Represent text chunks as JSON lists

README.md

@@ -188,6 +188,12 @@ All configurations contain:
   revision, URL, license, annotation provenance and syntax scheme;
 - `licenses`, `dedup_key`, and `split`.
 
+`text` is always a JSON list of strings. Atomic `_1` rows and training rows
+contain one string; validation and test rows in `_10` and `_100` contain
+exactly 10 or 100 separate sentence strings (`verse_sentence` and `prose`) or
+line strings (`verse_metre`). No whitespace separator is used to encode unit
+boundaries.
+
 The `_10` and `_100` configurations additionally contain `chunk_size`,
 `chunk_target_size`, ordered `constituent_ids`, JSON
 `constituent_provenance`, `chunk_work_ids`, `chunk_works`, and
@@ -197,8 +203,10 @@ rows always have `chunk_size` equal to the configuration suffix. Aggregated
 retains the original passage-level locations.
 
 `verse_sentence` additionally contains `metrical_line_ids` and a JSON list of
-the full overlapping `metrical_lines`; an overlap is marked `full` or
-`partial`. `verse_metre` contains the native list `parent_sentence_ids`,
+the overlapping `metrical_lines`. Each serialized line is strictly limited to
+`text`, `metre`, and `syllables`; identifying Hypotactic metadata and internal
+IDs are never included in this model-facing field. `verse_metre` contains the
+native list `parent_sentence_ids`,
 `book`, `poem_sequence`, `line_number`, raw Hypotactic `metre`, flattened
 syllable-level annotations, and the originating filename. Its `conllu` value is
 a valid multi-sentence CoNLL-U document when a line crosses a sentence

data/prose_1/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d578bf9a1fe783c69403b6d1d3015921dac27957d1a6abdbc1b5028e09fc4d5
-size 1223992
+oid sha256:e1427fca5d4efd1a843db94fe2c0f9476d3fcfada5c23f8b6c67f26e460ff64f
+size 1225487

data/prose_1/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8c5246a4aec0602ff62e655a3652b0aa9ba12a575193d6ebfd61bebb2f6ca92b
-size 11035971
+oid sha256:5eeaaba94135ecaa8190832d95d0154d908c36a8e2aee440970c1ea06eb1dcb8
+size 11047428

data/prose_1/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:611ffa1da35622bf6e36cff4d49bc7571919cb7f9528fa47e8171cf4b9916e39
-size 1226640
+oid sha256:b13263866f6ae938bd029e01ad90d4ffde49a3ec59b07907765ea80f9e1eb44a
+size 1228039

data/prose_10/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7aa337c0f9aab44f08718e7d514da09221a36fd6cf807ab5e966a82d028a5e4f
-size 1069490
+oid sha256:aa6a26cfeeb182cfae24790508d959af62bf031750df08524d6f617a7cdbdd97
+size 1070435

data/prose_10/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a8bb1944a4f3d22918f49535faf965b9ff3af520671c5c248d4d1db1e3d2dfe9
-size 11942841
+oid sha256:a9219a49732f1ca2cfe1e1a05fb355e68cae2bb3ded12e125c84588c2dc54e32
+size 11954298

data/prose_10/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad06c18e4718abc0103c6735dfeb570fad550cc6dc7a0a69cff05337abd22443
-size 1064917
+oid sha256:437257da9e82ea9f1146edae9057c6ebb7ea8f0519b8a316bec8e42384193c31
+size 1065929

data/prose_100/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:db1e91b761ba6f91398390acdf3f9341b7d127404a0f27b94ed153afb34ab8e7
-size 1017367
+oid sha256:c6ad6df02726e31a1be2be4bdb605d05a5bbee6b043d66211ff7eaef729115d7
+size 1017901

data/prose_100/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b7f0484ee95d436b0814b79419124cd3e8d0564c323b605b860de931564de6ad
-size 11942841
+oid sha256:c49570a2134db2995f410af8adf95f66fa50804f48b31dce7f13aef7dcf1888f
+size 11954298

data/prose_100/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:86afce4e0b6fd70b737f6dc507e940ab4bdf647dc976d836f91448700ba05ab9
-size 1018687
+oid sha256:46fbe75f312ca73a61056ed0a31acca8a3baa8cfac39740e4bd822f82bf83b9c
+size 1019329

data/verse_metre_1/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dd82437c2f1006c1f3e53ef2c1b777ec555654df52a932fbadb42dfa72ba0ccf
-size 1130983
+oid sha256:a1d00311cf5c03f8792854fe8c05d6a821833c281d8818a637dcfc1bd16b1bdf
+size 1131885

data/verse_metre_1/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:130a613b1111dbd68f2867ae7b9074b86901b1d4f22ded96b1fa4080ecac2c6d
-size 6041479
+oid sha256:2fae60f1ee2c79f3bb269e1e49eabf9c687591ec0deb89936194d1b17e05f455
+size 6048972

data/verse_metre_1/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:03df61d36ae0594b2fc4cc29868e31ccc0e1224d9b6ba2e821c434a4af73dfd2
-size 1141584
+oid sha256:15194b38cdce47ed7835fffaab53793651ac70ed0655e2ba63b57746ed81d7e3
+size 1142469

data/verse_metre_10/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a1399c7f64c54d53ff4d48f8c4db6480b49cd858e063561fe6fdbd9b706b4c69
-size 981467
+oid sha256:9857f0fc118df03093ccb481041bab1974eb7f5651abde10a2343d16a79f3fdd
+size 982348

data/verse_metre_10/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ad8c0298be774a8fba3bf3669c229e46d26b312a1d3fecb1556dac4e99116fd
-size 6772982
+oid sha256:a7d3f546988a9d8125fc51b5944fd0cbdaf37e5a5c503f655734d19a597550aa
+size 6780475

data/verse_metre_10/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f749520aad6612f42dbc2e2e34f8dd62af1a38f621fe96d25dcdfd0aef377337
-size 986891
+oid sha256:7f88672665bab7fcfa8bd41a8be85c1124cd58dc4ec07f7a6ac8f7e927091b2f
+size 987833

data/verse_metre_100/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c67c87ce3de759069fde5fa064145a4ba979bd12245faf9777b04b0a924aee91
-size 944200
+oid sha256:5e9258f3211faee3a51ed68e2e0681fc13ff56de608545ebf9be6cafbe144812
+size 944981

data/verse_metre_100/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d1a6c23a4eadf5fde6dd6f567a5d9d62f28abb6b7ebe696d285f198486ef6292
-size 6772982
+oid sha256:ff14d9c5c24e63abf9cf3e8c7ce259e296d63762be62b7915bc309c29cda9a29
+size 6780475

data/verse_metre_100/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2f65d9d8439b0492a82bd2a58eaaea996a2c3f1ff71d5e6d8cd1434ce355dbdf
-size 949898
+oid sha256:236bca4e8c7a0fde7e2aaa1b63257b2e39e258e3c5ea7eec77d4260223ee357a
+size 950741

data/verse_sentence_1/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:33c74306d623dfdfd9bd663430f0652c9ba2d531410f6e4359a3d391274bb847
-size 751790
+oid sha256:c37d439c0003c05757d3779595f9b57849d42879ba01d8b0089248a5f676eff4
+size 752329

data/verse_sentence_1/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:641bfec6257bc75d7656e7ebfc8747b7a26d9a599bdde900113a2676f5b3504c
-size 5641344
+oid sha256:926a5fee9d1fda5795b0db9f7b8169c22da89b6a0d5b33a687f76c1fdf5294cc
+size 5646102

data/verse_sentence_1/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:06be815e7e3df175b173bb87f0aba21519e9ce597c8ee5d7942bb56dac27f10c
-size 746317
+oid sha256:a283d6f59cc42b085ae8ce6a5e6e02c85bc5a543ac58606b8828282a7ebaf1d0
+size 746879

data/verse_sentence_10/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ab3b5a87efaf751e041cb18f0551b08ef8addb2f7dfbd293fbef2c5eb486095f
-size 661803
+oid sha256:8dbb0751a353a9fe62f5c07fae3b019de644252ecd9204b0bde77ad7879e2218
+size 662250

data/verse_sentence_10/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:93e425258f1d2c2bf790ee563be6012670ca63d3dd7d018fe5fa9f071dff4f35
-size 6026463
+oid sha256:e01495b59bd2a23623dcd18bfef762dda08b068c5a5248c9394f9dce4a348867
+size 6031221

data/verse_sentence_10/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ff1b4116dc8f16d591e76b808db8347a124938d576ca78ffe6f3897001bd5f49
-size 657981
+oid sha256:5f3b24962f00be97e3b9f13871a465dab201521b08bfffc3e85108e23a7dffab
+size 658460

data/verse_sentence_100/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c100cbe9407f9593755b1a5351fa93bfcb207ff04d38d29d8a1a32517bf264f1
-size 635007
+oid sha256:615703b0fcac44a20d138165448be6bfa446def447956d1a712182df48460fb0
+size 635279

data/verse_sentence_100/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:93b8856b340b83c4642211fcdccec9689b11f09e20c4633a2668f0a9dafcb115
-size 6026463
+oid sha256:f2efb7b94f86308c96aa5719f9c76b26969c2c1521b8a547aaa3df575d8a13e9
+size 6031221

data/verse_sentence_100/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ee5134006ddf4d83fe7f581ef9f700c3366cacaec6a56d8018ca80934ed092b7
-size 629189
+oid sha256:7ae48b0a6d93a860785280aed8420500a4f829100207e5320f96000291a265d8
+size 629496

inspection/README.md

@@ -9,11 +9,18 @@ Each configuration is a table named after the configuration, for example
 text. `_split_row_index` records the original physical row order within each
 split. `_mirror_manifest` records row counts, column lists, and content hashes.
 
-For example, retrieve the first validation row as follows:
+The top-level `text` column is a JSON list: one string for atomic rows and one
+string per constituent sentence or line for evaluation chunks. For example,
+retrieve the ten separate sentences in the first validation row as follows:
 
 ```sh
 sqlite3 -readonly -line inspection/sphragis.sqlite \
-  "SELECT text FROM verse_sentence_10 WHERE split = 'validation' ORDER BY _split_row_index LIMIT 1;"
+  "SELECT key AS sentence_index, value AS sentence
+   FROM json_each((
+     SELECT text FROM verse_sentence_10
+     WHERE split = 'validation'
+     ORDER BY _split_row_index LIMIT 1
+   ));"
 ```
 
 List the available tables with:

inspection/sphragis.sqlite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2056764115e1320dea9c2525cda4b9d9997e48514c7029768f980542d8d2f016
-size 1256095744
+oid sha256:8f4d0e67fae064366d2e854f42e2cb69ac39be2600e2f3d97f19f2f0b29e72f6
+size 1257635840

metadata/build_report.json

@@ -539,49 +539,49 @@
   },
   "data_files": {
     "prose_1": {
+      "test": 2800,
       "train": 25730,
-      "validation": 2800,
-      "test": 2800
+      "validation": 2800
     },
     "prose_10": {
+      "test": 280,
       "train": 25730,
-      "validation": 280,
-      "test": 280
+      "validation": 280
     },
     "prose_100": {
+      "test": 28,
       "train": 25730,
-      "validation": 28,
-      "test": 28
-    },
-    "verse_sentence_1": {
-      "train": 11914,
-      "validation": 1400,
-      "test": 1400
-    },
-    "verse_sentence_10": {
-      "train": 11914,
-      "validation": 140,
-      "test": 140
-    },
-    "verse_sentence_100": {
-      "train": 11914,
-      "validation": 14,
-      "test": 14
+      "validation": 28
     },
     "verse_metre_1": {
+      "test": 2600,
       "train": 21818,
-      "validation": 2600,
-      "test": 2600
+      "validation": 2600
     },
     "verse_metre_10": {
+      "test": 260,
       "train": 21818,
-      "validation": 260,
-      "test": 260
+      "validation": 260
     },
     "verse_metre_100": {
+      "test": 26,
       "train": 21818,
-      "validation": 26,
-      "test": 26
+      "validation": 26
+    },
+    "verse_sentence_1": {
+      "test": 1400,
+      "train": 11914,
+      "validation": 1400
+    },
+    "verse_sentence_10": {
+      "test": 140,
+      "train": 11914,
+      "validation": 140
+    },
+    "verse_sentence_100": {
+      "test": 14,
+      "train": 11914,
+      "validation": 14
     }
   },
   "deduplication": {
@@ -938,150 +938,151 @@
     }
   },
   "validation": {
+    "checks": {
+      "complete_chunk_text_and_syllable_aggregation": true,
+      "fixed_exact_evaluation_chunks": true,
+      "identical_source_rows_across_task_sizes": true,
+      "identifier_free_conllu_comments": true,
+      "identifier_free_conllu_misc": true,
+      "json_list_text_fields": true,
+      "official_conll18_loader": true,
+      "privacy_safe_metrical_lines": true,
+      "prose_exact_text_unique": true,
+      "scansion_column_absent": true,
+      "source_verse_component_exact_coverage": true,
+      "unique_conllu_documents_checked": 50734,
+      "unique_ids": true
+    },
     "prose_1": {
-      "rows": 31330,
       "authors": 12,
-      "works": 39,
+      "rows": 31330,
       "sources": {
+        "agdt": 568,
         "gorman": 17977,
-        "ud_perseus": 4306,
         "pedalion": 2447,
-        "ud_proiel": 6032,
-        "agdt": 568
+        "ud_perseus": 4306,
+        "ud_proiel": 6032
       },
       "splits": {
+        "test": 2800,
         "train": 25730,
-        "validation": 2800,
-        "test": 2800
-      }
+        "validation": 2800
+      },
+      "works": 39
     },
     "prose_10": {
-      "rows": 26290,
       "authors": 12,
-      "works": 65,
+      "rows": 26290,
       "sources": {
+        "agdt": 479,
         "gorman": 15088,
-        "ud_perseus": 3642,
+        "multiple": 26,
         "pedalion": 2059,
-        "ud_proiel": 4996,
-        "agdt": 479,
-        "multiple": 26
+        "ud_perseus": 3642,
+        "ud_proiel": 4996
       },
       "splits": {
+        "test": 280,
         "train": 25730,
-        "validation": 280,
-        "test": 280
-      }
+        "validation": 280
+      },
+      "works": 65
     },
     "prose_100": {
-      "rows": 25786,
       "authors": 12,
-      "works": 53,
+      "rows": 25786,
       "sources": {
+        "agdt": 473,
         "gorman": 14806,
-        "ud_perseus": 3578,
+        "multiple": 16,
         "pedalion": 2022,
-        "ud_proiel": 4891,
-        "agdt": 473,
-        "multiple": 16
+        "ud_perseus": 3578,
+        "ud_proiel": 4891
       },
       "splits": {
+        "test": 28,
         "train": 25730,
-        "validation": 28,
-        "test": 28
-      }
+        "validation": 28
+      },
+      "works": 53
     },
-    "verse_sentence_1": {
-      "rows": 14714,
+    "verse_metre_1": {
       "authors": 2,
-      "works": 2,
+      "rows": 27018,
       "sources": {
-        "agdt": 14714
+        "agdt": 27018
       },
       "splits": {
-        "train": 11914,
-        "validation": 1400,
-        "test": 1400
-      }
+        "test": 2600,
+        "train": 21818,
+        "validation": 2600
+      },
+      "works": 2
     },
-    "verse_sentence_10": {
-      "rows": 12194,
+    "verse_metre_10": {
       "authors": 2,
-      "works": 2,
+      "rows": 22338,
       "sources": {
-        "agdt": 12194
+        "agdt": 22338
       },
       "splits": {
-        "train": 11914,
-        "validation": 140,
-        "test": 140
-      }
+        "test": 260,
+        "train": 21818,
+        "validation": 260
+      },
+      "works": 2
     },
-    "verse_sentence_100": {
-      "rows": 11942,
+    "verse_metre_100": {
       "authors": 2,
-      "works": 2,
+      "rows": 21870,
       "sources": {
-        "agdt": 11942
+        "agdt": 21870
       },
       "splits": {
-        "train": 11914,
-        "validation": 14,
-        "test": 14
-      }
+        "test": 26,
+        "train": 21818,
+        "validation": 26
+      },
+      "works": 2
     },
-    "verse_metre_1": {
-      "rows": 27018,
+    "verse_sentence_1": {
       "authors": 2,
-      "works": 2,
+      "rows": 14714,
       "sources": {
-        "agdt": 27018
+        "agdt": 14714
       },
       "splits": {
-        "train": 21818,
-        "validation": 2600,
-        "test": 2600
-      }
+        "test": 1400,
+        "train": 11914,
+        "validation": 1400
+      },
+      "works": 2
     },
-    "verse_metre_10": {
-      "rows": 22338,
+    "verse_sentence_10": {
       "authors": 2,
-      "works": 2,
+      "rows": 12194,
       "sources": {
-        "agdt": 22338
+        "agdt": 12194
       },
       "splits": {
-        "train": 21818,
-        "validation": 260,
-        "test": 260
-      }
+        "test": 140,
+        "train": 11914,
+        "validation": 140
+      },
+      "works": 2
     },
-    "verse_metre_100": {
-      "rows": 21870,
+    "verse_sentence_100": {
       "authors": 2,
-      "works": 2,
+      "rows": 11942,
       "sources": {
-        "agdt": 21870
+        "agdt": 11942
       },
       "splits": {
-        "train": 21818,
-        "validation": 26,
-        "test": 26
-      }
-    },
-    "checks": {
-      "fixed_exact_evaluation_chunks": true,
-      "complete_chunk_text_and_syllable_aggregation": true,
-      "identifier_free_conllu_comments": true,
-      "identifier_free_conllu_misc": true,
-      "identical_source_rows_across_task_sizes": true,
-      "official_conll18_loader": true,
-      "privacy_safe_metrical_lines": true,
-      "prose_exact_text_unique": true,
-      "scansion_column_absent": true,
-      "source_verse_component_exact_coverage": true,
-      "unique_conllu_documents_checked": 50734,
-      "unique_ids": true
+        "test": 14,
+        "train": 11914,
+        "validation": 14
+      },
+      "works": 2
     }
   }
 }

scripts/build_dataset.py

@@ -31,10 +31,12 @@ import pyarrow.parquet as pq
 try:
     from scripts.dataset_variants import make_dataset_variants, variant_schema
     from scripts.metrical_lines import load_public_metrical_lines, public_metrical_line
+    from scripts.text_units import load_text_units
     from scripts.vendor.conll18_ud_eval import UDError, load_conllu
 except ModuleNotFoundError:  # Direct execution from the scripts directory.
     from dataset_variants import make_dataset_variants, variant_schema
     from metrical_lines import load_public_metrical_lines, public_metrical_line
+    from text_units import load_text_units
     from vendor.conll18_ud_eval import UDError, load_conllu
 
 
@@ -1411,6 +1413,9 @@ def validate(rows_by_config: dict[str, list[dict]]) -> dict:
         base_config = config.rsplit("_", 1)[0]
         assert all(row["genre"] == base_config for row in rows)
         assert all(row["text"] and row["author"] and row["work"] for row in rows)
+        for row in rows:
+            text_units = load_text_units(row["text"])
+            assert len(text_units) == row.get("chunk_size", 1)
         assert all(row["split"] in {"train", "validation", "test"} for row in rows)
         suffix = config.rsplit("_", 1)[1]
         if suffix != "1":
@@ -1488,9 +1493,12 @@ def validate(rows_by_config: dict[str, list[dict]]) -> dict:
                     constituents = [
                         atomic_by_id[row_id] for row_id in row["constituent_ids"]
                     ]
-                    assert row["text"] == (
-                        "\n" if base_config == "verse_metre" else "\n\n"
-                    ).join(item["text"].strip() for item in constituents)
+                    expected_text_units = [
+                        unit
+                        for item in constituents
+                        for unit in load_text_units(item["text"])
+                    ]
+                    assert load_text_units(row["text"]) == expected_text_units
                     if base_config == "verse_metre":
                         expected_syllables = [
                             syllable
@@ -1510,6 +1518,7 @@ def validate(rows_by_config: dict[str, list[dict]]) -> dict:
         "identifier_free_conllu_comments": True,
         "identifier_free_conllu_misc": True,
         "privacy_safe_metrical_lines": True,
+        "json_list_text_fields": True,
     }
     return report
 

scripts/dataset_variants.py

@@ -12,8 +12,10 @@ import pyarrow as pa
 
 try:
     from scripts.metrical_lines import load_public_metrical_lines
+    from scripts.text_units import encode_text_units, source_text_units
 except ModuleNotFoundError:  # Direct execution from the scripts directory.
     from metrical_lines import load_public_metrical_lines
+    from text_units import encode_text_units, source_text_units
 
 
 BASE_CONFIGS = ("prose", "verse_sentence", "verse_metre")
@@ -146,6 +148,7 @@ def _provenance(row: dict) -> dict:
 def _chunk_metadata(row: dict, target: int) -> dict:
     row = dict(row)
     row.update({
+        "text": encode_text_units(source_text_units(row["text"])),
         "chunk_size": 1,
         "chunk_target_size": target,
         "constituent_ids": [row["id"]],
@@ -205,9 +208,9 @@ def _aggregate_chunk(base_config: str, rows: list[dict], target: int, split: str
         "id": f"chunk-{base_config}-{target}-{digest}",
         "work": works[0] if not mixed_work else "Multiple works",
         "work_id": work_ids[0] if not mixed_work else f"multiple:{digest}",
-        "text": ("\n" if base_config == "verse_metre" else "\n\n").join(
-            row["text"].strip() for row in rows
-        ),
+        "text": encode_text_units([
+            unit for row in rows for unit in source_text_units(row["text"])
+        ]),
         "conllu": "\n\n".join(row["conllu"].strip() for row in rows) + "\n\n",
         "cts_urn": rows[0]["cts_urn"] if len({row["cts_urn"] for row in rows}) == 1 else None,
         "passage": (
@@ -314,7 +317,13 @@ def make_dataset_variants(
         shared_rows, retained_authors, bottleneck_discarded = select_bottleneck_rows(
             base_config, rows,
         )
-        variants[f"{base_config}_1"] = shared_rows
+        variants[f"{base_config}_1"] = [
+            {
+                **row,
+                "text": encode_text_units(source_text_units(row["text"])),
+            }
+            for row in shared_rows
+        ]
         report[f"{base_config}_1"] = {
             "authors": len(retained_authors),
             "retained_authors": sorted(retained_authors),

scripts/text_units.py

@@ -0,0 +1,32 @@
+"""Canonical JSON representation for model-facing text units."""
+
+from __future__ import annotations
+
+import json
+
+
+def load_text_units(encoded: str) -> list[str]:
+    """Load a published text field and enforce a non-empty string list."""
+    units = json.loads(encoded)
+    if not isinstance(units, list) or not units:
+        raise ValueError("text must encode a non-empty JSON list")
+    if not all(isinstance(unit, str) and unit.strip() for unit in units):
+        raise ValueError("every text unit must be a non-empty string")
+    return units
+
+
+def source_text_units(value: str) -> list[str]:
+    """Accept either an internal atomic string or published JSON text units."""
+    try:
+        return load_text_units(value)
+    except (json.JSONDecodeError, TypeError, ValueError):
+        if not isinstance(value, str) or not value.strip():
+            raise ValueError("source text must be a non-empty string")
+        return [value]
+
+
+def encode_text_units(units: list[str]) -> str:
+    """Serialize text units without losing their original strings."""
+    if not units or not all(isinstance(unit, str) and unit.strip() for unit in units):
+        raise ValueError("text units must be non-empty strings")
+    return json.dumps(units, ensure_ascii=False)

scripts/validate_dataset.py

@@ -26,6 +26,7 @@ try:
         variant_schema,
     )
     from scripts.metrical_lines import load_public_metrical_lines
+    from scripts.text_units import load_text_units
 except ModuleNotFoundError:  # Direct execution from the scripts directory.
     from dataset_variants import (  # type: ignore[no-redef]
         BOTTLENECK_TARGET,
@@ -33,6 +34,7 @@ except ModuleNotFoundError:  # Direct execution from the scripts directory.
         variant_schema,
     )
     from metrical_lines import load_public_metrical_lines
+    from text_units import load_text_units
 
 
 BASE_CONFIGS = ("prose", "verse_sentence", "verse_metre")
@@ -221,7 +223,9 @@ def validate_metrical_line_privacy(root: Path) -> int:
 def cumulative_lengths(rows: list[dict]) -> list[int]:
     boundaries = [0]
     for row in rows:
-        boundaries.append(boundaries[-1] + len(normalize(row["text"])))
+        boundaries.append(boundaries[-1] + sum(
+            len(normalize(unit)) for unit in load_text_units(row["text"])
+        ))
     return boundaries
 
 
@@ -247,8 +251,16 @@ def validate_alignment_components(configs: dict[str, list[dict]]) -> int:
                 continue
             if any(not set(row["parent_sentence_ids"]) <= sentence_ids for row in lines):
                 continue
-            sentence_text = "".join(normalize(row["text"]) for row in sentences)
-            line_text = "".join(normalize(row["text"]) for row in lines)
+            sentence_text = "".join(
+                normalize(unit)
+                for row in sentences
+                for unit in load_text_units(row["text"])
+            )
+            line_text = "".join(
+                normalize(unit)
+                for row in lines
+                for unit in load_text_units(row["text"])
+            )
             assert sentence_text == line_text
             complete_components += 1
             sentence_bounds = cumulative_lengths(sentences)
@@ -318,9 +330,12 @@ def validate_variants(configs: dict[str, list[dict]]) -> None:
                     constituents = [atomic_by_id[row_id] for row_id in chunk["constituent_ids"]]
                     assert all(row["author"] == chunk["author"] for row in constituents)
                     assert all(row["split"] == split for row in constituents)
-                    assert chunk["text"] == (
-                        "\n" if base == "verse_metre" else "\n\n"
-                    ).join(row["text"].strip() for row in constituents)
+                    expected_text_units = [
+                        unit
+                        for row in constituents
+                        for unit in load_text_units(row["text"])
+                    ]
+                    assert load_text_units(chunk["text"]) == expected_text_units
                     if base == "verse_metre":
                         expected_syllables = [
                             syllable
@@ -492,6 +507,13 @@ def main() -> None:
         base_config = config.rsplit("_", 1)[0]
         assert all(row["genre"] == base_config for row in rows)
         assert all(row["text"] and row["author"] and row["work"] for row in rows)
+        for row in rows:
+            text_units = load_text_units(row["text"])
+            expected_units = row.get("chunk_size", 1)
+            assert len(text_units) == expected_units, (
+                f"text-unit count mismatch in {config}, row {row['id']}: "
+                f"{len(text_units)} != {expected_units}"
+            )
         if base_config == "prose":
             keys = [row["dedup_key"] for row in rows]
             assert len(keys) == len(set(keys)), "prose is not exactly deduplicated"

tests/test_split_stratification.py

@@ -9,6 +9,7 @@ from scripts.metrical_lines import (
     load_public_metrical_lines,
     sanitize_metrical_lines,
 )
+from scripts.text_units import load_text_units
 
 
 DATA_ROOT = Path(__file__).resolve().parents[1] / "data"
@@ -71,7 +72,7 @@ def test_all_task_sizes_cover_exactly_the_same_atomic_rows() -> None:
         for split in ("train", "validation", "test"):
             path = DATA_ROOT / f"{base_config}_1" / f"{split}-00000-of-00001.parquet"
             atomic_rows[split] = {
-                row["id"]: row["text"]
+                row["id"]: load_text_units(row["text"])
                 for row in pq.read_table(path, columns=["id", "text"]).to_pylist()
             }
 
@@ -93,17 +94,33 @@ def test_all_task_sizes_cover_exactly_the_same_atomic_rows() -> None:
                     threshold,
                     split,
                 )
-                separator = "\n" if base_config == "verse_metre" else "\n\n"
                 for chunk in chunks:
                     expected_text = (
                         atomic_rows[split][chunk["constituent_ids"][0]]
                         if split == "train"
-                        else separator.join(
-                            atomic_rows[split][row_id].strip()
+                        else [
+                            unit
                             for row_id in chunk["constituent_ids"]
-                        )
+                            for unit in atomic_rows[split][row_id]
+                        ]
                     )
-                    assert chunk["text"] == expected_text
+                    assert load_text_units(chunk["text"]) == expected_text
+
+
+def test_text_is_json_list_with_one_entry_per_constituent() -> None:
+    for base_config in BASE_CONFIGS:
+        for suffix in ("1", "10", "100"):
+            for split in ("train", "validation", "test"):
+                path = (
+                    DATA_ROOT / f"{base_config}_{suffix}"
+                    / f"{split}-00000-of-00001.parquet"
+                )
+                columns = ["text"]
+                if suffix != "1":
+                    columns.append("chunk_size")
+                for row in pq.read_table(path, columns=columns).to_pylist():
+                    expected = row.get("chunk_size", 1)
+                    assert len(load_text_units(row["text"])) == expected
 
 
 def test_verse_metre_chunks_concatenate_every_constituent_syllable() -> None:

tests/test_verse_character_coverage.py

@@ -3,6 +3,8 @@ import unicodedata
 
 import pyarrow.parquet as pq
 
+from scripts.text_units import load_text_units
+
 
 DATA_ROOT = Path(__file__).resolve().parents[1] / "data"
 
@@ -11,7 +13,8 @@ def normalized_character_count(config: str) -> int:
     count = 0
     for path in sorted((DATA_ROOT / config).glob("*.parquet")):
         texts = pq.read_table(path, columns=["text"])["text"].to_pylist()
-        for text in texts:
+        for encoded in texts:
+            text = "".join(load_text_units(encoded))
             decomposed = unicodedata.normalize(
                 "NFD", text.lower().replace("ς", "σ")
             )