The Rapid Growth of Hugging Face Models: Trends and Insights

We are seeing almost 600,000 unique users downloading models from Hugging Face, which is a significant increase in recent times. This trend is quite notable, and it's essential to understand what's driving this growth. One plot that showcases this trend is the number of unique users downloading models. In the pre-plot, we can see the number of unique users unloading the model, while the current plot shows the number of Llama-based repositories being created on Hugging Face. The former indicates the popularity of certain models, whereas the latter highlights the proliferation of new Llama-based models.

Llama is an architecture that has gained immense popularity in recent times. However, it's essential to note that not all Llama-based models are official ones. Many people have been training and sharing fine-tuned models with similar architectures but without being officially affiliated with the Llama framework. This proliferation of new models is indeed impressive, with between 3,000 to 4,000 new Llama-based models being created every week. Moreover, the plot shows a Lama-shaped trend, indicating a growth curve that's not unlike a typical startup trend.

Another significant trend we're witnessing is the emergence of small but powerful models. Just a few years ago, there was a limited set of very small models available. However, now we see a plethora of powerful models at small scales. For instance, there are GMA-5 models in both text and vision spaces, as well as various smaller variants. Additionally, models like Whisper, PEAR, and MobileBERT have made significant strides in speech recognition and computer vision tasks. Furthermore, there are 20 different variations of MobileBERT alone, showcasing the vast range of smaller models that can be achieved while maintaining high performance.

While Hugging Face is renowned for its Transformers-based models, it's essential to note that the platform hosts a wide range of models from various libraries. These include reinforcement learning, computer vision, and more. The model download plot showcases this diversity, with a significant number of users opting for models from different libraries like FAIR (facebook ai research), SQuAD, and more. Furthermore, Hugging Face has an impressive collection of over 1.5 million models in total.

The creation of new models is another trend that's worth highlighting. According to the plot, every single week, between 30,000 and 40,000 new models are being shared on Hugging Face. This is a significant increase compared to when our founder joined Hugging Face about three-and-a-half years ago, when there were only a few thousand new models being created weekly. This growth in model creation is indeed remarkable.

The Architecture Life Cycle: Understanding the Rise and Fall of Models

When a new model architecture emerges, it often sparks excitement among researchers and practitioners alike. However, this excitement can be short-lived if the model doesn't gain widespread adoption or longevity. There are typically two possible outcomes for a new model's life cycle: it either gains long-term usage or adoption or fails to do so.

To help extend a model's life cycle and foster its adoption, there are several strategies that can be employed. Firstly, focusing on a single architecture rather than constantly changing or experimenting with different architectures can help maintain user interest. Llama is an excellent example of such an architecture, which has become synonymous with the rapid growth of generative text models.

Secondly, transparency and openness about the model's training data and usage are crucial for its adoption. Not all newly created models share their training data or provide clear information on how to use them. This lack of transparency can lead to a decrease in user interest and adoption over time.

Lastly, avoiding complex licenses that might deter users from adopting the model is essential. A simple and consistent API that aligns with other Hugging Face models can go a long way in making it easier for users to integrate the new model into their workflow. Avoiding complicated licensing agreements will also make it more likely for new models to be adopted, which is crucial for any open-source platform like Hugging Face.

In conclusion, understanding the trends and insights behind the rapid growth of Hugging Face models can provide valuable lessons for researchers and practitioners alike. By focusing on a single architecture, maintaining transparency, avoiding complex licenses, and promoting consistency, we can foster a more inclusive community around model creation and adoption.

"WEBVTTKind: captionsLanguage: enhi everyone so I Omar Sano I'm the chief L officer at hugin face and today I will talk about how to distribute a million open models in the wild ER and my goal with this short talk is just to share some of the lessons learned from the huging face Hub uh an alternative title for the talk is open model SC H because some of these strs are quite fun so just to begin with a quick quiz uh how many open models does huging face host so a 250k b c so who thinks that we have 1 million open models or D 1.5 so that's the title of the talk so it's a 1 million it's kind of a an easy question okay so how many py models do Hing face host so who thinks that we have a 300K 500 700 1 million okay so not all models on huging face are by torch uh but yeah have million models that number is quite impressive if you put it in like the whole context okay so how many terabytes do you think face Earth daily a 300 terabytes a day 500 1 petabyte six petabytes okay so people are looking at Twitter because H yeah it's six petabytes H so every single day in July nowadays it's a actually a bit higher we serve about six petabytes uh and over 1 billion request a day in big part thanks to Lama 400 5p it's a large Model H but yeah so to switch a bit let me share a bit about security H so if you know H many libraries including pych use pickle as a way to persist models as a way to share models and you might know that these models by default are unsafe so it pickle files con contain the weights but you can also execute arbitrary code so if you're downloading a model you might be running code that you don't want to run uh so there are a few ways you can do H to avoid this so the first one the easy is to just load models from people you trust so you might just want to download models from a couple of organizations a p now has this nice parameter called weights only equals through so that means that you will just load the tensors and then you will make sure that you're are not loading any potentially malicious code so that's nice as well on huging face side we have a thing called the pickle scanning so we scan pickle files for potential malware we look at the Imports within the pickle to see if there are potential things such as running any arbitrary code of course none of this is perfect uh but this helps you identify potentially malicious pickle files and alternatively you can use other calization uh formats so we have developed with a couple of other partners the safet and Source format that allows people to share models uh with a persistent uh format that don't that will not contain any arbitrary code uh so yeah that was more of a parenthesis about security but that's something that is quite quite good to know so now let me share a bit of interesting stats uh so these are a here you can see in the y axis a unique number of users that are downloading models for different tasks so this slide in particular is for uh for different computer vision tasks so P classification text to text and uh this is quite nice you can see the trend that for example uh model the the red line for example that's image text to text so those are Vision language models so models that take us input both images and text such as pixal or qu BL and you can see clearly this year uh the number of people downloading these models has grown significantly purple is a image classification so that's a kind of expected but you can see the overall trend for all of these different tasks is going up okay so Vision language model is go uh actually all modalities are going you can see for example NLP vision is also going quite well at one point it almost overtook the NLP models but the growth has been a bit less explosive uh but then we're also seeing AIO models uh multimodal models uh there has been uh very nice Trends here so this is a plot of Lama based models on the Hub so how many unique users are downloading Lama based models and you can clearly see like a very nice Trend here so we are seeing almost 600k H unique users downloading the model uh there's this other plot so in the pre in the last plot here I'm just showing number of unique users unloading the model here what I'm showing is the number of Lama based repositories being created on huging phas so as you know Lama is just an architecture there are many people training or sharing fine tune models of Lama or models with the same architecture but it's not a official Lama models right so here you can see how many new models are being created uh with the lamb architecture every single week uh in the last few weeks it has been between three and 4,000 models H if you think about it it that's crazy you we're having 3 to 4,000 new Lama based models every single week what is even more impressive is that if you look at the plot it's kind of a Lama shaped so I really like that H another Trend we're seeing is that the small models are also H going in a very nice Direction so few years ago we had a a very small set of very small models now we are seeing like very powerful models at very small scales right so there is GMA five model M quen that's in the text space in the vision space there is some death anything mobile vit for some there are 20 different variations that are smaller than the original ones there's whisper par TTS for the speech recognition and generative speech and finally in multimodal there's mobile CLE Nano lava Florence 2 Moon dream and there are a bunch of other models now huging face is very well known for Transformers uh but huging face does not host only Transformers models we actually host models for many different libraries that go from reinforcement learning uh computer vision I mean trans also has computer vision but there are other libraries such as team uh so yeah so here you can see again the number of unique people downloading models for different libraries ER okay safe tensors is the one at the top but you can see for example people are using the fusion models quite a bit people are using GF sentence Transformers H set fit team many different libraries so it's really a platform for all kinds of libraries and finally uh all models are really going uh so this again is a plot of how many new models we are having every single week so in total right now hugfest has 1.5 million models that's close Plus open models in total and if you see the trend in these last couple of weeks uh every single week there are between 30 and 40 sorry yeah 30 and 40,000 new models every single week so when I joined hogen phase three and a half years ago we had maybe a few thousand new models every week so it's quite fun quite quite impressive to really think there are 40,000 new models being shared every single week H just to close a bit h a bit the talk I want to share a bit more about the architecture life cycle so when there's a new model architecture usually people get extremely excited there are some tweets by famous people H people start to play with it maybe it doesn't perform as well as they expected and then usually the US the users the excitement the hype kind of goes down and there are kind of two options two pills so on one hand uh the model simply doesn't get the long-term uh usage or adoption that was expected on the other hand it might kind of recuperate and might have quite a bit of usage in the long term so there are a couple of things that can help uh this architectural life cycle go as the rest of the models so first of all H focus on a single architecture rather than trying to change the architecture two often uh Lama is one of the most common architectures and most of the new generative text models are actually very close to L the L architecture H the second thing that helps quite a bit in adoption is transparency and openness so not all of the new models uh that are open H are transparent with what which was the data that was used to train the models so in general being more transparent about the data set about how to use the model having very nice model car very nice data set car that kind of stuff can help quite a bit with a adoption of the model uh the third point is avoiding complex uh licenses so if your license is extremely confusing then the legal teams in many companies will not want to use the model just because it's a very different license if you use MIT Apache then everyone will uh will be able to use the model easily at least from a license perspective but if you have a very custom license that can deteriorate ad option of the model a simple API consistent with the rest of the model so try to avoid like a super weird chat template that will help quite a bit uh so yeah that's pretty much my talk thanks a lothi everyone so I Omar Sano I'm the chief L officer at hugin face and today I will talk about how to distribute a million open models in the wild ER and my goal with this short talk is just to share some of the lessons learned from the huging face Hub uh an alternative title for the talk is open model SC H because some of these strs are quite fun so just to begin with a quick quiz uh how many open models does huging face host so a 250k b c so who thinks that we have 1 million open models or D 1.5 so that's the title of the talk so it's a 1 million it's kind of a an easy question okay so how many py models do Hing face host so who thinks that we have a 300K 500 700 1 million okay so not all models on huging face are by torch uh but yeah have million models that number is quite impressive if you put it in like the whole context okay so how many terabytes do you think face Earth daily a 300 terabytes a day 500 1 petabyte six petabytes okay so people are looking at Twitter because H yeah it's six petabytes H so every single day in July nowadays it's a actually a bit higher we serve about six petabytes uh and over 1 billion request a day in big part thanks to Lama 400 5p it's a large Model H but yeah so to switch a bit let me share a bit about security H so if you know H many libraries including pych use pickle as a way to persist models as a way to share models and you might know that these models by default are unsafe so it pickle files con contain the weights but you can also execute arbitrary code so if you're downloading a model you might be running code that you don't want to run uh so there are a few ways you can do H to avoid this so the first one the easy is to just load models from people you trust so you might just want to download models from a couple of organizations a p now has this nice parameter called weights only equals through so that means that you will just load the tensors and then you will make sure that you're are not loading any potentially malicious code so that's nice as well on huging face side we have a thing called the pickle scanning so we scan pickle files for potential malware we look at the Imports within the pickle to see if there are potential things such as running any arbitrary code of course none of this is perfect uh but this helps you identify potentially malicious pickle files and alternatively you can use other calization uh formats so we have developed with a couple of other partners the safet and Source format that allows people to share models uh with a persistent uh format that don't that will not contain any arbitrary code uh so yeah that was more of a parenthesis about security but that's something that is quite quite good to know so now let me share a bit of interesting stats uh so these are a here you can see in the y axis a unique number of users that are downloading models for different tasks so this slide in particular is for uh for different computer vision tasks so P classification text to text and uh this is quite nice you can see the trend that for example uh model the the red line for example that's image text to text so those are Vision language models so models that take us input both images and text such as pixal or qu BL and you can see clearly this year uh the number of people downloading these models has grown significantly purple is a image classification so that's a kind of expected but you can see the overall trend for all of these different tasks is going up okay so Vision language model is go uh actually all modalities are going you can see for example NLP vision is also going quite well at one point it almost overtook the NLP models but the growth has been a bit less explosive uh but then we're also seeing AIO models uh multimodal models uh there has been uh very nice Trends here so this is a plot of Lama based models on the Hub so how many unique users are downloading Lama based models and you can clearly see like a very nice Trend here so we are seeing almost 600k H unique users downloading the model uh there's this other plot so in the pre in the last plot here I'm just showing number of unique users unloading the model here what I'm showing is the number of Lama based repositories being created on huging phas so as you know Lama is just an architecture there are many people training or sharing fine tune models of Lama or models with the same architecture but it's not a official Lama models right so here you can see how many new models are being created uh with the lamb architecture every single week uh in the last few weeks it has been between three and 4,000 models H if you think about it it that's crazy you we're having 3 to 4,000 new Lama based models every single week what is even more impressive is that if you look at the plot it's kind of a Lama shaped so I really like that H another Trend we're seeing is that the small models are also H going in a very nice Direction so few years ago we had a a very small set of very small models now we are seeing like very powerful models at very small scales right so there is GMA five model M quen that's in the text space in the vision space there is some death anything mobile vit for some there are 20 different variations that are smaller than the original ones there's whisper par TTS for the speech recognition and generative speech and finally in multimodal there's mobile CLE Nano lava Florence 2 Moon dream and there are a bunch of other models now huging face is very well known for Transformers uh but huging face does not host only Transformers models we actually host models for many different libraries that go from reinforcement learning uh computer vision I mean trans also has computer vision but there are other libraries such as team uh so yeah so here you can see again the number of unique people downloading models for different libraries ER okay safe tensors is the one at the top but you can see for example people are using the fusion models quite a bit people are using GF sentence Transformers H set fit team many different libraries so it's really a platform for all kinds of libraries and finally uh all models are really going uh so this again is a plot of how many new models we are having every single week so in total right now hugfest has 1.5 million models that's close Plus open models in total and if you see the trend in these last couple of weeks uh every single week there are between 30 and 40 sorry yeah 30 and 40,000 new models every single week so when I joined hogen phase three and a half years ago we had maybe a few thousand new models every week so it's quite fun quite quite impressive to really think there are 40,000 new models being shared every single week H just to close a bit h a bit the talk I want to share a bit more about the architecture life cycle so when there's a new model architecture usually people get extremely excited there are some tweets by famous people H people start to play with it maybe it doesn't perform as well as they expected and then usually the US the users the excitement the hype kind of goes down and there are kind of two options two pills so on one hand uh the model simply doesn't get the long-term uh usage or adoption that was expected on the other hand it might kind of recuperate and might have quite a bit of usage in the long term so there are a couple of things that can help uh this architectural life cycle go as the rest of the models so first of all H focus on a single architecture rather than trying to change the architecture two often uh Lama is one of the most common architectures and most of the new generative text models are actually very close to L the L architecture H the second thing that helps quite a bit in adoption is transparency and openness so not all of the new models uh that are open H are transparent with what which was the data that was used to train the models so in general being more transparent about the data set about how to use the model having very nice model car very nice data set car that kind of stuff can help quite a bit with a adoption of the model uh the third point is avoiding complex uh licenses so if your license is extremely confusing then the legal teams in many companies will not want to use the model just because it's a very different license if you use MIT Apache then everyone will uh will be able to use the model easily at least from a license perspective but if you have a very custom license that can deteriorate ad option of the model a simple API consistent with the rest of the model so try to avoid like a super weird chat template that will help quite a bit uh so yeah that's pretty much my talk thanks a lot\n"