I've actually been working on a visualizer for AI models. You prompt the model and it shows the tokens that were chosen, the weightings at each choice and the top X other tokens that could have been chosen at each node.
You can then click a different path and nudge the model towards that decision, like targeted fine-tuning.
It doesn't really have much practical use given the size of the models that actually get used, but I like it for demystifying how LLMs work internally and I'll probably use it for a presentation along those lines at some point in the future.
It uses React Flow (https://reactflow.dev/) for the visualization, which is pretty great for the purpose. I originally started out using Mermaid but I had to do a lot of hacking/abuse of Mermaid to get it to do what I wanted.
I've only really tried it against GPT2 so far...mostly because my hardware won't handle anything much bigger. Would be interesting to see a bigger model, but of course there's a lot more to the mainstream ones than just the model and basic token selection.
Comments
I've actually been working on a visualizer for AI models. You prompt the model and it shows the tokens that were chosen, the weightings at each choice and the top X other tokens that could have been chosen at each node.
You can then click a different path and nudge the model towards that decision, like targeted fine-tuning.
It doesn't really have much practical use given the size of the models that actually get used, but I like it for demystifying how LLMs work internally and I'll probably use it for a presentation along those lines at some point in the future.
It uses React Flow (https://reactflow.dev/) for the visualization, which is pretty great for the purpose. I originally started out using Mermaid but I had to do a lot of hacking/abuse of Mermaid to get it to do what I wanted.
I've only really tried it against GPT2 so far...mostly because my hardware won't handle anything much bigger. Would be interesting to see a bigger model, but of course there's a lot more to the mainstream ones than just the model and basic token selection.