Submitting an adapter

Community adapters are welcome, including several for the same task. Build your data set to the data guidelines first: adapters trained on v1.2 will need retraining on v1.3, and a data set that meets the guidelines carries over unchanged. Each community adapter is labelled so readers can judge it: who made it, how far its results were checked, and where its data came from.

Have an idea but no time to build it? Suggest it instead.

What to include

Required:

• The weights on your own Hugging Face repository, in the standard PEFT format, with the Jeff base as base_model. JeffHub never hosts weights.
• The manifest, adapters/<name>/manifest.yaml (below).
• The test set, or a link to it, so maintainers can reproduce your results.
• The shortcut report for your training data.
• A data card: every data source with its licence, which model generated each part of the data (if any), and how the test set was held out.

Encouraged: the training data itself, when its licence allows, and the code that built it.

1. Write the manifest

Fork firelex/jeffhub and add adapters/<name>/manifest.yaml. The folder name is the adapter's name. Copy an official manifest (for example adapters/support-intents/manifest.yaml), change every field, and set kind: community. Put the test set and shortcut report in the same folder, or link to them.

The manifest holds:

• What it is: name, title, a one-line summary, the task category, when to use it and when not to.
• The request format: the state fields in order (the changing one last), the questions, their exact instructions, and one example request.
• Results: for each test set, the untrained base, Jeff without the adapter and your adapter, each with the checkpoint it came from. Use null for anything not yet measured; never a guess.
• Where the data came from: a one-line provenance for the catalogue card, and under licence.data every source with its licence and generated_by (the model that wrote it, or null).
• The submission and data card:

provenance: "Generated by Qwen3.8-Flash-Next, plus the public HWU64 data set"
review:
  level: report-attached     # see "Checking levels" below
  reproduced: null
submission:
  test_set: test.jsonl                  # a file in this folder, or a web address
  shortcut_report: shortcut-report.md   # the shortcut report's output
  training_data: null                   # a web address, if you share it
  code: https://github.com/you/your-adapter
data_card:
  test_holdout: Messages from 10% of the generated companies, none of which appear in training.

2. Check it locally

pnpm install
pnpm check

The check fails loudly and names every problem: a name that clashes with the base model, the changing state field not last, number-like option keys, a result without its checkpoint, a data source without a licence or without generated_by, weights not yet on Hugging Face, or a test set or shortcut report that is not in the folder.

Checking levels

Every adapter shows one of three levels. They describe how far its numbers were checked, not how good it is.

• Self-reported: the numbers are the maintainers' own, and nothing else is attached.
• Report attached: the shortcut report and data card are included and pass pnpm check.
• Reproduced: a JeffHub maintainer re-ran your test set and got your numbers within a stated tolerance. The manifest then records who, when and the tolerance.

Readers can filter by level. Within the catalogue, reproduced adapters come first, then those with a report attached, then self-reported ones; results never decide the order.

3. Open the pull request

Every pull request is checked automatically. Before merging, a maintainer confirms that:

• the adapter loads on the base model and version it names;
• the example request works and gives the example response;
• the licences of the adapter and its data are clear and allow the use described.

A maintainer may then reproduce your results and raise the level to "reproduced".

After it is listed

Each adapter page shows its Hugging Face likes and downloads, read when the site is built, and a comment thread. Comments are GitHub Discussions in the registry repository, so readers sign in with GitHub and JeffHub stores no accounts. There are no star ratings.