LLMStead: Raise your own models

LLMStead builds a tiny transformer, the same kind of model behind modern AI chat assistants, and trains it from scratch right in your browser. Every step is open to inspect, so you can watch a language model go from random numbers to making real predictions.

1. Set up: pick a dataset, a tokenizer vocabulary size, and the model's architecture.
2. Train: watch the training and validation loss fall, then benchmark what it memorized versus what it learned.
3. Look inside: follow a prompt through tokenization, embeddings, attention, and next-token prediction.

All the machine-learning code (tensor math, backpropagation, the optimizer, the BPE tokenizer) is written from scratch with no libraries. LLMStead needs JavaScript to run. View the source on GitHub.