mschonhardt/consilia-baldus-venice1575-pagexml

Consilia: Baldus de Ubaldis, Venice 1575. PageXML ground truth for layout analysis and text recognition

PageXML for the 1,392 pages of the five volumes of Baldus de Ubaldis, Consiliorum sive responsorum volumen primum–quintum (Venice: Girolamo Polo, 1575), the largest printed collection of legal opinions of a medieval jurist. Every page carries its text regions, every one of the 194,654 text lines its polygon, baseline, structural type and the text of the Consilia edition. The data are the working basis of the Consilia portal and were made for two uses: training and evaluating layout and text recognition models for sixteenth-century Latin legal prints in roman and italic type, and reuse of the transcription with its coordinates.

Developed in the project Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL|Mainz), funded by HMWK Hessen, PI Prof. Dr. Michael Schonhardt. The kraken models trained on this dataset are published alongside it (see Related).

Volume 1, page 43 (test split): text lines coloured by type, their baselines, and the regions (header, page number, two columns).

Contents

path	what
Baldo_Cons_Print_Venice_1575_v1/ … _v5/	one folder per volume: <page>.xml (PAGE 2013 XML, image referenced by file name <page>.jpg; the images are not included, see Images)
lines.parquet	one row per text line: print, page, image, split, region_id, region_type, region_index, line_id, line_index, line_type, polygon, baseline (PAGE point strings), x0 y0 x1 y1, height, text, n_chars, readings_agree (0–3, see Text), human_decision, flags
pages.csv	one row per page: volume, page, image file name, size, split, counts, and the image source: image_url (the original file at the Internet Archive), image_sha1, image_bytes, iiif_manifest, ia_view (the page in the Internet Archive viewer)
fetch_images.py	downloads the page images next to the XML files and checks them against image_sha1 (Python standard library only)
splits/train.txt, val.txt, test.txt	the page lists of the three splits (paths relative to this folder)
stats.json	the counts quoted below
segmentation_example.jpg	the figure above

volume	pages	lines
v1	324	45,555
v2	268	37,443
v3	284	40,079
v4	236	32,645
v5	280	38,932
all	1,392	194,654 (194,570 with text; 1.96 M words, 11.8 M characters)

split	pages	lines
train	1,221	170,679
val	100	13,884
test	71	10,091

The split is by page (hash of volume and page id, 5 % test, 8 % validation), so no page contributes to two sets. The test pages were not used in training or model selection of the published kraken models.

Layout

Text regions (4,291) carry a structural type in custom="structure {type:…;}" and the matching PAGE @type:

region type	count	PAGE @type	what
column_1, column_2	1,376 / 1,375	paragraph	the two text columns, numbered from left to right
header	1,376	header	running title
page_number	160	page-number	the printed page number where it was annotated
footer	3	footer

Text lines carry their type the same way:

line type	lines	what
cons_body	177,686	the text of a consilium (the opinion itself, including the casus lines)
cons_argument	12,857	the numbered summaries (summaria) that precede each consilium
cons_title	2,515	the heading "CONSILIVM …"
header	1,512	running title
page_number	53	printed page number
catch-word	31	catchword at the foot of the page

Text

The text is the edition text of the Consilia portal, and based on its editorial conventions: long s as s, the ampersand as "et", ligatures resolved, ę kept, abbreviations marked with tildes or special signs expanded silently ("nõ" > "non", "ſnĩa" > "sententia"), abbreviations with a period in citations kept as printed ("ff. de uer. ob. l."), a marginal number as a prefix ("2 "), the structuring symbol in the text as "†". Visible hyphens at the line end as "¬". The character set has 99 characters, of which about 30 occur fewer than 20 times (left-over abbreviation glyphs and combining marks the normalisation did not catch).

The ground truth was proofread by three different models – PyLaia, kraken and a vision-language model (Gemini 3.8 Flash; Gemini 4 Argon for eight columns re-read on 29 October 2026) – and corrected based on their agreement. 17,210 lines were decided by hand or in a review with the images involving editor-in-the-loop. readings_agree in lines.parquet says how many of the three readings equal the final text after the edition's loose normalisation (no spaces, no punctuation, long s, ligatures, u/v, i/j):

readings_agree	lines
3	155,081
2	31,137
1	8,169
0	183

For training a recogniser, lines with readings_agree >= 2 or human_decision and without a flag (single_reading: only one reading exists; interpolated_geometry: the polygon was interpolated) are the safe choice: 191,250 lines. The published kraken recognition model was trained on that subset of the train pages as of 27 October 2026 (191,251 lines; eight columns were re-read afterwards, see Text).

Images

The page images are not part of this dataset; they are linked. They are JPEG, about 2,800 × 4,250 px at 300 dpi, the Google digitisation of the copy of the Universidad Complutense de Madrid (HathiTrust record 009264041), uploaded by the project to the Internet Archive (items baldus-consilia-venetiis-1575-vol1-ucm to -vol5-ucm, with IIIF manifests). The files there are the ones the PageXML was made on, byte for byte: image_url in pages.csv points to each original file, image_sha1 is its checksum, so the coordinates fit without rescaling. python fetch_images.py downloads them next to the XML files (about 5.2 GB; --split test or --print … for a part) and checks every file; it can be run again after an interruption.

The images are linked rather than copied because of their rights: the print is in the public domain, but the scans carry Google's watermark, and Google and HathiTrust ask that Google-digitised images be used for non-commercial purposes with attribution and not be re-hosted in bulk for other ends. The rights of the images are those of the digitisation, not of this dataset's licence.

Use

kraken (7.x) reads the folders directly once the images are there:

python fetch_images.py                     # the page images next to the XML files
# segmentation: fine-tune blla on the edition's line and region classes
ketos -d cuda:0 segtrain -f page --resize new -i blla.mlmodel -t splits/train.txt -e splits/val.txt -o seg
# recognition: compile the lines once, then train (the published model started from McCATMuS)
ketos compile -f page -F splits/train.txt -o train.arrow
ketos compile -f page -F splits/val.txt   -o val.arrow
ketos -d cuda:0 train -f binary -t train.lst -e val.lst -i McCATMuS_nfd_nofix_V1.mlmodel --resize union -u NFD -B 32 -o rec

(splits/*.txt list the XML files relative to the dataset root; *.lst are one-line manifests naming the arrow files.) eScriptorium and Transkribus import the XML with the downloaded images. In Python:

import pandas as pd
lines = pd.read_parquet("lines.parquet")
body = lines[(lines.line_type == "cons_body") & (lines.readings_agree >= 2)]

Licence and citation

Schonhardt, Michael: Consilia: Baldus de Ubaldis, Venice 1575. PageXML ground truth for layout analysis and text recognition, v1. Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL|Mainz), 2026. https://huggingface.co/datasets/mschonhardt/consilia-baldus-venice1575-pagexml

Related

• The edition: Consilia. A portal for medieval legal opinions, https://dekm-tud.github.io/coding-consilia/ ; code and data: https://github.com/dekm-tud/coding-consilia (this dataset is built by tools/make_pagexml_dataset.py from the repository's data/)
• kraken segmentation model trained on this dataset: mschonhardt/consilia-baldus-venice1575-kraken-seg
• kraken recognition model trained on this dataset: mschonhardt/consilia-baldus-venice1575-kraken-rec
• Base models: kraken blla (segmentation), McCATMuS (recognition, DOI 10.5281/zenodo.13788177)