JerichoWorld Dataset

Dataset corresponding to the paper Modeling Worlds in Text found here https://openreview.net/forum?id=Y1YtS9MZA75.

Dataset Organization

Unzip data.zip

data/train.json: The main dataset file containing a list of training examples.
data/test.json: The main dataset file containing a list of test examples.

The dataset is organized as:

{
"rom": "zork1",
"state": state,
"next_state": next_state,
"action": act,
"reward": rew,
}

Where state and next_state are two subsequent states with act being the corresponding action and rew being the corresponding reward. The rom names vary based on what is defined in the paper.

Example of a state:

{
    "walkthrough_act": "Open window",
    "walkthrough_diff": "((), ((235, 11),), ())",
    "obs": "Behind House\nYou are behind the white house. A path leads into the forest to the east. In one corner of the house there is a small window which is slightly ajar.\n\n",
    "loc_desc": "Behind House\nYou are behind the white house. A path leads into the forest to the east. In one corner of the house there is a small window which is slightly ajar.\n\n",
    "inv_desc": "You are carrying:\n  A jewel-encrusted egg\n\n",
    "inv_objs": {
      "The jewel encrusted egg is closed.": [
        "egg"
      ]
    },
    "inv_attrs": {
      "egg": []
    },
    "location": {
      "name": "Behind House",
      "num": 79
    },
    "surrounding_objs": {
      "The window is slightly ajar, but not enough to allow entry.": [
        "small",
        "window"
      ],
      "The house is a beautiful colonial house which is painted white. It is clear that the owners must have been extremely wealthy.": [
        "white",
        "house"
      ],
      "There's nothing special about the way.": [
        "path"
      ]
    },
    "surrounding_attrs": {
      "window": [],
      "house": [],
      "path": []
    },
    "state": "saves/f461488f-3085-4f5a-ac2f-bd424561e8c6.pkl",
    "valid_acts": {
      "((), ((235, 11),), ())": "open small",
      "(((86, 4),), (), ())": "take on egg",
      "(((87, 79),), (), ())": "put down egg",
      "(((4, 80),), ((80, 3),), ())": "south",
      "(((87, 79), (86, 79)), (), ())": "throw egg at small",
      "(((4, 81),), (), ())": "north",
      "(((4, 74),), ((74, 3),), ())": "east"
    },
    "graph": [
      [
        "you",
        "have",
        "jewel-encrusted egg"
      ],
      [
        "you",
        "in",
        "Behind House"
      ],
      [
        "Behind House",
        "east",
        "North House"
      ],
      [
        "golden clockwork canary",
        "in",
        "jewel-encrusted egg"
      ]
    ],
    "score": 5
  }

Fields

Each example defines the following fields:

rom: Name of the game that generated this example.
obs: Narrative text returned by the game as a result of the last action.
loc_desc: Text returned by look command from current location.
inv_desc: Text returned by inventory command from current step.
inv_objs: Dictionary of {obj_description : [obj_names]} containing detected objects in the player's inventory.
inv_attrs: Dictionary of {obj_name : [obj_attrs]} containing detected objects' attributes in the player's inventory.
surrounding_objs: Dictionary of {obj_description : [obj_names]} containing detected objects in the player's immediate surroundings.
surrounding_attrs: Dictionary of {obj_name : [obj_attrs]} containing detected objects' attributes in the player's immediate surroundings.
score: Current game score at this step.
location: Name and number for the world-object corresponding to the player's current location.
state: Path to pickle file containing saved game state.
walkthrough_act: Action taken by the data collection agent from the current state.
walkthrough_diff: world_diff corresponding to taking the taken action.
valid_acts: list of world_diff : action_str. The important part here is the world_diff since there are many action strings that can result in the same world diff.
graph: list of lists of subject, relation, object of knowledge graph for current step

Current Tasks

Predict knowledge graphs: Given the current observations and graph can we predict some/all of the next knowledge graph? Ground truth given by in graph.
Predict valid actions: Given a the current observations and graph can we predict some/all of the valid actions? Ground truth given by world-diffs in valid_acts.

Acknowledgements

We thank https://github.com/microsoft/jericho and its creators for the walkthroughs of the games and the environments.