Add HAKARI-Bench results for cl-nagoya/ruri-v3-reranker-310m

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

CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 hakari-bench evaluate from-model-card \
  --model-card config/model_cards/cl-nagoya__ruri-v3-reranker-310m.yaml \
  --all \
  --device cuda:0 \
  --dtype bf16 \
  --attn-implementation flash_attention_2 \
  --model-max-seq-length 8192 \
  --batch-size 16 \
  --candidate-ranking reranking_hybrid \
  --rerank-top-k 100 \
  --results-dir output/hakari-results

Submitter Notes

• Final full run used the tf4-fa2 dependency group with Transformers 4.57.6, Sentence Transformers 5.4.1, bf16, Flash Attention 2, trust_remote_code=false, and batch size 16 on a single RTX 5090. The model card records max_seq_length=8192.
• Reranker evaluation used the reranking_hybrid candidate ranking with rerank_top_k=100. The model README documents direct CrossEncoder pair scoring and does not document query/document prompt prefixes; no prompt names are recorded in the result JSON.
• Standard --all benchmark scope was completed: 551 .json.xz task result files, exit code 0, and cache_hit_count=0 in the final run log. No failed tasks remain in the submitted directory.

Checklist

• Result files are committed under hakari-results/cl-nagoya__ruri-v3-reranker-310m/.
• 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.