Investigating Alpaca 7B - Finetuned LLaMa LLM

[@ItsRyanStudios]

Open source models are ESSENTIAL Glad there are groups like Stanford and Big Science's BLOOM working to make this a reality. This technology is too important and too powerful for any single corporation to hold the keys to. Thanks for sharing Weighting on those weights!

[@samwitteveenai]

If you look in the Colab the weights are there you can grab them I have also uploaded at 13B Alpaca model as well.

[@Tymon0000]

I tried the demo for some programming prompts I used previously in chat-gpt. The alpaca didn't give me anything usable, although it kinda understood the question once.

[@brianmi40]

I predict this year will bring models that are trained for specific purposes. Imagine a model trained ONLY on working code...

[@priestesslucy]

@@brianmi40 I'm imagining an AI trained only on GMing a specific tabletop game. More complicated than code because there's a social aspect as well as rules. But it's definitely doable

[@Tymon0000]

@@brianmi40 Yea, I think so too. I would want to also give it full docs for a specific language, framework, and project. It doesn't need to be as generic as chat gpt.

[@brianmi40]

@@priestesslucy For sure, my son meets with a D&D group regularly, that should be pretty easy to do. Just like different dungeon master guides for different scenarios, there can be different datasets for the AI to use, based upon the desired types of stories or mythological basis. What will help drive these are the myriad business opportunities for people to create these specialized AIs, and with the advent of smaller and smaller models, like Stanford Alpaca being almost equal to ChatGPT, but running on a desktop, we'll soon have such simple ones such as for D&D running on a phone...

[@coderentity2079]

@@brianmi40 It has to understand language too to understand the question / give explanations, but sure it might be a better coder if it doesn't have to train on cooking or history. To understand the mechanics of the world still needed, otherwise it can't effetively help programmers who try to model the world and it's processes.

[@HomingPigeon]

Thanks for the hard work doing the research so we can learn. I appreciate this channel. Keep it up, this will grow. Quality channels always do.

[@samwitteveenai]

Much appreciated!

[@kaushikqravindran9792]

Awesome work!! Thank you so much.

[@wilfredomartel7781]

Great info Sam! I was reading that LLama is multilingual and can summerize long text.

[@Orangesnake221]

Incredible!

[@2happytimes]

The LLaMa weights are avalible via torrent now.

[@samwitteveenai]

yes they have been for a while but for legal reasons I don't want to share that. They are on HuggingFace though for people to get them there.

[@micbab-vg2mu]

Very interesting!!!

[@catyung1094]

Great Video , Sam 💪!

[@Leto2ndAtreides]

Now for someone to train it on GPT-4's outputs... Although the 7B model might be better optimized as a domain expert rather than trying to have it be a universal expert. And that's assuming that the 7B can represent certain complex ideas (one would assume so, but...). Like, some of the advancements in LLMs ability to comprehend things are just byproducts of them being bigger and capable of recognizing more patterns.

[@samwitteveenai]

optimizing for specific domains and tasks does make a lot of sense. Also using a model like this for retrieval with augmented generation tasks is where I think it can be very competitive.

[@Leto2ndAtreides]

@@samwitteveenai Yeah. Personally, I think they should at least release the 7B weights.

[@samwitteveenai]

@@Leto2ndAtreides agree it would win them a lot of credibility compared to OpenAI too.

[@_NoMercy_]

how to run alpaca on Local host?

[@samwitteveenai]

You will just need a decent GPU and then can wrap it in FastAPI etc for serving. Check out my video on running it in Colab.

[@puppergump4117]

@@samwitteveenai Does AMD count as decent

[@alzblb1417]

Know any good -weed- weights dealer?

[@kc-jm3cd]

Man I was talking to some guys in a back alley said they had the hookup but they just took me to an old boxing gym