Add results for Qwen/Qwen3-Embedding-8B

Summary

This PR adds HAKARI-Bench Nano-set dense retrieval results for Qwen/Qwen3-Embedding-8B.

• Result path: hakari-results/Qwen__Qwen3-Embedding-8B
• Result files: 551 .json.xz
• Evaluation method: dense embedding retrieval
• Primary metric: nDCG@10
• Nano-set macro mean over submitted task files: 0.6477265472
• Coverage audit: 551/551 expected task files present, missing 0
• Variant audit: all 551 files contain 55 embedding evaluations

Model And Runtime

• Model: Qwen/Qwen3-Embedding-8B
• Revision: 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af
• Backend: Hugging Face Text Embeddings Inference through examples.custom_backends.tei_embedding:load_model
• TEI image: ghcr.io/huggingface/text-embeddings-inference:120-1.9
• TEI dtype: float16
• Result metadata dtype: fp16
• Attention implementation recorded in results: flash_attention_2
• Similarity: cosine
• Batch size: 8
• Retrieval score device: CPU
• Candidate ranking: reranking_hybrid
• trust_remote_code: false
• Model max sequence length override: none recorded in result JSON
• TEI max batch tokens used for the served endpoints: 40960

Prompt Settings

The TEI custom loader records the following prompt settings in config.model_loader_kwargs:

• Query prompt: Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:
• Document prompt: empty string
• Query prompt name: null
• Document prompt name: null

Comparison With Other Models

The table below compares base/full-dimension nDCG@10 task-file macro means over the submitted Nano-set files. Quantized, binary, rescore, and truncate variants are excluded here so the comparison is on each model's native embedding output.

Existing perplexity-ai/pplx-embed-v1-4B full-coverage PR data reports official Overall nDCG@10 = 0.6586 over 551 files. It is not yet present in the local DuckDB snapshot used for the task-file macro table above, so it is listed here as a prior PR reference rather than mixed into the same aggregation table.

Selected Group Deltas

High-level readout: Qwen3-8B is only slightly above Qwen3-4B on the base task-file macro (+0.0022), but it is clearly above voyage-4-nano (+0.0354), jina-v5-nano (+0.0353), jina-v5-small (+0.0251), bge-m3 (+0.0736), and BM25 (+0.1713) under this aggregation. Against Qwen3-4B, the visible gains are concentrated in NanoLongEmbed, NanoMuPLeR, NanoMedical, NanoLaw, NanoJMTEB-v2, NanoVNMTEB, and several multilingual groups. Qwen3-8B is weaker than Qwen3-4B on this base aggregation for NanoCodeRAG, NanoCoIR, NanoRTEB, NanoRARb, and slightly on NanoBRIGHT, so the 8B model is not a uniform upgrade across all task families.

Embedding Variants

The run used the dense default quantized/rescore variants plus the documented Qwen3 truncation grid:

• Truncation dimensions: 3072,2048,1536,1024,768,512,256,128,64,32
• Dense default variants: full-dimension int8, binary, rescore:int8, and rescore:binary
• Expanded variant coverage: 55 embedding evaluations per task file

Environment

• Python: 3.12.11
• Platform: Linux-6.17.0-1012-oem-x86_64-with-glibc2.39
• torch: 2.9.0
• transformers: 5.12.1
• sentence-transformers: 5.4.1
• datasets: 4.8.4
• numpy: 2.4.4
• scipy: 1.17.1
• CUDA: 12.8
• cuDNN: 91002
• GPUs: 2 x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition

Reconstructed Command

The run was split across two TEI endpoints on separate GPUs. The command shape was:

PYTHONPATH=$PWD uv run hakari-bench evaluate dense \
  --model Qwen/Qwen3-Embedding-8B \
  --model-revision 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af \
  --model-loader examples.custom_backends.tei_embedding:load_model \
  --model-loader-kwargs-json '{"endpoint":"http://127.0.0.1:18101 or 18102","model":"Qwen/Qwen3-Embedding-8B","query_prompt":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:","document_prompt":"","query_prompt_name":null,"document_prompt_name":null,"prompts":{"query":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:","document":""},"similarity_fn_name":"cosine","timeout":300}' \
  --dtype fp16 \
  --attn-implementation flash_attention_2 \
  --retrieval-score-device cpu \
  --batch-size 8 \
  --evaluation-scope standard \
  --embedding-variant truncate:3072,2048,1536,1024,768,512,256,128,64,32 \
  --results-dir output/qwen3-embedding-8b-tei-32k-cosine-all-nano \
  --result-format json.xz \
  --dataset ...

The TEI containers were started with:

text-embeddings-router \
  --dtype float16 \
  --model-id Qwen/Qwen3-Embedding-8B \
  --revision 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af \
  --max-batch-tokens 40960 \
  --max-client-batch-size 128 \
  --payload-limit 64000000

Nano-Set Group Means

Notes

• The evaluation was run through TEI on two separate GPUs/endpoints: http://127.0.0.1:18101 and http://127.0.0.1:18102.

• Endpoint distribution recorded in result JSON: 311 files on 18101, 240 files on 18102.

• NanoLongEmbed was retried after prefetching the missing Hugging Face parquet files into the local cache; the final submitted coverage is complete.

• No DuckDB files, viewer artifacts, caches, HTML reports, or scratch files are included in this PR.

• Metadata correction: commit e9428c5fbbcddd4eab0f4d98ac08abd3374357f4 backfills config.query_prompt, config.document_prompt, and TEI backend prompt metadata from config.model_loader_kwargs; scores and evaluation artifacts are unchanged.

• Sanity check: NanoRTEB/NanoDS1000 was rerun once with direct SentenceTransformers inference using the same Qwen3 retrieval prompt, fp16, flash_attention_2, batch size 8, and the same truncate grid. The direct run reproduced the TEI score exactly: nDCG@10 = 0.0501675665326694.

Checklist

• Submitted path contains only .json.xz result files.
• Result count matches the coverage audit.
• Model revision is pinned in result metadata.
• Runtime settings and prompts are recorded in result metadata.
• Dense variant coverage was audited.