Publish AMTC dataset v1.0.0

README.md

@@ -0,0 +1,87 @@
+---
+pretty_name: African Multilingual Tokenizer Challenge
+language:
+- en
+- fr
+- ha
+- sw
+- yo
+- am
+license:
+- cc-by-sa-3.0
+- gfdl
+task_categories:
+- text-generation
+size_categories:
+- 100K<n<1M
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*.parquet
+  - split: validation
+    path: data/validation-*.parquet
+dataset_info:
+  features:
+  - name: language
+    dtype: string
+  - name: text
+    dtype: string
+  - name: source_id
+    dtype: string
+  - name: source_url
+    dtype: string
+  - name: source_config
+    dtype: string
+  - name: source_revision
+    dtype: string
+  - name: segment_index
+    dtype: int32
+  - name: content_sha256
+    dtype: string
+---
+
+# African Multilingual Tokenizer Challenge dataset
+
+The frozen public corpus for the African Multilingual Tokenizer Challenge. It contains one balanced multilingual `train` split and one balanced `validation` split.
+
+| Split | Per language | Total |
+| --- | ---: | ---: |
+| Train | 40,000 | 240,000 |
+| Validation | 4,000 | 24,000 |
+
+Languages are English (`en`), French (`fr`), Hausa (`ha`), Swahili (`sw`), Yoruba (`yo`) and Amharic (`am`). Official public-test and private-test text are deliberately absent from this repository.
+
+## Load
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset(
+    "Similoluwa/african-multilingual-tokenizer-challenge",
+    revision="v1.0.0",
+)
+```
+
+Tokenizer training needs only `language` and `text`. The remaining columns preserve page-level attribution and reproducibility.
+
+## Source and license
+
+All passages are derived from the `20231101.{language}` configurations of [`wikimedia/wikipedia`](https://huggingface.co/datasets/wikimedia/wikipedia), pinned to source revision `b04c8d1ceb2f5cd4588862100d08de323dccfbaa`. Wikipedia textual content is available under the Creative Commons Attribution-ShareAlike 3.0 license and GNU Free Documentation License; individual pages may have additional terms. Every row retains the source page URL and article ID. Reusers must comply with the source licenses, including attribution and share-alike obligations.
+
+## Construction
+
+- Unicode is normalized to NFC; language-bearing diacritics are retained.
+- Cleaned Wikipedia articles are greedily segmented around a 120-character target, with a hard 40 to 320 character range.
+- No article contributes to more than one split.
+- Exact duplicates after NFKC, case-folding and whitespace normalization are removed globally.
+- Each article contributes at most 64 passages.
+- Per-language split counts are exact and balanced.
+- Files are Parquet with Zstandard compression.
+- `manifest.json` records source revision, algorithm parameters, counts and SHA-256 checksums.
+
+## Limitations and required review
+
+Wikipedia is not a representative sample of everyday language and can contain bias, personal information, offensive text, markup residue or language-mixing. Configuration identity is used as the primary language label; no sentence-level human relabeling is claimed. Before a prize-bearing launch, proficient reviewers for every language must review samples from every split and the organizers must record acceptance criteria and remediation. The hidden test splits require the same review.
+
+This dataset is suitable for tokenizer experimentation, not for evaluating factual accuracy or demographic representativeness.

data/train-am.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfc9b3593f977d410182daba1b2d2f706e2e384d24c06c86a82670b07dee15cb
+size 5645429

data/train-en.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fd4568e9a57934baf64daceb34097e8fd9303bec7d2f5760eb50daf3d3e2702
+size 4343053

data/train-fr.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e4d18d745633e95e41a0a0708a6e458574e4f0c72f235544007b18c5429e3ae
+size 4485073

data/train-ha.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16c8a22c9590ca9d2a2e737f3e1ef37e2fb90d7d3f7f256ffe3367dc8237daa2
+size 4213822

data/train-sw.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc129a039f8e363043615897c0eb40b0f10ccf54a08f49822ed47d89b6ba2cd1
+size 3755496

data/train-yo.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eed86ead20ebe2193c79f6d0d05ba9b8f881ed65ebf5b159312bde42504d78f
+size 4690281

data/validation-am.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:638149069914fa578d1588259787c07eed9ce128e2f8d4879393853ce3ea88f5
+size 543192

data/validation-en.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8e93ffbf683f69ed722b0ecb84f864820d3b675f44b80745e3cae7be44e6f6
+size 432701

data/validation-fr.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa91286b9db816ded1d9d45f16a5a72fbe305adb3bd90378e2f218ac66951b8
+size 440286

data/validation-ha.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c4f938b4a7d5ed26c89ebaf2b7479613e47b08e78feb4e131337e215a734dd2
+size 412242

data/validation-sw.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36ea4136453ee6b558bb9fd1e2d83caef11cdea1b64f329d5c22bd750ee6f777
+size 369354

data/validation-yo.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2686fc7f5a08c4f00e688f861fd028515c93ee078f85c625aa5e3a5062bb0b80
+size 457289

manifest.json

@@ -0,0 +1,189 @@
+{
+  "dataset_version": "1.0.0",
+  "pipeline_version": "3.0.0",
+  "seed": "amtc-v1-2026",
+  "source_dataset": "wikimedia/wikipedia",
+  "source_revision": "b04c8d1ceb2f5cd4588862100d08de323dccfbaa",
+  "source_date": "20231101",
+  "targets_per_language": {
+    "train": 40000,
+    "validation": 4000,
+    "public_test": 4000,
+    "private_test": 8000
+  },
+  "languages": [
+    "en",
+    "fr",
+    "ha",
+    "sw",
+    "yo",
+    "am"
+  ],
+  "segmentation": {
+    "min_characters": 40,
+    "target_characters": 120,
+    "max_characters": 320,
+    "max_segments_per_article": 64,
+    "unicode_normalization": "NFC"
+  },
+  "language_reports": [
+    {
+      "language": "en",
+      "language_name": "English",
+      "articles_scanned": 1056,
+      "duplicates_removed": 41,
+      "segments_discarded_after_quota": 163,
+      "counts": {
+        "validation": 4000,
+        "public_test": 4000,
+        "private_test": 8000,
+        "train": 40000
+      }
+    },
+    {
+      "language": "fr",
+      "language_name": "French",
+      "articles_scanned": 1168,
+      "duplicates_removed": 171,
+      "segments_discarded_after_quota": 175,
+      "counts": {
+        "train": 40000,
+        "private_test": 8000,
+        "validation": 4000,
+        "public_test": 4000
+      }
+    },
+    {
+      "language": "ha",
+      "language_name": "Hausa",
+      "articles_scanned": 6494,
+      "duplicates_removed": 584,
+      "segments_discarded_after_quota": 55,
+      "counts": {
+        "train": 40000,
+        "private_test": 8000,
+        "public_test": 4000,
+        "validation": 4000
+      }
+    },
+    {
+      "language": "sw",
+      "language_name": "Swahili",
+      "articles_scanned": 5993,
+      "duplicates_removed": 1166,
+      "segments_discarded_after_quota": 46,
+      "counts": {
+        "public_test": 4000,
+        "validation": 4000,
+        "private_test": 8000,
+        "train": 40000
+      }
+    },
+    {
+      "language": "yo",
+      "language_name": "Yoruba",
+      "articles_scanned": 31002,
+      "duplicates_removed": 19211,
+      "segments_discarded_after_quota": 70,
+      "counts": {
+        "train": 40000,
+        "validation": 4000,
+        "public_test": 4000,
+        "private_test": 8000
+      }
+    },
+    {
+      "language": "am",
+      "language_name": "Amharic",
+      "articles_scanned": 13388,
+      "duplicates_removed": 1834,
+      "segments_discarded_after_quota": 60,
+      "counts": {
+        "public_test": 4000,
+        "train": 40000,
+        "private_test": 8000,
+        "validation": 4000
+      }
+    }
+  ],
+  "global_unique_content_hashes": 336000,
+  "visibility": "public",
+  "splits": [
+    "train",
+    "validation"
+  ],
+  "files": [
+    {
+      "path": "dataset_release/build/public/data/train-en.parquet",
+      "rows": 40000,
+      "bytes": 4343053,
+      "sha256": "4fd4568e9a57934baf64daceb34097e8fd9303bec7d2f5760eb50daf3d3e2702"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-en.parquet",
+      "rows": 4000,
+      "bytes": 432701,
+      "sha256": "5a8e93ffbf683f69ed722b0ecb84f864820d3b675f44b80745e3cae7be44e6f6"
+    },
+    {
+      "path": "dataset_release/build/public/data/train-fr.parquet",
+      "rows": 40000,
+      "bytes": 4485073,
+      "sha256": "9e4d18d745633e95e41a0a0708a6e458574e4f0c72f235544007b18c5429e3ae"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-fr.parquet",
+      "rows": 4000,
+      "bytes": 440286,
+      "sha256": "3fa91286b9db816ded1d9d45f16a5a72fbe305adb3bd90378e2f218ac66951b8"
+    },
+    {
+      "path": "dataset_release/build/public/data/train-ha.parquet",
+      "rows": 40000,
+      "bytes": 4213822,
+      "sha256": "16c8a22c9590ca9d2a2e737f3e1ef37e2fb90d7d3f7f256ffe3367dc8237daa2"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-ha.parquet",
+      "rows": 4000,
+      "bytes": 412242,
+      "sha256": "2c4f938b4a7d5ed26c89ebaf2b7479613e47b08e78feb4e131337e215a734dd2"
+    },
+    {
+      "path": "dataset_release/build/public/data/train-sw.parquet",
+      "rows": 40000,
+      "bytes": 3755496,
+      "sha256": "dc129a039f8e363043615897c0eb40b0f10ccf54a08f49822ed47d89b6ba2cd1"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-sw.parquet",
+      "rows": 4000,
+      "bytes": 369354,
+      "sha256": "36ea4136453ee6b558bb9fd1e2d83caef11cdea1b64f329d5c22bd750ee6f777"
+    },
+    {
+      "path": "dataset_release/build/public/data/train-yo.parquet",
+      "rows": 40000,
+      "bytes": 4690281,
+      "sha256": "1eed86ead20ebe2193c79f6d0d05ba9b8f881ed65ebf5b159312bde42504d78f"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-yo.parquet",
+      "rows": 4000,
+      "bytes": 457289,
+      "sha256": "2686fc7f5a08c4f00e688f861fd028515c93ee078f85c625aa5e3a5062bb0b80"
+    },
+    {
+      "path": "dataset_release/build/public/data/train-am.parquet",
+      "rows": 40000,
+      "bytes": 5645429,
+      "sha256": "bfc9b3593f977d410182daba1b2d2f706e2e384d24c06c86a82670b07dee15cb"
+    },
+    {
+      "path": "dataset_release/build/public/data/validation-am.parquet",
+      "rows": 4000,
+      "bytes": 543192,
+      "sha256": "638149069914fa578d1588259787c07eed9ce128e2f8d4879393853ce3ea88f5"
+    }
+  ]
+}