Hugging Face

Excited to share our latest piece on the double-edged sword of AI agents published in MIT Technology Review! 🤖 It builds on our research paper that's been making waves lately -- pretty cool to see all the attention it's getting! As these systems move beyond chat windows to navigate applications and execute complex tasks independently, we need to ask ourselves: how much control are we willing to surrender, and at what cost? In our recent op-ed, Margaret Mitchell, Avijit Ghosh, PhD, Dr. Sasha Luccioni, and I explore why the very feature being sold (reduced human oversight) is actually the primary vulnerability. When AI systems can control multiple information sources simultaneously, the potential for harm explodes exponentially. We imagine that "It wasn't me—it was my agent!!" will soon be a common refrain to excuse bad outcomes. The benefits of AI agents are undeniable, from assisting people with mobility challenges to coordinating emergency responses. But these benefits don't require surrendering complete human control. At Hugging Face, we're developing frameworks like smolagents that prioritize transparency and appropriate human oversight. Because human judgment, with all its imperfections, remains the fundamental component in ensuring these systems serve rather than subvert our interests.

Improving small model abilities with DeepSeek AI R1 few shots. A new experiment powered by Hugging Face Inference Providers and Data Studio! Last week, I shared the LIMO (Less is more for reasoning) dataset categorized with Llama 70B. I aimed to understand the topic distribution of this small but powerful dataset. This time, I wanted to experiment with a new idea: what if we could use few-shot examples (demonstrations) from R1 to improve Llama's ability to extract topics? Here's what I did: 1. Run R1 over a few rows. 2. Inspect and validate the results. 3. Use 5 validated R1 responses as part of the prompt to Llama. 4. Categorize the entire dataset. 5. Use Data Studio to extract the distribution of topics and compare it to my previous dataset. It all took under 30 minutes. It's amazing how quickly you can turn ideas into results using the Hub! As a bonus, you can share your results with the world. The open dataset, the prompts, and the pipeline config are in the first comment!

KDTalker is wild 🤯 Runs on one 4090 or 3090 🤩 Generates Talking Portraits which are audio-driven - Preserves character identity with fine facial details - Expressions change with the speech, giving high pose diversity - Diffusion based Play with their official gradio website: links in the comments! Coming very soon to the Hugging Face Spaces 🤗

So many open releases at Hugging Face past week 🤯 recapping all here ⤵️ link to collection in comments 👀 Multimodal > Mistral AI released a 24B vision LM, both base and instruction FT versions, sota 🔥 (OS) > with IBM we released SmolDocling, a sota 256M document parser with Apache 2.0 license (OS) > SpatialLM is a new vision LM that outputs 3D bounding boxes, comes with 0.5B (QwenVL based) and 1B (Llama based) variants > SkyWork released SkyWork-R1V-38B, new vision reasoning model (OS) 💬 LLMs > NVIDIA released new Nemotron models in 49B and 8B with their post-training dataset > LG released EXAONE, new reasoning models in 2.4B, 7.8B and 32B > Dataset: Glaive AI released a new reasoning dataset of 22M+ examples > Dataset: NVIDIA released new helpfulness dataset HelpSteer3 > Dataset: OpenManusRL is a new agent dataset based on ReAct framework (OS) > Open-R1 team released OlympicCoder, new competitive coder model in 7B and 32B > Dataset: GeneralThought-430K is a new reasoning dataset (OS) 🖼️ Image Generation/Computer Vision > Roboflow released RF-DETR, new real-time sota object detector (OS) 🔥 > YOLOE is a new real-time zero-shot object detector with text and visual prompts 🥹 > Stability AI released Stable Virtual Camera, a new novel view synthesis model > Tencent released Hunyuan3D-2mini, new small and fast 3D asset generation model > ByteDance released InfiniteYou, new realistic photo generation model > StarVector is a new 8B model that generates svg from images > FlexWorld is a new model that expands 3D views (OS) 🎤 Audio > Sesame released CSM-1B new speech generation model (OS) 🤖 Robotics > NVIDIA released GR00T, new robotics model for generalized reasoning and skills, along with the dataset *OS ones have Apache 2.0 or MIT license

Gradio Sketch in 2 Minutes 🔥

We introduce FAT5 (Flash Attention T5) ⚡ An implementation of T5 in PyTorch with UL2 objective optimized for GPGPU for both training and inference thanks to 13 differents optimizations. The main one is that we have designed a CUDA kernel to expand the Flash Attention by Tri Dao with RPE biases and supports other positional encodings such as RoPE, ALiBi or FIRE. The result kernel is 2 times faster than a SPDA implementation. We also use Triton kernels to optimize certain parts of the architecture, such as the cross-entropy and RMSNorm layer.   The various kernels have been carefully built to be compatible with BF16 and torch.compile to go even faster and achieve efficient pretraining. All other optimizations are described in a 📝 subsequent blog post available on Hugging Face 🤗: https://lnkd.in/eTTUqt7R This methodology enabled us to efficiently pretrain as a proof of concept a FAT5 with 147M parameters in French in a reasonable time (1,461H for 419B tokens), with limited resources (1 A100 i.e. a computational budget of ~ €1,900) and a low carbon footprint (13.5kg eq CO2).   The model's weights are also available on Hugging Face: https://lnkd.in/e-pcAMgN. Not very useful in practice, it's a PoC and not an instructed model (it's planned for later).   All the code is available on GitHub if you want to pretrain your own model in your own language or for a specific domain: https://lnkd.in/eY6e6jHJ ⭐   Ending by indicating that was a joint project with Boris Albar at CATIE 👨💻

Interior photographers spend HOURS editing lamp lighting. Finegrain just fixed that. Introducing Finegrain Light Switcher App -- an AI tool that instantly transforms off lamps to on in any image. Perfect for: - Real estate listings - Product photography - Interior design portfolios - E-commerce displays Kudos to Denis B. and the rest of the Finegrain team. Try it now on Hugging Face: Find the app link in the comments 💭

This is INSANELY cool 🤯 Chat ui + Orpheus = AI assistant with human-level emotions Type "@josh-tts Hey! [laugh] I'm surprised this works!" ↓ Get speech response that sounds genuinely happy and surprised OR Type "@josh-llm Hey! How is the weather in Amsterdam in July?" ↓ Get speech response from an LLM With this single interface, users can: - Generate human-quality emotional speech - Process images and videos - Create images with Stable Diffusion - Chat with multimodal AI models Orpheus and LLM powered Chat webui is on Hugging Face : https://lnkd.in/gxywSp_d

On March 14, Hugging Face submitted our response to the White House Office of Science and Technology Policy's request for information on the AI Action Plan. In our response (myself, coauthored with Yacine Jernite and Irene Solaiman), the overarching message is the following: Open AI systems and open science are fundamental to creating technology that is more performant, efficiently adopted, and secure. Our recommendations are centered on three interconnected pillars: 1️⃣ Recognize Open Source and Open Science as Fundamental to AI Success: The most advanced AI systems today stand on foundations of open research and software. Research has shown that investment in open systems have a strong economic multiplier effect. We need continued investment in public research infrastructure, compute access, and trusted open datasets. 2️⃣ Prioritize Efficiency and Reliability to Unlock Broad Innovation: Smaller, efficient models enable wider adoption across organizations with varying resources. Purpose-designed AI supports better in-context evaluation and resource utilization. This approach is especially critical in high-risk settings like healthcare. 3️⃣ Secure AI through Open, Traceable, and Transparent Systems: Information security history shows that open, transparent systems are essential for security. Different levels of openness can address various security requirements. Open-weight models that can run in air-gapped environments help manage information risks. Read our blog post (linked in comments) for a more detailed overview and the complete response document!

🪦 🌶️ Data annotation is dead! I just classified and open-sourced a dataset in under 10 minutes. Since the arrival of ChatGPT, many have declared the death of data labeling as we know it. 🛑 But: - Human feedback remains the cornerstone of reliable AI systems. - Pure synthetic data and AI feedback come with many known flaws: lack of diversity, accuracy, and bias amplification. - The alternative proposed by those claiming the end of data annotation is either generating entirely synthetic data or relying on AI-assisted labeling. 🟩 What if we flipped the script? Instead of helping humans with data work, why not have human experts assist AI with data work? Here’s a sneak peek at what we’re building for the open-source community —and an open dataset I created in under 10 minutes assisting Llama-3.3-70B-Instruct powered by Hugging Face Inference Providers. Link to the dataset and pipeline in the first comment.