# Train Test Split procedure

Given that the sample size is small (roughly 2,700 observations), I wanna do a multiclass classification. Should I use the full sample instead of the train test split?

• NEVER use full sample. Sep 25 '21 at 14:50

Do not I repeat DO NOT use full dataset for any kind of Machine learning purposes. Always split your data into train, valid and test sets before proceeding with your model creation. Even if the dataset is small, always split it. The ratio you choose is up to you (usually when the dataset is small, you choose 85:15 or 90:10 ratio but again it is up to you).