hylcool commited on
Commit
fdebe7c
1 Parent(s): 2be23de

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +76 -0
README.md CHANGED
@@ -1,3 +1,79 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - data-juicer
9
+ - pretraining
10
+ size_categories:
11
+ - 100M<n<1B
12
  ---
13
+
14
+ # RedPajama -- C4 (refined by Data-Juicer)
15
+
16
+ A refined version of C4 dataset in RedPajama by [Data-Juicer](https://github.com/alibaba/data-juicer). Removing some "bad" samples from the original dataset to make it higher-quality.
17
+
18
+ This dataset is usually used to pretrain a Large Language Model.
19
+
20
+ **Notice**: Here is a small subset for previewing. The whole dataset is available [here](https://dail-wlcb.oss-cn-wulanchabu.aliyuncs.com/LLM_data/our_refined_datasets/pretraining/redpajama-c4-refine-result.jsonl) (About 832GB).
21
+
22
+ ## Dataset Information
23
+
24
+ - Number of samples: 344,491,171 (Keep ~94.42% from the original dataset)
25
+
26
+ ## Refining Recipe
27
+ ```yaml
28
+ # global parameters
29
+ project_name: 'Data-Juicer-recipes-c4'
30
+ dataset_path: '/path/to/your/dataset' # path to your dataset directory or file
31
+ export_path: '/path/to/your/dataset.jsonl' # path to your dataset result file
32
+
33
+ np: 50 # number of subprocess to process your dataset
34
+ open_tracer: True
35
+
36
+ # process schedule
37
+ # a list of several process operators with their arguments
38
+ process:
39
+ - clean_email_mapper:
40
+ - clean_links_mapper:
41
+ - fix_unicode_mapper:
42
+ - punctuation_normalization_mapper:
43
+ - whitespace_normalization_mapper:
44
+
45
+ - alphanumeric_filter:
46
+ tokenization: false
47
+ min_ratio: 0.65 # <3sigma (0.740)
48
+ max_ratio: 0.9 # >3sigma (0.867)
49
+ - average_line_length_filter: # for code
50
+ max_len: 3000 # >3sigma (1277)
51
+ - character_repetition_filter:
52
+ rep_len: 10
53
+ max_ratio: 0.3 # >3sigma (0.168)
54
+ - language_id_score_filter:
55
+ min_score: 0.6
56
+ - maximum_line_length_filter: # for code
57
+ max_len: 4000 # >3sigma (2017)
58
+ - perplexity_filter:
59
+ lang: en
60
+ max_ppl: 6000 #(>3sigma 4543)
61
+ - special_characters_filter:
62
+ max_ratio: 0.4 # > 3sigma (0.303)
63
+ - words_num_filter:
64
+ tokenization: true
65
+ min_num: 20
66
+ max_num: 10000
67
+ - word_repetition_filter:
68
+ lang: en
69
+ tokenization: true
70
+ rep_len: 10
71
+ max_ratio: 0.231 # 3sigma
72
+ - document_simhash_deduplicator:
73
+ tokenization: space
74
+ window_size: 6
75
+ lowercase: true
76
+ ignore_pattern: '\p{P}'
77
+ num_blocks: 6
78
+ hamming_distance: 4
79
+ ```