research-publishing

Research Publishing Methodology

You are helping a researcher prepare their code and artifacts for public release alongside a paper submission.

Step 1: Repository Assessment

Before any changes, audit the current state:

1. Sensitive content scan:

   • API keys, tokens, credentials (grep for common patterns)
   • Hardcoded paths specific to the researcher's machine
   • Internal URLs or private infrastructure references
   • Personal identifiable information in comments or data

2. Dependency audit:

   • List all dependencies with pinned versions
   • Identify any proprietary or restricted-license dependencies
   • Check for abandoned/unmaintained dependencies
   • Verify all dependencies are pip/conda installable

3. Code organization:

   • Identify dead code, debugging artifacts, scratch files
   • Find duplicated code that should be unified
   • Check for overly complex code that can be simplified

Step 2: Repository Structure

A publishable research repository should have:

project/
  README.md            # Installation, usage, citation
  LICENSE              # Must have an explicit license
  requirements.txt     # or pyproject.toml with pinned deps
  setup.py / setup.cfg # Package installation
  src/                 # Source code
  scripts/             # Training, evaluation, inference scripts
  configs/             # Configuration files
  data/                # Sample data or download instructions
  checkpoints/         # Download instructions (not actual weights)
  results/             # Key result files referenced in paper

Step 3: Reproducibility Checklist

For each experiment in the paper:

• Configuration file exists and matches paper's hyperparameters
• Random seeds are set and documented
• Training command is documented end-to-end
• Evaluation command produces the reported numbers
• Data preprocessing steps are scripted (not manual)
• Hardware requirements are documented (GPU type, memory, time)
• Dependencies are version-pinned