hf_model_tree — Hugging Face model atlas (2025)

Description

The genealogy of the machine learning models on the Hugging Face Hub, as of March 11, 2025, from the "Model Atlas" of Horwitz, Kurer, Kahana, Amar and Hoshen. Each node is a model repository of the Hub (with weight files), and a directed edge from model u to model v means that v is a base model of u, i.e. that u was derived from v, e.g. by fine-tuning, as an adapter (e.g. LoRA), by quantization, or by merging several models. The network is a directed acyclic graph (mostly a forest of trees, since merged models have several base models).

The base models are those declared in the model cards (base_model), completed by the authors with base models extracted from the texts of the model cards with regular expressions, and with over 1,000 manual annotations; they removed loops and self-loops, and the models without weight files. Base models that were removed from the Hub but are still referenced by other models are included as "virtual" models (relation virtual_recovered, without attributes). The data derive from the hub-stats dataset of Hugging Face. Thirteen models appear twice in the file, with different (annotated) base models; their rows are merged.

The network has the following properties:

(Node) model_id: the id of the model repository (<author>/<name>; its page is https://huggingface.co/<model_id>).
(Node) author.
(Node) created: the time of creation (Unix time; -1 if unknown).
(Node) last_modified: the time of the last modification (Unix time; -1 if unknown).
(Node) downloads: the number of downloads of the last 30 days, and downloads_all_time (-1 if unknown).
(Node) pipeline_tag: the task of the model (e.g. text-generation).
(Node) license.
(Node) architectures: the model class (e.g. LlamaForCausalLM).
(Node) dataset: a training dataset given in the tags (unknown if none).
(Node) tags: the tags of the repository (as a Python list).
(Node) base_model_relation: the relation of the model to its base models (finetune, adapter, quantized, merge, finetune_adapter, finetune_quantized; source for some original models; unknown if not declared; unknown_readme_extracted and unknown_annotated for base models extracted from the model card or annotated by the authors; virtual_recovered for virtual models; two values separated by a comma for the merged duplicates). Missing values are unknown, as in the data.
(Edge) relation: the relation of the model to its base model (as base_model_relation of the model).
(Edge) base_index: the position of the base model in the list of base models of the model (from 0).

Tags

Informational Technological Software Online Unweighted Metadata

Added

2026-10-31

Citation

E. Horwitz, N. Kurer, J. Kahana, L. Amar, Y. Hoshen, "Charting and Navigating Hugging Face's Model Atlas." arXiv:2503.10633 (2025) arXiv:2503.10633

Upstream URL OK

https://huggingface.co/datasets/Eliahu/ModelAtlasData
https://horwitz.ai/model-atlas

Networks

Tip: hover over a column header for a description.

Column legend

$\left<k\right>$: Mean degree
$\sigma_k$: Standard deviation of the degrees
$\lambda_h$: Largest eigenvalue of the Hashimoto (non-backtracking) matrix
$\tau$: Random walk mixing time
$r$: Degree assortativity
$c$: Global clustering coefficient
$\oslash$: (Pseudo-) diameter
$S$: Relative size of the largest component
NPs: Node properties
EPs: Edge properties

Name	Nodes	Edges	$\left<k\right>$	$\sigma_k$	$\lambda_h$	$\tau$	$r$	$c$	$\oslash$	$S$	Kind	Mode	NPs	EPs	gt	GraphML	GML	csv
hf_model_tree	1,167,895	485,834	0.42	54.74	21.18	48861.63	-0.16	0.00	37	0.12	Directed	Unipartite	model_id author pipeline_tag license architectures dataset tags created last_modified downloads downloads_all_time base_model_relation	relation base_index	28.3 MiB	36.9 MiB	35.1 MiB	45.9 MiB