Forest

Forest (forestdb.org) is a community repository for generative models written in probabilistic programming languages, mainly Church and WebPPL. Most code boxes are editable and run directly in the browser. Forest is maintained as a teaching resource and archive; it has long served as the home for class projects from courses such as Probabilistic Languages Understanding.

Adding models

To add a new model, create a markdown file in the models directory. You can do this directly on GitHub: add model.

For an example of the file format, see example.md. A minimal model file looks like this:

---
layout: model
title: My Model
model-language: webppl
model-status: code
model-category: Program Induction and Concept Learning
model-tags: concepts, program induction
---

Prose describing the model.

~~~~
flip(0.5)
~~~~

Frontmatter fields:

• model-language (required to make code boxes runnable): webppl or church.
• model-language-version (optional, webppl only): pins a webppl version. Available: pre-v0.7, v0.9.6, v0.9.7, v0.9.9, v0.9.13, v0.9.15 (default).
• model-status (optional): code (code runs), static (code shown as a non-runnable listing; add a model-status-verbose explanation), link (page links to external code), or hidden (not listed on the front page).
• model-category (required): one of Probability and Bayesian Data Analysis, Graphical Models and Causality, Regression and Statistical Learning, Time Series and Stochastic Processes, Bayesian Nonparametrics, Program Induction and Concept Learning, Language and Pragmatics, Agents, Games, and Social Reasoning, or Scientific and Physical Models. Use tags for cross-cutting provenance and applications such as benchmark, PPAML, replication, or a scientific domain.
• model-tags (optional): arbitrary comma-separated words or phrases.

Choose the primary category by the model's teaching target. Language meaning and use take precedence over recursive-agent machinery. Nonlinguistic choice, teaching, games, and social reasoning belong under agents. Inferred programs, rules, grammars, and concepts belong under program induction. Use Bayesian nonparametrics when the nonparametric prior itself is the lesson; use time series for temporal latent states or event streams; use graphical models for factorization, interventions, or causality; and use scientific models when a specific domain mechanism is central. General inference and hierarchical Bayes belong under probability and Bayesian data analysis.

To display a code box without making it runnable (e.g. to show a model fragment), mark the fence with norun:

~~~~ norun
// this box renders as static code
~~~~

Model coverage

Forest contains 214 models spanning Bayesian data analysis, graphical and causal models, statistical learning, time series, Bayesian nonparametrics, program induction, language and pragmatics, agent reasoning, and scientific models. The collection includes drift diffusion, item-response theory, survival with censoring, phylogenetics, Hawkes processes, stochastic volatility, interventional structure learning, partial pooling, signal detection, and Gaussian-process regression alongside the original Church and WebPPL teaching corpus.

Machine-readable index

forestdb.org/models.json lists every model with title, URL, markdown source URL, language, version, status, category, and tags.

Local development

The site is plain Jekyll, built by GitHub Pages from the gh-pages branch. To preview locally:

docker run --rm -v "$PWD":/site -p 4000:4000 -w /site jekyll/jekyll:4.4.1 \
  jekyll serve --host 0.0.0.0

then open localhost:4000.

All JavaScript and CSS (jQuery, Bootstrap, the webppl runtimes, webchurch) is vendored under assets/, so the site has no runtime dependencies on external CDNs. To support a new webppl version, build or download the browser bundle, add it as assets/webppl/webppl-<version>.js, and add the version to the list in this README and to scripts/test-models/package.json.

Continuous checks

Automated checks keep the site healthy (see .github/workflows/):

• Validate models checks frontmatter, category membership, unique titles, concise introductions on repository-maintained pages, and body hashes for protected third-party pages. Run node scripts/validate-models.js locally.
• Test models runs every WebPPL model headless against its declared version and rejects failures outside the documented baseline. Run locally with cd scripts/test-models && npm install --install-strategy=nested && npm run check (the nested install is needed for WebPPL 0.6.1's hardcoded module paths).
• Check links runs lychee over all model pages and files a report issue when links break.

License

Site code is MIT-licensed; models belong to their authors, with new contributions accepted under CC-BY 4.0 (prose) and MIT (code). See LICENSE.md.