Talk:Benchmarks

Can we have a standard benchmark dataset and training/test file so that these numbers reported have actual relative value? Currently they do not explain much because the variables are too varied.