What does prominence mean?
A model with 2,000 citations and 50k monthly HuggingFace downloads from a top lab is clearly prominent. A single preprint with no external uptake is not, whatever its authors claim. The rubric below formalises this without pretending to be objective.
i.The rubric
Five signals, scored 0–2 each. Signals that don't apply (no public repo, no HuggingFace presence, no paper) are omitted, not penalised. Click a row for the rationale behind the threshold.
Bands
Two structural fixes sit on top:
- Clinical-deployment substitute. Clinical FMs often have no repo or HF presence. A documented hospital deployment counts as one signal worth 2 points.
- Recency carve-out. A model released in the last 12 months from a credentialed lab, with at least minimal cross-listing, gets a one-band bump. Counters the structural penalty new releases face on stars, citations, and downloads.
ii.The two gates before any of that
The rubric only runs on candidates that pass both:
- Must be a foundation model. Pretrained, general-purpose or multi-task substrate. Single-task classifiers are excluded.
- Must operate on HCLS data. DNA, RNA, protein, SMILES, whole-slide imaging, EHR text, medical imaging, surgical or endoscopy video, related modalities. General-purpose LLMs (GPT-4, Claude, Gemini) are excluded, even when used clinically.
iii.Flagship vs tracked
Every entry carries one of two tiers:
iv.What this map does not track
A few things people sometimes expect a "model map" to cover are out of scope.
- Detailed license terms. The captures whether weights are downloadable. I don't track MIT vs Apache vs commercial beyond that.
- Clinical validation or regulatory approval.
- Safety, bias, or fairness evaluations.
- Commercial availability or pricing.
- Whether I agree with the model's framing or stated use case.
v.The 10 sources behind cross-listing
A monthly scan checks 10 sources: 5 GitHub lists and 5 surveys from Nature and arXiv. They are enthusiast-curated and have their own blind spots; the rubric treats them as one input, not ground truth.