Paper‑Replications – From‑scratch PyTorch implementations of classic and SOTA AI/ML papers

What it is

• A collection of self‑contained PyTorch code that reproduces the results of many influential machine‑learning papers. Each sub‑directory focuses on a single model or technique (e.g., BERT, GPT, Llama, Diffusion, GAN variants, LoRA, Whisper, etc.) and contains the scripts, notebooks, and sometimes pretrained weights needed to run the experiments described in the original work.

Why it matters

• Re‑implementations are valuable for learning how a paper’s algorithm works, for verifying claims, and for providing a baseline that can be extended or fine‑tuned. By offering a wide‑range of models—from classic RNNs to the latest LLMs—this repo serves as a practical “lab notebook” for students and researchers who want to dive into the code behind the headlines.

Key contents

• Folder hierarchy mirrors major research areas (Attention Mechanisms, Transformers, GANs, Vision‑Language models, RLHF methods, etc.).
• Inside each folder you’ll typically find:
  • A README or notes summarising the original paper and any deviations.
  • Python scripts/notebooks that build the model, train it, and evaluate it.
  • Optional pretrained checkpoints or result logs for quick sanity checks.
• The repository also links to smolhub where some of the models are deployed as inference endpoints.

Getting started

1. Clone the repo and navigate to the folder of the paper you want to explore.
2. Follow the local README/notebook instructions – usually just pip install -r requirements.txt and then run the provided training or inference script.
3. For quick verification you can compare the reproduced metrics (often printed at the end of the script) with the numbers reported in the original publication.

Who maintains it

• Authored by Yuvraj Singh (GitHub @YuvrajSingr‑mist) and open to community contributions via issues or pull requests.
• Sponsored on GitHub and through a “Buy me a coffee” link, with additional thanks to two individual sponsors.

License

• Code and notebooks are released under the Apache 2.0 license.
• Documentation, diagrams, generated results and any published artifacts are under CC BY 4.0.

Cite if you use it

@software{singh_paper_replications_2025,
  title     = {Paper-Replications: From‑Scratch PyTorch Replications of Classic and SOTA AI/ML Papers},
  author    = {Singh, Yuvraj},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23198171},
  url       = {https://github.com/YuvrajSingh-mist/Paper-Replications}
}

Bottom line

• If you want concrete, runnable PyTorch code that mirrors the behavior of well‑known AI papers, this repo is a ready‑made toolbox. It’s not a library for production use, but an educational and research‑oriented resource for reproducing and extending state‑of‑the‑art models.