1. Overview
Repository location
Revised dataset version DOI: https://doi.org/10.5281/zenodo.22695369.
Dataset concept DOI: https://doi.org/10.5281/zenodo.21795493.
Repository: https://github.com/gsgalezowski/quantum_archival.
Context
The corpus is a small, collection-specific pilot, not a universal handwritten-text recognition (HTR) or named-entity recognition (NER) benchmark. It tests auditable weak token labelling, matched-budget radial basis function (RBF) and simulated quantum fidelity kernels, and documentation needs before scaling. Its size permits page-level inspection and error tracing.
The source is Akta stanu cywilnego Parafii Rzymskokatolickiej w Kraśniku, unit 35/1850/0/2.1/24 (1917–1921). The set mainly contains 1918 birth acts, plus front, index, and cross-reference material. Each act is formulaic: a declarant, usually the father or a midwife, registers a child before the Kraśnik civil registrar with named witnesses, recording the child’s and parents’ names, the birth place and date, and the parties’ occupations or roles. These recurring slots are what the weak PERSON, PLACE, MONTH, NUMBER, and ROLE categories target; isolated-token labels do not encode the relations between them, such as which place is the birthplace or who acts as registrar.
Shared HTR benchmarks provide documented datasets and evaluation protocols for comparing recognition systems (Mühlberger et al., 2019; Sánchez et al., 2019). Historical NER faces transcription noise, spelling variation, and scarce annotation; HIPE-2022 provides contextual multilingual NER and linking beyond these token labels (Ehrmann et al., 2022, 2023). Garzoni and the Sydney Stock Exchange illustrate richer structured resources (Dan & Sherratt, 2023; Ehrmann et al., 2024), while RISE shows the value of small verified humanities datasets (Hindermann et al., 2026). Computational archival science supplies the provenance perspective (Hedges et al., 2022).
2. Method
Steps
The author transcribed, corrected, and checked all 100 TXT files against the scans with a Transkribus/PyLaia-assisted workflow. PAGE-XML is a parallel export, not an independent reference; documentation records the serialisation differences. Only the archival unit is linked; scans are not deposited.
Scripts use gazetteers and heuristics to propose six categories. PERSON covers names; PLACE covers toponyms without semantic relations; MONTH and NUMBER cover date and numerical forms; ROLE covers selected occupational, social, religious, or administrative terms; OTHER is residual, not an entity type. Without contextual spans or BIO (beginning-inside-outside) boundaries, this is weak token-category classification, not full NER.
Character n-grams feed illustrative classifiers. The deposited file pilot2_archival_text/results/gold_qsvm_validation_results.json reports matched-budget RBF and simulated quantum-kernel comparisons using four or eight principal component analysis (PCA)-reduced features; a tuned 64-feature RBF is the fuller classical control. Quantum similarity is squared state fidelity (Havlíček et al., 2019), simulated exactly on classical hardware. Training used 588 balanced weak-label rows after excluding audited (page, token) pairs; testing used 480 reviewed rows. In the 480-row audit evaluation summarised here, accuracy was 94.2% for the tuned 64-feature RBF and 90.8% for the selected eight-feature simulated quantum configuration; exact-pair deduplication yielded 93.8% and 90.2%. Of 480 test rows, 382 reuse a case-folded training form. On 98 unseen-form rows, accuracy fell to 74.5% and 58.2%. Scores therefore chiefly reflect formulaic lexical reuse. The McNemar result (p = 0.005) is descriptive because the audit was also used to select the configuration and the rows are dependent. No quantum advantage or speed-up is claimed (Schnabel & Roth, 2025); statevector memory and kernel matrices limit scaling.
Sampling strategy
The deposit is the 100-page set supplied for this pilot. With no recorded probability-sampling protocol, it is purposive, not representative, and supports method design rather than population estimates. The audit has 80 proposed candidates per category and is candidate-stratified, not prevalence-weighted.
Quality control
The 480-row token audit records 469 agreements and 11 corrections (97.71%), representing 450 exact page-token pairs (449 after case folding). Visible labels, one annotator, repeats, and candidate-stratified sampling preclude independent-agreement or corpus-wide-recall claims. Public PAGE-XML passes the official 2013-07-15 XSD after documented export sanitisation. An earlier exploratory experiment used 144 word-image crops from the archival scans, but neither the crops nor their verification worklist is deposited; its retained code and result files document the project development only and are not treated as evidence for this paper’s claims. Checksums, manifest, dictionary, scripts, and changelog support auditing.
3. Dataset Description
Repository name
Zenodo, with a public development repository on GitHub.
Object name
Manually Verified Polish Civil-Register Transcriptions and Weak Token Labels, version 1.0.0.
Format names and versions
The deposit contains 100 UTF-8 TXT and 100 PAGE-XML 2013-07-15 files, CSV labels/manifests, JSON results, Markdown documentation, Python 3 scripts and dependencies. Charts are PNG/PDF. Source JPGs and raster excerpts are excluded.
Creation dates
PAGE-XML metadata record creation on 2025-04-28 and changes through 2026-06-16. The research package was assembled from 2026-07-03 and revised for version 1.0.0 through 2026-09-10. The archival records date from 1917–1921.
Dataset creators
Grzegorz Gałęzowski created and verified the transcriptions, annotations, software, validation, documentation, and release package.
Language
The records and transcriptions are Polish; code and documentation are primarily English, with additional Polish instructions.
License
Author-created transcriptions, labels, results, and documentation are CC BY 4.0; code is MIT. Source images remain governed by archival rights statements and are not relicensed or redistributed.
Publication date
2026-09-10
4. Reuse Potential
Within the digital humanities, the corpus is reusable not because it represents Polish civil registers at scale, but because it exposes an inspectable chain from archival provenance and transcription exports to weak labels, audit decisions, and baseline results. It therefore connects HTR, historical NER, and computational archival science perspectives introduced in the Context section: users can examine how transcription variation, formulaic language, metadata choices, and sampling assumptions propagate into computational outputs. The package can also serve as a small, transparent reference example when planning richer contextual annotation or structured resources, without treating its scores as general benchmarks.
A computational archivist with a small register batch can inspect TXT/PAGE-XML alignment, audit weak labels, and decide whether a low-cost classical baseline is sufficient before commissioning contextual annotation. Here, the tuned 64-feature RBF using the full feature representation is the practical default: it avoids the added cost of statevector simulation and performed better. The simulated quantum configurations remain useful for method comparison and feasibility exercises because the workflow runs on ordinary hardware. A historian planning a study of mobility or witness networks can use the files to design a sampling strategy and data schema, but must add independently validated contextual spans and relations: the present categories do not identify, for example, which place is a birthplace or which person is a witness.
For teaching, the release supports a staged practical exercise suitable for digital humanities, archival studies, or introductory text-analysis courses. Students can first compare plain TXT with layout-aware PAGE-XML and trace each file to its archival provenance; next, inspect ambiguous weak labels and the documented audit corrections; and finally, rerun the classical and simulated quantum-kernel examples while discussing accuracy, computational cost, lexical leakage, and responsible interpretation. Instructors can use the deliberately limited scale to make every processing stage reviewable within a class rather than presenting a large benchmark as a black box. Purposive sampling, lexical repetition, excluded scans, and the absence of an independent gold corpus still limit inference beyond this collection.
Ethics and Consent
The study uses publicly accessible historical records and recruited no participants. The deposit includes no source images or other third-party media, so it seeks no licence to redistribute them. It provides archival provenance and leaves access to and reuse of source images subject to the current item-level statements in Szukaj w Archiwach. Users should assess ethical and legal implications before linking names to contemporary records or living people.
AI Declaration
OpenAI ChatGPT/Codex and Anthropic Claude assisted during revision with code revision and review, deterministic validation runs, literature discovery, release packaging, and editorial restructuring, including the response to reviewers. They did not originate archival transcription content, annotations, or images; one empty TXT export was restored from the author’s verified PAGE-XML. The author reviewed the assisted outputs and accepts full responsibility for the submitted text, data, code, citations, and claims.
Data Accessibility Statement
The revised dataset is archived at https://doi.org/10.5281/zenodo.22695369; the release family is at https://doi.org/10.5281/zenodo.21795493 and the matching Git tag is v1.0.0. The manifest links the source unit in Szukaj w Archiwach. Source scans and word-image crops are not redistributed.
Author Contributions
Grzegorz Gałęzowski: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review and editing.