Nerdless

A conversational AI Telegram bot based on a language model finetuned on a custom dataset.

The bot randomly responds to random messages with a chance set by command.

You can talk to the bot by mentioning it, replying to its message, or texting it privately.

Model

Used raw imageboard dialogues data and processed it into a dataset (notebook available at notebooks/dataset.ipynb).

Finetuned ruDialoGPT-medium on the dataset (notebook available at notebooks/finetuning.ipynb).

The model card is available at HuggingFace 🤗.

Technologies

• transformers - for cleaning the data, finetuning and running the model
• aiogram 3 - for the Telegram bot
• asyncio - for non-blocking response generation
• PostgreSQL - for keeping records of history and settings for chats
• SQLAlchemy - for interacting with the database
• Docker - for a consistent and safe deployment

Launch

To run the Telegram bot application:

1. Clone the repository
2. Switch to the repository directory
3. Create a .env configuration and fill it in:

   # Telegram bot configuration
   TELEGRAM_TOKEN=
   MODE=polling  # 'webhook' for webhook mode
   PORT=
   
   # Language model configuration
   MODEL_MODE=local  # Running the model locally; to pull from HF, put in anything else
   HF_MODEL_NAME=rakepants/ruDialoGPT-medium-finetuned-toxic  # Used for pulling the model when running non-locally
   
   # Database configuration
   DATABASE_HOSTNAME=db
   DATABASE_NAME=
   DATABASE_USERNAME=
   DATABASE_PASSWORD=
   DB_CONTAINER_PORT=5454  # Exposed database container port

4. If running the model locally instead of pulling from HuggingFace, download the model weights and put them into ./app/ai/model/ruDialoGPT-medium-finetuned-toxic
5. Run docker-compose up --build.

Creators

Made by Artyom Eryomkin & Mikhail Kuznecov