News for hackers. Stuff that matters.
Re:LLaMA Now Goes Faster on CPUs by gpm on Sunday March 31, 2024 @11:16PM (#39890615)
If you want to download a backup of a large chunk of human knowledge... download wikipedia. It's a similar size to a small LLM and can actually distinguish between real life and fantasy: https://en.wikipedia.org/wiki/Wikipedia:Database_download

If you just want to play around with an LLM though, absolutely.

Re:LLaMA Now Goes Faster on CPUs by int_19h on Monday April 01, 2024 @02:41AM (#39891447)
Kiwix provides prepackaged highly compressed archives of Wikipedia, Project Gutenberg, and many other useful things: https://download.kiwix.org/zim/.

Between that and dirt cheap storage prices, it is possible to have a local, offline copy of more human knowledge than one can sensibly consume in a lifetime. Hell, it's possible to have it all on one's smartphone (just get one with an SD card slot and shove a 1+ Tb one in there).

Re:LLaMA Now Goes Faster on CPUs by claritise on Monday April 01, 2024 @08:38AM (#39893498)
Just create a RAG with wikipedia as the corpus and a low parameter model to run it and you can basically have an instantly queryable corpus of human knowledge runnable on an old raspberry pi.
Re:LLaMA Now Goes Faster on CPUs by CaptainOfCoit on Monday April 01, 2024 @10:06AM (#39894325)
> a low parameter model

> on an old raspberry pi

I bet the LLM responses will be great... You're better off just opening up a raw text dump of Wikipedia markup files in vim.

Re:LLaMA Now Goes Faster on CPUs by boywitharupee on Monday April 01, 2024 @10:20AM (#39894423)
but which model to tokenize with? is there a leaderboard for models that are good for RAG?
Re:LLaMA Now Goes Faster on CPUs by sroussey on Monday April 01, 2024 @01:36PM (#39896727)
“For RAG” is ambiguous.

First there is a leaderboard for embeddings. [1]

Even then, it depends how you use them. Some embeddings pack the highest signal in the beginning so you can truncate the vector, while most can not. You might want that truncated version for a fast dirty index. Same with using multiple models of differing vector sizes for the same content.

Do you preprocess your text? There will be a model there. Likely the same model you would use to process the query.

There is a model for asking questions from context. Sometimes that is a different model. [2]

Re:LLaMA Now Goes Faster on CPUs by Workaccount2 on Monday April 01, 2024 @11:17AM (#39895040)
Pretty neat to have laying around, thanks
Re:LLaMA Now Goes Faster on CPUs by CaptainOfCoit on Monday April 01, 2024 @10:03AM (#39894304)
> actually distinguish between real life and fantasy

Are LLMs unable to distinguish between real life and fantasy? What prompts have you thrown at them to make this determination? Sending a small fairy tale and asking the LLM if it thinks it's a real story or fake one?

Re:LLaMA Now Goes Faster on CPUs by gpm on Monday April 01, 2024 @10:27AM (#39894497)
... having them talk about events from sci fi stories in response to questions about the real world. Having them confidently lie about pretty much everything. Etc.
Re:LLaMA Now Goes Faster on CPUs by CaptainOfCoit on Monday April 01, 2024 @10:51AM (#39894731)
What are the specific prompts you're using? You might get those answers when you're not being specific enough (or use models that aren't state of the art).

"Shit in, shit out" as the saying goes, but applied to conversations with LLMs where the prompts often aren't prescriptive enough.