Stability AI releases StableVicuna, a RLHF LLM Chatbot | Hacker News Reader

Stability AI releases StableVicuna, a RLHF LLM Chatbot

stability.ai

52 pointsdavidbarker20 comments

At least in terms of self-reported perplexity all the fine-tuned models seem significantly stupider than the raw llama checkpoints. Better for business but not nearly as fun or useful for human persons wanting to talk to something that can seem to reason and respond like a human does. And since the llamas are not for profit you can draw your own conclusions.

In terms of getting work done 30B and under llamas aren't going to be very helpful. Maybe 65B behaves differently but the 30B and under are just amazing toys; not tools. And, again, if you fine-tune them they get significantly stupider. Think of fine tuning as a lobotomy. The model is more likely to stay on the rails but it loses emergent capabilities.

Based on Llama, so not open source (non-commercial only license)...

Agreed, it's starting to get really annoying when models that aren't get announced as being "open source".

For the tinkerers: here are some huggingface available models that are really open source (Apache or BSD license as far as I evaluated it):

- RWKV (raven 14B)

- GPT-NeoX-Chat-Base (20B)

- Flan-T5-xxl

- Fairseq Dense (13B)

- Pythia Chat Base (7B)

- Codegen (16B)

- Bloomz (7B)

Some are general, some are more specific. But I could get something out of all of them.

- why fine tune vicuna and not your own model that was released last week?

- WizardLM was released a few days ago and showed promising performance for 7b, are you going to fine tune it next?

A startup with $100M in funding shouldn't focus on these optimizations.

Why would they build this on top of Llama when we know Meta has been going after projects using Llama?

The ambiguous licensing poisons all usage up the stack...

This project was probably extremely affordable for them and there is value in these tests. They can play around with these different fine-tuning techniques and put them out there for people to mess with while gaining insight into better methodologies for when they spend the big time and cash on larger models.

I'm team lead at Carper. This is correct. It's also just a project one of our engineers did over the course of a few days, so very low risk.

We will swap the base model out for StableLM as soon as we can and iterate from there. We just thought the community would enjoy this research artifact :)

Hopefully RedPajama will change things when (if) they release models based on their independent LLaMa recreation. Until that happens, it might make sense to use the LLaMa model as a stopgap for internal development.

Exactly! It seems pretty clear that someone will reproduce llama and alpaca with a commercially viable model within the next few months to a year at the longest. So if you want to build something based on top of this technology, it makes sense to build it on top of llama/alpaca internally, and then swap it out with something commercially licensed when it’s available. Then you’ve got a huge head start on anyone who is waiting for the stars to align before they start trying to build something at all.

Is there anyone compiling an in-depth comparison between different models that are emerging? Are there standardized benchmarks?

Nope and I'd really like to see a yearly benchmark to create some sort of feature rating!

I feel like I see charts floating around r/LocalLLaMA/ from time to time. Might be some more info there.

Try it out https://huggingface.co/spaces/CarperAI/StableVicuna

I feel like there's something I'm missing about Hugging Face. Every time I try any demo of any model on the website I can never even get a response before it times out.

Why is it trained on GPT generated content? Won't it reduced the content quality instead of improving it? GPT generates lots of fluff which will only pollute the new models further.

The vicuna model has shown that there is value in this method.

https://vicuna.lmsys.org/

Interesting. Many releases from Stability in a few days.

Would love to see how it performs against ChatGPT though.

They desperately fundraising.... so need to get into the news cycle again

yet another model using the stanford alpaca dataset, what a waste of time.

Reply on news.ycombinator.com