clefourrier HF Staff commited on
Commit
11538ef
·
1 Parent(s): 8978a83

Create sacrebleu_manual.py

Browse files
Files changed (1) hide show
  1. sacrebleu_manual.py +69 -0
sacrebleu_manual.py ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ NOTE: This file implements translation tasks using datasets from WMT conferences,
3
+ provided by sacrebleu. Traditionally they are evaluated with BLEU scores. TER
4
+ and CHRF are other options.
5
+
6
+ We defer citations and descriptions of the many translations tasks used
7
+ here to the SacreBLEU repo from which we've obtained the datasets:
8
+ https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/dataset.py
9
+
10
+ Homepage: https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/dataset.py
11
+ """
12
+ import pycountry
13
+ from sacrebleu import sacrebleu
14
+ import datasets
15
+
16
+
17
+ _CITATION = """
18
+ @inproceedings{post-2018-call,
19
+ title = "A Call for Clarity in Reporting {BLEU} Scores",
20
+ author = "Post, Matt",
21
+ booktitle = "Proceedings of the Third Conference on Machine Translation: Research Papers",
22
+ month = oct,
23
+ year = "2018",
24
+ address = "Belgium, Brussels",
25
+ publisher = "Association for Computational Linguistics",
26
+ url = "https://www.aclweb.org/anthology/W18-6319",
27
+ pages = "186--191",
28
+ }
29
+ """
30
+
31
+
32
+ class Sacrebleu(datasets.GeneratorBasedBuilder):
33
+ """The Pile is a 825 GiB diverse, open source language modeling dataset."""
34
+
35
+ VERSION = datasets.Version("0.0.1")
36
+
37
+ BUILDER_CONFIGS = [
38
+ datasets.BuilderConfig(name=f"{name}/{langpair}", version=VERSION, description="")
39
+ for langpair in sacrebleu.get_langpairs_for_testset(name)
40
+ for name in sacrebleu.get_available_testsets()
41
+ ]
42
+
43
+ def _info(self):
44
+ features = datasets.Features(
45
+ {
46
+ "translation": datasets.Value("dict"),
47
+ }
48
+ )
49
+ return datasets.DatasetInfo(
50
+ description=f"{_DESCRIPTION}\n{self.config.description}",
51
+ features=features,
52
+ homepage="",
53
+ license="",
54
+ citation=_CITATION,
55
+ )
56
+
57
+ def _split_generators(self, dl_manager):
58
+ return [
59
+ datasets.SplitGenerator(
60
+ name=datasets.Split.TEST,
61
+ gen_kwargs={path: f"{self.config.name}.jsonl"},
62
+ )
63
+ ]
64
+
65
+ # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
66
+ def _generate_examples(self, path):
67
+ with open(path, encoding="utf-8") as f:
68
+ for key, row in enumerate(f):
69
+ yield key, json.loads(row)