Jargon:Training data
1 min read
Say this at the party
“The pile of examples an AI learns from. Whatever is in that pile, good or bad, shapes everything the AI does.”
In plain English
An AI model starts out knowing nothing. It learns by studying examples: text, photos, recordings, whatever the job needs. That collection is its training data.
For chatbots, the pile is enormous: huge amounts of text gathered from the internet, books and other places.
It's like a diet. If the examples are full of mistakes or one-sided views, the AI picks those up too.
Where it came from
The idea is as old as machine learning itself. Researchers have long split their examples into a training set to learn from and separate sets to test on, so they can check the model hasn't just memorised the answers.
See it in action
A hospital wants AI to spot broken bones on X-rays. It trains the model on thousands of X-rays that doctors have already marked "broken" or "fine". Those marked X-rays are the training data.
What people get wrong
- A chatbot doesn't keep a searchable copy of everything it was trained on. It learns patterns from it, though it can sometimes repeat bits word for word.
- More data isn't automatically better. Messy or biased examples make a worse model, however many there are.
Go deeper
- ArticleWikipediaTraining, validation, and test data setsExplains why examples are split into training and test sets, and what each one is for.
- GuideGoogle for DevelopersDatasets: Data characteristicsGoogle's crash course on what makes data good or bad for training.
- GuideGoogle for DevelopersMachine Learning GlossaryHandy plain definitions of training set and related terms.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.