Insider-Outsider Classification - A new paradigm for Identifying Conspiracy Theories
imports.py
: Change the variables here to access the model checkpoints, dataset, and intermediatepkl
files available on OSF. Pay particular attention to*DATA_PATH
as well as theOPT
option to switch between training and inference.- Fine-tuning NP2IO is performed on 1-2 Titan RTX GPUs for about 2-3 hours with one of the GPUs solely attributed to data augmentation via BERT. Set
REEXTRACT
andSAVE
toFalse
to use a single-GPU for training post-data augmentation. The post-data augmented input files are available on OSF. - To train, run
python train.py
. TODO: Dockerize the environment. Till then transformers, pytorch1.7 is sufficient with additional minor dependencies. - To test, run
python SCRIPT_test_against_baselines.py
. This call takes a while - currently does not generate real-time outputs so it is a little annoying to run. test_inline.ipynb
provides a table-top setup for qualitative eval.- Other files are helpers.
Contact me if you have any questions.