UVET-benchmark

UVET-benchmark is an 86.7-million-impression log of a randomized A/B test run by an online advertising platform, released with the NeurIPS 2026 paper UVET: A Partitioning Approach for Continuous Treatment Assignment At Scale. Each impression received one of five bid multipliers α ∈ {0.8, 0.9, 1.0, 1.1, 1.2} applied to the platform's reference bidding policy (α = 1), and both its value and its cost were recorded. The task is to learn, from five covariates, a policy that assigns a multiplier to each impression so that total value exceeds the reference while total cost stays within the reference budget.

During double-blind review the same data was distributed as anonadata/CUVET-policy; the files here are byte-identical.

Files and columns

File	Rows	Content
uvet_benchmark_train.csv	65,031,420	75% of the rows, at random; is_val marks the validation subset
uvet_benchmark_test.csv	21,677,141	the remaining 25% of the rows

Column	Type	Description
(unnamed first column)	int	row index
f0 – f4	float	five continuous covariates built from online-interaction signals, perturbed and standardized, without semantic labels
policy	float	bid multiplier α assigned at random
spend	float	cost of the impression
value	float	value of the impression for the advertising objective
is_val	int (train file only)	1 for the validation subset (21,677,140 rows), 0 for training (43,354,280 rows)

No column has missing values. value is positive on every row, and spend is zero on fewer than 0.01% of rows.

Splits

The rows of the two-week test were split at random into training (50% of all rows), validation (25%, flagged by is_val in the training file) and test (25%). All three splits cover the same two weeks.

Treatment levels

α	0.8	0.9	1.0 (reference)	1.1	1.2
share of training rows	8.5%	9.2%	60.2%	10.7%	11.3%
value total (training rows)	957	973	994	1009	1014
cost total (training rows, thousands)	211.0	240.3	271.2	299.8	330.5

Shares are the same in the training, validation and test rows to one decimal. The four non-reference levels received equal shares of the test traffic and the reference level a larger share. Rows are won impressions, so higher multipliers win more auctions and produce more rows. The reference level holds 60.2% of rows, with value and cost per row about six times lower than at the other levels, and its totals lie on the same smooth curve as the other four, consistent with reference outcomes scaled to the traffic of one level. Compare treatment levels through their totals (global log–log elasticities 0.15 for value and 1.11 for cost on the training rows); per-row means are not comparable across levels.

Privacy and anonymization

The original data contains no direct identifiers such as names, postal addresses or e-mail addresses. The covariates were selected by hand from online-interaction signals, perturbed with noise and standardized, and their original scales are not released. Value and cost were scaled and perturbed to protect business confidentiality: ordering and relative comparisons are preserved, absolute magnitudes are not, and no interpretation in currency is possible.

Intended use

The dataset supports offline policy learning and budget-constrained treatment assignment with an ordered treatment, predict-then-optimize methods, and the evaluation of uplift and heterogeneous-treatment-effect methods with two outcomes. Because treatments were randomized, offline evaluation of the logged treatment levels is unbiased under the usual assumptions (no interference, consistency). It is not suitable for estimating monetary return on investment.

Known limitations

• Value and cost are scaled and perturbed, which limits absolute interpretation.
• The multiplier is continuous, but only five levels were logged, so offline evaluation is restricted to those levels.
• Offline evaluation assumes that rows are independent; the random split does not test robustness to shifts over time.

Loading

from datasets import load_dataset

ds = load_dataset("criteo/uvet-benchmark")  # about 13.7 GB of CSV

License

CC BY-NC-SA 4.0.

Citation

@inproceedings{betlei2026uvet,
  title     = {{UVET}: A Partitioning Approach for Continuous Treatment Assignment At Scale},
  author    = {Betlei, Artem and Vladimirova, Mariia and Girou, Victor and Rahier, Thibaud},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}