NEW WizardCoder 15b - The Best Open-Source Coding Model?
Vložit
- čas přidán 19. 06. 2023
- In this video, we review WizardLM's WizardCoder, a new model specifically trained to be a coding assistant. It's completely open-source and can be installed locally. Let's test it out!
Enjoy :)
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewberman.com
Need AI Consulting? ✅
forwardfuture.ai/
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew-berman-youtube
USE CODE "MatthewBerman" for 50% discount
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Links:
TheBloke's Model - huggingface.co/TheBloke/Wizar...
WizardCoder HF - huggingface.co/WizardLM/Wizar...
TextGen WebUI - github.com/oobabooga/text-gen...
How to Install TextGen - • How To Install TextGen... - Věda a technologie
It's channels like this that makes paying for CZcams premium worth! All of this amazing content without a commercial and you still get paid in the process. Mad love from New Orleans
It will generate a snake game if you specify for it to use the pygame library. I noticed a lot of models default to using turtle and even chatGPT fails when trying to use turtle a lot of the time.
As always: Outstanding! Thanks for sharing it!
Great information as usual!
It's impressive how much more performant it is at coding compared to a normal 15b model.
Imagine how incredible it would be if GPT 3.5 or GPT 4 (which are already great at coding) were finetuned like this.
If such very large models can be finetuned for coding and that there is actually that kind of relative jump in performance, I think that it would create enormous economic value even if it can't generalize.
For real tho. When they make the gpt-4 code interpreter open for everyone for gpt plus users. It's gonna be massive.
I'm excited to see what IBM has... A leak there would be amazing.
Isn't that what Copilot does? GPT-4 Finetuned at coding.
"Imagine how incredible it would be if GPT 3.5 or GPT 4 (which are already great at coding) were finetuned like this."
They very well may be. For all we know, ChatGPT automatically switches between multiple fine-tuned models based on the type of query the user is asking. Seems like a no-brainer type of method to increase performance.
@@MoeShlomo GPT-4 is too slow for that to be true. However, that's a very economy approach. Model balancing lol the equivalent of Load Balancing but for AI models.
Great Stuff. Thank you!
Let's see Paul Allen's Coding LLM
What?
@@matthew_berman It's just a meme lol
@@SinanAkkoyun i got it dw
@@matthew_berman You should be highliighting that guanaco 33B is the best open source coding llm
Look at that subtle, off-white colouring…
I would love to see this against Orca whenever it's released.
dude these tools are coming out faster than I can install them on my computer!!!
hey brother - love ur videos; just curious which mic u use
The great thing about local models is, you can work without an internet connection, which is really helpful in those situations where your provider is having issues.
Sadly people don't seem to get how beneficial and POSSIBLE it is to have local models. Everyone's all about the "google colab." I just skip colab tuts now. It's just not what I'm looking for/what I want from AI...
Imagine not even NEEDING or WANTING internet... Doubt that is what the media companies want.
@@VioFaxI don't have a good pc, so Google Colab is my only choice, not everyone has an RTX
@@VioFax Thing is that you move the cost of buying and operating HW to the cloud computing provider.
So as much as i would loved to have personal private AI i can not come close to the performance/ possibilities i get from what huge multi billion companies can offer me for pennies on comparison.
It might change however with new AI GPU boom that is taking over nVidia and AMD as well. Hopefully we get to see some affordable AI HW to run private (offline) LLM.
@@Greenmarty I really wonder what GPUs will be like in the not so distant future, like imagine in 2026 we get 1TB GPU cards and then the price starts to drop for other cards. Like a 24GB GPU card might be really cheap then.
@@ZeroIQ2 For now these RTX 4090, RTX 3090 Ti, RTX 3090, or Titan RTX can do for 7B-13B models but they are still expensive . As I've seen multiple expensive > affordable HW over the years i don't see the reason why this would not happen. And there's always an aftermarket.
the im_sep is a command in the prompt. you can change the prompt to match the model in your chat app. i prefer lm studios myself
The limit of 2048 is in the models, text-generation-webui has a config file you can edit but that wouldnt change that the model cant do it. So asking it to do a snake game is silly, it would never fit. It isnt a limitation in the model "intelligence" but of the architecture.
You could try to ask for code that is minimised. Meaning it doesnt use full names for values, doesnt use newlines etc.
When will we get to play with that shiny new Orca model?
3:55 What LLM coding tool operates directly in VSCode(for example) so you DON'T have to "go back and forth" as you mention.
Copilot?
Looking for tools that work better than ChatGPT for Unity/C#
Would this be a good candidate?
Given a natural language description, it can generate code. But can it do the opposite - given a piece of code, can it describe in layman terms what that code does? What are the best LLMs for this code explaining task?
That's a very good questions which is definitely worth exploring with some tests.
which software was you using to acess openb source models, im using chatgpt4all and i can see code models on hugginface but dont seem to be in correct format, it require .bin files
any good models that actually execute python code and can self correct?
Is it better then simple , clean GPT4 ?
How about using it with GPT Engineer?
Why?
What's the difference between textgen web ui and oobabooga?
Is this better than guanaco 33b or Falcon?
What sort of consumer GPU is really needed to run a model like this? Do I need a 24GB beast, or can I run on something a little more practical for home/hobby use?
It's an AI with 15B parameters that uses a 4-bit storage value, so it can fit on a 12Gb GPU.
@@rom100main Thanks that's helpful.
I just installed it and it is giving me terrible answer. Like complete nonsense. I think i messed something up or it doesn't run properly on a GTX 1080.
Is Wizard coder suited for C/C++, too?
I have RTX 2060 ( 6GB) and an I7-10700 with 32GB ram. how far can I run models on my pc
What is gptq by bloke?
Has anyone tried Tree of thought prompts with this?
People like me
Amd GPU and GGLM doesn't run on textgenWebui plus not running on GPT4ALL
Any chances amd GPU will be even supported in future or not 😢
Yea, I keep reading that AMD GPU isn't supported. Sorry :(
so gpt engineer is still better?
is it better that replit AI coder?
I wonder how does it perform at Competitive Programming
Test it out and release a blog post or video on the topic.
"Just a Little Bit Longer" in the Training, and ChatGPT 4 and 5 might be taking the Back Seat with this type of progress.
What a Time for Tech !!!
You must be exhausted from the ....
"Thrill of the Chase"
Just when you thought you caught something, you let it go to Chase after something even Better.
It's a Marathon at this point on who can bring the most Powerful tools to the Social Market and have the most up to date solution for the future. We gotta be getting close to the Grand Finale right ? 😅
What are the local hardware requirements?
15 billion parameter model at 4 bits per parameter works out to 6.98 GB of Vram (there is a formula - ask chatgpt)
If it isn't better than GPT 3.5 it's really not that good. An improvement yes, but what's the point?
I don't think it's better than guanaco 33B at coding.
Unfortunately I don't have the right hardware for the situation... otherwise it would be my daily "pleasure" to test new IAs
TheBloke rocks, he is by orders of magnitude more important for AI progress than Ilya , who get's all the praise now, can't stand that because he is just a footnote in the grand scheme. Without Hinton, Google and M$ he is nothing, while the TheBloke needs no one.
Sorry, but that `format_number` - final code test - is nowhere near a 7/10.. for one it's just a function, I'd give it a 2/10. The real problem I find LLM's having is compound logic or making a complete project (like snake). Nothing against you, I love your videos, and info! In my experience, using AI as a code assistant is frustrating and more pain then it's worth. With that said however, I do find it's very good at assisted code reviews, summarizing, planning (steps), static unit tests, and boilerplate stuff. It'll be awesome to see where it goes. Best!
if its not better than gpt3.5 then its garbage. Cause GPT3.5 is garbage for coding. GPT4 is very useful for coding tho its night and day compared to GPT3.5.
A more powerful computer is needed and training must be of better quality.
Get some sleep.
That's just how I look lol
This is clickbait. You confirmed in your video GPT3.5 and above is much better. So watch it with your titles dude! You should also declare you were approached by Python Principles to inject a useless piece of information in this video. Not related at all. You are too sneaky!