Add HAKARI-Bench results for nvidia/Nemotron-3-Embed-8B-BF16

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 flash-attn hakari-bench evaluate dense \
  --model nvidia/Nemotron-3-Embed-8B-BF16 \
  --all \
  --results-dir output/nemotron-3-embed-8b-tf5121-fa283-full \
  --dtype bf16 \
  --attn-implementation flash_attention_2 \
  --query-prompt-name query \
  --document-prompt-name document \
  --batch-size 8 \
  --model-revision 2b29550c4ab0646bb6bb47032dda54ea11f6dfe2

Submitter Notes

• The run uses the model's reviewed SentenceTransformers query and document prompts (query: and passage: ), BF16, FlashAttention 2, the model's configured 32,768-token maximum sequence length, and no truncation variants because this checkpoint is not documented as Matryoshka. Dense defaults provide the full-dimension int8, binary, int8_rescore, and binary_rescore variants.
• This was a resumed standard evaluation across two machines. Existing outputs were retained without --overwrite; the earlier RTX 5090 run used batch sizes 1, 2, and 4, while the resumed RTX PRO 6000 Blackwell Max-Q run used batch size 8. The resume was split into disjoint dataset groups across two GPUs, with one process stopped after its Thai tasks completed to avoid duplicate writes when the groups converged. No completed result file was overwritten. All 551 files were subsequently decompressed and audited for the pinned revision, prompts, BF16, FlashAttention 2, max sequence length, package versions, and all five expected embedding variants.
• These are complete standard --all results: 551 evaluated task files, including the 13 NanoBEIR-en tasks retained for evaluation completeness. The leaderboard Overall summary excludes the configured 13 duplicate task copies and therefore uses 538 raw task results.

Checklist

• Result files are committed under hakari-results/nvidia__Nemotron-3-Embed-8B-BF16/.
• 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.