LocalLLaMA

3 readers

1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 1 year ago

MODERATORS

communick@poweruser.forum

How to give knowledge to a model (alien.top)

submitted 11 months ago by bullerwins@alien.top to c/localllama@poweruser.forum

4 comments fedilink hide all child comments

Hi!

I’m quite new to LLMs and want to use it to make training workouts. My idea would be to feed it scientific studies and a bunch of example workouts.

Is this what “training a model” is for? Any resource where I can start to learn how to train one?

Can I use and already fine tuned model like Mistral, or do I need to train a base model like LLama2?

Can I train a quantized model or do I need to use a vanilla one? And quantize it after training?

I have 2x3090, 5950x and 64GB of Ram. If that matters. If I can load a model for inference can I train? Are the resources needed the same?

Thanks!

you are viewing a single comment's thread
view the rest of the comments

[–] ThinkExtension2328@alien.top 1 points 11 months ago (2 children)

I have only recently found the correct awnswr which is take the information and use Sparse Priming Representations (SPR) to distil the information. Next feed this text to privateGPT to use as a vector db document. Since SPR condenses the text you will be able to use more items as part of the retrieval phase.

Now query the LLM using the vector db, due to the SPR encoded text you get highly detailed and accurate results with a small LLM that is easy to run.

[–] bullerwins@alien.top 1 points 11 months ago (1 children)

Hi! It’s the first time I’m seeing SPR, any resource where I can learn more about it? I’ve seen privateGPT, I believe it’s a front end that lets you upload files and I guess it build a database using something like chromaDB that learns what you feed it and takes it into consideration when giving answers, is that right?

[–] ThinkExtension2328@alien.top 1 points 11 months ago

SPR is not a technology it self it’s a methodology to “compress information” in a way that ai can effectively achieve larger context input for the same size, David does a great video explaining the methodology behind it. Iv found it to be useful as hell.