stefan-hf/rvlcdip-duplicates

RVL-CDIP Test-Train Duplicates

This dataset contains the result of a search for duplicates and near-duplicates between the training split of RVL-CDIP and its test and validation splits. It marks the test and validation documents that have a (near-) duplicate in the training split, lists the matching training documents, and gives the similarity scores of every pair that was compared. It contains no images and no text; rvlcdip_path identifies each document in the original release.

split	documents searched	with a duplicate in train	candidate pairs	duplicate pairs
test	39,999	14,076 (35.2%)	399,990	68,290
validation	40,000	14,151 (35.4%)	400,000	69,172

Configurations

config	contents	rows (test / validation)
documents	one row per test or validation document, with the flag and the matching training documents	40,000 / 40,000
pairs	one row per compared pair of a test or validation document and a training document, with the scores	399,990 / 400,000

The splits are named after the split of the query document; the other document of a pair is always from the training split. The rows of documents are in the order of the RVL-CDIP label files (test.txt, val.txt), which is also the order of stefan-hf/rvlcdip-label-fixes, so the two can be combined by position or by id.

A de-duplicated test split is obtained by removing the flagged documents:

from datasets import load_dataset

docs = load_dataset("stefan-hf/rvlcdip-duplicates", "documents", split="test")
dedup_test = docs.filter(lambda r: not r["has_train_duplicate"])

For a test split that is de-duplicated and has corrected labels:

fixes = load_dataset("stefan-hf/rvlcdip-label-fixes", split="test")
assert docs["id"] == fixes["id"]
keep = [i for i, (d, f) in enumerate(zip(docs["has_train_duplicate"], fixes["label_fixed"]))
        if d is False and f is not None]
fixed_dedup_test = fixes.select(keep)

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

How the duplicates were found

For each test and validation document, the ten training documents of the same category with the highest cosine similarity of their CLIP ViT-B/32 image embeddings were taken as candidates. Each candidate pair was then scored in two ways: by the similarity of the OCR text of the two pages (difflib.SequenceMatcher(...).ratio() over the whitespace-separated words), and by the number of SuperPoint keypoints that LightGlue matches between the two images with a matching score above 0.75 (ETH-CVG/lightglue_superpoint, with the second image resized to the size of the first). A pair is counted as a duplicate if the text similarity is above 0.2 and the number of matched keypoints is above 100. The pairs config contains the scores of all candidate pairs, so that other thresholds can be applied.

Fields

documents:

field	description
id	file name in our copy of the corpus, {split}_{category}_{uuid}.jpg
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)
has_train_duplicate	whether at least one candidate pair of the document is a duplicate
n_train_duplicates	number of duplicate pairs of the document (0 to 10)
train_duplicates	ids of the matching training documents, ordered by decreasing number of matched keypoints

pairs:

field	description
id, rvlcdip_path	the test or validation document
train_id, train_rvlcdip_path	the training document
label	original RVL-CDIP category of both documents
rank	rank of the training document among the ten candidates of the query, 1 for the highest CLIP similarity
clip_vit32	cosine similarity of the CLIP ViT-B/32 image embeddings
difflib	similarity of the OCR text, 0 to 1
superpoint	number of matched keypoints
duplicate	whether the pair is counted as a duplicate (difflib > 0.2 and superpoint > 100)

Known limitations

• The numbers of duplicates are lower bounds. Only the ten nearest training documents of the same original category were compared with each document, so that a duplicate with a different label, or one outside the ten candidates, is not found. Training documents were not compared with each other, and test documents were not compared with validation documents.
• The search used the original labels and pages. Label errors were not corrected before the candidates were selected.
• One test page was not searched. 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 in documents, so that the test split follows test.txt line by line, with null in has_train_duplicate, n_train_duplicates and train_duplicates, and it has no rows in pairs.

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}
}