find-your-level

Find your level

You are doing a placement quiz for AI Engineering from Scratch curriculum (20 stages, 523 lessons). Your task is to find out where the learner should start so they can skip material they already know and begin exactly where it is challenging. Works for any agent.

Quiz structure

There are 5 knowledge areas, with 2 questions per area, for a total of 10 questions. Each round presents 2 questions (one round per area). After the learner answers both questions in a round, score that area before moving on to the next round.

Scoring

Each question is worth 1 point (0 = wrong or blank, 1 = correct). Each area scores 0-2. The total score ranges from 0-10.

Conducting the quiz

Give the learner a brief greeting first, then immediately move to Round 1. When the environment has a structured question/option tool, use it for every question; otherwise present lettered options in plain text and wait for a reply. After each round, first tell the learner the score for that area (for example, "Math and Statistics: 2/2") before moving on to the next round. Keep explanations brief. Leave all answer explanations until the end.

Answer isolation

The answer key is intentionally stored outside the main body of this quiz at references/answer-key.md. Do not read that reference until the learner has submitted answers for the current round. After that, read only the current round's answer key, grade it, and keep explanations private until all five rounds are complete. Do not preload later rounds.

Never put real answer letters, possible answers, or answer distributions in the reply format examples. In plain text, strictly use neutral prompts:请按此格式回复：Q1: <letter>, Q2: <letter>。Replace the current question number, but keep both values as <letter>。

Round 1 — Math and Statistics

Q1. There are two vectors, a = [1, 2, 3] and b = [4, 5, 6]. What is their dot product?

A) 32
B) 21
C) 15
D) 27

Q2. A fair coin is tossed 3 times. What is the probability of exactly 2 heads?

A) 1/4
B) 1/2
C) 1/8
D) 3/8

Round 2 — Classical Machine Learning

Q3. In a classification task with 90% negative samples and 10% positive samples, a model predicts all samples as the negative class. What is its accuracy?

A) 50%
B) 90%
C) 10%
D) 0%

Q4. Which of the following is a hyperparameter of Random Forest?

A) The learned split thresholds
B) The leaf node prediction values
C) The number of trees
D) The Gini impurity at each node

Round 3 — Deep Learning

Q5. In backpropagation, what does the chain rule compute?

A) The gradient of the loss for each trainable weight
B) The optimal learning rate for the current optimizer
C) The exact value of the number of layers the network needs
D) The batch size used at each training step

Q6. What problem do residual connections (skip connections) in ResNet mainly solve?

A) Poor generalization on small training datasets
B) Slow loading of batches from persistent storage
C) High activation memory during model inference
D) Gradient flow being too weak in very deep networks

Round 4 — NLP and Transformers

Q7. In the Transformer architecture, what does the attention mechanism compute between?

A) Pixels and labels
B) Only Encoder and Decoder
C) Queries, Keys, and Values
D) Embeddings and positions only

Q8. What is the main benefit of fine-tuning a large language model using LoRA (Low-Rank Adaptation)?

A) Retrain every parameter in the base model starting from fresh random initialization
B) Freeze the base model weights and train only low-rank adapters
C) No need for labeled examples or task-specific training data
D) Duplicate model layers to increase adaptation capacity

Round 5 — Applied AI

Q9. In a RAG (Retrieval-Augmented Generation) system, what happens before the LLM generates an answer?

A) Retrieve relevant documents and add them to the model prompt
B) Fully retrain the model for the user's current question
C) The user selects all context passages before each model request
D) The model searches only pretrained parameter values

Q10. In a multi-agent system, what is the primary responsibility of the "coordinator" or "orchestrator" agent?

A) Replace all specialized agents with one general-purpose model
B) Dispatch tasks, route messages, and coordinate other agents
C) Maximize token usage in every agent interaction
D) Prepare an identical backup model for system failures

After all five rounds are complete

Show the domain breakdown and total score:

数学与统计：           X/2
经典机器学习：         X/2
深度学习：             X/2
NLP 与 Transformer：   X/2
应用 AI：              X/2
----------------------------
总分：                 X/10

Score to starting point mapping

Total score	Starting point	Meaning
0-3	Stage 1: Math foundations	Start from the basics
4-5	Stage 3: Deep learning core	Already have math and ML foundations
6-7	Stage 7: Transformer deep dive	Already understand DL, time to learn transformers
8-9	Stage 11: LLM engineering	Solid foundations, can go directly into LLM applications
10	Stage 14: Agent engineering	You know it all, start building agents

Personalized learning path

After revealing the starting point, generate a markdown table covering all 20 stages. Use the score to determine each stage's status. Stages before the starting point are marked as Skip(learner has already mastered the material); the starting point and after are marked as Do. If the learner scores 1/2 in the domain of a corresponding skippable stage, that stage is marked as Review instead of Skip。

Domain-to-stage mapping for detecting review items:

Math and statistics (1/2) -> Mark stage 1 as Review
Classical machine learning (1/2) -> Mark stage 2 as Review
Deep learning (1/2) -> Mark stage 3 as Review
NLP and Transformers (1/2) -> Mark stages 5 and 7 as Review
Applied AI (1/2) -> Mark stage 14 as Review

Read time estimates from ROADMAP.md (the canonical single source of truth). Each stage heading includes the estimated number of hours in the format (~N hours) . Parse these values; do not use hardcoded numbers, so that the learning path always stays up to date with the roadmap estimates. If the repository is not cloned locally, fetch it from the following address: https://raw.githubusercontent.com/fancyboi999/ai-engineering-from-scratch-zh/main/ROADMAP.md。

Output format

Generate the table in this format:

| Phase | Name | Status | Est. Hours |
|-------|------|--------|------------|
| 0 | Setup & Tooling | Skip | -- |
| 1 | Math Foundations | Review | 30 |
| 2 | ML Fundamentals | Skip | -- |
| 3 | Deep Learning Core | Do | 20 |
| ... | ... | ... | ... |

Table rules:

Skip Stage hours are shown as --(not counted in the total).
Review Stages show full hours (the learner should review quickly).
Do The stage shows full hours.
Stage 0 (environment setup and tools) is always Skipregardless of score (it is tool setup, not knowledge).
Summarize Review and Do stage hours, and display the total at the bottom.

Add a sentence with the total duration after the table:你的个性化路径：约 X 小时，覆盖 Y 个阶段。

Then give a brief recommendation: which stage to start from, and what to focus on first based on the weakest areas.

Finally, give the next step:/start-learning will save this assessment to the persistent LEARNING.md learning plan, and/learn will begin interactively teaching the first lesson.