3 ms·

I saved a full snapshot of Wikipedia (and Stack Overflow) in the weeks before ChatGPT launched, and every day I'm more glad that I did. They will become the Low

I saved a full snapshot of Wikipedia (and Stack Overflow) in the weeks before ChatGPT launched, and every day I'm more glad that I did. They will become the Low Background Steel of text.

jakderrida 3y ago
The thing is that the Wiki mods will need to be more diligent with uncited things. I also see 2 massive opportunities here. First is that they can have agents check the cited source and verify whether the source backs up what's said to a reasonable degree. Second opportunity is fitting in things only found in other language Wikis that either be incorporated into the english one or help introduce new articles. Believe it or not, LLMs can't generate english answers for things answered only in Russian (or any language) in the training data.
- groceryheist 3y ago
  > First is that they can have agents check the cited source and verify whether the source backs up what's said to a reasonable degree. This is a hard and tmk unsolved NLP/IR problem, and data access is an issue. > Second opportunity is fitting in things only found in other language Wikis that either be incorporated into the english one or help introduce new articles. This has been attempted via machine translation in the past, and it failed because you need native speakers to verify and correct the translations and this wasn't the sort of work that people were jumping to volunteer to do.
  - lazyasciiart 3y ago
    I speak multiple languages. It'd be fun to be given a link saying "this article has content present that is not in the English version" and see if I can update either of the articles using what's in the other copy. But I'm not going to go read wikipedia articles in two languages to find them.
- WhitneyLand 3y ago
  >>LLMs can't generate english answers for things answered only in Russian in the training data. For multilingual LLM’s? Why do you think that? An LLM can translate inputs of arbitrary Russian text. If there were an English question about something only in the training data as Russian, I would expect an answer - with the quality being on par with its general translation capabilities.
barbarr 3y ago
Good analogy! There's good reason to believe that web archives "uncontaminated" by LLM output will have some unique value in the future (if not now).
- cmcollier 3y ago
  For those wondering about the analogy: * https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel
pksebben 3y ago
That's gonna be a lot of fun to play with in a year or so. There's a concurrent explosion of 'veracity' analysis - it'll be fun to run those against wikipedia a year from now and your data. Incidentally, are you interested in mirroring your dataset and making it more robust? I'm sure I've got a few TB of storage lying around somewhere...
- Anon84 3y ago
  You can just download it yourself. Wikimedia publishes regular dumps in easily accessible formats: https://dumps.wikimedia.org/enwiki/20240320/ https://dumps.wikimedia.org/enwiki/20240320/ (the most recent for english Wikipedia)
  - pksebben 3y ago
    I don't see historical dumps. Am I just dumb?
    - Anon84 3y ago
      No, the website is just weird. The original link I posted is for the most recent dump... if you want older ones: https://dumps.wikimedia.org/enwiki/ https://dumps.wikimedia.org/enwiki/
  - _akhe 3y ago
    "Note that the data dumps are not backups, not consistent, and not complete."
- LeoPanthera 3y ago
  They are already on the Internet Archive as Kiwix archives.
tiptup300 3y ago
You know that wikipedia keeps revisions on all articles. I'm sure you could put together a script to make a copy any time of each page from a certain point of time.
schainks 3y ago
How do you browse this snapshot? I'm interested in this solution, too

I saved a full snapshot of Wikipedia (and Stack Overflow) in the weeks before ChatGPT launched, and every day I'm more glad that I did. They will become the Low Background Steel of text.

jakderrida 3y ago

The thing is that the Wiki mods will need to be more diligent with uncited things. I also see 2 massive opportunities here. First is that they can have agents check the cited source and verify whether the source backs up what's said to a reasonable degree. Second opportunity is fitting in things only found in other language Wikis that either be incorporated into the english one or help introduce new articles. Believe it or not, LLMs can't generate english answers for things answered only in Russian (or any language) in the training data.

groceryheist 3y ago
> First is that they can have agents check the cited source and verify whether the source backs up what's said to a reasonable degree. This is a hard and tmk unsolved NLP/IR problem, and data access is an issue. > Second opportunity is fitting in things only found in other language Wikis that either be incorporated into the english one or help introduce new articles. This has been attempted via machine translation in the past, and it failed because you need native speakers to verify and correct the translations and this wasn't the sort of work that people were jumping to volunteer to do.
- lazyasciiart 3y ago
  I speak multiple languages. It'd be fun to be given a link saying "this article has content present that is not in the English version" and see if I can update either of the articles using what's in the other copy. But I'm not going to go read wikipedia articles in two languages to find them.
WhitneyLand 3y ago
>>LLMs can't generate english answers for things answered only in Russian in the training data. For multilingual LLM’s? Why do you think that? An LLM can translate inputs of arbitrary Russian text. If there were an English question about something only in the training data as Russian, I would expect an answer - with the quality being on par with its general translation capabilities.

barbarr 3y ago

Good analogy! There's good reason to believe that web archives "uncontaminated" by LLM output will have some unique value in the future (if not now).

cmcollier 3y ago
For those wondering about the analogy: * https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel

pksebben 3y ago

That's gonna be a lot of fun to play with in a year or so. There's a concurrent explosion of 'veracity' analysis - it'll be fun to run those against wikipedia a year from now and your data. Incidentally, are you interested in mirroring your dataset and making it more robust? I'm sure I've got a few TB of storage lying around somewhere...

Anon84 3y ago
You can just download it yourself. Wikimedia publishes regular dumps in easily accessible formats: https://dumps.wikimedia.org/enwiki/20240320/ https://dumps.wikimedia.org/enwiki/20240320/ (the most recent for english Wikipedia)
- pksebben 3y ago
  I don't see historical dumps. Am I just dumb?
  - Anon84 3y ago
    No, the website is just weird. The original link I posted is for the most recent dump... if you want older ones: https://dumps.wikimedia.org/enwiki/ https://dumps.wikimedia.org/enwiki/
- _akhe 3y ago
  "Note that the data dumps are not backups, not consistent, and not complete."
LeoPanthera 3y ago
They are already on the Internet Archive as Kiwix archives.

tiptup300 3y ago

You know that wikipedia keeps revisions on all articles. I'm sure you could put together a script to make a copy any time of each page from a certain point of time.

schainks 3y ago

How do you browse this snapshot? I'm interested in this solution, too