villacu commited on
Commit
7859146
·
verified ·
1 Parent(s): 0ce0c1c

Add full English and Turkish text with three-way labels

Browse files
README.md CHANGED
@@ -31,27 +31,16 @@ configs:
31
 
32
  HateSpeechMix is a research-oriented normalization of four multilingual hate/offensive-speech resources. It contains 187,315 rows across Amharic, Chinese, English, Hindi, and Turkish with deterministic `train` and `test` splits.
33
 
34
- ## Critical limitation
35
-
36
- The English/Turkish source distributes tweet IDs and annotations but not tweet text, in accordance with Twitter/X terms. Its 128,907 rows therefore have `text = null` and `text_available = false`. Do not pass the full dataset directly to a tokenizer without filtering:
37
-
38
- ```python
39
- from datasets import load_dataset
40
-
41
- dataset = load_dataset("villacu/hatespeechmix")
42
- text_dataset = dataset.filter(lambda row: row["text_available"])
43
- ```
44
-
45
- The remaining 58,408 rows contain text. No tweet hydration or text reconstruction was attempted.
46
 
47
  ## Schema
48
 
49
  | Field | Meaning |
50
  |---|---|
51
- | `text` | Source text, or null when the source publishes metadata only |
52
- | `label` | Binary `ClassLabel`: 0 = `normal`, 1 = `abusive` |
53
  | `label_text` | Human-readable form of `label` |
54
- | `fine_label` | Best available normalized subtype: `normal`, `offensive`, `hate`, `profane`, or `abusive` |
55
  | `language` | ISO 639-1 language code |
56
  | `source` | Stable source-corpus identifier |
57
  | `source_id` | Identifier within the source corpus |
@@ -61,7 +50,16 @@ The remaining 58,408 rows contain text. No tweet hydration or text reconstructio
61
  | `target` | Source target type when available |
62
  | `text_available` | Whether `text` contains usable text |
63
 
64
- The primary label is deliberately called **abusive**, not **hate**. The Chinese train/dev and Hindi task-1 annotations conflate multiple types of harmful language, so a harmonized multiclass hate label would claim distinctions that those sources do not provide. Fine-grained labels remain available but are not uniformly annotated across sources or splits.
 
 
 
 
 
 
 
 
 
65
 
66
  ## Split policy
67
 
@@ -70,7 +68,7 @@ The primary label is deliberately called **abusive**, not **hate**. The Chinese
70
  - Amharic RANLP 2023: preserve test; merge upstream dev into train.
71
  - English/Turkish v2: create a deterministic 90/10 split, stratified by language, topic, and original three-way label. The source used 10-fold cross-validation and does not publish a fixed split.
72
 
73
- After assigning splits, the build removes training rows whose exact `(source, text)` pair also occurs in test. Test rows are retained. This removes 56 leaked training rows (54 Chinese, one Hindi, and one Amharic), including eight whose binary label conflicts with the test copy. Within-split duplicates are retained and reported rather than silently resolved because some have conflicting annotations.
74
 
75
  Rows are kept in source order and are not globally shuffled. Shuffle explicitly during training.
76
 
@@ -80,13 +78,15 @@ Rows are kept in source order and are not globally shuffled. Shuffle explicitly
80
  |---|---:|---:|---:|
81
  | Amharic (`am`) | 15,000 | 15,001 | 15,000 |
82
  | Chinese (`zh`) | 37,426 | 37,480 | 37,426 |
83
- | English (`en`) | 68,597 | 68,597 | 0 |
84
  | Hindi (`hi`) | 5,982 | 5,983 | 5,982 |
85
- | Turkish (`tr`) | 60,310 | 60,310 | 0 |
86
- | **Total** | **187,315** | **187,371** | **58,408** |
87
 
88
  Exact split and label counts are recorded in [`metadata/build_summary.json`](metadata/build_summary.json). Source revisions and checksums are recorded in [`metadata/source_manifest.json`](metadata/source_manifest.json).
89
 
 
 
90
  ## Licensing and redistribution
91
 
92
  This compilation does **not** have a single permissive license:
 
31
 
32
  HateSpeechMix is a research-oriented normalization of four multilingual hate/offensive-speech resources. It contains 187,315 rows across Amharic, Chinese, English, Hindi, and Turkish with deterministic `train` and `test` splits.
33
 
34
+ All published rows contain text. English and Turkish use the authors' official 2024 full-text v2 release rather than the older ID-only export.
 
 
 
 
 
 
 
 
 
 
 
35
 
36
  ## Schema
37
 
38
  | Field | Meaning |
39
  |---|---|
40
+ | `text` | Source text |
41
+ | `label` | Three-way `ClassLabel`: 0 = `neutral`, 1 = `offensive`, 2 = `hate` |
42
  | `label_text` | Human-readable form of `label` |
43
+ | `fine_label` | Best available normalized subtype; preserves Hindi `profane` separately |
44
  | `language` | ISO 639-1 language code |
45
  | `source` | Stable source-corpus identifier |
46
  | `source_id` | Identifier within the source corpus |
 
50
  | `target` | Source target type when available |
51
  | `text_available` | Whether `text` contains usable text |
52
 
53
+ ## Label normalization
54
+
55
+ | Source | Unified mapping |
56
+ |---|---|
57
+ | English/Turkish | Normal → `neutral`; Offensive → `offensive`; Hate → `hate` |
58
+ | Amharic | Normal → `neutral`; Offensive → `offensive`; Hate → `hate` |
59
+ | Hindi | NONE → `neutral`; OFFN or PRFN → `offensive`; HATE → `hate` |
60
+ | Chinese | Safe → `neutral`; every offensive/attack category → `offensive` |
61
+
62
+ Chinese has no unified `hate` examples. Its test-only individual/group distinction remains in `source_label` and `target`; mapping group attacks to hate would add an interpretation absent from the source's main binary annotation. Hindi `PRFN` remains visible as `fine_label = "profane"` and in `source_label` even though its unified label is `offensive`.
63
 
64
  ## Split policy
65
 
 
68
  - Amharic RANLP 2023: preserve test; merge upstream dev into train.
69
  - English/Turkish v2: create a deterministic 90/10 split, stratified by language, topic, and original three-way label. The source used 10-fold cross-validation and does not publish a fixed split.
70
 
71
+ After assigning splits, the build removes training rows whose exact `(source, text)` pair also occurs in test. Test rows are retained. This removes 56 leaked training rows (54 Chinese, one Hindi, and one Amharic), including nine whose unified label conflicts with the test copy. Within-split duplicates are retained and reported rather than silently resolved because some have conflicting annotations.
72
 
73
  Rows are kept in source order and are not globally shuffled. Shuffle explicitly during training.
74
 
 
78
  |---|---:|---:|---:|
79
  | Amharic (`am`) | 15,000 | 15,001 | 15,000 |
80
  | Chinese (`zh`) | 37,426 | 37,480 | 37,426 |
81
+ | English (`en`) | 68,597 | 68,597 | 68,597 |
82
  | Hindi (`hi`) | 5,982 | 5,983 | 5,982 |
83
+ | Turkish (`tr`) | 60,310 | 60,310 | 60,310 |
84
+ | **Total** | **187,315** | **187,371** | **187,315** |
85
 
86
  Exact split and label counts are recorded in [`metadata/build_summary.json`](metadata/build_summary.json). Source revisions and checksums are recorded in [`metadata/source_manifest.json`](metadata/source_manifest.json).
87
 
88
+ The official English/Turkish TSV contains bare carriage returns inside some tweet texts. The builder normalizes these formatting characters before parsing, validates required fields and categorical values, and reports any dropped records under `source_quality` in the build summary. The pinned release retains all 128,907 English/Turkish rows.
89
+
90
  ## Licensing and redistribution
91
 
92
  This compilation does **not** have a single permissive license:
data/test-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9b8ca26ee47e49bf848de4fbf3245e2d991966b96fe9320a4f936058dc7ef5b9
3
- size 1304572
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:378f35e5094853d3276bc2cac013f5cc5fd5a578b44feab3754e836b03d6cdb8
3
+ size 3010553
data/train-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ede1a97f70ba86789865281d1d5d9bc2679a95687e04bbe8e612269d75f25d01
3
- size 7715691
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f3e43cb95833ff04930fb8c53f02c921938b395cd006f3dd7673c43035803971
3
+ size 22749897
metadata/build_summary.json CHANGED
@@ -1,9 +1,9 @@
1
  {
2
- "build_version": "1.0.0",
3
  "split_seed": "hatespeechmix-v1",
4
  "num_rows": 187315,
5
- "num_text_rows": 58408,
6
- "num_metadata_only_rows": 128907,
7
  "split_counts": {
8
  "train": 166282,
9
  "test": 21033
@@ -32,8 +32,9 @@
32
  }
33
  },
34
  "label_counts": {
35
- "abusive": 72720,
36
- "normal": 114595
 
37
  },
38
  "source_counts": {
39
  "amharic_ranlp2023": 15000,
@@ -44,7 +45,7 @@
44
  "cross_split_text_leakage": {
45
  "policy": "remove train rows with exact source/text match in test",
46
  "removed_train_rows": 56,
47
- "removed_rows_with_binary_label_conflict": 8,
48
  "removed_by_source": {
49
  "amharic_ranlp2023": 1,
50
  "cold_zh": 54,
@@ -59,5 +60,14 @@
59
  "within_split_duplicate_text_counts": {
60
  "train": 102,
61
  "test": 3
 
 
 
 
 
 
 
 
 
62
  }
63
  }
 
1
  {
2
+ "build_version": "2.0.0",
3
  "split_seed": "hatespeechmix-v1",
4
  "num_rows": 187315,
5
+ "num_text_rows": 187315,
6
+ "num_metadata_only_rows": 0,
7
  "split_counts": {
8
  "train": 166282,
9
  "test": 21033
 
32
  }
33
  },
34
  "label_counts": {
35
+ "hate": 21847,
36
+ "neutral": 114595,
37
+ "offensive": 50873
38
  },
39
  "source_counts": {
40
  "amharic_ranlp2023": 15000,
 
45
  "cross_split_text_leakage": {
46
  "policy": "remove train rows with exact source/text match in test",
47
  "removed_train_rows": 56,
48
+ "removed_rows_with_label_conflict": 9,
49
  "removed_by_source": {
50
  "amharic_ranlp2023": 1,
51
  "cold_zh": 54,
 
60
  "within_split_duplicate_text_counts": {
61
  "train": 102,
62
  "test": 3
63
+ },
64
+ "source_quality": {
65
+ "toraman2022_en_tr_v2": {
66
+ "source_rows": 128907,
67
+ "retained_rows": 128907,
68
+ "dropped_rows": 0,
69
+ "dropped_reasons": {},
70
+ "normalized_bare_carriage_returns": 22
71
+ }
72
  }
73
  }
metadata/source_manifest.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "build_version": "1.0.0",
3
  "sources": [
4
  {
5
  "name": "HASOC 2019 Hindi",
@@ -11,11 +11,11 @@
11
  },
12
  {
13
  "name": "Large-Scale Hate Speech Dataset v2",
14
- "source": "https://github.com/avaapm/hatespeech",
15
- "revision": "15e72305dd6cec705009624e21933fd0176c15e8",
16
- "sha256": "6829aa3bdff8f739a941efd6968fb2824641356747f4c3283ae3318a7ccf4224",
17
  "license": "CC BY-NC-SA 4.0",
18
- "redistribution_note": "The source contains tweet IDs and annotations, not tweet text"
19
  },
20
  {
21
  "name": "COLDataset",
@@ -37,4 +37,3 @@
37
  }
38
  ]
39
  }
40
-
 
1
  {
2
+ "build_version": "2.0.0",
3
  "sources": [
4
  {
5
  "name": "HASOC 2019 Hindi",
 
11
  },
12
  {
13
  "name": "Large-Scale Hate Speech Dataset v2",
14
+ "source": "https://github.com/metunlp/hate-speech/blob/f27c847b34e68c40ac654219b5824f8750b284d8/dataset_v2/Toraman22_hate_speech_v2.tsv",
15
+ "revision": "f27c847b34e68c40ac654219b5824f8750b284d8",
16
+ "sha256": "c6c110de8e9656577b5781c8173d7974268c8e716adf5d804f98a50b2f4c0a95",
17
  "license": "CC BY-NC-SA 4.0",
18
+ "redistribution_note": "Official 2024 full-text release; only tweet ID, text, language, topic, and final label are retained"
19
  },
20
  {
21
  "name": "COLDataset",
 
37
  }
38
  ]
39
  }
 
scripts/build_dataset.py CHANGED
@@ -3,11 +3,10 @@
3
 
4
  The conversion is intentionally conservative:
5
 
6
- * ``label`` is binary because not every source distinguishes hate from other
7
- offensive content.
8
  * ``fine_label`` and ``source_label`` retain the available source semantics.
9
- * The English/Turkish source publishes tweet IDs, not tweet text. Those rows
10
- are retained with ``text=None`` and ``text_available=False``.
11
  * Official test splits are preserved. Upstream dev splits are folded into
12
  train. The unsplit English/Turkish data receives a deterministic,
13
  stratified 90/10 split.
@@ -30,7 +29,7 @@ from typing import Iterable, Iterator
30
  from datasets import ClassLabel, Dataset, DatasetDict, Features, Value
31
 
32
 
33
- BUILD_VERSION = "1.0.0"
34
  SPLIT_SEED = "hatespeechmix-v1"
35
 
36
  REPO_ROOT = Path(__file__).resolve().parents[1]
@@ -39,12 +38,13 @@ WORKSPACE_ROOT = REPO_ROOT.parent
39
  SOURCE_SPECS = {
40
  "toraman_en_tr": {
41
  "url": (
42
- "https://github.com/avaapm/hatespeech/archive/"
43
- "15e72305dd6cec705009624e21933fd0176c15e8.zip"
 
44
  ),
45
- "sha256": "6829aa3bdff8f739a941efd6968fb2824641356747f4c3283ae3318a7ccf4224",
46
- "filename": "toraman_en_tr.zip",
47
- "revision": "15e72305dd6cec705009624e21933fd0176c15e8",
48
  },
49
  "cold_zh": {
50
  "url": (
@@ -85,11 +85,12 @@ SOURCE_SPECS = {
85
  }
86
 
87
  HINDI_SHA256 = "d419780fe825f9946e3a03da4cf3fdf41699b188932d3662ca304693829994ad"
 
88
 
89
  FEATURES = Features(
90
  {
91
  "text": Value("string"),
92
- "label": ClassLabel(names=["normal", "abusive"]),
93
  "label_text": Value("string"),
94
  "fine_label": Value("string"),
95
  "language": Value("string"),
@@ -187,11 +188,11 @@ def record(
187
  topic: str | None = None,
188
  target: str | None = None,
189
  ) -> dict[str, object]:
190
- if label_text not in {"normal", "abusive"}:
191
- raise ValueError(f"Invalid binary label: {label_text}")
192
  return {
193
  "text": text,
194
- "label": 0 if label_text == "normal" else 1,
195
  "label_text": label_text,
196
  "fine_label": fine_label,
197
  "language": language,
@@ -207,7 +208,8 @@ def record(
207
 
208
  def load_hindi(hindi_zip: Path) -> dict[str, list[dict[str, object]]]:
209
  verify_file(hindi_zip, HINDI_SHA256)
210
- fine_labels = {"NONE": "normal", "OFFN": "offensive", "HATE": "hate", "PRFN": "profane"}
 
211
  targets = {"NONE": None, "TIN": "targeted", "UNT": "untargeted"}
212
  members = {
213
  "train": "hindi_dataset/hindi_dataset.tsv",
@@ -217,7 +219,7 @@ def load_hindi(hindi_zip: Path) -> dict[str, list[dict[str, object]]]:
217
  for split, member in members.items():
218
  for row in dict_rows_from_zip(hindi_zip, member, delimiter="\t"):
219
  task_1 = row["task_1"]
220
- label_text = "normal" if task_1 == "NOT" else "abusive"
221
  splits[split].append(
222
  record(
223
  text=row["text"],
@@ -270,38 +272,86 @@ def deterministic_stratified_split(
270
  return splits
271
 
272
 
273
- def load_english_turkish(archive_path: Path) -> dict[str, list[dict[str, object]]]:
 
 
 
 
 
 
 
 
 
 
 
 
 
274
  lang = {"0": "tr", "1": "en"}
275
  topics = {"0": "religion", "1": "gender", "2": "race", "3": "politics", "4": "sports"}
276
- fine_labels = {"0": "normal", "1": "offensive", "2": "hate"}
277
  rows: list[dict[str, object]] = []
278
- for row in dict_rows_from_zip(
279
- archive_path, "dataset_v2/hate_speech_dataset_v2.csv"
280
- ):
281
- fine_label = fine_labels[row["HateLabel"]]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
282
  rows.append(
283
  record(
284
- text=None,
285
- label_text="normal" if fine_label == "normal" else "abusive",
286
- fine_label=fine_label,
287
- language=lang[row["LangID"]],
288
  source="toraman2022_en_tr_v2",
289
- source_id=row["TweetID"],
290
- source_label=row["HateLabel"],
291
  original_split="unsplit",
292
- topic=topics[row["TopicID"]],
293
  )
294
  )
295
- return deterministic_stratified_split(
296
- rows,
297
- strata=("language", "topic", "fine_label"),
298
- test_fraction=0.1,
 
 
 
 
 
 
299
  )
 
300
 
301
 
302
  def load_chinese(archive_path: Path) -> dict[str, list[dict[str, object]]]:
303
  splits: dict[str, list[dict[str, object]]] = {"train": [], "test": []}
304
- fine_test = {"0": "normal", "1": "offensive", "2": "hate", "3": "normal"}
305
  target_test = {"0": None, "1": "individual", "2": "group", "3": None}
306
  for upstream_split in ("train", "dev", "test"):
307
  output_split = "test" if upstream_split == "test" else "train"
@@ -311,7 +361,7 @@ def load_chinese(archive_path: Path) -> dict[str, list[dict[str, object]]]:
311
  source_binary = row["label"]
312
  source_fine = row.get("fine-grained-label")
313
  if source_fine is None:
314
- fine_label = "normal" if source_binary == "0" else "abusive"
315
  target = None
316
  source_label = source_binary
317
  else:
@@ -321,7 +371,7 @@ def load_chinese(archive_path: Path) -> dict[str, list[dict[str, object]]]:
321
  splits[output_split].append(
322
  record(
323
  text=row["TEXT"],
324
- label_text="normal" if source_binary == "0" else "abusive",
325
  fine_label=fine_label,
326
  language="zh",
327
  source="cold_zh",
@@ -347,16 +397,19 @@ def load_amharic(paths: dict[str, Path]) -> dict[str, list[dict[str, object]]]:
347
  raise ValueError(
348
  f"Expected two Amharic TSV columns in {upstream_split} row {index}, got {len(row)}"
349
  )
350
- fine_label, text = row
 
 
 
351
  splits[output_split].append(
352
  record(
353
  text=text,
354
- label_text="normal" if fine_label == "normal" else "abusive",
355
- fine_label=fine_label,
356
  language="am",
357
  source="amharic_ranlp2023",
358
  source_id=f"{upstream_split}:{index:05d}",
359
- source_label=fine_label,
360
  original_split=upstream_split,
361
  )
362
  )
@@ -378,8 +431,8 @@ def remove_cross_split_text_leakage(
378
  ) -> tuple[dict[str, list[dict[str, object]]], dict[str, object]]:
379
  """Remove train rows whose exact source/text pair occurs in test.
380
 
381
- Test rows are authoritative. Metadata-only rows are unaffected because
382
- there is no text on which to detect leakage.
383
  """
384
  raw_language_counts = Counter(
385
  str(row["language"]) for rows in splits.values() for row in rows
@@ -396,14 +449,14 @@ def remove_cross_split_text_leakage(
396
 
397
  kept_train: list[dict[str, object]] = []
398
  removed: list[dict[str, object]] = []
399
- conflicting_binary = 0
400
  for row in splits["train"]:
401
  key = (str(row["source"]), str(row["text"]))
402
  matching_test_rows = test_by_key.get(key, []) if row["text_available"] else []
403
  if matching_test_rows:
404
  removed.append(row)
405
  if any(test_row["label"] != row["label"] for test_row in matching_test_rows):
406
- conflicting_binary += 1
407
  else:
408
  kept_train.append(row)
409
 
@@ -411,7 +464,7 @@ def remove_cross_split_text_leakage(
411
  report = {
412
  "policy": "remove train rows with exact source/text match in test",
413
  "removed_train_rows": len(removed),
414
- "removed_rows_with_binary_label_conflict": conflicting_binary,
415
  "removed_by_source": dict(
416
  sorted(Counter(str(row["source"]) for row in removed).items())
417
  ),
@@ -446,7 +499,7 @@ def validate_records(
446
  row["text"] is not None and row["text"] != ""
447
  ):
448
  raise ValueError("text_available disagrees with text")
449
- expected_label = 0 if row["label_text"] == "normal" else 1
450
  if row["label"] != expected_label:
451
  raise ValueError("label disagrees with label_text")
452
 
@@ -552,16 +605,22 @@ def parse_args() -> argparse.Namespace:
552
  def main() -> None:
553
  args = parse_args()
554
  source_paths = download_sources(args.cache_dir)
 
 
 
555
  splits = merge_sources(
556
  [
557
  load_hindi(args.hindi_zip),
558
- load_english_turkish(source_paths["toraman_en_tr"]),
559
  load_chinese(source_paths["cold_zh"]),
560
  load_amharic(source_paths),
561
  ]
562
  )
563
  splits, leakage_report = remove_cross_split_text_leakage(splits)
564
  summary = validate_records(splits, leakage_report)
 
 
 
565
  dataset = to_dataset_dict(splits)
566
  write_outputs(dataset, summary, args.output_dir)
567
  print(json.dumps(summary, indent=2, ensure_ascii=False))
 
3
 
4
  The conversion is intentionally conservative:
5
 
6
+ * ``label`` uses the shared three-way space: neutral, offensive, and hate.
 
7
  * ``fine_label`` and ``source_label`` retain the available source semantics.
8
+ * Chinese positives map to offensive because its main labels do not
9
+ distinguish offensive content from hate.
10
  * Official test splits are preserved. Upstream dev splits are folded into
11
  train. The unsplit English/Turkish data receives a deterministic,
12
  stratified 90/10 split.
 
29
  from datasets import ClassLabel, Dataset, DatasetDict, Features, Value
30
 
31
 
32
+ BUILD_VERSION = "2.0.0"
33
  SPLIT_SEED = "hatespeechmix-v1"
34
 
35
  REPO_ROOT = Path(__file__).resolve().parents[1]
 
38
  SOURCE_SPECS = {
39
  "toraman_en_tr": {
40
  "url": (
41
+ "https://raw.githubusercontent.com/metunlp/hate-speech/"
42
+ "f27c847b34e68c40ac654219b5824f8750b284d8/"
43
+ "dataset_v2/Toraman22_hate_speech_v2.tsv"
44
  ),
45
+ "sha256": "c6c110de8e9656577b5781c8173d7974268c8e716adf5d804f98a50b2f4c0a95",
46
+ "filename": "toraman22_hate_speech_v2.tsv",
47
+ "revision": "f27c847b34e68c40ac654219b5824f8750b284d8",
48
  },
49
  "cold_zh": {
50
  "url": (
 
85
  }
86
 
87
  HINDI_SHA256 = "d419780fe825f9946e3a03da4cf3fdf41699b188932d3662ca304693829994ad"
88
+ LABEL_NAMES = ["neutral", "offensive", "hate"]
89
 
90
  FEATURES = Features(
91
  {
92
  "text": Value("string"),
93
+ "label": ClassLabel(names=LABEL_NAMES),
94
  "label_text": Value("string"),
95
  "fine_label": Value("string"),
96
  "language": Value("string"),
 
188
  topic: str | None = None,
189
  target: str | None = None,
190
  ) -> dict[str, object]:
191
+ if label_text not in LABEL_NAMES:
192
+ raise ValueError(f"Invalid normalized label: {label_text}")
193
  return {
194
  "text": text,
195
+ "label": LABEL_NAMES.index(label_text),
196
  "label_text": label_text,
197
  "fine_label": fine_label,
198
  "language": language,
 
208
 
209
  def load_hindi(hindi_zip: Path) -> dict[str, list[dict[str, object]]]:
210
  verify_file(hindi_zip, HINDI_SHA256)
211
+ fine_labels = {"NONE": "neutral", "OFFN": "offensive", "HATE": "hate", "PRFN": "profane"}
212
+ labels = {"NONE": "neutral", "OFFN": "offensive", "HATE": "hate", "PRFN": "offensive"}
213
  targets = {"NONE": None, "TIN": "targeted", "UNT": "untargeted"}
214
  members = {
215
  "train": "hindi_dataset/hindi_dataset.tsv",
 
219
  for split, member in members.items():
220
  for row in dict_rows_from_zip(hindi_zip, member, delimiter="\t"):
221
  task_1 = row["task_1"]
222
+ label_text = labels[row["task_2"]]
223
  splits[split].append(
224
  record(
225
  text=row["text"],
 
272
  return splits
273
 
274
 
275
+ def normalized_tsv_rows(path: Path) -> tuple[list[dict[str, str]], int]:
276
+ """Read the official TSV after repairing bare carriage returns in text."""
277
+ with path.open(encoding="utf-8-sig", newline="") as handle:
278
+ raw = handle.read()
279
+ crlf_count = raw.count("\r\n")
280
+ bare_cr_count = raw.count("\r") - crlf_count
281
+ normalized = raw.replace("\r\n", "\n").replace("\r", " ")
282
+ rows = list(csv.DictReader(io.StringIO(normalized, newline=""), delimiter="\t"))
283
+ return rows, bare_cr_count
284
+
285
+
286
+ def load_english_turkish(
287
+ tsv_path: Path,
288
+ ) -> tuple[dict[str, list[dict[str, object]]], dict[str, object]]:
289
  lang = {"0": "tr", "1": "en"}
290
  topics = {"0": "religion", "1": "gender", "2": "race", "3": "politics", "4": "sports"}
291
+ labels = {"0": "neutral", "1": "offensive", "2": "hate"}
292
  rows: list[dict[str, object]] = []
293
+ dropped_reasons: Counter[str] = Counter()
294
+ source_rows, normalized_bare_cr = normalized_tsv_rows(tsv_path)
295
+ seen_ids: set[str] = set()
296
+ required_columns = {"tweet_id", "topic", "language", "text", "label"}
297
+
298
+ for row in source_rows:
299
+ if not required_columns.issubset(row) or any(
300
+ row.get(column) is None for column in required_columns
301
+ ):
302
+ dropped_reasons["malformed_columns"] += 1
303
+ continue
304
+ if not row["tweet_id"].isdigit():
305
+ dropped_reasons["invalid_tweet_id"] += 1
306
+ continue
307
+ if row["tweet_id"] in seen_ids:
308
+ dropped_reasons["duplicate_tweet_id"] += 1
309
+ continue
310
+ if row["language"] not in lang:
311
+ dropped_reasons["invalid_language"] += 1
312
+ continue
313
+ if row["topic"] not in topics:
314
+ dropped_reasons["invalid_topic"] += 1
315
+ continue
316
+ if row["label"] not in labels:
317
+ dropped_reasons["invalid_label"] += 1
318
+ continue
319
+ if not row["text"].strip():
320
+ dropped_reasons["empty_text"] += 1
321
+ continue
322
+
323
+ seen_ids.add(row["tweet_id"])
324
+ label_text = labels[row["label"]]
325
  rows.append(
326
  record(
327
+ text=row["text"],
328
+ label_text=label_text,
329
+ fine_label=label_text,
330
+ language=lang[row["language"]],
331
  source="toraman2022_en_tr_v2",
332
+ source_id=row["tweet_id"],
333
+ source_label=row["label"],
334
  original_split="unsplit",
335
+ topic=topics[row["topic"]],
336
  )
337
  )
338
+
339
+ quality_report = {
340
+ "source_rows": len(source_rows),
341
+ "retained_rows": len(rows),
342
+ "dropped_rows": sum(dropped_reasons.values()),
343
+ "dropped_reasons": dict(sorted(dropped_reasons.items())),
344
+ "normalized_bare_carriage_returns": normalized_bare_cr,
345
+ }
346
+ splits = deterministic_stratified_split(
347
+ rows, strata=("language", "topic", "fine_label"), test_fraction=0.1
348
  )
349
+ return splits, quality_report
350
 
351
 
352
  def load_chinese(archive_path: Path) -> dict[str, list[dict[str, object]]]:
353
  splits: dict[str, list[dict[str, object]]] = {"train": [], "test": []}
354
+ fine_test = {"0": "neutral", "1": "offensive", "2": "offensive", "3": "neutral"}
355
  target_test = {"0": None, "1": "individual", "2": "group", "3": None}
356
  for upstream_split in ("train", "dev", "test"):
357
  output_split = "test" if upstream_split == "test" else "train"
 
361
  source_binary = row["label"]
362
  source_fine = row.get("fine-grained-label")
363
  if source_fine is None:
364
+ fine_label = "neutral" if source_binary == "0" else "offensive"
365
  target = None
366
  source_label = source_binary
367
  else:
 
371
  splits[output_split].append(
372
  record(
373
  text=row["TEXT"],
374
+ label_text="neutral" if source_binary == "0" else "offensive",
375
  fine_label=fine_label,
376
  language="zh",
377
  source="cold_zh",
 
397
  raise ValueError(
398
  f"Expected two Amharic TSV columns in {upstream_split} row {index}, got {len(row)}"
399
  )
400
+ source_fine_label, text = row
401
+ label_text = (
402
+ "neutral" if source_fine_label == "normal" else source_fine_label
403
+ )
404
  splits[output_split].append(
405
  record(
406
  text=text,
407
+ label_text=label_text,
408
+ fine_label=label_text,
409
  language="am",
410
  source="amharic_ranlp2023",
411
  source_id=f"{upstream_split}:{index:05d}",
412
+ source_label=source_fine_label,
413
  original_split=upstream_split,
414
  )
415
  )
 
431
  ) -> tuple[dict[str, list[dict[str, object]]], dict[str, object]]:
432
  """Remove train rows whose exact source/text pair occurs in test.
433
 
434
+ Test rows are authoritative. The `text_available` guard keeps this routine
435
+ safe if a future source introduces metadata-only rows.
436
  """
437
  raw_language_counts = Counter(
438
  str(row["language"]) for rows in splits.values() for row in rows
 
449
 
450
  kept_train: list[dict[str, object]] = []
451
  removed: list[dict[str, object]] = []
452
+ conflicting_labels = 0
453
  for row in splits["train"]:
454
  key = (str(row["source"]), str(row["text"]))
455
  matching_test_rows = test_by_key.get(key, []) if row["text_available"] else []
456
  if matching_test_rows:
457
  removed.append(row)
458
  if any(test_row["label"] != row["label"] for test_row in matching_test_rows):
459
+ conflicting_labels += 1
460
  else:
461
  kept_train.append(row)
462
 
 
464
  report = {
465
  "policy": "remove train rows with exact source/text match in test",
466
  "removed_train_rows": len(removed),
467
+ "removed_rows_with_label_conflict": conflicting_labels,
468
  "removed_by_source": dict(
469
  sorted(Counter(str(row["source"]) for row in removed).items())
470
  ),
 
499
  row["text"] is not None and row["text"] != ""
500
  ):
501
  raise ValueError("text_available disagrees with text")
502
+ expected_label = LABEL_NAMES.index(str(row["label_text"]))
503
  if row["label"] != expected_label:
504
  raise ValueError("label disagrees with label_text")
505
 
 
605
  def main() -> None:
606
  args = parse_args()
607
  source_paths = download_sources(args.cache_dir)
608
+ english_turkish, english_turkish_quality = load_english_turkish(
609
+ source_paths["toraman_en_tr"]
610
+ )
611
  splits = merge_sources(
612
  [
613
  load_hindi(args.hindi_zip),
614
+ english_turkish,
615
  load_chinese(source_paths["cold_zh"]),
616
  load_amharic(source_paths),
617
  ]
618
  )
619
  splits, leakage_report = remove_cross_split_text_leakage(splits)
620
  summary = validate_records(splits, leakage_report)
621
+ summary["source_quality"] = {
622
+ "toraman2022_en_tr_v2": english_turkish_quality,
623
+ }
624
  dataset = to_dataset_dict(splits)
625
  write_outputs(dataset, summary, args.output_dir)
626
  print(json.dumps(summary, indent=2, ensure_ascii=False))
tests/test_build_dataset.py CHANGED
@@ -15,12 +15,12 @@ build_dataset = importlib.util.module_from_spec(SPEC)
15
  SPEC.loader.exec_module(build_dataset)
16
 
17
 
18
- def test_binary_mapping_rejects_unknown_label() -> None:
19
- with pytest.raises(ValueError, match="Invalid binary label"):
20
  build_dataset.record(
21
  text="sample",
22
- label_text="hate",
23
- fine_label="hate",
24
  language="en",
25
  source="fixture",
26
  source_id="1",
@@ -32,9 +32,9 @@ def test_binary_mapping_rejects_unknown_label() -> None:
32
  def test_stratified_split_is_deterministic_and_disjoint() -> None:
33
  rows = [
34
  build_dataset.record(
35
- text=None,
36
- label_text="normal" if index % 2 == 0 else "abusive",
37
- fine_label="normal" if index % 2 == 0 else "offensive",
38
  language="en",
39
  source="fixture",
40
  source_id=str(index),
@@ -67,8 +67,34 @@ def test_built_dataset_schema_and_counts() -> None:
67
  assert len(dataset["train"]) + len(dataset["test"]) == 187_315
68
  label = dataset["train"].features["label"]
69
  assert isinstance(label, ClassLabel)
70
- assert label.names == ["normal", "abusive"]
71
- assert sum(dataset[split]["text_available"].count(False) for split in dataset) == 128_907
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72
 
73
 
74
  def test_no_exact_source_text_leakage() -> None:
 
15
  SPEC.loader.exec_module(build_dataset)
16
 
17
 
18
+ def test_three_way_mapping_rejects_unknown_label() -> None:
19
+ with pytest.raises(ValueError, match="Invalid normalized label"):
20
  build_dataset.record(
21
  text="sample",
22
+ label_text="abusive",
23
+ fine_label="abusive",
24
  language="en",
25
  source="fixture",
26
  source_id="1",
 
32
  def test_stratified_split_is_deterministic_and_disjoint() -> None:
33
  rows = [
34
  build_dataset.record(
35
+ text=f"sample {index}",
36
+ label_text="neutral" if index % 2 == 0 else "offensive",
37
+ fine_label="neutral" if index % 2 == 0 else "offensive",
38
  language="en",
39
  source="fixture",
40
  source_id=str(index),
 
67
  assert len(dataset["train"]) + len(dataset["test"]) == 187_315
68
  label = dataset["train"].features["label"]
69
  assert isinstance(label, ClassLabel)
70
+ assert label.names == ["neutral", "offensive", "hate"]
71
+ assert all(all(dataset[split]["text_available"]) for split in dataset)
72
+ assert sum(
73
+ source == "toraman2022_en_tr_v2"
74
+ for split in dataset.values()
75
+ for source in split["source"]
76
+ ) == 128_907
77
+
78
+
79
+ def test_source_specific_label_mapping() -> None:
80
+ local_dataset = ROOT / "local_dataset"
81
+ if not local_dataset.exists():
82
+ pytest.skip("Run scripts/build_dataset.py before the integration test")
83
+
84
+ dataset = load_from_disk(local_dataset)
85
+ rows = (row for split in dataset.values() for row in split)
86
+ chinese_labels = set()
87
+ profanity_rows = 0
88
+ for row in rows:
89
+ if row["language"] == "zh":
90
+ chinese_labels.add(row["label_text"])
91
+ if "|PRFN|" in row["source_label"]:
92
+ profanity_rows += 1
93
+ assert row["label_text"] == "offensive"
94
+ assert row["fine_label"] == "profane"
95
+
96
+ assert chinese_labels == {"neutral", "offensive"}
97
+ assert profanity_rows > 0
98
 
99
 
100
  def test_no_exact_source_text_leakage() -> None: