Good toy parameters?

[bittlingmayer]

For teaching, baselines and proofs of concept, it would be good to have toy parameters that have trained something in, for example, 1 hour on a CPU, on a relatively simple dataset. (We have parallel datasets for language pairs that are similar, or the same language but with different alphabets.) Can be word-level or character-level.

Can you suggest something?

[bittlingmayer]

By the way, some parameters in the example at https://aws.amazon.com/en/blogs/machine-learning/train-neural-machine-translation-models-with-sockeye/ are out of date (for example --attention-type).

Right now it is hard to know which combinations of parameters are valid, and which values are the default (for example rnn).

[mjpost]

You might find it helpful to start with the sequence copy model tutorial. This uses a very small vocabulary of integers and the settings there should work fine.

If you want to update the tutorial (or add another) and issue it as a PR against docs/, that would likely be welcomed!

[bittlingmayer]

That works, thank you!