Skip to content

Add AdamW–MuonClip nanoGPT WeightWatcher campaign - #115

Merged
charlesmartin14 merged 1 commit into
mainfrom
feature/nanogpt-adamw-muonclip-5seed
Aug 23, 2026
Merged

Add AdamW–MuonClip nanoGPT WeightWatcher campaign#115
charlesmartin14 merged 1 commit into
mainfrom
feature/nanogpt-adamw-muonclip-5seed

Conversation

@charlesmartin14

Copy link
Copy Markdown
Member

Summary

  • add the reproducible one-head FineWeb-Edu nanoGPT campaign
  • compare exactly AdamW and MuonClip over five paired seeds
  • run seed 1337 first, with provisional analysis before the other seeds
  • record WeightWatcher alpha and raw_alpha from one fix_fingers="clip_xmax" call
  • retain ERG gap, correlation traps, detX, performance metrics, strict restart/provenance, and Mac /tmp workflow

Validation

  • 151 passed in the complete baseline/nanogpt_one_head/tests suite
  • production doctor passed AdamW and MuonClip checkpoint/resume
  • real WeightWatcher smoke passed with six corrected alphas and six raw alphas

@charlesmartin14
charlesmartin14 merged commit 54419d9 into main Aug 23, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant