Fix Qwen3-Embedding-8B bf16 code-task results
Fix Qwen3-Embedding-8B code-task results with bf16 SentenceTransformers reruns
Summary
This PR replaces 10 Qwen/Qwen3-Embedding-8B task results that were affected by fp16 numerical instability.
The previous results were generated with dtype=fp16 through the TEI custom backend. Follow-up checks showed that Qwen/Qwen3-Embedding-8B produces NaN document embeddings on the affected code-oriented tasks when run in fp16. For NanoDS1000, all document embeddings became NaN and the top-100 retrieval list collapsed to corpus-id order for every query.
These replacement files were rerun with SentenceTransformers directly using dtype=bf16, flash_attention_2, and the model's query / document prompt names.
Replacement Files
| Benchmark | Task | Old 8B | New 8B bf16/ST | 4B reference | Best distance |
|---|---|---|---|---|---|
| NanoRTEB | NanoDS1000 | 0.050168 | 0.846710 | 0.831132 | dot |
| NanoRTEB | NanoMBPP | 0.127167 | 0.894907 | 0.898194 | dot |
| NanoRTEB | NanoHumanEval | 0.316456 | 0.985156 | 0.981163 | cosine |
| NanoRTEB | NanoApps | 0.317500 | 0.872072 | 0.869746 | cosine |
| NanoRTEB | NanoFreshStack | 0.110894 | 0.478514 | 0.450492 | dot |
| NanoCodeRAG | NanoCodeRAGProgrammingSolutions | 0.263605 | 0.870437 | 0.878626 | cosine |
| NanoCoIR | NanoApps | 0.312500 | 0.872072 | 0.869746 | cosine |
| NanoCoIR | NanoCodeTransOceanDL | 0.189886 | 0.538475 | 0.553990 | dot |
| NanoRARb | NanoRARbCode | 0.446010 | 0.851631 | 0.936625 | cosine |
| NanoBRIGHT | NanoBrightLeetcode | 0.180871 | 0.411045 | 0.400893 | cosine |
Scores are nDCG@10. Old 8B and 4B reference scores are from the latest leaderboard DuckDB used for this audit.
Validation Notes
- Reproduced the bad
NanoDS1000behavior with SentenceTransformers direct loading underfp16: query embeddings were finite, but all 997 document embeddings contained NaN values with bothflash_attention_2andsdpa. - Re-ran the affected tasks with
bf16; document embeddings were finite and retrieval rankings varied by query. - Checked additional low-scoring 8B tasks (
NanoBrightPony,NanoSpartQA,NanoWinoGrande); bf16 did not materially improve them, so they are not included here. - Excluded
NanoBrightTheoremQATheoremsandlembpasskeyfrom this PR because bf16 reruns did not materially improve those results.
Reproducibility
| Field | Value |
|---|---|
| Model | Qwen/Qwen3-Embedding-8B |
| Model revision | 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af |
| Evaluation method | dense |
| Backend | SentenceTransformers |
| dtype | bf16 |
| attention implementation | flash_attention_2 |
| query prompt name | query |
| document prompt name | document |
| batch size | 8 |
| retrieval score device | cuda |
| embedding variants | disabled for these targeted replacement runs |
| candidate ranking | reranking_hybrid |
| Python | 3.12.11 |
| torch | 2.9.0 |
| transformers | 5.12.1 |
| sentence-transformers | 5.4.1 |
| datasets | 4.8.4 |
| GPU | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
Representative command shape:
CUDA_VISIBLE_DEVICES=0 uv run hakari-bench evaluate dense \
--model Qwen/Qwen3-Embedding-8B \
--model-revision 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af \
--dtype bf16 \
--attn-implementation flash_attention_2 \
--query-prompt-name query \
--document-prompt-name document \
--batch-size 8 \
--retrieval-score-device cuda \
--no-default-embedding-variants \
--result-format json \
--overwrite \
--dataset NanoRTEB \
--split NanoDS1000
The same options were used for the listed task groups, split across two GPUs.
Checklist
- Result files are under
hakari-results/Qwen__Qwen3-Embedding-8B/. - Result files are compressed
.json.xz. - No DuckDB files, HTML reports, caches, or scratch artifacts are included.
- JSON metadata records model revision, dataset revision, runtime configuration, and package versions.