Pre-training data review checklist
(Fine-Tune Your Own Model with Unsloth - starter kit)

Read every pair end to end. The PII scan catches patterns, not names. A human
still reads every pair before it trains.

1. I read every question and every answer in the file, start to finish.
2. No member names appear in any question or answer.
3. No email addresses appear anywhere in the file.
4. No phone numbers appear anywhere in the file.
5. No postal addresses appear anywhere in the file.
6. No account numbers, member IDs, or ID numbers (SSN, EIN, license numbers)
   appear anywhere in the file.
7. De-identification rule: if a real exchange would let someone identify a
   member, I rewrote it as a generic version or I dropped the pair.
8. No internal-only policy text that must never be quoted back to members.
9. I held out 10 pairs for evaluation and they are NOT in the training file.
10. I re-ran the PII scan after my last edit, and I reviewed every hit by hand.
11. I saved the final reviewed file with a version number, so I can prove
    exactly which pairs trained the model I keep.
12. I confirmed the publishability decision: if the file holds sensitive data,
    training happens on a local GPU only, never in a cloud notebook.