Documentation

[Bachstelze]

Is there somewhere a documentation about this project? How are the attention weights aggregated?

[mattneary]

Here's the code that aggregates. Nothing fancy, just averaging across layers and attention heads.

If you're curious about more advanced aggregation, I started exploring learned aggregators a few months ago in MIA. If you can produce training data that links input text to output text in some meaningful way, you can train a classifier on top of the attention values to predict that relationship.