#time-series-split

Ordinary $K$-fold cross-validation shuffles away the time order, so for a series of 100 observations you use forward-chaining splits instead (scikit-learn's TimeSeriesSplit with n_splits=4). The data are cut into 5 consecutive blocks of 20; split $i$ trains on all blocks before block $i+1$ and tests on block $i+1$. How many observations does the third split train on?