Add HAKARI-Bench results for voyageai/voyage-4-nano

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

# Two long-running workers evaluated disjoint shards of the standard --all target
# set. Each worker used one visible GPU, so CUDA_VISIBLE_DEVICES=1 is recorded as
# cuda:0 inside that worker's runtime metadata.
CUDA_VISIBLE_DEVICES=<0-or-1> uv run --group tf4-fa2 hakari-bench evaluate dense \
  --model voyageai/voyage-4-nano \
  --model-revision 67fabc9bef010dabc5f6024aa1b1b6b93410426f \
  --all \
  --dtype bf16 \
  --trust-remote-code \
  --attn-implementation flash_attention_2 \
  --embedding-variant truncate:1024,512,256 \
  --candidate-ranking reranking_hybrid \
  --batch-size <32 initially; 8 for resumed long-context shards>

Submitter Notes

• Evaluated as a SentenceTransformers dense model with trust_remote_code=True, bf16, flash_attention_2, the model's default 32768 max sequence length, and the model's stored query / document retrieval prompts.
• Truncation variants use the model-card-supported Matryoshka dimensions below the native 2048-dimensional output: 1024, 512, and 256. Default HAKARI dense quantization and rescore variants are included in each result JSON.
• The run was resumed safely from existing outputs. Result metadata records both batch sizes because initial shards used batch size 32 and resumed long-context shards used batch size 8 to avoid memory pressure without shortening max sequence length.
• These are standard built-in --all results. The submitted directory contains 551 compressed per-task result files and no intentionally narrowed benchmark subset.

Checklist

• Result files are committed under hakari-results/voyageai__voyage-4-nano/.
• 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.