llm-from-scratch 0.1.0

Building Language Models from Scratch — tokenizer, Transformer, training, inference, and fine-tuning in clean, tested PyTorch.

Building Language Models from Scratch

From the chain rule to a deployable model — no black boxes.

Instead of calling a library and treating the model as magic, you build every piece yourself: byte-level BPE, causal self-attention, the Transformer block, a training loop with warmup-cosine and diagnostics, a KV-cache inference engine, an evaluation harness, and a fine-tuning stack (SFT, LoRA, DPO). Each chapter is paired with a notebook that runs end-to-end on a single GPU and produces every figure the chapter cites — you see the real outputs, not idealized diagrams. The library is small (~5,000 LOC), inspectable, and backed by 190 tests; the GRU is verified against PyTorch's reference, and DPO comes with a proof.

Part of the "Beyond … and Pray" series: governed agents · trustworthy RAG · test & validate · LLMs from scratch

What's inside

• Tokenization — character / word / byte + byte-level BPE from scratch
• Transformers — attention, multi-head, RoPE / ALiBi / sinusoidal / learned, TinyGPT
• Training — Trainer, scaling laws, mixed precision, diagnostics
• Inference — greedy / beam / top-k / top-p + a KV-cache engine
• Evaluation — perplexity, calibration, LLM-as-judge, contamination checks
• Fine-tuning — SFT (loss-masked), LoRA from scratch, DPO with a proof
• 190 tests, 33 notebooks, one per chapter

Install

pip install llm-from-scratch              # core: torch, numpy, scikit-learn
pip install "llm-from-scratch[notebooks]" # + jupyterlab, matplotlib, pandas
pip install "llm-from-scratch[all]"       # everything

Python 3.12+. A CUDA-capable GPU is recommended for the training chapters.

License

Apache-2.0. © 2026 Knowlytix.