Well, put your data in a matrix and take the first 2000 data and train your model on it(I am assuming this is a supervised learning). Then you can test your model on the 1000 data and compute your error. This would be in matlab(if your data is row vector).
training_data = first_data(1:2000,:);
test_data = first_data(2001:end,:);
Alternative approach would be to split the data into k-sections and train on the K-1 dataset and test on the what you have left. Doing this repeatedly is helpfully to avoid over-fitting. For much detail read about bias-variance dilemma and cross-validation.