← AI Terminology

Weak Supervision

Weak supervision builds training labels from noisy programmatic sources (heuristics, distant supervision, other models) instead of only hand-labeled gold data.

Snorkel-style systems popularised combining labeling functions.
Why It Matters in AI
Gold labels do not scale. Weak methods create large silver datasets for classifiers and sometimes LLM SFT. Managing noise via label models is a practical data-centric skill.
Key Points
Aspect Description
Con Systematic biases; leakage
Pro Scale; rapid iteration
Tools Snorkel, programmatic ETL label pipelines
Combine Label models estimate source accuracies
Related Distant supervision, synthetic data
Sources Rules, regexes, KB alignment, teacher models
Simple Analogy
Many imperfect TAs marking homework with different rubrics, then a system estimating whom to trust on each question.
Common Usage Examples
  • Write labeling functions over tickets
  • Snorkel generative label model
  • Train classifier on probabilistic labels
  • Audit errors for rule blind spots
Summary
In short: Weak supervision creates noisy programmatic labels at scale — trading perfection for coverage when gold annotation is scarce.