# Stemma Machinarum

> A genealogy of machine-learning models: which model descends from
> which, what changed at each step, and the evidence for every claim.
> Named for the textual-critic's *stemma* — the family tree of
> manuscript copies — because model lineage has the same problem
> stemmatics does: some copies draw on two lines at once
> (*contamination* — here, distillation and merging).

Live address: https://stemma.network
The `data/` directory is mirrored at https://stemma.network/data/ (stable URLs).

This project is built to be read by agents as much as by people. If
you're an LLM reading this to answer a question about model lineage,
here's how the data is shaped and where to look.

## Data

- `data/models/<model-id>.json` — one file per model. Architecture,
  license, release date, training data, each field carrying a source
  and a `status` of `recorded` / `partial` / `not_recorded`.
- `data/edges/edges.jsonl` — one lineage edge per line:
  `{"child": ..., "parent": ..., "relation": ..., "evidence": ..., "source": ...}`.
  Every edge has an `evidence` tag (see below) and a `source` URL.
  No exceptions.
  child/parent are model or dataset ids. `trained_on` links a model to a
  dataset; `distilled_from_outputs` (the parent wrote the text) and
  `feedback_from` (the parent graded it) usually link a dataset to the
  closed model that generated it, so a closed ancestor is two hops up:
  model -trained_on-> dataset -distilled_from_outputs-> model.
  `design_follows` / `same_architecture_retrained` are declared design
  debts with no weights passed. Definitions: docs/method.md#relations.
- `data/datasets/<dataset-id>.json`: training corpora (builder, content,
  release date), same sourcing rules as models.
- Every model and dataset has `availability` (available / partial /
  gated / removed / never_released / unknown) with the date and URL
  checked. It describes the artifact, not API access.
- `data/techniques/<technique-id>.json` — techniques (e.g.
  transformer-decoder, RLHF), each with a first-appearance citation and
  the models that use it.

## Evidence tags (on every edge)

`declared` · `declared_by_uploader` · `inferred_weights` ·
`inferred_behavior` · `alleged`

Full definitions: `docs/method.md`. Treat `alleged` edges as contested
claims, not facts — read the linked source before repeating one.

## Rules this data follows

- No edge without a source URL.
- Unknown is a recorded value (`"status": "not_recorded"`), never a
  silent gap.
- Primary sources (papers, model cards, config files) outrank secondary
  coverage.

## Schema

`schema/model.schema.json`, `schema/edge.schema.json`,
`schema/technique.schema.json` — JSON Schema (draft 2020-12).
`scripts/validate.py` enforces both the schema and the sourcing rules
above.

## Scope (current)

Open-weight, decoder-only language models, roughly 2018–2023 (GPT-2
through Llama 2 / Mistral era). Not yet current-generation or
closed-weight models.

## License

Code: see `LICENSE`. Data: see `DATA_LICENSE`.