Dataset
We encourage the participants to read the MMFM Data Download & Process for more data details.
Phase 1
For this phase, we build up a comprehensive data suite comprising publically available datasets including DocVQA, FUNSD, IconQA, InfogrpahicVQA, Tabfact, TextbookVQA, WebSrc, Wildreceipt, WTQ. All these datasets are aligned with the goal of the challenge of image document understanding in very specific domains like table fact-checking, answering questions about receipts, question-answering about infographics, and visual question-answering about documents, etc. This collection contains the train set and the test set.
Phase 2
For this phase, an alien test set does not include answers The original intent of this dataset is to prevent people from overfitting on the publicly available datasets. This test set will consist of data that is similar to the distribution of the Phase 1 dataset collection but will not consist of any publically available data sources.