TerminologytermStep 4: Grounding & specializingLegalHRRiskAll

Term: Training data

~6 min read

Estimated time: ~6 min read — for the in-app brief plus opening the primary source.

What this is

Training data is the set of examples a model learned from. Rights, bias, and staleness travel with those examples.

Everyday example

A model fine-tuned on five years of “who we hired” will learn those patterns — including ones you no longer want. The examples are the training data.

Training data is the examples the model learned from — quality and rights travel with those examples.

  • Bias, stale process, and copyright all start here.
  • You rarely see the full set for a foundation model; you can see it for a model you fine-tune.
  • “We trained on the internet” is not a due-diligence answer.
  • For your own models, data lineage is a Legal and Risk file.

Next action: For any model you customize, write down the source, rights, and refresh date of the examples.

What changes in how you lead

How decision rights, process, and ownership should change.

  • Legal reviews training-data rights before a fine-tune, not after a complaint.
  • HR does not train people-decisions on unexamined history.

Compare related ideas

Training data vs Fine-tuning

Training data is the examples. Fine-tuning is the extra training step that uses a new set of examples to shift the model.

Open Fine-tuning

Deep dive

You rarely see the full set for a foundation model; you can control it when you fine-tune.

“We trained on the internet” is not due diligence.

For your own models, data lineage is a Legal and Risk file.

HR does not train people-decisions on unexamined history.

Related terms

Related weekly lessons

terminologydatarights