Ashsinha1/hf-spaces-clones

Unmodified copies among Hugging Face Spaces

104,791 public Spaces that are copies of another Space, grouped into the 25,183 source histories they came from. Derived from a full census of the 1,458,692 public Spaces taken in August 2026.

This is not a similarity score. Members of a family are byte-identical repositories, and the check below is what establishes that.

How a copy is identified

A Space is a git repository. Pushing an existing history into a new repository carries the original commit dates with it, so the copy ends up with a lastModified that predates its own createdAt. That ordering cannot occur in a Space whose files were authored in place, which makes it usable as a marker.

Spaces carrying the marker are grouped by that inherited timestamp. One group is one source history.

Validation

The marker was checked against HEAD commit SHAs, which hash the entire tree and history, by re-querying the live API:

• 60 of 60 sampled family members shared their family's HEAD commit.
• 12 randomly chosen Spaces had 12 distinct HEAD commits.

Two unrelated accounts holding VoiceFixer were confirmed to share all 15 commit SHAs, not merely the same file names.

What this is not

It is a lower bound. A copy taken from a source that was edited after the copy was made inherits a newer timestamp, shows no inversion, and is invisible here. The true number of copies is higher than 104,791.

It is not the "Duplicate this Space" button. That was the first hypothesis and it failed: cardData.duplicated_from is set on 3.3% of these Spaces against 1.7% of a random control, which is no real difference. These are git-history copies, which that field does not track.

The population is not representative of Spaces at large. It is dominated by course assignments and one-click deploy templates. The largest family is 9,490 copies of the Hugging Face Agents Course practice project; 13,776 copies exist across that template's versions.

stage is a snapshot from August 2026 and drifts. SLEEPING means the Space wakes on request, not that it is broken.

Columns

column	meaning
id	Space identifier, author/name
author	account that created the copy
sdk	gradio, docker, streamlit (static Spaces are excluded: no runtime)
stage	runtime stage at census time
created_at	when this copy was created
inherited_last_modified	timestamp carried from the source; the family key
family_id	0-based, ordered by family size descending
family_size	number of copies sharing this source history
family_common_name	most frequent base name within the family

One thing it was used for

Deduplicating these copies moves the share of non-static Spaces in an error state from 38.57% to 38.90%, a change of a third of a percentage point. The copies were not inflating that figure.

Holding code constant is also the one comparison a cross-sectional census cannot make. Within 552 families of 20+ copies, the older half is not more broken than the newer half (sign test z = -2.1), even though breakage rises sharply with age across the catalogue as a whole and inside every SDK.

Reproducing

python clone_detect.py --data spaces.csv
python clone_experiment.py --data spaces.csv

Code: https://github.com/ashishsinha1602/dataset-integrity-audit

Source census: https://huggingface.co/datasets/Ashsinha1/hf-spaces-census

Citation

@misc{sinha2026spacesclones,
  author = {Sinha, Ashish},
  title  = {Unmodified copies among Hugging Face Spaces},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Ashsinha1/hf-spaces-clones}
}