R-Net

Why we included this project

R-Net reproduces the Microsoft Research reading-comprehension architecture in TensorFlow, tuned for the SQuAD benchmark. The repository walks through the whole workflow: downloading and preprocessing SQuAD with GloVe embeddings, training, and running the official evaluation script, which makes it a practical reference for seeing how self-matching attention and answer-pointer decoding fit together in real code. It also documents engineering choices that go beyond the paper, including scaled multiplicative attention, variational dropout, and CudnnGRU bucketing to cut training time, with the F1 trade-off measured. Since it lands almost exactly on the paper's published EM and F1 scores, it works well as a baseline when you are comparing newer reading-comprehension models or checking your own implementation against a known result.