stefan-hf/rvlcdip-label-fixes

RVL-CDIP Label Fixes

This dataset contains the result of a manual review of the category labels of all 400,000 documents of RVL-CDIP. For each document it gives the original label, whether the reviewers judged that label correct, the corrected label where one was assigned, and the reviewers' notes. It contains no images and no text; the rows are in the order of the RVL-CDIP label files (train.txt, val.txt, test.txt), and rvlcdip_path identifies each document in the original release.

review outcome	documents	label_correct	label_fixed
the label is correct	354,067	true	the original label
the label is wrong and a corrected label was assigned	15,366	false	the corrected label
the label is wrong and no corrected label was assigned	26,661	false	null
the reviewers could not decide whether the label is correct	3,906	null	null

The corrected version of RVL-CDIP consists of the documents with a label_fixed (296,606 train / 36,380 validation / 36,447 test), with label_fixed as the label. Documents without one are removed from it; this includes the undecided documents.

from datasets import load_dataset

fixes = load_dataset("stefan-hf/rvlcdip-label-fixes")
fixed_test = fixes["test"].filter(lambda r: r["label_fixed"] is not None)

The pages and their OCR are available in stefan-hf/rvlcdip-redact, which shares the id and rvlcdip_path fields with this dataset.

The test and validation documents that have a duplicate or near-duplicate in the training split are marked in stefan-hf/rvlcdip-duplicates, whose rows are in the same order as the rows of this dataset.

Fields

field	description
id	file name in our copy of the corpus, {split}_{category}_{uuid}.jpg, where the category is the original label
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_correct	whether the reviewers judged label correct; null if they could not decide
label_fixed	the corrected category, equal to label where it was judged correct; null if the document is removed from the corrected version
mixed	whether the reviewers marked the page as a mixture of two categories (5,977 documents)
mixed_with	the second category of a mixed page, the first being label; null if the page is not mixed or no category was named (25 documents)
note	the reviewers' note, in lower case and otherwise as written, or null (35,225 documents have one)

Nearly all mixed pages are among the documents with a wrong label and no corrected one, and are therefore not part of the corrected version. The most frequent second categories are form (1,988), handwritten (1,387), memo (949), letter (767) and news_article (405).

The notes are working notes and were not normalized. Most of them name what a document is when no category fits, for example press_release (6,508), specification data_sheet (4,845) and leo_burnett_estimate (2,841); unsure and check mark documents that the reviewers found hard to judge. The same thing can be written in several ways (press_release, press release), and the mixed pages appear here as mixed:<category> as well as in the two fields above.

Known limitations

• Notes that mark personal data are left out. Where a note said that a page contains personal data, that part of the note was removed, together with a small number of abbreviations whose meaning could not be established. The notes of 1,946 documents were removed completely for this reason.
• One test page cannot be read. imagese/e/j/e/eje42e00/2500126531_2500126536.tif is corrupt in RVL-CDIP, which is a known problem (see the discussion on the RVL-CDIP dataset page). It has a row here so that the test split follows test.txt line by line, but it has no image in stefan-hf/rvlcdip-redact.
• 33 documents were judged correct although a different category had also been entered for them; label_fixed is the original label for these.

License

The annotations in this dataset are released under CC BY 4.0. The documents themselves are not part of this dataset; see RVL-CDIP for their terms.

Citation

If you use this dataset, please cite:

@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}
}