Grover

UPDATE, Sept 17 2019. We got into NeurIPS (camera ready coming soon!) and we've made Grover-Mega publicly available without you needing to fill out the form. You can download it using download_model.py.

(aka, code for Defending Against Neural Fake News)

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

Visit our project page at rowanzellers.com/grover, the AI2 online demo, or read the full paper at arxiv.org/abs/1905.12616.

What's in this repo?

We are releasing the following:

Code for the Grover generator (in lm/). This involves training the model as a language model across fields.
Code for the Grover discriminator in discrimination/. Without much changing, you can run Grover as a discriminator to detect Neural Fake News.
Code for generating from a Grover model, in sample/.
Code for making your own RealNews dataset in realnews/.
Model checkpoints freely available online for all of the Grover models. For using the RealNews dataset for research, please submit this form and message me on contact me on Twitter or through email. You will need to use a valid account that has google cloud enabled, otherwise, I won't be able to give you access 😢

Scroll down 👇 for some easy-to-use instructions for setting up Grover to generate news articles.

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 ~/miniconda.sh -O  https://repo.continuum.io/miniconda/Miniconda3-4.5.4-Linux-x86_64.sh  && \
     chmod +x ~/miniconda.sh && \
     ~/miniconda.sh -b -p ~/conda && \
     rm ~/miniconda.sh && \
     ~/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!

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
Download the model using python download_model.py base
Now generate: PYTHONPATH=$(pwd) python sample/contextual_generate.py -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.

FAQ: What's the deal with the release of Grover?

Our core position is that it is important to release possibly-dangerous models to researchers. At the same time, we believe Grover-Mega isn't particularly useful to anyone who isn't doing research in this area, particularly as we have an online web demo available and the model is computationally expensive. We previously were a bit stricter and limited initial use of Grover-Mega to researchers. Now that several months have passed since we put the paper on arxiv, and since several other large-scale language models have been publicly released, we figured that there is little harm in fully releasing Grover-Mega.

Bibtex

@inproceedings{zellers2019grover,
    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},
    booktitle={Advances in Neural Information Processing Systems 32},
    year={2019}
}