Datasets:
corpus.json doesn't contain all the txt files
Hi,
It seems that the corpus.json doesn't contain all the txt files, for example cmn-cmn-cu89s.txt. For language cmn it only contains cmn-cmn-cu89t.txt but three versions of txt are listed in your github repository. Not sure about what is the problem, maybe the current version is outdated?
Many thanks
Yes, the corpus has been updated with some additional editions since the script to load to Hugging Face was last run. We're looking at revising how we've got things structured to allow users finer-grained control over which things they select and load, so we've held off on loading all of the updates. If there's a need for the new versions sooner, we can try running the update script again.
Yes, the corpus has been updated with some additional editions since the script to load to Hugging Face was last run. We're looking at revising how we've got things structured to allow users finer-grained control over which things they select and load, so we've held off on loading all of the updates. If there's a need for the new versions sooner, we can try running the update script again.
Thanks for your reply! I just checked the corpus.json and actually cmn-cmn-cu89t.txt is inside. However, when I do language=["cmn"], pair="single", cmn-cmn-cu89t.txt doesn't appear. And when I do language=["cmn"], pair="range", cmn-cmn-cu89t.txt appears but the total number of data piece is only 65. I did not understand well the usage of pair, could you please elaborate on that?
Ah, I think part of the problem is that we originally designed this for translation, so it expects at least a pair of languages, not one language. If you want to get all copies of cmn, you might pair it with a language that has every verse, but only one version (Somali "som" should work). Pair is designed to handle the differences in versification across versions. Some Bibles might translate John 1:1-3 as one block, for example, while others translate each verse separately. So, let's say you have only one translation in each language you're looking at and they handle these differently. If you want to only consider verses that are translated individually, you'd use "single" which completely ignores longer ranges (which would mean John 1:1-3 is omitted). "range" will join the three verses together on the one side so that they all match to the longest common range (so ... John 1:1-3 is included in this example). The default (if you don't specify a value for "range") will give you both. This isn't a problem if you have one version in each language. But if you have multiple versions with different versifications, the single ones will be represented twice. Once in the verse to verse pairings, and then again as it is paired to a range. Now ... this won't cause a problem with train, validation, and test sets, as those are designed to make sure they don't contain duplicates. But, it might give you problems with bias as the content isn't deduped.
Ah, I think part of the problem is that we originally designed this for translation, so it expects at least a pair of languages, not one language. If you want to get all copies of cmn, you might pair it with a language that has every verse, but only one version (Somali "som" should work). Pair is designed to handle the differences in versification across versions. Some Bibles might translate John 1:1-3 as one block, for example, while others translate each verse separately. So, let's say you have only one translation in each language you're looking at and they handle these differently. If you want to only consider verses that are translated individually, you'd use "single" which completely ignores longer ranges (which would mean John 1:1-3 is omitted). "range" will join the three verses together on the one side so that they all match to the longest common range (so ... John 1:1-3 is included in this example). The default (if you don't specify a value for "range") will give you both. This isn't a problem if you have one version in each language. But if you have multiple versions with different versifications, the single ones will be represented twice. Once in the verse to verse pairings, and then again as it is paired to a range. Now ... this won't cause a problem with train, validation, and test sets, as those are designed to make sure they don't contain duplicates. But, it might give you problems with bias as the content isn't deduped.
Thanks a lot for your help! I think I have understood, though when I pair cmn with som it still gives me only one version of cmn... I did language=["cmn", "som"], pair="single".
Well, I've updated it now to include everything in the GitHub. I did a test with "cmn" and it was returning three versions. You may need to force Hugging Face to retrieve the new version, but it should work.