Joseph Pollack
AI & ML interests
Articles
Organizations
Tonic's activity
stepfun-ai/GOT-OCR2_0 is in top trending and spaces of the week for the second week straight !!
This is madness ๐ฑ
๐๐check out my demo here : Tonic/GOT-OCR
the Math one is absolutely incredible , the demo is great :-)
i'm doing something with desktop software about this , if you would like you can join me and discuss it here : http://chat.tonic-ai.com (discord)
thanks for the large_folder_upload it was not necessary in general , but i, as well as the public, kinda needed to be spoon fed , and for this i thank you ๐ค
Nvidia just released a small 4B Nemotron-mini model , and it works surprisingly well !
you can check it out here :
base : nvidia/Minitron-4B-Base
instruct : nvidia/Nemotron-Mini-4B-Instruct
demo : Tonic/Nemotron-Mini-4B
hoep you like it ๐ค๐ค
... and BIG THANKS for the cool PR on friday night ;-)
examples welcome if you have a cool one to show off for folks ;-)
this is AWESOME ! congrats on a cool release and an amazing collaboration ๐
@ucaslcl released a new OCR model , that's๐๐ป๐๐ป fantastic : https://huggingface.co/ucaslcl/GOT-OCR2_0
GPU : Tonic/GOT-OCR
Gradio Demo (Image Edit) : Tonic1/ImageEdit-GOT-OCR
Model : https://huggingface.co/ucaslcl/GOT-OCR2_0
Official demo : https://huggingface.co/spaces/ucaslcl/GOT_online
github : https://github.com/Ucas-HaoranWei/GOT-OCR2.0
made an image similarity demo to test out the mistral-community/pixtral-12b-240910 model .
If anyone knows how to generate captions with it , please do let me know x ๐
here's the demo : Tonic/Pixtral
hope you like it ๐ค
Did you see the new coding model from @01-ai ?
collection : 01-ai/yi-coder-66bdb00f5bdd611f9a008f30
demo : Tonic/Yi-Coder-9B
achieves SOTA on benchmarks , 125K context window , 55 languages including Docker, Js and many more ๐
โ๏ธInkubaLM has been trained from scratch using 1.9 billion tokens of data for five African languages, along with English and French data, totaling 2.4 billion tokens of data. It is capable of understanding and generating content in five African languages: Swahili, Yoruba, Hausa, isiZulu, and isiXhosa, as well as English and French.
model lelapa/InkubaLM-0.4B
demo Tonic/Inkuba-0.4B
just published a demo for Salesforce's new Function Calling Model
Salesforce/xLAM
- Tonic/Salesforce-Xlam-7b-r
- Tonic/On-Device-Function-Calling
just try em out, and it comes with
on-device
version too ! cool ! ๐lightning.ai
huggingface.co
many others have jupyterlab available and can scale ... for a price.
hope this helps !
May i ask : has the huggingface image been updated accordingly ?
your posts and demos are always sooooo cool
I found this cool (new?) thing by Docker called Testcontainers , and there's an @ollama object that you can use to programmatically serve ephemeral containers and LLMs.
I made a post about it here : https://huggingface.co/blog/Tonic/localai-testcontainers
It's really useful, powerful and fun !
Demo coming soon ๐ค
made a demo for Nvidia Minitron on an A100.
Minitron is a family of small language models (SLMs) obtained by pruning NVIDIA's Nemotron-4 15B model. We prune model embedding size, attention heads, and MLP intermediate dimension, following which, we perform continued training with distillation to arrive at the final models.
Deriving the Minitron 8B and 4B models from the base 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch; this results in compute cost savings of 1.8x for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. Please refer to our arXiv paper for more details.
Minitron models are for research and development only.
source : nvidia/Minitron-8B-Base
demo : Tonic/Minitron
https://huggingface.co/voyageai
i'm very happy for the huggingface team , it really makes sense to get closer together and not only for data ;-)
love the link github/git idea (not sure if it's actually possible)
it would be so much nicer if we kept posts about open source, machine learning and releases ... not like promotions and stuff... just my perspective on what kind of technical content i prefer to see on hugginface.
i was visting your v1 webapp this week , i cant wait for v2 now !
- what's your number for real ?
+ and did it work at parties for you ?
it's not a bad idea, would be nice to have a bridge with git based on verified email, but i guess you know that already.
would be nice to track datasets and models more than demos . something like a design that's not an exact copy of micro$$oftgithub would be nice... but i dont have a solution for you...
my main ask about the hub is better control over notifications , that would be tremendously useful...
interesting ! i'm quite curious ... i was already struggling to keep the pace with gradio v4+ and now i'm looking forward to v.1 once again , meanwhile i really really think there depth and breadth of gradio deserves a comprehensive tutorial/course , and no the docs (they are fine, they are good) are not enough about it ... not a criticism , just a request from a big fan :-)
@tiiuae released Falcon 11B Vision Model !
๐ฆ ๐ฆ ๐๐
it's quite good , and you can try it here : Tonic/Falcon-Vision
this is actually amazing + very cool/interesting i'm very happy i found the paper and the models.
congratulations on the tencent collaboration , i'm looking forward to the future.
i got an email my space was down so now my space is back up ! StarCoder2 (Raw) on A100 , for your enjoyment and apache2 research purposes ๐๐ป๐๐ป
Tonic/starcoder2
check my profile for more cool GPUzero demos, i'll cycle them with some new overlooked models soon ๐๐ค
https://huggingface.co/spaces/Tonic/Yi-9B
it's true ! and a lot of successful base models too !
lol i gave a follow
i've had "best" results mushing everything into a single context window with a single "final"/"next" answer , i think i remember @teknium saying they often do that and they may have published that research , but i cant speak for them, i just remember them saying that and feeling validated :-)
After today it's gone !
actually not - just joking ! it's <3 open source !
just trying to get folks' attention to my featured "Spaces of the Week" :
Tonic/starcoder2
drop a like for your boy and join us next week for making fine tunes !
and we made it to last place on trending.
i really thought it couldnt get any better, but i'm crying ! ๐ญ
The thing i like the most about ZeroGPU ,
import spaces
, is that i dont have to always check to see if someone decided to test if i have hard character limits , and it reloads the application flawlessly . drop a like on my spaces here :
Spaces of the Week : https://huggingface.co/spaces/tonic/starcoder2
9 other ZeroGPU demos : https://huggingface.co/tonic
it's just a base model but you can check it out here : Tonic/Yi-9B
cant wait to fine tune this one ๐ค๐
Star coder came out and it's really fascinating in more ways than one !
first off it codes well already. but secondly it's reported to "know" 101 programming languages !
that actually means it's ripe for fine tunes, so if you're like me you've been bookmarking cool datasets and cant wait to get started !
that said , here's a cool demo where you can try it out now : Tonic/starcoder2
turns out it can program a T5 demo using gradio !
๐คAya has been released ! It's an absolutely massive undertaking to create a huge multilingual dataset and multilingual model of very high quality.
Papers :
https://cohere.com/research/papers/aya-dataset-paper-2024-02-13
https://cohere.com/research/papers/aya-model-paper-2024-02-13
Model : CohereForAI/aya-101
Dataset : CohereForAI/aya_dataset
I am proud to be one of 3,000 humans who built Aya - a new massively multilingual, generative LLM that outperforms existing open-source models and covers 101 different languages. Together, we are accelerating multilingual AI. ๐ค
i'm so impressed by the reception that https://huggingface.co/collabora/whisperspeech has recieved !
Check out the cool demo here : collabora/WhisperSpeech
- Open issue : how do we provide MPS support ? cc. @intel :-) looking into this now , any leads welcome!
check out also [collabora/whisperfusion](https://github.com/collabora/WhisperFusion)
hope you enjoy ! ๐ค
here's one until i make the PR later today : https://discord.gg/QCYXNAkGxV
hey thanks for pointing that out, there's so much to organise and build for this, help is really welcome if you like the subject :-)
check out this Chest X-Ray model from AIMIStanford : Tonic/CheXRay
thanks to @lunarflu for kicking me a bit to get the examples in there !
would be great to get even more examples and even more downstream functions , so contributions are very welcome, or if you have a dataset source, please do share it in the discussions !
whisperspeech
just :
pip install whisperspeech
to get started and check out my demo to do multilingual text to speech including making voice prints using
whisperspeech
reverse engineering of whisper here :
Tonic/whisperspeechand the model card here : https://huggingface.co/collabora/whisperspeech
i met collabora on LAION check out LAION here :
https://huggingface.co/laion
I launched my first competition !
Goal : Use AI to beat the Math Olympics within the set time
Basically we're looking for adventurous teams and individuals to make a common submission to the AI Math Olympics by the MLCommons.
Althought the ultimately there can only be one winner and there must always be a winner, the ultimate goal is to get together for a common solution.
check it out here :
Tonic1/mathathon
the recent e5mistral7B embeddings model uses prompts to specifically tailor embeddings to specific use cases. check it out : https://huggingface.co/spaces/Tonic/e5
i quite like using specialized models to test them out too https://huggingface.co/spaces/TeamTonic/hallucination-test
i'm a fan of this community project : to train sector-specific 32K-context BERT embedding models ๐ค
cant wait to get this one working but right now i'm getting an error ๐ค
cant wait to participate in yours and host some of mine :-)
Everyone's๐ฃ๏ธtalking about microsoft's new e5mistral embeddings model
๐ค๐ค but did you actually try it yet ?
Well , now you can, just check it out. it's a new way to serve and create embeddings.
try it hosted on GPUZero : Tonic/e5
or served on an A10G : Tonic/e5
you get best results actually building with it though, so use it in your app !
Our demo is coming soon too so let's work together if you want :-)
our target is to pursue LowRes animated waifus and husbandos and be the leading frontrunners of anime related content ( โขฬ ฯ โขฬ )y
So right now they're gathering cool datasets, soon we'll make and serve some LORAs, then we'll build with these for a little bit more interesting and simple anime applications , actually we already started at least that part :-)
you should check it out, it's a wild and massive community of i think of a quarter million folks on facebook - if you add up all the parts.
actually i should tag @not-lain because i basically take directions from him, perhaps he can say more :-)
i wanted to share with you a really cool new organisation called
https://huggingface.co/lowres
In just one week it has gathered almost 150 members !
Check them out if you love anime , SDLX, LORAs and cool datasets.
can we make this one reach 200 members? ๐
- facebook's Seamless M4T for the audio and voice to make multilingual , there's an ondevice model demo here: https://huggingface.co/spaces/Tonic/SeamlessOnDevice
- one for YI-200K , but it actually doesnt quite fit on a GPUZero... U_U
- one for SDXL style align, but omg i didnt even realize at the time it wasnt my demo of it (lol)
- one for texify (which works great btw, keep an eye on texify, it's about to blow up... in a couple of months!)
so yeah, i ran back and tried to get my demos working at least for sdxl which i love , but i simply couldnt get the CPU stuff working, or the refactored code working. no wonder i was thinking "wow this is so easy" on @osanseviero 's demo : yeah , it's not my code that's why it works ๐ ๐๐ป
anyway spent the day unsuccessfully experimenting, but starting tomorrow i'll try to serve some cool and overlooked models so ๐คhuggingface appreciators can try them out ๐
- just a ๐ ๏ธbuilder from ๐ผParis !
Everyone is making something special for their first post , so since i got access to **GPUZero** , well, my first post is about **GPUZero**
### GPUZero is here !
This one's great for builders like me that are often making and serving models to their community.
- demos get popular then fade away
- they retain interest over the next three months as folks have questions
**GPUZero** lets you serve demos to your community over time while optimizing for costs .
Believe it or not it's actually impossible to pay for everything over a whole month if you have even one GPU running at a time.
I'm so excited for this because it lets me serve a complete stack of specialized models and to build with them too.
- all optimized for efficiency in dollar cost.
check out some demos that are available on GPUZero :
- Tonic/marker-texify : this one is the first one i made it's for an image to latex formula model.
- https://huggingface.co/spaces/Tonic/YI-6B-200k : this one probably actually works better on GPUZero than on a standard A10, but dont take my word for it , try it out ๐ค
- https://huggingface.co/spaces/Tonic/style-aligned_sdxl : this one was my greatest technical achievement, check the dates and times on it too, there's a backstory to this one so i'll maybe tell it in another post
i love the "posting" from arguilla , what a fantastic way to share ๐ค