01-03 Random Choice for Generation

Why this matters

Weighted sampling turns static token probabilities into diverse generation behavior.

Intuition first (no jargon)

Higher-probability tokens should be selected more often, but not deterministically every time.

Code walkthrough

js
export function weightedRandom(items, weights, rng = Math.random) {
  // return one item based on weights
}

Your task

Implement weightedRandom(items, weights, rng).

• Validate equal lengths and non-empty arrays.
• Treat weights as non-negative.
• Use cumulative sums and one random draw.

Hints

• Sum weights first.
• Draw r = rng() * total.
• Walk cumulative weight until r < cumulative.

Check your thinking

1. Why pass rng as an argument?
2. What happens if all weights are zero?
3. Why do tests prefer deterministic randomness?

Stretch (optional)

Handle tiny floating-point errors robustly when selecting the last item.

Likely test focus

• Returns only from items.
• Deterministic with seeded rng.
• Correct behavior on edge inputs.

What should improve

Your generator can now sample plausible alternatives instead of always taking argmax.

Bridge to next lesson

Next lesson: combine cleaning, counting, and sampling into one baseline generator.