skunkworks

RobotArena is live today! Led by @sumo43_, RobotArena is an ELO-based 🤖 Robot-Action Model benchmark that allows you to test models directly in your browser. Check it out here: robotarena.ai Special thanks to @HIVEDigitalTech for their compute support!

@sumo43_ @HIVEDigitalTech Do give us feedback/ideas on how to expand RobotArena further👇

We’d like to thank @HIVEDigitalTech for their reliable compute support in enabling open source AI research @skunkworks_ai for this project and beyond. Feel free to reach out to us if you want to contribute or accelerate your open source AI research with compute!

Reasoning is an integral component of capable and competent AI models. We’re releasing one of our first reasoning data subsets for the benefit of the open source community. More data to be released along with some of our reasoning models. huggingface.co/datasets/Skunk…

Made a tiny-frankenstein of TinyLlama @LeopolisDream Works great even at 942Mb size. Full model, secret sauce and custom gguf quants included. @skunkworks_ai @far__el huggingface.co/SkunkworksAI/t…

quick question- should we open source failed projects too?

@arbius_ai Amazing!

Amica supports vision as well. Use your camera to take pictures and Amica will be able to see them. This is powered by BakLLaVa (github.com/SkunkworksAI/B…) which feels completely from the future. @skunkworks_ai thank you for making such an amazing alternative to GPT-4V.

Introducing... Amica! 👩‍🦰🤖 Amica is an open source interface for interactive communication with 3D characters with voice synthesis, speech recognition, visual understanding, and an emotion system. 🔗 heyamica.com

Good tutorial for using @skunkworks_ai BakLLaVA!

Mirror: A hackable AI-powered Mirror on Your Laptop. Mirror is a simple yet powerful web app that runs 100% locally, where the AI (@ggerganov's llama.cpp + @skunkworks_ai's Bakllava) constantly watches your webcam feed and sends you messages. And you can try it too, right now.

👀bakLLaVA vision AI can read xrays with only 6Gb of RAM @far__el @skunkworks_ai Easy code colab to run it on your own via @ggerganov ./server below👊 Looking to train fully opensource model for cancer tissue detection next 🩺 If anyone can clear the gunk of bureacracy & just provide that data dear god please do. Crazy part is.. model wasn't even trained for this purpose, general image data was limited to 384:384px, yet it can read🤯words.. and reason. Colab server: colab.research.google.com/drive/1uHqkj1y… (gotta make a free ngrok account to open gguf webserver remotely ) Github: github.com/SkunkworksAI/B… Hugginface: huggingface.co/SkunkworksAI/B… Quantized ggml downloads: huggingface.co/mys/ggml_bakll… Thank you @imhaotian for building and opensourcing the orignal LLaVA vision-llama code that we used to train on top of @MistralAI 7b with.

@togethercompute @ontocord @Sentdex @laion_ai @alignment_lab @ClementDelangue @huggingface Please do check out the OG LLaVA implementation by @imhaotian! Here: llava-vl.github.io

As always, thank you to our compute sponsor @togethercompute and to our collaborators @ontocord @Sentdex @laion_ai @alignment_lab . We can't wait to show you what we're up to with the next updates! Demo for BakLLaVA-1 coming soon. (@ClementDelangue @huggingface 👀)

BakLLaVA-1 is now available to everyone. We show Mistral-7B with vision outperforms Llama-2 13B on several tasks, but we've only scratched the surface. BakLLaVA-2 is cooking with a larger corpus and significant architectural improvements/innovations. huggingface.co/SkunkworksAI/B…

@AdarrabYoussef Not hard for BakLLaVA-1! BakLLaVA-2 will be a different story 😉

@AdarrabYoussef BakLLaVA-1's code is a fork of llava-vl.github.io adapted for Mistral-7B. Here's our repo (although some updates coming over the next day to clean up and incorporate some new features we added) github.com/SkunkworksAI/B…

Our first large H100 cluster (4424 GPUs) starts coming online today! Available as sub-clusters configured for pretraining models. We plan to prioritize the startups and research institutions. 🚀👾 Request access: together.ai/training-reque…

We're getting hundreds of access requests every day since we announced BakLLaVA v1 🚀 We'll open the flood gates shortly. Stay tuned.

@felix_red_panda @thukeg @vilm_hq V1 is out in the next day 🚀🚀🚀

List of contemporary GPT4V-like open source models - Blip2 - Llava1.5 - CogVLM by @thukeg - StableVL by @vilm_hq (not out yet) - @skunkworks_ai 's model (not out yet) Did I miss any?

This model was forged thanks to the generous support from @togethercompute. We'll continue to push the multimodal model further in collaboration with @ontocord @cliangyu_ @laion_ai @alignment_lab with an even more powerful v2 coming soon.

BakLLaVA 1 can successfully see! And we're not even half way through our run 👀 The base model is a Mistral 7B, so this is possibly the first @MistralAI vision model! Model is hosted here and will be fully released soon: huggingface.co/SkunkworksAI/B…