queries contains `�`

#2
by Samoed - opened

Hi! I wanted to use this dataset to pretrain a model and found that the quries (titles of web pages) in most subsets contain � (0xFFFD), likely caused by old encoding (e.g., CP1251), and they can’t be recovered. I briefly checked, and the cmn, rus, and arb subsets contain these issues (probably more than 5% of examples), and such queries will degrade model performance.

Samoed changed discussion title from Titles contains `�` to queries contains `�`

Yeah I noticed that as well. I’m updating this with more data this weekend and will look. I assume I had a UTF-8 conversion problem somewhere

Looks like it starts even in my first gathering from CC. It'll take another 2-6 weeks to re-gather it with my academic compute :(. Seems like the problem maybe my usage of jq

Looks like it starts even in my first gathering from CC

I think so too. I tried to debug some examples and get it working only with plain bytes.

Seems like the problem maybe my usage of jq

So you have your script in some repo?

Yeah a private repo. The only way I could figure out how to get this was to basically re-download streaming wise all of CC in parts (since we can’t store it on disk given the massive size in PB) and grep for the ids so I could extract the title. It’s a bit more optimized than that (I can localize the chunks needed so we don’t need every chunk) but not that much more as the ids we want are spread out among CC files.

I just reached out to some friends that are more skilled with CC, maybe there’s a better way. Otherwise I’ll slowly stream all of CC again to collect these which takes a while given their rate limits

think so too. I tried to debug some examples and get it working only with plain bytes.

Unless you’re saying you can recover the information when going to bytes? I could do that

Unless you’re saying you can recover the information when going to bytes

No, this can't be done with current texts, sadly. I tried a lot of methods, but nothing works

Sign up or log in to comment