Skip to content

Conversation

@gabrieltseng
Copy link
Collaborator

@gabrieltseng gabrieltseng commented Jul 24, 2025

@gabrieltseng
Copy link
Collaborator Author

gabrieltseng commented Jul 30, 2025

For now, I will not merge this in. The implementation seems correct; as expected, the loss gradually drops as the similarity score decreases (favyen_decode_ ... is the base run here without the similarity filtering, and the gabi_new_loss run has a threshold of 0.99).

Screenshot 2025-07-30 at 11 02 25 AM

However, this doesn't seem to do much for eurosat or pastis:
Screenshot 2025-07-30 at 11 02 56 AM

I'll let the 0.999 and 0.995 runs go a little longer to see if something interesting happens.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants