mathiascreutz
commited on
Commit
•
ee6c327
1
Parent(s):
daced4d
More comments
Browse files- opusparcus.py +31 -14
opusparcus.py
CHANGED
@@ -41,15 +41,16 @@ _HOMEPAGE = ""
|
|
41 |
|
42 |
_LICENSE = ""
|
43 |
|
44 |
-
# The HuggingFace dataset library doesn't host the datasets but only
|
45 |
-
# This can be an arbitrary nested
|
46 |
-
|
47 |
_URLs = {
|
48 |
"validation": "validation.jsonl",
|
49 |
"test": "test.jsonl",
|
50 |
"validation.full": "validation.jsonl",
|
51 |
"test.full": "test.jsonl",
|
52 |
-
# NB: the "train" split file is defined dynamically inside the
|
|
|
53 |
}
|
54 |
|
55 |
_VERSION = datasets.Version("1.0.0", "")
|
@@ -73,8 +74,22 @@ class OpusparcusConfig(datasets.BuilderConfig):
|
|
73 |
self.lang = lang
|
74 |
self.quality = quality
|
75 |
|
|
|
|
|
|
|
76 |
LANGS = [ "de", "en", "fi", "fr", "ru", "sv" ]
|
77 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
QUALITIES = [ 100, 95, 90, 85, 80, 75, 70, 65, 60 ]
|
79 |
|
80 |
class Opusparcus(datasets.GeneratorBasedBuilder):
|
@@ -95,7 +110,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
95 |
]
|
96 |
|
97 |
# There is no default configuration. User always needs to specify one:
|
98 |
-
#DEFAULT_CONFIG_NAME = None
|
99 |
|
100 |
def _info(self):
|
101 |
# This method specifies the datasets.DatasetInfo object which
|
@@ -139,7 +154,8 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
139 |
# self.config.lang and self.config.quality.
|
140 |
|
141 |
if self.config.lang is None:
|
142 |
-
# This is an error
|
|
|
143 |
return []
|
144 |
|
145 |
# Select which file of the training data contains the matching data:
|
@@ -212,7 +228,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
212 |
|
213 |
# If the desired quality value is 100, no subset of the
|
214 |
# training set is good enough, and we only produce validation
|
215 |
-
# and test sets, in order to save space and time
|
216 |
|
217 |
if self.config.quality <= 95:
|
218 |
# In this case there is matching training data, so we produce
|
@@ -243,7 +259,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
243 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
244 |
|
245 |
if split == datasets.Split.TRAIN:
|
246 |
-
# Training sets are in compressed
|
247 |
# They contain a field "quality" missing from the validation and test sets.
|
248 |
# We also know that this file only contains the desired language,
|
249 |
# because for the training sets the languages are in separate
|
@@ -264,10 +280,11 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
264 |
}
|
265 |
else:
|
266 |
# The validation and test sets are in jsonl files.
|
267 |
-
# They contain the fields "lang" and "annot_score" that we
|
268 |
-
# If we ask for the full sets, we will keep
|
269 |
-
#
|
270 |
-
#
|
|
|
271 |
keep_all = (split == "validation.full" or split == "test.full")
|
272 |
with open(filepath, encoding="utf-8") as f:
|
273 |
for id_, row in enumerate(f):
|
@@ -276,8 +293,8 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
276 |
if keep_all or data["annot_score"] >= 3.0:
|
277 |
# for full sets keep all;
|
278 |
# for standard test and validation sets, keep only
|
279 |
-
# the paraphrases (annot_score >= 3.0 means
|
280 |
-
# or mostly good example of paraphrases")
|
281 |
yield id_, {
|
282 |
"lang": data["lang"],
|
283 |
"sent1": data["sent1"],
|
|
|
41 |
|
42 |
_LICENSE = ""
|
43 |
|
44 |
+
# The HuggingFace dataset library doesn't host the datasets but only
|
45 |
+
# points to the original files. This can be an arbitrary nested
|
46 |
+
# dict/list of URLs (see below in `_split_generators` method):
|
47 |
_URLs = {
|
48 |
"validation": "validation.jsonl",
|
49 |
"test": "test.jsonl",
|
50 |
"validation.full": "validation.jsonl",
|
51 |
"test.full": "test.jsonl",
|
52 |
+
# NB: the "train" split file is defined dynamically inside the
|
53 |
+
# `_split_generators` method
|
54 |
}
|
55 |
|
56 |
_VERSION = datasets.Version("1.0.0", "")
|
|
|
74 |
self.lang = lang
|
75 |
self.quality = quality
|
76 |
|
77 |
+
|
78 |
+
# Languages in Opusparcus: German (de), English (en), Finnish (fi),
|
79 |
+
# French (fr), Russian (ru), Swedish (sv):
|
80 |
LANGS = [ "de", "en", "fi", "fr", "ru", "sv" ]
|
81 |
|
82 |
+
# The training sets (train splits) come in eight sizes (95 .. 60),
|
83 |
+
# where the number indicates the estimated proportion [%] of true
|
84 |
+
# paraphrases in the set. The higher the number the smaller (but
|
85 |
+
# ideally cleaner) the set. The lower the number, the larger (but
|
86 |
+
# noisier) the set is. The smaller sets are included as subsets of
|
87 |
+
# larger sets. The special value 100 matches no training data at all,
|
88 |
+
# so if you are only interested in validation and test sets, you can
|
89 |
+
# use the value 100 in order to save time and space. (The quality
|
90 |
+
# value is irrelevant for the validation and test sets, which have
|
91 |
+
# been annotated manually, and each example has an annotation score
|
92 |
+
# attached to it.)
|
93 |
QUALITIES = [ 100, 95, 90, 85, 80, 75, 70, 65, 60 ]
|
94 |
|
95 |
class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
|
110 |
]
|
111 |
|
112 |
# There is no default configuration. User always needs to specify one:
|
113 |
+
# DEFAULT_CONFIG_NAME = None
|
114 |
|
115 |
def _info(self):
|
116 |
# This method specifies the datasets.DatasetInfo object which
|
|
|
154 |
# self.config.lang and self.config.quality.
|
155 |
|
156 |
if self.config.lang is None:
|
157 |
+
# This is an error: nothing to do here if no language
|
158 |
+
# has been defined:
|
159 |
return []
|
160 |
|
161 |
# Select which file of the training data contains the matching data:
|
|
|
228 |
|
229 |
# If the desired quality value is 100, no subset of the
|
230 |
# training set is good enough, and we only produce validation
|
231 |
+
# and test sets, in order to save space and time:
|
232 |
|
233 |
if self.config.quality <= 95:
|
234 |
# In this case there is matching training data, so we produce
|
|
|
259 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
260 |
|
261 |
if split == datasets.Split.TRAIN:
|
262 |
+
# Training sets are in jsonl files that have been compressed using bzip2.
|
263 |
# They contain a field "quality" missing from the validation and test sets.
|
264 |
# We also know that this file only contains the desired language,
|
265 |
# because for the training sets the languages are in separate
|
|
|
280 |
}
|
281 |
else:
|
282 |
# The validation and test sets are in jsonl files.
|
283 |
+
# They contain the fields "lang" and "annot_score" that we
|
284 |
+
# filter on. If we ask for the full sets, we will keep
|
285 |
+
# all data entries for the desired language, also the
|
286 |
+
# sentence pairs that were not considered paraphrases by
|
287 |
+
# the annotators:
|
288 |
keep_all = (split == "validation.full" or split == "test.full")
|
289 |
with open(filepath, encoding="utf-8") as f:
|
290 |
for id_, row in enumerate(f):
|
|
|
293 |
if keep_all or data["annot_score"] >= 3.0:
|
294 |
# for full sets keep all;
|
295 |
# for standard test and validation sets, keep only
|
296 |
+
# the actual paraphrases (annot_score >= 3.0 means
|
297 |
+
# "good or mostly good example of paraphrases")
|
298 |
yield id_, {
|
299 |
"lang": data["lang"],
|
300 |
"sent1": data["sent1"],
|