We just released an open-source framework that makes it easy to build visual reasoning agents (with GRPO).
github.com/groundlight/r1…
github.com/groundlight/r1…
I’ve used it to train a tiny model to solve cryptograms (a task large pre-trained vision language models can’t solve), but you can use it to train on whatever task you want! Personally, I’d love it if someone trained a model that could solve mazes.
My collaborators and I published a blog post with our initial learnings from this project where we discuss our learnings, including critical insights for reward design: groundlight.ai/blog/visual-re…
Thank you to my collaborators for their contributions to this project: @leopd and @BowenROIM and a shoutout to @willccbb for his work on verifiers, which really improved the ergonomics of the code.
Happy to answer any questions here (just reply!) or on GitHub.
Happy to answer any questions here (just reply!) or on GitHub.
@leopd @BowenROIM @willccbb PS: we’re working on multi turn conversations and tool use. Stay tooned!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
