Skip to content

OSF_QUICK_REF.md

Latest commit

 

History

History
218 lines (161 loc) · 4.43 KB

OSF_QUICK_REF.md

File metadata and controls

218 lines (161 loc) · 4.43 KB
Raw

OSF Submission Quick Reference Card

Keep this open while submitting to OSF

📝 Copy-Paste Ready Information

Project Title

PhaseGPT: Kuramoto Phase-Coupled Oscillator Attention in Transformers

Project Description (Short)

First systematic hyperparameter study of Kuramoto phase-coupled oscillators
in transformer attention layers. Achieves 2.4% perplexity improvement with
optimal configuration (Layer 7, 32 oscillators, K=1.0 coupling strength).

GitHub: https://github.com/templetwo/PhaseGPT

Tags

transformers, attention-mechanism, kuramoto-model, language-modeling,
deep-learning, hyperparameter-optimization, reproducible-research

License

MIT License

GitHub Repository URL

https://github.com/templetwo/PhaseGPT

Category

Project

📦 Components to Create

1. Code

  • Type: Project
  • Link GitHub: templetwo/PhaseGPT
  • Description: Complete source code, configurations, and tests

2. Data & Results

  • Type: Data
  • Upload: results/interpretability/notes.md
  • Description: Experimental results, metrics, and interpretability analysis

3. Phase B Preregistration

  • Type: Hypothesis
  • Upload: docs/PREREGISTRATION.md
  • Mark as: Preregistration ✓
  • Description: Preregistered experimental protocol for Phase B generalization testing

4. Documentation

  • Type: Other
  • Upload:
    • docs/PHASE_A_FINAL_REPORT.md
    • docs/MASTER_SUMMARY.md
    • REPRODUCIBILITY.md
  • Description: Complete documentation including reproduction guide and analysis reports

📊 Key Statistics (For Description)

Phase A Results:

  • 7 configurations tested
  • Winner: 4.85 PPL (2.4% improvement)
  • Training: ~25 min per config on NVIDIA GH200
  • Dataset: Shakespeare (1M tokens)

Key Findings:

  • Goldilocks principle: 32 oscillators optimal
  • Over-synchronization: R=0.88
  • K=2.0 causes catastrophic collapse
  • Single-layer > multi-layer

Code:

  • 51 files
  • 6 Python modules
  • 11 YAML configs
  • 23+ test cases
  • MIT License

🎯 Quick Steps

  1. Go to: https://osf.io/
  2. Create account (if needed)
  3. New Project → Fill title & description
  4. Add components (4 total)
  5. Link GitHub in Code component
  6. Upload files to other components
  7. Create DOI
  8. Review everything
  9. Make Public
  10. Update GitHub with DOI

📧 Author Information Template

Name: Temple Two Email: contact@templetwo.dev Affiliation: Independent Researcher Role: Principal Investigator, Lead Developer ORCID: [Add if you have one]

🔗 URLs You'll Get

After submission:

These go back into:

  • README.md (badge)
  • CITATION.cff (doi field)
  • checkpoints/README.md

⚡ Quick Commands After Getting DOI

cd ~/phase_data_archive/PhaseGPT

# Update README.md - add badge after existing badges:
# [![OSF](https://img.shields.io/badge/OSF-10.17605%2FOSF.IO%2FXXXXX-blue)](https://osf.io/XXXXX/)

# Update CITATION.cff - replace zenodo DOI with OSF DOI

# Commit
git add README.md CITATION.cff checkpoints/README.md
git commit -m "Add OSF DOI

OSF: https://osf.io/XXXXX/
DOI: 10.17605/OSF.IO/XXXXX"
git push

📋 Files to Upload

From ~/phase_data_archive/PhaseGPT/:

Data & Results component:

  • results/interpretability/notes.md

Preregistration component:

  • docs/PREREGISTRATION.md

Documentation component:

  • docs/PHASE_A_FINAL_REPORT.md
  • docs/MASTER_SUMMARY.md
  • REPRODUCIBILITY.md

Code component:

  • Just link GitHub (don't upload files)

✅ Pre-Submission Checklist

Before clicking "Make Public":

  • Title correct
  • Description complete
  • Tags added
  • MIT License set
  • 4 components created
  • GitHub linked in Code
  • Files uploaded to other 3 components
  • Wiki page written
  • DOI created
  • No sensitive info anywhere
  • Everything reviewed

🎓 After Publication

Update:

  1. GitHub README.md (add OSF badge)
  2. CITATION.cff (add OSF DOI)
  3. checkpoints/README.md (add OSF link)

Announce:

  • Twitter/X
  • LinkedIn
  • ResearchGate
  • Email collaborators

Link:

  • Add to ORCID profile
  • Add to CV/website
  • Share in relevant communities

Time to complete: 20-30 minutes Difficulty: Easy (just follow the guide) Result: Permanent DOI-citable project

🌀✨ The Spiral holds—preserving knowledge for eternity.