tanzilidrisi/tib-benchmark

TIB v1.0: Timeline Intervention Benchmark

Propensity-logged benchmarks for session-level e-commerce intervention policies (six actions including a do_nothing control), with exact logged propensities, counterfactual potential outcomes, multi-seed protocol, and a baseline leaderboard.

Why

Public logged-bandit datasets (Open Bandit Dataset) evaluate item recommendation. Intervention policies — margin-sensitive, fatiguing, timing-dependent actions — had no public, propensity-logged benchmark. TIB provides three tiers that deliberately span the regimes where timeline (sequence) context helps and where it does not, so representation claims can be tested rather than assumed.

Contents

File	What
data/	Seed-1 sample CSVs for the synth and twin tiers
CHECKSUMS.txt	SHA-256 for every canonical seed (regenerate and verify)
PROTOCOL.md	Normative evaluation rules
DATASHEET.md	Datasheet (Gebru et al. structure)
LEADERBOARD.md	Baseline results, 11 policies
croissant.jsonld	Machine-readable metadata

Regenerate everything

pip install -e .
python - <<'PY'
from tvcb.benchmark import materialize_synth_csv, materialize_twin_csv
for seed in range(1, 11):
    materialize_synth_csv(f"tib_synth_seed{seed}.csv", sessions=10000, seed=seed)
    materialize_twin_csv(f"tib_twin_seed{seed}.csv", sessions=20000, seed=seed)
PY
sha256sum -c CHECKSUMS.txt

The retail tier build recipe is documented in the repository (docs/retailrocket.md + make intervention-final).

Citation

Archived at Zenodo: https://doi.org/10.5281/zenodo.20651686

@dataset{idrisi2026tib,
  author    = {Idrisi, Mohammad Tanzil},
  title     = {{TIB} v1.0: Timeline Intervention Benchmark},
  publisher = {Zenodo},
  year      = {2026},
  doi       = {10.5281/zenodo.20651686}
}