Add HAKARI-Bench results for hotchpotch/bekko-embedding-v1-a8m

Summary

Core nDCG@10

Reproducibility

Command

CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 hakari-bench evaluate dense \
  --model hotchpotch/bekko-embedding-v1-a8m \
  --all \
  --dtype bf16 \
  --device cuda:0 \
  --flash-attn2 \
  --query-prompt-name query \
  --document-prompt-name passage \
  --embedding-variant truncate:384,256,128,64 \
  --batch-size 16

Submitter Notes

• Standard dense --all run with the model card's retrieval prompts:
  query for queries and passage for documents. The passage prompt name is
  an alias for the documented passage: document prefix.
• Run used Transformers 4.x with Flash Attention 2 through the tf4-fa2 uv
  dependency group, bf16, cuda:0, max sequence length 8192, and batch size
• The model supports truncation dimensions 384,256,128,64. The 384-dimensional
  truncate variant matches the base embedding dimension and was skipped by
  HAKARI-Bench as a no-op duplicate; base rows represent the 384-dimensional
  result.
• Dense default embedding variants were left enabled, so full-dimension
  quantized/rescore variants and truncate x quantized/rescore variants were
  generated where applicable.

Checklist

• Result files are committed under hakari-results/hotchpotch__bekko-embedding-v1-a8m/.
• 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.
• Core 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.