One fork for every 6.9 stars — and what the ratio tells you

Every repository page shows two counters side by side, and almost everyone reads only the first. A star costs one click, applies retroactively to work you have never run, and never expires. A fork is a copy: someone decided the code was worth having under their own name. The two numbers measure different acts, so the ratio between them measures something neither does alone.

Across the whole catalogue — 10,840 projects, 58,331,228 stars — there are 8,427,454 forks. One fork for every 6.9 stars. What surprised us is how little that ratio moves.

The ratio barely moves

Restricting the count to the 2,001 projects above 5,000 stars, where both numbers are large enough to be meaningful, the median is 0.115 forks per star. The distribution is tight.

That stability is what makes the ratio useful. A measure that clusters this tightly turns its own outliers into signal.

Near 1.0: repositories you are supposed to copy

At the top of the range, the ratio approaches one fork per star — meaning very nearly everyone who liked the project also took a copy. Look at what those projects are.

Two categories, one behaviour. The courses — a coding-agent curriculum, an LLM engineering workshop, a Udacity exercise set — exist to be forked: the fork is the product, because you work through the notebooks in your own copy. The autopilot projects are forked because flying hardware means a build configured for your airframe, which starts with a copy of the tree.

In both cases a high ratio is healthy. It is also a warning about reading those star counts as adoption of software: nobody depends on a course the way they depend on a library.

By language, the ratio tracks how the software is delivered

Group the same cohort by primary language and the pattern is about packaging, not popularity.

The two ends of that table are the two ends of how software reaches you. A notebook is cloned and edited; C++ inference code is cloned and compiled against your own CUDA version. A Rust or Go project ships a binary you download, or a crate you declare as a dependency — you never need a copy of the tree, so the fork count stays low while the star count does not. Rust sits 2.2× below C++ on this measure, and that is a statement about distribution rather than quality.

The practical consequence: never compare fork ratios across languages. Compare a project to its own neighbourhood.

Below 0.02, there are eight

The low end is the interesting one, and it is nearly empty. Eight projects above 5,000 stars — 0.4 % of the cohort, 75,546 stars between them — have fewer than one fork per fifty stars. As a group they are young: median age 240 days, five of the eight created in 2026. They are also thinly staffed where we have a reading, with a median of 7 contributors against 49 for the cohort as a whole.

There are honest explanations for every one of those traits. A wallet, an SDK or an agent runtime is something you install, not something you fork. A repository published four months ago has had four months to collect forks and the same four months to collect stars, and stars arrive faster. And a small team is not a defect.

So this is not a fraud detector, and we are not presenting it as one. It is a cheap question: why does a project with tens of thousands of admirers have almost no copies? Sometimes the answer is “because you install it”. Occasionally the answer is that the star count grew for reasons unrelated to the code. Either way, eight candidates out of 10,840 projects is a short list to check by hand.

Method

Figures cover the 10,840 projects in the olud.ai catalogue, snapshot of 24 September 2026: 58,331,228 stars and 8,427,454 forks in total. “Forks per star” is the fork count divided by the star count, both as reported by each project’s forge; neither is compared across forges. Band and language figures are restricted to the 2,001 projects above 5,000 stars so that small denominators cannot distort a ratio — at 200 stars, a dozen forks moves it by 0.06. Languages are shown where the cohort holds at least 40 projects; the primary language is the one the forge reports, and 166 projects in the cohort report none. Contributor counts, where quoted, come from our health scores and were checked between 20 and 22 August 2026.

Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.

Keep reading

• 459 repos hold half the stars — how steep the popularity curve really is.
• 127 repos above 10,000 stars are held up by fewer than ten people — the other number worth reading next to the stars.
• One in five AI projects has no licence a machine can read — the check that comes before either of them.