froggeric commited on
Commit
db3151d
1 Parent(s): 717ca43

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +29 -16
README.md CHANGED
@@ -20,41 +20,54 @@ You will find below descriptions for the various input files provided, to help y
20
 
21
  ## Community provided files
22
 
23
- ### 8k_random_data
24
 
25
- ### 20k_random_data
26
 
27
- ### groups_merged
28
 
29
- ### group_10_merged
30
 
31
- ### ptb.train
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32
 
33
  ## exllamav2 calibration data
34
 
35
  https://github.com/turboderp/exllamav2/tree/master/conversion/standard_cal_data
36
 
37
- ### c4
38
-
39
- ### code
40
 
 
41
  Programming
42
 
43
- ### multilingual
44
-
45
  English, Arabic, Chinese, French, German, Japanese, Polish, Russian, Spanish, Swedish, Turkish, Hebrew,
46
  Macedonian, Norwegian, Lithuanian, Greek, Italian, Afrikaans, Dutch, Danish.
47
 
48
- ### technical
49
-
50
  Technical writing.
51
 
52
- ### tiny
53
-
54
  Very short stories.
55
 
56
- ### wiki
57
-
58
  Wikipedia dump.
59
 
60
  ## How to quantize with an imatrix in llama.cpp
 
20
 
21
  ## Community provided files
22
 
23
+ **8k_random_data**\
24
 
25
+ **20k_random_data**\
26
 
27
+ **groups_merged**\
28
 
29
+ **group_10_merged**\
30
 
31
+ **ptb.train**\
32
+ Penn Treebank (PTB) is a widely used preprocessed large dataset designed for language training. Casing,
33
+ punctuation and numbers have been removed from the training data. Recently it has kind of been superseeded
34
+ by WikiText which does not have these removals, features a larger vocabulary and full articles (better
35
+ suited for models that can take advantage of long term dependencies). However, for importantce matrix training,
36
+ PTB is still a valid dataset, which has the advantage of being manually curated, and similar to WikiText,
37
+ without being WikiText; this can help against bias.
38
+
39
+ **WikiText**\
40
+ The WikiText language modeling dataset is a collection of over 100 million tokens extracted from the set of
41
+ verified Good and Featured articles on Wikipedia. Compared to PTB, WikiText-2 is over 2 times larger and
42
+ WikiText-103 is over 110 times larger. As it is composed of full articles, the dataset is well suited for models
43
+ that can take advantage of long term dependencies.\
44
+ https://huggingface.co/datasets/wikitext
45
+
46
+ **WikiText_FR**\
47
+ 70 million tokens extracted from the set of french Wikipedia articles that are classified as "quality articles"
48
+ or "good articles".\
49
+ https://huggingface.co/datasets/asi/wikitext_fr
50
 
51
  ## exllamav2 calibration data
52
 
53
  https://github.com/turboderp/exllamav2/tree/master/conversion/standard_cal_data
54
 
55
+ **c4**\
 
 
56
 
57
+ **code**\
58
  Programming
59
 
60
+ **multilingual**\
 
61
  English, Arabic, Chinese, French, German, Japanese, Polish, Russian, Spanish, Swedish, Turkish, Hebrew,
62
  Macedonian, Norwegian, Lithuanian, Greek, Italian, Afrikaans, Dutch, Danish.
63
 
64
+ **technical**\
 
65
  Technical writing.
66
 
67
+ **tiny**\
 
68
  Very short stories.
69
 
70
+ **wiki**\
 
71
  Wikipedia dump.
72
 
73
  ## How to quantize with an imatrix in llama.cpp