🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Want to know if a ML model was trained on your dataset? Introducing ✨Data Taggants✨! We use data poisoning to leave a harmless and stealthy signature on your dataset that radiates through trained models. Learn how to protect your dataset from unauthorized use... A 🧵

Alright actual serious post. Lingua := super simple codebase + torch.compile for speed --> clean, hackable, but still efficient *It can train a 7B >llama2 in 24h*. Crazy. If you got the gpus, not only can you train a good 7B, you can *iterate* on it. You can do *research*

It's easy to add your weird new attention idea and see if it *scales*. Does it beat llama2? If two 2nd/3rd year phd students can do it, that means *you* can do it! You don't need a big team! *Just do stuff* Also Badr Youbi Idrissi is on the market in May and Mathurin next Oct. Just sayin.

Meta Lingua: a minimal, fast LLM codebase for training and inference. By researchers, for researchers. Easily hackable, still reproducible. Built-in efficiency, profiling (cpu, gpu and mem) and interpretability (automatic activation and gradient statistics) Joint work w/ Badr Youbi Idrissi

🔒Image watermarking is promising for digital content protection. But images often undergo many modifications—spliced or altered by AI. Today at AI at Meta, we released Watermark Anything that answers not only "where does the image come from," but "what part comes from where." 🧵

🎉Exciting news from AI at Meta FAIR! We've released a Watermark Anything Model under the MIT license! It was announced yesterday: ai.meta.com/blog/meta-fair… Great project with Pierre Fernandez et al. ! We're close to hitting 1,000 stars on GitHub. Give it a try: github.com/facebookresear… 🚀

Outstanding Finalist 2: “DINOv2: Learning Robust Visual Features without Supervision," by Maxime Oquab, Timothée Darcet (TimDarcet), Théo Moutakanni (Théo Moutakanni) et al. 5/n x.com/AIatMeta/statu…

Very happy to say DINOv2 got outstanding certification finalist at TMLR! The models had an amazing reception already, but this kind of award is the cherry on top 😁

📢Exciting announcement today 📢 #EndoDINO is a foundation model for #GI #Endoscopy. We believe this represents a new paradigm for #AI development in the field! EndoDINO - inspired by AI at Meta - will democratize/accelerate new AI research in gastroenterology. More below 👇

🔥 The DINO team is looking for a PostDoc! 🔥 If you are about to graduate, and want to be part of what’s next for SSL, don’t hesitate to reach out! Link to job offer : metacareers.com/jobs/502476149…

Want strong SSL, but not the complexity of DINOv2? CAPI: Cluster and Predict Latents Patches for Improved Masked Image Modeling.

Somewhat poorly advertised MS program in Paris: MVA master-mva.com is a 1-year research Master's, with (mostly English) courses across Maths, Vision, Learning, and an internship (that can lead to PhD). Applications open May 1 - June 30 for 2025-2026. Grants possible.

This is my first paper done at FAIR. We show that adaptive visual token segmentation, especially in subobject-level (i.e., subwords in images), enables VLMs to have a better and faster learning of image understanding! arxiv.org/pdf/2402.14327

Does your LLM truly comprehend the complexity of the code it generates? 🥰   Introducing our new non-saturated (for at least the coming week? 😉) benchmark:   ✨BigO(Bench)✨ - Can LLMs Generate Code with Controlled Time and Space Complexity?   Check out the details below !👇

🔥Very happy to introduce BigO(Bench) dataset on Hugging Face 🤗 ✨3,105 coding problems and 1,190,250 solutions from CodeContests ✨Time/Space Complexity labels and curve coefficients ✨Up to 5k Runtime/Memory Footprint measures for each solution  huggingface.co/datasets/faceb…

🚨 Your RL only improves 𝗽𝗮𝘀𝘀@𝟭, not 𝗽𝗮𝘀𝘀@𝗸? 🚨 That’s not a bug — it’s a 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲 you’re optimizing. You get what you optimize for. If you want better pass@k, you need to optimize for pass@k at training time. 🧵 How?

Happy to share our latest work: V-JEPA 2, a world model trained from millions of videos, that enables visual understanding, planning and physical reasoning!

DINOv2 meets text at #CVPR 2025! Why choose between high-quality DINO features and CLIP-style vision-language alignment? Pick both with dino.txt 🦖📖 We align frozen DINOv2 features with text captions, obtaining both image-level and patch-level alignment at a minimal cost. [1/N]

🚨New AI Security paper alert: Winter Soldier 🥶🚨 In our last paper, we show: -how to backdoor a LM _without_ training it on the backdoor behavior -use that to detect if a black-box LM has been trained on your protected data Yes, Indirect data poisoning is real and powerful!

I'm attending ICML'25 in Vancouver. Will present: 1) Subobject-level adaptive image token segmentation (main conference) arxiv.org/abs/2402.14327 2) WorldPrediction benchmark for world modeling and procedual planing (in Assessing World Models workshop) arxiv.org/abs/2506.04363