Wikidata talk:WikiProject Large Language Models

From Wikidata
Jump to navigation Jump to search

Burning hot interest[edit]

Thanks for setting up this project. I want to discuss this topic thoroughly with my Master's seminar on Semantic Modeling this coming summer term (starting April 4, 2023). I think a lot will hinge on lookups, being able to look up tokens and get the Q-value. But then of course, LLMs can't do inference (and let's face it, there are true statements that cannot be proven correct within the logic of a system). Still, great to think about this! WiseWoman (talk) 21:25, 21 March 2023 (UTC)[reply]

I mean not just collect information *about* LLMs, but about connecting LLMs and Wikidata! WiseWoman (talk) 21:27, 21 March 2023 (UTC)[reply]

LLM Applications to Wikidata[edit]

I recently built a tool using LLM (well a modestly sized LM, DistilBERT). It's called psychiq and it predicts P31 statements from Wikipedia text. I'm interested in doing more though.

Things I've considered include:

  • Vandalism detection using LLMs (I know the WMF's ML team has models in progress for vandalism detection on wikidata but I have no idea what their status/timeline is).
  • (partially-)Automating mix-n-match using LLMs
  • Expanding Pyschiq to handle more properties, more obscure statements and parsing more of the article's text.
  • Expanding psychiq to more languages

BrokenSegue (talk) 02:03, 22 March 2023 (UTC

Nice! Feel free to add information about it on the WikiProject main page, or perhaps open a Subpage for "statement prediction tools" or similar TiagoLubiana (talk) 18:43, 24 March 2023 (UTC)[reply]

Hugging Face Property[edit]

Maybe a good thing to work on is adding a property for huggingface.com (Hugging Face (Q108943604)) model pages. We could link deep learning models to their official huggingface repo page. BrokenSegue (talk) 02:05, 22 March 2023 (UTC)[reply]

Would support. -wd-Ryan (Talk/Edits) 19:03, 22 March 2023 (UTC)[reply]
@BrokenSegue: Looks useful! If you propose it, I'd support it too. TiagoLubiana (talk) 18:39, 24 March 2023 (UTC)[reply]
It might be preferable to get HF to add an identifier to their model pages? Unclear whether a model can have many pages, whether it can be renamed / merged / split, &c. Sj (talk) 20:24, 12 March 2024 (UTC)[reply]

Artificial intelligence field VS model[edit]

artificial intelligence (Q11660) has been used to refer to the computer science field and an individual AI model simultaneously. Should I create a new "AI model" item and update all items that link to it?

-wd-Ryan (Talk/Edits) 17:46, 30 March 2023 (UTC)[reply]

TiagoLubiana (talk) 19:30, 21 March 2023 (UTC) John Samuel (talk) 20:07, 21 March 2023 (UTC) WiseWoman (talk) 21:22, 21 March 2023 (UTC) -wd-Ryan (Talk/Edits) 21:42, 21 March 2023 (UTC) BrokenSegue (talk) 01:58, 22 March 2023 (UTC) Kippelboy (talk) 14:33, 22 March 2023 (UTC) Premeditated (talk) 18:36, 22 March 2023 (UTC) LydiaPintscher (talk) 09:38, 23 March 2023 (UTC) Stevenliuyi (talk) 12:40, 24 March 2023 (UTC) Fuzheado (talk) 08:42, 30 March 2023 (UTC) Waldyrious (talk) 12:27, 6 April 2023 (UTC) LizeLive (talk) 21:32, 10 April 2023 (UTC) BEANS X2[reply]
RShigapov (talk) 10:10, 2 June 2023 (UTC) Vladimir Alexiev (talk) 17:15, 19 November 2023 (UTC)[reply]
Notified participants of WikiProject Large Language Models -wd-Ryan (Talk/Edits) 17:46, 30 March 2023 (UTC)[reply]
yes I think a model is distinct from the field. maybe convert GPT-4 or something into a model item. one problem is that many of these models are not really a single thing. for example there are 5 different versions of LLAMA with different param counts. Maybe we should call the item model-family? I don't really know. Things would get real complicated if we made properties for "param count" or "BLEU score" but maybe we never intend to do that. BrokenSegue (talk) 18:34, 30 March 2023 (UTC)[reply]
honestly we should probably fix all of these to point elsewhere https://w.wiki/6Wnc BrokenSegue (talk) 19:10, 30 March 2023 (UTC)[reply]
@BrokenSegue: I will be fixing these all. I'll make a model and model family item. Thanks! -wd-Ryan (Talk/Edits) 23:13, 30 March 2023 (UTC)[reply]
I don't know which are individual models and which are families of models, but please look over them: https://w.wiki/6WpD
I replaced the ones I thought were appropriate with new artificial intelligence model (Q117349473) and artificial intelligence model type (Q117349475). Better than what we had before. -wd-Ryan (Talk/Edits) 23:36, 30 March 2023 (UTC)[reply]

Number of Parameters[edit]

One of the key characteristic of Large language models is the number of parameters. For example, GPT-3 has 175 billion parameters. Currently, we are tracking this number using data size (as a qualifier). Do we need a new property? Or any existing property could be used? If the usage of data size is agreed, we may need to correctly document its usage in this project: as qualifier to software identifier or as property. John Samuel (talk) 09:51, 3 April 2023 (UTC)[reply]

Good points. I think that using "data size" should be good enough, but no strong opinions. Perhaps changing the property name to "software size", so it is a bit more general?
I'd use it as a main property, but that is only a first impression. TiagoLubiana (talk) 12:25, 6 April 2023 (UTC)[reply]
Personally, I also agree with the use of "data size" as main property and use "software version" as quantifier. John Samuel (talk) 13:05, 6 April 2023 (UTC)[reply]

Merge with WikiProject Artificial Intelligence?[edit]

I'd love to have one project that includes discussions of classifiers, translators, datasets, and many other elements not specific to LLMs. There's an unused AI wikiproject that might make a better umbrella. Sj (talk) 20:25, 12 March 2024 (UTC)[reply]

Retrieved from "https://www.wikidata.org/w/index.php?title=Wikidata_talk:WikiProject_Large_Language_Models&oldid=2100786699"