Closed thomasgauthier closed 2 weeks ago
Implemented @ElliotStein weighted overlap loss function.
I trained a merge (devout-firebrand-278) with it on the japanese/math task: https://wandb.ai/arcee-ai/Dynamic%20Adaptive%20Merging/runs/sllygkhu
Unfortunately it does not seem to beat similarity loss (proud-snowflake-238)
Implemented @ElliotStein weighted overlap loss function.
I trained a merge (devout-firebrand-278) with it on the japanese/math task: https://wandb.ai/arcee-ai/Dynamic%20Adaptive%20Merging/runs/sllygkhu
Unfortunately it does not seem to beat similarity loss (proud-snowflake-238)