Add benchmark results for codefuse-ai/F2LLM-v2-0.6B

#50
HAKARI-Bench org

Add HAKARI-Bench results for codefuse-ai/F2LLM-v2-0.6B

Summary

Field Value
Model codefuse-ai/F2LLM-v2-0.6B
Result directory codefuse-ai__F2LLM-v2-0.6B
Target path hakari-results/codefuse-ai__F2LLM-v2-0.6B
Result files 563 total, 563 .json.xz
Evaluation method dense
Overall nDCG@10 0.5969
Overall score units 381 grouped units from 550 raw task results

DuckDB Nano-set Comparison

Computed from DuckDB task_results with the same Overall grouping as this PR body. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Overall component codefuse-ai/F2LLM-v2-0.6B Qwen/Qwen3-Embedding-0.6B (1024 dims) jinaai/jina-embeddings-v5-text-small (1024 dims) BAAI/bge-m3 (1024 dims) intfloat/multilingual-e5-small (384 dims) bm25
Overall 0.5969 0.5944 0.6296 0.5820 0.5149 0.4806
NanoMMTEB-v2 0.5103 0.5581 0.5590 0.4846 0.4455 0.4550
NanoRTEB 0.6792 0.6713 0.7005 0.5365 0.4711 0.3553
MNanoBEIR 0.5589 0.5509 0.6077 0.5575 0.5117 0.4646
NanoBIRCO 0.3050 0.3070 0.3526 0.2617 0.1613 0.2693
NanoMLDR 0.5315 0.6239 0.5384 0.6621 0.3920 0.7396
NanoLongEmbed 0.7017 0.7232 0.6680 0.6527 0.5014 0.8217
NanoDAPFAM 0.2940 0.3018 0.3179 0.2406 0.2380 0.2400
NanoCoIR 0.8999 0.8601 0.8777 0.6924 0.6915 0.5436
NanoIFIR 0.3311 0.3364 0.3893 0.2391 0.2152 0.2761
NanoLaw 0.5570 0.6075 0.6370 0.5597 0.4790 0.6854
NanoMedical 0.5154 0.5694 0.5803 0.5371 0.5055 0.4145
NanoRARb 0.2998 0.2689 0.2889 0.2343 0.2240 0.1359
NanoBRIGHT 0.3685 0.3885 0.4284 0.2941 0.1758 0.2790
NanoCodeRAG 0.8727 0.8712 0.9139 0.7155 0.7464 0.5823
NanoChemTEB 0.7685 0.8035 0.7980 0.7777 0.8081 0.7012
NanoR2MED 0.3509 0.3180 0.3630 0.2088 0.1099 0.2094
NanoBuiltBench 0.4993 0.5129 0.5277 0.4248 0.4291 0.3958
NanoCMTEB 0.7799 0.7982 0.8052 0.7591 0.6999 0.6003
NanoIndicQA 0.6627 0.6413 0.7056 0.7586 0.7009 0.5653
NanoMuPLeR 0.7435 0.7122 0.8388 0.8912 0.7837 0.7994
NanoMTEB-v2 0.6041 0.6372 0.6450 0.5726 0.5348 0.5028
NanoMTEB-Dutch 0.5925 0.5686 0.6213 0.5863 0.5287 0.4673
NanoMTEB-French 0.6008 0.5771 0.6377 0.5527 0.4702 0.4261
NanoMTEB-German 0.5997 0.6298 0.6536 0.6189 0.5711 0.5522
NanoJMTEB-v2 0.7803 0.7732 0.8008 0.7906 0.7165 0.7465
NanoMTEB-Korean 0.7625 0.7792 0.8246 0.8183 0.7668 0.6743
NanoFaMTEB-v2 0.6685 0.6338 0.6882 0.6652 0.6135 0.5651
NanoMTEB-Polish 0.5111 0.4738 0.5316 0.4999 0.4365 0.3424
NanoMTEB-BR 0.6208 0.6247 0.6595 0.6540 0.5224 0.5178
NanoSSRB 0.3878 0.3464 0.4340 0.2665 0.2511 0.2821
NanoRuMTEB 0.8796 0.8622 0.9121 0.9169 0.8643 0.7089
NanoMTEB-Scandinavian 0.7532 0.6981 0.7596 0.7740 0.7029 0.6091
NanoMTEB-Spanish 0.6023 0.5662 0.6292 0.5624 0.4848 0.3679
NanoMTEB-Thai 0.7590 0.7455 0.7670 0.7672 0.7107 0.5216
NanoVNMTEB 0.5782 0.5717 0.6066 0.5616 0.5197 0.4571
NanoMTEB-Misc 0.7550 0.7629 0.8011 0.7766 0.6423 0.4939
NanoMIRACL 0.7873 0.7879 0.8351 0.8475 0.7871 0.5715

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.5103 18 18
NanoRTEB 0.6792 14 14
MNanoBEIR 0.5589 13 182
NanoBIRCO 0.3050 5 5
NanoMLDR 0.5315 13 13
NanoLongEmbed 0.7017 6 6
NanoDAPFAM 0.2940 12 12
NanoCoIR 0.8999 10 10
NanoIFIR 0.3311 4 4
NanoLaw 0.5570 4 4
NanoMedical 0.5154 7 7
NanoRARb 0.2998 14 14
NanoBRIGHT 0.3685 20 20
NanoCodeRAG 0.8727 4 4
NanoChemTEB 0.7685 3 3
NanoR2MED 0.3509 8 8
NanoBuiltBench 0.4993 2 2
NanoCMTEB 0.7799 8 8
NanoIndicQA 0.6627 11 11
NanoMuPLeR 0.7435 14 14
NanoMTEB-v2 0.6041 10 10
NanoMTEB-Dutch 0.5925 27 27
NanoMTEB-French 0.6008 8 8
NanoMTEB-German 0.5997 5 5
NanoJMTEB-v2 0.7803 11 11
NanoMTEB-Korean 0.7625 5 5
NanoFaMTEB-v2 0.6685 17 17
NanoMTEB-Polish 0.5111 14 14
NanoMTEB-BR 0.6208 6 6
NanoSSRB 0.3878 6 6
NanoRuMTEB 0.8796 3 3
NanoMTEB-Scandinavian 0.7532 7 7
NanoMTEB-Spanish 0.6023 7 7
NanoMTEB-Thai 0.7590 9 9
NanoVNMTEB 0.5782 26 26
NanoMTEB-Misc 0.7550 12 12
NanoMIRACL 0.7873 18 18

Reproducibility

Field Value
Model source codefuse-ai/F2LLM-v2-0.6B
Model revision 8786315a8711c242ee03ec67c74dd9ad0a61e2cf
Dataset revision(s) 00541a0fce4048057fb7ddec30d37155a5c23d95, 01736efbaa96f020c2a4d996efdacc18071e2fcb, 017849a95097eea984680cbab35972f8d3812376, 0f3a6f43b8a26a9b8c8d5f31b09bd60dc4cd572d, 1726763179e1e114ad9ffcdc7262923471e8ecc8, ... (50 total)
Evaluated at UTC 2026-11-04T10:38:02.835810+00:00 to 2026-11-04T15:29:20.958287+00:00
Generated at UTC 2026-11-04T10:38:02.835835+00:00 to 2026-11-04T15:29:20.958308+00:00
dtype bf16
device cuda:0
batch size 1, 8
attention implementation flash_attention_2
trust remote code True
max sequence length 40960
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.15 (main, Oct 21 2026, 17:47:06) [Clang 23.1.3 ]
Platform Linux-6.8.0-147-generic-x86_64-with-glibc2.39
torch 2.9.0
transformers 4.57.6
sentence-transformers 5.4.1
datasets 4.8.4
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

Command

# Python argv manifest (one call per pinned dataset/task; two GPU workers):
command=['evaluate','dense','--model',plan['model'],'--model-revision',plan['model_revision'],
  '--dataset',item['dataset'],'--split',item['task'],'--dataset-revision',plan['dataset_revisions'][item['dataset']],
  '--dtype','bf16','--trust-remote-code','--attn-implementation','flash_attention_2','--model-max-seq-length','40960',
  '--device','cuda:0','--batch-size',str(batch),'--query-prompt',plan['query_prompt'],
  '--embedding-variant','truncate:512,256,128,64,32,16,8','--results-dir',str(out)]

Submitter Notes

  • Official retrieval prompt (actual newline), BF16/FlashAttention 2, native 40960-token context. Native dimension 1024 plus MRL 512/256/128/64/32/16/8, with default quantized/rescore variants (40 categories). Batch 8 for short tasks, 1 for long tasks. Official multilingual example validation max absolute difference 0.00593 from rounded reference scores.
  • Both GPUs used a shared small-first task queue; all tasks completed without errors. Context was not shortened. Model reused and warmed in each worker.
  • Standard full scope: 563 task files, including 13 overlapping NanoBEIR-en results retained for completeness. All 50 dataset revisions pinned and recorded in per-task metadata.

Checklist

  • Result files are committed under hakari-results/codefuse-ai__F2LLM-v2-0.6B/.
  • 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.
  • Overall 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.

Code, model cards and reusable loader: https://github.com/hakari-bench/hakari-bench/pull/19

hotchpotch changed pull request status to merged

Sign up or log in to comment