Labels pair one refined line crop (training/line_proposals_v2.json, images in
training/proposals_v2/) with the exact text visible in it. Batches go in
training/labels/batch_NNN_<row>.json and are merged with
pipeline/merge_labels.py. See that script’s docstring for the format.
pipeline/refine_lines.py extends each cell line so words that run across a
printed column rule stay whole. Ink belonging to neighbouring cells is painted with
the recorded paper colour (mask_fill); the export verifies that only those pixels
differ from the scan. A crop never contains half a glyph, but it can contain:
note.Reject (rejected_crop) strips with only rules, flourishes, junk, two merged text
lines, or a glyph visibly cut by the top/bottom edge in a way that hides its identity.
=.три-, -бли, Прихода. -).12го, 23го. Underlining is not transcribed.1888. if dotted.image_checked: every character read with confidence. Exported for training and
evaluation. One reader only: this is not independent verification.draft: any doubt. Mark it inline: [Г?]елена, [Бекешіовъ?]. Never exported. Names,
villages and townships are usually drafts until a second source confirms them.baseline_seen). Treat
it as a hint only. It modernizes spelling (drops ъ, writes е for ѣ).Baptism: Тысяча восемьсотъ восемьдесятъ восьмаго года [месяц] [день] дня, въ
Плунгянскомъ Р.К. Приходскомъ Костелѣ Викарнымъ Кс: [priest] окрещенъ младенецъ
именемъ [NAME] со всѣми обрядами Таинства.
Parents: Крестьянъ [township] волости [father] и [mother] изъ [maiden] [surname]
законныхъ супруговъ сынъ/дочь родившійся/родившаяся 1888 года [month] [day] дня въ
деревнѣ [village] сего Прихода.
Godparents: Воспріемниками были [name surname] съ [name surname] [status].
The clerk’s actual wording varies (e.g. деревни, Яна и Францишки). Transcribe what is
written, never what the formula predicts.