Contributing
This knowledge base is the shared place to learn about Buddhist tasks: what each task is, how it is measured, the workflows it comes from, and what our experiments found. Everyone who does Buddhist work is welcome here: translators, teachers, editors, artists, librarians, center staff, developers. You do not need to write code or use git. Please follow the code of conduct.
Ways to take part
| You can give | How | Credit |
|---|---|---|
| A skill, or any file you give to AI (Claude file, project instructions, style guide, glossary, prompt) | Share your files (goes to the prompt library), or for a model paste a link on the homepage | authors on the entry; leaderboard row |
| Test material (exam papers, checked translations, corrected OCR) | Propose a benchmark | Dataset card; benchmark page |
| An expert review of a benchmark | Review a benchmark: check items and answers in a field you know | Named as reviewer in the dataset card |
| Reading in your language | Volunteer as a language reader | Named as reader |
| Compute or API credits | Donate compute or API credits: pays for evaluations of everyone’s submissions | “Supported by” on the homepage and org profile, if you wish |
| How your project works | Share a workflow | Owner of the workflow page |
| A project or dataset we do not list | Add to the Directory | — |
| A fix or a new page | “Edit this page” on any page | contributors on the page |
| Feedback on our sites, or a review of a design | UX reviews: tell us what was hard, review a design as a persona, write a design, or suggest a persona | Named in the design or the persona |
New to this? The tutorials teach the steps by doing them. Organizations that contribute often can get their own GitHub team: see Governance.
Improve a task page
Task pages belong to everyone. Click “Edit this page” on the site (or the pencil icon on GitHub) and propose a change. The most useful additions are:
- a real example with its correct answer, and where it comes from;
- an edge case: a hard case where people or models often make mistakes;
- a language or tradition note: how the task differs in Pali, Classical Chinese, Tibetan and so on;
- a sharper definition, especially where two tasks overlap;
- a link to a tool, data set, paper or benchmark.
Add your name to contributors in the page’s frontmatter. A maintainer reviews the change; a disagreement about a definition goes to the page’s discussion (an issue or a Discussions thread linked from the page) and both views can be recorded on the page until it is settled.
Who owns a workflow page
Every workflow has an owner: the person or organization whose practice it describes. Choose one of two modes.
| Mode | Where the document lives | What lives here | Who can change it |
|---|---|---|---|
linked |
Your own place (Google Doc, website, PDF) | A short page in our own words: summary, credit, a link to your version, and a map from your steps to tasks | You change your document whenever you like. Our page follows it; we update the version and the step map. |
hosted |
This repo | The full workflow | You review every change to your page (listed in .github/CODEOWNERS). Others suggest edits by pull request. |
Rules for both modes:
- We never copy a linked document wholesale. We summarize it and link to it.
- The owner is credited at the top of the page and in the frontmatter (
owner,contributors). Name the people who helped you incontributors. - A page records the owner’s own version number (
source_version) so readers know which version the step map describes. - An owner may ask us to remove or change their page at any time.
Share a workflow
- Easiest: open a “Share a workflow” issue and paste a link. A maintainer writes the
linkedpage, maps your steps to tasks and asks you to check it. - By pull request: copy
templates/workflow.mdtoworkflows/<id>.md, fill it in, runpython3 scripts/build.pyandpython3 scripts/check.py, and open a pull request.
Good workflow pages:
- Describe what people really do, including the steps no AI does today. A step with no task is useful information: it shows what is missing.
- Map each step to a task ID from
tasks/where one fits. If none fits, writenew task?in the table and say what the step needs; a maintainer will propose a task. - Stay one workflow per page. A variant (another language, another tradition, another team’s order of steps) gets its own page that names the one it adapts.
Propose a task
A task is one step that a model or skill can be measured on, such as segmentation or translation review. Open a “Suggest a task” issue, or copy templates/task.md. Task IDs are shared with buddhist-benchmarks tasks/index.json and skill names in buddhist-skills (<task>-<language>), so a new task gets added there too.
Help a task get a benchmark
Most tasks have no benchmark yet. A benchmark starts small: twenty to fifty real items from work you already do, with the answer a careful human would give. See buddhist-benchmarks SUBMIT.md and the task page’s “Help wanted” section.
Propose or run an experiment
Start with an Idea issue. How an experiment is planned, logged and reported is described in experiments/README.md.
For maintainers
- Frontmatter fields are checked by
scripts/check.py. The index tables inREADME.md, the “Found in workflows” lists on task pages and the directory tables indirectory/README.mdare generated byscripts/build.py; never edit them by hand. Directory entries are edited indirectory/projects.ymlanddirectory/datasets.yml(the field list is at the top of each file). - When a benchmark or a best skill changes, update the task page’s
benchmarks,leaderboardandskillsfields. - Version pages with semver in
version: a new step or task mapping is minor, a wording fix is patch, a restructure is major.source_versionis the owner’s own version and is separate.

BuddhistAI Knowledge base