biglam/on_the_books

On the Books: Jim Crow and Algorithms of Resistance — Training Set

This dataset is the labeled training set from On the Books, a collections-as-data project at UNC Chapel Hill Libraries that used machine learning to identify Jim Crow laws within North Carolina session laws passed between Reconstruction and the Civil Rights era (1866–1967).

Each row is a chapter/section pair from a North Carolina session law, labeled by experts as either a Jim Crow law or not. The set was used to train a supervised classifier that was then applied to the full ~century of session laws.

• Project site: https://onthebooks.lib.unc.edu
• Source code: https://github.com/UNC-Libraries-data/OnTheBooks
• Full corpus & ML-identified subset (DOI): https://doi.org/10.17615/5c4g-sd44
• Original creators: UNC Chapel Hill Libraries (project team incl. Frank Baumgartner, Megan Winget, Hannah Jacobs)
• Hub upload: BigLAM

Dataset Summary

• Task: binary text classification (jim_crow vs. no_jim_crow)
• Domain: U.S. state legislation, North Carolina, 1866–1967
• Size: 1,785 rows (single train split) — 512 positive (jim_crow) / 1,273 negative
• Language: English (legal/legislative register, with period orthography from OCR)

Data Fields

Field	Type	Description
id	string	Identifier for the chapter/section item (e.g. 1911_public laws_42_2).
source	string	Label provenance — see Label Sources below.
jim_crow	ClassLabel (2)	0 = no_jim_crow, 1 = jim_crow.
type	string	Law type: session laws, public laws, private laws, public local laws.
chapter_num	int32	Chapter number within the session (1–1,460).
section_num	int32	Section number within the chapter (1–1,100).
chapter_text	string	Full OCR'd text of the chapter.
section_text	string	Full OCR'd text of the section that the label applies to.

Label Sources

The source field records which expert-curated reference list each labeled item came from. Three values appear:

source	Rows	jim_crow=1	jim_crow=0
project experts	1,673	403	1,270
paschal	74	74	0
murray	38	35	3

• murray — laws drawn from Pauli Murray's States' Laws on Race and Color (1951), the landmark compilation of U.S. segregation laws.
• paschal — laws drawn from a Paschal-attributed compilation of North Carolina segregation laws.
• project experts — items labeled directly by the On the Books project team (the bulk of the set, including most of the negative class).

The exact bibliographic reference for paschal is not documented in this upload. Users who need to cite provenance precisely should consult the On the Books project documentation and the GitHub repository.

Law-type Distribution

type	Rows
public laws	749
private laws	527
public local laws	304
session laws	205

How the Data Was Created

1. Digitization & OCR. Session-law volumes were OCR'd from Internet Archive scans using Tesseract via Python scripts in the project repo.
2. Segmentation. OCR output was segmented into chapters and sections.
3. Expert labeling. Items appearing in named, expert-compiled lists of Jim Crow laws (Murray, Paschal) were labeled jim_crow; the project team labeled additional positives and the bulk of the negatives.
4. Training. The labeled set was used to train a supervised classifier, which was then applied to the full corpus to flag laws likely to be Jim Crow. Both the full corpus and the ML-identified subset are deposited under DOI 10.17615/5c4g-sd44.

This Hub upload is the training set only, not the full ~century corpus or the model's predictions over it.

Intended Uses

• Training and evaluating text classifiers for historical legal language.
• Benchmarks for OCR-tolerant classification on period legal text.
• Research on segregation-era legislation, computational legal history, and collections-as-data methods.
• Teaching materials for digital humanities, library/archives data science, and ML-for-cultural-heritage courses.

Out-of-Scope / Limitations

• Jurisdiction. North Carolina only. Patterns will not transfer cleanly to other states without adaptation.
• Period coverage. 1866–1967. Modern legal text differs substantially.
• OCR noise. Texts contain OCR errors typical of late-19th / early-20th-century print; tokenisation and downstream metrics should account for this.
• Label coverage. "Jim Crow" labels reflect what was identified by named expert sources or by project staff. Laws with discriminatory effect that those sources did not catalogue may appear as no_jim_crow. Treat the negative class as "not flagged by the project's labeling process," not "verified non-discriminatory."
• Class imbalance. Roughly 29% positive — adjust loss / sampling accordingly.
• Source × label imbalance. paschal is 100% positive and murray is 92% positive; using source as a feature would leak the label.

Ethical Considerations

The dataset documents racist legislation. The texts include slurs and dehumanising language that appear verbatim in the historical record. The project's framing — Algorithms of Resistance — is that surfacing this material is part of confronting it; downstream users should preserve that framing in any derived work, dashboards, or models, and should not strip the historical context.

Licensing

• Dataset license: CC BY 3.0 (as per Hub metadata).
• Source code (UNC repo): GPL-3.0.
• The underlying session laws are public records.

Citation

Please cite the original project. A general-purpose form:

On the Books: Jim Crow and Algorithms of Resistance.
University of North Carolina at Chapel Hill Libraries.
https://onthebooks.lib.unc.edu
DOI: https://doi.org/10.17615/5c4g-sd44

For the data deposit specifically, use the DOI above; for code, cite the GitHub repository UNC-Libraries-data/OnTheBooks.

Acknowledgements

All credit for corpus construction, labeling, and methodology belongs to the On the Books project team at UNC Chapel Hill Libraries. This Hub upload by BigLAM is a redistribution to make the training set conveniently loadable via datasets.