SageMaker Qwen Platform Architecture Quiz
Multiple Choice Questions
How do model output and Python processing relate in this design?
- A) The model certainly removes all PII
- B) The model emits candidate TSV and code validates/replaces it; missed detection is evaluated separately
- C) A round trip proves anonymity
- D) The model deletes projects
Show Answer
Answer: B
Deterministic replacement does not automatically recover entities the model missed.
How should the diagram and validation record be interpreted?
- A) Both GPU paths were validated
- B) It is a target design; distinguish local tests from historical AWS observations
- C) It shows live residual-project counts
- D) It reports trained F1 and costs
Show Answer
Answer: B
The 2026-09-01 record stopped before training. Current AWS state and GPU outcomes are not proven by local tests.
Do identical model IDs and seeds fully reproduce results across environments?
- A) Always
- B) No; also verify revisions, data hashes, images/dependencies, CUDA/hardware and execution conditions
- C) Only EKS needs a seed
- D) The tokenizer does not matter for QLoRA
Show Answer
Answer: B
Direct package pins are not a full transitive lock, and GPU determinism has additional requirements.
What should be excluded from ordinary MLflow logs/public reports?
- A) Dataset hashes
- B) Reviewed LoRA settings
- C) Raw source/completions, token mappings and presigned URLs
- D) Non-sensitive aggregates
Show Answer
Answer: C
Distinguish necessary private inventory IDs/ARNs from public disclosure, and verify autologging/tracing.
Which statement about governance and model size is correct?
- A) QLoRA always requires a Unified Studio project
- B) Governance is this experiment's chosen procedure; 3.3B active does not represent total model memory
- C) Owner assignment atomically rolls back every partial failure
- D) A three-hour job limit caps all experiment costs
Show Answer
Answer: B
The model card lists 30.5B total/3.3B active. Verify readiness/ownership and actual GPU memory/cost separately.