Deck LLM

Deck LLM (Language Model)

• Purpose

  Machine learning utilities for Yu-Gi-Oh! deck analysis and modeling using PyTorch. Provides neural network components, card embeddings, and training examples for deck-based language models.

• Core Components

  • nn.py — Neural network utilities and building blocks

    • Custom PyTorch modules and tensor operations
    • Attention mechanisms (Flash Attention support)
    • Embedding functions with Qwen3-Embedding integration
    • PCA dimensionality reduction
    • Various activation functions (SwiGLU, SoftmaxGLU)

  • ygonn.py — YGO-specific neural network features

    • Card database integration (SQLite cards.cdb)
    • Card embedding generation
    • Deck loading and preprocessing
    • Feature extraction for cards
    • Bindings and aliases management

  • bindx5c.py — Bind card ID prediction model

    • Lightning-based training example
    • Predicts 5-card bind combinations
    • Uses attention mechanism for sequence prediction

  • open.py — Open vocabulary model

    • Alternative model architecture
    • Card feature embedding model
    • Training example for deck generation

• Requirements

  • PyTorch with CUDA support
  • Flash Attention (optional, for performance)
  • Lightning AI framework
  • SQLite database (cards.cdb)
  • YGOPro installation with card data

• Usage Examples

  from ygonn import cardsfeat, decks
  
  # Load card features
  features = cardsfeat()
  
  # Load deck data
  deck_tensors = decks()
  
  # Train model (example)
  model = Model()
  trainer = L.Trainer(precision=16)
  trainer.fit(model, DataLoader(decks(), ...))
  

• Integration Skills

  Used by:

  • ygo/card-build/SKILL.md — For ML-assisted card design
  • ygo/deck-compare/SKILL.md — For deck pattern recognition
  • ygopro/ranking.md — For meta trend prediction

• Customization

  • Modify embedding models in ygonn.py
  • Adjust model architectures in examples
  • Add custom training objectives
  • Integrate with other YGO data sources

• Performance Notes

  • Models benefit from GPU acceleration
  • Use mixed precision (FP16) for training
  • Flash Attention significantly speeds up attention layers
  • PCA reduces feature dimensionality for efficiency

Source: ZisIsNotZis/ygoskill — distributed by TomeVault.