question about SPM and BPE

[jemesome]

Noob here, running the option2 using docker to install the tornado web server with several language models as listed in the readme here. I ran the dockerfile after cloning from here .

What is the difference between models in OPUS-MT-train and models in the Tatoeba-Challenge ? Looking at the benchmark on test set for for the Spanish to english model, the ES-EN model trained on tatoeba dataset has lower metric compared to the same model in Tatoeba- challenge (Spa-eng).

are these 2 the same model trained on different dataset? is the model based on Tatoeba newer and better compared to the one in OPUS-MT-train as suggested by the benchmark metric?

[jorgtied]

Most of the Tatoeba models are better than the earlier opus-mt models. The training data is a bit cleaner, sometimes also a bit bigger. Models are also sometimes bigger than the original OPUS-MT models (transformer-big instead of transformer-base). But there are different variants and you need to check the model card / README.