Add HAKARI-Bench results for Qwen/Qwen3-Embedding-4B

Summary

Core nDCG@10

Reproducibility

Command

PYTHONPATH=$PWD uv run hakari-bench evaluate dense \
  --model Qwen/Qwen3-Embedding-4B \
  --model-loader examples.custom_backends.tei_embedding:load_model \
  --model-loader-kwargs-json '{"endpoint":"http://127.0.0.1:8080","model":"Qwen/Qwen3-Embedding-4B","timeout":600}' \
  --all \
  --dtype fp16 \
  --attn-implementation flash_attention_2 \
  --query-prompt $'Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:' \
  --document-prompt '' \
  --embedding-variant truncate:2048,1536,1024,768,512,256,128,64,32 \
  --batch-size 16 \
  --results-dir tmp/tei-qwen3-embedding-4b-full-hakari

Submitter Notes

• These are standard --all dense results generated through a Text Embeddings Inference OpenAI embeddings-compatible backend.
• TEI served Qwen/Qwen3-Embedding-4B locally at http://127.0.0.1:8080 with image ghcr.io/huggingface/text-embeddings-inference:120-1.9, dtype float16, --max-batch-tokens 40960, --max-client-batch-size 128, and --payload-limit 64000000.
• Prompt settings follow the Qwen3 embedding retrieval guidance: query prefix Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: and no document prefix.
• The run included full-dimension results plus default int8,binary and rescore:int8,binary variants, and explicit truncation variants for 2048, 1536, 1024, 768, 512, 256, 128, 64, and 32 dimensions.
• The evaluation was resumed. Earlier attempts used batch sizes 64, 8, and 1 before increasing the TEI payload limit for long-document tasks; the final resumed command used batch size 16. Existing result files were reused, so per-task JSON preserves the batch size used when each task file was produced.
• Final CLI summary: primary metric ndcg@10, primary metric mean across all evaluated result files 0.6455653533036781, evaluated count 339, cache hit count 212.

Checklist

• Result files are committed under hakari-results/Qwen__Qwen3-Embedding-4B/.
• 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.