Publish structured columns instead of JSON inside strings

Remove annotation-scheme leakage from model-facing columns

Add exact 50-sentence benchmark task

Rebuild Sphragis from the complete treebank sentence pool

Chunk training splits at fixed task sizes

Enforce exact verse boundaries and nested CoNLL-U

Represent text chunks as JSON lists

Remove identifying metadata from metrical lines

Fix metrical chunk aggregation

Unify source text across task sizes

Remove CoNLL-U provenance leakage

Validate all CoNLL-U with official loader

Freeze exact evaluation chunks

Add chunk-size dataset variants

Rebuild work-stratified splits with seed 776

Restore core Hesiodic works

Curate conservative authorship benchmark

Align verse across sentence boundaries

Add Sphregis 0.1 beta dataset