SEE7 commited on
Commit
bc24c8c
·
verified ·
1 Parent(s): e0951ef

Publish 5000 real globally deduplicated translation pairs with provenance and checksums

Browse files
README.md ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: AstheralDataset
3
+ language:
4
+ - en
5
+ - de
6
+ - fr
7
+ - es
8
+ - it
9
+ - pt
10
+ task_categories:
11
+ - translation
12
+ size_categories:
13
+ - 1K<n<10K
14
+ license: other
15
+ license_name: source-corpus-terms
16
+ license_link: https://huggingface.co/datasets/Helsinki-NLP/opus-100
17
+ configs:
18
+ - config_name: default
19
+ data_files:
20
+ - split: train
21
+ path: data/train.parquet
22
+ - split: validation
23
+ path: data/validation.parquet
24
+ - split: test
25
+ path: data/test.parquet
26
+ ---
27
+ # AstheralDataset
28
+
29
+ **5,000 real English translation pairs**, curated by SEE7 for German, French, Spanish, Italian and Portuguese.
30
+
31
+ ## Contents
32
+
33
+ | Target | Train | Validation | Test | Total |
34
+ | --- | ---: | ---: | ---: | ---: |
35
+ | German | 900 | 50 | 50 | 1,000 |
36
+ | French | 900 | 50 | 50 | 1,000 |
37
+ | Spanish | 900 | 50 | 50 | 1,000 |
38
+ | Italian | 900 | 50 | 50 | 1,000 |
39
+ | Portuguese | 900 | 50 | 50 | 1,000 |
40
+ | **Total** | **4,500** | **250** | **250** | **5,000** |
41
+
42
+ ## Record fields
43
+
44
+ Each row contains a unique SHA-256 ID, English source text, translated target text, source and target language codes, a translation instruction, and the original dataset/config/split/revision/row index.
45
+
46
+ ## Curation
47
+
48
+ Text is drawn from real parallel sentences in [OPUS-100](https://huggingface.co/datasets/Helsinki-NLP/opus-100). Selection preserves the original train, validation and test split membership. All 10,000 text values are globally unique after Unicode normalization, case folding, punctuation removal and whitespace normalization. Source-target identity, short fragments, markup and placeholder patterns are filtered. No model-generated translations are included.
49
+
50
+ The complete build script, exact source revision, split counts, rejection counts and Parquet checksums are published alongside the data.
51
+
52
+ ## Load
53
+
54
+ ```python
55
+ from datasets import load_dataset
56
+
57
+ dataset = load_dataset("SEE7/AstheralDataset")
58
+ example = dataset["train"][0]
59
+ print(example["source_text"])
60
+ print(example["target_text"])
61
+ ```
62
+
63
+ ## Provenance and attribution
64
+
65
+ Source: Helsinki-NLP/opus-100, revision 805090dc28bf78897da9641cdf08b61287580df9.
66
+ OPUS-100: Zhang et al. (2020), “Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation.” Original parallel text comes from the OPUS collection. The original source corpus terms remain applicable.
67
+
68
+ Curated dataset repository: SEE7/AstheralDataset.
build_dataset.py ADDED
@@ -0,0 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json,hashlib,re,unicodedata,collections,time
2
+ from pathlib import Path
3
+ from datasets import load_dataset,Dataset
4
+ from huggingface_hub import HfApi
5
+ api=HfApi();source='Helsinki-NLP/opus-100';revision=api.dataset_info(source).sha
6
+ out=Path('/tmp/AstheralDataset');(out/'data').mkdir(parents=True,exist_ok=True)
7
+ langs={'de':'German','fr':'French','es':'Spanish','it':'Italian','pt':'Portuguese'}
8
+ seen_values=set();ids=set();splits={'train':[],'validation':[],'test':[]};rejected=collections.Counter()
9
+ def normalize(s):
10
+ s=unicodedata.normalize('NFKC',s).casefold()
11
+ return ' '.join(''.join(c if c.isalnum() else ' ' for c in s).split())
12
+ for split,n in [('test',50),('validation',50),('train',900)]:
13
+ for lang,name in langs.items():
14
+ config='-'.join(sorted(['en',lang]));ds=load_dataset(source,config,split=split,revision=revision).shuffle(seed=20261113)
15
+ added=0
16
+ for index,row in enumerate(ds):
17
+ a=row['translation']['en'].strip();b=row['translation'][lang].strip()
18
+ if not(20<=len(a)<=500 and 12<=len(b)<=600) or len(a.split())<4 or len(b.split())<3:
19
+ rejected['length']+=1;continue
20
+ if any(re.search(r'lorem ipsum|placeholder|\bTODO\b|https?://|<[^>]+>|\{\{|\ufffd',s,re.I) for s in [a,b]):
21
+ rejected['non_sentence_content']+=1;continue
22
+ x,y=normalize(a),normalize(b)
23
+ if x==y or x in seen_values or y in seen_values:
24
+ rejected['duplicate_or_identical_text']+=1;continue
25
+ uid=hashlib.sha256((lang+'\\0'+x+'\\0'+y).encode()).hexdigest()
26
+ assert uid not in ids
27
+ ids.add(uid);seen_values.update([x,y])
28
+ record={'id':uid,'source_language':'en','target_language':lang,'source_text':a,'target_text':b,'instruction':'Translate from English to '+name+'. Return only the translation.','source_dataset':source,'source_config':config,'source_split':split,'source_revision':revision,'source_row_index':int(ds._indices.column(0)[index].as_py()) if ds._indices is not None else index}
29
+ splits[split].append(record);added+=1
30
+ if added==n:break
31
+ assert added==n,(lang,split,added)
32
+ print('COLLECTED',split,lang,added,flush=True)
33
+ for split,rows in splits.items():
34
+ Dataset.from_list(rows).shuffle(seed=42).to_parquet(str(out/'data'/f'{split}.parquet'))
35
+ total=sum(map(len,splits.values()));assert total==5000 and len(ids)==5000 and len(seen_values)==10000
36
+ counter={split:dict(collections.Counter(r['target_language'] for r in rows)) for split,rows in splits.items()}
37
+ manifest={'name':'AstheralDataset','version':'1.0','total_examples':total,'unique_example_ids':len(ids),'unique_normalized_text_values':len(seen_values),'split_counts':{k:len(v) for k,v in splits.items()},'language_counts':counter,'source_dataset':source,'source_revision':revision,'normalization':'Unicode NFKC, casefold, punctuation removed, whitespace collapsed','deduplication':'Global source/target text uniqueness across all languages and splits; identical source-target pairs excluded','rejected':dict(rejected),'file_sha256':{p.name:hashlib.sha256(p.read_bytes()).hexdigest() for p in (out/'data').glob('*.parquet')}}
38
+ (out/'manifest.json').write_text(json.dumps(manifest,indent=2))
39
+ (out/'build_dataset.py').write_text(Path(__file__).read_text())
40
+ readme='''---
41
+ pretty_name: AstheralDataset
42
+ language:
43
+ - en
44
+ - de
45
+ - fr
46
+ - es
47
+ - it
48
+ - pt
49
+ task_categories:
50
+ - translation
51
+ size_categories:
52
+ - 1K<n<10K
53
+ license: other
54
+ license_name: source-corpus-terms
55
+ license_link: https://huggingface.co/datasets/Helsinki-NLP/opus-100
56
+ configs:
57
+ - config_name: default
58
+ data_files:
59
+ - split: train
60
+ path: data/train.parquet
61
+ - split: validation
62
+ path: data/validation.parquet
63
+ - split: test
64
+ path: data/test.parquet
65
+ ---
66
+ # AstheralDataset
67
+
68
+ **5,000 real English translation pairs**, curated by SEE7 for German, French, Spanish, Italian and Portuguese.
69
+
70
+ ## Contents
71
+
72
+ | Target | Train | Validation | Test | Total |
73
+ | --- | ---: | ---: | ---: | ---: |
74
+ | German | 900 | 50 | 50 | 1,000 |
75
+ | French | 900 | 50 | 50 | 1,000 |
76
+ | Spanish | 900 | 50 | 50 | 1,000 |
77
+ | Italian | 900 | 50 | 50 | 1,000 |
78
+ | Portuguese | 900 | 50 | 50 | 1,000 |
79
+ | **Total** | **4,500** | **250** | **250** | **5,000** |
80
+
81
+ ## Record fields
82
+
83
+ Each row contains a unique SHA-256 ID, English source text, translated target text, source and target language codes, a translation instruction, and the original dataset/config/split/revision/row index.
84
+
85
+ ## Curation
86
+
87
+ Text is drawn from real parallel sentences in [OPUS-100](https://huggingface.co/datasets/Helsinki-NLP/opus-100). Selection preserves the original train, validation and test split membership. All 10,000 text values are globally unique after Unicode normalization, case folding, punctuation removal and whitespace normalization. Source-target identity, short fragments, markup and placeholder patterns are filtered. No model-generated translations are included.
88
+
89
+ The complete build script, exact source revision, split counts, rejection counts and Parquet checksums are published alongside the data.
90
+
91
+ ## Load
92
+
93
+ ```python
94
+ from datasets import load_dataset
95
+
96
+ dataset = load_dataset("SEE7/AstheralDataset")
97
+ example = dataset["train"][0]
98
+ print(example["source_text"])
99
+ print(example["target_text"])
100
+ ```
101
+
102
+ ## Provenance and attribution
103
+
104
+ Source: Helsinki-NLP/opus-100, revision '''+revision+'''.
105
+ OPUS-100: Zhang et al. (2020), “Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation.” Original parallel text comes from the OPUS collection. The original source corpus terms remain applicable.
106
+
107
+ Curated dataset repository: SEE7/AstheralDataset.
108
+ '''
109
+ (out/'README.md').write_text(readme)
110
+ api.upload_folder(repo_id='SEE7/AstheralDataset',repo_type='dataset',folder_path=str(out),commit_message='Publish 5000 real globally deduplicated translation pairs with provenance and checksums')
111
+ # Download the published splits through the public dataset interface and validate the real release.
112
+ published=load_dataset('SEE7/AstheralDataset')
113
+ assert {k:len(v) for k,v in published.items()}=={'train':4500,'validation':250,'test':250}
114
+ values=[];row_ids=[]
115
+ for split in published:
116
+ for row in published[split]:
117
+ values.extend([normalize(row['source_text']),normalize(row['target_text'])]);row_ids.append(row['id'])
118
+ assert len(set(values))==10000 and len(set(row_ids))==5000
119
+ print('PUBLISHED_VERIFIED',json.dumps(manifest),flush=True)
data/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ef098abdabd64007cb91451f25e9d452f076b897bc11ce464fc9157b7b8fffa3
3
+ size 60081
data/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:db2ccfff2cb499647cd134bdbf03840b9d72c98ccbadf5af25572a189cb2bcc5
3
+ size 938671
data/validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7e65a4bc923eadaafa3cf92b8af54aeadb1aafdd2265547c52be078f6131db55
3
+ size 60972
manifest.json ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "AstheralDataset",
3
+ "version": "1.0",
4
+ "total_examples": 5000,
5
+ "unique_example_ids": 5000,
6
+ "unique_normalized_text_values": 10000,
7
+ "split_counts": {
8
+ "train": 4500,
9
+ "validation": 250,
10
+ "test": 250
11
+ },
12
+ "language_counts": {
13
+ "train": {
14
+ "de": 900,
15
+ "fr": 900,
16
+ "es": 900,
17
+ "it": 900,
18
+ "pt": 900
19
+ },
20
+ "validation": {
21
+ "de": 50,
22
+ "fr": 50,
23
+ "es": 50,
24
+ "it": 50,
25
+ "pt": 50
26
+ },
27
+ "test": {
28
+ "de": 50,
29
+ "fr": 50,
30
+ "es": 50,
31
+ "it": 50,
32
+ "pt": 50
33
+ }
34
+ },
35
+ "source_dataset": "Helsinki-NLP/opus-100",
36
+ "source_revision": "805090dc28bf78897da9641cdf08b61287580df9",
37
+ "normalization": "Unicode NFKC, casefold, punctuation removed, whitespace collapsed",
38
+ "deduplication": "Global source/target text uniqueness across all languages and splits; identical source-target pairs excluded",
39
+ "rejected": {
40
+ "duplicate_or_identical_text": 66,
41
+ "length": 1939,
42
+ "non_sentence_content": 42
43
+ },
44
+ "file_sha256": {
45
+ "train.parquet": "db2ccfff2cb499647cd134bdbf03840b9d72c98ccbadf5af25572a189cb2bcc5",
46
+ "validation.parquet": "7e65a4bc923eadaafa3cf92b8af54aeadb1aafdd2265547c52be078f6131db55",
47
+ "test.parquet": "ef098abdabd64007cb91451f25e9d452f076b897bc11ce464fc9157b7b8fffa3"
48
+ }
49
+ }