A minimal gpt training environment
As a research demo I've built Eça persona, a simple gpt model trained from scratch on Eça de Queirós books,
Eça was a well known Portuguese writer, so this persona rambles only in Português from Portugal.
The model is tiny compared with LLM's so it gets wierd very quickly.
clip.mp4
- basic tokenizer trainer
- a simple gpt model based on Andrej Karpathy lectures
- model trainer
- python inferencer
- python sampler
- onnx model export
- Web demo located on
web
- javascript compatible tokenizer
- javascript model inferencer using onnx
- javascript model sampler
python3 -m venv .venv
source .venv/bin/activate
pip3 install regex torch matplotlib
# export needs
pip3 install onnx onnxruntime
python main.py --train_tokenizer
python main.py --tokenize_data
python main.py --train
python main.py --complete "hello world"
use the parameter --project
to specify a working project, by default it will use tinyshakespeare
python main.py --project "eca_queiroz" --train_tokenizer
python main.py --project "eca_queiroz" --tokenize_message "os marinheiros encontraram o tesouro."
python main.py --project "eca_queiroz" --tokenize_data
python main.py --project "eca_queiroz" --train
python main.py --project "eca_queiroz" --complete "o mundo estava escuro" --temperature 0.3
python main.py --project "eca_queiroz" --export
# create a new folder under web with `ckpt.onnx` `project.json` and `vocabulary.json`
# update `app.js` to use your newly created folder
# run a local webserver i.e:
cd web
npx light-server -s . -p 8080