Llava 34B Released! Exceeding Gemini Pro in Performance Benchmarks?

Key Takeaways

this is amazing now we have lava 34 billion parameter model this version exceeds Gemini Pro on several benchmarks it has improved reasoning OCR and World Knowledge increasing the input image resolution to four times more pixels than before better visual reasoning and OCR capability better visual conversation for more scenarios efficient deployment and inference that is lava 1.6 that's exactly what we're going to see today let's get [Music] started hi everyone I'm really excited to show you about lava 34 billion parameter model it is a multimodal model where we are able to enter a text and also an image and ask questions based on that I'm going to take you through step by step on how to set this lava interface on your local computer but before that I regularly create videos in regards to Artificial Intelligence on my my YouTube channel so do subscribe and click the Bell icon to stay tuned make sure you click the like button so this video can be helpful for many others like you here's the table for our comparison you can see gerini pro version here and lava 34 billion parameter here you can clearly see this is 47.9 this is 51.1 this is 45.2 46.5 73.6 and 79.3 so in most of the cases this is better performance than Gemini pro multimodal model the main highlights are lava 1.6 achieves the best performance compared to open source llms such as these it has zero short Chinese capability and low training cost it costs 100 to thousand times smaller than others we can see the comparison here compared to the previous model 7 billion parameter and 13 billion parameter now we're going to see how we can set up this interface locally on your computer step by step for this purpose we are going to follow this repo I will put all the commands in the description below and also you can find that in this git repo you can find the spec required to run this model in the documentation page this is the configuration I'm using so the first step is get clone and then lava and then click enter next navigate to the folder now Conta create hyphen in lava python equal 3.11 and then click enter once that is done type cond activate lava to activate the virtual enrollments now pip install hyphen iPhone upgrade to upgrade the PIP package next pip install hyphen e and then click dot to install all the required packages once that is installed we need to set these three components first we are going to set up controller then we are going to set up the worker and finally we are going to set up the gradio interface to set up the controller python henm lava serve controller and it's going to run in port number 10,000 and then click enter keep this terminal running now we're going to to open a new terminal in the terminal navigate to the lava folder and cond activate lava to activate the lava the second step is to run the worker so we're going to type python hym lava. serve. model worker and this is going to run in port number 40,000 and we are connecting this worker to the controller by calling the controller command and the control is running in port number 10,000 here we're going to mention the lava 34 billion parameter model and we are going to load the 4bit quanti version to make this demo quicker next click enter this will automatically download the 34 billion parameter model and it might take some time because it's a large file you can see now it's loading the checkpoints now worker is ready now we have created the controller and the worker and finally we are going to start the gradio user interface so now open a completely new terminal keep the both other terminal running same like before navigate to the lava folder and then activate the virtual environment in the terminal python hym lava serve gradia web server we are providing the controller URL model list mode is reload and at the end we are providing the share so that we get the URL now we have the UI running so I'm going to open this URL the user interface is loading and we can see 34 billion parameter model in the models drop down list I'm going to select that choose example image and going to off ask what is unusual about this image about this image and then click sent the image depicts a man ironing a shirt while standing on a street which is unusual so that is correct next going to show this image and then going to ask what are the things I should be cautious about when I visit here and then click enter when visiting a location like one shown in the image which appears to be a lake with docks surrounded by mountains and Forest there are several things to be cautious about water safety weather condition Wildlife terrain personal safety leave no Tres equipment regulations altitude sickness emergency preparedness that is very detailed now I'm going to upload this image and try to ask if it can read the text in this image what is in this image the image shows a framed cardboard with the phrase I see a light in the darkness and that is correct next going to upload this image a bit more harder what is in this image the image shows a piece of graphity with the phrase take it easy written in a stylized handwritten font that is correct next going to upload this image what is in this image and then click enter the image shows a page from a planner or journal with a motivational quote written on it the quote reads every day is a fresh start that is amazing finally I'm going to test if it's a multilingual o C I'm going to upload this handwritten Tamil language text and check if it can identify this what is in this image enter the image shows a piece of paper with handwritten text in Tamil script the text appears to be a poem or piece of literature there are decorative elements such as flowers and a border around the text the writing done in blue ink and the paper has lined background I'm going to ask again what is written in this image I can see that it was not able to read Tamil language but that's fine it at least found out this is a Tamil script and for most of the images we saw in-depth in its details I'm really excited about this you can also run this lava using ol Lama run lava we can see the various versions here I hope the 34 billion parameter will be available soon here as well I'm going to create more videos similar to this so stay tuned I hope you like this video do like share and subscribe and thanks for watching