manu commited on
Commit
d8da048
1 Parent(s): 4a38dd0

Upload config.yaml with huggingface_hub

Browse files
Files changed (1) hide show
  1. config.yaml +69 -0
config.yaml ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data_mix:
2
+ (): dataset_construction.DataMix
3
+ name: "code_30b"
4
+ shuffle: false
5
+ local_save_dir: data/
6
+ load_from_local_save_dir: false
7
+ compute_dataset_stats: true
8
+ keep_separated_datasets_in_dataset_dict: false
9
+ deduplicate_test_set: false
10
+ ngram_path_for_extra_deduplication: null
11
+ max_shard_size: "2GB"
12
+ datasets:
13
+ # Code
14
+ - (): dataset_construction.DatasetConfig
15
+ dataset_path: 'bigcode/starcoderdata'
16
+ train_split: "train[:20%]"
17
+ text_column: "content"
18
+ dataset_kwargs:
19
+ data_dir: python
20
+ build_test_set_from_train: true
21
+ - (): dataset_construction.DatasetConfig
22
+ dataset_path: 'bigcode/starcoderdata'
23
+ train_split: "train[:20%]"
24
+ text_column: "content"
25
+ dataset_kwargs:
26
+ data_dir: markdown
27
+ build_test_set_from_train: true
28
+ - (): dataset_construction.DatasetConfig
29
+ dataset_path: 'bigcode/starcoderdata'
30
+ train_split: "train[:20%]"
31
+ text_column: "content"
32
+ dataset_kwargs:
33
+ data_dir: jupyter-scripts-dedup-filtered
34
+ build_test_set_from_train: true
35
+ - (): dataset_construction.DatasetConfig
36
+ dataset_path: 'bigcode/starcoderdata'
37
+ train_split: "train[:20%]"
38
+ text_column: "content"
39
+ dataset_kwargs:
40
+ data_dir: jupyter-structured-clean-dedup
41
+ build_test_set_from_train: true
42
+ - (): dataset_construction.DatasetConfig
43
+ dataset_path: 'bigcode/starcoderdata'
44
+ train_split: "train[:15%]"
45
+ text_column: "content"
46
+ dataset_kwargs:
47
+ data_dir: json
48
+ build_test_set_from_train: true
49
+ - (): dataset_construction.DatasetConfig
50
+ dataset_path: 'teven/code_contests'
51
+ train_split: "train"
52
+ # build_test_set_from_train: true
53
+ build_test_set_from_train: false
54
+ test_split: "valid"
55
+ # needs to be debugged
56
+ filtering_function:
57
+ (): dataset_collection.code.deepmind_code_contest.CodeContestFilter
58
+ preprocessing_function:
59
+ (): dataset_collection.code.deepmind_code_contest.CodeContestMapper
60
+ - (): dataset_construction.DatasetConfig
61
+ dataset_path: 'vikp/pypi_clean'
62
+ train_split: "train"
63
+ build_test_set_from_train: true
64
+ text_column: code
65
+ id_column: path
66
+ # we could add synthetic textbook datasets or code with explanations !
67
+
68
+
69
+ tokenizer: "mistralai/Mistral-7B-v0.1"