NEW WizardCoder Python 34B LLM is AMAZING!!!

Key Takeaways

coder pipes on 34 billion parameter model is the latest model to join the family of wizard coded models the model has been in the news because there have been claims that this model has knocked gpt4 out of the park it is partly true it is partly not true and in this video I'm going to explore why it is partly true and why it is not partly true but moreover we're going to also explore why this model has been discussed so much about the first thing to start with this wizard coder is a family of models every time there is a new model or a new data set comes in Wizard coder manages to kind of update their existing model and then stay always at the top and now wizard coder has got the latest version which is a python specific 34 billion parameter model and that is based on code Lama so they've used code Lama and they've fine-tuned their existing wizard decoder model and that model is beating gpd4 how is it beating gpd4 before I even get into the model in itself let's first clear this thing so what the news is that jeep wizard coder python 34 billion parameter model achieved 73.2 on a human evil plus one and that is surpassing gpt4 but how it is surpassing gpt4 is something that you need to keep in mind the first thing for you to know is when GPT 4 was launched open AI reported a particular human evil score and that score was 62 is the GPT Force code then you know for sure that wizard coder and lot of other models have beaten gpd4 but is that the entire score no when the wizard coder team try to replicate the gpd4 or replicate the benchmarks for the latest API of gpd4 that has code 82 so have wizard coder python 34 billion parameter model scored more than 82 no absolutely not it has not scored more than 82. it has surpassed what openai reported at the start so there are two things to it let me clear it when openai launch gpd4 there was a score that they reported and wizard coder and a bunch of other models have overcome it when these researchers tried to replicate or calculate the Benchmark score for gpt4 with the latest API then that is code 82 which is like way above every other open source model that is available having said this if you take this headline just simply out that wizard coder 34 billion parameter python model has beaten open AI gpd4 if you just remove the title wizard coder 34 billion parameter model python is still the best the best python or coding model that you could see primary reason is because this has been fine-tuned based on the Llama 2 or the code Lama model that was released couple of days back much more promising aspect of this entire piece is that this is a 34 billion parameter model this is not a 70 billion parameter model this is not a mixture of experts this is just a single single 34 billion parameter model that is doing extremely well on the human eval benchmarks now if you are going to ask me hey do you trust every single Benchmark that is available I would say probably no I don't trust benchmarks I would expect people to give the models in the hands of other people for them to try it out and that is exactly what the team has done they have uploaded the model on hugging face modeler the model is available for you to directly download the checkpoint is available you can go here click files and versions there is no terms of services there is no form that you have to fill in directly go there and download and if you have got a big enough GPU like I think you need a really good GPU then you can directly check this model and then try it out I've heard from couple of places like Hacker News where people have felt that this model and every other code llama derivative model is doing really good and they're happy with the performance that they've gotten when you compare it with gpd4 but unfortunately I could not make the comparison because the radio application was running forever and in fact like I waited for a long time to see if this task would finish but didn't finish so if you happen to run this model locally please let me know what you feel about it but as a matter of fact that we have an open source model that is way way above every other open source model is first of all a great advantage and in fact if you see this open source model which is code 73.2 this is a really good score so overall the point is very simple wizard coder python 34 billion parameter model is really one of the best open source models that you could use today and that that is like scoring much more than code llama python model I think that's the power of fine tuning here but does it beat gpt4 no it doesn't beat gpd4 in its current format the fact that this comes closer to gpd4 is quite an amazing achievement so having said that like having that we have cleared all these questions if you want to use this model yourself all you have to do is go download here I think this clears up the entire Buzz around wizard coder model or any other beating gpd4 whenever you see a news that some model has beaten gpd4 make sure that you check which version of gpt4 they are talking about what is a benchmark what is a benchmark that they have used have they trained the model has there been a data leakage I mean this is a good thing for you to try it out but the fact that this research is very honestly put out that they got 82 and 72.5 when they tested their self with the latest API I think this is a commendable thing I wanted to really appreciate them for that and as a matter of fact that you can use this model that is available open source for you to try it out is also another cherry on the top of the game and I just wanted to release this video and then explain you why everybody's been talking about beating gpt4 and why it may not be entirely true if you have any questions let me know in the comment section but I'm definitely looking forward to try out this model and compare it with gbt4 and give you some insights see in another video happy prompting