BEiT3 based Korean VQA Model # (https://github.com/SeanJeonghwanLee/KoBEiT3)

Basic information

Model

• Base Model : beit3_large_indomain_patch16_224 (https://github.com/microsoft/unilm/tree/master/beit3)

  • best epoch : 8
  • learning rate : 2e-5
  • fixed seed : 42

• Tokenizer

  • korean sentencepiece tokenizer trained on korean wikipedia

Dataset

• KoBEiT3
  • aihub 시각정보 기반 질의응답 (https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=104)
    • Only Korean can access to the dataset
• Tokenizer
  • kowiki-latest-pages-articles.xml.bz2 (https://dumps.wikimedia.org/kowiki/latest/)