LoginSign Up
← Back to all sparks
P

posteriordb

ANALYTICS
Velocity0.0

Database with posteriors of interest for Bayesian inference

github.com/stan-dev/posteriordb ↗

A reference posterior database that hit 1.0 with a paper, and is now graded on the statistics it ships.

bayesianbenchmarkingreference-datastanopen-science
◆Current state
posteriordb distributes Bayesian models with data and reference posterior draws so inference algorithms can be benchmarked against a common target. It reached 1.0.0 alongside a published paper, and ships both R and Python access. Recent work is about the metadata around the draws — licences, machine-readable dataset descriptors, and additional summary statistics.
◆Where it's heading
The database is maturing from a model collection into a citable benchmark asset: licence information per model, a Croissant metadata file for dataset discovery, and summary statistics like mean squared value and lag-1 autocorrelation that let users judge whether reference draws are good enough for their comparison. Earlier releases were about content and correctness; current ones are about making the content machine-readable and verifiable.
◆Prediction
Further work should continue on draw-quality diagnostics and metadata rather than model count, since the last two releases both added ways to assess the reference draws instead of adding posteriors.

◆Recent moves

  1. 1y ago

    1.0.0: licences, Croissant metadata, and draw diagnostics

    The 1.0 release is additive rather than a redesign: per-model licences, a Croissant descriptor, and two new summary statistics. The accompanying paper is what makes the version number meaningful — the database is now citable as a benchmark.

    View source ↗
  2. 2y ago

    Stan code updated to 2.26 syntax; posterior tags cleaned

    Moving every model to modern Stan syntax raises the floor on what version users need but keeps the database compiling against current toolchains. The autocorrelation-based guidance here is the precursor to the diagnostics formalised in 1.0.0.

    View source ↗
  3. 3y ago

    New posteriors and a corrected dogs model

    Content growth plus a fix to a model that was wrong — the latter matters more for a reference database, where an incorrect target silently invalidates every comparison made against it.

    View source ↗
  4. 5y ago

    Python module gains GitHub-backed and env-var database paths

    Establishes that the database is consumed from Python as well as R, with a GitHub-backed client so users need not vendor the data. Cross-language access is what lets the benchmark span inference ecosystems.

    View source ↗