Upload README.md with huggingface_hub

README.md

@@ -1,19 +1,69 @@
 ---
-dataset_info:
-  features:
-  - name: macronized
-    dtype: string
-  - name: plain
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 833127461
-    num_examples: 2234599
-  download_size: 399705065
-  dataset_size: 833127461
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+language:
+- grc
+license: other
+task_categories:
+- text-generation
+- token-classification
+tags:
+- ancient-greek
+- vowel-length
+- macronization
+- classics
+pretty_name: OGA Macronized (Ancient Greek Vowel Length)
+size_categories:
+- 1M<n<10M
 ---
+
+# OGA Macronized
+
+Automatically macronized Ancient Greek text: the phonemic length of the three
+"dichrona" (ambiguous-length letters) — alpha (α), iota (ι), upsilon (υ) — has
+been marked wherever the [grc-macronizer](https://github.com/Urdatorn/grc-macronizer)
+rule-based system could disambiguate it with confidence. A trailing `^` marks a
+short dichronon, a trailing `_` marks a long one; dichrona the macronizer could
+not resolve are left unmarked (never guessed).
+
+## Source
+
+Text is the [Opera Graeca Adnotata](https://doi.org/10.5281/zenodo.14206061)
+corpus (Celano 2024) — ~40M tokens of Ancient Greek aggregating First1KGreek,
+the Patristic Text Archive, and Perseus' canonical-greekLit, with lemma/POS/
+morphology already annotated by Celano's own parser. That existing annotation
+was converted directly into the macronizer's internal representation (bypassing
+any separate NLP pipeline) and run through the rule-based macronizer.
+
+- 2,234,599 sentences
+- Overall dichrona-in-open-syllables macronization ratio: 69.20%
+  (11,627,591 / 16,802,080 disambiguated)
+
+## Fields
+
+- `macronized`: the sentence with `^`/`_` markup, e.g. `χρὴ γι^νώσκειν ὅτι^ πά_σης τῆς γῆς`
+- `plain`: the same sentence with markup stripped (i.e. ordinary Greek orthography)
+
+## Intended use
+
+Training data for sequence models that predict Ancient Greek vowel length
+(e.g. a char-level tagger), and as an intermediate resource for downstream
+prosody-aware NLP (verse scansion, metrical classification, morphology
+disambiguation of the dichrona-bearing minimal pairs like ἄν/ἂν).
+
+Note: labels are *silver*, not gold — they come from a rule-based system that
+is deliberately conservative (does not guess), not from human annotation. For
+gold-standard evaluation, see
+[Norma Syllabarum Graecarum](https://github.com/Urdatorn/norma-syllabarum-graecarum).
+
+## License
+
+The underlying OGA corpus is distributed by its authors under the terms shown
+on its [Zenodo record](https://doi.org/10.5281/zenodo.14206061); check that
+record for the authoritative license before redistribution. The macronization
+itself is produced by software licensed under GNU GPL v3
+([grc-macronizer](https://github.com/Urdatorn/grc-macronizer)).
+
+## Citation
+
+> Giuseppe G. A. Celano. Opera Graeca Adnotata: Building a 34M+ Token Multilayer Corpus for Ancient Greek. arXiv:2404.00739.
+
+> Thörn Cleland, Albin and Eric Cullhed (forthcoming). Automatic Annotation of Ancient Greek Vowel Length.