Hugging Face

Due to a security incident, we strongly suggest you rotate any tokens or keys you use in secrets for HF Spaces: https://lnkd.in/eEpQQVBB. We have already proactively revoked a number of HF tokens and are working with outside cybersecurity forensic specialists to investigate the issue as well as review our security policies and procedures. You can find more initial information at https://lnkd.in/et4vWVPg.

We are soon releasing a feature that will enable you to preview images in your VLM Chatbots! 😍 😉 In case you'd like to try it now, you can use the Gradio wheel mentioned in this PR: https://lnkd.in/gh9XNJRS 🙌 Thanks, Dawood Khan, for the amazing feature!

FiftyOne Datasets 🤝 🤗 Hub Over the past year and a half working in the open-source machine learning community, I've repeatedly seen and heard the following pain points: ❌ It's hard to find and download high-quality datasets for specific tasks ❌ It's hard to apply state-of-the-art models to my data ❌ It's hard to share my data with others These frustrations prevent cutting-edge ML work in academia and industry, especially on unstructured visual data! Our team at Voxel51 has been listening, and over the past few months, we've been working closely with Hugging Face to help you streamline these workflows!! We've connected FiftyOne visual datasets with Hugging Face's Hub and Transformers libraries, so now you can do any/all of the following with a single line of code: ✅ Load Image, Video, and 3D datasets in Parquet or FiftyOne format from the 🤗 Hub directly into FiftyOne ✅ Apply 🤗 Transformer models like CLIP, Grounding DINO, and Depth Anything to your entire dataset ✅ Share your visual dataset with others by pushing to the 🤗 Hub I just published an article on Hugging Face highlighting this integration. Check it out! Major shoutout to everyone from the Hugging Face team who was integral in making this happen: Daniel van Strien Sylvain Lesage Quentin Lhoest Omar Sanseviero Julien Chaumond Merve Noyan #opensource #ml #computervision #transformers #mlops #ai #huggingface #fiftyone 🔗 https://lnkd.in/gaWHXymt

Do you need a dataset to train a custom sentence transformer model? There are many use cases where a more customized sentence transformer model could be valuable, but creating the training data for fine-tuning a sentence transformer model can be expensive and time-consuming. I've created a pipeline using Argilla Distilabel, which allows you to use an LLM to create a synthetic dataset you can directly use to fine-tune/train a Setence Transformers model. The steps are roughly: - Starting from data for the domain/task you are interested in - Use an LLM to generate a set of candidate positive/negative pairs based on the input data (the anchor) - Mine the hard examples from the negative pair Using Outlines for structured generation gives you extra control over generating negative/positive pairs. - The pipeline is covered in a tutorial in the Awesome Synthetic datasets repository (https://lnkd.in/e5rJBeCR). - You can find an example that uses a BigCode dataset as a starting point for creating a new dataset/model for detecting the similarity of code prompts in this Hugging Face Collection: https://lnkd.in/ehVMRYjJ

FiftyOne Datasets 🤝 🤗 Hub Over the past year and a half working in the open-source machine learning community, I've repeatedly seen and heard the following pain points: ❌ It's hard to find and download high-quality datasets for specific tasks ❌ It's hard to apply state-of-the-art models to my data ❌ It's hard to share my data with others These frustrations prevent cutting-edge ML work in academia and industry, especially on unstructured visual data! Our team at Voxel51 has been listening, and over the past few months, we've been working closely with Hugging Face to help you streamline these workflows!! We've connected FiftyOne visual datasets with Hugging Face's Hub and Transformers libraries, so now you can do any/all of the following with a single line of code: ✅ Load Image, Video, and 3D datasets in Parquet or FiftyOne format from the 🤗 Hub directly into FiftyOne ✅ Apply 🤗 Transformer models like CLIP, Grounding DINO, and Depth Anything to your entire dataset ✅ Share your visual dataset with others by pushing to the 🤗 Hub I just published an article on Hugging Face highlighting this integration. Check it out! Major shoutout to everyone from the Hugging Face team who was integral in making this happen: Daniel van Strien Sylvain Lesage Quentin Lhoest Omar Sanseviero Julien Chaumond Merve Noyan #opensource #ml #computervision #transformers #mlops #ai #huggingface #fiftyone 🔗 https://lnkd.in/gaWHXymt

Following up on our efforts to expand the presence of DPO-style datasets for Arabic, we are excited to release 4 new DPO datasets extracted from NoRobots, a dataset we published earlier this year. We encourage you all to experiment with each of these datasets and report your results, we would gladly share your findings with the world. We are eager to see how these datasets can advance our understanding and modeling of our beloved language. For easy access, here are the links to the DPO datasets collections we have so far : NoRobots-DPO collection: https://lnkd.in/eCx4hrhF Previous Aya-DPO collection: https://lnkd.in/eP4Qij4D Share your findings using these datasets and stay tuned for more exciting updates! On a final note, we are immensely grateful to Hugging Face for the compute. Without their unconditional support, this work wouldn’t have seen the light of day. We are also deeply grateful to Daniel van Strien for his continuous support 🤗

Uploaded a new YouTube video! In this tutorial, I explain how one could fine-tune PaliGemma, an open vision-language model from Google on a custom dataset. This can be done at home since I'm using an L4 GPU available on Google Colab. I explain various things like how the model actually works, LoRa (low-rank adaption), quantization and more! Also feel free to leave some feedback :) https://lnkd.in/eWQ8HCAN #youtube #artificialintelligence #huggingface

AutoTrain now supports finetuning embedding models using Sentence Transformers! With other words: embedding models for your data without having to write any code. Details: ☁️ AutoTrain runs locally, in the cloud, in Hugging Face Spaces, or in Google Colab 💾 Provide JSON/CSV data, or a Hugging Face Dataset 🔔 Supports numerous dataset formats: (question, answer) pairs, (anchor, positive) pairs, pair + class, pair + similarity score, (anchor, positive, negative) triplets. Learn about it here: https://lnkd.in/eJ8GkAkj Great job Abhishek Thakur for shipping this so quickly.

🚨 NEW BLOG: How to Fine-Tune Custom Embedding Models Using AutoTrain Learn: - what should be the data format - how to map columns properly - example datasets - custom configs - train locally - train on hugging face spaces https://lnkd.in/dG4NnutY

🚀 Scribble SDXL ControlNet with Gradio ImageEditor component works like magic! Check out the model and cool Spaces👇 Video and Demo credit: Linoy Tsaban Gradio Demo link: https://t.co/3r4s3zfUWN Model on 🤗Hugging Face: https://lnkd.in/gGfih7s5