Skip to content

Matched Canon vs no-Canon rerun for the methods record #116

Open

Description

@briney

Why: The methods evidence record for "Canon layers are on" is marked needs matched rerun. The no-Canon arm's learning rate is not bracketed (still improving at the highest rate tested, 0.01), and the comparison is at equal steps with step times measured before the fused Canon kernel.
Context: docs/methods/evidence/canon-ablation/ (PR #115). Original runs: W&B oplm-mup-4M-nocanon, 170M, 20k stable + 4k decay, commit 6e8b277, cluster ~/projects/oplm/mup_redo/nocanon.sh and sweeps-nocanon/. Fused kernel: PR #96.
Done when: no-Canon cells above lr 0.01 bracket its optimum, step time is re-measured on main for both arms, the record restates the comparison at equal wall-clock, and its status is settled.

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

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions