Measuring Bias of Web-filtered Text Datasets and Bias Propagation Through Training