Training data can include text, images, audio, video, code, labels, demonstrations, preferences, rewards, or simulated interactions. Its composition, quality, provenance, duplication, and weighting influence what the model learns and where it fails.
Large models combine information from many examples, so one output rarely has a single simple source. Dataset documentation, filtering, attribution tools, and targeted evaluations help teams investigate whether problematic examples contributed to unexpected behavior.
ELI5
Training data is the collection of examples an AI model learns from. Those examples shape the patterns, facts, styles, and mistakes the model is likely to reproduce later.
For example, a model trained to recognize plants needs many correctly labeled plant images from varied conditions. If its examples include only bright close-up photos, it may struggle with the same plants in shade or at a distance.
What forms can training data take?
It can include text, images, video, audio, code, labels, demonstrations, preferences, feedback, rewards, synthetic examples, and interaction records.
Why does training data quality matter?
Errors, bias, duplication, missing coverage, unsafe content, and poor provenance can shape model behavior and make failures harder to diagnose.
