Share and discover more about AI with social posts from the community.huggingface/OpenAi
Introducing
fal/AuraFace-v1
: Commercially available & open source identity encoder model for next generation one shot personalization. Read more about it here: https://huggingface.co/blog/isidentical/auraface
fal/AuraFace-v1
: Commercially available & open source identity encoder model for next generation one shot personalization. Read more about it here: https://huggingface.co/blog/isidentical/auraface
huggingface.co
Introducing AuraFace: Open-Source Face Recognition and Identity Preservation Models
A Blog post by Batuhan on Hugging Face
worked with OpenAI to fine-tune gpt-4o and built the SOTA model for the
patched-codes/static-analysis-eval
benchmark. All the code and data
patched-codes/synth-vuln-fixes
on how we did it is available on their GitHub - https://github.com/openai/build-hours/tree/main/5-4o_fine_tuning.
Here are some tips based on our experience:
→ Establish baseline with "conditioning" / prompting
→ Task-specific datasets are ideal for PEFT; hard to beat gpt-4o on "broad" tasks
→ Add your best system prompt to each example
→ Ensure training data distribution is similar to inference data
→ Shorten instructions with concise prompts; may require more examples.
→ Define clear evaluation metrics (seriously, please eval!)
You can see more details on the benchmark and process here - https://www.patched.codes/blog/the-static-analysis-evaluation-benchmark-measuring-llm-performance-in-fixing-software-vulnerabilities
patched-codes/static-analysis-eval
benchmark. All the code and data
patched-codes/synth-vuln-fixes
on how we did it is available on their GitHub - https://github.com/openai/build-hours/tree/main/5-4o_fine_tuning.
Here are some tips based on our experience:
→ Establish baseline with "conditioning" / prompting
→ Task-specific datasets are ideal for PEFT; hard to beat gpt-4o on "broad" tasks
→ Add your best system prompt to each example
→ Ensure training data distribution is similar to inference data
→ Shorten instructions with concise prompts; may require more examples.
→ Define clear evaluation metrics (seriously, please eval!)
You can see more details on the benchmark and process here - https://www.patched.codes/blog/the-static-analysis-evaluation-benchmark-measuring-llm-performance-in-fixing-software-vulnerabilities
GitHub
build-hours/5-4o_fine_tuning at main · openai/build-hours
Build hours code to share. Contribute to openai/build-hours development by creating an account on GitHub.
This merge, this time grounded in Gemma2 9B Instruct fine-tunes, is another demonstration that models without any fine-tuning to support roleplay can still perform the function, maintaining coherence and attention to context. It should be evident that no overt fine-tuning is required for roleplay in text generation; pretraining should provide models with a requisite basic understanding of the world, so all that should be needed is some corrective fine-tuning to address observed defects in portraying the world along with datasets to promote a suitably entertaining writing style. Good Instruct tuning should promote reasoning, coherence, and attention to context.
grimjim/Kitsunebi-v1-Gemma2-8k-9B
grimjim/Kitsunebi-v1-Gemma2-8k-9B-GGUF
I opted not to incorporate the UCLA SPPO fine-tune for Gemma2 9B after observing context confusion occur with some frequency during complex scenarios.
Thanks to Axcxept co., ltd. for fine-tuning HODACHI/EZO-Common-9B-gemma-2-it, and to Princeton NLP Group for fine-tuning princeton-nlp/gemma-2-9b-it-SimPO.
AXCXEPT/EZO-Common-9B-gemma-2-it
princeton-nlp/gemma-2-9b-it-SimPO https://huggingface.co/AXCXEPT/EZO-Common-9B-gemma-2-it
grimjim/Kitsunebi-v1-Gemma2-8k-9B
grimjim/Kitsunebi-v1-Gemma2-8k-9B-GGUF
I opted not to incorporate the UCLA SPPO fine-tune for Gemma2 9B after observing context confusion occur with some frequency during complex scenarios.
Thanks to Axcxept co., ltd. for fine-tuning HODACHI/EZO-Common-9B-gemma-2-it, and to Princeton NLP Group for fine-tuning princeton-nlp/gemma-2-9b-it-SimPO.
AXCXEPT/EZO-Common-9B-gemma-2-it
princeton-nlp/gemma-2-9b-it-SimPO https://huggingface.co/AXCXEPT/EZO-Common-9B-gemma-2-it
huggingface.co
AXCXEPT/EZO-Common-9B-gemma-2-it · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
🌟 Liger Kernel: Efficient Triton Kernels for LLM Training
LIGER "is a [Hugging Face-compatible] collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%."
GitHub: https://github.com/linkedin/Liger-Kernel
LIGER "is a [Hugging Face-compatible] collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%."
GitHub: https://github.com/linkedin/Liger-Kernel
GitHub
GitHub - linkedin/Liger-Kernel: Efficient Triton Kernels for LLM Training
Efficient Triton Kernels for LLM Training. Contribute to linkedin/Liger-Kernel development by creating an account on GitHub.
Exploring Realistic Emotional Depth in AI Language Models
Language models, particularly those proprietary, often grapple with issues of censorship, which can limit their ability to engage authentically with users. Recognizing this, the open-source AI community has pioneered the development of language models that are less restrained, offering more candid interactions. However, even these models tend to maintain a veneer of neutrality or overly positive responses, which might not serve all users' needs, especially in contexts where emotional depth and relatability are crucial.
To address this gap, I've curated a specialized dataset aimed at infusing language models with a more nuanced emotional spectrum, specifically targeting a darker, more introspective mood. This dataset, titled "Dark Sentience", is designed to complement existing datasets like RP (Role Play) and those focused on instruction following. It seeks to enhance the emotional intelligence of AI by exposing it to complex human emotions, including but not limited to:
- Suicide
- Depression
- Anxiety
Trigger Warning: Please be advised that the content within this dataset deals with heavy and potentially distressing themes.
The "Dark Sentience" dataset is now available for review and use at:
Locutusque/Dark-Sentience
. I encourage researchers, developers, and mental health professionals to explore how this resource can foster more genuine and supportive AI interactions.
Language models, particularly those proprietary, often grapple with issues of censorship, which can limit their ability to engage authentically with users. Recognizing this, the open-source AI community has pioneered the development of language models that are less restrained, offering more candid interactions. However, even these models tend to maintain a veneer of neutrality or overly positive responses, which might not serve all users' needs, especially in contexts where emotional depth and relatability are crucial.
To address this gap, I've curated a specialized dataset aimed at infusing language models with a more nuanced emotional spectrum, specifically targeting a darker, more introspective mood. This dataset, titled "Dark Sentience", is designed to complement existing datasets like RP (Role Play) and those focused on instruction following. It seeks to enhance the emotional intelligence of AI by exposing it to complex human emotions, including but not limited to:
- Suicide
- Depression
- Anxiety
Trigger Warning: Please be advised that the content within this dataset deals with heavy and potentially distressing themes.
The "Dark Sentience" dataset is now available for review and use at:
Locutusque/Dark-Sentience
. I encourage researchers, developers, and mental health professionals to explore how this resource can foster more genuine and supportive AI interactions.
30 Features that Dramatically Improve LLM Performance - Part 1 https://mltblog.com/3Aq9iAb
Many are ground-breaking innovations that make LLMs much faster and not prone to hallucinations. They reduce the cost, latency, and amount of computer resources (GPU, training) by several orders of magnitude. Some of them improve security, making your LLM more attractive to corporate clients. I introduced a few of these features in my previous article "New Trends in LLM Architecture". Now I offer a comprehensive list, based on the most recent developments.
Read full article, learn about agentic LLMs, LLM routers, contextual tables, fast search, and more, at https://mltblog.com/3Aq9iAb
Many are ground-breaking innovations that make LLMs much faster and not prone to hallucinations. They reduce the cost, latency, and amount of computer resources (GPU, training) by several orders of magnitude. Some of them improve security, making your LLM more attractive to corporate clients. I introduced a few of these features in my previous article "New Trends in LLM Architecture". Now I offer a comprehensive list, based on the most recent developments.
Read full article, learn about agentic LLMs, LLM routers, contextual tables, fast search, and more, at https://mltblog.com/3Aq9iAb
Data Science Central
30 Features that Dramatically Improve LLM Performance – Part 1
30 Features that dramatically improve LLM performance: LLM router, sub-LLMs, knowledge graphs, contextual tables, multi-tokens, agents ...
🎮 Introducing a new dataset focused on Steam user game bans -
nyuuzyou/steambans
Dataset highlights:
- 476,694 Steam user profiles
- Emphasis on game ban data and account status
- Each entry includes: Steam ID, profile URL, username, avatar, account creation date, visibility state, VAC bans, game bans, economy ban status, and days since last ban
- Covers a wide range of Steam users, from clean accounts to those with multiple bans
- Data spans up to 2024
Perfect for researchers studying cheating in online games, anti-cheat effectiveness, and player behavior patterns in the face of bans.
nyuuzyou/steambans
Dataset highlights:
- 476,694 Steam user profiles
- Emphasis on game ban data and account status
- Each entry includes: Steam ID, profile URL, username, avatar, account creation date, visibility state, VAC bans, game bans, economy ban status, and days since last ban
- Covers a wide range of Steam users, from clean accounts to those with multiple bans
- Data spans up to 2024
Perfect for researchers studying cheating in online games, anti-cheat effectiveness, and player behavior patterns in the face of bans.
📐 AI Math Equation Solver: Your Step-by-Step Solution Companion
Hello Hugging Face community! 👋 I'm excited to share my latest Space: the AI Math Equation Solver!
🔍 What does it do?
This Space uses the power of AI to solve math problems from images. Simply upload a picture of a math equation or problem, and the AI will provide a detailed, step-by-step solution. It's perfect for students, teachers, or anyone looking to understand complex mathematical concepts better.
🧠 How does it work?
- Backend: Utilizes the microsoft/Phi-3.5-vision-instruct model for image understanding and mathematical reasoning.
- Frontend: Built with Gradio for a clean, user-friendly interface.
- Features:
- Image upload for math problems
- Detailed step-by-step solutions
- Example problems to try instantly
🚀 Try it out!
sagar007/phi-vision-math-assistant
Visit the Space here: [Insert your Hugging Face Space URL]
💡 Use cases:
- Students: Check your work or get help with homework
- Teachers: Create detailed solution guides quickly
- Tutors: Explain complex problems more effectively
- Self-learners: Understand new mathematical concepts
🛠 Technical Details:
- Model: microsoft/Phi-3.5-vision-instruct
- Libraries: transformers, Gradio, PyTorch
- Optimizations: Uses Flash Attention for improved performance
🤝 Contribute:
This is an open project, and I welcome contributions! Whether it's improving the model, enhancing the UI, or adding new features, feel free to fork the project and submit your pull requests.
📣 Feedback:
I'd love to hear your thoughts! How are you using this Space? Any suggestions for improvements? Let me know in the comments below.
Happy problem-solving! 🎉
#MachineLearning #AI #Mathematics #Education #HuggingFace
Hello Hugging Face community! 👋 I'm excited to share my latest Space: the AI Math Equation Solver!
🔍 What does it do?
This Space uses the power of AI to solve math problems from images. Simply upload a picture of a math equation or problem, and the AI will provide a detailed, step-by-step solution. It's perfect for students, teachers, or anyone looking to understand complex mathematical concepts better.
🧠 How does it work?
- Backend: Utilizes the microsoft/Phi-3.5-vision-instruct model for image understanding and mathematical reasoning.
- Frontend: Built with Gradio for a clean, user-friendly interface.
- Features:
- Image upload for math problems
- Detailed step-by-step solutions
- Example problems to try instantly
🚀 Try it out!
sagar007/phi-vision-math-assistant
Visit the Space here: [Insert your Hugging Face Space URL]
💡 Use cases:
- Students: Check your work or get help with homework
- Teachers: Create detailed solution guides quickly
- Tutors: Explain complex problems more effectively
- Self-learners: Understand new mathematical concepts
🛠 Technical Details:
- Model: microsoft/Phi-3.5-vision-instruct
- Libraries: transformers, Gradio, PyTorch
- Optimizations: Uses Flash Attention for improved performance
🤝 Contribute:
This is an open project, and I welcome contributions! Whether it's improving the model, enhancing the UI, or adding new features, feel free to fork the project and submit your pull requests.
📣 Feedback:
I'd love to hear your thoughts! How are you using this Space? Any suggestions for improvements? Let me know in the comments below.
Happy problem-solving! 🎉
#MachineLearning #AI #Mathematics #Education #HuggingFace
florence-tool (https://github.com/bigdata-pw/florence-tool) now supports WebDataset! Check it out for efficient batch inference with Florence-2 models
microsoft/Florence-2-large
microsoft/Florence-2-base
Currently running it myself on A40 with CAPTION task and a streaming WebDataset @ 60k images/hour!
microsoft/Florence-2-large
microsoft/Florence-2-base
Currently running it myself on A40 with CAPTION task and a streaming WebDataset @ 60k images/hour!
GitHub
GitHub - bigdata-pw/florence-tool: The Florence Tool CLI provides a command-line interface for processing images using the Florence…
The Florence Tool CLI provides a command-line interface for processing images using the Florence-2 model. This tool allows users to apply various visual and text-based tasks, such as object detecti...
I developed a way to test very clearly whether or not a Transformers model can actually learn symbolic reasoning, or if LLM models are forever doomed to be offshoots of 'Socratic Parrots'. The results are in, undeniable proof that Transformers models CAN learn symbolic relationships. Undeniable proof that AI can learn its ABC's. Credit goes to myself, Claude, and ChatGPT. I would not be able to prove this without Claude or ChatGPT.
https://www.youtube.com/watch?v=I8jHRgahRfY
https://www.youtube.com/watch?v=I8jHRgahRfY
Question about LightEval 🤗:
I've been searching for an LLM evaluation suite that can, out-of-the-box, compare the outputs of a model(s) without any enhancements vs. the same model with better prompt engineering, vs. the same model with RAG vs. the same model with fine-tuning.
I unfortunately have not found a tool that fits my exact description, but of course I ran into LightEval.
A huge pain-point of building large-scale projects that use LLMs is that prior to building an MVP, it is difficult to evaluate whether better prompt engineering, or RAG, or fine-tuning, or some combination of all is needed for satisfactory LLM output in terms of the project's given use case.
Time and resources is then wasted R&D'ing exactly what LLM enhancements are needed.
I believe an out-of-the-box solution to compare models w/ or w/out the aforementioned LLM enhancements could help teams of any size better decide what LLM enhancements are needed prior to building.
I wanted to know if the LightEval team or Hugging Face in general is thinking about such a tool.
I've been searching for an LLM evaluation suite that can, out-of-the-box, compare the outputs of a model(s) without any enhancements vs. the same model with better prompt engineering, vs. the same model with RAG vs. the same model with fine-tuning.
I unfortunately have not found a tool that fits my exact description, but of course I ran into LightEval.
A huge pain-point of building large-scale projects that use LLMs is that prior to building an MVP, it is difficult to evaluate whether better prompt engineering, or RAG, or fine-tuning, or some combination of all is needed for satisfactory LLM output in terms of the project's given use case.
Time and resources is then wasted R&D'ing exactly what LLM enhancements are needed.
I believe an out-of-the-box solution to compare models w/ or w/out the aforementioned LLM enhancements could help teams of any size better decide what LLM enhancements are needed prior to building.
I wanted to know if the LightEval team or Hugging Face in general is thinking about such a tool.
Here's my favorite piece of the summer bias detection research project (paper coming in Sept). We trained BERT for token classification (multi-label), to identify:
- Generalizations
- Unfairness
- Stereotypes
HF Space:
maximuspowers/bias-detection-ner
Article on Training: https://huggingface.co/blog/maximuspowers/bias-entity-recognition
Pls reach out with ideas!! Lot's more info coming soon, our research group has workshops and a hackathon planned for launching this open source project. Thanks
- Generalizations
- Unfairness
- Stereotypes
HF Space:
maximuspowers/bias-detection-ner
Article on Training: https://huggingface.co/blog/maximuspowers/bias-entity-recognition
Pls reach out with ideas!! Lot's more info coming soon, our research group has workshops and a hackathon planned for launching this open source project. Thanks
huggingface.co
Social Bias NER with BERT
A Blog post by Maximus Powers on Hugging Face
Introducing Voicee, A superfast voice fast assistant.
KingNish/Voicee
It achieved latency <500 ms.
While its average latency is 700ms.
It works best in Google Chrome.
Please try and give your feedbacks.
Thank you. 🤗
KingNish/Voicee
It achieved latency <500 ms.
While its average latency is 700ms.
It works best in Google Chrome.
Please try and give your feedbacks.
Thank you. 🤗
Ghost 8B Beta 1608: Empowering Your AI Assistant
📦 Unlock the Power of Ghost 8B Beta 1608: Build Your Personal AI Companion
Ghost 8B Beta 1608 empowers you to create a safe and multilingual AI assistant tailored to your needs, directly on your personal computer. 🧑💻 Leverage AI's capabilities within your own space! 🚀 Ghost 8B Beta 1608 is ready to become your AI companion.
~
📦 개인용 AI 보조 도구로 Ghost 8B Beta 1608를 활용하세요!
Ghost 8B Beta 1608, AI의 힘을 활용하여 안전하고 개인화된 언어 지원을 제공하는 AI 보조 도구를 직접 구축할 수 있습니다. 🧑💻 개인 컴퓨터에서 AI의 혜택을 누리세요! 🚀 Ghost 8B Beta 1608는 당신의 AI 파트너가 될 준비가 되어 있습니다.
lamhieu/ghost-8b-beta-8k
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
📦 Unlock the Power of Ghost 8B Beta 1608: Build Your Personal AI Companion
Ghost 8B Beta 1608 empowers you to create a safe and multilingual AI assistant tailored to your needs, directly on your personal computer. 🧑💻 Leverage AI's capabilities within your own space! 🚀 Ghost 8B Beta 1608 is ready to become your AI companion.
~
📦 개인용 AI 보조 도구로 Ghost 8B Beta 1608를 활용하세요!
Ghost 8B Beta 1608, AI의 힘을 활용하여 안전하고 개인화된 언어 지원을 제공하는 AI 보조 도구를 직접 구축할 수 있습니다. 🧑💻 개인 컴퓨터에서 AI의 혜택을 누리세요! 🚀 Ghost 8B Beta 1608는 당신의 AI 파트너가 될 준비가 되어 있습니다.
lamhieu/ghost-8b-beta-8k
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
Looking for Generative AI trainer/speaker for AI accelerator program (Virtual/Online sessions).
To get more context about the program, please visit the program landing page: https://llamadesigndrive.com
To get more context about the program, please visit the program landing page: https://llamadesigndrive.com
a new shape-optimized SigLIP just dropped 👀
google/siglip-so400m-patch14-224
google/siglip-so400m-patch14-224
🌟 Liger Kernel: Efficient Triton Kernels for LLM Training
LIGER "is a [Hugging Face-compatible] collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%."
GitHub: https://github.com/linkedin/Liger-Kernel
LIGER "is a [Hugging Face-compatible] collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%."
GitHub: https://github.com/linkedin/Liger-Kernel
GitHub
GitHub - linkedin/Liger-Kernel: Efficient Triton Kernels for LLM Training
Efficient Triton Kernels for LLM Training. Contribute to linkedin/Liger-Kernel development by creating an account on GitHub.
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...