Code Llama: Meta's State-of-the-Art LLM for Coding

--- type: transcript date: 2023-08-25 youtube_id: U24mPUHXuv8 --- # Transcript: Code Llama: Meta's State-of-the-Art LLM for Coding in this video I'm going to be showing you code llama which is a new state-of-the-art large language model capable of generating code and natural language about code so there's three different variants of code llama there's Cod llama which is the foundation model for all different programming languages there is code llama python specifically for python there is cod llama instruct which is further fine-tuned on generating responses from natural language and in this video I'm going to be showing you everything from benchmarking where to get resources if you're looking to pull this down locally and use your own compute I'm going to be showing you a little bit about how the model was trained and some pieces that I found interesting and then finally I'm going to point you to where you can go ahead and play around with the model right now without having to pay for an account or anything and just get up and running within a web app so the first thing that I wanted to go through was a piece of the paper that I found interesting so instead of opting for using professional developers to get uh data to further fine-tune and create this model what they decided to do was they decided to generate 62,000 interview style programming questions from prompting llama 270b from there they removed all the duplicates then what they decided to do was they generated unit tests from there they generated 10 Solutions then from there they ran those unit tests on the 10 Solutions and the first solution that passed the test was passed to the data set with both the question and the test so that's sort of how they built this which I thought was an interesting and creative approach so actually asking the large language model for interview style questions as sort of that first step which I thought was pretty pretty an interesting approach for this self-instructed to show you was just where you can go ahead and find some more information on this so I found the blog post to be a little bit more digestible than the white paper but I do encourage you if you're interested in this stuff there is a lot of interesting stuff in here on how they chose to approach this so within the blog post there is a nice table where they do list out uh both how they fine-tune and broke out those three main models so there's the path on how they went about uh creating the python model which it was further trained on more tokens and then it was uh broken out into three different sizes so the 7B 13B and 34b and then finally the stream for the general model so code llama uh for coding specific tasks and then there's also the natural language variant so there's nine in total across these three different uh sort of umbrella models within um code llama now the benchmarks themselves I'll just leave it up on screen for a moment here now the one that everyone does strive for is gp4 now well it doesn't surpass gp4 uh within its uh large largest variants or anything like that uh the one thing to note is that it does perform exceptionally well and even uh if you look at GPT 3.5 uh some of the models do perform better so pretty interesting that an an open- Source model is on some metrics performing better than a Clos Source model like GPT 3.5 so I encourage you to read through the blog post I'll put all these links in the description of the video if you're curious to look through this further so finally I wanted to just show you uh easy way to get up and running with playing around with this so on labs. perplexity doai you can go ahead and just hit this web interface and they already have up the 34b instru struct model where you can go ahead and just ask a questions so I can say uh I'll just ask some basic ones generate me a react app boilerplate now most people wouldn't go in here for that but I'm just sort of demonstrating how fast it is and what the interface looks like and all of that but I'd encourage you to just come over here throw your tough questions at it and see how it responds see how it uh performs in comparison to your experience with GPD 4 GPD 3.5 I'd be really curious uh individuals that play around with us please leave comments below for all of us it's always helpful when you leave comments uh good feedback or or not uh it's all helpful um to helping us understand these these new models as they come out so that's pretty much it for this video I just wanted to point you in the direction of a number of different sources about code llama and just sort of put it on your radar but that's it for this one so if you found this video useful please like comment share and subscribe and otherwise until the next one