PCEI Exam Prep — Mock Exam & Review

• Know the PCEI format and topic spread.
• Take a timed 10-question mock.
• Score yourself and find weak areas.
• Follow a one-week plan to the real exam.

PCEI (Python Certified Entry-Level AI/ML practitioner)

Format     : multiple-choice (check the latest spec for count/time)
Pass mark  : ~70%
Topics     :
  - ML concepts: supervised/unsupervised, classification/regression
  - the train/test/evaluate workflow + metrics
  - overfitting, cross-validation
  - NumPy / pandas for ML data prep
  - neural network basics (layers, activations, training)
  - NLP basics (tokens, TF-IDF)
  - responsible AI (bias, fairness)

Set a timer for 15 minutes. No notes.

Q1.

A model scores 99% on training data and 62% on test data. This is most likely:

(A) underfitting · (B) overfitting · (C) data leakage · (D) good generalisation

Q2.

Predicting a house's price in RM is a:

(A) classification task · (B) clustering task · (C) regression task · (D) reinforcement task

Q3.

Which metric is misleading when 99% of samples are one class?

(A) precision · (B) recall · (C) accuracy · (D) F1

Q4.

Why do we split data into train and test sets?

(A) to save memory · (B) to measure generalisation to unseen data · (C) to speed up training · (D) to remove outliers

Q5.

K-Means is an example of:

(A) supervised classification · (B) supervised regression · (C) unsupervised clustering · (D) reinforcement learning

Q6.

Which activation is standard for hidden layers in a neural net?

(A) softmax · (B) ReLU · (C) sigmoid on every layer · (D) none

Q7.

KNN performs poorly until you do what to the features?

(A) one-hot encode · (B) scale them · (C) shuffle them · (D) remove the labels

Q8.

TF-IDF gives a LOW weight to a word that:

(A) appears in almost every document · (B) appears in one document only · (C) is very long · (D) is capitalised

Q9.

A hiring model trained on biased historical data will most likely:

(A) automatically correct the bias · (B) reproduce the bias · (C) ignore the data · (D) become more accurate for everyone

Q10.

Cross-validation is better than a single train/test split because it:

(A) trains faster · (B) uses less data · (C) gives a more reliable average score over several splits · (D) removes the need for a test set

Q1.  (B) overfitting     huge train-test gap = memorised, didn't generalise
Q2.  (C) regression      a continuous number output
Q3.  (C) accuracy        "always predict majority" scores 99% yet is useless
Q4.  (B) generalisation   the test set estimates real-world performance
Q5.  (C) clustering       finds groups with no labels
Q6.  (B) ReLU            fast, no vanishing gradient for positives
Q7.  (B) scale them      KNN is distance-based; unscaled features dominate
Q8.  (A) appears everywhere  low IDF → low TF-IDF weight
Q9.  (B) reproduce the bias  models copy patterns, including unfair ones
Q10. (C) reliable average    averages over folds, less luck-dependent

Scoring

• 9-10: exam-ready.
• 7-8: pass-likely; review your wrong topics.
• 5-6: another week; redo Try-Its from the relevant lessons.
• ≤4: re-walk Lessons 9-26 slowly — the core ML + NN concepts.

For each missed question, do the matching drill:

• Overfitting / CV (Q1, Q10) → redo Lessons 10 & 25.
• Task types (Q2, Q5) → redo Lessons 1, 3, 17.
• Metrics (Q3) → redo Lesson 11.
• Train/test (Q4) → redo Lesson 9.
• Neural nets (Q6) → redo Lessons 21-22.
• Scaling / KNN (Q7) → redo Lesson 12.
• NLP (Q8) → redo Lesson 35.
• Ethics (Q9) → redo Lesson 46.

Re-type the worked examples from memory — concepts stick when your fingers do them, not just your eyes.

Level 5 complete. You can:

• Frame a problem as ML; prep data with NumPy & pandas.
• Train, cross-validate and honestly evaluate classic ML models.
• Build and train neural nets in Keras; fight overfitting.
• Classify images (CNN + transfer learning) and do computer vision with OpenCV.
• Process text, build NLP classifiers, and run sentiment analysis.
• Call LLM APIs with prompting, tools, streaming, and RAG; ship an AI web app.
• Reason about bias, fairness and responsibility — and document it in a model card.

That is a genuine entry-level AI/ML skill set. Level 6 (Testing & Quality) turns you from someone who builds models into an engineer who ships reliable software.