🌍 Open-Source Research

Releasing your research code is no longer optional — it is the baseline expectation at top venues. A paper without code is increasingly treated as a second-class contribution.

1. Why Open-Source Your Research

The Scientific Argument

Reproducibility requires access to your implementation. Without it, other researchers cannot:

  • Verify your results
  • Build on your method correctly
  • Identify bugs in your evaluation
  • Reuse your components in their work

The Career Argument

A well-maintained research repository is a career accelerator:

  • High GitHub stars signal community recognition
  • Open-source users become paper citers and future collaborators
  • Industry recruiters specifically look for research code quality
  • It demonstrates software engineering competence alongside research depth

The Evidence

Studies consistently show that papers releasing code receive significantly more citations than papers without code, even controlling for paper quality.

2. What to Release

Minimum Viable Release

  • All code used to produce results in the paper
  • Pre-trained model weights (for applicable papers)
  • Data preprocessing scripts
  • Evaluation scripts
  • A README.md with installation and quick-start instructions

Ideal Release Package

  • Clean, documented code (not your messy experiment branch)
  • Training script with all hyperparameters configurable from CLI or config file
  • Inference script / demo notebook
  • Pre-trained weights for all model sizes/variants
  • Example data or synthetic demo data
  • Comprehensive README with method overview, installation, training, evaluation, results
  • Docker image for full reproducibility
  • Colab or Kaggle notebook for interactive demo
  • CITATION.cff or BibTeX entry for easy citation

3. Repository Quality Standards

README Template

# [Method Name]: [One-sentence description]

[Paper badge] [License badge] [Stars badge]

> Brief description of what this repository contains.

## Abstract
[1-2 sentence summary of the paper's contribution]

## Key Results
| Dataset | Metric | Ours | Previous SOTA |
|---------|--------|------|---------------|
| CIFAR-10 | Top-1 Acc | 97.8% | 97.2% |

## Installation
\`\`\`bash
git clone https://github.com/username/method-name
cd method-name
pip install -e .
\`\`\`

## Quick Start
\`\`\`python
from method import Model
model = Model.from_pretrained("username/method-name-base")
result = model.predict(input_data)
\`\`\`

## Training
\`\`\`bash
python train.py --config configs/cifar10.yaml --seed 42
\`\`\`

## Evaluation
\`\`\`bash
python evaluate.py --checkpoint checkpoints/best.pt --split test
# Expected output: Accuracy: 97.8%
\`\`\`

## Pretrained Models
| Model | Dataset | Accuracy | Download |
|-------|---------|----------|----------|
| Method-Base | CIFAR-10 | 97.8% | [HuggingFace](https://huggingface.co/...) |

## Citation
\`\`\`bibtex
@inproceedings{yourname2024method,
  title={Your Paper Title},
  author={Your Name and Co-author},
  booktitle={Conference Name},
  year={2024}
}
\`\`\`

## License
MIT License

Code Quality Checklist

  • Consistent code style (use black for Python formatting)
  • Type hints on all public functions
  • Docstrings on all modules, classes, and non-trivial functions
  • Unit tests for core components (aim for >60% coverage)
  • No hardcoded paths or secrets
  • Clear separation of configuration from code
# Code formatting
pip install black isort
black src/
isort src/

# Linting
pip install flake8 pylint
flake8 src/ --max-line-length 100

# Type checking
pip install mypy
mypy src/

4. Licensing Your Research Code

Choose the right license before releasing:

License Allows Commercial Use Requires Attribution Copyleft Best For
MIT ✓ ✓ ✗ Most research code
Apache 2.0 ✓ ✓ ✗ Research with patent protection
GPL v3 ✓ ✓ ✓ Code you want kept open-source
CC BY 4.0 ✓ ✓ ✗ Datasets and models
CC BY-NC 4.0 ✗ ✓ ✗ Non-commercial research only

[!TIP] MIT is the community default for most CS research code. It maximizes adoption and places minimal burden on users, encouraging citation and contribution.

5. Publishing on Hugging Face Hub

Hugging Face Hub has become the de-facto standard for releasing ML models and datasets.

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("checkpoints/best")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# Upload model
model.push_to_hub("username/my-research-model")
tokenizer.push_to_hub("username/my-research-model")

# Create a model card (README on the hub)
# IMPORTANT: Write a model card at the model repo on HuggingFace

Model Card Template

---
language: en
license: mit
tags:
  - text-classification
  - research
datasets:
  - my_dataset
metrics:
  - accuracy
---

# Model Name

## Model Description
This model was released with the paper "[Title](arxiv link)".

## Intended Uses and Limitations

## Training Details

## Evaluation Results

## Citation

6. Papers With Code Submission

Papers With Code is the standard platform for linking papers to code and benchmark results.

  1. Go to paperswithcode.com
  2. Search for your paper (it may already be indexed from arXiv)
  3. Add your GitHub repository link
  4. Submit your results to the relevant benchmark leaderboards

This dramatically increases discoverability — researchers searching for implementations of your method find it immediately.

7. Maintaining Your Repository

Releasing code is the beginning, not the end:

Post-Release Checklist

  • Set up GitHub Issues for bug reports and feature requests
  • Add a CONTRIBUTING.md if you want community contributions
  • Respond to issues within 1–2 weeks (builds community trust)
  • Tag releases with version numbers (v1.0.0)
  • Update the README as the API evolves
  • Add a CHANGELOG.md

Further Reading