dotime 0.1.1

Synthetic temporal structural causal models with interventions for causal foundation models

DoTime

A synthetic benchmark generator for interventional and counterfactual time series.

dotime samples multivariate temporal structural causal models (SCMs), applies interventions to them, and produces paired observational / interventional trajectories together with exact counterfactual targets.

Where most time-series causal benchmarks are observational (recover a graph from passive dynamics), DoTime is interventional / counterfactual: it answers do(...) queries over time and ships frozen suites to measure how well a model estimates interventional effects.

Highlights

• Four intervention regimes — hard, soft, time-varying, and continuous-time intervention windows.
• Four counterfactual sampling modes — prior, observed_discrete, observed_normal, observed_uniform, with a positivity guard.
• Regime-switching SCMs as a strict generalization of interrupted-time-series.
• Eight named identification structures (back-door, front-door, IV, …) with exact counterfactual ground truth.
• Reference baselines and four frozen evaluation suites (released with Zenodo DOIs and Croissant metadata).

Install

pip install dotime            # core generator + suite loaders (CPU, no GPU needed)
pip install 'dotime[baselines]'   # classical / Bayesian baselines
pip install 'dotime[all]'         # everything except dev tooling

Requires Python ≥ 3.10. The core install runs on CPU with no GPU dependency.

Quickstart

from dotime import DoTime

prior = DoTime(seed=42)
X_obs, X_int, intervention, scm = prior.generate_pair(T=100)
# X_obs, X_int: (T, N) tensors;  intervention: InterventionSpec;  scm: TemporalSCM

Command-line:

dotime-generate -n 1000 -T 200 -o data/sample.pt   # sample paired trajectories
dotime-benchmark --list                            # list frozen benchmark suites

Hosted artifacts

Four frozen suites, on the Hugging Face Hub (discovery) and Zenodo (citable archive). load_benchmark("dot-Identifiability-v1") pulls and md5-verifies them.

Suite	Episodes	Hugging Face	Zenodo DOI
dot-Identifiability-v1	10,800	thummd/dot-Identifiability-v1	10.5281/zenodo.20846064
dot-RegimeSwitch-v1	9,999	thummd/dot-RegimeSwitch-v1	10.5281/zenodo.20846074
dot-Continuous-v1	9,999	thummd/dot-Continuous-v1	10.5281/zenodo.20845981
dot-Generic-100k	100,000	thummd/dot-Generic-100k	10.5281/zenodo.20845983

Trained Do-Over-Time-PFN checkpoints: thummd/do-over-time-pfn.

Documentation

Full docs at https://dotime.readthedocs.io. See docs/quickstart.md and docs/benchmarks.md to get started locally.

Citation

If you use DoTime, please cite the paper and the software (see CITATION.cff).

License

Apache-2.0 (code). Released benchmark suites are distributed under CC-BY-4.0.