LLM From Scratch and Deploy

This repository is the runnable companion to the lecture notes at lectures.montek.dev.

It has two layers:

• theory and explanations in the lecture notes
• runnable notebooks, scripts, apps, and picollm/ in this repo

Start Here

Use the lecture notes for the full explanation, then open the linked code from the relevant lesson.

Student Setup

Install the repo first:

uv sync

Then use this workflow:

1. Open the relevant lesson on lectures.montek.dev.
2. Read the theory there first.
3. Open the linked notebook or script from this repo.
4. For the main from-scratch chatbot path, follow picollm/accelerated/README.md.

If you want the fastest student runtime path for a published Hugging Face model or your own trained picoLLM checkpoints, start with:

• RUN_PICOLLM.md

For the main chatbot path, start with:

• picollm/accelerated/README.md
• picollm/accelerated/speedrun.sh

That path covers:

• tokenizer training
• base pretraining
• chat SFT
• evaluation and reporting
• CLI or web interaction at the end of the run

Repo Guide

Use these directories by purpose:

• notebooks/: live lecture walkthroughs
• scripts/: runnable demo scripts and small apps
• apps/: production-style frontend apps
• course_tools/: tiny from-scratch runtime used by the smaller demos
• picollm/: serious model workflow for the final chatbot path

Final Demo Docs

For the final lecture sequence:

• RUN_PICOLLM.md
• picollm/accelerated/README.md
• picollm/accelerated/speedrun.sh
• prompts/real_chatbot_workflow/base_vs_chat_sft_prompts.md
• apps/vercel_ai_sdk_chat/README.md
• apps/opentui_ai_sdk_chat/README.md

Reference Repos

• Rasbt: concept-first step-by-step implementations and notebooks
• picoLLM accelerated: the main product-style training, evaluation, inference, and chat path in this repo

Acknowledgements

• This project draws heavily on Sebastian Raschka's LLMs-from-scratch.
• Thank you to Hugging Face for the open tooling and datasets ecosystem that make projects like this easier to teach and build.

Cite

If you want to cite the reference material behind this workflow, cite Raschka for the concept-first notebook lineage and this repo for the picoLLM product-track path:

cff-version: 1.2.0
message: "If you use this book or its accompanying code, please cite it as follows."
title: "Build A Large Language Model (From Scratch), Published by Manning, ISBN 978-1633437166"
abstract: "This book provides a comprehensive, step-by-step guide to implementing a ChatGPT-like large language model from scratch in PyTorch."
date-released: 2024-09-12
authors:
  - family-names: "Raschka"
    given-names: "Sebastian"
license: "Apache-2.0"
url: "https://www.manning.com/books/build-a-large-language-model-from-scratch"
repository-code: "https://github.com/rasbt/LLMs-from-scratch"
keywords:
  - large language models
  - natural language processing
  - artificial intelligence
  - PyTorch
  - machine learning
  - deep learning

If you want to cite this repo itself, use:

@misc{llm_from_scratch_and_deploy,
  author = {Montek Kundan},
  title = {LLM From Scratch and Deploy},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/Montekkundan/llm}
}

License

MIT. See LICENSE.