Training set
Also called Training Set / Train Split
The data the model actually learns from — the examples it sees over and over.
Think of it like
The textbook chapters you study from, as opposed to the exam that tests you on them.
Example
A sentiment model is fit on 80,000 labeled reviews in the training set, leaving the rest for validation and test.
How it actually works
This is the largest split, the material the optimizer minimizes loss over. Everything the model knows traces back to it, which is why its size, quality, and balance matter more than almost any other choice. Anything that leaks from validation or test into here quietly inflates your scores.
For product teams
The quality ceiling of your model is mostly set here — garbage in, garbage out.
For engineers
The split gradients are computed over; guard it against contamination from val/test data.
Related
- Validation set — The honest check the model never learns from.
- Test set — The final untouched exam.
- Pretraining corpus — The huge unlabeled version for base models.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome