Hugging Face

I talk a lot about small models for GLAM institutions, but I realise there aren't enough concrete examples to point people to. So with William James Mattingly, Ph.D., I'm trying to do something about it: a small-models-for-glam on Hugging Face org! The idea: - Cultural heritage institutions have specific AI needs - They have data that is often quite different from the kinds of data general-purpose LLMs / VLMs are optimised for. - They lack extensive compute infrastructure. Small specialised models (~0.6B-4B parameters) trained on domain-specific use cases and data can often:   - Outperform general LLMs on domain tasks - Run on modest hardware (even laptops) - Fine-tune quickly on institutional data First models: ICONCLASS classification, Getty AAT materials, historical name parsing (with lots of ideas for more models!) We're slowly starting to add some models to the org and will start curating other examples already on the Hub (let me know if there are some we should include). Early days, but excited to see if this helps make specialised models more visible and accessible for cultural heritage AI work. Follow the org for updates! 👉 https://lnkd.in/e5d8BdJu

ICCV'25 will be a blast 💥 Will kick things off with a talk at the LongVid workshop. Talk is titled: "Controllable Video Diffusion for Length" 📹 Then I have got two works to flaunt: 1. SANA Sprint: The recipe to train SOTA few-step image generation models 2. ReflectionFlow: Enabling reasoning capabilities in diffusion models for image generation (dataset, evaluation, training, inference-time scaling) Very thankful to all the collaborators here! Make sure to check them and their works 🤗 If you're attending and working on these topics, let's meet and talk about stuff!

Day 2 on nanochat and it’s starting to get real! We now have a full set of demos and models for folk to try out: 🔗 https://lnkd.in/e39nJb4Q - We've published pretrained models from base training, mid training, sft, and rl. Which means you can drop into the process at any of these points without investing in pre-training. …working on post training demos! - There are demo spaces for base and chat examples. These are just fun and educational to help folk see how the models performs. You can also reuse them if you want to demo your own checkpoints - There's inference code for transformers for base and chat models. All the model repos have working snippets and notebook examples. ... still need to port this to a transformer fork. - As before, we’ve got the community on the hub to share work, a share trackio space for logs, and all the models and demos in one place. > > next step is to get this model inferring on CPU and in JS. Then move on to some longer form tutorials and notebooks on post training options. I’ll share everything in the thread.