Write your response in your own notes, then compare it with the sample answer. This practice is separate from the scored exam and does not affect your certificate.
Practice 1
A colleague is deciding whether to try a local model for a task involving material already on their computer. Write a short decision note naming one reason a local run may suit that task and one trade-off they should still consider.
Check your answer against these points:
- Names a task-related reason for considering local use.
- Names a relevant limitation or resource trade-off.
- Avoids treating local use as an automatic privacy or capability certainty.
Compare with a sample answer
A local model may suit the task if working offline or keeping the material off another server matters. The colleague should also consider whether the model’s capability is sufficient for the task and the computer has enough memory. This is a comparison, not a claim that local processing is automatically private or best.
Ungraded paper exercise based on the chapter’s local-versus-hosted comparison. It does not assess a real setup or its network configuration.
Practice 2
Create a tool-choice checklist for a first local-model attempt. Compare the kinds of user each of Ollama, LM Studio and llama.cpp may suit, then state why learning one does not require mastering all three at once.
Check your answer against these points:
- Describes the interface or workflow distinction in the source.
- Explains that one tool is enough to begin.
- Treats exact commands and features as subject to current documentation.
Compare with a sample answer
A concise checklist can distinguish a graphical browser and downloader, a command-line workflow or local API, and a more hands-on engine for unusual setups. A learner can pick one approach first because common concepts such as GGUF and quantisation can carry across tools, while interfaces differ. Exact features and commands should be checked against current documentation.
Ungraded planning exercise based on the tool comparison chapter, not a current feature comparison or installation test.
Practice 3
A model is described as a mixture of experts. In your own words, explain why its total parameters matter for memory while its active parameters help explain per-token computation.
Check your answer against these points:
- Separates total parameters from active parameters.
- Connects total parameters to memory.
- Connects the active subset to per-token computation.
Compare with a sample answer
All parameters must be held in memory because any expert may be used next. Only some experts run for a given token, so the active count helps explain computation per token. Memory and speed therefore follow different counts in this example.
Ungraded written explanation based on the chapter’s MoE callout. No benchmark or hardware performance is being assessed.
Practice 4
Draft a model-card inspection note for one candidate you find. List the licence, base model if it is a fine-tune, available GGUF files, intended use and stated limitations. Add one question you would resolve before relying on the candidate.
Check your answer against these points:
- Includes the licence and treats custom terms as requiring reading.
- Includes base model where relevant and format availability.
- Records intended use and limitations plus an unresolved question.
Compare with a sample answer
A useful note records the candidate’s licence and reads any custom terms, identifies a base model where relevant, checks whether GGUF files exist, and captures intended use and limitations. An unresolved licence or limitation is a reason to pause and investigate, not infer permission.
Ungraded checklist-writing practice using the chapter’s model-card guidance. It is not legal advice or a review of current licensing terms.
Practice 5
Use the troubleshooting table to write a short response for two symptoms: a model will not load, and replies are garbled or ignore the system prompt. For each, name a likely cause and a first suggested check or adjustment.
Check your answer against these points:
- Pairs each symptom with a cause stated in the table.
- Suggests the matching first response for each symptom.
- Keeps the troubleshooting advice within the chapter’s scope.
Compare with a sample answer
For a load failure, memory may be insufficient, so try a smaller model or lower quantisation and close memory-heavy applications. For garbled replies, check whether the chat template or prompt format is wrong and use the tool’s default chat mode for that model rather than raw completion mode.
Ungraded paper exercise based on the troubleshooting table. It does not establish that any specific device or model has these faults.
Practice 6
Plan a fair comparison between two quantisations of the same model. State what you would keep constant, what observations you would record, and what limit to put on your conclusion.
Check your answer against these points:
- Keeps the model and prompts comparable.
- Records answer differences and speed differences.
- Limits conclusions to the prompts and setup tested.
Compare with a sample answer
Use the same three real questions on both versions. Record whether answer quality appears different for those tasks and compare time to first word and streaming speed. The result describes this small prompt set and setup, not every task or learner.
Ungraded experiment-planning exercise from the comparison activity in chapter 4. It is not a runtime test.