microgpt.lua (300 lines) — a Lua port of Andrej Karpathy's microgpt. Runs standalone, or with alignment scripts to verify exact numerical match: same loss, same samples. Lua interpreter + microgpt.lua + alignment scripts + training data — everything needed, combined under 0.5MB.
lua microgpt.luaTrains for 1,000 steps on input.txt (a names dataset, included), then samples 20 generated names.
Takes a few minutes on a modern laptop.
Requirements: Lua 5.1+, no external dependencies.
| File | Description |
|---|---|
microgpt.lua |
300 lines, self-contained. Autograd, model, training, inference |
pyalign.lua |
Alignment test against the Python reference |
pyrand.lua |
Python-compatible RNG used by pyalign.lua |
input.txt |
Training data: 32K common names (via microgpt) |
- Small: Lua binary (~184K) + script (~10K) + training data (~223K) = ~417KB total
- Simple: Lua has only one composite data structure — the table — which serves as array, map, and object
- Sufficient: tables are enough to implement autograd — both forward and backward passes for microgpt (see Autograd in the microgpt blog)
pyalign.lua verifies that outputs match the Python reference implementation:
lua pyalign.luaLoss values and sampled names align with the microgpt blog.
MIT