The "Logit Lens" is a technique for decoding intermediate hidden states into predictions. It makes it easy to see the gradual change of predictions from the first to last layer.
It's commonly done alongside visualising attention patterns and other activations (which CircuitsVis excels at). The Top K visualisation looks similar to the logit lens, so the visual capabilities are there, it just needs to be implemented.
I'm happy to make a PR for this if others are interested. Otherwise I will likely fork and implement for my own use.
Thanks 🙏
The "Logit Lens" is a technique for decoding intermediate hidden states into predictions. It makes it easy to see the gradual change of predictions from the first to last layer.
It's commonly done alongside visualising attention patterns and other activations (which CircuitsVis excels at). The Top K visualisation looks similar to the logit lens, so the visual capabilities are there, it just needs to be implemented.
I'm happy to make a PR for this if others are interested. Otherwise I will likely fork and implement for my own use.
Thanks 🙏