RevolutionCrossroads/RevX_NER_Entities

Dataset Card for RevX Revolutionary War Era Named Entities

Dataset Summary

This dataset contains 24,035,075 machine-extracted named entity occurrences from materials brought together for Revolution Crossroads: Smithsonian Institution collection records and object images, National Archives and Records Administration (NARA) Revolutionary War pension files, and Library of Congress Chronicling America newspapers.

Each row records an occurrence of a name, place, organization, date, event, or vessel in the processed text, along with its assigned type and information that helps locate it in the source material. These are raw extraction results. Repeated mentions have not been reconciled into distinct people, places, or other identities, and the counts should not be interpreted as counts of unique entities.

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

Dataset Description

For Revolution Crossroads, the Smithsonian Institution prepared this dataset by extracting named entities from text associated with three collections: Smithsonian Institution Revolutionary era collections, NARA Revolutionary War pension files, and Library of Congress Chronicling America newspapers. The results bring entity occurrences from these collections into a common structure while retaining information about where each occurrence was found.

Dataset Details

• Prepared by: Smithsonian Institution, Office of Digital & Innovation
• Shared by: Revolution Crossroads
• Language: Primarily English
• License: CC0 1.0

Collection	Records	Pages	Entity occurrences
Smithsonian Institution collections	12,667	14,829	129,163
NARA Revolutionary War pension files	49,065	1,291,962	14,794,537
Library of Congress Chronicling America newspapers	13,871	53,917	9,111,375
Total	75,603	1,360,708	24,035,075

The figures describe the source records and pages represented in this extraction and the number of entity occurrence rows produced. They are not counts of distinct people, places, or other entities.

Entity Types

The model assigned one of the following types to each extracted occurrence. These definitions describe the intended categories; individual assignments have not been manually validated.

Type	Definition
person	A person, including one named by initials or surname alone.
geographic_location	A settlement or governed area, such as a village, town, city, county, township, parish, state, territory, or country.
named_place	A named feature, structure, or landholding, such as a building, tavern, church, mill, street, farm, river, or hill.
organization	A named business or organization, including companies, regiments, courts, churches, government bodies, and newspapers.
calendar_date	A date or date range with a determinable year that can be represented in ISO 8601.
temporal_event	A time reference without a determinable year, such as a season, day of the week, or month without a year.
named_event	A named event, such as a battle, treaty, Act of Congress, or constitutional amendment.
named_vessel	A named ship or other vessel, such as a brig, schooner, sloop, or packet.

Relationship to Source Datasets

This dataset contains entity occurrences extracted from text associated with three Revolution Crossroads source collections:

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

For information about the scope of each collection, its original records, metadata, provenance, and rights, refer to the linked source dataset cards. This card describes the subsequent text and entity extraction work and the resulting entity rows.

The extraction used positioned text blocks from RevX_OCR_results for all three collections. For Smithsonian records, it also used catalogue text and EDAN documents from the Smithsonian source dataset. That catalogue text is written metadata, rather than text read from an object image.

Extraction input	Commit used
RevolutionCrossroads/RevX_OCR_results	54d32993728511c65299721c7c7664b4a7f73a42
RevolutionCrossroads/si_us_revolutionary_era_collections	bd53b6822470f7ad398637801a2b44ee41cb7147

Curation Rationale

The three collections are held by different institutions and have no shared index of the people, places, and events they describe. Extracting named entities into one structure makes it possible to search for mentions across the collections and return to the text in which each mention appeared. A shared name does not, by itself, establish that two mentions refer to the same person or place.

The entity types were chosen to reflect what these materials contain: people, geographic locations, named places, organizations, dates, events, and vessels. Titles of books, pamphlets, almanacs, plays, and songs were not included as entity types because that material is already recorded in bibliographic sources. Newspaper titles are an exception and were extracted as organizations.

Dataset Creation

This dataset was created from three Revolution Crossroads collection datasets: Smithsonian Revolutionary Era Collections, NARA Revolutionary War Pension Files, and Library of Congress Chronicling America, 1770–1810. The NARA dataset, for example, provides the page-level pension records from which the pension text and source identifiers can be traced. The linked source dataset cards describe how those collections were assembled.

For entity extraction, the project used positioned text blocks in RevX_OCR_results, produced from the collections’ page images. Smithsonian catalogue text and EDAN documents from the Smithsonian source dataset were also included. Those records contain written catalogue metadata rather than text read from an object image.

Consecutive text blocks were grouped into windows of about 1,000 characters. Each window was sent to gemma-4-12b-it in a request constrained by a JSON schema. The preceding and following windows, along with the record title, were supplied as context; entities were extracted from the target window only.

The model returned entity text and an assigned type. The processing code then located each returned string in the target window and mapped it to its source text block and character offsets where available. If the string could not be found, it was discarded rather than assigned a guessed position.

Notes on Completeness

26,026 requests returned no answer in the first pass. Every one of those windows was retried, and 18,988 succeeded. 7,038 windows remain unextracted, or 0.22 percent of all windows. They are listed in errors_12b_retry.jsonl; entity occurrences in those windows are absent from this dataset.

Failure in the first pass	Windows	Recovered on retry
501 Not Implemented	6,129	99.0%
Connection reset	4,416	99.0%
Malformed HTTP response	1,713	99.4%
400 Bad Request	13,768	49.8%

Nearly all requests in the first three categories succeeded on retry. 400 Bad Request behaved differently: about half failed again, and all 7,038 remaining failures returned a 400. The reason for those recurring failures is unknown. The extractor retained only the first 200 characters of each exception, which recorded the status code but not the server's explanation.

Another 421,378 requests completed without returning an entity. These are distinct from failed requests: the model returned an answer, but no entity occurrence was recorded for the window.

Smithsonian coverage differs from the other collections. All 12,667 Smithsonian records contribute catalogue text, but only about 2,200 have a page image processed by the layout service. Block identifiers and bounding boxes are therefore present on about 12 percent of Smithsonian rows, compared with 100 percent of rows from the pension files and newspapers.

Data Collection and Processing

Processing Steps

1. Source records were drawn from the three Revolution Crossroads collection datasets described above. Smithsonian catalogue text and EDAN documents were also included.
2. Positioned text blocks for the page images were read from RevX_OCR_results.
3. Consecutive blocks were grouped into target windows of about 1,000 characters. The preceding and following windows, along with the record title, were supplied as context.
4. Each target window was sent to gemma-4-12b-it for entity extraction using a constrained JSON schema. Windows that returned no answer in the first pass were retried.
5. Each returned entity was located in the target window text and attributed to its source block, with character offsets where available. Repeated occurrences in the text were retained as separate rows.
6. Operational fields were removed, dates and entity text were normalised, and duplicate extraction outputs within a window were removed.
7. Rows from both passes were merged and written in document order. See Row Order.

Supporting Files Available

File	Contents
parquet/collection=<name>/	The entity table, with one directory per collection.
run_manifest_12b_full.json	Provenance for the first pass, including source repositories and commits, model, measured block and window maxima, and counts for each collection.
run_manifest_12b_retry.json	The corresponding provenance for the retry pass.
errors_12b_full.jsonl	The 26,026 requests that returned no answer in the first pass. Most were recovered on retry.
errors_12b_retry.jsonl	The 7,038 requests still unanswered after the retry. These identify the windows from which entity occurrences are absent.

Accessing a Specific Data Version

This dataset may be updated as source data is revised or extraction is re-run. Accessing a tagged version keeps a workflow stable and reproducible.

from datasets import load_dataset

dataset_name = "RevolutionCrossroads/RevX_NER_Entities"
commit_tag = "NER_v1"

ds = load_dataset(dataset_name, "all", revision=commit_tag, split="train")

One collection can be requested on its own:

pension = load_dataset(dataset_name, "pension", revision=commit_tag, split="train")

The Parquet files can also be scanned directly for analysis:

import polars as pl
from huggingface_hub import snapshot_download

local = snapshot_download(dataset_name, revision=commit_tag, repo_type="dataset")
df = pl.scan_parquet(f"{local}/parquet/**/*.parquet")

people = (
    df.filter((pl.col("entity_type") == "person") & (pl.col("occurrence") == 0))
      .group_by("entity_text_normalised")
      .len()
      .sort("len", descending=True)
      .head(50)
      .collect()
)

This example ranks normalised person-name strings by the number of extraction windows in which they occur. It does not count distinct people. The snapshot_download call downloads the repository files locally before Polars scans them.

Date Normalisation

Dates were standardised after entity extraction where the text provided enough information to determine a year. The iso_date_method field records whether an ISO date came from the model or was derived or repaired during post-processing. Relative dates that could not be anchored, such as “the 3d instant,” were left unresolved rather than assigned a guessed date.

The model classified some text without a determinable year as calendar_date. During processing, 2,357 such rows were changed to temporal_event. No rows were changed in the other direction.

The date_out_of_era field flags 231,234 rows with a resolved year outside 1600 to 1860. These rows were retained because a flagged value may be a date in the source or an extraction error. Users can review or exclude them according to their purpose.

Quality Considerations

This is a raw, experimental entity extraction dataset. The named entities have not been manually reviewed against the source documents or evaluated against a reference set. Precision and recall are unknown. The assigned types, explanations, normalised strings, and dates may contain errors and should not be treated as authoritative descriptions of the source material.

The extraction was designed to keep results tied to the text supplied to the model. The model was instructed to return entities found in the target window, and processing code located each returned string in that window before recording its occurrence and source text. A returned string that could not be located was discarded rather than given a guessed position. These checks limit unsupported outputs, but they do not establish that the text was read correctly or that the model classified it correctly.

Most image-derived text used for this extraction came from AI-assisted OCR. An OCR error can therefore be carried into an entity row, and the entity model can introduce further errors when identifying or classifying that text. The block_text field lets users inspect the text supplied for extraction, but an important finding should also be checked against the original image or institutional record. Smithsonian catalogue fields are a separate source of written metadata and were not read from an image by OCR.

The checks completed for this release were internal consistency checks, not an evaluation of historical or extraction accuracy. Derived columns were recalculated and compared, row counts were checked across the published formats, and character offsets were checked against block_text. On 23 rows, the offsets do not reproduce entity_text because of line breaks across formatting boundaries. Case-sensitive text matching may also have discarded valid entities whose model response differed from the supplied text only in capitalisation; the effect on recall is unknown.

Some correctly extracted text can still distort aggregate counts. In the pension files, printed introductions and form instructions recur across many files. A set of repeated text blocks accounts for 496,009 entity rows, or 3.4 percent of pension rows. These mentions appear in the scanned material, but they often describe the publication or form rather than an applicant. Counts of organizations, dates, and events are particularly affected.

Text blocks also vary greatly in length. The longest observed blocks contain 1,499 characters for Smithsonian, 23,825 for pension files, and 27,388 for Chronicling America. Code that displays or highlights block_text should allow for long values. The per-collection maxima are recorded in the run manifests.

Dataset Structure

Subsets

The default all configuration contains rows from all three collections. The si, pension, and chronam configurations each contain one collection. Use a single-collection configuration when only that material is needed; the collection field is also available for filtering within all.

Some fields are populated differently across the collections:

Field	Smithsonian	Pension files	Chronicling America
block_id, block_bbox	12%	100%	100%
image_id	12%	0%	0%
iso_date, iso_date_method, date_out_of_era	29%	14%	5%
Other fields	100%	100%	100%

Most Smithsonian entity rows come from catalogue text rather than text read from an object image. Those rows have no image-based block identifier or bounding box. Pension and newspaper rows come from page images and have positioned text blocks. The image_id field applies only to Smithsonian images. Date fields are populated where a date could be recorded or assessed.

Row Order

Rows are stored in document order: by record_id, then page, then window number within the page. Within a window, they retain the order returned by the model. This order is deterministic for a given dataset revision.

Rows from the initial and retry passes are merged into that order. A retry row appears alongside other rows from the same page, while run_id identifies which pass produced it.

Keeping rows from the same record together also improves compression of repeated block_text values. The dataset occupies 939 MB in document order, compared with 1,372 MB when the same rows are stored as two separate groups in extraction order.

Entity Occurrences and Source Context

Each row records one occurrence of an extracted entity in a text window. If the same entity text appears three times in that window, it produces three rows. occurrence numbers those appearances from 0, and n_occurrences gives their total. To count an extracted entity once per window, filter to occurrence = 0. That still does not produce a count of distinct people, places, or other identities across documents.

Each row retains the entity text as it appeared in the text supplied for extraction, its assigned type, and a short model-generated rationale for that type. It also carries the source record and page identifiers, the surrounding block_text, and character offsets within that text. For text read from images, a block identifier and bounding box locate the text on the page. Smithsonian catalogue fields have no image-based block position.

These fields allow users to inspect an extraction in context. Because image-derived text came through OCR, the original image or institutional record should be consulted when the reading matters.

Data Fields

Each row has 26 fields. The examples below come from a Smithsonian catalogue-field row unless otherwise noted. Fields may contain null values when they do not apply or no value is available.

Identity and Provenance Fields

• run_id (large_string)
  Extraction pass that produced the row: 12b_full or 12b_retry. The corresponding run manifest records details of that pass.
  Example: 12b_full

• llm_model (large_string)
  Model used to extract the entity.
  Example: gemma-4-12b-it

• collection (large_string)
  Source collection: si, pension, or chronam.
  Example: si

• record_id (large_string)
  Identifier of the source record. Smithsonian uses an EDAN identifier; pension files use a NARA identifier; and newspapers use an identifier incorporating the serial number, issue date, and edition.
  Example: edanmdm:nmah_1000590

• page (int32)
  Page number within the processing unit. Smithsonian catalogue text can have page = 0 because it is record metadata rather than text read from a numbered page image.
  Example: 0

• unit_id (large_string)
  Identifier combining record_id and page with a vertical bar.
  Example: edanmdm:nmah_1000590|0

• window_id (large_string)
  Identifier of the target text window submitted for extraction. Rows sharing a window_id came from the same request.
  Example: edanmdm:nmah_1000590|p0|w0

• bucket_path (large_string)
  Location of the layout output from which the block was read.
  Example: smithsonian/

Entity Fields

• entity_text (large_string)
  Entity string as it appeared in the text supplied for extraction. Spelling and capitalisation are preserved; image-derived text may contain OCR errors.
  Example: United States

• entity_text_normalised (large_string)
  Lowercase version of entity_text, with repeated whitespace reduced to a single space. This supports grouping by string but does not reconcile identities.
  Example: united states

• entity_type (large_string)
  Type assigned by the model. The eight available types are defined below.
  Example: geographic_location

• confidence (float32)
  Confidence value reported by the model for its assigned type, from 0 to 1. It has not been calibrated as a measure of accuracy.
  Example: 1.0

• rationale (large_string)
  Short model-generated explanation for the assigned type.
  Example: Country

• occurrence (int32)
  Index of this appearance of an entity within its target window, starting at 0.
  Example: 0

• n_occurrences (int32)
  Number of appearances of that entity in the target window.
  Example: 1

Date Fields

• iso_date (large_string)
  Date expressed in ISO 8601 where a year could be determined. Values may be a year, month, full date, or interval. Null where no date was resolved.
  Examples: 1809, 1809-06, 1809-06-24, 1793/1814

• iso_date_method (large_string)
  Records whether iso_date was supplied by the model or derived or repaired during post-processing. See Date Normalisation. Null where no method applies.
  Example: model

• date_out_of_era (bool)
  True when the resolved year falls outside 1600 to 1860. Null where no date was assessed.
  Example: null for the sample place-name row

Source Text and Position Fields

• block_id (large_string)
  Identifier of the positioned text block containing the entity. Null for Smithsonian catalogue fields that were not read from an image.
  Example: /page/0/Text/23 for an image-derived block

• block_type (large_string)
  Classification of the source text block, such as Text, Table, SectionHeader, or ListGroup. field identifies a Smithsonian catalogue field.
  Example: field

• block_bbox (large_list)
  Bounding box of an image-derived block, expressed as [x1, y1, x2, y2]. Null for Smithsonian catalogue fields.
  Example: null for the sample catalogue-field row

• image_id (large_string)
  Identifier of the Smithsonian image from which a block was read. Null for catalogue fields and for pension and newspaper rows.
  Example: null for the sample catalogue-field row

• field_name (large_string)
  image for text read from an image; for Smithsonian catalogue text, the name of the catalogue field.
  Example: place

• char_start (int32)
  Starting character position of the entity within block_text, counted from 0.
  Example: 18

• char_end (int32)
  Ending character position of the entity within block_text.
  Example: 31

• block_text (large_string)
  Source text block or catalogue-field text containing the entity. This is the text supplied for extraction, not necessarily a verified transcription of the original image.
  Example: place paper made: United States:

Source Data

Source Data

Original Data Producers

The materials represented here come from three institutions. The linked Revolution Crossroads source dataset cards provide more detail about each collection’s scope, metadata, provenance, and rights.

Collection	Original institution	Materials
Smithsonian Revolutionary Era Collections	Smithsonian Institution	Catalogue records and 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 collection datasets linked above contain the underlying records and links to materials made available by their institutions. Entity rows in this dataset retain collection, record_id, and page so users can locate the corresponding source record and examine the extraction in context. Consult the original institutional record or image when a name, date, or other reading is important to an interpretation.

Personal and Sensitive Information

This dataset contains names and other information drawn from publicly available historical records. No personal or sensitive information was intentionally added from sources outside those records. The materials may contain outdated or offensive terminology reflecting the periods in which they were created.

A person row records a string identified by a model in the text supplied for extraction. It does not establish that the named person existed, that two mentions identify the same person, or that a statement in the source is accurate. OCR and entity extraction errors can also produce incorrect names. Users should check the original record before publishing claims about an individual.

Considerations for Using the Data

Risks and Limitations

• Extraction accuracy has not been measured. There is no reference set for evaluating these entity rows, so precision and recall are unknown. Both the AI-assisted OCR and the entity extraction can introduce errors. These rows are experimental outputs, separate from the institutions’ source records.

• Mentions are not resolved identities. The same name can appear in many rows and may refer to one person, several people, or an OCR or extraction error. entity_text_normalised groups strings by simple text normalisation; it does not determine whether mentions refer to the same individual, place, or organization.

• Some windows are absent. After the retry, 7,038 extraction requests remained unanswered. Their contents are not represented by entity rows. See Notes on Completeness.

• Model confidence is limited as a filter. Many confidence values are near the top of the range, and the values have not been calibrated against verified results. A high value should not be read as confirmation that an extraction is correct.

• Source positions have known exceptions. On 23 rows, the character offsets do not reproduce entity_text. Formatting within some text blocks can also affect precise highlighting. The surrounding text and original image remain available for inspection.

Recommendations

• Use occurrence = 0 when counting each extracted entity string once per window. Count all rows when studying individual mentions. Neither measure counts distinct real-world people, places, or other identities.
• Use entity_text_normalised to group strings that differ only in case or spacing, and retain entity_text when displaying how a name appeared in the extracted text.
• Review iso_date_method alongside iso_date. Consider the date_out_of_era flag when building timelines or filtering by period.
• Inspect block_text and, where available, the source image or institutional record before relying on a particular extraction. A model-generated rationale or high confidence value does not replace that check.

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 three source dataset cards and the originating institutions.

Citation Information

BibTeX

@dataset{RevolutionCrossroads_NER_Entities_2026,
  author = {Revolution Crossroads Project Team},
  title = {RevX Revolutionary War Era Named Entities},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RevolutionCrossroads/RevX_NER_Entities}
}

APA

Revolution Crossroads Project Team. (2026). RevX Revolutionary War Era Named Entities [Data set]. Hugging Face. https://huggingface.co/datasets/RevolutionCrossroads/RevX_NER_Entities

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

Glossary

Term	Meaning
Block	A unit of source text used for extraction. An image-derived block may have a type and position on the page. A Smithsonian catalogue field is recorded with the type field and has no image-based position.
Window	The target unit of extraction, formed from consecutive text blocks of about 1,000 characters. Neighboring windows and the record title were supplied as context.
Occurrence	One appearance of an extracted entity string within a target window. A string appearing three times produces three rows.
Out of era	A resolved date with a year outside 1600 to 1860, as indicated by date_out_of_era.

Dataset Card Contact

revolutioncrossroads@si.edu