Synthetic PII Data and Augmentation Quiz
Last Updated: September 15, 2026
Multiple-choice questions
What is the correct order for paraphrases and OCR variants of the same source?
- A) Generate every variant, then split random record IDs
- B) Assign source families to splits, then augment training only
- C) Assign each variant to a different split
- D) Copy test failures into training first
Show Answer
Answer: B
All descendants must remain with their source family to prevent cross-split source/variant leakage.
How many training records result from 24 families with one base and two variants each?
- A) 24
- B) 48
- C) 72
- D) 120
Show Answer
Answer: C
24 × 3 = 72. Each of the eight validation/test families retains only its base.
What is required when substituting a synthetic name?
- A) Change only the record ID
- B) Change the source but keep TSV
- C) Change TSV but keep the source
- D) Update source, entities, and TSV together and record lineage
Show Answer
Answer: D
Regenerate gold against the transformed source. String presence alone still does not prove semantic correctness or complete annotation.
What does the lab establish through family and exact-source separation?
- A) Families and identical NFC sources do not cross splits
- B) All people and identifiers are split-disjoint
- C) All templates are split-disjoint
- D) Real-world generalization is validated
Show Answer
Answer: A
Some names, templates, and domains are shared. Entity/template/domain holdouts need separate design.
After transforming a negative document containing only a switchboard excluded by this policy, what should its target be?
- A) Add the switchboard as PHONE
- B) Keep entities and target_tsv empty
- C) Add an UNKNOWN type
- D) Add all numbers as ACCOUNT
Show Answer
Answer: B
No-entity examples matter. Define negatives for the domain and verify EOS supervision for empty answers in actual training.
Why can this lab reject an omitted label even after file hashes are recomputed?
- A) Hashes authenticate the producer
- B) It finds every PII item in arbitrary customer documents
- C) It also rechecks content against the seed and fixed family recipe
- D) It automatically measures model F1
Show Answer
Answer: C
This checks reproducibility of the fixed synthetic exercise, not arbitrary annotations or producer identity.
What should be held constant or recorded when comparing original and augmented training?
- A) Select the best result by repeatedly inspecting test
- B) Compare only file counts
- C) Compare only generation success
- D) Use the same validation/decoding and record example/token/step budgets
Show Answer
Answer: D
Equal steps can still expose different lengths and amounts of data. Use validation for candidate selection.
Why does summing historical pairwise overlaps 22, 36, and 9 differ from 51 unique cross-split source groups?
- A) Sources in all three splits appear in multiple pairwise intersections
- B) NFC sources always identify different documents
- C) All record IDs are identical
- D) 51 is the trained model error count
Show Answer
Answer: A
The sum of pairwise intersections is not the unique-source-group count. These are CPU data observations, not model-performance numbers.