biglam/AmericanStories-parquet

AmericanStories (parquet)

A parquet-native reformat of dell-research-harvard/AmericanStories — article-level full text of ~20 million U.S. newspaper scans (1774–1963) from the Library of Congress's Chronicling America collection, originally extracted by Dell et al. (arXiv:2308.12477).

This repo exists so the dataset loads in one line with the standard datasets / polars / pyarrow / dask stack, with no custom loading script and full Dataset Viewer support on the Hub.

What's different from the original

• Article-level only. This reformat does not include the scan-level (*_content_regions) view. If you need raw scan metadata, bounding boxes, and per-region OCR, use the upstream dataset directly.
• Plain parquet, no custom loading script. Files are organised as {year}/train-*.parquet. You can load them with datasets, polars, pyarrow, dask, duckdb, etc.
• Each year is a config, not a split. Use load_dataset("biglam/AmericanStories-parquet", "1870", split="train") — the upstream API was dataset["1870"].
• October 2023 snapshot. This reformat does not include the additions made to the upstream dataset on 2025-03-25. For the comparable upstream cut, pin the original to revision="v0.1.0".
• Schema is unchanged from the upstream article-level config: article_id, newspaper_name, edition, date, page, headline, byline, article.

For the authoritative version and the scan-level data, please cite and use the upstream dataset: https://huggingface.co/datasets/dell-research-harvard/AmericanStories.

Quick start

from datasets import load_dataset

# A single year
ds = load_dataset("biglam/AmericanStories-parquet", "1870", split="train")
print(ds[0])

# Stream — no local download
ds = load_dataset("biglam/AmericanStories-parquet", "1870", split="train", streaming=True)
for example in ds.take(5):
    print(example["headline"], example["date"])

With polars (skip datasets entirely):

import polars as pl

df = pl.read_parquet(
    "hf://datasets/biglam/AmericanStories-parquet/1870/train-*.parquet"
)

With pyarrow:

import pyarrow.dataset as pds
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
table = pds.dataset(
    "datasets/biglam/AmericanStories-parquet/1870", filesystem=fs, format="parquet"
).to_table()

DuckDB can read directly from the parquet URLs over httpfs.

Dataset summary

Reproduced from the upstream dataset card:

The American Stories dataset is a collection of full article texts extracted from historical U.S. newspaper images. It includes nearly 20 million scans from the public domain Chronicling America collection maintained by the Library of Congress. The dataset is designed to address the challenges posed by complex layouts and low OCR quality in existing newspaper datasets.

It was created using a novel deep learning pipeline that incorporates layout detection, legibility classification, custom OCR, and the association of article texts spanning multiple bounding boxes. It employs efficient architectures specifically designed for mobile phones to ensure high scalability.

The dataset offers high-quality data that can be utilized for various purposes. It can be used to pre-train large language models and improve their understanding of historical English and world knowledge. The dataset can also be integrated into retrieval-augmented language models, making historical information more accessible, including interpretations of political events and details about people's ancestors. Additionally, the structured article texts in the dataset enable the use of transformer-based methods for applications such as detecting reproduced content. This significantly enhances accuracy compared to relying solely on existing OCR techniques.

Data fields (article level)

Field	Type	Description
article_id	string	Unique identifier for an article.
newspaper_name	string	Newspaper name.
edition	string	Edition number.
date	string	Date of publication (YYYY-MM-DD).
page	string	Page number (e.g. p1).
headline	string	Headline text (may be empty).
byline	string	Byline text (may be empty).
article	string	Article body text.

Example row:

{
  "article_id":      "1_1870-01-01_p1_sn82014899_00211105483_1870010101_0773",
  "newspaper_name":  "The weekly Arizona miner.",
  "edition":         "01",
  "date":            "1870-01-01",
  "page":            "p1",
  "headline":        "",
  "byline":          "",
  "article":         "PREyors 10 leaving San Francisco for Wash ington City, our Governor, A. r. K. Saford. called upon Generals Thomas and Ord and nt the carrying out of what …"
}

Source data

Reproduced from the upstream dataset card:

The dataset is drawn entirely from image scans in the public domain that are freely available for download from the Library of Congress's website. The source language was produced by people — newspaper editors, columnists, and other sources. See the upstream paper for pipeline details (layout detection, legibility classification, custom OCR, and article-text association across bounding boxes).

Considerations for using the data

This dataset contains unfiltered content composed by newspaper editors, columnists, and other sources. In addition to other potentially harmful content, the corpus may contain factual errors and intentional misrepresentations of news events. All content should be viewed as individuals' opinions and not as a purely factual account of events of the day.

OCR quality varies. The Dell et al. pipeline targets higher quality than prior Chronicling America OCR, but errors remain — especially for early years, blurry scans, and dense layouts.

Licensing

CC-BY-4.0, inherited from the upstream dataset. Please attribute Melissa Dell et al. when using or redistributing this data.

Citation

If you use this dataset, please cite the original paper:

@misc{dell2023american,
      title={American Stories: A Large-Scale Structured Text Dataset of Historical U.S. Newspapers},
      author={Melissa Dell and Jacob Carlson and Tom Bryan and Emily Silcock and Abhishek Arora and Zejiang Shen and Luca D'Amico-Wong and Quan Le and Pablo Querubin and Leander Heldring},
      year={2023},
      eprint={2308.12477},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Dataset DOI (upstream): 10.57967/hf/0757

Acknowledgements

All credit for the dataset itself belongs to Melissa Dell, Jacob Carlson, Tom Bryan, Emily Silcock, Abhishek Arora, Zejiang Shen, Luca D'Amico-Wong, Quan Le, Pablo Querubin, and Leander Heldring. This repository is purely a format conversion to make the dataset easier to load and explore on the Hub. If you find this data useful, please cite the paper above and ⭐ the upstream dataset.