Appearance
Topic Drill 3.1.4: Pretrained Models, Fine-Tuning, and RAG
About this drill
LO AI-3.1.4 (K2): Summarize the use of pretrained models, fine-tuning, and retrieval-augmented generation.
Lesson: Ch 3, section 3.1.4
This drill covers every point in section 3.1.4:
- Why reuse pretrained models (cost and time); fine-tuning with additional task-specific training and less data
- Which layers are tuned; the evaluate-and-repeat loop; task similarity (cat → dog breeds vs. accents; ISTQB® BVA)
- RAG: searchable data sources, enhanced prompt, no change to the model, more precise responses
- Combining RAG and fine-tuning; inherited biases and vulnerabilities, and the need to test
Topic Drill: 3.1.4
Q1. What does the syllabus give as a key benefit of fine-tuning a pretrained neural network instead of training a new model from scratch?
AIt removes the need for testing
BIt requires much less training data and training effort
CIt guarantees that the model is free of bias
DIt works equally well for any new task
Q2. Which statement about WHERE fine-tuning is applied in a neural network is CORRECT?
AFine-tuning must always retrain every layer
BFine-tuning only changes the input layer
CFine-tuning changes the prompt, not the network
DTuning can be applied to the entire network, only to specific layers (typically near the output end), or to additional layers
Q3. An online shop wants its LLM chatbot to answer questions using the CURRENT returns policy, which changes every month. The shop does not want to retrain the model. Which approach fits BEST?
ARetrieval-augmented generation (RAG)
BFine-tuning all layers every month
CTraining a new model from scratch
DReinforcement learning from customer complaints
Q4. Which sequence BEST describes how RAG works, according to the syllabus?
AAdd layers to the network → store documents in the new layers → retrain
BFine-tune the model on documents → deploy → answer the prompt
CTransform data sources into a searchable format → compare them to the subject of the prompt → put relevant documents into an enhanced prompt → pass it to the LLM
DCollect user feedback → reward the model → update its weights
Q5. A pretrained model classifies chest X-rays. Which adaptation is MOST likely to succeed with fine-tuning?
AClassifying bird songs from audio recordings
BDetecting fractures in hand X-rays
CTranslating medical reports from German to English
DPredicting hospital share prices
Q6. A pretrained LLM known to show gender bias is fine-tuned with high-quality data for a recruitment assistant. What should the tester assume?
AThe bias has been removed by the high-quality data
BBias cannot occur in fine-tuned models
CBias only matters if RAG is also used
DBiases in the pretrained model typically carry over, so the new model must be tested for reliable and fair performance in the new task
Q7. Can a pretrained model use RAG and fine-tuning at the same time?
AYes: a pretrained model can use RAG, fine-tuning, or both together to enhance performance
BNo: RAG and fine-tuning are mutually exclusive
COnly if all layers are retrained
DOnly for image models
Q8. Which TWO statements about pretrained models, fine-tuning and RAG are CORRECT?
Select 2 options.
AFine-tuning always retrains all layers of the network
BFine-tuning removes any vulnerabilities inherited from the pretrained model
CWith RAG, no change is made to the pretrained model
DAfter fine-tuning, ML functional performance is evaluated, and further fine-tuning may follow until the acceptance criteria are met
ERAG needs more training data than fine-tuning
Q9. What does the syllabus say about fine-tuning an LLM to apply ISTQB®-defined boundary value analysis?
AIt is impossible, because test techniques cannot be learned
BIt requires training a new LLM from scratch
CIt requires only a small change to the LLM and is readily achievable with good training data
DIt only works with RAG
0 / 9 answered