Skip to content

Datasets

Four public datasets ship with loaders. Two are auto-downloaded; the other two require a free login on Ods.ai and a manual file placement.

Hillstrom (MineThatData)

Kevin Hillstrom's classic e-mail marketing experiment. ~64k rows, three treatment arms (no email / mens email / womens email), three outcomes (visit, conversion, spend).

  • Auto-download. uv run uplift-bench download hillstrom
  • URL. http://www.minethatdata.com/Kevin_Hillstrom_MineThatData_E-MailAnalytics_DataMiningChallenge_2008.03.20.csv (HTTP only; HTTPS fallback via a GitHub mirror).
  • Default contrast. Womens E-Mail vs No E-Mail — has the strongest measured ATE in published analyses.
  • Default outcome. visit — conversion is too rare to give a meaningful Qini at this n.

Criteo Uplift Prediction Dataset v2.1

The standard uplift benchmark. ~13.9M rows, 12 anonymised features, binary treatment, two outcomes (visit, conversion).

  • Auto-download. uv run uplift-bench download criteo
  • URL. https://huggingface.co/datasets/criteo/criteo-uplift/resolve/main/criteo-research-uplift-v2.1.csv.gz (~297 MB compressed, ~1.5 GB uncompressed).
  • Cache. First load decodes the CSV.gz and caches a parquet next to it; subsequent loads parse in seconds.
  • Subsample. For local iteration use dataset.loader_params.subsample=1_000_000 to grab a fixed-seed one-million-row slice.

Lenta uplift dataset

Russian grocery loyalty campaign published by Lenta and packaged into the scikit-uplift library.

  • Auto-download. uv run uplift-bench download lenta
  • URL. https://sklift.s3.eu-west-2.amazonaws.com/lenta_dataset.csv.gz (~138 MB compressed, ~567 MB uncompressed). Same source-of-truth URL that sklift.datasets.fetch_lenta uses, so this loader stays in lock-step with upstream without depending on it at runtime.
  • Schema. ~687k rows, binary treatment (group = test/control, ratio ~0.75/0.25), binary outcome (response_att, ~10% positive). We retain ~14 most informative numeric features plus a one-hot gender block; the full ~190-column raw schema is documented at https://www.uplift-modeling.com/en/v0.5.1/api/datasets/fetch_lenta.html.
  • Cache. First load decodes the CSV.gz and caches a parquet next to it; subsequent loads parse in seconds.
  • Sample fixture. data/sample/lenta/lenta_dataset.csv.gz for offline tests / smoke runs.

X5 RetailHero Uplift Modeling Contest

Russian retail-loyalty contest dataset. Login-walled on Ods.ai.

  • Manual download. Sign in at https://ods.ai/competitions/x5-retailhero-uplift-modeling/data and place 
    data/raw/retailhero/uplift_train.csv
data/raw/retailhero/clients.csv
  • Loader. RetailHeroLoader joins clients onto the uplift table, median-imputes missing numerics, one-hots gender.
  • Sample fixture. A tiny synthetic stand-in lives at data/sample/retailhero/ so CI smoke runs work offline.

MegaFon Uplift Competition

Telco dataset, ~600k rows, 50 anonymised numeric features.

  • Manual download. Sign in at https://ods.ai/competitions/megafon-uplift-competition/data and place 
    data/raw/megafon/train.csv
  • Loader. MegaFonLoader discovers X_* features at read time; normalises both treatment-column variants ("treatment_group" string and numeric).
  • Sample fixture. data/sample/megafon/.

Sample fixtures

Every loader can read from data/sample/<dataset>/ instead of data/raw/<dataset>/. Sample files match the production schemas but contain only a few thousand synthetic rows. Useful for tests and confirming the install without downloading anything.