Upload folder using huggingface_hub

README.md

@@ -1,74 +1,54 @@
 ---
-language:
-- kab
 license: apache-2.0
-tags:
-- kabyle
-- taqbaylit
-- berber
-- amazigh
-- punctuation-restoration
-- truecasing
-- token-classification
-- asr-post-processing
-- low-resource
+language:
+  - kab
+language_bcp47:
+  - kab-Latn
+size_categories:
+  - 1M<n<10M
 task_categories:
-- token-classification
+  - token-classification
 task_ids:
-- part-of-speech
-pretty_name: KabPunct
-size_categories:
-- 1M<n<10M
+  - part-of-speech
+pretty_name: KabPunct — Kabyle punctuation and casing restoration corpus
+tags:
+  - kabyle
+  - taqbaylit
+  - berber
+  - tamazight
+  - low-resource
+  - punctuation-restoration
+  - truecasing
+  - asr-post-processing
 configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: default/train.jsonl
-  - split: dev
-    path: default/dev.jsonl
-  - split: test
-    path: default/test.jsonl
-- config_name: ood
-  data_files:
-  - split: ood
-    path: ood/ood.jsonl
+  - config_name: default
+    data_files:
+      - split: train
+        path: default/train.jsonl
+      - split: dev
+        path: default/dev.jsonl
+      - split: test
+        path: default/test.jsonl
+  - config_name: ood
+    data_files:
+      - split: ood
+        path: ood/ood.jsonl
 ---
 
 # KabPunct
 
-**Punctuation and capitalisation restoration corpus for Kabyle** (Taqbaylit, `kab`, Latin
-script). Every row is one sentence broken into words with two label sequences: which mark
-follows each word (`NONE`, `COMMA`, `PERIOD`, `QUESTION`, `COLON`) and how the word is cased
-(`LOWER`, `UPPER_INIT`).
+A punctuation and capitalisation restoration corpus for Kabyle (Taqbaylit, `kab`, Latin
+script), from the [AƔBALU](https://huggingface.co/agbalu) project.
 
-It is the training and evaluation corpus for
+**1,318,707 word-labelled sentences** drawn from the full AƔBALU-Text v1 corpus and
+speaker-disjoint Common Voice Kabyle splits. Every sentence is broken into lowercased ASR
+tokens with two parallel label sequences: which punctuation mark follows each word, and how
+the word is capitalised. It is the training and evaluation corpus for
 [`agbalu/Belaid-31M`](https://huggingface.co/agbalu/Belaid-31M), which restores punctuation
-at **0.793 macro-F1** and non-initial proper nouns at **0.933 F1** on the `test` split. It is,
-as far as we can establish, **the first labelled punctuation restoration corpus for Kabyle or
-any Berber language.**
+at **0.793 macro-F1** on the held-out `test` split.
 
-## Splits
-
-| split | config | rows | words | source |
-|---|---|---|---|---|
-| `train` | `default` | 1,262,922 | 12,101,835 | Text corpus + Common Voice train |
-| `dev` | `default` | 5,597 | 29,280 | Common Voice dev, decontaminated |
-| `test` | `default` | 5,160 | 26,969 | Common Voice test, decontaminated |
-| `ood` | `ood` | 45,028 | 1,001,370 | Long-form prose (HCA), held out of training |
-
-## Format
-
-Each row has four fields:
-
-| field | type | description |
-|---|---|---|
-| `words` | `list[str]` | Lowercased, unpunctuated word tokens — what an ASR system emits |
-| `punctuation` | `list[str]` | Per-word punctuation label: `NONE`, `COMMA`, `PERIOD`, `QUESTION`, or `COLON` |
-| `case` | `list[str]` | Per-word casing label: `LOWER` or `UPPER_INIT` |
-| `source` | `str` | Provenance tag — which corpus this sentence came from |
-
-Every list has the same length (one entry per word). The labels align to words, not subwords:
-`words[i]` takes mark `punctuation[i]` and initial casing `case[i]`.
+It is, as far as we can establish, **the first labelled punctuation restoration corpus for
+Kabyle or any Berber language.**
 
 ```python
 from datasets import load_dataset
@@ -83,7 +63,6 @@ row = ds["train"][0]
 #   "source":      "hf.abdelhaqueidali.kab-latn-tfng"
 # }
 
-# Reconstruct the original sentence from the labels
 def restore(row):
     mark_map = {"NONE": "", "COMMA": ",", "PERIOD": ".", "QUESTION": "?", "COLON": ":"}
     parts = []
@@ -96,60 +75,93 @@ print(restore(row))
 # "Tecfiḍ fell-i?"
 ```
 
-## Label scheme
+## Splits
 
-### Punctuation
+| split | config | rows | words | description |
+|---|---|---|---|---|
+| `train` | `default` | 1,262,922 | 12,101,835 | Text corpus + Common Voice train |
+| `dev` | `default` | 5,597 | 29,280 | Common Voice dev, decontaminated |
+| `test` | `default` | 5,160 | 26,969 | Common Voice test, decontaminated |
+| `ood` | `ood` | 45,028 | 1,001,370 | Long-form prose (HCA), held out of training entirely |
 
-Five classes derived from what survives decontamination at usable support:
+## Schema
 
-| label | character | notes |
+| field | type | description |
 |---|---|---|
-| `NONE` | — | No mark follows this word |
-| `COMMA` | `,` | |
-| `PERIOD` | `.` | `!` and `;` fold here: together 0.41% of tokens, and the corpus carries identical sentences under both, so the distinction is not recoverable |
-| `QUESTION` | `?` | |
-| `COLON` | `:` | 43 examples in test; support is low and the score should be quoted with that named |
+| `words` | `list[str]` | Lowercased, unpunctuated word tokens — the form an ASR system emits |
+| `punctuation` | `list[str]` | Per-word punctuation label: `NONE`, `COMMA`, `PERIOD`, `QUESTION`, or `COLON` |
+| `case` | `list[str]` | Per-word casing label: `LOWER` or `UPPER_INIT` |
+| `source` | `str` | Provenance tag: which corpus this sentence came from |
 
-### Casing
+Every list is the same length (one entry per word). Labels align to words, not subwords:
+`words[i]` takes mark `punctuation[i]` and initial casing `case[i]`.
 
-Two classes:
+## Label Scheme
+
+### Punctuation
+
+| label | character | support in `test` |
+|---|---|---|
+| `NONE` | — | 20,904 |
+| `PERIOD` | `.` | 4,357 |
+| `QUESTION` | `?` | 861 |
+| `COMMA` | `,` | 804 |
+| `COLON` | `:` | 43 |
+
+`!` and `;` fold into `PERIOD`. Together they account for 0.41% of tokens, and the corpus
+contains identical sentences under both marks — the distinction is not recoverable from text
+alone, and collapsing them costs less than carrying a class that cannot be learned.
+
+### Casing
 
 | label | meaning |
 |---|---|
 | `LOWER` | Word begins lowercase |
 | `UPPER_INIT` | First character is uppercase (sentence-initial or proper noun) |
 
-`ALL_CAPS` was a third class. It occurs once in dev and three times in test; the model that
-carried it emitted fourteen spurious all-capitals words for the three real ones. Restoring an
-acronym to full capitals is out of scope and the class is absent.
+`ALL_CAPS` was trialled as a third class. It occurs once in dev and three times in test —
+too few to learn without spurious generalisation. The model that carried it emitted fourteen
+all-capitals words for the three real ones. Restoring an acronym to full capitals is out
+of scope, and the class is absent.
 
 ## Decontamination
 
-**58.2% of Common Voice Kabyle transcripts also appear in AƔBALU-Text v1**, because most are
-Tatoeba sentences and Tatoeba is in the corpus. Every clip whose text appears in the text
-corpus is excluded from `dev` and `test`, which decontaminates the `Masinissa-31M` encoder
-used by `Belaid-31M` at the same time (it was pretrained on the same file). The rows scored
-in the `Belaid-31M` card are unseen by both the punctuation heads and the backbone.
+**58.2% of Common Voice Kabyle transcripts also appear in AƔBALU-Text v1**, because the
+majority are Tatoeba sentences and Tatoeba is in the corpus. Every clip whose text is found
+in the text corpus is removed from `dev` and `test`. This simultaneously decontaminates the
+`Masinissa-31M` encoder used by `Belaid-31M`, which was pretrained on the same file. The
+rows in the `Belaid-31M` evaluation are unseen by both the punctuation heads and the backbone.
 
-A transcript carrying no final mark is dropped rather than labelled `NONE`. Reading them
-settles what they are: no punctuation and no capitals — contributor typing habit, not Kabyle
-grammar. Training on them teaches transcriber style, not language.
+A transcript without a sentence-final mark is dropped rather than labelled `NONE`. Those
+transcripts carry neither punctuation nor capitals — transcriber habit, not Kabyle grammar.
 
 `opus.nllb-kab` (871,663 records) is excluded from training entirely. Its record boundaries
-were chosen by an NLLB bitext miner rather than a writer; a record can end mid-sentence, and
-sentence-final punctuation is exactly the label being predicted.
+were chosen by a bitext miner rather than a writer; a record can end mid-sentence, and
+sentence-final punctuation is the predicted label.
+
+## Out-of-Domain Split (`ood`)
 
-## Out-of-domain split (`ood`)
+The `ood` split is `hf.imsidag.kabyle-corpus-hca` — 45,028 records of continuous long-form
+Kabyle prose, held out of training entirely. Common Voice records are one sentence each;
+HCA records average 22 words across several sentences.
 
-The `ood` split is `hf.imsidag.kabyle-corpus-hca` — 45,028 records of long-form Kabyle
-prose, held out of training entirely. Common Voice records are one sentence each by
-construction; HCA records average 22 words across several sentences. The model's performance
-degradation on `ood` (macro-F1 0.635 vs 0.793 on test) is documented in the `Belaid-31M`
-card and explained there: it learned *one period at the end* from single-sentence training
-examples, and that pattern breaks on multi-sentence prose.
+The `Belaid-31M` card reports this split explicitly alongside `test`. A model evaluated only
+in its training domain has not been evaluated.
 
-Both splits are reported because a model evaluated only in its training domain has not been
-evaluated.
+## Composition
+
+| source id | type | train rows |
+|---|---|---|
+| `hf.abdelhaqueidali.kab-latn-tfng` | text corpus | 647,039 |
+| `hf.imsidag.kabyle-corpus-ummto` | text corpus | 136,799 |
+| `hf.imsidag.kabyle-raw-text` | text corpus | 121,338 |
+| `hf.imsidag.kabyle-corpus-ubouira` | text corpus | 76,336 |
+| `hf.boffire.kab-en-toponyms-sentences` | text corpus | 30,584 |
+| `speech.train` | Common Voice 22.0 train | 151,100 |
+| others (28 sources) | text corpus | 99,726 |
+| `opus.nllb-kab` | **excluded** (miner-cut boundaries) | — |
+
+The `ood` config contains `hf.imsidag.kabyle-corpus-hca` exclusively (45,028 rows).
 
 ## Reproduction
 
@@ -157,32 +169,30 @@ evaluated.
 make punctuation TASK=corpus
 ```
 
-The corpus builder reads AƔBALU-Text v1 and the Common Voice speech splits, applies the
-decontamination procedure, and writes the four files to `data/processed/punctuation/`.
-
-## Limitations
+Reads AƔBALU-Text v1 and the Common Voice speech splits, applies decontamination, and writes
+the four files to `data/processed/punctuation/`.
 
-**No inter-annotator agreement has been measured for Kabyle punctuation.** Nobody has
-established what two fluent readers agree on, so no figure here can be read as a fraction of
-what is attainable. Comma placement in particular has no codified convention in Kabyle and
-the corpus writers do not agree with each other.
+## Citation
 
-**One language.** Trained and evaluated on Kabyle. Related Berber languages (Tashelhit,
-Tarifit, Central Atlas Tamazight) were not tested and should not be assumed.
-
-**`ood` is one source.** Long-form prose generalisation is estimated from a single corpus.
+```bibtex
+@misc{agbalu_kabpunct,
+  title  = {KabPunct: a Kabyle punctuation and casing restoration corpus},
+  author = {AƔBALU},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/agbalu/KabPunct}
+}
+```
 
 ## Licence
 
-**Apache-2.0** for this dataset release. The underlying sentences carry their upstream
-licences:
+**Apache-2.0** for the label annotations and the dataset release. The underlying sentences
+retain their upstream licences:
 
-- Common Voice clips: CC-0 (Mozilla).
-- Text corpus: mixed composition — see
-  [`agbalu/Masinissa-31M`](https://huggingface.co/agbalu/Masinissa-31M) for the full
-  breakdown. The published splits contain no `opus.nllb-kab` rows (excluded during corpus
-  construction) and no rows from share-alike-only sources that would impose obligations on
-  redistribution. Apache-2.0 applies to the label annotations; the sentences keep their
-  upstream licence.
+- Common Voice train/dev/test clips: **CC0-1.0** (Mozilla Foundation).
+- Text corpus sentences: **mixed** — see the composition table above. Every share-alike source
+  (`opus.nllb-kab`, ODbL-licensed rows) was excluded before publication. No row in any
+  published split imposes a redistribution obligation on downstream users.
+- The label sequences (`punctuation`, `case`) are this project's annotation and are
+  **Apache-2.0**.
 
 Part of [AƔBALU](https://huggingface.co/agbalu), a Kabyle corpus and model collection.