mackelab/benchmarking_sbi_runs

Benchmarking SBI Runs

This dataset contains the raw, per-run results underlying the manuscript "Benchmarking Simulation-Based Inference" (Lueckmann, Boelts, Greenberg, Goncalves & Macke, AISTATS 2021).

It is a direct migration of the Git LFS data from mackelab/benchmarking_sbi_runs on GitHub. For compiled, ready-to-use dataframes built from these raw results (and the code that produced them), see the companion repository:

https://github.com/sbi-benchmark/results/tree/main/benchmarking_sbi

Dataset Summary

The benchmark evaluates simulation-based inference (SBI) algorithms — classical Approximate Bayesian Computation (ABC) methods, Synthetic Likelihood, and neural density-estimation methods that approximate likelihoods, posteriors, or likelihood ratios — across 10 inference tasks with known ground-truth reference posteriors.

This dataset contains 21,261 individual run folders (~147,877 files, ~17.5 GB total), each holding the full output of one algorithm/task/observation/simulation-budget configuration, including posterior samples, posterior-predictive samples, and evaluation metrics.

Dataset Structure

Each run is stored under <run-id>/ (a UUID) at the repo root, containing:

File	Description
run.yaml	Run configuration: algorithm name, device, hyperparameters, task name, observation index, and simulation budget
metrics.csv	Evaluation metrics for this run (see below)
posterior_samples.csv.bz2	Samples drawn from the approximate posterior (bz2-compressed CSV, columns parameter_1..parameter_D)
predictive_samples.csv.bz2	Posterior-predictive samples (bz2-compressed CSV)
log_prob_true_parameters.csv	Log-probability the approximate posterior assigns to the true parameters (where applicable)
num_simulations_simulator.csv	Actual number of simulator calls used
runtime.csv	Wall-clock runtime (seconds) for the run

run.yaml schema

algorithm:
  device: cpu
  name: SMC-ABC          # algorithm identifier, see below
  params: {...}          # algorithm-specific hyperparameters
  run: pyabc.smcabc.run  # entry point used
compute_metrics: true
device: cpu
seed: 4023017558
task:
  name: bernoulli_glm     # task identifier, see below
  num_observation: 6      # 1-10, which of the task's reference observations
  num_simulations: 100000 # simulation budget for this run

metrics.csv columns

C2ST, C2ST_Z, MMD, MMD_Z, KSD_GAUSS, MEDDIST, C2ST_1K, C2ST_1K_Z, MMD_1K, MMD_1K_Z, KSD_GAUSS_1K, MEDDIST_1K, NLTP, RT

(C2ST = classifier two-sample test accuracy vs. the reference posterior; MMD = maximum mean discrepancy; KSD_GAUSS = kernel Stein discrepancy; MEDDIST = median distance; NLTP = negative log true-parameter probability; RT = runtime; the _1K suffix denotes the metric recomputed on a 1,000-sample subset; _Z denotes a z-scored variant.)

Tasks

10 tasks spanning statistical and applied simulator models, each with 10 fixed reference observations and known/estimated reference posteriors:

bernoulli_glm, bernoulli_glm_raw, gaussian_linear, gaussian_linear_uniform, gaussian_mixture, lotka_volterra, sir, slcp, slcp_distractors, two_moons

Algorithms

• ABC: REJ-ABC (rejection ABC), SMC-ABC (sequential Monte Carlo ABC), RF-ABC (random forest ABC)
• Synthetic Likelihood: SL
• Neural, amortized: NPE-{MAF,NSF}, NLE-{MAF,NSF}, NRE-{MLP,RES}
• Neural, sequential: SNPE-{MAF,NSF}, SNLE-{MAF,NSF}, SNRE-{MLP,RES}
• Baselines: Prior/Prior baseline, Posterior/Posterior baseline

Simulation budgets used: 1,000 / 10,000 / 100,000 (all methods), and 10,000,000 (SL only).

How to use

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="mackelab/benchmarking_sbi_runs",
    repo_type="dataset",
)

Or download a subset with --include, e.g. */metrics.csv, to just pull the lightweight metrics without the sample files.

Citation

@InProceedings{lueckmann2021benchmarking,
  title     = {Benchmarking Simulation-Based Inference},
  author    = {Lueckmann, Jan-Matthis and Boelts, Jan and
               Greenberg, David and Goncalves, Pedro and Macke, Jakob},
  booktitle = {Proceedings of The 24th International Conference
               on Artificial Intelligence and Statistics},
  pages     = {343--351},
  year      = {2021}
}

Paper: PMLR v130 · arXiv:2101.04653