Skip to content

HySonLab/ViDeBERTa

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ViDeBERTa: A powerful pre-trained language model for Vietnamese, EACL 2023

Paper: https://aclanthology.org/2023.findings-eacl.79.pdf

ViDeBERTa

Contributors

  • Tran Cong Dao
  • Pham Nhut Huy
  • Nguyen Tuan Anh
  • Hy Truong Son (Correspondent / PI)

Main components

  1. Pre-training
  2. Model
  3. Fine-tuning

Pre-training

Code architecture

  1. bash: bash scripts to run the pipeline
  2. config: model_config (json files)
  3. dataset: datasets folder (both store original txt dataset and the pointer to memory of datasets.load_from_disk)
  4. source: main python files to run pre-training tokenizers
  5. tokenizer: folder to store tokenizers

Pre-tokenizer

  • Split the original txt datasets into train, validation and test sets with 90%, 5%, 5%.
  • Using the PyVi library to segment the datasets
  • Save datasets to disk

Pre-train_tokenizer

  • Load datasets
  • Train the tokenizers with SentencePiece models
  • Save tokenizers

Pre-train_model

  • Load datasets
  • Load tokenizers
  • Pre-train DeBERTa-v3

Model

Fine-tuning

Code architecture

  1. POS tagging and NER (POS_NER)
  2. Question Answering (QA and QA2)
  3. Open-domain Question Answering (OPQA)