Basic example of SAE

[annxingyuan]

Changes:

• Adds button to model viewer for downloading activations, to be used as training data for SAE
• Adds dedicated page for training SAE

The basic workflow is:

• Go to the animated transformers page
• Select the boundary decision task
• Train it until accuracy is good (~80% max)
• Click "Download activations"
• Go to /#/sae page (click on the logo, then click "SAE")
• Upload activations, then click train

Next step:

• Actually browse some learned dictionary features to see whether they're meaningful

[iislucas]

FYI: If you wanted to save all the computation, I think you could do:

const data = getData(decoderOutputs);

I guess it depends a bit on where/how much you want to serialse vs edit this code later.