Pull requests, discussions and being useful

You can open a pull request on a model

Because a repository is git, anyone can propose a change to anyone’s model. On the Hub these live as refs like refs/pr/4, so a proposal can be downloaded and tested before it is merged:

model = AutoModel.from_pretrained("some-org/some-model", revision="refs/pr/4")

The contributions that are welcome and take under an hour:

• A missing licence or pipeline_tag. Without the tag the model is invisible to every filtered search, which the author usually does not realise.
• A safetensors conversion, so nobody has to unpickle a stranger’s file.
• A missing chat_template, which is the difference between the model working and not.
• A working code snippet in the card, tested.
• A translation of the card, which for Hindi, Bengali, Tamil or Marathi widens who can use the model.
• A named limitation you found, with the inputs that produced it.

That last one is the most valuable and the least common. A card that says “fails on inputs over 400 tokens, see discussion #12” is worth more than a benchmark table.

hf upload some-org/some-model ./README.md --create-pr

Discussions are where the real documentation lives

The Community tab on a model or dataset frequently contains the information the card does not: the actual prompt format, the version that broke, the workaround for a tokenizer bug, the licence clarification from the author. Read the discussions before concluding a model does not work. A large fraction of the problems in this course have a thread about them.

When you open one, make it answerable. A good report contains the model id with revision, library versions, the exact rendered prompt, the generation parameters, what you expected and what happened. Most unanswered threads on the Hub are missing the rendered prompt, and the rendered prompt is the answer surprisingly often.

The norms

The Hub is a public square with an unusually mixed population: researchers, hobbyists, students, companies. A few things hold:

• Assume the maintainer is unpaid. Most are. A model released for free carries no support obligation.
• Report security issues privately first. A malicious file or a leaked secret goes to the platform and the author, not to a public thread.
• Do not reupload somebody’s model under your own name. Fine-tune it, quantize it, convert it, and say what it came from, with a base_model field and a link. Republishing weights as your own is both a licence problem and the thing that damages the commons everybody here depends on.
• Credit the dataset as carefully as you credit the model.

Publishing is contributing

The most useful things a learner can publish are not state-of-the-art models:

• A quantized conversion of a model that has none, done carefully and documented.
• A Space that demonstrates a technique with a readable app.py. People learn from working code far more than from documentation.
• A dataset in a language or domain that is thin on the Hub. Indic-language datasets are scarce and valuable, and a few thousand well-documented rows is a real contribution.
• An evaluation of existing models on a task nobody has measured, with the set published.
• A card translation.

Each of these is achievable in a weekend and each is more useful than another fine-tune of a popular model on a public instruction set.

The reciprocity that makes this work

Everything in this course exists because people published things they could have kept. The libraries are free, the models are free, the hosting is free, the compute for small things is free. That arrangement is not guaranteed and it is not charity: it is sustained by people putting back more than they take out.

You do not owe anyone a model. Opening one pull request that adds a missing licence tag takes ten minutes and makes a repository usable by everyone who finds it afterwards. That is a reasonable rate of exchange.

Do this now

Find one model you have used whose card is missing something you know: a licence, a template, a limitation, a working snippet. Open a pull request. It is the fastest way to stop being only a consumer of this ecosystem.