Understanding Hugging Face Model Cards: A Comprehensive Guide

hi everyone this is Fahad mza and I welcome you to the channel this video has been sparked by one of the comments

I received on one of the videos where I was showing you the installation of large language models locally the

question was very much relevant as it was coming from someone who is just starting out in Ai and the person was

wondering as once we go to the model card of hugging face model what exactly is meant by all of these

files so in this video I am going to explain each one of these files in as simple words as possible in fact I'm

just going to explain this whole page which is related to this Lama 3.2 3 billion instruct model but you can of

course go to any model page on hugging face and more or less it will be similar and if you don't know how to do that

just click on this model on the top and then you will see that there are thousands of models already hosted on

hugging fist maybe hundreds of thousands now so if you just open any model it should be similar one way or another

going back to our this model card now let's try to decifer and dissect and try to understand in as simple words as what

exactly is meant by this page so this is a model card which provides an overview of the model including its architecture

performance benchmark marks license and other relevant details the most important part here is this files and

version the middle tab the third tab is where you can ask questions you can uh and this is where

the model creator gives lot of updates as um what exactly is happening there are also some closed issues so model is

hosted just like if you have seen the GitHub repo where you can raise issues you can ask questions and that sort of

stuff one of the problem is that I don't see that this gets updated as much as it should be so don't expect the answer

immediately or at all for most of the model providers I believe anyway the first step model card is just

a read me file and you can find all the related files here in files and version and that is the main I would

say focus of this video so this section contains the actual model files let's break down each file's purpose so I will

just simply start from the Top If You see there is a original folder in this original folder there is py torch format

of the model not only it comes in pytorch but also it comes in safe tensor as you can see here and that is

primarily what we use in order to do the entrance and I will explain it shortly what exactly is meant by

that the next file in the sequence is this dogit attribute it is a standard configuration file from the git that

tells git how to handle specific files or directories in a repository it defines attributes such as line endings

merch strategies and more and you can also include lot of ignore statements here so this is related to um for

example I would say different packages and different stuff you within the GitHub

repo and now in summary if you look at this content so in this case this dogit attribute file is configuring git to use

large file storage or lfs for various file types including archives models and data files lfs is used to efficiently

store and manage large files in a repository because mod model files are large and this file is specifying that

lfs should be used for filtering diffing and also merging these file types okay let's go back I'm just clicking on this

files and versions and then the next important file is here is license.txt as the name suggests this is the licensing

file which describe all the licensing of this model now if you are looking to use it in production environment it is

extremely important to read and understand this file so that you don't violate the licensing and I cannot

stress it enough don't take it easy because some companies are pretty precious about their licensing and if

they find out that you are using the model without permission and you're violating the license the legal

penalties could be very high especially in Europe so make sure that you are aware of it okay so now we know what

that is and then we have this readme.txt it is in the markdown format so the first file which you see here this is

what this read me file is in this first model card and you can also open it and check it out it will all be the markdown

format here okay let's go back and then we have this use policy

this is um and this is not the standard not every model has all of these files some some of them just combine this with

the licensing but Lama has this acceptable use policies as where you can use this that model and where you cannot

use this model it is also important to read okay and then we have some of the technical files starting from config.js

now if I click on this config.js let me explain what is happening in this file so this file contains metadata about the

models architecture including layer activations layer activations means mathematical functions that introduce

non linearity into the model sizes vocabulary size vocabulary size means the number of unique words or tokens in

the models vocabulary and then number of attention heads which is a mechanism that allows the model to focus on

specific parts of the input data and the model Precision which is a numerical Precision used to represent the models

weights and activations the Transformer Library uh uses this whole config.js file to build

the model architect Ure once we download it so pretty important file I would say then we have this

generation config.js now this is similar to what we just saw above but this is more related to I would say the

inference this file contains metadata for inference inference simply means the process of using a model to generate

output for new unseen input data now this contains things like temperature and top P topk thresholds which control

the output and I already have covered it in lot of videos on my channel now you can also specify context

window size which is amount of input data the model considers when generating output and token ID for special

tokens okay let's go back and then we have the safe tensor files before I tell you about them let me also quickly

introduce you to the sponsors of the video who are entp ENT bot lets you effortlessly deploy a personalized

knowledge bot across platforms like Discord slack and others it is ideal for open source Tech communities and

startups that provide user support and I will drop the link to their website in videos

description okay let's go back here so now we have at this safe tensor files what exactly these are now in short

these are the model weights so due to the large size of llm which could be billions of P

parameters the model weights are split into smaller files for safe downloading so you see we have two safe tensors here

these files come in either bin or do safe tensor format by the way safe tensor is a newer more secure format

proposed by hugging phase as an alternative to the default py toch format because there are some security

vulnerabilities when it comes to py toch format okay then we have model save tensor index do Json if I just open this

file now this file stores a sequential map of the models architecture specifying which part file each layer's

weights are stored in layer is simply a component of model that processes input data through a set of mathematical

operations allowing the network to learn complex patterns and a weight is a numerical value that determines the

strength of connection between two neurons in the model that influences how much each neur neuron's output

contributes to the next layer input so this is what con contains that very important file by the way for the inner

workings of the model and then we have something this special token map let me open it

now so again this file is related to the tokenizer and tokenizer is a component of a model that breaks down input text

into individual token or words so this contains like beginning of sentence end of sentence so that the model would know

when a sentence is beginning or ending and then going forward we have again this tokenizer do Json file again

two it's it contains the whole tokenizer so it can uh similar to other file it contains the whole tokenizer and

it also gets the information from this special token file and this tokenizer config.js on this contains the tokenizer

class name layer names input processing configuration and all the related information to the Token which is

required for it to work properly let's go back to the files and versions and these are all the files

which we uh normally get in a model there are some specific models in image or in other where we also find some

other files like for example if you go to a model like flux Dev which is a text image model you might get some um text

encoders you might get some vae files which are the variational auto encoder to convert our image from latent space

to pixel space so and I already have covered that on the channel so if you encounter any

file just search it on the channel you should be able to find a video around it if not let me know happy to cover it for

you but overall these are the major standard files which we use I hope that this was useful if you still have any

questions feel free to ask if you like the content please consider subscribing to the channel if you're already

subscribed please share it among your network as it helps a lot thank you for watching