A dataset split separates examples for model fitting from examples reserved for evaluation. In a toxic-comment classifier, an 80/20 split provides substantial training data while preserving a held-out check on performance. section 9.3.4
Toxic-comment classification
Keep comment text in X and six toxicity labels in y, then use a fixed random seed to create reproducible training and evaluation subsets. The resulting data contains 127,656 training rows and 31,915 evaluation rows.
