wenzhuoxu / civilcommentswpds_accuracy_with_distillbert

A Very Brief Implementation of Classifying CivilCommentsWPDS Dataset as in (R. Zhai, 2022)

Overview

Implements the ultilities functions in [civilcomments_utils.py]. Initialize training by running [civilcomments_train.py]. Currently don't support using command lines to change training parameters.

Can do training series in [civilcomments_pds.py]. Change the fraction of dataset by frac, and seek to monitor a smaller distance as frac gets closer to testing frac. If using batch as testing dataset, uncomment the corresponding lines in train().

Install dependent packages by

pip install pip --upgrade
pip install -r requirements.txt

Work division

Wenzhuo

(1) Worked on recreating training of FMOW dataset; (2) Developed code for Epsilon-KL Divergence function; (3) Assist in development of new divergence metric; (4) Report: Incorporate findings from new divergence functions;

Yihang

(1) Formatted and tested benchmark code from Runtian; (2) Lead team’s coding processes; (3) Developed new divergence metric; (4) Theoritical analysis and numerical simulation on Gaussian distributions; (5) Report: Incorporate findings from new divergence functions;

Martha

(1) Determined the methods to implement as part of the baseline testing and found code to assist in the development of the baseline divergence metrics (2) Report: Outlined the formatting of the report, Wrote the Data section, Background section; (3) Testing: Developed a function to output a Gaussian distribution and a shifted Gaussian. (4) Presentation: Presented, recorded, and edited the presentation that provides an overview of this work.