+ sm model no external features baseline
+ sm model with IDF weights
+ sm model with IDF weights without removing punctuation --> barely better than df/idf (a la Pytorch).
+ sm model with stemming before computing IDF weights
^ all on the TrecQA dataset
