ML: Model Evaluation & Selection

Question 1

What is the bias–variance trade-off?

Answer 1

Bias = error from oversimplification. Introduced by approximating real-world problems with an overly simplistic model. Wrong assumptions about data.

Variance = error from oversensitivity. Introduced by a model’s sensitivity to small noise in training data. Model too complex and overfits training data.

Increasing complexity ↓ bias but ↑ variance.

Question 2

Why does overfitting occur?

Answer 2

Model learns noise or spurious patterns due to:

High complexity

Too few examples

Poor regularization

Data leakage

Question 3

What is the difference between training error and generalization error?

Answer 3

Training error = performance on training data.

Generalization error = expected performance on unseen data.

Question 4

What is cross-validation and why is it used?

Answer 4

Resampling method to estimate out-of-sample performance.

Reduces variance and avoids relying on a single validation split.

Question 5

Explain k-fold vs stratified k-fold CV.

Answer 5

Stratified preserves class proportions → better stability in classification tasks.

Question 6

What is regularization? Give examples.

Answer 6

Techniques to reduce overfitting by penalizing complexity:

L1 (sparsity)

L2 (weight shrinkage)

Dropout

Data augmentation

Early stopping

Question 7

How does L1 differ from L2 regularization?

Answer 7

L1: promotes sparse weights; feature selection

L2: distributes weights; smoother optimization; prevents large weights

Question 8

What is the purpose of a validation set?

Answer 8

Hyperparameter tuning and model selection without contaminating the test set.

Question 9

What is data leakage, and how do you prevent it?

Answer 9

Using information from the test/validation set during training.

Prevent by applying preprocessing inside pipelines after splitting data.

Question 10

How do you detect overfitting in practice?

Answer 10

Training loss ↓, validation loss ↑

Gap between training and validation metrics

Poor performance shift after deployment

Question 11

Why is early stopping a form of regularization?

Answer 11

Stops training before the model fits noise; limits effective complexity.

Question 12

When would AUC be preferred over accuracy?

Answer 12

Class imbalance

Different decision thresholds

When ranking performance matters

Question 13

What evaluation metric do you use for class imbalance?

Answer 13

F1-score

Balanced accuracy

ROC-AUC

Precision–Recall AUC

Question 14

Why do ML engineers care about calibration?

Answer 14

Probabilities should reflect true likelihood → critical for risk-sensitive tasks (credit scoring, medical decisions, agentic planning).

Question 15

How do you choose between a simpler and more complex model?

Answer 15

Trade-off accuracy vs interpretability, training cost, robustness, risk of overfitting, and deployment constraints.

Question 16

What is hyperparameter search, and what are common strategies?

Answer 16

Exploration of config space:

Grid search

Random search

Bayesian optimization

Hyperband / ASHA

Population-based training

Question 17

Why does dropout reduce overfitting?

Answer 17

Forces networks to not rely on specific neurons; ensemble averaging effect.

Question 17

Why is validation loss often noisier than training loss?

Answer 17

Validation set is smaller; no gradient smoothing; stochasticity in regularization (dropout, augmentation).

Question 18

How does batch normalization affect model evaluation?

Answer 18

Different behavior in train vs eval mode (uses running averages in eval).

Incorrect mode → inflated or broken metrics.

Question 19

What is the impact of label noise on model selection?

Answer 19

Increases variance

Causes overfitting

Makes validation metrics unreliable
Need robust losses or cleaning strategies.

Question 20

What is the double descent phenomenon?

Answer 20

Deep models can generalize well even far past the interpolation threshold where classical bias–variance predicts poor performance.

Question 21

Why is the test set used exactly once?

Answer 21

Repeated evaluation leaks information into training → overly optimistic performance.

Question 22

What is the difference between global and local interpretability?

Answer 22

Global: model-wide behavior (feature importance, coefficients)

Local: individual prediction explanations (SHAP, LIME)

Question 23

How does SHAP compute feature attributions?

Answer 23

Uses Shapley values from cooperative game theory → average marginal contribution of each feature.

Question 24

Why is regularization related to interpretability?

Answer 24

L1 creates sparse models; discourages overly complex decision boundaries.

Question 25

Name 3 reasons interpretability matters in ML + agentic systems.

Answer 25

Debugging and trust Safety and compliance Detecting hallucinations or unsafe agent actions

Question 26

What is permutation feature importance?

Answer 26

Measure drop in performance after randomly permuting one feature → assesses sensitivity.

Question 27

Why might SHAP or LIME give misleading results?

Answer 27

Correlated features Extrapolation outside data manifold Instability for small perturbations Model interactions not captured well

Question 28

How do we evaluate an LLM-based agent vs a static ML model?

Answer 28

Agents require behavioral evaluation, including: Task success rate Consistency, robustness Tool-use accuracy Safety constraints Planning reliability Hallucination rates

Question 29

Why is standard accuracy insufficient for evaluating agentic systems?

Answer 29

Agents generate sequences of actions, not single outputs → must measure process quality, not just final predictions.

Question 30

What are key evaluation metrics for modern LLM systems?

Answer 30

Response correctness Faithfulness Reasoning depth Retrieval precision/recall Latency and token efficiency Tool invocation success

Question 31

How does bias–variance relate to LLM fine-tuning?

Answer 31

Overly specialized fine-tuning → low bias, high variance → poor generalization LoRA rank controls effective model complexity Needs proper validation tasks

Question 32

How do we evaluate retrieval systems in agent memory?

Answer 32

Recall@k MRR (Mean Reciprocal Rank) Embedding drift over time Query latency Relevance consistency

Question 33

What is reward hacking in agent systems and why is it a model selection issue?

Answer 33

The agent exploits loopholes in evaluation or reward signals. Model selection must detect gaming behavior rather than relying solely on numeric metrics.

Question 34

Why does interpretability matter more for autonomous agents?

Answer 34

Agents make multi-step decisions → need visibility into: Planning chain Memory usage Tool routing Risk of hallucination-driven actions

Question 35

What is “offline evaluation” for LLM agents, and why is it difficult?

Answer 35

Evaluating agents without executing live actions (e.g., simulated tools). Difficult because behavior depends heavily on environment feedback.

Question 36

What is precision?

Answer 36

Fraction of predicted positives that are actually positive. Precision = 𝑇𝑃/ (𝑇𝑃 + 𝐹𝑃) ​

Question 37

What is recall?

Answer 37

Fraction of actual positives that were correctly identified. Recall = 𝑇𝑃/ (𝑇𝑃 + 𝐹𝑁)​

Question 38

When is high precision more important than recall?

Answer 38

When the cost of a false positive is high, e.g.: Fraud detection reviewing bank accounts Medical test with harmful interventions Spam filters blocking important emails

Question 39

When is high recall more important than precision?

Answer 39

When missing positive cases is expensive/dangerous, e.g.: Cancer screening Safety anomaly detection Search systems (don’t miss relevant documents)

Question 40

What is the F1-score and when is it preferred?

Answer 40

Harmonic mean of precision and recall. Used when: Class imbalance exists Precision and recall are both important

Question 41

What is Precision@K?

Answer 41

Fraction of top-K results that are relevant. Used in ranking, retrieval, and recommender systems.

Question 42

Why is accuracy a poor metric for imbalanced datasets?

Answer 42

Majority class dominates the score; a naive model can achieve high accuracy without learning anything about minority classes.

Question 43

What is the goal of retrieval evaluation in a RAG pipeline?

Answer 43

To measure how well retrieved documents support answering the query.

Question 44

What metrics are used to evaluate retrieval quality?

Answer 44

Recall@k → proportion of relevant docs retrieved Precision@k nDCG (discounted cumulative gain) MRR (mean reciprocal rank) Embedding similarity drift (advanced)

Question 45

When is Recall@k more important than Precision@k in RAG?

Answer 45

When coverage matters—LLM can filter noise but cannot hallucinate missing facts. E.g., fact retrieval, medical Q&A, legal reasoning.

Question 46

When is Precision@k more important?

Answer 46

When poor docs confuse or mislead the LLM → reduces generation quality. E.g., domain-specific assistive tools.

Question 47

How do you detect retrieval issues in a RAG system?

Answer 47

Ground-truth retrieval evaluation Similarity histograms Retrieval latency + recall Inspecting LLM output for hallucinations caused by retrieval gaps

Question 48

What metrics evaluate the generation stage?

Answer 48

Faithfulness (does it stick to retrieved doc?) Correctness / factuality Relevance Context utilisation Token-level overlap metrics (BLEU, ROUGE) LLM-as-a-judge metrics (modern)

Question 49

Why are BLEU/ROUGE not enough for RAG evaluation?

Answer 49

They measure surface similarity → fail on: Paraphrasing Correct answers with different wording Reasoning tasks

Question 50

How do you measure hallucinations in a RAG system?

Answer 50

Compare output against retrieved docs (faithfulness) Use LLM-judge scoring Entropy or perplexity-based heuristics Retrieval mismatch detection

Question 51

How do you evaluate end-to-end performance of a RAG system?

Answer 51

Combination of: Retrieval recall Generation correctness Faithfulness to retrieved content Latency + cost metrics (tokens, retrieval time)

Question 52

Why is context compression important for RAG evaluation?

Answer 52

Long contexts dilute relevant info → reduces faithfulness and raises hallucination risk. Evaluations should measure how well context is used.

Question 53

What are failure modes unique to RAG systems?

Answer 53

Retrieval misses relevant docs Over-retrieval causes noise Hallucinations even with correct retrieval Query reformulation errors Embedding drift over time