Stemma Machinarum

Stemma Machinarum

A genealogy of machine-learning models, starting with open-weight language models. It records which model descends from which, what changed at each step, and the evidence for every claim.

The name comes from textual criticism: a stemma is the family tree of manuscript copies, and stemmatics already has a term, contamination, for a copy that draws on two lines at once. Distillation and model merging are contamination in exactly that sense, so the structure here is a network, not a strict tree.

It is built for people and for LLM agents alike. More about the project.

2022-11 2022-12 2023-01 2023-02 2023-03 2023-04 2023-05 2023-06 2023-07 2023-08 2023-09 Closed · known by outputs Datasets Open weights alpaca-7b ← fine_tuned_from ← llama-7b (declared) llama-2-7b-chat ← fine_tuned_from ← llama-2-7b (declared) llama-2-7b ← successor_in_series ← llama-7b (declared) code-llama-7b ← fine_tuned_from ← llama-2-7b (declared) vicuna-7b-v1-3 ← fine_tuned_from ← llama-7b (declared) alpaca-52k ← distilled_from_outputs ← text-davinci-003 (declared) alpaca-7b ← trained_on ← alpaca-52k (declared) sharegpt-vicuna ← distilled_from_outputs ← chatgpt (declared) vicuna-7b-v1-3 ← trained_on ← sharegpt-vicuna (declared) codellama-7b-instruct ← fine_tuned_from ← code-llama-7b (declared) codellama-7b-python ← fine_tuned_from ← code-llama-7b (declared) llama-2-13b ← successor_in_series ← llama-7b (declared) llama-2-70b ← successor_in_series ← llama-7b (declared) nous-hermes-llama-2-7b ← fine_tuned_from ← llama-2-7b (declared_by_uploader) open-llama-7b ← design_follows ← llama-7b (declared) vicuna-7b-v1-5 ← fine_tuned_from ← llama-2-7b (declared) vicuna-7b-v1-5 ← trained_on ← sharegpt-vicuna (declared) vicuna-13b-v1-5 ← fine_tuned_from ← llama-2-13b (declared) vicuna-13b-v1-5 ← trained_on ← sharegpt-vicuna (declared) wizardlm-13b-v1-2 ← fine_tuned_from ← llama-2-13b (declared_by_uploader) text-davinci-003closedtext-davinci-003 ChatGPTclosedChatGPT (Nov. 2022 launch model) Alpaca 52K2023-03-13Alpaca instruction data (52K) Alpaca 7B2023-03-13Alpaca 7B Vicuna 7B v1.32023-06-22Vicuna 7B v1.3 Vicuna 7B v1.52023-08-01vicuna-7b-v1.5 LLaMA 7B2023-02-24LLaMA 7B Llama 2 7B2023-07-18Llama 2 7B Code Llama 7B2023-08-24Code Llama 7B … Instruct2023-08-24CodeLlama-7b-Instruct-hf … Python2023-08-24CodeLlama-7b-Python-hf Llama 2-Chat 7B2023-07-18Llama 2-Chat 7B Nous-Hermes 7B2023-07-25Nous-Hermes-llama-2-7b OpenLLaMA 7B2023-06-07open_llama_7b Llama 2 13B2023-07-18Llama-2-13b-hf WizardLM 13B2023-07-25WizardLM-13B-V1.2 Vicuna 13B v1.52023-08-01vicuna-13b-v1.5 Llama 2 70B2023-07-18Llama-2-70b-hf ShareGPTundatedShareGPT conversations (LMSYS collection) alpaca-52k ← distilled_from_outputs ← text-davinci-003 (declared) sharegpt-vicuna ← distilled_from_outputs ← chatgpt (declared) alpaca-7b ← trained_on ← alpaca-52k (declared) vicuna-7b-v1-3 ← trained_on ← sharegpt-vicuna (declared) vicuna-7b-v1-5 ← trained_on ← sharegpt-vicuna (declared) vicuna-13b-v1-5 ← trained_on ← sharegpt-vicuna (declared) llama-2-7b ← successor_in_series ← llama-7b (declared) llama-2-13b ← successor_in_series ← llama-7b (declared) llama-2-70b ← successor_in_series ← llama-7b (declared) open-llama-7b ← design_follows ← llama-7b (declared) alpaca-7b ← fine_tuned_from ← llama-7b (declared) llama-2-7b-chat ← fine_tuned_from ← llama-2-7b (declared) code-llama-7b ← fine_tuned_from ← llama-2-7b (declared) vicuna-7b-v1-3 ← fine_tuned_from ← llama-7b (declared) codellama-7b-instruct ← fine_tuned_from ← code-llama-7b (declared) codellama-7b-python ← fine_tuned_from ← code-llama-7b (declared) nous-hermes-llama-2-7b ← fine_tuned_from ← llama-2-7b (declared_by_uploader) vicuna-7b-v1-5 ← fine_tuned_from ← llama-2-7b (declared) vicuna-13b-v1-5 ← fine_tuned_from ← llama-2-13b (declared) wizardlm-13b-v1-2 ← fine_tuned_from ← llama-2-13b (declared_by_uploader) text-davinci-003 closedtext-davinci-003 Alpaca 52K 2023-03-13Alpaca instruction data (52K) ChatGPT closedChatGPT (Nov. 2022 launch model) ShareGPT undatedShareGPT conversations (LMSYS collection) LLaMA 7B2023-02-24LLaMA 7B Alpaca 7B2023-03-13Alpaca 7B OpenLLaMA 7B2023-06-07open_llama_7b Vicuna 7B v1.32023-06-22Vicuna 7B v1.3 Llama 2 7B2023-07-18Llama 2 7B Llama 2-Chat 7B2023-07-18Llama 2-Chat 7B Nous-Hermes 7B2023-07-25Nous-Hermes-llama-2-7b Vicuna 7B v1.52023-08-01vicuna-7b-v1.5 Code Llama 7B2023-08-24Code Llama 7B … Instruct2023-08-24CodeLlama-7b-Instruct-hf … Python2023-08-24CodeLlama-7b-Python-hf Llama 2 13B2023-07-18Llama-2-13b-hf WizardLM 13B2023-07-25WizardLM-13B-V1.2 Vicuna 13B v1.52023-08-01vicuna-13b-v1.5 Llama 2 70B2023-07-18Llama-2-70b-hf See the whole family →

The LLaMA family, drawn as a manuscript stemma. Dashed lines are contamination: text written by closed models (hollow) that became training data for open ones.

69models
8datasets
79edges

Last data change: 2026-09-24. Counts are computed from the data at build time.

The Graph

Every model and dataset, every lineage edge, each with an evidence tag and a source. Raw JSON is served as-is.

The stemma · Browse records · edges.jsonl

The Notebook

The reading path through the primary sources, with dated notes and hands-on exercises. Each station links to the records it produced.

Follow the path · 0 of 6 stations read · 0 notes

Evidence tags

Every edge carries one. Definitions are in the method.