MicroGPT.jl

A minimal, dependency-light Julia translation of Andrej Karpathy's microgpt.py, preserving the same core ideas:

• scalar autograd (Value)
• tiny GPT forward pass (attention + MLP)
• Adam training loop
• autoregressive character-level sampling

This repository currently contains:

• microgpt.jl — standalone Julia translation/implementation

Background

This repository is based on Andrej Karpathy's microgpt.py and follows his MicroGPT write-up:

• https://karpathy.github.io/2026/02/12/microgpt/

Run

julia microgpt.jl

Running the script will:

1. Download input.txt (if missing)
2. Train for 1000 steps on the names dataset
3. Print sampled generated names

Notes

• The Julia version uses column-major-friendly matrix layout.