This repository is home to the DOOM Mistral project, crafted during the CV Mistral AI Hackathon in San Francisco, on March 23-24 2024. Our team (Bhav, Umut, and Paul) developed a deep learning model capable of playing DOOM through visual input alone, utilizing the ViZDoom engine, a prominent tool in visual reinforcement learning. We transformed each game frame into a 32x64 grid, representing game elements in each cell, enabling the LLM to interpret the game via ASCII text. For a glimpse into how the model views the game, refer to this video (note: contains strobing effects):

https://twitter.com/i/status/1772075251297550457

We generated training data by playing DOOM, then fine-tuned a LoRA model on Mistral-7B, achieving reasonable results. While not an expert, the model can navigate the map and engage enemies when they appear.

During the inference-loop, the model takes in a prompt and a grid representing the current game screen and produces actions which are then asyncronously fed as input into the game.

Watch the model in action here:

https://twitter.com/i/status/1772166935410532709

You can find our Weights & Biases dashboard here.

To initialize, create a virtual environment and install the ViZDoom dependencies:

pip install -r requirements.txt

Rename .env.example to .env and input your FIREWORKS_API_KEY.

To start the game, execute:

python llm_game.py

Capture gameplay for training with: python user_game.py

Access the training dataset at HuggingFace 🤗, download it, and place it in the root directory.

Upload your dataset to Fireworks.ai for model training:

firectl create dataset doom-mistral doom_mistral.jsonl

Training settings are listed in train_settings.yaml. To train, use:

firectl create fine-tuning-job --settings-file train_settings.yaml --display-name "DOOM-Mistral"

Access the trained model on fireworks.ai, named doom-mistral.

To play the game with any LLM, modify model_id in line 16, then run: python llm_game.py

• E: Enemy
• P: Player
• B: Bullet
• W: Wall
• F: Floor
• A: ArmorBonus
• Z: Zombieman
• H: HealthBonus
• S: Stimpack

ViZDoom's original code is under the MIT license. ZDoom incorporates various sources, each with distinct licensing.