Skip to content

Latest commit

 

History

History
55 lines (40 loc) · 2.54 KB

howto_slurk.md

File metadata and controls

55 lines (40 loc) · 2.54 KB
Raw

The Lightweight Dialogue Game framework

How to integrate with Slurk

Multiplayer games between a human player and a LM using slurk

In order to have a human player interact with a language model via the slurk chat interface, the game master is also a slurk bot that connects to a running slurk server. The game loop is then controlled via incoming slurk events:

class ChatMaster(GameMaster, APIChatBot):
    def __init__(self, experiment, model_name, token, user, task, host, port):
        GameMaster.__init__(self, GAME_NAME, experiment, model_name)
        APIChatBot.__init__(self, token, user, int(task), host, port)
        self.sio.on("text_message", self.incoming_message())

In order to start the game loop you can either directly call the ChatMaster object's run() method or its play() method (that then calls run()).

The core logic is driven by text_message events that are handled by the incoming_message() method. The bot checks that the message came from the relevant user, sets the Player message to the incoming string and creates the new dialog object to pass on to the LM just like before.

In the example, the slurk bot also implements general game logic. Before the interaction properly starts, the human player has to type the /ready command and when the dialog ends, the room is closed.

Running the GameMaster as a slurk bot

The repository contains a setup that lets human players enter an entry room that dynamically creates individual rooms for each player (ConciergeBot).

In order to run the whole setup locally, you need:

All three repositories should be located at the same level:

.
├── 101_clembench
├── slurk
└── slurk-bots

Navigate into the slurk-bots repository and run

../101_clembench/scripts/slurk.sh

This script will set up a local slurk server, build and run both bots, and generate an access token to log into slurk at localhost:5000, like so (you can choose any name):

TOKEN for logging into slurk interface at localhost:5000
64211e7b-8e90-4cd8-a07a-d24130e0c81d

In the mock version, this will just repeat "I don't know anything about that." for the LLM. To connect to a model, change "mock" to a model name in games/chatgame/master.py's main(). Running slurk.sh will then let you interact with the LM via a local slurk server.