Implementation guide OWASP AISVS v1.01

Training Data Integrity & Traceability · Training Data Origin & Data Security

Keep an inventory of training data sources

AISVS 1.1.2 Level 2

Verify that an up-to-date inventory is kept of every training-data source, including its origin, responsible party, license, collection method, intended use constraints, and processing history.

OWASP AISVS v1.01, requirement 1.1.2

In plain English

Keep a current list of every source your training data comes from. For each one, record where it came from, who is responsible for it, its licence, how it was collected, what it may be used for and what has been done to it since. When a licence dispute, a deletion request or a poisoning report arrives, this list is how you find out which models are affected.

What good looks like

  • Every dataset used in training, fine-tuning or evaluation has an entry, including scraped, bought, synthetic and user-feedback data.
  • Each entry names a person or team who can answer questions about it.
  • Licence and use restrictions are recorded, and someone checks them before the dataset is used for a new model or purpose.
  • The processing history (filtering, deduplication, labelling, augmentation) is recorded per dataset version.

How teams usually do it

  • A catalogue entry per source using a standard template. "Datasheets for Datasets" and Hugging Face dataset cards are common starting points.
  • A pipeline check that refuses to start a training run if any input dataset has no catalogue entry.
  • Larger teams record lineage automatically, for example with OpenLineage or the lineage features of their data platform, so the processing history writes itself.

Suggested evidence

  • The inventory, with owner, licence and collection method filled in for every entry.
  • The dataset list from a recent training run, matched line by line against the inventory.
  • A record of a licence check carried out before a dataset was reused.

Common mistakes

  • An inventory written for the first model and never touched again.
  • Leaving out the fine-tuning or evaluation set because it was "only a small one".
  • Recording the licence of the download site and missing that the content inside it belongs to someone else.

Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.

Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.