DUDE_loader / README.md
jordyvl's picture
Update README.md
fffc312
|
raw
history blame
1.99 kB
metadata
license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
pretty_name: DUDE
size_categories:
  - 10K<n<100K

Loading the dataset with a specific configuration

There are 3 different OCR versions to choose from with their original format or standardized DUE format, as well as the option to load the documents as filepaths or as binaries (PDF). To load a specific configuration, pass a config from one of the following:

#{bin_}{Amazon,Azure,Tesseract}_{original,due}
['Amazon_due', 'Amazon_original', 'Azure_due', 'Azure_original', 'Tesseract_due', 'Tesseract_original', 
'bin_Amazon_due', 'bin_Amazon_original', 'bin_Azure_due', 'bin_Azure_original', 'bin_Tesseract_due', 'bin_Tesseract_original']

Loading the dataset:

from datasets import load_dataset

ds = load_dataset("jordyvl/DUDE_loader", 'Amazon_original')

This dataset repository contains helper functions to convert the dataset to ImDB (image database) format. We advise to clone the repository and run it according to your preferences (OCR version, lowercasing, ...). When running the above data loading script, you should be able to find the extracted binaries under the HF_CACHE:
HF_CACHE/datasets/downloads/extracted/<hash>/DUDE_train-val-test_binaries, which can be reused for the data_dir argument.

For example:

python3 DUDE_imdb_loader.py \
--data_dir ~/.cache/huggingface/datasets/downloads/extracted/7adde0ed7b0150b7f6b32e52bcad452991fde0f3407c8a87e74b1cb475edaa5b/DUDE_train-val-test_binaries/

For baselines, we recommend having a look at the MP-DocVQA repository

We strongly encourage you to benchmark your best models and submit test set predictions on the DUDE competition leaderboard To help with test set predictions, we have included a sample submission file RRC_DUDE_testset_submission_example.json.