Top-K Sampling
Domain: NLP Fundamentals
🎯 Motivation
Only samples from the K most likely next tokens, ignoring the long tail of unlikely ones.
📋 Concrete Example
Top-K = 40 means the model only considers the top 40 candidates and samples among them.
🔗 Analogy
Like a talent show where only the top 40 contestants get to perform - the rest are cut.