The experimental results have a large gap with the one in README

[Sun-Happy-YKX]

I can not get the corresponding result in readme, and does anyone else has once get the corresponding result(that is, BLUE ≈ 26)

I wonder if author's language pack version used is different from mine, which leads to a large gap between the experimental results and the one in README .

[tanjeffreyz]

Are you evaluating the saved model without re-training? This could be happening because the code doesn't seem to be seeded using torch.manual_seed, so the token embeddings are randomized every time you load the model.