Datasets:

---

hakari-bench

/

results

like 0

Follow

HAKARI-Bench 3

Summary

This PR updates only NanoBEIR-{lang}/nfcorpus result JSON files for models whose previous rows were produced against older NanoNFCorpus data revisions.

The re-evaluation uses the restored NanoNFCorpus data revisions and keeps the output limited to the affected nfcorpus task files.

Scope

• Repository path: hakari-results/**/hakari-bench__NanoBEIR-*/nfcorpus.json.xz
• Models: 53 affected models
• Datasets: 14 NanoBEIR-{lang} datasets (ar, de, en, es, fr, it, ja, ko, no, pt, sr, sv, th, vi)
• Files: 742 .json.xz files

Validation

• Re-evaluation jobs completed: 53/53
• Output files present: 742/742
• Score comparisons against prior rows: 6720
• Rows below 90% of the previous same model/dataset/variant nDCG@10: 0
• Staging directory audit: no non-.json.xz files, no DuckDB files

Runtime note

Qwen/Qwen3-Reranker-0.6B failed with OOM at the original batch size and was completed with --batch-size 4. Its max sequence length was not shortened.