Complete multilingual HAKARI-Bench coverage for dmis-lab/UltraBERT-Embed-150M

Summary

Adds 398 missing standard tasks to the 165 English/code tasks merged in PR #44. The resulting model coverage is 563 evaluation tasks, or 550 tasks in the Overall leaderboard after duplicate exclusions. The existing 165 files remain byte-for-byte unchanged.

Overall nDCG@10

Reproducibility

Command

CUDA_VISIBLE_DEVICES=1 PYTHONUNBUFFERED=1 uv run --group flash-attn \
  --with transformers==5.20.1 --with sentence-transformers==6.1.0 --with torch==2.9.1 \
  hakari-bench evaluate dense \
  --model dmis-lab/UltraBERT-Embed-150M \
  --model-revision b06b36724c5417a1ec1b7863dd4fa542b74e4899 \
  --all --dtype bf16 --attn-implementation flash_attention_2 \
  --query-prompt-name query --document-prompt-name document \
  --batch-size 32 --device cuda:0 --results-dir output/hakari-results

Submitter Notes

• Evaluation code: hakari-bench/hakari-bench remote main ed765c5 (full checkout SHA recorded locally). The model remains a dense encoder; this is a coverage extension, not a sparse evaluation.
• Physical GPU 1 was isolated with CUDA_VISIBLE_DEVICES=1; the new process therefore records cuda:0. The reused English results record cuda:1 from the original run.
• Kept the model's default 32,768-token maximum length, bf16, Flash Attention 2, native 768-dimensional vectors, and configured query/document prompt names. No trust-remote-code or context shortening was used. No documented Matryoshka truncation dimensions were found.
• All 563 tasks contain base, int8, binary, int8_rescore, and binary_rescore results. Coverage and model/runtime metadata audits passed with no missing tasks or variants.
• Before extending coverage, NanoArguAna was evaluated in the new environment: all five variant scores matched the published English result exactly.
• Standard --all coverage: 398 newly evaluated tasks plus 165 cache hits. Batch size stayed at 32; no OOM fallback was needed. A transient Hugging Face DNS error recovered through the built-in retry.
• Dataset revisions, full runtime details, and ranking artifacts are retained in each result file. The post-run local audit wrapper initially failed to import the package; the evaluation itself exited successfully, and the corrected audit passed separately.

Checklist

• Result files are committed under hakari-results/dmis-lab__UltraBERT-Embed-150M/.
• Result files are compressed .json.xz; no caches, DuckDB files, HTML reports, or local scratch artifacts are included.
• The result JSON records model revision, dataset revision, runtime configuration, and package versions.
• Overall nDCG@10 above was generated from the submitted result files.
• Any non-default prompt, sequence length, attention implementation, candidate ranking, or reranker setting is documented above.