Hugging Face: a public model warehouse teachers can verify

On a Saturday morning the IT group forwards a sales message: "We use open source models, so your data never goes overseas." The panel head replies: "Fine — send me the model name. I want to read its card online first." Five minutes later the thread goes quiet, because the vendor could name a brand but not a link.

Hugging Face is the first stop in module 3. Once the conversation shifts from cloud subscription to open weights and on-premises deployment, you need an entrance you can verify yourself instead of trusting a slide deck.

---

A cabinet of labelled specimens

Picture a natural history gallery. Each specimen sits beside a label: collection date, origin, preservation method, and a "do not touch" warning. You need not take the specimen home. The label alone tells you whether it belongs in your classroom demo.

Hugging Face is that kind of cabinet for AI models. It is currently the main public place where models are stored and shared, with over two million models publicly available. Engineers call it "the GitHub of AI" because it offers version history, branches, downloads and collaboration. For a teacher the point is simpler: when somebody says a model is good, you can open its page instead of relying on what you were told.

Three shelves to tell apart

It looks chaotic at first because three categories live side by side:

• Models: the actual model files plus a documentation page. This is the main part.
• Datasets: material used for training or evaluation. What matters most for schools is realising what kind of world a model learned from, which helps explain where bias comes from.
• Spaces: small web apps others have built, letting you try a model in the browser with nothing to install. This is the easiest area for teachers and the best one for live demonstrations at a staff workshop.

If you only have ten minutes, start with Spaces. Find a text or speech demo and try it yourself. That teaches more than ten articles.

Four fields to check before you buy

Open any model page and read in this order. No coding required:

1. The licence. This is the one item you cannot skip. Some models are free for commercial use, some are restricted to research, some cap the size of organisation that may use them. Using a model in class, or letting pupils reach it through the school website, needs the licence checked first.
2. The size. The page shows the parameter count, for example 8B (eight billion). That decides what hardware a school needs to run it. See the rise of small language models.
3. Whether anyone maintains it. Look at the last update date and the discussion tab. A model untouched for two years with unanswered questions does not belong in a school workflow.
4. How thorough the documentation is. That page is called the model card, and it is the subject of the next lesson, reading a model card and what GGUF and Safetensors mean.

Three layers of "open"

The most common misunderstanding is treating "open source" as a single state. There are at least three layers:

• Open weights: you can download the model files and run them, but the training data and training code are not published. Most models described as open source fall into this category, including the main models from Meta and Alibaba.
• Open with conditions: downloadable, but the licence adds restrictions, such as forbidding certain uses or requiring larger organisations to apply separately.
• Fully open: weights, data, code and training records all published. There are few of these, and they come mostly from research institutions.

When writing a funding application or a school policy, it is more accurate to write "open weight model" than "open source model".

Leaderboards report exam scores; a model card tells you whether the pupil fits your class. School selection should start with the second.

A weekend trial for General Studies

A primary school's General Studies panel wanted AI help summarising supplementary reading for Primary 5 on renewable energy, but the principal insisted any trial needed verifiable documents. The subject teacher did three things.

First, she filtered Hugging Face for models that support Chinese, sit under 8B parameters, and carry a licence permitting institutional use, then shortlisted three.

Second, she used Spaces demos to paste three old reading passages with names removed, comparing whether each model's summary stayed faithful or invented facts beyond the textbook. No installation was required, and the whole step fit into one Saturday morning.

Third, she printed the chosen model's page together with its licence terms and submitted both to the principal as the trial approval appendix. Three months later that model ran on a school server through Ollama, and the trial material never left the campus network. The whole chain started with someone willing to open the site and read the labels first.

Common misreadings

• The claim: The top leaderboard entry must be the right model for your school. What actually happens: Benchmarks test general tasks, not Primary 5 Chinese feedback or the tone of your school circulars. Self-check: have you run ten real pieces of your own material through it?
• The claim: A Spaces demo is the same as a formal deployment. What actually happens: Third parties run the demo and may log what you paste. Self-check: did you use only de-identified or invented content? See defending against jailbreaks and prompt injection.
• The claim: "Open source" means zero cost. What actually happens: Weights may be free; running them is not. Self-check: have you counted hardware, maintenance and IT time?

Try it once

Time: about ten minutes. Materials: a browser with access to Hugging Face.

Task: find any text-generation Space, or use two different models in tools you already have, with the same prompt:

You are a Hong Kong primary Chinese teacher. Below is a Primary 4 pupil's diary entry. Give feedback in three sentences: one praising something specific that works, one naming a single thing to improve, one suggesting what to practise next.
Do not rewrite the entry. Where you are unsure, write "teacher to verify".
[paste a short de-identified sample]

Hand in: three lines of notes: (1) did the two models differ noticeably in tone or rigour; (2) did either use non-Hong-Kong wording or rewrite the whole entry; (3) if a vendor claims a particular open model, can you find its page on Hugging Face and read the licence.

Done right: you can explain why the same input and instruction produce different answers (training and alignment differ, see alignment with RLHF and DPO), and you confirmed the trial used no real pupil data.

Continue with reading a model card and what GGUF and Safetensors mean. Previous lesson: knowledge cutoffs and model versioning. To compare the current line-ups of the major providers, see the model database.