Papers with Code

Papers With Code was a free, open-access website that organized machine learning research by linking papers to their source-code implementations and maintaining community-edited evaluation tables that tracked state-of-the-art results for tasks and datasets. Its core operating team was based in Meta AI Research, and all content and data were openly licensed under CC-BY-SA ^[3].

The service stopped operating in 2025. Its underlying dataset was frozen at a final public snapshot retrieved on July 28, 2025 and republished as static archives explicitly marked as never being updated, its GitHub organization ceased activity, and by September 3, 2025 the paperswithcode.com domain began redirecting visitors to Hugging Face's Trending Papers feed, which is where the domain still leads today ^[1][5][6][7].

For quantitative and alternative-data researchers, Papers With Code had been the de facto provenance layer connecting research claims to runnable code and benchmark numbers; after the shutdown it exists only as frozen archives, and paper-to-code discovery has migrated to platforms such as Hugging Face.

What It Is

The site described its mission as creating "a free and open resource with Machine Learning papers, code, datasets, methods and evaluation tables." Every paper page listed linked GitHub repositories, and every task carried evaluation tables where implementations were ranked by benchmark metrics, making the site a combined literature index, code registry, and leaderboard system ^[3].

Content covered computer vision, natural language processing, and other ML fields, supplemented by directories of methods and datasets, trend charts, and specialized portals for astronomy, physics, computer science, mathematics, and statistics ^[3]. The homepage surfaced trending research ranked by community stars-per-hour, mixing arXiv preprints with conference publications ^[4].

Organizationally it presented itself as "a community project" whose core team sat inside Meta AI Research, stating that no data was shared with any Meta Platforms product; day-to-day operations included hiring, a newsletter, and a public Slack where every edit was announced ^[3].

History

The site was online well before the generative-AI boom: the Internet Archive's earliest preserved capture of the paperswithcode.com homepage dates to June 22, 2018, and an About page existed by February 2019 ^[4][3]. Through 2019 to 2022 it became a standard reference for reproducibility, publishing its full database as downloadable JSON dumps and releasing tooling such as the sota-extractor pipeline and the GALACTICA model interface on its GitHub organization ^[2].

By 2022 the About page confirmed that the core team worked from Meta AI Research, while emphasizing community ownership and CC-BY-SA licensing identical to Wikipedia's ^[3]. Repositories under the organization show activity continuing until mid-2025, with the paperswithcode-data repository last updated on September 8, 2025 ^[7].

The shutdown followed in mid-2025 without the site itself remaining online to announce it: the final data snapshot was captured on July 28, 2025, the archive datasets were uploaded to Hugging Face with a caution notice that they would never be updated, and archived requests to the domain recorded an HTTP 301 redirect to huggingface.co/papers/trending on September 3, 2025 ^[5][6][7].

Listing Model

There was no commercial listing model: nothing was sold, and inclusion was earned rather than purchased. Entries originated from community edits, submitted through Edit buttons on paper and task pages, plus automated extraction pipelines whose logic was open-sourced ^[3].

An explicit inclusion policy governed benchmarks: a result could be listed once its paper was published as a preprint, conference paper, or journal article; accompanying code was strongly encouraged but not required, so tables captured results before implementations appeared. All contributions were licensed under CC BY-SA, and all edits were monitored through a public Slack review channel ^[3].

Distribution mirrored the openness: complete data dumps covering papers with abstracts, paper-to-code links, evaluation tables, methods, and datasets were downloadable by anyone. After the shutdown this distribution model persisted one last time, with the same five dumps re-hosted as static Hugging Face datasets under the pwc-archive organization, still CC-BY-SA but permanently frozen ^[2][5].

For Alt-Data Buyers

Buyers used the site primarily for verification: when a vendor or researcher claimed performance on a named benchmark, the corresponding evaluation table showed whether the result was reproducible, whether code existed, and how far the method trailed the actual state of the art. This made it cheap diligence infrastructure before contracting for models or data derived from academic work ^[3][4].

Quantitative finance was a visible use case on the platform: trending listings included reinforcement-learning libraries for automated stock trading, and finance-flavored papers regularly circulated through its newsletter and rankings, giving analysts a map from publication to implementation ^[4].

Post-shutdown workflows change shape. The frozen pwc-archive dumps allow offline lookups of historical paper-code-benchmark mappings, but nothing newer than July 2025 will appear there; ongoing discovery has shifted to Hugging Face's Trending and Daily Papers feeds, which list GitHub and arXiv links for each entry, and to community venues such as Kaggle ^[1][5]. Indexed remnants sit under Free & Open.

Limitations

Even at peak operation the model had structural limits: inclusion required only a preprint, so unreviewed results could appear in leaderboards; quality control consisted of volunteer review over a Slack channel; and coverage depended on contributors noticing relevant papers, which skewed toward popular domains ^[3].

The shutdown converts these weaknesses into permanent ones. The archived datasets carry an explicit caution that they will not be updated and represent only the July 28, 2025 snapshot, so every leaderboard position, repository link, and method entry risks staleness or rot, and deep links into paperswithcode.com no longer resolve to their original pages ^[5][6].

Continuity is partial at best: the redirect lands users on a general trending-papers feed rather than task-specific leaderboards, and no successor currently reproduces the combination of evaluation tables, methods taxonomy, and code linkage that made the original useful for systematic verification ^[1][2].

Landscape

Papers With Code is the one definitively dead hub among the major data and research platforms: unlike Kaggle, Hugging Face, AWS Data Exchange, or Snowflake Marketplace, it no longer operates, and its domain functions purely as a funnel into Hugging Face's papers product ^[1][6].

Its functional inheritance is visible in that destination: Hugging Face's paper pages now bundle the arXiv abstract, community discussion, and GitHub links that Papers With Code pioneered, though without open evaluation tables ^[1][5]. For the alternative-data industry the practical consequence is that verification of academic claims now requires either the frozen CC-BY-SA archives or manual search across GitHub and arXiv, with no single authoritative index filling the gap ^[2][5].