Skip to content

Alternative Losses #54

Description

@ClashLuke

Currently, we're using only the softmax classification/cross-entropy loss to create a language-modeling loss for next-token prediction. However, other works such as T-Few showed that adding alternative losses for external benefits such as length explicit penalties during training can help downstream-task performance. Additionally, other works like DCL and InfoLOOB demonstrated that changing the fundamental structure of the loss from softmax classification to something different can help speed up convergence. That's why a similar approach could be beneficial for us.
In this issue, we'll explore whether InfoLOOB's classification loss for the language-modeling objective helps or if we should change the entire objective.

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    MLRequires machine-learning knowledge (can be built up on the fly)coreImproves core model while keeping core idea intactresearchCreative project that might fail but could give high returns

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions