Share and discover more about AI with social posts from the community.huggingface/OpenAi
Neural Network (1 Byte explainer for everybody)
Just like our brain, a Neural Network is made up of interconnected "neurons". These neurons work together by learning from (input) data and getting better at tasks (in the hidden layer) to give (output) predictions or decisions.
Just like our brain, a Neural Network is made up of interconnected "neurons". These neurons work together by learning from (input) data and getting better at tasks (in the hidden layer) to give (output) predictions or decisions.
𝗔𝗜𝟮𝟭 𝗶𝘁𝗲𝗿𝗮𝘁𝗲𝘀 𝘄𝗶𝘁𝗵 𝗻𝗲𝘄 𝗝𝗮𝗺𝗯𝗮 𝟭.𝟱 𝗿𝗲𝗹𝗲𝗮𝘀𝗲: 𝗡𝗲𝘄 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗳𝗼𝗿 𝗹𝗼𝗻𝗴-𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝘂𝘀𝗲-𝗰𝗮𝘀𝗲𝘀!🏅
@ai21labs used a different architecture to beat the status-quo Transformers models: Jamba architecture combines classic Transformers layers with the new Mamba layers, for which the complexity is a linear (instead of quadratic) function of the context length.
What does this imply?
➡️ Jamba models are much more efficient for long contexts: faster (up to 2.5x faster for long context), takes less memory, and also performs better to recall everything in the prompt.
That means it’s a new go-to model for RAG or agentic applications!
And the performance is not too shabby: Jamba 1.5 models are comparable in perf to similar-sized Llama-3.1 models! The largest model even outperforms Llama-3.1 405B on Arena-Hard.
✌️ Comes in 2 sizes: Mini (12B active/52B) and Large (94B active/399B)
📏 Both deliver 256k context length, for low memory: Jamba-1.5 mini fits 140k context length on one single A100.
⚙️ New quanttization method: Experts Int8 quantizes only the weights parts of the MoE layers, which account for 85% of weights
🤖 Natively supports JSON format generation & function calling.
🔓 Permissive license *if your org makes <$50M revenue*
Available on the Hub 👉
ai21labs/jamba-15-66c44befa474a917fcf55251
Read their release blog post 👉 https://www.ai21.com/blog/announcing-jamba-model-family
@ai21labs used a different architecture to beat the status-quo Transformers models: Jamba architecture combines classic Transformers layers with the new Mamba layers, for which the complexity is a linear (instead of quadratic) function of the context length.
What does this imply?
➡️ Jamba models are much more efficient for long contexts: faster (up to 2.5x faster for long context), takes less memory, and also performs better to recall everything in the prompt.
That means it’s a new go-to model for RAG or agentic applications!
And the performance is not too shabby: Jamba 1.5 models are comparable in perf to similar-sized Llama-3.1 models! The largest model even outperforms Llama-3.1 405B on Arena-Hard.
✌️ Comes in 2 sizes: Mini (12B active/52B) and Large (94B active/399B)
📏 Both deliver 256k context length, for low memory: Jamba-1.5 mini fits 140k context length on one single A100.
⚙️ New quanttization method: Experts Int8 quantizes only the weights parts of the MoE layers, which account for 85% of weights
🤖 Natively supports JSON format generation & function calling.
🔓 Permissive license *if your org makes <$50M revenue*
Available on the Hub 👉
ai21labs/jamba-15-66c44befa474a917fcf55251
Read their release blog post 👉 https://www.ai21.com/blog/announcing-jamba-model-family
Ai21
The Jamba 1.5 Open Model Family: The Most Powerful and Efficient Long Context Models
The new family of open models from AI21, offering unrivaled speed, efficiency, and quality and the longest context window among open models.
The latest timm validation & test set results are now viewable by a leaderboard space:
timm/leaderboard
As of yesterday, I updated all of the results for ImageNet , ImageNet-ReaL, ImageNet-V2, ImageNet-R, ImageNet-A, and Sketch sets. The csv files can be found in the GH repo https://github.com/huggingface/pytorch-image-models/tree/main/results
Unfortunately the latest benchmark csv files are not yet up to date, there are some gaps in dataset results vs throughput/flop numbers impact the plots.
h/t to @MohamedRashad for making the first timm leaderboard.
timm/leaderboard
As of yesterday, I updated all of the results for ImageNet , ImageNet-ReaL, ImageNet-V2, ImageNet-R, ImageNet-A, and Sketch sets. The csv files can be found in the GH repo https://github.com/huggingface/pytorch-image-models/tree/main/results
Unfortunately the latest benchmark csv files are not yet up to date, there are some gaps in dataset results vs throughput/flop numbers impact the plots.
h/t to @MohamedRashad for making the first timm leaderboard.
GitHub
pytorch-image-models/results at main · huggingface/pytorch-image-models
The largest collection of PyTorch image encoders / backbones. Including train, eval, inference, export scripts, and pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (V...
Huge updates and improvements for FLUX LoRA training : https://www.patreon.com/posts/kohya-flux-lora-110293257
10 GB, 16 GB, 24 GB and 48 GB GPU configs added - 10 GB config is like 3x to 5x slower sadly
Massed Compute, RunPod and Windows Kohya SS GUI LoRA installers added to the zip file
Also right now testing new 16 GB FLUX LoRA training config and new way of regularization images. Moreover testing Apply T5 Attention Mask too. Lets see if Kohya FLUX LoRA workflow will become even better or not
Also massive grids comparisons shared here : https://www.reddit.com/r/StableDiffusion/comments/1eyj4b8/kohya_ss_gui_very_easy_f
10 GB, 16 GB, 24 GB and 48 GB GPU configs added - 10 GB config is like 3x to 5x slower sadly
Massed Compute, RunPod and Windows Kohya SS GUI LoRA installers added to the zip file
Also right now testing new 16 GB FLUX LoRA training config and new way of regularization images. Moreover testing Apply T5 Attention Mask too. Lets see if Kohya FLUX LoRA workflow will become even better or not
Also massive grids comparisons shared here : https://www.reddit.com/r/StableDiffusion/comments/1eyj4b8/kohya_ss_gui_very_easy_f
Patreon
Kohya FLUX LoRA and Fine Tuning Training Full Tutorial For Local Windows and Cloud RunPod and Massed Compute for Research & Development…
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Shoutout to everyone who participated in BigScience! Doesn't get enough credit but IMO paved the way for open-source LLMs!
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (2211.05100)
bigscience/bloom
bigscience/bloomz
https://huggingface.co/bigscience/bloom
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (2211.05100)
bigscience/bloom
bigscience/bloomz
https://huggingface.co/bigscience/bloom
huggingface.co
bigscience/bloom · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
ACL 2024: The Missing Papers
Apparently, some papers from the ACL 2024 are still not listed in the ACL Anthology. While this issue will hopefully be fixed soon, we should give those papers additional spotlight.
Some of my favorites:
1. Dolma is an English corpus that encompasses 3 trillion tokens. Additionally, it is accompanied by an exceptional software package that consdierably advances the state-of-the-art in preparing data for LLM pretraining. (Source: I am currently using Dolma.)
Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research (2402.00159)
2. In the paper "Same Task, More Tokens: the Impact of Input Length on
the Reasoning Performance of Large Language Models", the authors show how extending the context length impacts an LLM's reasoning performance. I asked myself a similar question a few months ago, and therefore this paper is highly interesting to me.
Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models (2402.14848)
This was brought to my attention through a Linkedin post by @ShayeghB, who is also affected:
Ensemble-Based Unsupervised Discontinuous Constituency Parsing by Tree Averaging (2403.00143)
View all the missing papers here:
https://theshayegh.github.io/ACL2024MissingPapers/
Apparently, some papers from the ACL 2024 are still not listed in the ACL Anthology. While this issue will hopefully be fixed soon, we should give those papers additional spotlight.
Some of my favorites:
1. Dolma is an English corpus that encompasses 3 trillion tokens. Additionally, it is accompanied by an exceptional software package that consdierably advances the state-of-the-art in preparing data for LLM pretraining. (Source: I am currently using Dolma.)
Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research (2402.00159)
2. In the paper "Same Task, More Tokens: the Impact of Input Length on
the Reasoning Performance of Large Language Models", the authors show how extending the context length impacts an LLM's reasoning performance. I asked myself a similar question a few months ago, and therefore this paper is highly interesting to me.
Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models (2402.14848)
This was brought to my attention through a Linkedin post by @ShayeghB, who is also affected:
Ensemble-Based Unsupervised Discontinuous Constituency Parsing by Tree Averaging (2403.00143)
View all the missing papers here:
https://theshayegh.github.io/ACL2024MissingPapers/
v2ray/deepgelbooru
A Danbooru tag image tagger, maybe better than WD14 at some images.
Training code, inference code, dataset included.
:3
https://huggingface.co/v2ray/deepgelbooru
A Danbooru tag image tagger, maybe better than WD14 at some images.
Training code, inference code, dataset included.
:3
https://huggingface.co/v2ray/deepgelbooru
huggingface.co
v2ray/deepgelbooru · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
I can't believe this... Phi-3.5-mini (3.8B) running in-browser at ~90 tokens/second on WebGPU w/ Transformers.js and ONNX Runtime Web! 🤯 Since everything runs 100% locally, no messages are sent to a server — a huge win for privacy!
- 🤗 Demo:
webml-community/phi-3.5-webgpu
- 🧑💻 Source code: https://github.com/huggingface/transformers.js-examples/tree/main/phi-3.5-webgpu
- 🤗 Demo:
webml-community/phi-3.5-webgpu
- 🧑💻 Source code: https://github.com/huggingface/transformers.js-examples/tree/main/phi-3.5-webgpu
GitHub
transformers.js-examples/phi-3.5-webgpu at main · huggingface/transformers.js-examples
A collection of 🤗 Transformers.js demos and example applications - huggingface/transformers.js-examples
You can now use DoRA for your embedding layers!
PR: https://github.com/huggingface/peft/pull/2006
I have documented my journey of this specific PR in a blog post for everyone to read. The highlight of the PR was when the first author of DoRA reviewed my code.
Blog Post: https://huggingface.co/blog/ariG23498/peft-dora
Huge thanks to @BenjaminB for all the help I needed.
PR: https://github.com/huggingface/peft/pull/2006
I have documented my journey of this specific PR in a blog post for everyone to read. The highlight of the PR was when the first author of DoRA reviewed my code.
Blog Post: https://huggingface.co/blog/ariG23498/peft-dora
Huge thanks to @BenjaminB for all the help I needed.
🌐 Check out the new dataset sourced from Fishki.net, one of the popular entertainment and news portals in the Russian Internet, known for its diverse content including humor, interesting facts, and viral stories -
nyuuzyou/fishkinet-posts
.
📊 Dataset highlights:
- 369,180 posts
- Includes original posts with titles, content, images, and metadata
- Each entry contains URL, title, author, date, tags, content, and image URLs
- Primarily in Russian language
- Covers a wide range of topics in entertainment, news, and social media content
- Spans nearly two decades of posts, likely from early 2000s to 2024
- Dedicated to public domain under Creative Commons Zero (CC0) license
nyuuzyou/fishkinet-posts
.
📊 Dataset highlights:
- 369,180 posts
- Includes original posts with titles, content, images, and metadata
- Each entry contains URL, title, author, date, tags, content, and image URLs
- Primarily in Russian language
- Covers a wide range of topics in entertainment, news, and social media content
- Spans nearly two decades of posts, likely from early 2000s to 2024
- Dedicated to public domain under Creative Commons Zero (CC0) license
Alan Turing's mind-bender: "Can machines think?" in its glorified form. This 74yr old paper laid the foundation for how we think about AI and machine intelligence today. The level of detail, clarity and foresight is just phenomenal - he was way ahead of his time 🧠🤖
Original copy: https://archive.org/details/MIND--COMPUTING-MACHINERY-AND-INTELLIGENCE
Original copy: https://archive.org/details/MIND--COMPUTING-MACHINERY-AND-INTELLIGENCE
g/ - /ldg/ - Local Diffusion General - Technology
>Beginner UI
EasyDiffusion: https://easydiffusion.github.io
Fooocus: https://github.com/lllyasviel/fooocus
Metastable: https://metastable.studio
>Advanced UI
Automatic1111: https://github.com/automatic1111/stable-diffusion-webui
ComfyUI: https://github.com/comfyanonymous/ComfyUI
InvokeAI: https://github.com/invoke-ai/InvokeAI
SD.Next: https://github.com/vladmandic/automatic
SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI
>Beginner UI
EasyDiffusion: https://easydiffusion.github.io
Fooocus: https://github.com/lllyasviel/fooocus
Metastable: https://metastable.studio
>Advanced UI
Automatic1111: https://github.com/automatic1111/stable-diffusion-webui
ComfyUI: https://github.com/comfyanonymous/ComfyUI
InvokeAI: https://github.com/invoke-ai/InvokeAI
SD.Next: https://github.com/vladmandic/automatic
SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI
GitHub
GitHub - lllyasviel/Fooocus: Focus on prompting and generating
Focus on prompting and generating. Contribute to lllyasviel/Fooocus development by creating an account on GitHub.
Introducing HelpingAI2-9B, an emotionally intelligent LLM.
Model Link :
OEvortex/HelpingAI2-9B
Demo Link:
Abhaykoul/HelpingAI2
This model is part of the innovative HelpingAI series and it stands out for its ability to engage users with emotional understanding.
Key Features:
-----------------
* It gets 95.89 score on EQ Bench greather than all top notch LLMs, reflecting advanced emotional recognition.
* It gives responses in empathetic and supportive manner.
Must try our demo:
Abhaykoul/HelpingAI2
Model Link :
OEvortex/HelpingAI2-9B
Demo Link:
Abhaykoul/HelpingAI2
This model is part of the innovative HelpingAI series and it stands out for its ability to engage users with emotional understanding.
Key Features:
-----------------
* It gets 95.89 score on EQ Bench greather than all top notch LLMs, reflecting advanced emotional recognition.
* It gives responses in empathetic and supportive manner.
Must try our demo:
Abhaykoul/HelpingAI2
NEW math-instruct model + dataset!
ValiantLabs/Llama3.1-8B-Cobalt
is our new math-instruct model.
Trained using a synthetic math-instruct dataset generated with Llama 3.1 405b. Find the dataset here:
sequelbox/Polytope
More to come soon :)
ValiantLabs/Llama3.1-8B-Cobalt
is our new math-instruct model.
Trained using a synthetic math-instruct dataset generated with Llama 3.1 405b. Find the dataset here:
sequelbox/Polytope
More to come soon :)
Supercool Weekend Read🤖
Nvidia researchers achieved SOTA LLM compression metrics using pruning and knowledge distillation techniques.
Details on Techniques (Simplified):
They started off with a large pre-trained language model (15B params), then:
1. Estimated the importance of different parts of the model (neurons, attention heads, layers) using activation-based metrics on a small calibration dataset.
Nvidia researchers achieved SOTA LLM compression metrics using pruning and knowledge distillation techniques.
Details on Techniques (Simplified):
They started off with a large pre-trained language model (15B params), then:
1. Estimated the importance of different parts of the model (neurons, attention heads, layers) using activation-based metrics on a small calibration dataset.
What Happens When RAG System Become Fully Vision-Language Model-Based?
HF Demo:
bokesyo/MiniCPMV-RAG-PDFQA
Multimodal Dense Retriever:
RhapsodyAI/minicpm-visual-embedding-v0
Generation Model:
openbmb/MiniCPM-V-2_6
Github: https://github.com/RhapsodyAILab/MiniCPM-V-Embedding-v0-Train
The Vision-Language Model Dense Retriever MiniCPM-Visual-Embedding-v0 reads PDFs directly -- no OCR required. With strong OCR capability and visual understanding capability, it generates multimodal dense representations, allowing you to build and search through your personal library with ease.
Ask a question, it retrieves the most relevant pages. Then, MiniCPM-V-2.6 provides answers based on the retrieved pages, with strong multi-image understanding capabilities.
Whether you’re working with a visually-intensive or text-oriented PDF, it helps you quickly find the information you need. You can also build a personal library with it.
It operates just like a human: reading, storing, retrieving, and answering with full visual comprehension.
Currently, the online demo supports PDFs with up to 50 pages due to GPU time limits. For longer PDFs or entire books, you can deploy it on your own machine.
https://cdn-uploads.huggingface.co/production/uploads/6415818a986557e8cac252bf/sjtQD7CFgox46h9EVHCG_.png
HF Demo:
bokesyo/MiniCPMV-RAG-PDFQA
Multimodal Dense Retriever:
RhapsodyAI/minicpm-visual-embedding-v0
Generation Model:
openbmb/MiniCPM-V-2_6
Github: https://github.com/RhapsodyAILab/MiniCPM-V-Embedding-v0-Train
The Vision-Language Model Dense Retriever MiniCPM-Visual-Embedding-v0 reads PDFs directly -- no OCR required. With strong OCR capability and visual understanding capability, it generates multimodal dense representations, allowing you to build and search through your personal library with ease.
Ask a question, it retrieves the most relevant pages. Then, MiniCPM-V-2.6 provides answers based on the retrieved pages, with strong multi-image understanding capabilities.
Whether you’re working with a visually-intensive or text-oriented PDF, it helps you quickly find the information you need. You can also build a personal library with it.
It operates just like a human: reading, storing, retrieving, and answering with full visual comprehension.
Currently, the online demo supports PDFs with up to 50 pages due to GPU time limits. For longer PDFs or entire books, you can deploy it on your own machine.
https://cdn-uploads.huggingface.co/production/uploads/6415818a986557e8cac252bf/sjtQD7CFgox46h9EVHCG_.png
GitHub
GitHub - RhapsodyAILab/MiniCPM-V-Embedding
Contribute to RhapsodyAILab/MiniCPM-V-Embedding development by creating an account on GitHub.
So turns out I've been spreading a bit of misinformation when it comes to imatrix in llama.cpp
It starts true; imatrix runs the model against a corpus of text and tracks the activation of weights to determine which are most important
However what the quantization then does with that information is where I was wrong.
I think I made the accidental connection between imatrix and exllamav2's measuring, where ExLlamaV2 decides how many bits to assign to which weight depending on the goal BPW
Instead, what llama.cpp with imatrix does is it attempts to select a scale for a quantization block that most accurately returns the important weights to their original values, ie minimizing the dequantization error based on the importance of activations
The mildly surprising part is that it actually just does a relatively brute force search, it picks a bunch of scales and tries each and sees which one results in the minimum error for weights deemed important in the group
But yeah, turns out, the quantization scheme is always the same, it's just that the scaling has a bit more logic to it when you use imatrix
Huge shoutout to @compilade for helping me wrap my head around it - feel free to add/correct as well if I've messed something up
It starts true; imatrix runs the model against a corpus of text and tracks the activation of weights to determine which are most important
However what the quantization then does with that information is where I was wrong.
I think I made the accidental connection between imatrix and exllamav2's measuring, where ExLlamaV2 decides how many bits to assign to which weight depending on the goal BPW
Instead, what llama.cpp with imatrix does is it attempts to select a scale for a quantization block that most accurately returns the important weights to their original values, ie minimizing the dequantization error based on the importance of activations
The mildly surprising part is that it actually just does a relatively brute force search, it picks a bunch of scales and tries each and sees which one results in the minimum error for weights deemed important in the group
But yeah, turns out, the quantization scheme is always the same, it's just that the scaling has a bit more logic to it when you use imatrix
Huge shoutout to @compilade for helping me wrap my head around it - feel free to add/correct as well if I've messed something up
How good are you at spotting AI-generated images?
Find out by playing Fake Insects 🐞 a Game where you need to identify which insects are fake (AI generated). Good luck & share your best score in the comments!
victor/fake-insects
Find out by playing Fake Insects 🐞 a Game where you need to identify which insects are fake (AI generated). Good luck & share your best score in the comments!
victor/fake-insects
I'm excited to share a really cool milestone in my AI/LLM journey.
Brief backstory: Before diving into AI, I spent over a decade working in ecological fields such as the conservation corps, biodynamic farming, and natural habitat restoration. This background instilled in me a deep concern about the environmental impact of scaling AI without sustainable practices.
Driven by this concern, I've spent months planning and experimenting to make my AI work more eco-friendly. I'm thrilled to announce that I've successfully transitioned my entire operation to run on 100% sustainable solar power!
Brief backstory: Before diving into AI, I spent over a decade working in ecological fields such as the conservation corps, biodynamic farming, and natural habitat restoration. This background instilled in me a deep concern about the environmental impact of scaling AI without sustainable practices.
Driven by this concern, I've spent months planning and experimenting to make my AI work more eco-friendly. I'm thrilled to announce that I've successfully transitioned my entire operation to run on 100% sustainable solar power!
🚀 We’re excited to launch Ghost 8B Beta (1608), a top-performing language model with unmatched multilingual support and cost efficiency.
Key Highlights:
- Superior Performance: Outperforms Llama 3.1 8B Instruct, GPT-3.5 Turbo, Claude 3 Opus, GPT-4, and more in winrate scores.
- Expanded Language Support: Now supports 16 languages, including English, Vietnamese, Spanish, Chinese, and more.
- Enhanced Capabilities: Improved math, reasoning, and instruction-following for better task handling.
Key Highlights:
- Superior Performance: Outperforms Llama 3.1 8B Instruct, GPT-3.5 Turbo, Claude 3 Opus, GPT-4, and more in winrate scores.
- Expanded Language Support: Now supports 16 languages, including English, Vietnamese, Spanish, Chinese, and more.
- Enhanced Capabilities: Improved math, reasoning, and instruction-following for better task handling.