Add dataset card (source credit, schema, caveats)

README.md

@@ -78,4 +78,74 @@ configs:
   data_files:
   - split: train
     path: data/train-*
+license: cc-by-4.0
+language:
+- akk
+task_categories:
+- token-classification
+tags:
+- akkadian
+- assyriology
+- cuneiform
+- ancient-near-east
+- dependency-parsing
+- conllu
+- glam
+- cultural-heritage
+size_categories:
+- 1K<n<10K
 ---
+
+# State Archives of Assyria — Morphosyntactic Letter Corpus
+
+7,355 fully annotated sentences from **2,586 Neo-Assyrian letters** — the royal
+correspondence of the Assyrian empire, 8th–7th century BCE.
+
+Every sentence carries full morphosyntactic annotation (lemma, UPOS, morphological
+features, dependency head and relation) **and** the letter's archival context: who wrote it,
+who received it, which archive it was excavated from, and where.
+
+Row 0 is Adda-hati writing to the king, from the Northwest Palace at Nimrud:
+
+> `ana šarri bēlīya urdaka Adda-hati` — *"To the king, my lord: your servant Adda-hati."*
+
+## Fields
+
+- `tokens` — `{id, form, lemma, upos, xpos, feats, head, deprel}` (CoNLL-U)
+- `text` — the normalised sentence
+- `text_id` (ORACC/CDLI P-number), `sent_id`, `volume` (`saa01`–`saa21`), `sentence_idx`
+- Letter metadata joined from the ORACC catalogue: `ancient_author`, `ancient_recipient`,
+  `archive`, `collection`, `provenience`, `genre`, `dialect`, `language`, `period`, `ruler`,
+  `script`, `object_type`, `material`, `museum_no`, `date_of_origin`, `designation`,
+  `vol_title`, `primary_publication`
+
+Covers 10 of the 21 SAA volumes. All 2,586 letters join to a catalogue record.
+
+## Notes on the conversion
+
+Two quirks of the deposit are handled here, and are worth knowing if you go back to the source:
+
+- The deposit ships a `concatenated/` directory holding 10 per-volume merged files — **the
+  same letters again**, in UD-release form. They are excluded; including them would double
+  the corpus.
+- The `*Renumbered` files open with a provenance pseudo-sentence (`saao / saa01 / P224395`,
+  tokenised but unlemmatised). These are dropped — they are not sentences of the letters.
+
+Filenames in the source are not uniform (`P<id>lookup`, `P<id>Renumbered`,
+`P<id>lookupRenumbered`, `P<id>_lookup`, `P<id>_intermediate`, plus a few `X`/`Q` ids), so
+text ids are extracted by pattern rather than by stripping suffixes.
+
+## Caveats
+
+- Akkadian is a **broken-text** tradition. Damaged/illegible signs appear as `x` runs in the
+  text; they are preserved, not cleaned.
+- Annotation quality descends from ORACC's lemmatisation; `xpos` and `feats` are sparse in
+  some volumes.
+
+## Source & credit
+
+Matthew Ong. *A Full Morphosyntactic Annotation of the State Archives of Assyria Letter
+Corpus.* Zenodo, 2024-02-06. <https://zenodo.org/records/10622984> — deposit CC-BY-4.0.
+
+Underlying text and lemmatisation from **ORACC** (<http://oracc.org/saao/>), released CC0;
+adapted from the State Archives of Assyria print volumes. Please cite both.