stefan-hf/rvlcdip-redact

RVL-CDIP, De-identified, With and Without Identification Codes

This dataset contains two versions of the RVL-CDIP document classification corpus, each as page images and as OCR output from two engines:

• redact: personal data on the pages (such as social security numbers) has been replaced with synthetic values of the same kind, drawn onto the page in place of the original. About 15,171 of the 400,000 pages were changed; the other pages are those of RVL-CDIP.
• redact_noid: the redact pages with the identification codes (Bates numbers and similar stamps) removed as well. Most RVL-CDIP pages contain such a code, and its position and format are associated with the category label, so a classifier can use it as a shortcut. The codes were located with stefan-hf/yolov8n-rvlcdip-idcodes; their boxes and transcriptions are in stefan-hf/rvlcdip-id-codes.

Training on redact and on redact_noid and comparing the two measures how much a model relies on the codes. The classifiers in stefan-hf/rvlcdip-classifiers were trained on these files.

Configurations

config	contents	size
images_redact	page images, redact	38.1 GB
images_redact_noid	page images, redact_noid	37.5 GB
textract_redact	Amazon Textract OCR of the redact images	1.85 GB, 84.3 million words
textract_redact_noid	Amazon Textract OCR of the redact_noid images	1.84 GB, 84.0 million words
tesseract_redact	Tesseract 4.1.1 OCR of the redact images	1.34 GB, 69.2 million words
tesseract_redact_noid	Tesseract 4.1.1 OCR of the redact_noid images	1.33 GB, 68.6 million words

Every config has the RVL-CDIP splits (320,000 train / 40,000 validation / 39,999 test documents) and one row per document. The rows are in the same order in all six configs, and the id field is shared, so an image config and a text config can be joined by position or by id.

from datasets import load_dataset

images = load_dataset("stefan-hf/rvlcdip-redact", "images_redact_noid", split="test")
words = load_dataset("stefan-hf/rvlcdip-redact", "textract_redact_noid", split="test")
assert images[0]["id"] == words[0]["id"]

Downloading a single configuration

Each config is stored in its own folder of Parquet files, so one config can be downloaded without the others. Naming the config in load_dataset fetches only the files of that config, and adding split restricts the download to the files of that split:

from datasets import load_dataset

# Tesseract OCR of the redact_noid pages, all three splits (1.33 GB)
ocr = load_dataset("stefan-hf/rvlcdip-redact", "tesseract_redact_noid")

# test split of the redact images only (3.8 GB instead of 38.1 GB)
test_images = load_dataset("stefan-hf/rvlcdip-redact", "images_redact", split="test")

# rows read over the network without storing the files
stream = load_dataset("stefan-hf/rvlcdip-redact", "images_redact", split="train", streaming=True)

images_redact is the default config, so load_dataset("stefan-hf/rvlcdip-redact") without a config name downloads the 38.1 GB of redact images.

To get the Parquet files themselves, restrict the download to one folder:

hf download stefan-hf/rvlcdip-redact --repo-type dataset --include "textract_redact/*" --local-dir rvlcdip_redact

from huggingface_hub import snapshot_download

snapshot_download("stefan-hf/rvlcdip-redact", repo_type="dataset", allow_patterns="textract_redact/*")

Fields

All configs:

field	description
id	file name in our copy of the corpus, {split}_{category}_{uuid}.jpg; the same as in stefan-hf/rvlcdip-id-codes
rvlcdip_path	path of the page in the original RVL-CDIP release, e.g. imagesb/b/w/o/bwo50e00/92704843.tif
label	category in the original RVL-CDIP labels, as a string (16 categories)
label_fixed	category in our corrected label set, or null if the document was removed from it (296,605 / 36,380 / 36,446 documents remain)

The corrected labels come from a manual review of all RVL-CDIP labels, which is available with the reviewers' judgments and notes in stefan-hf/rvlcdip-label-fixes; the two datasets share the id and rvlcdip_path fields.

The test and validation documents that have a duplicate or near-duplicate in the training split are marked in stefan-hf/rvlcdip-duplicates, which also shares these two fields.

Image configs:

field	description
image	the page as a grayscale JPEG, 1,000 pixels tall, with the bytes of the file the classifiers were trained on (34 pages were re-encoded, see Known limitations)
width, height	size of the image in pixels

OCR configs:

field	description
text	the text of the page, one line of text per detected line
tokens	the words of the page in the order the engine returned them
bboxes	one box per word, [x0, y0, x1, y1], scaled to 0–1000 in both directions (the LayoutLM convention); multiply by width / 1000 and height / 1000 of the image for pixels
confidence	the engine's confidence for each word (0–100)
line	for each word, the index of the line it belongs to, counted from 0 in text
n_words	number of words
handwritten	for each word, whether Textract classified it as handwriting (Textract configs only)
n_handwritten	number of words classified as handwriting (Textract configs only)

The raw responses of the OCR engines are not included. They are about 290 KB per page for Textract, and the fields above keep the words, boxes, confidences, line membership, and the handwriting flag from them.

Known limitations

• Residual personal data. The de-identification was automatic and is not complete. After the automatic pass we searched the OCR text of the redact pages for social security numbers that it had left in place, reviewed the candidates by hand, and covered those that remained on 34 pages with white boxes. The same boxes are applied in redact and redact_noid, and the words under them are left out of the four OCR configs. These 34 images are therefore re-encoded and differ from the files our classifiers were trained on. The search cannot find numbers that the OCR engines did not read, and other kinds of personal data were not reviewed. Please report any page that needs further masking.
• Names and business contact details printed on the pages, such as authors, addressees, and company telephone numbers, were not targeted and are present as in RVL-CDIP.
• Identification codes missed by the detector (about 3% of codes) are still present in redact_noid, and a small share of the removed regions did not contain a code.
• OCR of handwriting is poor for both engines, and Tesseract gives no handwriting flag. The engines returned no words for 3.5% to 5.7% of pages, depending on the config; these rows have empty lists.
• One test page of RVL-CDIP is missing. imagese/e/j/e/eje42e00/2500126531_2500126536.tif is not in this dataset, so the test split has 39,999 documents and every config has 399,999 rows. This is a known problem of RVL-CDIP, in which the file of this page is corrupt and cannot be read (see the discussion on the RVL-CDIP dataset page).
• label contains the known label errors of RVL-CDIP; label_fixed is our corrected set.

License

The pages are derived from RVL-CDIP, which is a subset of IIT-CDIP, which came from the Legacy Tobacco Document Library, for which license information can be found here. The same terms apply to the images and the OCR output in this dataset.

Citation

If you use this dataset, please cite:

@inproceedings{larson-etal-2025-spurious,
  author    = {Larson, Stefan and Duwal, Sharad and Vilnrotter, Brian and Chakkithara, Gayatri and Padwal, Vedant and Leach, Kevin},
  title     = {Spurious Cues in {RVL-CDIP} and {Tobacco3482} Document Classification: The Case of {ID} Codes},
  booktitle = {Proceedings of the 2025 ACM Symposium on Document Engineering},
  series    = {DocEng '25},
  year      = {2025},
  month     = aug,
  pages     = {1--4},
  publisher = {ACM},
  address   = {New York, NY, USA},
  doi       = {10.1145/3704268.3748683},
  url       = {https://doi.org/10.1145/3704268.3748683}
}

@inproceedings{larson-etal-2026-revising,
  author    = {Larson, Stefan and Nagy, Attila and Desai, Sam and Desai, Cyrus and Lima, Nicole and Yuan, Yixin and Betala, Siddharth and Prajapati, Kaushal Kumar and Suleiman, Jamiu T. and Duwal, Sharad and Leach, Kevin},
  title     = {Revising {RVL-CDIP}: Quantifying Errors and Test-Train Overlap},
  booktitle = {Proceedings of the 2026 ACM Symposium on Document Engineering},
  series    = {DocEng '26},
  year      = {2026},
  articleno = {23},
  numpages  = {10},
  isbn      = {9798400727863},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  doi       = {10.1145/3820755.3821486},
  url       = {https://doi.org/10.1145/3820755.3821486}
}