Training Dataset Consent and Split Auditor

How to use Training Dataset Consent and Split Auditor

A dataset can meet a size target and still leak across splits or contain missing permission records. Review those checks independently before any external training pilot.

1. Use pseudonymous sample metadata, not private images or identities.
2. Declare purpose/date, reviewed permission records and actual splits.
3. Resolve source-document holds and evaluate any external model separately; no upload/training occurs.

Good to know

• Permission labels are not a legal determination; exact hashes miss near-duplicates.
• No training, image inspection or prediction of output quality.
• Independent local review of supplied records. No account/API access, AI, publication, outreach, source-truth verification, legal certification or income forecast.
• Each field is limited to 8,000 UTF-16 characters and each CSV to 100 data rows with the exact displayed header. Quoted commas/newlines and doubled quotes are accepted. Unsupported controls, unpaired surrogates, duplicate identities and invalid records are refused rather than silently removed. Ordinary decimal inputs have magnitude at most 999,999,999; the smallest supported nonzero magnitude is 0.000000001. Exponents are refused; whole counts use exact integer checks before numeric conversion. Results use exact rational arithmetic, with at most 18 decimal digits and an ellipsis when further fractional digits remain.
• Markdown and CSV are editable review exports. CSV retains unresolved records and apostrophe-prefixes spreadsheet-like formula text; review your real import settings. Tool outputs can contain your entered pseudonymous records, so inspect them before sharing.
• Independent deterministic ToolJovia workflow. Inputs stay in this browser. No vendor account, model call, upload, publication, outreach, payment, saved input or outside service is used. A local reconciliation does not validate source truth, customers, income or legal compliance.