Methodology: how a model page is made
Every /models/ page on this site is assembled from exactly two kinds of content.
This page says what each kind is, where it comes from, and what we do not do.
The About page covers who operates the site and the general content-production statement;
this page is specific to the model pages.
1. Facts copied from the model card
The section titled "From the model card" is a definition list built from the fields our pipeline
fetches from the public HuggingFace Hub API for each repository: pipeline_tag, the repository's
tags, and the downloads and likes counters. The rows are derived mechanically:
- Publisher — the namespace part of the repository id (the org or user that owns the repo). We do not know or state whether it is a company, a lab, or an individual.
- Pipeline tag, library, framework and weight-format tags — matched against fixed allowlists of HuggingFace tag names (for example
transformers,safetensors,gguf). A tag that is not on the allowlist is simply not shown as a fact; the raw tag list is still printed at the bottom of the page. - License tag — the value of the
license:tag as set by the uploader. It is a tag, not the license text; the license file in the repository is authoritative. - Lineage — the
base_model:tags, including the relation HuggingFace records (fine-tune, adapter, merge, quantized). - Language tags — two- and three-letter tags that are valid language codes, expanded to their English names. Library shorthands such as
tforjaxare excluded by validation, not by hand. - Papers cited and datasets declared — the
arxiv:anddataset:tags, linked to arXiv and to the HuggingFace dataset page. - Downloads and likes — HuggingFace's counters at the time of our last fetch.
A row is omitted when the card has no such tag. We do not fill gaps with guesses: this block has no row for a parameter count, a model size, a context length or a benchmark score, because none of those are in the data we fetch. (Such a number can still appear inside the generated text described in section 2, where it comes from the generation step rather than from a fetched field.) Tags are set by whoever uploaded the repository and can be wrong; we do not verify them.
2. Generated text: summary, use cases, pros, cons
The one-paragraph summary under the title and the Use cases, Pros and
Cons lists are stored per model in our dataset. They were produced by an automated pipeline
from the model card and the metadata above — either by a scheduled Claude Code routine run under the operator's
account, or, for entries the routine has not reached, by a deterministic template that expands the metadata.
Every row carries a label naming its generator, and since September 2026 only the two labels that mean the text was
written for that entry — claude-routine and human-edited — get a page at all. Rows produced
by a template (metadata-derived, polish-py-template, fallback-template) stay in
the dataset without a page, so their /models/… URLs return 404 instead of showing template prose. The
labels were re-derived by matching each row's stored text against the deterministic generators, which makes
claude-routine the residual class: text that matches none of the templates. The
About page carries the current counts.
This text is not individually reviewed. We do not benchmark, download or run these models ourselves, and nothing on a model page is a first-hand test result. Numbers inside the generated text (a parameter count, an accuracy figure, a speed-up) come from the generation step, not from the fact rows, and are not verified by us — our September 2026 audit found entries where digits in the repository name had been misread as a parameter count. If you need such a number, confirm it against the repository before relying on it.
3. What we removed, and why
Until September 2026 each model page also carried sections titled "When does it fit?", "Real-world usage signals", "How we look at … models" and a "Frequently asked questions" block, plus FAQ structured data. That text was produced by the page template from the pipeline tag and from download and like thresholds, so the same paragraphs appeared on thousands of pages with only the model name swapped in, and it made claims that were not true of individual models (for example advising a "fine-tuned classifier head" on zero-shot classification models, or asserting engagement percentiles we had never computed). We removed all of it. A model page now carries only the two kinds of content described above, plus the two monetisation boxes described in section 6 — a set of Amazon search links, and a sidebar of GPU-provider referral links — each of which is rendered only when the corresponding affiliate or referral code is configured for the site. Neither code is configured in the build this page ships with, so neither box appears on any page today.
4. What gets a page
The dataset behind this site is a fetch of the most-downloaded public models on HuggingFace. CI fixture repositories
(internal-testing namespaces, tiny-random and dummy repos) are excluded from publishing.
In the build this page ships with, the export holds 1,859 fetched model rows and
240 of them are built as pages; the other 1,619 are kept as data only, and
their /models/… URLs return 404. A row is built only when its text was written for that entry
(section 2), it is not a CI fixture, and there is evidence someone is looking for the page — the full rule, and
the count behind each part of it, is on the About page.
Rows are never deleted from the dataset when they stop being published; the publishing rule is applied at build time.
5. Freshness
New-model ingestion for this site is currently switched off, so the download and like counters shown are those of the last fetch and are not live. A model page does not currently show the fetch date; every page links the HuggingFace repository, which always has the current counters, files and README. Pages are rebuilt automatically when the site is deployed, and a rebuild is not a review.
6. Related models and sort orders
"Related models" lists the entries that share the same pipeline tag and have a page on this site (section 4), ordered by the HuggingFace download counter; an entry that is kept as data only never appears there. Tag pages use the same ordering. No paid placement affects any list. Money never enters an ordering: the only monetised elements are a sidebar box of GPU-provider referral links and a box of Amazon search links, both labelled as such, both rendered only when a referral or affiliate code is configured for the site. No such code is configured in the build this page ships with, so neither box appears on any page today.
7. Corrections
If a fact row disagrees with the model card, or generated text is wrong, use the contact page. Verified errors are fixed in the dataset, so they stay fixed on the next rebuild.