LiteSeed
Back

Use Case

Train models on scenarios that rarely happen

Simulate rare but critical events at the frequency your model needs — not the frequency they occur in the real world.

Explore Platform

The challenge

Models learn to ignore what rarely happens

Fraud, equipment failure, medical emergencies, security incidents — the events that matter most for AI models are the ones that happen least often. Real-world datasets systematically underrepresent these scenarios. Models trained on real data learn to ignore them.

Control rare event frequency

Set the exact probability of rare events in the training dataset — independent of real-world occurrence rates.

Reproducible simulation

Reproduce the same rare event distribution exactly for every training run using the seeded RNG.

Blend with common cases

Use mixture distributions to blend rare events with normal behaviour at configurable ratios.

How LiteSeed helps

rare_event distribution

Define a field that returns a rare value with configurable probability p and a base value otherwise.

  • Parameters: base_value, rare_value, p (probability)
  • Reproducible: same seed always produces the same rare event pattern
  • Combinable with constraints for multi-field rare scenarios

mixture distribution

Blend multiple distributions with configurable weights — for example, 95% normal behaviour + 5% anomalous behaviour.

  • Parameters: components (array of distributions), weights (array of probabilities)
  • Supports any combination of the 8 base distributions
  • Weights must sum to 1.0

Edge case injection

Use Training mode edge case injection to ensure rare scenarios appear at a minimum frequency across the dataset.

  • Edge case injection configurable per Blueprint
  • Minimum frequency threshold per rare scenario
  • Tracked in distribution summary and run report

Constraint-based rare scenarios

Define multi-field rare scenarios using the constraint system — for example, high transaction amount + new account + unusual location.

  • Formula constraints for multi-field conditions
  • Soft constraints for rare scenario tracking
  • Violation rate tracked in run report