Machine Learning Engineer question bank
100 questions. Filter by level and skill, track what you've solved, and see how many people get each one right.
- Big gap between train and validation accuracy
- Why keep a separate test set
- Which model needs feature scaling
- Definition of precision
- Symptom of a learning rate that is too high
- Encoding a city feature
- Bias and variance as complexity grows
- Accuracy on imbalanced data
- Shape of a matrix product
- Cosine similarity by hand
- Chain rule on a squared term
- What sampling temperature does
- Standard error of a sample mean
- What logistic regression outputs
- k-nearest neighbours vote
- What k-means minimises
- Variance kept by two components
- Gini impurity of a node
- Why random forests average trees
- ReLU applied to a vector
- Recall from confusion counts
- RMSE from four errors
- Broadcasting a column and a row
- Dropout after eval()
- Why export a model to ONNX
- Why LLMs use subword tokens
- What text embeddings capture
- Standardising a single value
- When early stopping triggers
- Optimizer steps in one epoch
- Expected starting loss, 10 classes
- What retrieval-augmented generation adds
- What data drift means
- What prompt injection is
- Scaling before splitting
- Validating a forecasting model
- Different results every inference run
- Handling class imbalance
- Which regularisation gives sparse weights
- Early layers stop learning
- Probabilities that cannot be trusted
- Millions of product IDs as a feature
- F1 score from a confusion matrix
- Log loss of a poor prediction
- Positive test for a rare condition
- AUC from four scores
- Conv output size with stride 2
- Parameters in a conv layer
- Reading a logistic regression weight
- Gradient of a sum of squares
- Training loop without zero_grad
- Softmax before CrossEntropyLoss
- Shape after a batched matmul
- Rows after a merge on duplicate keys
- Precision at 5 for a ranked list
- NDCG for a three-item ranking
- Effective batch size across GPUs
- Why use learning rate warmup
- Random search versus grid search
- Initialisation for a deep ReLU network
- Raising C in a soft-margin SVM
- Loss turns NaN in an RNN
- Scaling factor in attention
- Trainable parameters in a LoRA adapter
- Memory for 4-bit weights
- RAG misses exact part numbers
- Target encoding without leakage
- Several scans per patient
- Great offline, bad in production
- Why fp16 needs loss scaling but bf16 usually does not
- Model does not fit in GPU memory
- ROC-AUC after calibration
- Better offline metric, worse A/B test
- L2 penalty versus AdamW
- A feature that is too good
- Tuning an HNSW vector index
- KV cache size for one sequence
- Parameters in a multi-head attention block
- view() after a transpose
- Fine-tuning a CNN with batch size 2
- Why Adam divides by 1 - beta^t
- What ZeRO stage 3 shards
- Activation checkpointing trade-off
- Why RoPE encodes relative position
- What FlashAttention changes
- Speculative decoding and output quality
- Why serving engines page the KV cache
- INT8 activations hurt accuracy
- Why scale distillation loss by T squared
- How DPO differs from PPO-based RLHF
- Fine-tuned model forgets general skills
- Filtered vector search returns too few
- Inputs stable but accuracy falls
- Retraining on your own recommendations
- Random ID tops feature importance
- SMOTE applied before cross-validation
- LLM judge prefers the first answer
- Instructions hidden in a retrieved page
- Daily cost of an LLM feature
- Negative R-squared on test data