Contextual Action Language Model (CALM) and the ClubFloyd Dataset

Code and data for paper Keep CALM and Explore: Language Models for Action Generation in Text-based Games at EMNLP 2020.

Overview

Our ClubFloyd dataset (calm/lm_data.zip) is crawled from the ClubFloyd website and contains 426 human gameplay transcripts, which cover 590 text-based games of diverse genres and styles.

The data consists of 223,527 context-action pairs in the format [CLS] observation [SEP] action [SEP] next observation [SEP] next action [SEP]. We use [CLS] observation [SEP] action [SEP] next observation [SEP] as the context to train language models (n-gram, GPT-2) to predict next action [SEP], and show that this action generation ability generalizes to unseen games and supports gameplay when combined with reinforcement learning.

Getting Started

Clone repo and install dependencies:

bash
pip install torch==1.4 transformers==2.5.1 jericho fasttext wandb importlib_metadata
git clone https://github.com/princeton-nlp/calm-textgame && cd calm-textgame
ln -s ../lm calm && ln -s ../lm drrn

(If the pip installation fails for fasttext, try the build steps here: https://github.com/facebookresearch/fastText#building-fasttext-for-python)

Train CALM:

bash
cd calm
unzip lm_data.zip
python train.py

Trained model weights can be downloaded here for both GPT-2 and n-gram models.

Then train DRRN using the trained CALM:

bash
cd ../drrn
python train.py --rom_path ../games/${GAME} --lm_path ${PATH_TO_CALM} --lm_type ${gpt_or_ngram}

To quickly try out the GPT-2 CALM model:

python
from lm import GPT2LM
model = GPT2LM("model_weights/gpt2")
print(model.generate("[CLS] observation [SEP] action [SEP] next observation [SEP]", k=30))

Citation

@inproceedings{yao2020calm,
    title={Keep CALM and Explore: Language Models for Action Generation in Text-based Games},
    author={Yao, Shunyu and Rao, Rohan and Hausknecht, Matthew and Narasimhan, Karthik},
    booktitle={Empirical Methods in Natural Language Processing (EMNLP)},
    year={2020}
}