights_annealing=(
260
+ DataMix:
261
+ mc4_parallel_uniform=(
262
+ Dataset:
263
+ en=0
264
+ es=0
265
+ de=0
266
+ fr=0
267
+ nl=0
268
+ pt=0
269
+ it=0
270
+ ru=0
271
+ zh=0
272
+ ko=0
273
+ en_de=833
274
+ de_en=833
275
+ en_fr=833
276
+ fr_en=833
277
+ en_es=833
278
+ es_en=833
279
+ en_it=833
280
+ it_en=833
281
+ en_nl=833
282
+ nl_en=833
283
+ en_pt=833
284
+ pt_en=833
285
+ en_ru=833
286
+ ru_en=833
287
+ en_zh=833
288
+ zh_en=833
289
+ en_ko=833
290
+ ko_en=833
291
+ en_synth=0
292
+ es_synth=0
293
+ de_synth=0
294
+ fr_synth=0
295
+ nl_synth=0
296
+ pt_synth=0
297
+ it_synth=0
298
+ ru_synth=0
299
+ zh_synth=0
300
+ ko_synth=0
301
+ instructions=85000
302
+ en_de_pre_annealing=0
303
+ de_en_pre_annealing=0
304
+ en_fr_pre_annealing=0
305
+ fr_en_pre_annealing=0
306
+ en_es_pre_annealing=0
307
+ es_en_pre_annealing=0
308
+ en_it_pre_annealing=0
309
+ it_en_pre_annealing=0
310
+ en_nl_pre_annealing=0
311
+ nl_en_pre_annealing=0
312
+ en_pt_pre_annealing=0
313
+ pt_en_pre_annealing=0
314
+ en_ru_pre_annealing=0
315
+ ru_en_pre_annealing=0
316
+ en_zh_pre_annealing=0
317
+ zh_en_pre_annealing=0
318
+ en_ko_pre_annealing=0
319
+ ko_en_pre_annealing=0
320
+ )
321
+ )
322
+
323
+
324
+ # number such that final tokens for each language are around 1B
325
+ n_tokens=(Dataset:
326
+ en=1000000000
327
+ es=833333330
328
+ de=833333330
329
+ fr=833333330
330
+ nl=833333330
331
+ pt=833333330
332
+ it=833333330
333
+ ru=500000000
334
+ zh=13888888
335
+ ko=250000000
336
+ en_de=20000000
337
+ de_en=20000000
338
+ en_fr=20000000
339
+ fr_en=20000000
340
+ en_es=20000000
341
+ es_en=20000000
342
+ en_it=20000000
343
+ it_en=20000000
344
+ en_nl=20000000
345
+ nl_en=20000000
346
+ en_pt=20000000
347
+ pt_en=20000000
348
+ en_ru=20000000
349
+ ru_en=20000000
350
+ en_zh=20000000
351
+ zh_en=20000000
352
+ en_ko=20000000
353
+ ko_en=20000000
354
+ en_synth=20000000
355
+ es_synth=20000000
356
+ de_synth=20000000
357
+ fr_synth=20000000
358
+ nl_synth=20000000
359
+ pt_synth=20000000
360
+ it_synth=20000000
361
+ ru_synth=20000000
362
+ zh_synth=20000000
363
+ ko_synth=20000000
364
+ instructions="oi"
365
+ en_de_pre_annealing="oi"
366
+ de_en_pre_annealing="oi"
367
+ en_fr_pre_annealing="oi"
368
+ fr_en_pre_annealing="oi"
369
+ en_es_pre_annealing="oi"
370
+ es_en_pre_annealing="oi"
371
+ en_it_pre_annealing="oi"
372
+ it_en_pre_annealing="oi"
373
+ en_nl_pre_annealing="oi"
374
+ nl_en_pre_annealing="oi"
375
+ en_pt_pre_annealing="oi"
376
+ pt_en_pre_annealing="oi"
377
+ en_ru_pre_annealing="oi"
378
+ ru_en_pre_annealing="oi"
379
+ en_zh_pre_annealing="oi"
380
+ zh_en_pre_annealing="oi"
381
+ en_ko_pre_annealing="oi"
382
+ ko_en_pre_annealing="oi"
383
+ )
384
+
385
+ is_parallel=(Dataset:
386
+ en=False
387
+ es=False
388
+ de=False
389
+ fr=False
390
+ nl=False
391
+ pt=False
392
+ it=False
393
+ ru=False
394
+ zh=False
395
+ ko=False
396
+ en_de=True
397
+ de_en=True
398
+ en_fr=True
399
+ fr_en=True
400
+ en_es=True
401
+ es_en=True
402
+ en_it=True
403
+ it_en=True
404
+ en_nl=True
405
+ nl_en=True
406
+ en_pt=True
407
+ pt_en=True
408
+ en_ru=True
409
+ ru_en=True
410
+ en_zh=True
411
+ zh_en=True
412
+ en_ko=True
413
+ ko_en=True
414
+ en_synth=False
415
+ es_synth=False
416
+ de_synth=False
417
+ fr_synth=False
418
+ nl_synth=False
419
+ pt_synth=False
420
+ it_synth=False
421
+ ru_synth=False
422
+ zh_synth=False
423
+ ko_synth=False
424
+ instructions="oi"
425
+ en_de_pre_annealing="oi"
426
+ de_en_pre_annealing="oi"
427
+ en_fr_pre_annealing="oi"
428
+ fr_en_pre_annealing="oi"
429
+ en_es_pre_annealing="oi"
430
+ es_en_pre_annealing="oi"
431
+ en_it_pre_annealing="oi"
432
+ it_en_pre_annealing="oi"
433
+ en_nl_pre_annealing="oi"
434
+ nl_en_pre_annealing="oi"
435
+ en_pt_pre_annealing="oi"
436
+ pt_en_pre_annealing="oi"
437
+ en_ru_pre_annealing="oi"
438
+ ru_en_pre_annealing="oi"
439
+ en_zh_pre_annealing="oi"
440
+ zh_en_pre_annealing="oi"
441
+ en_ko_pre_annealing="oi"
442
+ ko_en_pre_annealing="oi"
443
+ )
444
+
445
+ lp=(Dataset:
446
+ en=""
447
+ es=""
448
+ de=""
449
+ fr=""
450
+ nl=""
451
+ pt=""
452
+ it=""
453
+ ru=""
454
+ zh=""
455
+ ko=""
456
+ en_de="en-de"
457
+ de_en="de-en"
458
+ en_fr="en-fr"
459
+ fr_en="fr-en"
460
+ en_es="en-es"
461
+ es_en="es-en"
462
+ en_it="en-it"
463
+ it_en="it-en"
464
+ en_nl="en-nl"
465
+ nl_en="nl-en"
466
+ en_pt="en-pt"
467
+ pt_en="pt-en"
468
+ en_ru="en-ru"
469
+ ru_en="ru-en"
470
+ en_zh="en-zh"
471
+ zh_en="zh-en"
472
+ en_ko="en-ko"
473
+ ko_en="ko-en"
474
+ en_synth=""
475
+ es_synth=""
476
+ de_synth=""
477
+ fr_synth=""
478
+ nl_synth=""
479
+ pt_synth=""
480
+ it_synth=""
481
+ ru_synth=""
482
+ zh_synth=""
483
+ ko_synth=""
484
+ instructions="oi"
485
+ en_de_pre_annealing="oi"
486
+ de_en_pre_annealing="oi"
487
+ en_fr_pre_annealing="oi"
488
+ fr_en_pre_annealing="oi"
489
+ en_es_pre_annealing="oi"
490
+ es_en_pre_annealing="oi"
491
+ en_it_pre_annealing="oi"
492
+ it_en_pre_annealing="oi"
493
+ en_nl_pre_annealing="oi"
494
+ nl_en_pre_annealing="oi"
495
+ en_pt_pre_annealing="oi"
496
+ pt_en_pre_annealing="oi"
497
+ en_ru_pre_annealing="oi"
498
+ ru_en_pre_annealing="oi"
499
+ en_zh_pre_annealing="oi"
500
+ zh_en_pre_annealing="oi"
501
+ en_ko_pre_annealing="oi"
502
+ ko_en_pre_annealing="oi"
503
+ )
504
+
505
+ min_perplexity=50
506
+
507
+ size=(Size: 8)
508
+
509
+ log_interval=1
510
+ save_interval=635
511
+ eval_interval=635
512
+ train_steps=11430
513
+ train_steps_annealing=1270
514
+
515
+ lr_scheduler=constant
516
+ warmup_steps=32
517
+ lr=3e-5
518
+ lr_min=3e-6
519
+ weight_decay=0.1
520
+
521
+ lr_scheduler_annealing=linear
522
+ warmup_steps_annealing=0
523
+ lr_annealing=3e-5
524
+ lr_min_annealing=3e-6
525
+
526
+ n_gpus=8
527
+ gpu_ids=0,1,2,3,4,5,6,7
528
+ tp=(TP: 1 2 3 4 5 6 7 8)
529
+ pp=(PP: 1 2 3 4)
530
+ micro_batch_size=4
531
+ grad_accum_steps=12
532
+ vocab_size=128256
533
+
534
+ cpu_workers=16
535
+ wikipedia=False
536
+ freeze_layers=""
537
+ posterior_tokens=False
538
+ n_posterior_tokens=0
539
+ eval_iters=1
540
+
541
+ seq_length=4096
542
+
543
+ glu_activation=swiglu
544
+ kv_channels=""
545
+ layernorm_epsilon=1e-5
546
+ }
multilinguality_megatron/ducttape/llama_3_flavio_wmt_annealing.tconf ADDED
@@ -0,0 +1,570 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ global {
2
+ model_type="llama3"
3
+ ducttape_output=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio
4
+ repo=/mnt/data/pmartins/code/multilinguality_megatron
5
+
6
+ external_model_dir=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio/pre_annealing_checkpoints
7
+ external_model_dir_annealing=/mnt/data/shared/multilingual_llm/experiments_megatron/continue_pretraining_llama3_flavio/checkpoints_annealed_wmt
8
+ model_path=/mnt/data_2/cache/models--meta-llama--Meta-Llama-3-8B/snapshots/cd892e8f4da1043d4b01d5ea182a2e8412bf658f/
9
+ tokenizer_path=/mnt/data_2/cache/models--meta-llama--Meta-Llama-3-8B/snapshots/cd892e8f4da1043d4b01d5ea182a2e8412bf658f/
10
+
11
+ tokenizer_type=PretrainedFromHF
12
+
13
+ dataset=(Dataset: en de fr es it nl pt ru zh ko en_de de_en en_fr fr_en en_es es_en en_it it_en en_nl nl_en en_pt pt_en en_ru ru_en en_zh zh_en en_ko ko_en en_de_pre_annealing de_en_pre_annealing en_fr_pre_annealing fr_en_pre_annealing en_es_pre_annealing es_en_pre_annealing en_it_pre_annealing it_en_pre_annealing en_nl_pre_annealing nl_en_pre_annealing en_pt_pre_annealing pt_en_pre_annealing en_ru_pre_annealing ru_en_pre_annealing en_zh_pre_annealing zh_en_pre_annealing en_ko_pre_annealing ko_en_pre_annealing en_synth es_synth de_synth fr_synth nl_synth pt_synth it_synth ru_synth zh_synth ko_synth instructions en_de_wmt en_ru_wmt en_zh_wmt)
14
+ dataset_path=(Dataset:
15
+ en=/mnt/data_2/shared/tower_llm_data/en/data
16
+ en_synth=""
17
+ es=/mnt/data_2/shared/tower_llm_data/es/3/0000.json.gz
18
+ es_synth=""
19
+ de=/mnt/data_2/shared/tower_llm_data/de/2/0000.json.gz
20
+ de_synth=""
21
+ fr=/mnt/data_2/shared/tower_llm_data/fr/1/0000.json.gz
22
+ fr_synth=""
23
+ nl=/mnt/data_2/shared/tower_llm_data/nl/0000.json.gz
24
+ nl_synth=""
25
+ pt=/mnt/data_2/shared/tower_llm_data/pt/0000.json.gz
26
+ pt_synth=""
27
+ it=/mnt/data_2/shared/tower_llm_data/it/0000.json.gz
28
+ it_synth=""
29
+ ru=/mnt/data_2/shared/tower_llm_data/ru/6/0000.json.gz
30
+ ru_synth=""
31
+ zh=/mnt/data_2/shared/tower_llm_data/zh/0000.json.gz
32
+ zh_synth=""
33
+ ko=/mnt/data_2/shared/tower_llm_data/ko/0000.json.gz
34
+ ko_synth=""
35
+ en_de="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
36
+ de_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-de/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
37
+ en_fr="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
38
+ fr_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-fr/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
39
+ en_es="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
40
+ es_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-es/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
41
+ en_it="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
42
+ it_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-it/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
43
+ en_nl="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
44
+ nl_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-nl/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
45
+ en_pt="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
46
+ pt_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-pt/bicleaner_0.6_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
47
+ en_ru="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
48
+ ru_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ru/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
49
+ en_zh="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
50
+ zh_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-zh/no_bicleaner_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
51
+ en_ko="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
52
+ ko_en="/mnt/data/shared/tower_llm_data/bilingual_data/v1/en-ko/bicleaner_0.5_cometkiwi-wmt22-cometkiwi-da/threshold_0.75"
53
+ instructions="oi"
54
+ en_de_pre_annealing="oi"
55
+ de_en_pre_annealing="oi"
56
+ en_fr_pre_annealing="oi"
57
+ fr_en_pre_annealing="oi"
58
+ en_es_pre_annealing="oi"
59
+ es_en_pre_annealing="oi"
60
+ en_it_pre_annealing="oi"
61
+ it_en_pre_annealing="oi"
62
+ en_nl_pre_annealing="oi"
63
+ nl_en_pre_annealing="oi"
64
+ en_pt_pre_annealing="oi"
65
+ pt_en_pre_annealing="oi"
66
+ en_ru_pre_annealing="oi"
67
+ ru_en_pre_annealing="oi"
68
+ en_zh_pre_annealing="oi"
69
+ zh_en_pre_annealing="oi"
70
+ en_ko_pre_annealing="oi"
71
+ ko_en_pre_annealing="oi"
72
+ en_de_wmt="oi"
73
+ en_ru_wmt="oi"
74
+ en_zh_wmt="oi"
75
+ )
76
+
77
+ is_hf_dataset=(Dataset:
78
+ en=True
79
+ es=False
80
+ de=False
81
+ fr=False
82
+ nl=False
83
+ pt=False
84
+ it=False
85
+ ru=False
86
+ zh=False
87
+ ko=False
88
+ en_de=False
89
+ de_en=False
90
+ en_fr=False
91
+ fr_en=False
92
+ en_es=False
93
+ es_en=False
94
+ en_it=False
95
+ it_en=False
96
+ en_nl=False
97
+ nl_en=False
98
+ en_pt=False
99
+ pt_en=False
100
+ en_ru=False
101
+ ru_en=False
102
+ en_zh=False
103
+ zh_en=False
104
+ en_ko=False
105
+ ko_en=False
106
+ en_synth=False
107
+ es_synth=False
108
+ de_synth=False
109
+ fr_synth=False
110
+ nl_synth=False
111
+ pt_synth=False
112
+ it_synth=False
113
+ ru_synth=False
114
+ zh_synth=False
115
+ ko_synth=False
116
+ instructions="oi"
117
+ en_de_pre_annealing="oi"
118
+ de_en_pre_annealing="oi"
119
+ en_fr_pre_annealing="oi"
120
+ fr_en_pre_annealing="oi"
121
+ en_es_pre_annealing="oi"
122
+ es_en_pre_annealing="oi"
123
+ en_it_pre_annealing="oi"
124
+ it_en_pre_annealing="oi"
125
+ en_nl_pre_annealing="oi"
126
+ nl_en_pre_annealing="oi"
127
+ en_pt_pre_annealing="oi"
128
+ pt_en_pre_annealing="oi"
129
+ en_ru_pre_annealing="oi"
130
+ ru_en_pre_annealing="oi"
131
+ en_zh_pre_annealing="oi"
132
+ zh_en_pre_annealing="oi"
133
+ en_ko_pre_annealing="oi"
134
+ ko_en_pre_annealing="oi"
135
+ en_de_wmt="oi"
136
+ en_ru_wmt="oi"
137
+ en_zh_wmt="oi"
138
+ )
139
+
140
+ threshold=(Dataset:
141
+ en=516
142
+ es=275
143
+ de=611
144
+ fr=322
145
+ nl=649
146
+ pt=257
147
+ it=332
148
+ ru=334
149
+ zh=2041
150
+ ko=198
151
+ en_de=100000
152
+ de_en=100000
153
+ en_fr=100000
154
+ fr_en=100000
155
+ en_es=100000
156
+ es_en=100000
157
+ en_it=100000
158
+ it_en=100000
159
+ en_nl=100000
160
+ nl_en=100000
161
+ en_pt=100000
162
+ pt_en=100000
163
+ en_ru=100000
164
+ ru_en=100000
165
+ en_zh=100000
166
+ zh_en=100000
167
+ en_ko=100000
168
+ ko_en=100000
169
+ en_synth=100000
170
+ es_synth=100000
171
+ de_synth=100000
172
+ fr_synth=100000
173
+ nl_synth=100000
174
+ pt_synth=100000
175
+ it_synth=100000
176
+ ru_synth=100000
177
+ zh_synth=100000
178
+ ko_synth=100000
179
+ instructions="oi"
180
+ en_de_pre_annealing="oi"
181
+ de_en_pre_annealing="oi"
182
+ en_fr_pre_annealing="oi"
183
+ fr_en_pre_annealing="oi"
184
+ en_es_pre_annealing="oi"
185
+ es_en_pre_annealing="oi"
186
+ en_it_pre_annealing="oi"
187
+ it_en_pre_annealing="oi"
188
+ en_nl_pre_annealing="oi"
189
+ nl_en_pre_annealing="oi"
190
+ en_pt_pre_annealing="oi"
191
+ pt_en_pre_annealing="oi"
192
+ en_ru_pre_annealing="oi"
193
+ ru_en_pre_annealing="oi"
194
+ en_zh_pre_annealing="oi"
195
+ zh_en_pre_annealing="oi"
196
+ en_ko_pre_annealing="oi"
197
+ ko_en_pre_annealing="oi"
198
+ en_de_wmt="oi"
199
+ en_ru_wmt="oi"
200
+ en_zh_wmt="oi"
201
+ )
202
+
203
+ # rougly 67% for mc4, 33% for total parallel data
204
+ datamix_weights=(
205
+ DataMix:
206
+ mc4_parallel_uniform=(
207
+ Dataset:
208
+ en=603
209
+ es=603
210
+ de=603
211
+ fr=603
212
+ nl=603
213
+ pt=603
214
+ it=603
215
+ ru=603
216
+ zh=603
217
+ ko=603
218
+ en_de=0
219
+ de_en=0
220
+ en_fr=0
221
+ fr_en=0
222
+ en_es=0
223
+ es_en=0
224
+ en_it=0
225
+ it_en=0
226
+ en_nl=0
227
+ nl_en=0
228
+ en_pt=0
229
+ pt_en=0
230
+ en_ru=0
231
+ ru_en=0
232
+ en_zh=0
233
+ zh_en=0
234
+ en_ko=0
235
+ ko_en=0
236
+ en_synth=67
237
+ es_synth=67
238
+ de_synth=67
239
+ fr_synth=67
240
+ nl_synth=67
241
+ pt_synth=67
242
+ it_synth=67
243
+ ru_synth=67
244
+ zh_synth=67
245
+ ko_synth=67
246
+ instructions=0
247
+ en_de_pre_annealing=183
248
+ de_en_pre_annealing=183
249
+ en_fr_pre_annealing=183
250
+ fr_en_pre_annealing=183
251
+ en_es_pre_annealing=183
252
+ es_en_pre_annealing=183
253
+ en_it_pre_annealing=183
254
+ it_en_pre_annealing=183
255
+ en_nl_pre_annealing=183
256
+ nl_en_pre_annealing=183
257
+ en_pt_pre_annealing=183
258
+ pt_en_pre_annealing=183
259
+ en_ru_pre_annealing=183
260
+ ru_en_pre_annealing=183
261
+ en_zh_pre_annealing=183
262
+ zh_en_pre_annealing=183
263
+ en_ko_pre_annealing=183
264
+ ko_en_pre_annealing=183
265
+ en_de_wmt=0
266
+ en_ru_wmt=0
267
+ en_zh_wmt=0
268
+ )
269
+ )
270
+
271
+ datamix_weights_annealing=(
272
+ DataMix:
273
+ mc4_parallel_uniform=(
274
+ Dataset:
275
+ en=0
276
+ es=0
277
+ de=0
278
+ fr=0
279
+ nl=0
280
+ pt=0
281
+ it=0
282
+ ru=0
283
+ zh=0
284
+ ko=0
285
+ en_de_wmt=833
286
+ en_ru_wmt=833
287
+ en_zh_wmt=833
288
+ en_de=0
289
+ de_en=833
290
+ en_fr=833
291
+ fr_en=833
292
+ en_es=833
293
+ es_en=833
294
+ en_it=833
295
+ it_en=833
296
+ en_nl=833
297
+ nl_en=833
298
+ en_pt=833
299
+ pt_en=833
300
+ en_ru=0
301
+ ru_en=833
302
+ en_zh=0
303
+ zh_en=833
304
+ en_ko=833
305
+ ko_en=833
306
+ en_synth=0
307
+ es_synth=0
308
+ de_synth=0
309
+ fr_synth=0
310
+ nl_synth=0
311
+ pt_synth=0
312
+ it_synth=0
313
+ ru_synth=0
314
+ zh_synth=0
315
+ ko_synth=0
316
+ instructions=85000
317
+ en_de_pre_annealing=0
318
+ de_en_pre_annealing=0
319
+ en_fr_pre_annealing=0
320
+ fr_en_pre_annealing=0
321
+ en_es_pre_annealing=0
322
+ es_en_pre_annealing=0
323
+ en_it_pre_annealing=0
324
+ it_en_pre_annealing=0
325
+ en_nl_pre_annealing=0
326
+ nl_en_pre_annealing=0
327
+ en_pt_pre_annealing=0
328
+ pt_en_pre_annealing=0
329
+ en_ru_pre_annealing=0
330
+ ru_en_pre_annealing=0
331
+ en_zh_pre_annealing=0
332
+ zh_en_pre_annealing=0
333
+ en_ko_pre_annealing=0
334
+ ko_en_pre_annealing=0
335
+ )
336
+ )
337
+
338
+
339
+ # number such that final tokens for each language are around 1B
340
+ n_tokens=(Dataset:
341
+ en=1000000000
342
+ es=833333330
343
+ de=833333330
344
+ fr=833333330
345
+ nl=833333330
346
+ pt=833333330
347
+ it=833333330
348
+ ru=500000000
349
+ zh=13888888
350
+ ko=250000000
351
+ en_de_wmt="oi"
352
+ en_ru_wmt="oi"
353
+ en_zh_wmt="oi"
354
+ en_de=20000000
355
+ de_en=20000000
356
+ en_fr=20000000
357
+ fr_en=20000000
358
+ en_es=20000000
359
+ es_en=20000000
360
+ en_it=20000000
361
+ it_en=20000000
362
+ en_nl=20000000
363
+ nl_en=20000000
364
+ en_pt=20000000
365
+ pt_en=20000000
366
+ en_ru=20000000
367
+ ru_en=20000000
368
+ en_zh=20000000
369
+ zh_en=20000000
370
+ en_ko=20000000
371
+ ko_en=20000000
372
+ en_synth=20000000
373
+ es_synth=20000000
374
+ de_synth=20000000
375
+ fr_synth=20000000
376
+ nl_synth=20000000
377
+ pt_synth=20000000
378
+ it_synth=20000000
379
+ ru_synth=20000000
380
+ zh_synth=20000000
381
+ ko_synth=20000000
382
+ instructions="oi"
383
+ en_de_pre_annealing="oi"
384
+ de_en_pre_annealing="oi"
385
+ en_fr_pre_annealing="oi"
386
+ fr_en_pre_annealing="oi"
387
+ en_es_pre_annealing="oi"
388
+ es_en_pre_annealing="oi"
389
+ en_it_pre_annealing="oi"
390
+ it_en_pre_annealing="oi"
391
+ en_nl_pre_annealing="oi"
392
+ nl_en_pre_annealing="oi"
393
+ en_pt_pre_annealing="oi"
394
+ pt_en_pre_annealing="oi"
395
+ en_ru_pre_annealing="oi"
396
+ ru_en_pre_annealing="oi"
397
+ en_zh_pre_annealing="oi"
398
+ zh_en_pre_annealing="oi"
399
+ en_ko_pre_annealing="oi"
400
+ ko_en_pre_annealing="oi"
401
+ )
402
+
403
+ is_parallel=(Dataset:
404
+ en=False
405
+ es=False
406
+ de=False
407
+ fr=False
408
+ nl=False
409
+ pt=False
410
+ it=False
411
+ ru=False
412
+ zh=False
413
+ ko=False
414
+ en_de_wmt="oi"
415
+ en_ru_wmt="oi"
416
+ en_zh_wmt="oi"
417
+ en_de=True
418
+ de_en=True
419
+ en_fr=True
420
+ fr_en=True
421
+ en_es=True
422
+ es_en=True
423
+ en_it=True
424
+ it_en=True
425
+ en_nl=True
426
+ nl_en=True
427
+ en_pt=True
428
+ pt_en=True
429
+ en_ru=True
430
+ ru_en=True
431
+ en_zh=True
432
+ zh_en=True
433
+ en_ko=True
434
+ ko_en=True
435
+ en_synth=False
436
+ es_synth=False
437
+ de_synth=False
438
+ fr_synth=False
439
+ nl_synth=False
440
+ pt_synth=False
441
+ it_synth=False
442
+ ru_synth=False
443
+ zh_synth=False
444
+ ko_synth=False
445
+ instructions="oi"
446
+ en_de_pre_annealing="oi"
447
+ de_en_pre_annealing="oi"
448
+ en_fr_pre_annealing="oi"
449
+ fr_en_pre_annealing="oi"
450
+ en_es_pre_annealing="oi"
451
+ es_en_pre_annealing="oi"
452
+ en_it_pre_annealing="oi"
453
+ it_en_pre_annealing="oi"
454
+ en_nl_pre_annealing="oi"
455
+ nl_en_pre_annealing="oi"
456
+ en_pt_pre_annealing="oi"
457
+ pt_en_pre_annealing="oi"
458
+ en_ru_pre_annealing="oi"
459
+ ru_en_pre_annealing="oi"
460
+ en_zh_pre_annealing="oi"
461
+ zh_en_pre_annealing="oi"
462
+ en_ko_pre_annealing="oi"
463
+ ko_en_pre_annealing="oi"
464
+ )
465
+
466
+ lp=(Dataset:
467
+ en=""
468
+ es=""
469
+ de=""
470
+ fr=""
471
+ nl=""
472
+ pt=""
473
+ it=""
474
+ ru=""
475
+ zh=""
476
+ ko=""
477
+ en_de_wmt="oi"
478
+ en_ru_wmt="oi"
479
+ en_zh_wmt="oi"
480
+ en_de="en-de"
481
+ de_en="de-en"
482
+ en_fr="en-fr"
483
+ fr_en="fr-en"
484
+ en_es="en-es"
485
+ es_en="es-en"
486
+ en_it="en-it"
487
+ it_en="it-en"
488
+ en_nl="en-nl"
489
+ nl_en="nl-en"
490
+ en_pt="en-pt"
491
+ pt_en="pt-en"
492
+ en_ru="en-ru"
493
+ ru_en="ru-en"
494
+ en_zh="en-zh"
495
+ zh_en="zh-en"
496
+ en_ko="en-ko"
497
+ ko_en="ko-en"
498
+ en_synth=""
499
+ es_synth=""
500
+ de_synth=""
501
+ fr_synth=""
502
+ nl_synth=""
503
+ pt_synth=""
504
+ it_synth=""
505
+ ru_synth=""
506
+ zh_synth=""
507
+ ko_synth=""
508
+ instructions="oi"
509
+ en_de_pre_annealing="oi"
510
+ de_en_pre_annealing="oi"
511
+ en_fr_pre_annealing="oi"
512
+ fr_en_pre_annealing="oi"
513
+ en_es_pre_annealing="oi"
514
+ es_en_pre_annealing="oi"
515
+ en_it_pre_annealing="oi"
516
+ it_en_pre_annealing="oi"
517
+ en_nl_pre_annealing="oi"
518
+ nl_en_pre_annealing="oi"
519
+ en_pt_pre_annealing="oi"
520
+ pt_en_pre_annealing="oi"
521
+ en_ru_pre_annealing="oi"
522
+ ru_en_pre_annealing="oi"
523
+ en_zh_pre_annealing="oi"
524
+ zh_en_pre_annealing="oi"
525
+ en_ko_pre_annealing="oi"
526
+ ko_en_pre_annealing="oi"
527
+ )
528
+
529
+ min_perplexity=50
530
+
531
+ size=(Size: 8)
532
+
533
+ log_interval=1
534
+ save_interval=635
535
+ eval_interval=635
536
+ train_steps=11430
537
+ train_steps_annealing=1270
538
+
539
+ lr_scheduler=constant
540
+ warmup_steps=32
541
+ lr=3e-5
542
+ lr_min=3e-6
543
+ weight_decay=0.1
544
+
545
+ lr_scheduler_annealing=linear
546
+ warmup_steps_annealing=0
547
+ lr_annealing=3e-5
548
+ lr_min_annealing=3e-6
549
+
550
+ n_gpus=8
551
+ gpu_ids=0,1,2,3,4,5,6,7
552
+ tp=(TP: 1 2 3 4 5 6 7 8)
553
+ pp=(PP: 1 2 3 4)
554
+ micro_batch_size=4
555
+ grad_accum_steps=12
556
+ vocab_size=128256
557
+
558
+ cpu_workers=16
559
+ wikipedia=False
560
+ freeze_layers=""
561
+ posterior_tokens=False
562
+ n_posterior_tokens=0
563
+ eval_iters=1
564
+
565
+ seq_length=4096
566
+
567
+ glu_activation=swiglu
568
+ kv_channels=""
569
+ layernorm_epsilon=1e-5
570
+ }