Grover is a model for Neural Fake News -- both generation and detection. However, it probably can also be used for other generation tasks.

Setting up your environment

NOTE: If you just care about making your own RealNews dataset, you will need to set up your environment separately just for that, using an AWS machine (see realnews/.)

There are a few ways you can run Grover:

  • Generation mode (inference). This requires a GPU because I wasn't able to get top-p sampling, or caching of transformer hidden states, to work on a TPU.
  • LM Validation mode (perplexity). This could be run on a GPU or a TPU, but I've only tested this with TPU inference.
  • LM Training mode. This requires a large TPU pod.
  • Discrimination mode (training). This requires a TPU pod.
  • Discrimination mode (inference). This could be run on a GPU or a TPU, but I've only tested this with TPU inference.

I used Python3.6 for everything. Usually I set it up using the following commands:

curl -o ~/ -O  && \
     chmod +x ~/ && \
     ~/ -b -p ~/conda && \
     rm ~/ && \
     ~/conda/bin/conda install -y python=3.6

Then pip install -r requirements-gpu.txt if you're installing on a GPU, or pip install requirements-tpu.txt for TPU.

Misc notes/tips:

  • If you have a lot of projects on your machine, you might want to use an anaconda environment to handle them all. Use conda create -n grover python=3.6 to create an environment named grover. To enter the environment use source activate grover. To leave use source deactivate.
  • I'm using tensorflow 1.13.1 which requires Cuda 10.0. You'll need to install that from the nvidia website. I usually install it into /usr/local/cuda-10.0/, so you will need to run export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64 so tensorflow knows where to find it.
  • I always have my pythonpath as the root directory. While in the grover directory, run export PYTHONPATH=$(pwd) to set it.

Quickstart: setting up Grover for generation!

  1. Set up your environment. Here's the easy way, assuming anaconda is installed: conda create -y -n grover python=3.6 && source activate grover && pip install -r requirements-gpu.txt
  2. Download the model using python base
  3. Now generate: PYTHONPATH=$(pwd) python sample/ -model_config_fn lm/configs/base.json -model_ckpt models/base/model.ckpt -metadata_fn sample/april2019_set_mini.jsonl -out_fn april2019_set_mini_out.jsonl

Congrats! You can view the generations, conditioned on the domain/headline/date/authors, in april2019_set_mini_out.jsonl.


    title={Defending Against Neural Fake News},
    author={Zellers, Rowan and Holtzman, Ari and Rashkin, Hannah and Bisk, Yonatan and Farhadi, Ali and Roesner, Franziska and Choi, Yejin},
    journal={arXiv preprint arXiv:1905.12616},