RevolutionCrossroads/RevX_OCR_results

Dataset Card for Revolution Crossroads OCR Results

Dataset Summary

This dataset contains machine-generated optical character recognition (OCR) results from materials brought together for Revolution Crossroads: Smithsonian Institution object images, National Archives and Records Administration (NARA) Revolutionary War pension files, and Library of Congress Chronicling America newspapers.

It provides three representations of the processed text: individual text blocks with layout information, text grouped by page, and text grouped by source file. The results were generated using Chandra OCR 2, a vision language model (VLM) for text extraction and document understanding, and transformed into tabular data to support searching, analysis, and further processing across the collections.

These are automated processing outputs, not verified transcriptions or replacements for the institutions' source records. They may contain missing, misread, or generated text.

The dataset was prepared as part of the Revolution Crossroads project: https://www.si.edu/revolution-crossroads

Dataset Description

The Revolution Crossroads project team prepared this dataset from AI-assisted OCR of images associated with three collection datasets. It brings the OCR results into a common structure while retaining source filenames, collection labels, and, at block level, text and layout information.

NOTE: This dataset has been compiled using carefully selected AI and machine learning tools and is derived from thousands of historic object records and digitized images. Evaluation and refinement of this data is ongoing, and the dataset will be updated accordingly. Users who encounter errors in this dataset should record them under the Community Tab to be logged for future remediation. More detail about the nature of this data is provided under Quality Considerations. This dataset is intended to expand access to primary source materials, not to replace them, and users are encouraged to consult source collections wherever possible. More information about the limitations of this dataset is provided under Risks and Limitations. Guidelines for using and citing this data in research, publications, and presentations are provided under Licensing Information and Citation Information.

Dataset Details

• Prepared by: Smithsonian Institution, Office of Digital & Innovation
• Shared by: Revolution Crossroads
• OCR service: Datalab, Chandra OCR 2
• Language: Primarily English; extracted text may contain other languages
• License: CC0 1.0 Universal

Configuration	What one row represents	Rows
block_level	One extracted text block	11,849,534
page_level	One processed page	1,367,839
file_level	One source file	65,423

These counts describe three representations of the processed material. They should not be added together as a count of distinct source records.

Relationship to Source Datasets

The processed material comes from three Revolution Crossroads source collections:

• Smithsonian Revolutionary Era Collections
• NARA Revolutionary War Pension Files
• Library of Congress Chronicling America, 1770–1810

For collection scope, original records, metadata definitions, provenance, and rights, refer to those source dataset cards. This card describes the OCR outputs and their transformation into block-, page-, and file-level tables.

Document-level representations are also available for NARA pension files and Chronicling America newspaper issues. Those dataset cards describe how the source pages were assembled into files for processing.

Page- and file-level OCR from this dataset was used to add RevX_OCR_text fields to the collection datasets. Complete processing artifacts are available through the Revolution Crossroads VLM OCR bucket.

Block-level text also supplied an input to the Revolution Crossroads raw named entities dataset. The entity dataset card describes that subsequent processing, including its use of Smithsonian catalogue text in addition to OCR.

Curation Rationale

The collections include printed newspapers, handwritten pension documents, and text on museum objects. A shared OCR table makes the processed text available for searching and analysis across those different materials.

The three configurations support different uses. Block-level data preserves the extracted text and its layout context. Page-level data provides text for individual images or document pages. File-level data brings page text together for work with complete source files.

Dataset Creation

Source images were prepared for processing by resizing where needed, converting formats, and assembling pages into multi-page PDFs. Pension pages were grouped into pension files, and newspaper pages into issues. Smithsonian object images were also converted into processing files as needed. Image preparation was limited to these steps; it did not include correcting or enhancing the source content.

The prepared files were processed using Datalab’s Chandra OCR models. For further information about the source images and PDF assembly, refer to the source dataset cards linked above.

Datalab's per-document chunk JSON supplied the text blocks, block types, page numbers, formatted text, and bounding boxes. The project flattened those outputs into rows and extracted plain text from each block's HTML. It then grouped the block text into page-level and file-level tables.

The tabular conversion did not create a separate set of layout blocks directly from the images. It used the blocks returned in Datalab's outputs. Extracting plain text and grouping it by page or file changed its representation; these steps did not manually correct its content.

Processing Runs

OCR processing was completed in two rounds during 2026:

Round	Processing dates	Model and version	OCR_run_ID
1	April 2026	Chandra 2	VLM_OCR_1
2	September 2026	Chandra 2.3	VLM_OCR_2

The OCR_run_ID field identifies the processing run associated with each row.

Notes on Completeness

This dataset contains the material represented in the available OCR outputs, not all material in the three source collections. Source dataset record counts and OCR table counts therefore need not match. The source cards provide further information about collection scope and OCR coverage.

A page or file appearing in the table does not establish that every word or region was captured. OCR can omit content even when it returns text for a page. Blank pages, processing failures, and unprocessed images should not be inferred solely from the number of rows in these tables.

Data Collection and Processing

Processing Steps

1. Datalab processed source images and PDFs and returned per-document chunk JSON.
2. The chunk outputs were flattened into one row per block, retaining the filename, collection label, page number, block identifier, type, Markdown, HTML, and bounding box.
3. Plain text was extracted from block_html using an HTML parser (selectolax) and stored as block_text. Formatting information, model comments, and other non-text information was removed.
4. Blocks were grouped by filename, collection, and page number. Their text was retained as a list and joined with newlines to create page-level text, skipping blank entries in the joined text.
5. Pages were grouped by filename and collection, ordered by page number, and stored as a list of page texts at file level.
6. The resulting tables were published as separate configurations. Page- and file-level text was also joined to the collection datasets as RevX_OCR_text.

Supporting Files Available

Directory	Contents
block_level/train-*.parquet	Sharded block-level table used by the block_level configuration.
page_level/train-*.parquet	Sharded page-level table used by the page_level configuration.
file_level/train-*.parquet	Sharded file-level table used by the file_level configuration.

For complete per-document processing artifacts, including additional layout and quality information, refer to the Revolution Crossroads VLM OCR bucket and its README. These tables contain selected fields from those outputs.

Accessing a Specific Data Version

This dataset may be updated as more OCR results are added or processing is revised. Use a specific tagged version to keep an analysis reproducible.

from datasets import load_dataset

dataset_name = "RevolutionCrossroads/RevX_OCR_results"
revision = "new_OCR_v2"

pages = load_dataset(
    dataset_name,
    "page_level",
    revision=revision,
    split="train",
)

Use block_level or file_level in place of page_level to load another representation.

Available Data Versions

• new_OCR_v1: Initial release with OCR from Round 1, processed in April 2026 using Chandra 2.
• new_OCR_v2: Updated release incorporating Round 2 OCR, processed in September 2026 using Chandra 2.3.

Quality Considerations

These are machine-generated OCR outputs. The published text has not been manually corrected throughout the dataset and should not be treated as a verified transcription. AI-assisted OCR can misread characters and names, omit content, repeat text, or generate text that is not present in the image. Dense newspaper layouts, handwriting, damaged scans, and text on objects can pose different difficulties.

Project OCR evaluation has included visual inspection and comparison with human transcriptions where available. Those checks help identify patterns of error; they do not establish the accuracy of every row or amount to a complete evaluation of all three configurations.

The Markdown, HTML, and plain-text fields are representations of the same OCR output. Agreement between them is not independent confirmation of a reading. Similarly, a bounding box locates a region identified during processing but does not establish that all text in that region was read correctly.

Grouping text into pages and files preserves the processing sequence, which may differ from the intended reading order in the source. Removing HTML formatting can also lose information conveyed by layout or table structure. Important readings should be checked against the original image or institutional record.

Dataset Structure

Configurations

All three configurations have a split named train. The split name is a storage convention; it does not indicate that the text is a verified training set or that separate evaluation data is provided.

Configuration	Structure
block_level	One row per block, with formatted text, plain text, block type, and bounding box.
page_level	One row per page, with a list of block texts and the joined page text.
file_level	One row per file, with a list of page texts.

The meaning of new_ocr_text_list changes by configuration: at page level it holds block texts; at file level it holds page texts. The three configurations describe the same processed material at different levels, so their rows should not be combined as independent observations.

Collection Labels and Source Identifiers

collection value	Source collection
smithsonian	Smithsonian Revolutionary Era Collections
rev_war_pension	NARA Revolutionary War Pension Files
chronicling_america	Library of Congress Chronicling America, 1770–1810

filename identifies the source file used for processing. Pension filenames incorporate the NARA identifier, such as 111403760.pdf. Newspaper filenames incorporate the serial number (LCCN), issue date, and edition, such as sn82014385_1809-07-08_1.pdf. Smithsonian processing files are associated with image identifiers rather than necessarily identifying a complete catalogue record.

Use collection together with filename to identify a file, and include page_number for page-level work. The linked source dataset cards describe the institutional identifiers and image fields needed to trace those processing inputs back to the collections.

Page Numbering and Text Order

page_number is zero-indexed: 0 identifies the first page of a processed file. When joining to a source dataset with one-indexed page numbers, account for that difference. A processing page index need not match a printed page number in the document.

Page-level text was assembled by ordering blocks by the numeric suffix of block_id. File-level text was assembled by ordering pages numerically. These are processing orders, not a guarantee of correct historical reading order. File-level lists contain the page texts represented in the processing output; the list position alone should not be assumed to establish a source page number if pages are missing.

Data Fields

Field types below match the published schemas. string and large_string are distinct schema types. Null values may occur where a value is unavailable.

Shared Fields

These three fields occur in all configurations.

• filename (string)
  Name of the source file used for OCR processing.
  Example: 111403760.pdf

• collection (string)
  Source collection label: smithsonian, rev_war_pension, or chronicling_america.
  Example: rev_war_pension

• OCR_run_ID (string)
  Identifier of the OCR run associated with the output.
  Example: VLM_OCR_1

Block-Level Fields

The block_level configuration has the three shared fields above and these seven fields, for ten fields in total.

• page_number (int64)
  Zero-indexed page number within the processed file.
  Example: 0

• block_id (large_string)
  Identifier assigned to the text block in the layout output.
  Example: /page/0/SectionHeader/0

• block_type (string)
  Layout classification assigned during processing.
  Example: SectionHeader

• block_markdown (large_string)
  Markdown representation of the block returned by the OCR service.
  Example: # THE DELAWARE GAZETTE.

• block_html (large_string)
  HTML representation of the block returned by the OCR service.
  Example: <h1>THE DELAWARE GAZETTE.</h1>

• block_bbox (large_string)
  Bounding box stored as a string containing four coordinates, rather than as a list-valued field. Coordinate order is [x1, y1, x2, y2].
  Example: [233.0, 212.0, 2197.0, 338.0]

• block_text (large_string)
  Plain text extracted from block_html during tabular processing.
  Example: THE DELAWARE GAZETTE.

Page-Level Fields

The page_level configuration has the three shared fields above and these three fields, for six fields in total.

• page_number (int64)
  Zero-indexed page number within the processed file.
  Example: 0

• new_ocr_text_list (large_list of large_string)
  Ordered list of block texts for the page.

• new_ocr_text (large_string)
  Page text assembled by joining the ordered block texts with newline characters, skipping blank entries.

File-Level Fields

The file_level configuration has the three shared fields above and this field, for four fields in total.

• new_ocr_text_list (large_list of string)
  Ordered list of page texts for the processed file. Each element contains the joined text of one processed page, rather than one block. This field has a different element type and meaning from the field of the same name in page_level.

Source Data

Original Data Producers

The materials come from three institutions. Refer to the linked source dataset cards for further information about selection, digitization, metadata, provenance, and rights.

Collection	Original institution	Materials
Smithsonian Revolutionary Era Collections	Smithsonian Institution	Object images from the National Museum of American History, National Postal Museum, Smithsonian American Art Museum, and National Portrait Gallery.
NARA Revolutionary War Pension Files	National Archives and Records Administration	Pension and bounty-land warrant application files related to Revolutionary War service.
Library of Congress Chronicling America, 1770–1810	Library of Congress	Digitized newspaper pages published from 1770 through 1810.

Data Access

The source datasets contain the underlying metadata and links to institutional records and images. Use the collection label, filename, and processing page number to locate the relevant input, accounting for the identifier and page-number conventions described above.

The complete VLM OCR bucket provides additional processing artifacts. Consult the original image when the accuracy of a reading matters.

Personal and Sensitive Information

The text comes from publicly available historical records and images. No personal or sensitive information was added from sources outside those materials. They may contain outdated or offensive terminology reflecting the periods in which they were created.

OCR can produce incorrect names or statements. A name or statement in this dataset records a machine-generated reading, not a verified fact about an individual.

Considerations for Using the Data

Risks and Limitations

• OCR can make errors. Text may be missing, misread, repeated, or generated without support in the image. The output is separate from the institutions' authoritative records.
• Coverage is limited to processed material. The tables do not represent all records in the contributing collections or guarantee that all content on a represented page was captured.
• Layout and sequence can affect interpretation. Block grouping, reading order, and conversion from formatted output to plain text may affect how passages appear in the page- and file-level tables.
• The source may include supporting material. Scanned files can include archival front matter, microfilm publication information, and other pages beyond the historical document itself. Such content is present in sampled pension OCR outputs and can affect text searches and aggregate counts.
• Downstream processing can compound errors. Entity extraction or other analysis based on these OCR outputs can carry forward existing errors and introduce additional ones. Refer to the raw named entities dataset card for the limits of that subsequent extraction.

Recommendations

• Choose the configuration that fits the question: blocks for layout and local context, pages for page-level analysis, and files for document-level work.
• Keep collection, filename, and, where available, page_number with text used in an analysis so readings can be traced back to their inputs.
• Account for zero-indexed page numbers when joining to source metadata. Confirm the relevant page dimensions and coordinate conventions before displaying bounding boxes on images.
• Inspect the source image or institutional record before relying on a particular name, date, or passage.
• Use a specific repository revision for research or other work that needs stable inputs.

Additional Information

Licensing Information

This dataset is released under CC0 1.0 Universal. For rights and reuse information about the underlying collection records and images, consult the source dataset cards and originating institutions.

Citation Information

BibTeX

@dataset{RevolutionCrossroads_OCR_Results_2026,
  author = {Revolution Crossroads Project Team},
  title = {Revolution Crossroads OCR Results},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RevolutionCrossroads/RevX_OCR_results}
}

APA

Revolution Crossroads Project Team. (2026). Revolution Crossroads OCR Results [Data set]. Hugging Face. https://huggingface.co/datasets/RevolutionCrossroads/RevX_OCR_results

A DOI can be added to both forms if one is assigned to this dataset.

Glossary

Term	Meaning
OCR	Optical character recognition: automated extraction of text from images.
VLM	Vision language model: an AI model that processes images and text.
Block	A unit of text and layout returned in Datalab's processing output.
Page-level	A representation that groups the extracted blocks of one processed page.
File-level	A representation that groups the processed page texts of one source file.
Bounding box	Four coordinates describing a region on the processed page image.

Dataset Card Contact

revolutioncrossroads@si.edu