Ericu950/swedish-print-ocr-training-data

Swedish Print OCR — training and evaluation data

Training mixture and benchmarks for Ericu950/swedish-print-ocr-3b (production) and Ericu950/swedish-print-ocr-3b-benchmark (held-out evaluation model). Each training row is a chat-format example:

{"messages": [{"role": "user", "content": "<image>Transkribera sidans brödtext normaliserat: ..."},
              {"role": "assistant", "content": "..."}],
 "images": ["<absolute path on the training cluster>"]}

Images are included under images/ as tar shards of roughly 10 GB each (337,115 files, ~65 GB in total — every image referenced by the training, validation and test JSONLs). Inside the tars, paths are relative to the training project root: a JSONL row's absolute path /nobackup/proj/flash/dionysus/shared/litteraturbanken/OCR_modell/external/runeberg/img/x.png corresponds to the tar member external/runeberg/img/x.png. images/images_manifest.json lists the shards and their file counts. Unpack all shards into one directory and strip the prefix from the JSONL paths (or point the prefix at your unpack directory) to train.

Files

File	What it is
train/train_allt.jsonl	419,999 examples — the production model's training mixture (held-out works included)
train/train_rent.jsonl	398,356 examples — the evaluation model's training mixture (held-out works excluded; original had 398,310 rows, see below)
val/train_allt_val.jsonl	the 402-page validation set used during training (original file)
val/valset_v2.jsonl	work-disjoint validation pages (original file)
test/testset_v2.jsonl	the principal benchmark: 1,182 pages from 214 works, 1600–1999 (original file)
test/wikisource_v3.jsonl	the external benchmark: 542 proofread Swedish Wikisource pages (original file)
held_out_works_v2.json	the list of 338 held-out works defining the train/test split
images/images_*.tar	all 337,115 page/line images referenced above (~65 GB, see Images)
train/train_blandat.jsonl	page + line-crop augmentation mixture (598,356 rows; line rows reference a page image plus a polygon to crop)
lines/radpar.jsonl	4,358,192 segment-to-ground-truth line pairs used to mine line training data

Provenance: original vs. reconstructed

The evaluation and validation files are the original files used in the study. The two training JSONLs are deterministic reconstructions: the exact files were deleted in a disk-quota cleanup, then rebuilt with the same scripts, seeds and source pools. The reconstruction machinery was verified byte-exact against a surviving validation file, and train_allt matches every number in the original build log (example counts per group, typeface census, language distribution). One caveat is stated for honesty: the instruction line optionally names the century of publication, and this is omitted at random from one third of examples; the surviving evidence shows the random omission pattern of the original files differed from the reconstruction's (same seed, slightly different code path at the time). The example pool, transcriptions, image assignments and mixture proportions are identical; which individual rows omit the century line is not guaranteed to be. For train_rent the same drift shifts which examples enter the 400,000-example sample: the reconstruction has 398,356 rows against the original build log's 398,310 (a 0.012% difference), while the cleaning filter itself reproduces the original exactly (154,862 pages scored, 1,982 below threshold).

Sources

The mixture draws on: Litteraturbanken (normalised page transcriptions), Projekt Runeberg (proofread pages), NewsEye Swedish newspapers (lines), GT4HistOCR (German blackletter lines, diplomatic), a Swedish fraktur corpus (lines, diplomatic), and Språkbanken regions. 34% Swedish roman print, 44% Swedish blackletter, 22% German blackletter. Each source retains its own licence; this compilation is distributed gated for research use. Consult the individual source collections before any redistribution or commercial use.