MTG AI

Overview

Magic the Gathering LLM with RAG for general magic questions and creating card combos.

Trained using question and answer data generated from cards and popular combinations.

Installation

pixi install --all

Usage

To start development shell

pixi shell -e dev

Training

# options can be set in the yaml file or passed as arguments
# --help to list all options

# for multi-gpu training
accelerate launch train.py --config ./llama-3.3-70B-Instruct-fsdp-qlora.yaml

# for single gpu training
accelerate launch train.py --config ./llama-3.3-70B-Instruct-unsloth.yaml

from mtg_ai.ai import ModelAndTokenizer, MTGCardAITrainer
model = ModelAndTokenizer.UNSLOTH_LLAMA_3_2_3B_INSTRUCT_Q8
ai_trainer = MTGCardAITrainer(
    model_name=model.value, 
    gguf_file=gguf_file, 
    dataset_name="cards"
)
ai_trainer.train()

# Modify the config file to change the training parameters or resume from checkpoint

To use the model for inference

from mtg_ai.ai import MTGAIRunner

ai_model_name: str = "./results"
rag_embedding_model_name: str = "sentence-transformers/all-MiniLM-L6-v2"

runner = MTGAIRunner(
    ai_model_name=ai_model_name,
    rag_embedding_model_name=rag_embedding_model_name
)

runner.run(
    "What is the converted mana cost of Acquisitions Expert?"
    max_new_tokens=100, 
    top_k=3
)

Inference Server

The easiest way to use this model for model inference is to run it with the huggingface/text-generation-inference server.

model=unsloth/Llama-3.3-70B-Instruct
adapter=FringeFields/mtg-ai-llama-3.2-3b-adapter
docker run --gpus all --shm-size 1g -p 8080:80 \
    ghcr.io/huggingface/text-generation-inference:latest \
    --model-id $model \
    --lora-adapters $adapter \
    --quantize bitsandbytes-nf4

Acknowledgements

Card data is from MTGJSON

Combo data is from CommanderSpellbook

Disclaimer

This is an unofficial project and is not affiliated with Wizards of the Coast.

Permitted under the Wizards of the Coast Fan Content Policy