Datasets:

---

hakari-bench

/

results

like 0

Follow

HAKARI-Bench 3

This PR reverts the result-file deletion performed by #20.

NanoBEIR-en is the English component of MNanoBEIR and should remain part of the complete evaluation set. A complete evaluation preserves 551 task results. The leaderboard excludes 13 other overlapping Nano task copies only when calculating Overall, producing the intended 538-task Overall score.

Changes

• Restore 1,001 historical hakari-bench__NanoBEIR-en/*.json.xz files.
• Restore 13 English NanoBEIR task results for each of 77 models.
• Make no changes outside the restored NanoBEIR-en result paths.

Validation

• Changed files: 1,001
• Model directories: 77
• Files per model: 13
• Unexpected paths: 0
• DuckDB, cache, report, and unrelated files: 0

Follow-up

The benchmark code repository must separately restore the NanoBEIR-en dataset definition and its membership in MNanoBEIR. Models added after #20, including Sarashina Embedding v2, do not have historical English results in this revert and require a separate 13-task evaluation/upload.