The LLM Project: Start Here
Since late 2025 I've been building a language model from scratch in my home office: my own tokenizer, my own architecture, my own data pipeline, trained on my own hardware. Not a fine-tune of somebody else's weights. From-scratch, on a bounded corpus (301 Wikipedia articles about World War II), because a bounded corpus is the only place you can actually measure what a model learned versus what you taught it.
The posts below tell that story in order. If you only read one, read the full story from February. Everything here happened on personal time with personal resources, and everything is real, including the failures, which get named honestly.
The Main Line
From N-Grams to Micro-Models: Training AI on a 1080 Ti
Where it started: how far can you get teaching a machine language on an old gaming card? Further than I expected, and not nearly far enough.
Does Quantization Make Models "Stupid"? A Home Lab Experiment
Before trusting quantized models as tools in my pipeline, I measured what quantization actually costs. Numbers, not vibes.
When Bad Shuffling Breaks Training: A Home Lab Lesson
A sawtooth in my loss curve taught me more about data loading than any tutorial ever did. The bug was mine; the lesson is general.
Building an LLM From Scratch
Five months, 35,000 lines of Python, 711 commits: the synthetic data pipeline, the architecture decisions, the failures with honest names, and the model that finally started answering questions.
D-Day Quiz: A From-Scratch LLM Passes a Historical Exam
The payoff moment: the model sits a real exam on its domain and passes.
Using the LLM Last: Deterministic-First Typography Cleaning
A data-quality philosophy in miniature: exhaust the deterministic tools first, spend inference only where judgment is genuinely needed.
My First Dataset on Hugging Face: A 21-Million-Pair WWII Synthetic Corpus
The training data goes public: 21+ million conversational QA pairs across 29 subsets, CC BY-SA, built entirely with permissively licensed models.
Automating AI Quality Checks: A Self-Improving Model Loop
The current frontier: a 27B model interrogates my model topic by topic, triangulates every weakness, and writes the problem reports that become the next round of training data. QA that runs while I sleep.
Artifacts
WWII Synthetic Corpus (Hugging Face dataset)
21+ million conversational QA pairs, 29 subsets, CC BY-SA 4.0: huggingface.co/datasets/wayneworkman2012/ww2-synthetic-corpus
The model itself
A from-scratch release is in the works. The current training cycle (probe fleet, failure taxonomy, targeted remediation) feeds it. Subscribe or check back for the release post.
Side Quests
EmailLM: How I Solved My Spam Problem with Local AI
The same local-AI toolbox pointed at a household problem.
Understanding Prompt Injection: A Hands-On Demo
Adversarial thinking about LLMs, and the mindset that later became the probe fleet.
YAML vs JSON: The Hidden Token Tax
Tokenizers bill by their own arithmetic. Measuring the difference paid for itself many times over.