nanoGPT

Tracked

A minimal, fast repository for training and fine-tuning medium-sized GPT models, suitable for teaching and experiments.

Author Andrej Karpathy Open Sourced 2022-12-28 Last Commit Unknown

nanoGPT, created by Andrej Karpathy, is the simplest and fastest repository for training and fine-tuning medium-sized GPT models. With a compact, highly readable codebase and minimal dependencies, it makes Transformer training workflows, data preprocessing, and optimization techniques accessible to learners and practitioners alike.

Minimal Training Pipeline

  • Strips the GPT training pipeline to its essentials, making every line of code understandable without sacrificing correctness
  • Both training from scratch and fine-tuning on smaller datasets are supported out of the box for quick experimentation
  • Example configurations and training scripts ensure reproducibility, letting anyone replicate published training workflows

Education and Prototyping

  • Widely used as a teaching tool for building deep, hands-on understanding of GPT architecture and the full training pipeline
  • Researchers rely on it for rapid prototyping of training experiments and benchmarking optimization techniques in controlled settings
  • Small teams explore model capabilities and data processing strategies without the overhead of large-scale training frameworks

Clean and Readable Codebase

  • The Python codebase prioritizes readability and experimentability, approachable from beginner to intermediate levels
  • Released under the MIT License with an active community that contributes improvements and extensions
  • One of the most widely referenced repositories in AI education and small-scale model exploration