Papers with Code,
retired.

On July 24–25, 2025, Meta sunsetted Papers with Code without prior notice. Users reported “Bad Gateway 502” errors and garbled text; GitHub issues went unanswered. Within days, the domain began redirecting to Hugging Face's Trending Papers feature, and the 9,327 benchmark leaderboards, 79,817 paper-to-code linkages and 5,628 datasets that had been tracked were no longer served from the canonical URL.

The site had been acquired by Meta (then Facebook AI) in December 2019 from its founders Robert Stojnic and Ross Taylor, under a public commitment that it would “remain a neutral, open and free resource.” That commitment held for five and a half years.

Historical data was rescued by the community — as JSON dumps on GitHub, as the pwc-archive on Hugging Face, and as 2021-era imports into the Open Research Knowledge Graph. The data survives; the integrated experience does not.

A small piece of infrastructure, used daily.

Papers with Code was, in effect, the Wikipedia of machine-learning research: a wiki of papers linked to implementations, benchmarks organised as <Task, Dataset, Metric> tuples, and leaderboards that showed, for each tuple, who currently held the top score and with what code.

Before PWC, establishing a baseline meant searching GitHub with keyword guesses, hoping the authors had released code, and debugging frameworks that had since been deprecated. After PWC, a researcher could read the abstract, click “Code,” and start. That small convenience raised the rate at which results were reproduced — and papers with linked code were cited, by community estimates, noticeably more often.

For engineers, it collapsed the “find a reference implementation” step of a project from days to minutes. For decision-makers, it made it possible to distinguish actual state-of-the-art scores from press-release claims. For the field, it made reproducibility an expectation rather than an aspiration.

None of the successors Meta pointed at — Hugging Face Trending Papers, Semantic Scholar, Connected Papers — replicate the unified paper-code-benchmark-dataset-method linkage. The data survived; the rhythm didn't.

Three ordinary ways Codesota is different.

We inherit PWC's best habits; we correct its worst. The difference is small and, we hope, boring — which is what you want from a registry.

First, every benchmark here is traceable. Reproducible code where the model is open and the eval is deterministic — verified by a signed container hash on rerun. Where it isn’t — closed APIs whose behaviour drifts behind the same name, sampled outputs that vary across seeds — we publish the dated source link, the prompt template, and the vendor model-card snapshot, and we say so. What never publishes is a number with no source. That single rule eliminates most of the drift that made late-stage PWC tables unreliable.

Second, every score is dated — publication date, verification date, and the date of the model checkpoint it refers to. Regressions stay visible. The table never silently forgets.

Third, the registry is routable. Every benchmark row corresponds to an API endpoint; we are slowly building the “router” that lets a caller send a prompt or a document to the current SOTA without rewriting their code when the SOTA changes. See the roadmap page for what that looks like at scale.

Everything else — the wiki-style task pages, the dataset cards, the method explanations — we are rebuilding gradually, in the open, one modality at a time.