Reduce monolingual data for en-lt to investigate distillation performance

[gregtatum]

In #771 I tested the effects of reducing the distillation data to understand that expensive part of our pipeline. However, we should do it again for the base student model, as the other one was done for a tiny model too see if there is a difference. Also, I want to test it on a morphologically more complex language like Lithuanian.

[gregtatum]

In #771 @ZJaume commented:

This results seem very interesting to me. I believe the fact that NLLB and Paracrawl are full of redundant and repetitive data has something to do with this. If there is interest in finding a better way to sample, I think n-gram saturation (rank lower the sentences that have a significant portion of the 2-grams or 3-grams already present in the corpus) could be something worth to explore.

After some light searching I found this paper with has an approach we could use if we wanted to go this route, which seems like a reasonable balance of cost vs quality: STACC, OOV Density and N-gram Saturation: Vicomtech’s Participation in the WMT 2018 Shared Task on Parallel Corpus Filtering