stefan-hf/rvlcdip-n

RVL-CDIP-N

RVL-CDIP-N is a test set for document classifiers trained on RVL-CDIP. Its 1,002 page images belong to RVL-CDIP categories but come from a different distribution: they were collected from DocumentCloud and from web search, and are largely born-digital, more recent than 2006, and unrelated to the tobacco industry, where RVL-CDIP consists of scanned tobacco-litigation documents. A classifier that has learned the categories, and not the peculiarities of the RVL-CDIP scans, should keep its accuracy on these pages.

The set was introduced, together with the out-of-scope set RVL-CDIP-O, in Evaluating Out-of-Distribution Performance on Document Image Classifiers (NeurIPS 2022 Datasets and Benchmarks).

The set has been publicly available since that paper was published; the original release is at github.com/gxlarson/rvl-cdip-ood. This repository contains the same pages as JPEG files together with their OCR.

Contents

• 1,002 JPEG images in a single test split, each with a label and a text field holding the OCR text of the page (Amazon Textract, one line of text per detected line).
• The raw Textract responses, with word and line bounding boxes, are in textract/, one JSON file per page with the same file stem as the image.
• Pages are in color and 1,000 pixels tall, converted from the TIFF files of the original release.
• Twelve of the sixteen RVL-CDIP categories are present; advertisement, file_folder, presentation, and scientific_report are not.

category	pages	category	pages
budget	58	memo	47
email	33	news_article	86
form	70	questionnaire	39
handwritten	176	resume	184
invoice	57	scientific_publication	39
letter	152	specification	61

The label field is the category name as a string (for example news_article), so predictions of a sixteen-way RVL-CDIP classifier can be scored by name.

RVL-CDIP is grayscale and these images are in color. Convert them to grayscale (image.convert("L")) before giving them to a model trained on RVL-CDIP.

Zero-shot classification with CLIP

The example below classifies every page into the sixteen RVL-CDIP categories with CLIP, using one prompt per category of the form "a resume document", and needs no training. It reaches 80.6% accuracy on this dataset with openai/clip-vit-large-patch14-336.

import torch
from datasets import load_dataset
from transformers import CLIPModel, CLIPProcessor

CATEGORIES = ["advertisement", "budget", "email", "file folder", "form", "handwritten", "invoice", "letter",
              "memo", "news article", "presentation", "questionnaire", "resume", "scientific publication",
              "scientific report", "specification"]
prompts = [f"{'an' if c[0] in 'aeiou' else 'a'} {c} document" for c in CATEGORIES]  # "a resume document", ...

ds = load_dataset("stefan-hf/rvlcdip-n", split="test")

device = "cuda" if torch.cuda.is_available() else "cpu"
model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14-336").to(device).eval()
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14-336")

correct = 0
for start in range(0, len(ds), 32):
    batch = ds[start:start + 32]
    inputs = processor(text=prompts, images=[im.convert("RGB") for im in batch["image"]],
                       return_tensors="pt", padding=True).to(device)
    with torch.no_grad():
        pred = model(**inputs).logits_per_image.argmax(dim=1)  # index into the 16 prompts
    for p, y in zip(pred.tolist(), batch["label"]):
        correct += CATEGORIES[p].replace(" ", "_") == y  # labels are category names, e.g. "news_article"
print(f"zero-shot accuracy: {correct / len(ds):.3f}")

Classifiers trained on RVL-CDIP are available at stefan-hf/rvlcdip-classifiers; their accuracy on RVL-CDIP-N ranges from 70% to 87%.

License

The labels and the selection of the documents are released under CC BY 4.0. The documents themselves were collected from DocumentCloud and from web search, and they and the OCR text derived from them remain the property of their original authors and publishers; they are redistributed here for research on document classification. The pages have not been reviewed for personal data. If you hold the rights to a document, or find a page that should not be public, please open a discussion on the Community tab of this dataset and the page will be removed.

Citation

@inproceedings{larson-etal-2022-rvlcdipn,
  author    = {Larson, Stefan and Lim, Yi Yang Gordon and Ai, Yutong and Kuang, David and Leach, Kevin},
  title     = {Evaluating Out-of-Distribution Performance on Document Image Classifiers},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {35},
  pages     = {11673--11685},
  publisher = {Curran Associates, Inc.},
  year      = {2022},
  url       = {https://proceedings.neurips.cc/paper_files/paper/2022/file/4c0986bd04d747745beba3752bdf4d9d-Paper-Datasets_and_Benchmarks.pdf}
}