Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

vicuna.lmsys.org

Post reply on HN
Print view View original

161–170 of 185 posts

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#161

Post by nico »

evanchisholm wrote: post #77
nico wrote: post #36

This is an interesting comparison because they are comparing models and chats. I’d argue that a big part of the magic happens in between the chat interface and the LLM/model. We had GPT3 for a long time and most people didn’t really care too much. But then when they added Chat on top and whatever they are doing to provide previous context in each prompt… that was magic. So then I wonder how much better could we make…

ChatGPT isn't just a ux change over gpt3 a lot of work was put into RLHF that makes the output so much better.

Thank you for that insight. That supports the idea that it's not just the base LLM model, but it's all the pieces together that make ChatGPT such an incredible experience for us to use.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#162

Post by HanClinto »

zhisbug wrote: post #120
visarga wrote: post #103

Earlier quoted context omitted.

Turns out LLMs are not so inaccessible for individual users.

Exactly. but if you want some commercial use, still lacking good quality models.

At the moment. It's only been a few weeks.

What happens when we get something LLaMa quality that is actually free of restriction?

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#163

Post by hyperopt »

The demo for this is great. It's the best non-corporation assistant I've used so far. I suspect most of the gains here relative to the Alpaca model might have to do with the fact that the ShareGPT data are full conversations. They allow for the assistant to respond to the earlier messages in a cohesive way. As opposed to Alpaca, the data was a single question and answer, so the model seems to lose context of earlier information. Also, the coding abilities of Vicuna are significantly improved relative to Alpaca, to the point that I began to suspect they might be calling out to OpenAI in the backed. Please, release model weights and finetuning training data.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#164

Post by Anunayj »

jerojero wrote: post #42
shp0ngle wrote: post #7

Earlier quoted context omitted.

I don’t understand… GPT is itself trained on copyrighted text?

Yes it probably is. However chatGPT's term of service are explicit about disallowing people to use code generation from the model to train other models. Now, how enforceable that might be... that probably remains a matter for the courts. But doing what these people have done is breaking the terms of service of the application.

OpenAI persuing that would likely end up setting precedent for it's own demise, I'd like to see them try.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#165

Post by blacksoil »

toxik wrote: post #157
jerojero wrote: post #47

The work being done with Llama and Alpaca and so on is what I've found to be most interesting in the past few weeks. The fact that we are capable of producing such high quality conversational models that we can actually get to freely experiment with is really encouraging. I'm particularly excited about applications such as viper[1] and other forms of plugins as well as model ensembles. [1] https://viper.cs.columbia.e…

I'm not trying to be negative, but there is a rather large gap between ChatGPT and what Vicuna is showing in its online demo. Here's an example I just had: Q: What should I do for my wife today? A: There are many things you can End of transaction.

Perhaps there was connection issue? I am getting this:

The best thing you can do for your wife today is to show her love and affection in a way that she will appreciate. Every person is different, so it's important to consider her preferences and interests when deciding what to do. Here are a few ideas: 1. Cook her favorite meal: If your wife has a favorite dish, why not cook it for her today? 2. Plan a special outing: Whether it's a trip to the spa, a shopping excursion, or a day at the park, planning a special outing can be a great way to show your wife that you care. 3. Do something together: Whether it's playing a game, watching a movie, or going for a walk, doing something together can be a great way to spend quality time with your wife. 4. Show your appreciation: Take a moment to tell your wife how much you appreciate her and why. It can be as simple as a kind word or a heartfelt message. 5. Help with a task: Offer to help your wife with a task that she needs to do, whether it's folding laundry, washing the dishes, or taking the dog for a walk. 6. Give her a gift: Even a small token of your affection, such as a bouquet of flowers or a piece of jewelry, can be a meaningful way to show your wife that you care.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#166

Post by leobg »

covi wrote: post #88
sliken wrote: post #80

Earlier quoted context omitted.

Anyone know exactly config is needed for training, presumably the $300 is on some GPU heavy instance on ec2? Is $300 of Ec2 p4de.24xlarge which is $40.96 an hour? Or maybe 7-8 nodes for an hour? Something else?

From the post: > The training was done with PyTorch FSDP on 8 A100 GPUs in one day. > We employ SkyPilot managed spot to reduce the cost by leveraging the cheaper spot instances with auto-recovery for preemptions and auto zone switch. This solution slashes costs for training the 7B model from $500 to around $140 and the 13B model from around $1K to $300. So, this is using for example a2-ultragpu-8g (8x A100-80GB) on…

What’s wrong with vast.ai? You’d be probably looking at $2/hr. So like $50 for the whole fine-tuning.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#167

Post by DeusExMachina »

alexvoda wrote: post #140
DeusExMachina wrote: post #133

Earlier quoted context omitted.

I was doing some market research with ChatGPT some time ago, and I started asking about some of my competitors. The output matched what I saw people say in conversations online, so I got curious about what it knew about myself. I asked and the answer was completely false.

But did it match what people were saying about yourself?

No, it just invented some expertise I don't have and things I never did, like books I didn't publish.

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#168

Post by pazimzadeh »

gpm wrote: post #101
pazimzadeh wrote: post #11

> Can you name all of the soccer players whose names contain a Z > Zlatan Ibrahimovic Zinedine Zidane Zoran Banović Zlatko Zahovič Zoran Matić Zayed Khoury Zé Mario Zé Douala Zé Elias Zé Luís Zé Manuel Zé Roberto Zé Xavier Zé Abel Zé Eduardo Zé Tomás Zé Luís Zé Henrique Zé Jorge Zé Breno Zé Manuel Zé Luís Zé Sérgio Zé Maria Zé Roberto Zé Ricardo Zé Álvaro Zé Luís Zé Carlos Zé Luís Zé Luís Zé Luís Zé Luís Zé Luís Zé L…

Considering words are mostly tokenized I'm not really surprised that questions about spelling don't work so well.

If that's the case then how is it so good at anagrams?

Re: Vicuna: An open-source chatbot impressing GPT-4 with 90% ChatGPT quality

#170

Post by nathanasmith »

DeusExMachina wrote: post #167
alexvoda wrote: post #140

Earlier quoted context omitted.

But did it match what people were saying about yourself?

No, it just invented some expertise I don't have and things I never did, like books I didn't publish.

Sounds like you have some work to do!
Post reply on HN