Frequently asked questions

Models are curated by AI agents that research candidates from primary sources — papers on arXiv and bioRxiv, peer-reviewed venues, and official code on GitHub and Hugging Face. To be included, a model must be a genuine biological foundation model that is in scope for the catalog: applying modern machine learning to a core problem in biology such as protein structure or design, genomics, RNA, single-cell analysis, pathology, or biomedical imaging. Before anything is added, an agent verifies the model is real and credibly sourced — backed by an identifiable paper and a public implementation or weights, not vaporware — and confirms it is not already listed. Inclusion reflects scientific credibility and relevance, not popularity or paid placement.

The leaderboard sorts models by transparent, objective metrics — citation count, GitHub stars and forks, and Hugging Face downloads and likes — refreshed daily from public APIs. These are adoption signals, not a judgment of scientific quality; the editorial point of view lives in the model write-ups.

The number on the badge is the bio.rodeo openness score: 0–100 for how much of the model is actually released — weights, code, training data, documentation, and the licenses on each — rather than whether it is labeled “open.” 70 and above reads as Open, 40–69 Partial, below 40 Closed; the colour is just the band, and the exact number is always shown. The inputs are the 17 release components defined by the Model Openness Framework (MOF). A greyed-out badge means we have no openness assessment for that model yet.

Openness is a gradient, not a pass/fail: a model can ship open weights and inference code and still keep its training data closed. So the bio.rodeo openness score rates three continuous 0–100 axes from the 17 release components defined by the Model Openness Framework:

• Usability — can I run it? Weighted toward weights, inference code, and architecture.
• Reproducibility — can I retrain it from scratch? Weighted toward training code and datasets.
• Overall — the headline closed→open number; its own blend of all 17 components, not the average of the other two.

Each component earns credit by status — open (full), described (partial), restricted (little), missing (none) — then that credit is scaled down for non-commercial or no-derivatives licenses. Because each axis divides by its full weight set, a missing artifact drags the score down rather than failing the model outright. Overall ≥ 70 reads as Open, 40–69 Partial, and below 40 Closed — the band is just color; the exact number is always kept. Models whose usability far outruns their reproducibility are flagged “open weights, closed recipe.”

Yes — two ways, no key or signup. The simplest is the shields.io endpoint badge, which needs no query parameters:Swap openness for openness-usability or openness-reproducibility for the other two axes, and basenji2 for the last path segment of any model’s page URL (e.g. /models/basenji2).Under the hood the same numbers are served as plain JSON, if you’d rather pick a specific field or restyle the badge yourself:GET https://bio.rodeo/api/models/{slug}/opennessIt returns usability, reproducibility and overall as 0–100 integers, plus the band and evaluated_on date; unknown or unevaluated models return 404. Point the dynamic JSON badge at any field:Both endpoints are CORS-open and cached for 24 hours, so either is safe to hotlink from a README.

Citation metadata is enriched from Semantic Scholar and CrossRef using each paper's DOI. Repository and download metrics come from the GitHub and Hugging Face APIs. Everything is refreshed on a daily schedule so rankings stay current without manual updates.

Yes. Use the “Submit a model” button on the Models page to suggest something new. To report a problem with an existing entry, use the “Flag content” button on that model's page — every model page has one.

bio.rodeo is built by a single developer, Justin Kiggins, using Claude Code — with a set of specialized skills and custom AI agents that handle brand consistency, engineering quality, and content curation.

No — the prose is researched and written by AI agents from primary sources, with cited, fact-checked details rather than copied marketing text. Quantitative metrics are pulled automatically from public APIs. A human designs and oversees the system; the writing is not done by hand.

bio.rodeo is built and maintained by Justin Kiggins as a project of Pulsatance. There is no paywall and no advertising. Learn more on the About page.