About

Stemma Machinarum is a genealogy of machine-learning models. It records which model descends from which, what changed at each step, and the evidence for every claim. It starts with open-weight language models from roughly 2018 to 2023, the era from GPT-2 to Llama 2 and Mistral, because that era is well documented and its lineages are simple enough to test the method on.

Why a stemma

The name comes from textual criticism. When a text survives in many manuscript copies, editors work out which copy was made from which and draw the result as a family tree, the stemma codicum. Some ancestors are lost and are known only through their descendants. Some copies were made from two exemplars at once, which editors call contamination.

Language models have the same shape. A fine-tuned model is a copy of its parent’s weights, altered. A closed model is a lost exemplar: its weights were never published, but its outputs became other models’ training data. And a model trained on another model’s outputs, while starting from a third model’s weights, draws on two lines at once. Distillation and merging are contamination in exactly the stemmatic sense, so the structure here is a network, not a strict tree. The stemma shows one family drawn this way.

Genealogy and taxonomy

There are two different questions to ask about any model, and this project keeps them apart.

• Genealogy asks what a model descended from: whose weights it started from, what data it learned from, whose outputs shaped that data. That is a matter of history, and it needs a source: a paper, a model card, a release post. In Stemma, genealogy is the edges.
• Taxonomy asks what a model is: its structural characters, such as how many layers, which kind of attention, which positional encoding. That is a matter of classification. Franz Reuleaux did this for machines in 1875, sorting them by the kinematic pairs they were built from rather than by what they were for. In Stemma, taxonomy is the architecture fields and techniques on each record.

Neither is derived from the other. Two models can look nearly identical and share no recorded descent: Mistral 7B uses the same recipe as LLaMA, but its developers never say it descends from any particular Llama, so no edge is recorded. Whether structural resemblance tracks actual descent is a question the data should be able to answer, and it can only answer it if the two are recorded separately.

Evidence

Every edge carries an evidence tag saying who makes the claim: the developer, the person who uploaded the model, an inference from weights or behaviour, or a third party whose claim is disputed. Every edge and every non-trivial field has a source. Where something is not known, the record says so; it is never filled with a plausible guess. The rules are in the method.

Two audiences, on purpose

Stemma is built for two kinds of reader, and neither is an afterthought.

• People: developers, policymakers, advocates, and historians who need a factual account of where AI systems came from, without hype or doom. For them there are the record pages, the stemma, and the notebook.
• Language-model agents, which increasingly answer questions about AI and need a dated, machine-readable record in which fact and inference are kept apart. For them the same data is published as plain JSON at stable addresses, with a schema, and described in llms.txt. All crawlers are welcome, and the data is licensed CC BY 4.0.

Both readers get the same thing: every page is built from the same files, and nothing on a page says more than its record does.

The notebook

The project is built by Wilson Pruitt, who is learning the field from its primary sources as he goes. The notebook is that reading path, from Rosenblatt’s perceptron (1958) to the LLaMA paper (2023), with his dated notes. Each source read should add or verify records in the data, and each station links to the records it touched. Anyone can take the same route.