amazon/music-off-policy-evaluation-benchmark

Music Off-Policy Evaluation Dataset

Music Off-Policy Evaluation Dataset is a dataset designed for Off-Policy Evaluation (OPE) research. It contains logged interactions from the home page of Amazon Music.

Use cases:

• Benchmarking OPE estimators
• Evaluating counterfactual ranking policies offline

License

Music Off-Policy Evaluation Benchmark © 2026 by Amazon is licensed under Creative Commons Attribution-NonCommercial 4.0 International.

Code

Please refer to our GitHub repository for the code to reproduce the benchmark.

Citation

Paper under review.

Schema

The following table provides an overview of the dataset schema. Let it be $kk$ the number of available actions for a given observation.

Column	Type	Dimension	Description
actions	List[List[float]]	$L\times 129L\; \backslash times\; 129$	Context vectors of size $129129$ for each action in the observation.
rewards	List[float]	$min(L,50)min(L,\; 50)$	Observed binary rewards per action.
logging_selected_actions	List[int]	$min(L,50)min(L,\; 50)$	Selected actions of the logging policy ${\pi }_0\backslash pi\_0$.
target_selected_actions	List[int]	$min(L,50)min(L,\; 50)$	Selected actions of the target policy $\pi \backslash pi$.
propensities	List[List[float]]	$min(L,50)\times min(L,50)min(L,\; 50)\; \backslash times\; min(L,\; 50)$	Squared matrix of propensities.

Actions

Vectors of size 129 describing a context vector. The number of available actions $LL$ differs across observations.

Example:

Assuming for a given observation there are $L=3L\; =\; 3$ actions, then actions contains $33$ vectors each of size $129129$ as following:

actions = [
  [1.         0.         0.28867126 ... 0.09611709 0.32168165 0.        ],
  [1.         0.         0.32252225 ... 0.19847895 0.24163522 0.        ],
  [1.         0.         0.57331926 ... 0.15339291 0.57508302 0.        ]
]

Logging selected actions

Vector of action indices selected by ${\pi }_0\backslash pi\_0$ in the order they are displayed.

Example:

Let's assume in a given observation there are $L=3L\; =\; 3$ actions such that:

logging_selected_actions = [2, 1, 0]

It indicates that the third action (index 2) is ranked by ${\pi }_0\backslash pi\_0$ in the first position, the second action (index 1) is ranked in second position and first action (index 0) ranked in third position.

Target selected actions

Vector of action indices selected by $\pi \backslash pi$.

Example:

Let's assume in a given observation there are $L=3L\; =\; 3$ actions such that:

logging_selected_actions = [2, 1, 0]
target_selected_actions = [1, 2, 0]

It indicates that $\pi \backslash pi$ ranks in first position the action with index 1 (while ranked in second position under ${\pi }_0\backslash pi\_0$); action with index 2 is ranked by $\pi \backslash pi$ in second position (first position under ${\pi }_0\backslash pi\_0$). Action index 0 is being ranked in third position under both ${\pi }_0\backslash pi\_0$ and $\pi \backslash pi$.

Rewards

Vector of binary rewards ordered by logging_selected_actions.

Example:

Assuming we have $33$ actions in the actions set:

logging_selected_actions = [2, 1, 0]
rewards = [1.0, 0.0, 0.0]

It indicates that a positive reward of $11$ has been observed for the action at index $22$ in actions vector which is ranked by ${\pi }_0\backslash pi\_0$ in first position.

Propensities

Propensity matrix $P\in R^{d\times d}P\; \backslash in\; R^\{d\; \backslash times\; d\}$ where $d=min(L,50)d\; =\; min(L,\; 50)$ that describes the probabilities with which actions are ranked in different positions under the logging policy ${\pi }_0\backslash pi\_0$. Rows correspond to actions (in the order they were ranked) and columns correspond to positions in the ranking. Specifically, $P_{ij}P\_\{ij\}$ describes:

Example:

Let it be the action space $\mathcal{A}=\{A,B,C\}\backslash mathcal\{A\}\; =\; \backslash \{A,\; B,\; C\backslash \}$ with $L=3L=3$ and assume:

1. The stochastic logging policy ${\pi }_0\backslash pi\_0$ produces the ranking $[B,A,C][B,\; A,\; C]$, and

2. The probability (as determined by the policy's selection mechanism) for

   • action $BB$ to end up in the first position was $0.50.5$ (it could have ended up in the second position with probability $0.40.4$ or in the third position with probability $0.10.1$),

   • action $AA$ to end up in the second position was $0.30.3$ (it could have ended up in the first position with probability $0.30.3$ or in the third position with probability $0.40.4$),

   • action $CC$ to end up in the third position was $0.50.5$ (it could have ended up in the first position with probability $0.20.2$ or in the second position with probability $0.30.3$).

Then, the propensities matrix is: where the rows correspond to the actions as ranked (i.e., $[B,A,C][B,\; A,\; C]$), and columns correspond to positions the actions could have been ranked (first, second and third).