ma2za commited on
Commit
abc625e
·
1 Parent(s): 925dd0d

Publish unsplit Parquet corpus

Browse files
.gitignore ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ /build/
2
+ __pycache__/
README.md CHANGED
@@ -3,6 +3,8 @@ pretty_name: Many Emotions
3
  license: other
4
  task_categories:
5
  - text-classification
 
 
6
  multilinguality:
7
  - multilingual
8
  source_datasets:
@@ -18,8 +20,12 @@ language:
18
  size_categories:
19
  - 1M<n<10M
20
  configs:
21
- - config_name: split
22
  default: true
 
 
 
 
23
  data_files:
24
  - split: train
25
  path: data/split_dataset_train.jsonl.gz
@@ -33,6 +39,7 @@ configs:
33
  path: data/many_emotions.json.gz
34
  tags:
35
  - emotion
 
36
  - multilingual
37
  ---
38
 
@@ -40,9 +47,9 @@ tags:
40
 
41
  Many Emotions is a multilingual text-classification dataset for recognizing seven emotion categories. It combines examples from Emotion, DailyDialog, and GoEmotions and provides text in English, French, Italian, Spanish, and German.
42
 
43
- The prepared split contains 2,710,741 rows derived from 550,123 source IDs. Every row records its source dataset and source-specific license.
44
 
45
- > **Evaluation warning:** translations of the same source example are distributed across train, validation, and test. In total, 221,851 source IDs occur in more than one split. Regroup by the source ID before using this dataset for held-out evaluation.
46
 
47
  ## Labels
48
 
@@ -54,7 +61,7 @@ The prepared split contains 2,710,741 rows derived from 550,123 source IDs. Ever
54
  | 3 | love | 186,061 | 6.9% |
55
  | 4 | sadness | 615,520 | 22.7% |
56
  | 5 | surprise | 94,298 | 3.5% |
57
- | 6 | neutral | 500,392 | 18.5% |
58
 
59
  ## Load the dataset
60
 
@@ -64,25 +71,41 @@ Install the required library:
64
  pip install datasets
65
  ```
66
 
67
- Load the prepared train, validation, and test splits:
68
 
69
  ```python
70
  from datasets import load_dataset
71
 
72
- dataset = load_dataset("ma2za/many_emotions", "split")
73
  ```
74
 
75
- `split` is the default configuration, so `load_dataset("ma2za/many_emotions")` is equivalent. Load the source records and their five text columns with:
 
 
76
 
77
  ```python
78
  raw = load_dataset("ma2za/many_emotions", "raw")
79
  ```
80
 
81
- Both configurations load directly from compressed JSON Lines files and do not execute repository code.
 
 
 
 
 
 
82
 
83
  ## Dataset structure
84
 
85
- ### Prepared splits
 
 
 
 
 
 
 
 
86
 
87
  | Split | Rows |
88
  |---|---:|
@@ -91,7 +114,7 @@ Both configurations load directly from compressed JSON Lines files and do not ex
91
  | test | 135,537 |
92
  | **Total** | **2,710,741** |
93
 
94
- These are row-level splits, not source-grouped splits. Removing the final language suffix from `id` gives the source ID. Its membership across the provided splits is:
95
 
96
  | Source ID membership | Source IDs |
97
  |---|---:|
@@ -105,14 +128,14 @@ These are row-level splits, not source-grouped splits. Removing the final langua
105
 
106
  Pairwise, train shares 122,325 source IDs with validation and 122,579 with test. Validation and test share 23,131 source IDs.
107
 
108
- The prepared files contain approximately the same number of rows for each language:
109
 
110
  | Language | Code | Rows |
111
  |---|---|---:|
112
  | English | `en` | 550,122 |
113
  | German | `de` | 542,598 |
114
  | Italian | `it` | 541,145 |
115
- | Spanish | `es` | 538,700 |
116
  | French | `fr` | 538,176 |
117
 
118
  ### Data fields
@@ -120,8 +143,10 @@ The prepared files contain approximately the same number of rows for each langua
120
  | Field | Type | Description |
121
  |---|---|---|
122
  | `id` | string | Example identifier derived from the source and language. |
 
123
  | `text` | string | Text to classify. |
124
  | `label` | integer | Emotion class ID from 0 through 6. |
 
125
  | `dataset` | string | Source dataset: `emotion`, `daily_dialog`, or `go_emotions`. |
126
  | `license` | string | License or usage designation inherited from the source row. |
127
  | `language` | string | ISO 639-1 language code: `de`, `en`, `es`, `fr`, or `it`. |
@@ -131,8 +156,10 @@ Example:
131
  ```json
132
  {
133
  "id": "emotion_199458_en",
 
134
  "text": "i live with a big wonderful family that wants me to feel welcomed so sitting alone in my room is not an option",
135
  "label": 2,
 
136
  "dataset": "emotion",
137
  "license": "educational/research",
138
  "language": "en"
@@ -149,29 +176,42 @@ The prepared data retains 550,123 of the raw source IDs; 48,175 raw IDs are abse
149
 
150
  ## Source composition
151
 
152
- Counts in the prepared train, validation, and test files:
153
 
154
  | Source | Rows | Share | License value stored in rows |
155
  |---|---:|---:|---|
156
  | [Emotion](https://huggingface.co/datasets/dair-ai/emotion) | 2,055,918 | 75.8% | `educational/research` |
157
  | [DailyDialog](https://aclanthology.org/attachments/I17-1099.Datasets.zip) | 414,020 | 15.3% | `CC BY-NC-SA 4.0` |
158
- | [GoEmotions](https://huggingface.co/datasets/google-research-datasets/go_emotions) | 240,803 | 8.9% | `apache license 2.0` |
159
 
160
  ## Data quality and limitations
161
 
162
- The following results come from a complete scan of the checked-in compressed files:
163
 
164
- - All 2,710,741 prepared rows have the documented schema, unique IDs, valid language codes, and labels between 0 and 6.
165
- - One prepared row has empty text: `go_emotions_1_48976_es` in validation. The corresponding Spanish field is also empty in the raw record.
166
- - There are no repeated `(language, text, label)` triples in the prepared data.
167
  - There are 159,897 exact `(language, text)` values associated with multiple labels, involving 323,049 rows. The same text can have as many as six labels.
168
- - Exact same-language text crosses train and validation for 14,535 values, train and test for 14,660 values, and validation and test for 888 values. None of these overlaps has the same label on both sides of the measured pair.
169
- - Source-level leakage is more extensive than exact-text leakage because different-language translations of one raw example are frequently placed in different splits. Create grouped splits using the source ID if independence matters.
170
  - Labels are imbalanced and strongly associated with source. For example, 83.2% of DailyDialog rows and 64.7% of GoEmotions rows are neutral, while the Emotion contribution contains no neutral rows.
171
  - The translation method and translation-quality evaluation are not documented in this repository. Inspect the languages relevant to your use case before training or evaluation.
172
  - Median text length is 15 to 16 whitespace-delimited words, depending on language; the 95th percentile is 39 to 42 words.
173
 
174
- Source labels are normalized to anger, fear, joy, love, sadness, surprise, and neutral. The prepared labels exactly match the normalization implemented in `many_emotions.py` for every retained source ID.
 
 
 
 
 
 
 
 
 
 
 
 
 
175
 
176
  ## Licensing
177
 
 
3
  license: other
4
  task_categories:
5
  - text-classification
6
+ task_ids:
7
+ - multi-class-classification
8
  multilinguality:
9
  - multilingual
10
  source_datasets:
 
20
  size_categories:
21
  - 1M<n<10M
22
  configs:
23
+ - config_name: all
24
  default: true
25
+ data_files:
26
+ - split: full
27
+ path: data/all/*.parquet
28
+ - config_name: split
29
  data_files:
30
  - split: train
31
  path: data/split_dataset_train.jsonl.gz
 
39
  path: data/many_emotions.json.gz
40
  tags:
41
  - emotion
42
+ - emotion-classification
43
  - multilingual
44
  ---
45
 
 
47
 
48
  Many Emotions is a multilingual text-classification dataset for recognizing seven emotion categories. It combines examples from Emotion, DailyDialog, and GoEmotions and provides text in English, French, Italian, Spanish, and German.
49
 
50
+ The default unsplit corpus contains 2,710,740 non-empty rows derived from 550,123 source IDs. Every row records its source dataset and source-specific license.
51
 
52
+ > **No train/test split is implied.** The default `all` configuration intentionally exposes one `full` partition. Create splits appropriate for your task and group them by `source_id`.
53
 
54
  ## Labels
55
 
 
61
  | 3 | love | 186,061 | 6.9% |
62
  | 4 | sadness | 615,520 | 22.7% |
63
  | 5 | surprise | 94,298 | 3.5% |
64
+ | 6 | neutral | 500,391 | 18.5% |
65
 
66
  ## Load the dataset
67
 
 
71
  pip install datasets
72
  ```
73
 
74
+ Load the complete corpus as one dataset:
75
 
76
  ```python
77
  from datasets import load_dataset
78
 
79
+ dataset = load_dataset("ma2za/many_emotions", split="full")
80
  ```
81
 
82
+ `all` is the default configuration. Calling `load_dataset("ma2za/many_emotions")` without `split="full"` returns a `DatasetDict` containing that partition.
83
+
84
+ Load the source records and their five text columns with:
85
 
86
  ```python
87
  raw = load_dataset("ma2za/many_emotions", "raw")
88
  ```
89
 
90
+ The deprecated row-level splits remain available for reproducibility:
91
+
92
+ ```python
93
+ legacy = load_dataset("ma2za/many_emotions", "split")
94
+ ```
95
+
96
+ All configurations load directly from data files and do not execute repository code.
97
 
98
  ## Dataset structure
99
 
100
+ ### Default unsplit corpus
101
+
102
+ | Partition | Rows |
103
+ |---|---:|
104
+ | full | 2,710,740 |
105
+
106
+ The default corpus is stored in 11 Zstandard-compressed Parquet shards. It adds explicit `source_id` and `label_name` columns and omits the single empty record found in the legacy files.
107
+
108
+ ### Deprecated legacy splits
109
 
110
  | Split | Rows |
111
  |---|---:|
 
114
  | test | 135,537 |
115
  | **Total** | **2,710,741** |
116
 
117
+ These splits are retained only for reproducibility. They are row-level splits, not source-grouped splits. Removing the final language suffix from `id` gives the source ID. Its membership across the legacy splits is:
118
 
119
  | Source ID membership | Source IDs |
120
  |---|---:|
 
128
 
129
  Pairwise, train shares 122,325 source IDs with validation and 122,579 with test. Validation and test share 23,131 source IDs.
130
 
131
+ The default corpus contains approximately the same number of rows for each language:
132
 
133
  | Language | Code | Rows |
134
  |---|---|---:|
135
  | English | `en` | 550,122 |
136
  | German | `de` | 542,598 |
137
  | Italian | `it` | 541,145 |
138
+ | Spanish | `es` | 538,699 |
139
  | French | `fr` | 538,176 |
140
 
141
  ### Data fields
 
143
  | Field | Type | Description |
144
  |---|---|---|
145
  | `id` | string | Example identifier derived from the source and language. |
146
+ | `source_id` | string | Shared identifier for all retained language variants of a source example. Available in `all`. |
147
  | `text` | string | Text to classify. |
148
  | `label` | integer | Emotion class ID from 0 through 6. |
149
+ | `label_name` | string | Human-readable emotion label. Available in `all`. |
150
  | `dataset` | string | Source dataset: `emotion`, `daily_dialog`, or `go_emotions`. |
151
  | `license` | string | License or usage designation inherited from the source row. |
152
  | `language` | string | ISO 639-1 language code: `de`, `en`, `es`, `fr`, or `it`. |
 
156
  ```json
157
  {
158
  "id": "emotion_199458_en",
159
+ "source_id": "emotion_199458",
160
  "text": "i live with a big wonderful family that wants me to feel welcomed so sitting alone in my room is not an option",
161
  "label": 2,
162
+ "label_name": "joy",
163
  "dataset": "emotion",
164
  "license": "educational/research",
165
  "language": "en"
 
176
 
177
  ## Source composition
178
 
179
+ Counts in the default unsplit corpus:
180
 
181
  | Source | Rows | Share | License value stored in rows |
182
  |---|---:|---:|---|
183
  | [Emotion](https://huggingface.co/datasets/dair-ai/emotion) | 2,055,918 | 75.8% | `educational/research` |
184
  | [DailyDialog](https://aclanthology.org/attachments/I17-1099.Datasets.zip) | 414,020 | 15.3% | `CC BY-NC-SA 4.0` |
185
+ | [GoEmotions](https://huggingface.co/datasets/google-research-datasets/go_emotions) | 240,802 | 8.9% | `apache license 2.0` |
186
 
187
  ## Data quality and limitations
188
 
189
+ The following results come from complete build-time and post-build scans:
190
 
191
+ - All 2,710,740 default rows have the documented typed schema, unique IDs, non-empty text, valid language codes, and labels between 0 and 6.
192
+ - The empty legacy record `go_emotions_1_48976_es` is excluded from `all`. Its corresponding Spanish field is also empty in the raw record.
193
+ - There are no repeated `(language, text, label)` triples in the default corpus.
194
  - There are 159,897 exact `(language, text)` values associated with multiple labels, involving 323,049 rows. The same text can have as many as six labels.
195
+ - The deprecated legacy files place the same-language text across train and validation for 14,535 values, train and test for 14,660 values, and validation and test for 888 values. None has the same label on both sides of the measured pair.
196
+ - The deprecated legacy files also distribute translations of the same source example across splits. The default `full` partition has no split boundaries and makes no evaluation claim.
197
  - Labels are imbalanced and strongly associated with source. For example, 83.2% of DailyDialog rows and 64.7% of GoEmotions rows are neutral, while the Emotion contribution contains no neutral rows.
198
  - The translation method and translation-quality evaluation are not documented in this repository. Inspect the languages relevant to your use case before training or evaluation.
199
  - Median text length is 15 to 16 whitespace-delimited words, depending on language; the 95th percentile is 39 to 42 words.
200
 
201
+ Source labels are normalized to anger, fear, joy, love, sadness, surprise, and neutral. The default corpus preserves the normalized labels from the legacy prepared files.
202
+
203
+ ## Reproducibility
204
+
205
+ `reports/all_validation.json` records input and shard checksums, row counts, feature definitions, language counts, label counts, source counts, and license counts.
206
+
207
+ Rebuild into a separate directory:
208
+
209
+ ```bash
210
+ python scripts/build_all.py --output-dir build/all --report build/all_validation.json
211
+ python scripts/validate_all.py --report build/all_validation.json
212
+ ```
213
+
214
+ The builder concatenates the three legacy files into one partition, derives `source_id`, adds `label_name`, excludes empty text, and writes deterministic 250,000-row Parquet shards. It does not assign new train, validation, or test splits.
215
 
216
  ## Licensing
217
 
data/all/part-00000.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4af344ca136b03b5524aba5eb42bb4ab34d8f5918ca329b59f55f08f24129749
3
+ size 13348676
data/all/part-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6428bd49dcb8087cba478000860f24a5e9910b609ea56e06d11837fba40f719f
3
+ size 13351647
data/all/part-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d7dfc4a8239cac1aff98a9b2b1acfcdd7cf92829aa968e7ef801e713005aa14
3
+ size 13360167
data/all/part-00003.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71c7aa1340780023f57bccfbd37ecb2054183c9477a318e7f1c34bb9c74589e8
3
+ size 13337982
data/all/part-00004.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12a0457a7af6210ab84fe1812b4df598a041f6c82b43ec29f2fe312d75c16c5f
3
+ size 13333900
data/all/part-00005.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cfd0c00708b7f763247c271214448fa4da22bcf4c795ce88ffe8fcbfd75e2d33
3
+ size 13325351
data/all/part-00006.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2ec738b8486708f14f7325492d3cb3afed3c539177c9d62352072c0b89b28271
3
+ size 13360140
data/all/part-00007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5366a33b945df50d0fa855d435d6b2efd6dcf01a321b22962c86a948455c9f94
3
+ size 13334655
data/all/part-00008.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d22610d780922a42b910354bf0b79ffdef6748920ba2e7ee93cf5ea07bc26b8a
3
+ size 13356232
data/all/part-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:837d1651bcb9abdc4769089206fdd745cabda60462a7b35b1d20c4b46ba9c59a
3
+ size 13351262
data/all/part-00010.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:09e0bfe169f287eab6ad8620a0c196816cee13f28361293c4cd427108c0670df
3
+ size 11262857
reports/all_validation.json ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": "all",
3
+ "counts": {
4
+ "dropped_empty_text": 1,
5
+ "input_rows": 2710741,
6
+ "input_unique_ids": 2710741,
7
+ "labels": {
8
+ "anger": 306529,
9
+ "fear": 242462,
10
+ "joy": 765479,
11
+ "love": 186061,
12
+ "neutral": 500391,
13
+ "sadness": 615520,
14
+ "surprise": 94298
15
+ },
16
+ "languages": {
17
+ "de": 542598,
18
+ "en": 550122,
19
+ "es": 538699,
20
+ "fr": 538176,
21
+ "it": 541145
22
+ },
23
+ "licenses": {
24
+ "CC BY-NC-SA 4.0": 414020,
25
+ "apache license 2.0": 240802,
26
+ "educational/research": 2055918
27
+ },
28
+ "output_rows": 2710740,
29
+ "output_unique_ids": 2710740,
30
+ "retained_source_ids": 550123,
31
+ "sources": {
32
+ "daily_dialog": 414020,
33
+ "emotion": 2055918,
34
+ "go_emotions": 240802
35
+ }
36
+ },
37
+ "features": {
38
+ "dataset": {
39
+ "_type": "Value",
40
+ "dtype": "string"
41
+ },
42
+ "id": {
43
+ "_type": "Value",
44
+ "dtype": "string"
45
+ },
46
+ "label": {
47
+ "_type": "ClassLabel",
48
+ "names": [
49
+ "anger",
50
+ "fear",
51
+ "joy",
52
+ "love",
53
+ "sadness",
54
+ "surprise",
55
+ "neutral"
56
+ ]
57
+ },
58
+ "label_name": {
59
+ "_type": "Value",
60
+ "dtype": "string"
61
+ },
62
+ "language": {
63
+ "_type": "Value",
64
+ "dtype": "string"
65
+ },
66
+ "license": {
67
+ "_type": "Value",
68
+ "dtype": "string"
69
+ },
70
+ "source_id": {
71
+ "_type": "Value",
72
+ "dtype": "string"
73
+ },
74
+ "text": {
75
+ "_type": "Value",
76
+ "dtype": "string"
77
+ }
78
+ },
79
+ "format_version": 1,
80
+ "input_files": [
81
+ {
82
+ "bytes": 143575307,
83
+ "path": "data/split_dataset_train.jsonl.gz",
84
+ "rows": 2439666,
85
+ "sha256": "b420aba48b9cd72b9363a9794dad3346cd572ae7418b77962b14fd44b725eeb2"
86
+ },
87
+ {
88
+ "bytes": 7969332,
89
+ "path": "data/split_dataset_validation.jsonl.gz",
90
+ "rows": 135538,
91
+ "sha256": "24c46037cc2be059040caf83bfd8fb76110111517857602acdc1caa1d41848c5"
92
+ },
93
+ {
94
+ "bytes": 7981677,
95
+ "path": "data/split_dataset_test.jsonl.gz",
96
+ "rows": 135537,
97
+ "sha256": "a26f085aa3157ffe3f8817d94d64040050072341d6395a12e7b6b31ad5b338f7"
98
+ }
99
+ ],
100
+ "partition": "full",
101
+ "shards": [
102
+ {
103
+ "bytes": 13348676,
104
+ "path": "data/all/part-00000.parquet",
105
+ "rows": 250000,
106
+ "sha256": "4af344ca136b03b5524aba5eb42bb4ab34d8f5918ca329b59f55f08f24129749"
107
+ },
108
+ {
109
+ "bytes": 13351647,
110
+ "path": "data/all/part-00001.parquet",
111
+ "rows": 250000,
112
+ "sha256": "6428bd49dcb8087cba478000860f24a5e9910b609ea56e06d11837fba40f719f"
113
+ },
114
+ {
115
+ "bytes": 13360167,
116
+ "path": "data/all/part-00002.parquet",
117
+ "rows": 250000,
118
+ "sha256": "9d7dfc4a8239cac1aff98a9b2b1acfcdd7cf92829aa968e7ef801e713005aa14"
119
+ },
120
+ {
121
+ "bytes": 13337982,
122
+ "path": "data/all/part-00003.parquet",
123
+ "rows": 250000,
124
+ "sha256": "71c7aa1340780023f57bccfbd37ecb2054183c9477a318e7f1c34bb9c74589e8"
125
+ },
126
+ {
127
+ "bytes": 13333900,
128
+ "path": "data/all/part-00004.parquet",
129
+ "rows": 250000,
130
+ "sha256": "12a0457a7af6210ab84fe1812b4df598a041f6c82b43ec29f2fe312d75c16c5f"
131
+ },
132
+ {
133
+ "bytes": 13325351,
134
+ "path": "data/all/part-00005.parquet",
135
+ "rows": 250000,
136
+ "sha256": "cfd0c00708b7f763247c271214448fa4da22bcf4c795ce88ffe8fcbfd75e2d33"
137
+ },
138
+ {
139
+ "bytes": 13360140,
140
+ "path": "data/all/part-00006.parquet",
141
+ "rows": 250000,
142
+ "sha256": "2ec738b8486708f14f7325492d3cb3afed3c539177c9d62352072c0b89b28271"
143
+ },
144
+ {
145
+ "bytes": 13334655,
146
+ "path": "data/all/part-00007.parquet",
147
+ "rows": 250000,
148
+ "sha256": "5366a33b945df50d0fa855d435d6b2efd6dcf01a321b22962c86a948455c9f94"
149
+ },
150
+ {
151
+ "bytes": 13356232,
152
+ "path": "data/all/part-00008.parquet",
153
+ "rows": 250000,
154
+ "sha256": "d22610d780922a42b910354bf0b79ffdef6748920ba2e7ee93cf5ea07bc26b8a"
155
+ },
156
+ {
157
+ "bytes": 13351262,
158
+ "path": "data/all/part-00009.parquet",
159
+ "rows": 250000,
160
+ "sha256": "837d1651bcb9abdc4769089206fdd745cabda60462a7b35b1d20c4b46ba9c59a"
161
+ },
162
+ {
163
+ "bytes": 11262857,
164
+ "path": "data/all/part-00010.parquet",
165
+ "rows": 210740,
166
+ "sha256": "09e0bfe169f287eab6ad8620a0c196816cee13f28361293c4cd427108c0670df"
167
+ }
168
+ ]
169
+ }
requirements.txt CHANGED
@@ -1 +1,2 @@
1
- datasets
 
 
1
+ datasets==5.0.0
2
+ pyarrow==25.0.0
scripts/build_all.py ADDED
@@ -0,0 +1,206 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import collections
3
+ import gzip
4
+ import hashlib
5
+ import json
6
+ from pathlib import Path
7
+
8
+ import pyarrow as pa
9
+ import pyarrow.parquet as pq
10
+ from datasets import ClassLabel, Features, Value
11
+
12
+
13
+ LABELS = ["anger", "fear", "joy", "love", "sadness", "surprise", "neutral"]
14
+ LANGUAGES = {"de", "en", "es", "fr", "it"}
15
+ INPUTS = [
16
+ "split_dataset_train.jsonl.gz",
17
+ "split_dataset_validation.jsonl.gz",
18
+ "split_dataset_test.jsonl.gz",
19
+ ]
20
+ FEATURES = Features(
21
+ {
22
+ "id": Value("string"),
23
+ "source_id": Value("string"),
24
+ "text": Value("string"),
25
+ "label": ClassLabel(names=LABELS),
26
+ "label_name": Value("string"),
27
+ "dataset": Value("string"),
28
+ "license": Value("string"),
29
+ "language": Value("string"),
30
+ }
31
+ )
32
+
33
+
34
+ def file_sha256(path):
35
+ digest = hashlib.sha256()
36
+ with path.open("rb") as stream:
37
+ for chunk in iter(lambda: stream.read(1024 * 1024), b""):
38
+ digest.update(chunk)
39
+ return digest.hexdigest()
40
+
41
+
42
+ def parse_args():
43
+ parser = argparse.ArgumentParser()
44
+ parser.add_argument("--data-dir", type=Path, default=Path("data"))
45
+ parser.add_argument("--output-dir", type=Path, default=Path("data/all"))
46
+ parser.add_argument("--report", type=Path, default=Path("reports/all_validation.json"))
47
+ parser.add_argument("--rows-per-shard", type=int, default=250_000)
48
+ parser.add_argument("--batch-size", type=int, default=25_000)
49
+ return parser.parse_args()
50
+
51
+
52
+ def main():
53
+ args = parse_args()
54
+ if args.output_dir.exists() and any(args.output_dir.iterdir()):
55
+ raise SystemExit(f"output directory is not empty: {args.output_dir}")
56
+
57
+ args.output_dir.mkdir(parents=True, exist_ok=True)
58
+ args.report.parent.mkdir(parents=True, exist_ok=True)
59
+
60
+ schema = FEATURES.arrow_schema
61
+ required = {"id", "text", "label", "dataset", "license", "language"}
62
+ ids = set()
63
+ source_ids = set()
64
+ counts = collections.Counter()
65
+ label_counts = collections.Counter()
66
+ language_counts = collections.Counter()
67
+ source_counts = collections.Counter()
68
+ license_counts = collections.Counter()
69
+ input_files = []
70
+ shards = []
71
+ writer = None
72
+ shard_index = 0
73
+ rows_in_shard = 0
74
+ batch = []
75
+
76
+ def open_writer():
77
+ nonlocal writer, shard_index, rows_in_shard
78
+ path = args.output_dir / f"part-{shard_index:05d}.parquet"
79
+ writer = pq.ParquetWriter(
80
+ path,
81
+ schema,
82
+ compression="zstd",
83
+ compression_level=9,
84
+ use_dictionary=True,
85
+ write_statistics=True,
86
+ write_page_index=True,
87
+ )
88
+ shards.append(path)
89
+ shard_index += 1
90
+ rows_in_shard = 0
91
+
92
+ def write_rows(rows):
93
+ nonlocal writer, rows_in_shard
94
+ offset = 0
95
+ while offset < len(rows):
96
+ if writer is None:
97
+ open_writer()
98
+ available = args.rows_per_shard - rows_in_shard
99
+ chunk = rows[offset : offset + available]
100
+ writer.write_table(pa.Table.from_pylist(chunk, schema=schema), row_group_size=65_536)
101
+ rows_in_shard += len(chunk)
102
+ offset += len(chunk)
103
+ if rows_in_shard == args.rows_per_shard:
104
+ writer.close()
105
+ writer = None
106
+
107
+ for filename in INPUTS:
108
+ path = args.data_dir / filename
109
+ file_rows = 0
110
+ input_files.append(
111
+ {
112
+ "path": path.as_posix(),
113
+ "bytes": path.stat().st_size,
114
+ "sha256": file_sha256(path),
115
+ }
116
+ )
117
+ with gzip.open(path, "rt", encoding="utf-8") as stream:
118
+ for line_number, line in enumerate(stream, 1):
119
+ counts["input_rows"] += 1
120
+ file_rows += 1
121
+ row = json.loads(line)
122
+ missing = required - row.keys()
123
+ if missing:
124
+ raise ValueError(f"{path}:{line_number}: missing fields {sorted(missing)}")
125
+ if row["id"] in ids:
126
+ raise ValueError(f"{path}:{line_number}: duplicate id {row['id']}")
127
+ ids.add(row["id"])
128
+ if row["language"] not in LANGUAGES:
129
+ raise ValueError(f"{path}:{line_number}: invalid language {row['language']}")
130
+ if not isinstance(row["label"], int) or not 0 <= row["label"] < len(LABELS):
131
+ raise ValueError(f"{path}:{line_number}: invalid label {row['label']}")
132
+ if not isinstance(row["text"], str) or not row["text"].strip():
133
+ counts["dropped_empty_text"] += 1
134
+ continue
135
+
136
+ suffix = "_" + row["language"]
137
+ if not row["id"].endswith(suffix):
138
+ raise ValueError(f"{path}:{line_number}: id/language mismatch {row['id']}")
139
+ source_id = row["id"][: -len(suffix)]
140
+ source_ids.add(source_id)
141
+ output = {
142
+ "id": row["id"],
143
+ "source_id": source_id,
144
+ "text": row["text"],
145
+ "label": row["label"],
146
+ "label_name": LABELS[row["label"]],
147
+ "dataset": row["dataset"],
148
+ "license": row["license"],
149
+ "language": row["language"],
150
+ }
151
+ batch.append(output)
152
+ counts["output_rows"] += 1
153
+ label_counts[row["label"]] += 1
154
+ language_counts[row["language"]] += 1
155
+ source_counts[row["dataset"]] += 1
156
+ license_counts[row["license"]] += 1
157
+ if len(batch) == args.batch_size:
158
+ write_rows(batch)
159
+ batch.clear()
160
+ input_files[-1]["rows"] = file_rows
161
+
162
+ if batch:
163
+ write_rows(batch)
164
+ if writer is not None:
165
+ writer.close()
166
+
167
+ shard_files = []
168
+ shard_rows_total = 0
169
+ for path in shards:
170
+ rows = pq.ParquetFile(path).metadata.num_rows
171
+ shard_rows_total += rows
172
+ shard_files.append(
173
+ {
174
+ "path": path.as_posix(),
175
+ "rows": rows,
176
+ "bytes": path.stat().st_size,
177
+ "sha256": file_sha256(path),
178
+ }
179
+ )
180
+ if shard_rows_total != counts["output_rows"]:
181
+ raise ValueError(f"shard row mismatch: {shard_rows_total} != {counts['output_rows']}")
182
+
183
+ report = {
184
+ "format_version": 1,
185
+ "config": "all",
186
+ "partition": "full",
187
+ "features": FEATURES.to_dict(),
188
+ "input_files": input_files,
189
+ "shards": shard_files,
190
+ "counts": {
191
+ **dict(sorted(counts.items())),
192
+ "input_unique_ids": len(ids),
193
+ "output_unique_ids": counts["output_rows"],
194
+ "retained_source_ids": len(source_ids),
195
+ "labels": {LABELS[key]: value for key, value in sorted(label_counts.items())},
196
+ "languages": dict(sorted(language_counts.items())),
197
+ "sources": dict(sorted(source_counts.items())),
198
+ "licenses": dict(sorted(license_counts.items())),
199
+ },
200
+ }
201
+ args.report.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8")
202
+ print(json.dumps(report["counts"], indent=2, sort_keys=True))
203
+
204
+
205
+ if __name__ == "__main__":
206
+ main()
scripts/validate_all.py ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import hashlib
3
+ import json
4
+ from pathlib import Path
5
+
6
+ import pyarrow.parquet as pq
7
+ from datasets import ClassLabel, Features, Value
8
+
9
+
10
+ LABELS = ["anger", "fear", "joy", "love", "sadness", "surprise", "neutral"]
11
+ LANGUAGES = {"de", "en", "es", "fr", "it"}
12
+ FEATURES = Features(
13
+ {
14
+ "id": Value("string"),
15
+ "source_id": Value("string"),
16
+ "text": Value("string"),
17
+ "label": ClassLabel(names=LABELS),
18
+ "label_name": Value("string"),
19
+ "dataset": Value("string"),
20
+ "license": Value("string"),
21
+ "language": Value("string"),
22
+ }
23
+ )
24
+
25
+
26
+ def file_sha256(path):
27
+ digest = hashlib.sha256()
28
+ with path.open("rb") as stream:
29
+ for chunk in iter(lambda: stream.read(1024 * 1024), b""):
30
+ digest.update(chunk)
31
+ return digest.hexdigest()
32
+
33
+
34
+ def parse_args():
35
+ parser = argparse.ArgumentParser()
36
+ parser.add_argument("--report", type=Path, default=Path("reports/all_validation.json"))
37
+ return parser.parse_args()
38
+
39
+
40
+ def main():
41
+ args = parse_args()
42
+ report = json.loads(args.report.read_text(encoding="utf-8"))
43
+ ids = set()
44
+ rows = 0
45
+ source_ids = set()
46
+
47
+ for shard in report["shards"]:
48
+ path = Path(shard["path"])
49
+ if file_sha256(path) != shard["sha256"]:
50
+ raise ValueError(f"checksum mismatch: {path}")
51
+ parquet = pq.ParquetFile(path)
52
+ if parquet.schema_arrow != FEATURES.arrow_schema:
53
+ raise ValueError(f"schema mismatch: {path}")
54
+ if parquet.metadata.num_rows != shard["rows"]:
55
+ raise ValueError(f"row count mismatch: {path}")
56
+ for batch in parquet.iter_batches(batch_size=65_536):
57
+ columns = batch.to_pydict()
58
+ for index, row_id in enumerate(columns["id"]):
59
+ language = columns["language"][index]
60
+ source_id = columns["source_id"][index]
61
+ label = columns["label"][index]
62
+ if row_id in ids:
63
+ raise ValueError(f"duplicate id: {row_id}")
64
+ ids.add(row_id)
65
+ source_ids.add(source_id)
66
+ if language not in LANGUAGES:
67
+ raise ValueError(f"invalid language: {language}")
68
+ if row_id != f"{source_id}_{language}":
69
+ raise ValueError(f"id/source mismatch: {row_id}")
70
+ if not columns["text"][index].strip():
71
+ raise ValueError(f"empty text: {row_id}")
72
+ if not 0 <= label < len(LABELS):
73
+ raise ValueError(f"invalid label: {row_id}")
74
+ if columns["label_name"][index] != LABELS[label]:
75
+ raise ValueError(f"label name mismatch: {row_id}")
76
+ rows += 1
77
+
78
+ expected = report["counts"]
79
+ if rows != expected["output_rows"]:
80
+ raise ValueError(f"total row mismatch: {rows} != {expected['output_rows']}")
81
+ if len(ids) != rows:
82
+ raise ValueError(f"unique id mismatch: {len(ids)} != {rows}")
83
+ if len(ids) != expected["output_unique_ids"]:
84
+ raise ValueError(f"reported unique id mismatch: {len(ids)} != {expected['output_unique_ids']}")
85
+ if len(source_ids) != expected["retained_source_ids"]:
86
+ raise ValueError(
87
+ f"source id mismatch: {len(source_ids)} != {expected['retained_source_ids']}"
88
+ )
89
+ print(
90
+ json.dumps(
91
+ {
92
+ "status": "valid",
93
+ "rows": rows,
94
+ "unique_ids": len(ids),
95
+ "source_ids": len(source_ids),
96
+ "shards": len(report["shards"]),
97
+ },
98
+ indent=2,
99
+ )
100
+ )
101
+
102
+
103
+ if __name__ == "__main__":
104
+ main()