[kaz] Kazakh "small" data is mis-sized

[kylebgorman]

Hello from the future @jkodner05 and team. I noticed that the file part1/suprise_languages/kaz_small.train is the same size (7,000 exemplars) as kaz_large.train in that directory.

1. I assume this is in error. Can you confirm?
2. What is the best way to evaluate/compare to prior results in the lower-resource ("small") setting given this finding? Should I just use kaz_small.train as training data in the lower-resource setting, even though it's not "small" (700 examples) like, say, hye_small.train?

[ivri]

1. You are right, I see that the file indeed contains 7000 samples instead of 700 (the maximum in the "small" setting). Maybe Jordan and Salam (@jkodner05 and @slkh) could also have a look?
2. I would suggest using the same files (for the sake of reproducibility) but would be interesting to see the difference in the model performance if we also train the models on 700 samples

Warm regards,
Kat

[slkh]

Hello Kyle and Kat,

Apologies for the delayed response!

I checked with Jordan, and indeed, this was a mistake.
Like what Kat had suggested and also Kyle's intuition, we should just use kaz_small.train as is for backward compatibility.
It would be nice to fix everything an update the reports, but currently we don't have the bandwidth for that unfortunately!

Happy New Year!

Salam