Published April 9, 2026 | Version v0.26.1

sbi reloaded: a toolkit for simulation-based inference workflows

Description

⚠️ Breaking Changes

  • Make PyMC and Pyro optional dependencies (#1835): pip install sbi no longer installs pymc or pyro-ppl. Users who need Pyro or PyMC MCMC samplers should install extras:
    • pip install "sbi[pyro]" for Pyro samplers (hmc_pyro, nuts_pyro)
    • pip install "sbi[pymc]" for PyMC samplers (slice_pymc, hmc_pymc, nuts_pymc)
    • pip install "sbi[all]" for both
    • Using a Pyro/PyMC method without the dependency installed raises a clear ImportError with install instructions.

🐛 Bug Fixes

  • Fix TARP z-scoring bug (#1832): Reference points are now z-scored alongside thetas and posterior_samples when z_score_theta=True, fixing incorrect distance calculations that masked bias detection.
  • Fix broken biased_toy_gaussian test helper: Rewrote to create actual location bias (posterior mean shifted from truth) instead of the previous NaN-producing formula.

📖 Documentation

  • Streamlined README installation section, recommend uv as default.
  • Added optional dependency install instructions to README, installation guide, and relevant tutorials.

🔧 Improvements

  • Change default num_bins in TARP: run_tarp and _run_tarp now default to num_bins=None (auto-scales to num_sims // 10) instead of the hardcoded 30, improving KS test power for larger sample sizes.

🔒 Security

  • Added SLSA build provenance attestations to the release workflow.
  • Added Dependabot for automated GitHub Actions version updates.
  • Added SECURITY.md with vulnerability disclosure policy.
  • Fixed release uploads with --clobber flag.

Notes

If you use this software, please cite our article in the Journal of Open Source Software.

Files

sbi-dev/sbi-v0.26.1.zip

Files (7.3 MB)

Name Size Download all
md5:39f04bd14a4a88ea3b9ee56a3c5b5508
7.3 MB Preview Download

Additional details

Related works

Is supplement to
Software: https://github.com/sbi-dev/sbi/tree/v0.26.1 (URL)

Software