Synthetic PII Data and Tokenization Quiz
Multiple Choice Questions
What is the model output contract and its limit?
- A) Always generate a correct anonymous document
- B) Every source string is PII
- C) Candidate TYPE/TAB/ORIGINAL rows; evaluate omissions/misclassification separately
- D) Return project membership
Show Answer
Answer: C
Source matching establishes occurrence, not semantic correctness or complete detection.
How are differently spaced names and a preexisting [PERSON_1] handled?
- A) Overwrite every spelling with one mapping value
- B) Use reversible per-spelling tokens and avoid existing marker names
- C) Always restore the old marker as a new name
- D) Arbitrarily shorten the source
Show Answer
Answer: B
Identical type/spelling reuses a token; different spellings get separate tokens mapped to actual source text.
Which statement about splits and validation scope is correct?
- A) There are 1,600/200/400 records; templates/names may be shared across splits
- B) Different hashes prove real-workload generalization
- C) Checksum failure proves official non-assignment
- D) Every PHONE is an officially reserved number
Show Answer
Answer: A
Existing generator-1.0.0 hashes are retained; distinguish synthetic-template tests from real-workload evaluation.
How does corrected leakage evaluation handle masking only Alpha in gold Alpha Beta?
- A) Always report zero because the full string disappeared
- B) Detect an uncovered part of the gold source span
- C) Search only placeholder names
- D) Override leakage to zero after a successful round trip
Show Answer
Answer: B
It checks complete source-span coverage by the union of replacements. Restoration and complete detection are separate metrics.
What does the 2,200-record oracle check using answer TSV establish?
- A) Measured fine-tuned model F1
- B) GPU throughput
- C) Evaluator/restoration/hash sanity, not model-output quality
- D) Guaranteed detection of all real PII
Show Answer
Answer: C
Fifty local tests and the oracle check passed; no GPU training or fine-tuned evaluation ran.