Skip to content

We implemented QANet from scratch and improved baseline BiDAF. We also used an ensemble of BiDAF and QANet models to achieve EM/F1 of 69.47/71.96, ranking #3 on the leaderboard as of Mar 4, 2022.

License

Notifications You must be signed in to change notification settings

jerrylzy/SQuAD-QANet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CS224N default final project (2022 IID SQuAD track)

Setup

  1. Make sure you have Miniconda installed

    1. Conda is a package manager that sandboxes your project’s dependencies in a virtual environment
    2. Miniconda contains Conda and its dependencies with no extra packages by default (as opposed to Anaconda, which installs some extra packages)
  2. cd into src, run conda env create -f environment.yml

    1. This creates a Conda environment called squad
  3. Run conda activate squad

    1. This activates the squad environment
    2. Do this each time you want to write/test your code
  4. Run python setup.py

    1. This downloads SQuAD 2.0 training and dev sets, as well as the GloVe 300-dimensional word vectors (840B)
    2. This also pre-processes the dataset for efficient data loading
    3. For a MacBook Pro on the Stanford network, setup.py takes around 30 minutes total
  5. Browse the code in train.py

    1. The train.py script is the entry point for training a model. It reads command-line arguments, loads the SQuAD dataset, and trains a model.
    2. You may find it helpful to browse the arguments provided by the starter code. Either look directly at the parser.add_argument lines in the source code, or run python train.py -h.

About

We implemented QANet from scratch and improved baseline BiDAF. We also used an ensemble of BiDAF and QANet models to achieve EM/F1 of 69.47/71.96, ranking #3 on the leaderboard as of Mar 4, 2022.

Topics

Resources

License

Stars

Watchers

Forks