fffoivos/greek-ocr-lines

Greek OCR lines (v0.5)

Printed text-line images with transcriptions: the exact lines used to train and evaluate the Greek OCR models (OCR v0.10, widths ×1 / ×1.5 / ×2). The lines are mostly polytonic and modern (monotonic) Greek, with Latin, Cyrillic and mixed Greek-Latin lines, numbers and page furniture.

Part of the Greek OCR collection. Code: fffoivos/greek-ocr-toolkit.

Configs and splits

Config / split	Lines	Role
real / train	793,227	real half of the training mixture (pools non_parliament 735,976 and parliament 57,251; see pool)
real / validation	13,088	training-time validation (loss curves)
real / harvest_dev	16,350	dev; with seal04_dev it selected the hyperparameters in Stage B
real / seal04_dev	437	dev; selected each released checkpoint (best of the 8 epoch-end evaluations)
real / harvest_test	52,847	test, report only; never used for any selection
real / seal04_test	2,253	test, report only; never used for any selection
synthetic / train	190,486	synthetic half of the training mixture

from datasets import load_dataset
real = load_dataset("fffoivos/greek-ocr-lines", "real", split="harvest_test")
synth = load_dataset("fffoivos/greek-ocr-lines", "synthetic", split="train")

Columns (both configs)

Column	Meaning
image	line image (PNG), exactly the payload the trainer read
gt_nfc / gt_nfd	transcription in NFC and NFD (the model predicts NFD internally)
label_ids, codec_version	the NFD transcription as codec v2.1 class ids
width	image width in pixels
script_class	greek_poly, greek_mono, greek_latin_cs, latin, cyrillic, furniture
pool, split	training pool and the original split tag
line_id	stable line id; for real lines it encodes harvest wave, source collection, document, page and line where available
source_dataset, source_doc_id	source collection and source document (book) id; for synthetic lines, where the rendered text came from
applied_ops	synthetic: the ordered degradation operations applied to this line (ink bleed, blur, noise, binarize, rotate, perspective, baseline wobble, crop variants…); real: the import resize step
severity, font, polarity, bg_type, accent_survival, reuse_index	synthetic generation record for the line (font = real for real lines)

Synthetic generator distribution: synthetic/RECIPE.json holds the configured distribution (content-class marginals, background weights, crop mix, fonts and held-out fonts, ink and length settings). synthetic/DISTRIBUTION.json holds the realised counts in these rows (per op, font, severity, background, polarity, script class, text source). The factory's own audit and selection reports are in synthetic/factory_*.json. SOURCES_BY_SPLIT.json counts real lines per split and source collection.

The licence column is intentionally removed (owner decision).

Status of v0.5

This is the first public upload: the per-line view of the exact training/evaluation files. Still being built, and added to main before the v0.5 tag:

• the relational tables: pages and page images, every line box with its position on the page, segmenter results, row decisions, label versions with corrections, features, training membership, book registry;
• source fields that resolve each line to its exact book: provider, title, year, page, catalogue/manifest/source URL;
• the original (un-resized) crops.

Selection roles above follow the release plan; seal04 dev/test are public by owner decision. Label-quality sample on the base harvest: 33 of 412 labels (8 %) do not match the print.