This is the official repository for the GPT-2 models described in the EMNLP 2020 paper [STORIUM: A Dataset and Evaluation Platform for Machine-in-the-Loop Story Generation]. It has all the code necessary to reproduce the models and analysis from the paper.

A high-level outline of our dataset and platform. In this example from a real STORIUM game, the character ADIRA MAKAROVA uses the strength card DEADLY AIM to DISRUPT THE GERMANS, a challenge card. Our model conditions on the natural language annotations in the scene intro, challenge card, strength card, and character, along with the text of the previous scene entry (not shown) to generate a suggested story continuation. Players may then edit the model output, by adding or deleting text, before publishing the entry. We collect these edits, using the matched text as the basis of our USER metric. New models can be added to the platform by simply implementing four methods: startup, shutdown, preprocess, and generate.

This repository contains the code that makes our GPT-2 story generation models deployable on our evaluation platform, so it serves as a great template for how to structure your code. Please see the file figmentate.py for the simple API required for making your model deployable on our platform. You will also need to provide a json file with any properties needed to pass to your startup method. See for example the properties below:

{
  "scene_entry":
  {
    "properties": {
      "checkpoint_path": "/var/lib/figmentator/checkpoint",
      "sample": {
	"top_p": 0.9,
	"temperature": 0.9,
	"repetition_penalty": 1.2
      }
    },
    "requires": ["torch==1.3.0", "transformers==2.2.0", "kiwisolver==1.1.0"],
    "cls": "model=figmentate:GPT2Figmentator"
  }
}

The key scene_entry defines the type of model being created. Currently, we only support models that generate the text of a scene entry, though we might support other types of prediction models in the future, like suggesting cards or narrator actions.

The properties object will be passed to your startup method. It allows for defining any parameters needed for sampling from your model.

The requires list, is simply a list of python packages that need to be installed for your model to run. These will be automatically installed when your model is deployed. If you notice, we specify the deep learning package torch as a requirement. That's because our code is agnostic to the underlying deep learning framework being used by your model. That means it should support models using other frameworks like tensorflow or jax.

Finally, the cls string is the class that wraps your model. It is specified using Python's entry points syntax.

@inproceedings{akoury2020storium,
  Author = {Nader Akoury, Shufan Wang, Josh Whiting, Stephen Hood, Nanyun Peng and Mohit Iyyer},
  Booktitle = {Empirical Methods for Natural Language Processing},
  Year = "2020",
  Title = {{STORIUM}: {A} {D}ataset and {E}valuation {P}latform for {S}tory {G}eneration}
}