Add HAKARI-Bench results for sbintuitions/sarashina-embedding-v2-1b

Summary

DuckDB Nano-set Comparison

Computed from DuckDB task_results with the same Overall grouping as this PR body. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Overall nDCG@10

Reproducibility

Command

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \\
CUDA_VISIBLE_DEVICES=0 \\
uv run hakari-bench evaluate from-model-card \\
  --model-card config/model_cards/sbintuitions__sarashina-embedding-v2-1b.yaml \\
  --dataset "$DATASET_GROUP" --device cuda:0 --batch-size 16 --show-progress

# A second non-overlapping dataset group was run concurrently with
# CUDA_VISIBLE_DEVICES=1 (also addressed as cuda:0 inside that process).

Submitter Notes

• Evaluated the standard --all scope: 538 tasks across 47 datasets. The model was pinned to revision 1f3408afaa7b617e3445d891310a9c26dd0c68a5, with bf16, SDPA, and its configured 8192-token maximum sequence length.
• Used the model-author retrieval prefixes: query task: 質問を与えるので、その質問に答えるのに役立つ関連文書を検索してください。\nquery: and document text: . No truncate/Matryoshka variant was requested because the model card does not document truncation support. The standard full-dimension int8/binary and rescore variants are included.
• An initial batch-size-32 attempt encountered memory pressure on long documents. The canonical workers were restarted at batch size 16; 22 early files retain batch size 8 and one completed smoke result retains batch size 32. These settings do not alter the model, prompts, or sequence length.
• The model is released under the Sarashina Model NonCommercial License Agreement; these results do not imply commercial-use permission.

Checklist

• Result files are submitted under hakari-results/sbintuitions__sarashina-embedding-v2-1b/.
• 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.