HAKARI-Bench Leaderboard

A lightweight benchmark for multilingual retrieval and reranking

HAKARI-Bench evaluates dense embeddings, learned sparse retrieval, late interaction, and cross-encoder rerankers, alongside BM25, across 37 dataset groups and 550 tasks. Models are evaluated using common tasks and metrics on compact datasets (Nano-sets). Results include base model performance and changes in retrieval quality from dimensionality reduction, int8/binary quantization, and float rescoring. Filter by language or dataset to compare results relevant to your use case.

Retrieval 79 models

Dense, sparse, late interaction, and BM25 search over the full corpus.

Reranking 81 models

Cross-encoders and other models reorder a shared candidate set.

Evaluation guide Paper GitHub Datasets
Loading leaderboard...
Loading leaderboard...