Discord bot/companion for the group chatte, powered by the GPT-J Neo language model and modified with a soft prompt to understand all of our esoteric, elaborate inside jokes.

Setup

Python 3.8+ and PyTorch required. CUDA strongly recommended.

The c1-1.3B model used in development should work without needing beefy hardware. It has been tested on a 1050 Ti 4 GB.

Set up a virtual environment:

Linux/MacOS

python3 -m venv venv
source venv/bin/activate

Windows

py -3 -m venv venv
.\venv\Scripts\activate

Install required packages

pip install -r requirements.txt

Copy .env.example to .env and fill in the bot's TOKEN.

For chat scraping, you will also need to get your own USER_TOKEN.

In Discord, hit Ctrl+Shift+I to open up developer tools
Go to the Network tab and filter by XHR requests
Open a new channel, or scroll up, or do something else that will trigger an authenticated request
Click on one that looks suitable (e.g. messages?limit=50)
Under the "Request" tab, copy the contents of the Authorization request header.

Usage

Scrape the messages from the chat channel you wish to use for a soft prompt. You will be prompted for the channel ID, which you can get by having developer mode on in Discord and right-clicking, or copying the last part of the URL in the browser.

cd src
python -m scraper

Train the soft prompt (TODO)

Run the Hatsune Miku bot. The first time you do this, it will download the model, which is ~5 GB.

python -m miku

Final Remarks

sukima nuts