Add results for Qwen/Qwen3-Embedding-8B
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.
| Model | Files | Macro nDCG@10 | Delta vs Qwen3-8B |
|---|---|---|---|
| Qwen/Qwen3-Embedding-8B | 551 | 0.6477265472 | +0.0000000000 |
| Qwen/Qwen3-Embedding-4B | 551 | 0.6455653533 | -0.0021611939 |
| Qwen/Qwen3-Embedding-0.6B | 551 | 0.5816140254 | -0.0661125218 |
| voyageai/voyage-4-nano | 551 | 0.6123061303 | -0.0354204169 |
| jinaai/jina-embeddings-v5-text-small | 551 | 0.6225883504 | -0.0251381968 |
| jinaai/jina-embeddings-v5-text-nano | 551 | 0.6124699163 | -0.0352566308 |
| BAAI/bge-m3 | 551 | 0.5741723723 | -0.0735541749 |
| bm25 | 551 | 0.4764072012 | -0.1713193460 |
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
| Group | Qwen3-8B | Qwen3-4B | voyage-4-nano | jina-v5-small | jina-v5-nano | BM25 | 8B - 4B | 8B - jina-nano |
|---|---|---|---|---|---|---|---|---|
| NanoMIRACL | 0.8496211685 | 0.8473776577 | 0.8014773444 | 0.8351453281 | 0.8315929532 | 0.5715253529 | +0.0022435108 | +0.0180282153 |
| NanoMLDR | 0.6912418219 | 0.6899641105 | 0.6339616956 | 0.5384091806 | 0.5671462117 | 0.7396439004 | +0.0012777114 | +0.1240956102 |
| NanoLongEmbed | 0.7462750879 | 0.7089644932 | 0.7355714525 | 0.6679654548 | 0.6266947358 | 0.8216763353 | +0.0373105947 | +0.1195803521 |
| NanoCoIR | 0.8022052403 | 0.8912617925 | 0.9027434884 | 0.8776845925 | 0.8600881419 | 0.5436265129 | -0.0890565522 | -0.0578829016 |
| NanoCodeRAG | 0.7575261020 | 0.9082610773 | 0.8766514961 | 0.9139396518 | 0.8704432248 | 0.5822803815 | -0.1507349753 | -0.1129171228 |
| NanoCMTEB | 0.8475203191 | 0.8435121275 | 0.7225281484 | 0.8052126706 | 0.7925984714 | 0.6002869462 | +0.0040081916 | +0.0549218477 |
| NanoLaw | 0.7080011798 | 0.6941212219 | 0.6494951431 | 0.6368341830 | 0.6230414396 | 0.5865332646 | +0.0138799579 | +0.0849597402 |
| NanoMedical | 0.6341795944 | 0.6177364913 | 0.5791765294 | 0.5803253721 | 0.5630679925 | 0.4364373540 | +0.0164431031 | +0.0711116019 |
| NanoBRIGHT | 0.4431184334 | 0.4438583327 | 0.4197619420 | 0.4283726302 | 0.4067276145 | 0.2790230035 | -0.0007398993 | +0.0363908189 |
| NanoRARb | 0.2794227075 | 0.2980498703 | 0.2945694488 | 0.2783487666 | 0.2695256005 | 0.1535945993 | -0.0186271628 | +0.0098971070 |
| NanoRTEB | 0.5249224280 | 0.7366341242 | 0.7450913667 | 0.7005405582 | 0.6764494008 | 0.3552568205 | -0.2117116962 | -0.1515269728 |
| NanoMuPLeR | 0.8961995559 | 0.8699589473 | 0.9127428408 | 0.8388165609 | 0.8342510146 | 0.7993851488 | +0.0262406086 | +0.0619485413 |
| NanoJMTEB-v2 | 0.8394951843 | 0.8273576341 | 0.7614611998 | 0.8007533856 | 0.7919040836 | 0.7465132991 | +0.0121375502 | +0.0475911007 |
| NanoRuMTEB | 0.9294912833 | 0.9244229431 | 0.8901581727 | 0.9120929665 | 0.9051484406 | 0.7089462998 | +0.0050683402 | +0.0243428427 |
| NanoVNMTEB | 0.6455178040 | 0.6301309550 | 0.5896446890 | 0.6066202208 | 0.5965730661 | 0.4570613577 | +0.0153868490 | +0.0489447379 |
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, andrescore: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
| Group | Files | Mean nDCG@10 |
|---|---|---|
| NanoBEIR-ar | 13 | 0.5930131434 |
| NanoBEIR-de | 13 | 0.6294201237 |
| NanoBEIR-en | 13 | 0.7039536137 |
| NanoBEIR-es | 13 | 0.6345901361 |
| NanoBEIR-fr | 13 | 0.6287032571 |
| NanoBEIR-it | 13 | 0.6357913379 |
| NanoBEIR-ja | 13 | 0.6259177034 |
| NanoBEIR-ko | 13 | 0.6202590470 |
| NanoBEIR-no | 13 | 0.6145164113 |
| NanoBEIR-pt | 13 | 0.6334985261 |
| NanoBEIR-sr | 13 | 0.6085335591 |
| NanoBEIR-sv | 13 | 0.6235139721 |
| NanoBEIR-th | 13 | 0.6117256052 |
| NanoBEIR-vi | 13 | 0.6389758423 |
| NanoBIRCO | 5 | 0.4041566142 |
| NanoBRIGHT | 20 | 0.4431184334 |
| NanoBuiltBench | 2 | 0.5727234024 |
| NanoCMTEB | 8 | 0.8475203191 |
| NanoChemTEB | 3 | 0.8316830009 |
| NanoCoIR | 10 | 0.8022052403 |
| NanoCodeRAG | 4 | 0.7575261020 |
| NanoDAPFAM | 12 | 0.3340586946 |
| NanoFaMTEB-v2 | 17 | 0.7166151250 |
| NanoIFIR | 7 | 0.5605404173 |
| NanoIndicQA | 11 | 0.7726695999 |
| NanoJMTEB-v2 | 11 | 0.8394951843 |
| NanoLaw | 8 | 0.7080011798 |
| NanoLongEmbed | 6 | 0.7462750879 |
| NanoMIRACL | 18 | 0.8496211685 |
| NanoMLDR | 13 | 0.6912418219 |
| NanoMMTEB-v2 | 18 | 0.5869471054 |
| NanoMTEB-Dutch | 27 | 0.6517744934 |
| NanoMTEB-French | 8 | 0.6778236240 |
| NanoMTEB-German | 5 | 0.6720045912 |
| NanoMTEB-Korean | 5 | 0.8376122401 |
| NanoMTEB-Misc | 12 | 0.8139304451 |
| NanoMTEB-Polish | 14 | 0.5913100219 |
| NanoMTEB-Scandinavian | 7 | 0.7786082929 |
| NanoMTEB-Spanish | 7 | 0.6731971282 |
| NanoMTEB-Thai | 9 | 0.8080316515 |
| NanoMTEB-v2 | 10 | 0.6936121378 |
| NanoMedical | 10 | 0.6341795944 |
| NanoMuPLeR | 14 | 0.8961995559 |
| NanoR2MED | 8 | 0.4663063223 |
| NanoRARb | 17 | 0.2794227075 |
| NanoRTEB | 14 | 0.5249224280 |
| NanoRuMTEB | 3 | 0.9294912833 |
| NanoVNMTEB | 26 | 0.6455178040 |
Notes
The evaluation was run through TEI on two separate GPUs/endpoints:
http://127.0.0.1:18101andhttp://127.0.0.1:18102.Endpoint distribution recorded in result JSON: 311 files on
18101, 240 files on18102.NanoLongEmbedwas 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
e9428c5fbbcddd4eab0f4d98ac08abd3374357f4backfillsconfig.query_prompt,config.document_prompt, and TEI backend prompt metadata fromconfig.model_loader_kwargs; scores and evaluation artifacts are unchanged.Sanity check:
NanoRTEB/NanoDS1000was 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.xzresult 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.