dataeval.core.rerank_stratified¶

dataeval.core.rerank_stratified(result, num_bins=50)¶

Rerank by stratified sampling across score bins.

Takes a RankResult (expected to be in easy_first order) and applies stratified sampling to balance selection across score bins. This encourages diversity by de-weighting samples with similar scores.

The output is in hard_first order to maintain priority while balancing.

Parameters:¶

result : RankResult¶

Ranking result with scores (must be from rank_knn or rank_kmeans_distance).

num_bins : int, default 50¶

Number of bins for stratification.

Returns:¶

Dictionary containing:

• indices: NDArray[np.intp] - Reranked indices in hard_first order

• scores: NDArray[np.float32] | None - Scores in original order (unchanged)

• method: str - Same as input

• policy: str - “stratified”

Return type:¶

RankResult

Raises:¶

ValueError – If result does not contain scores (e.g., from rank_kmeans_complexity).

Examples

>>> from dataeval.core import rank_knn, rerank_stratified
>>> import numpy as np
>>> embeddings = np.random.rand(100, 64).astype(np.float32)
>>> result = rank_knn(embeddings, k=5)
>>> result = rerank_stratified(result, num_bins=20)