Nucleus Sampling From Scratch | Dataford Interview Questions

Implement nucleus sampling (top-p sampling) from scratch in Python, ensuring efficient token selection and probability normalization.

Implement sample_top_p(probabilities, p, seed), where probabilities is a non-empty list of nonnegative numeric token probabilities, p is in (0, 1], and seed makes sampling reproducible. Return the selected token index as an integer. Normalize inputs that do not already sum to one, retain the smallest highest-probability set whose cumulative normalized probability is at least p, renormalize that set, and sample one index from it.

Constraints

1 <= len(probabilities) <= 500
Each probability is a nonnegative integer or floating-point value
sum(probabilities) > 0
0 < p <= 1
seed is an integer
Return the original token index, not its rank among sorted probabilities

Function Signature

def sample_top_p(probabilities, p, seed):

Your question is Nucleus Sampling From Scratch. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem

Implement nucleus sampling (top-p sampling) from scratch in Python, ensuring efficient token selection and probability normalization.

Constraints

1 <= len(probabilities) <= 500
Each probability is a nonnegative integer or floating-point value
sum(probabilities) > 0
0 < p <= 1
seed is an integer
Return the original token index, not its rank among sorted probabilities

Function Signature

def sample_top_p(probabilities, p, seed):

Your question is Nucleus Sampling From Scratch. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.