#numpy

17 numpy interview questions, asked at Amazon, Meta, OpenAI and 10 more companies. Last asked Sep 2026.

+1
Vectorized nearest-neighbour classifier as a linear layerCodingImplement a 1-nearest-neighbour classifier in NumPy with no Python loops. Then rewrite it as the forward pass of a small network, a matrix product plus a bias followed by an…
Last asked Sep 2026
+1
Find noisy annotators and retrain a classifier without themCodingA classifier is trained on labels from several human annotators, and some of them label carelessly. Find which annotators are unreliable, drop their labels, retrain, and show…
Last asked Sep 2026
+2
Implement k-means clusteringCodingImplement k-means from scratch: initialize k centroids, assign every point to its nearest centroid, move each centroid to the mean of its points, and repeat until the assignments…
Last asked Sep 2026
+1
Greedy, top-k and top-p decoding for a language modelCodingYou get a toy language model with a small vocabulary whose forward pass returns next-token logits for the tokens so far. Using NumPy, write the generation loop and three ways to…
Last asked Sep 2026
+2
Linear regression with hand-written gradient descentCodingFit a one-feature linear regression by gradient descent on mean squared error, without an ML library. Then discuss choosing the learning rate, when to stop, and how to spot a…
Last asked Sep 2026
Softmax cross-entropy forward and backward, then a training loopCodingImplement from scratch, with NumPy but no autograd, the forward and backward pass of a softmax layer followed by cross-entropy loss. Keep it numerically stable. Then use it in a…
Last asked Sep 2026
Cross-entropy loss with masking and label smoothing in NumPyCodingImplement cross-entropy loss over a batch of logits in NumPy, stable for large values. Add a mask so padded positions do not count, then add label smoothing. Be ready to explain…
Last asked Jul 2026
Stable softmax and multi-head attention in NumPyCodingIn NumPy, implement a numerically stable sigmoid and softmax, then scaled dot-product attention, then multi-head attention from scratch.
Last asked Jul 2026
Nearest-center assignment without an N x K x D distance tensorCodingGiven N points and K centers in D dimensions, assign each point to its nearest center, as in the assignment step of k-means. The obvious broadcast builds an N x K x D array of…
Last asked Jul 2026
Top-p (nucleus) sampling in NumPyCodingImplement top-p sampling in NumPy: from a vector of logits, keep the smallest set of most likely tokens whose probabilities add up to at least p, renormalize them, and sample one…
Last asked Jun 2026
Remaining path length at every waypoint of a batch of trajectoriesCodingYou get a tensor of trajectories of shape batch, numwaypoints, 2 (x, y per waypoint). Return a tensor of shape batch, numwaypoints with a suffix sum for every waypoint: the…
Last asked Feb 2026
GRPO in NumPy: group-relative advantages, process and shaped rewardsML designIn a notebook, implement the core math of Group Relative Policy Optimization with NumPy. For each prompt the policy samples a group of completions, each completion gets a score…
Last asked Jan 2026
2D convolution forward pass in NumPy, with parameter countsCoding1. Given the input tensor shape, input channels, output channels, kernel size, stride and padding, how many parameters does a Conv2D layer have, with and without bias? What is the…
Last asked Dec 2025
Write a 2-D convolution, a biased attention step and a training loopCodingImplement small ML building blocks from scratch: a 2-D convolution in NumPy, an attention step with an added bias term before the softmax, and the missing pieces of a standard…
Last asked Dec 2025
Scaled dot-product attention from scratchCodingWrite scaled dot-product attention by hand: from queries, keys and values, compute the scores, scale them by the square root of the key size, apply an optional mask, take a…
Last asked Sep 2025
Attention and a small Transformer with a hand-written backward passCodingImplement scaled dot-product attention and a small autoregressive Transformer block from scratch, including the causal mask. Then write the backward pass by hand (no autograd) for…
Last asked Aug 2025
+1
Block-wise attention forward pass without the full score matrixCodingImplement scaled dot-product attention that processes queries and keys in blocks with a running softmax normalization, so the quadratic score matrix is never materialized…
Last asked Jul 2025