AIMentor — AI Expertise Learning Platform

14 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
In the standard decision ladder, which order should you try tools in?
Fine-tuning → RAG → promptingPrompting → RAG → fine-tuningRAG → fine-tuning → promptingFine-tuning first, always
Question 2easy
What does supervised fine-tuning (SFT) primarily change?
The facts stored in the model's memoryThe model's default behavior — format, tone, task styleThe size of the context windowThe tokenizer used by the model
Question 3easy
The clearest one-line rule for RAG vs. fine-tuning is:
RAG is cheaper, fine-tuning is more accurateRAG changes what the model knows; fine-tuning changes how it behavesRAG is for text, fine-tuning is for imagesThey are interchangeable
Question 4medium
A team wants the model to always use their company's current, frequently-changing pricing. Best tool?
Fine-tuning on the price listRAG — retrieve the current prices at query timeFull fine-tuning weeklyLarger context window only
Question 5medium
What is LoRA (Low-Rank Adaptation)?
A method that retrains all of the model's weights from scratchA technique that freezes the base weights and trains small low-rank adapter matrices (often <1% of params)A prompting strategy for few-shot examplesA vector database optimized for embeddings
Question 6medium
Which are advantages of LoRA/PEFT over full fine-tuning? (Select all that apply)
Far lower GPU memory and costSmall adapters you can keep many of and swap per taskGuaranteed to always beat full fine-tuning on qualityCan run on smaller hardware, especially with quantization (QLoRA)
Select all that apply.
Question 7medium
What file format do hosted fine-tuning APIs expect for training data?
A single JSON array of examplesJSONL — one JSON object per lineCSV with a text columnPlain .txt with examples separated by blank lines
Question 8medium
Which most improves fine-tuning results?
Maximizing the number of epochsA few hundred clean, consistent, representative examples over tens of thousands of noisy onesUsing the largest possible base modelRaising the learning rate
Question 9hard
What is 'catastrophic forgetting' in the context of fine-tuning?
The API losing your uploaded training fileOver-training on a narrow task degrading the model's general capabilitiesThe model forgetting the system prompt at inferenceAdapters overwriting each other when swapped
Question 10hard
Which are POOR fits for fine-tuning? (Select all that apply)
Injecting up-to-date facts you'd want a citation forEnforcing a consistent output format the model keeps drifting fromPrivate knowledge that changes weeklyA narrow classification task with labelled examples
Select all that apply.
Question 11hard
You have an expensive model that behaves well on a narrow task and want a cheaper model to imitate it. What is this called?
QuantizationDistillation (fine-tuning the small model on the big model's outputs)Retrieval augmentationPrompt caching
Question 12medium
Why keep a held-out validation split when fine-tuning?
To increase the total training dataTo measure whether the model generalizes rather than memorizes, and detect regressionsBecause the API rejects datasets without oneTo speed up training
Question 13medium
Name the parameter-efficient fine-tuning technique that trains small low-rank adapters while freezing the base weights (acronym is fine).
Question 14hard
In one word: does fine-tuning change a model's knowledge or its behavior?

0 / 14 answered

Fine-Tuning Basics — Final Exam