muddsmasher/dvce-event-grammar-model

DVCE Event Grammar Model

Universal Event Grammar Model — predicts the next event in any sequential system.

Trained on 1.3M+ real-world events across 30+ domains.

What it does

Given a sequence of past events, predicts:

1. What happens next (291 event types)
2. When it happens (inter-event time)
3. How severe (0-1 severity score)

Domains trained on

Geopolitics (GDELT), earthquakes (USGS), sports (StatsBomb), commodities, cybersecurity, weather, clinical trials, logistics, financial markets, manufacturing, healthcare, e-commerce, energy grid, IT incidents, DeFi/blockchain, agriculture, construction, and more.

Architecture

• Transformer Decoder (GPT-style)
• d_model=256, n_layers=4, n_heads=8
• 4.5M parameters
• Continuous time encoding
• Multi-task output (type + time + severity)

Usage

Training

• 1.3M events from 30+ domains
• 100 epochs on balanced dataset with domain tokens
• Trained on AWS SageMaker (g5.xlarge)
• Total training cost: ~

License

MIT