Add HAKARI-Bench results for e5-v2 dense models and mxbai reranker

#6
by hotchpotch - opened

Add HAKARI-Bench results for e5-v2 dense models and mxbai reranker

This PR adds .json.xz benchmark result files for four models under hakari-results/. It intentionally excludes local model-card YAML, DuckDB files, caches, HTML reports, and scratch artifacts.

Submitted models

Model Method Result files Notes
intfloat/e5-large-v2 dense 551 query prompt query: and document prompt passage: ; batch size 128; dtype bf16; attention sdpa; max sequence length 512.
intfloat/e5-base-v2 dense 551 query prompt query: and document prompt passage: ; batch size 192; dtype bf16; attention sdpa; max sequence length 512.
intfloat/e5-small-v2 dense 551 query prompt query: and document prompt passage: ; batch size 256; dtype bf16; attention sdpa; max sequence length 512.
mixedbread-ai/mxbai-rerank-base-v2 reranker 551 reranker evaluation; candidate ranking reranking_hybrid; batch size varied across resumed runs (1, 4, 8); dtype bf16; attention sdpa; max sequence length 32768; evaluated with sentence-transformers 5.7.0.dev0 from git for the long-token scoring fix.

Reconstructed commands

uv run hakari-bench evaluate dense --model intfloat/e5-large-v2 --all --dtype bf16 --device cuda:0 --batch-size 128 --attn-implementation sdpa --query-prompt "query: " --document-prompt "passage: "
uv run hakari-bench evaluate dense --model intfloat/e5-base-v2 --all --dtype bf16 --device cuda:0 --batch-size 192 --attn-implementation sdpa --query-prompt "query: " --document-prompt "passage: "
uv run hakari-bench evaluate dense --model intfloat/e5-small-v2 --all --dtype bf16 --device cuda:0 --batch-size 256 --attn-implementation sdpa --query-prompt "query: " --document-prompt "passage: "
uv run hakari-bench evaluate reranker --model mixedbread-ai/mxbai-rerank-base-v2 --all --dtype bf16 --device cuda:0 --batch-size 4 --attn-implementation sdpa --model-max-seq-length 32768 --candidate-ranking reranking_hybrid

The mxbai reranker result set includes resumed runs with batch sizes 1, 4, and 8 as recorded in the JSON metadata. Dense result JSON records candidate_ranking=reranking_hybrid and five embedding evaluation entries per task, including the default embedding variants.

Validation

  • 2204 staged files total: 551 .json.xz files per model.
  • Non-.json.xz files under the submitted result directories: 0.
  • DuckDB / .duckdb.wal artifacts under the submitted result directories: 0.
  • Dense prompts checked in JSON: query_prompt="query: ", document_prompt="passage: ".
  • mxbai checked in JSON as method="reranker", max_seq_length=32768, and sentence-transformers=5.7.0.dev0.

Per-model details


Add HAKARI-Bench results for intfloat/e5-large-v2

Summary

Field Value
Model intfloat/e5-large-v2
Result directory intfloat__e5-large-v2
Target path hakari-results/intfloat__e5-large-v2
Result files 551 total, 551 .json.xz
Evaluation method dense
Overall nDCG@10 0.3627
Overall score units 369 grouped units from 538 raw task results

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.4060 18 18
NanoRTEB 0.5336 14 14
MNanoBEIR 0.3540 13 182
NanoBIRCO 0.2659 5 5
NanoMLDR 0.2457 13 13
NanoLongEmbed 0.5348 6 6
NanoDAPFAM 0.2810 12 12
NanoCoIR 0.7130 10 10
NanoIFIR 0.2449 4 4
NanoLaw 0.3476 4 4
NanoMedical 0.2994 7 7
NanoRARb 0.2921 14 14
NanoBRIGHT 0.2456 20 20
NanoCodeRAG 0.8246 4 4
NanoChemTEB 0.8100 3 3
NanoR2MED 0.1741 8 8
NanoBuiltBench 0.5158 2 2
NanoCMTEB 0.1595 8 8
NanoIndicQA 0.1111 11 11
NanoMuPLeR 0.5268 14 14
NanoMTEB-v2 0.5707 10 10
NanoMTEB-Dutch 0.4421 27 27
NanoMTEB-French 0.4257 8 8
NanoMTEB-German 0.5055 5 5
NanoJMTEB-v2 0.3114 11 11
NanoMTEB-Korean 0.1552 5 5
NanoFaMTEB-v2 0.1644 17 17
NanoMTEB-Polish 0.3011 14 14
NanoRuMTEB 0.4373 3 3
NanoMTEB-Scandinavian 0.6065 7 7
NanoMTEB-Spanish 0.3973 7 7
NanoMTEB-Thai 0.0934 9 9
NanoVNMTEB 0.3272 26 26
NanoMTEB-Misc 0.5027 12 12
NanoMIRACL 0.3465 18 18

Reproducibility

Field Value
Model source intfloat/e5-large-v2
Model revision f169b11e22de13617baa190a028a32f3493550b6
Dataset revision(s) 017849a95097eea984680cbab35972f8d3812376, 0a6b8e4feaac801f0748d2f77291e93ceb2cfdc1, 0c8fdb149eee31b8dd5dc17fc82e6795dd1e8681, 158ceac28e2468e55a56b3d056ccbe33e13aa8d8, 193d979abe245c7e7e6dec6e9ad6360cf98edbf9, ... (48 total)
Evaluated at UTC 2026-06-16T23:26:25.195908+00:00 to 2026-06-17T02:04:08.199915+00:00
Generated at UTC 2026-06-16T23:26:25.393192+00:00 to 2026-06-17T02:04:08.199929+00:00
dtype bf16
device cuda:0
batch size 128
attention implementation sdpa
trust remote code False
max sequence length 512
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.12 (main, Dec 9 2025, 19:02:36) [Clang 21.1.4 ]
Platform Linux-6.8.0-107-generic-x86_64-with-glibc2.39
torch 2.9.0
transformers 5.3.0
sentence-transformers 5.4.1
datasets 4.8.4
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

Checklist

  • Result files are committed under hakari-results/intfloat__e5-large-v2/.
  • 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.

Add HAKARI-Bench results for intfloat/e5-base-v2

Summary

Field Value
Model intfloat/e5-base-v2
Result directory intfloat__e5-base-v2
Target path hakari-results/intfloat__e5-base-v2
Result files 551 total, 551 .json.xz
Evaluation method dense
Overall nDCG@10 0.3411
Overall score units 369 grouped units from 538 raw task results

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.3993 18 18
NanoRTEB 0.5147 14 14
MNanoBEIR 0.3336 13 182
NanoBIRCO 0.2458 5 5
NanoMLDR 0.2253 13 13
NanoLongEmbed 0.5124 6 6
NanoDAPFAM 0.2785 12 12
NanoCoIR 0.6885 10 10
NanoIFIR 0.2254 4 4
NanoLaw 0.3111 4 4
NanoMedical 0.2851 7 7
NanoRARb 0.2698 14 14
NanoBRIGHT 0.2486 20 20
NanoCodeRAG 0.8139 4 4
NanoChemTEB 0.7830 3 3
NanoR2MED 0.1492 8 8
NanoBuiltBench 0.4916 2 2
NanoCMTEB 0.1516 8 8
NanoIndicQA 0.0853 11 11
NanoMuPLeR 0.4450 14 14
NanoMTEB-v2 0.5741 10 10
NanoMTEB-Dutch 0.4304 27 27
NanoMTEB-French 0.4037 8 8
NanoMTEB-German 0.4820 5 5
NanoJMTEB-v2 0.2792 11 11
NanoMTEB-Korean 0.0929 5 5
NanoFaMTEB-v2 0.1655 17 17
NanoMTEB-Polish 0.2783 14 14
NanoRuMTEB 0.3806 3 3
NanoMTEB-Scandinavian 0.5694 7 7
NanoMTEB-Spanish 0.3625 7 7
NanoMTEB-Thai 0.0907 9 9
NanoVNMTEB 0.2913 26 26
NanoMTEB-Misc 0.4551 12 12
NanoMIRACL 0.3253 18 18

Reproducibility

Field Value
Model source intfloat/e5-base-v2
Model revision f52bf8ec8c7124536f0efb74aca902b2995e5bcd
Dataset revision(s) 017849a95097eea984680cbab35972f8d3812376, 0a6b8e4feaac801f0748d2f77291e93ceb2cfdc1, 0c8fdb149eee31b8dd5dc17fc82e6795dd1e8681, 158ceac28e2468e55a56b3d056ccbe33e13aa8d8, 193d979abe245c7e7e6dec6e9ad6360cf98edbf9, ... (48 total)
Evaluated at UTC 2026-06-16T23:26:56.005431+00:00 to 2026-06-17T01:55:42.609508+00:00
Generated at UTC 2026-06-16T23:26:56.201238+00:00 to 2026-06-17T01:55:42.609522+00:00
dtype bf16
device cuda:0
batch size 192
attention implementation sdpa
trust remote code False
max sequence length 512
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.12 (main, Dec 9 2025, 19:02:36) [Clang 21.1.4 ]
Platform Linux-6.8.0-107-generic-x86_64-with-glibc2.39
torch 2.9.0
transformers 5.3.0
sentence-transformers 5.4.1
datasets 4.8.4
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

Checklist

  • Result files are committed under hakari-results/intfloat__e5-base-v2/.
  • 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.

Add HAKARI-Bench results for intfloat/e5-small-v2

Summary

Field Value
Model intfloat/e5-small-v2
Result directory intfloat__e5-small-v2
Target path hakari-results/intfloat__e5-small-v2
Result files 551 total, 551 .json.xz
Evaluation method dense
Overall nDCG@10 0.3094
Overall score units 369 grouped units from 538 raw task results

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.3874 18 18
NanoRTEB 0.4854 14 14
MNanoBEIR 0.3060 13 182
NanoBIRCO 0.2109 5 5
NanoMLDR 0.1955 13 13
NanoLongEmbed 0.5095 6 6
NanoDAPFAM 0.2689 12 12
NanoCoIR 0.6604 10 10
NanoIFIR 0.2117 4 4
NanoLaw 0.3048 4 4
NanoMedical 0.2779 7 7
NanoRARb 0.2256 14 14
NanoBRIGHT 0.1918 20 20
NanoCodeRAG 0.7724 4 4
NanoChemTEB 0.7711 3 3
NanoR2MED 0.1228 8 8
NanoBuiltBench 0.4704 2 2
NanoCMTEB 0.1465 8 8
NanoIndicQA 0.0666 11 11
NanoMuPLeR 0.3961 14 14
NanoMTEB-v2 0.5600 10 10
NanoMTEB-Dutch 0.3975 27 27
NanoMTEB-French 0.3646 8 8
NanoMTEB-German 0.4466 5 5
NanoJMTEB-v2 0.2527 11 11
NanoMTEB-Korean 0.0401 5 5
NanoFaMTEB-v2 0.0679 17 17
NanoMTEB-Polish 0.2664 14 14
NanoRuMTEB 0.1248 3 3
NanoMTEB-Scandinavian 0.5148 7 7
NanoMTEB-Spanish 0.3452 7 7
NanoMTEB-Thai 0.0835 9 9
NanoVNMTEB 0.2867 26 26
NanoMTEB-Misc 0.4096 12 12
NanoMIRACL 0.2949 18 18

Reproducibility

Field Value
Model source intfloat/e5-small-v2
Model revision ffb93f3bd4047442299a41ebb6fa998a38507c52
Dataset revision(s) 017849a95097eea984680cbab35972f8d3812376, 0a6b8e4feaac801f0748d2f77291e93ceb2cfdc1, 0c8fdb149eee31b8dd5dc17fc82e6795dd1e8681, 158ceac28e2468e55a56b3d056ccbe33e13aa8d8, 193d979abe245c7e7e6dec6e9ad6360cf98edbf9, ... (48 total)
Evaluated at UTC 2026-06-16T23:26:48.162430+00:00 to 2026-06-17T01:17:48.957333+00:00
Generated at UTC 2026-06-16T23:26:48.342123+00:00 to 2026-06-17T01:17:48.957351+00:00
dtype bf16
device cuda:0
batch size 256
attention implementation sdpa
trust remote code False
max sequence length 512
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.12 (main, Dec 9 2025, 19:02:36) [Clang 21.1.4 ]
Platform Linux-6.8.0-107-generic-x86_64-with-glibc2.39
torch 2.9.0
transformers 5.3.0
sentence-transformers 5.4.1
datasets 4.8.4
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

Checklist

  • Result files are committed under hakari-results/intfloat__e5-small-v2/.
  • 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.

Add HAKARI-Bench results for mixedbread-ai/mxbai-rerank-base-v2

Summary

Field Value
Model mixedbread-ai/mxbai-rerank-base-v2
Result directory mixedbread-ai__mxbai-rerank-base-v2
Target path hakari-results/mixedbread-ai__mxbai-rerank-base-v2
Result files 551 total, 551 .json.xz
Evaluation method reranker
Overall nDCG@10 0.5963
Overall score units 369 grouped units from 538 raw task results

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.5455 18 18
NanoRTEB 0.6076 14 14
MNanoBEIR 0.5567 13 182
NanoBIRCO 0.3729 5 5
NanoMLDR 0.6115 13 13
NanoLongEmbed 0.6754 6 6
NanoDAPFAM 0.2855 12 12
NanoCoIR 0.7916 10 10
NanoIFIR 0.3934 4 4
NanoLaw 0.5082 4 4
NanoMedical 0.5595 7 7
NanoRARb 0.3360 14 14
NanoBRIGHT 0.3552 20 20
NanoCodeRAG 0.8117 4 4
NanoChemTEB 0.8297 3 3
NanoR2MED 0.3351 8 8
NanoBuiltBench 0.5178 2 2
NanoCMTEB 0.7859 8 8
NanoIndicQA 0.6120 11 11
NanoMuPLeR 0.8906 14 14
NanoMTEB-v2 0.6270 10 10
NanoMTEB-Dutch 0.5434 27 27
NanoMTEB-French 0.6116 8 8
NanoMTEB-German 0.5963 5 5
NanoJMTEB-v2 0.7760 11 11
NanoMTEB-Korean 0.8325 5 5
NanoFaMTEB-v2 0.6362 17 17
NanoMTEB-Polish 0.4546 14 14
NanoRuMTEB 0.8741 3 3
NanoMTEB-Scandinavian 0.6537 7 7
NanoMTEB-Spanish 0.5879 7 7
NanoMTEB-Thai 0.7410 9 9
NanoVNMTEB 0.5856 26 26
NanoMTEB-Misc 0.7328 12 12
NanoMIRACL 0.7946 18 18

Reproducibility

Field Value
Model source mixedbread-ai/mxbai-rerank-base-v2
Model revision 3ea9d4dffa7d12a4f366be8e275c349de9fc9865
Dataset revision(s) 017849a95097eea984680cbab35972f8d3812376, 0a6b8e4feaac801f0748d2f77291e93ceb2cfdc1, 0c8fdb149eee31b8dd5dc17fc82e6795dd1e8681, 158ceac28e2468e55a56b3d056ccbe33e13aa8d8, 193d979abe245c7e7e6dec6e9ad6360cf98edbf9, ... (48 total)
Evaluated at UTC 2026-06-17T00:14:29.735639+00:00 to 2026-06-19T13:11:54.832614+00:00
Generated at UTC 2026-06-17T00:14:29.946677+00:00 to 2026-06-19T13:11:54.832635+00:00
dtype bf16
device cuda:0
batch size 1, 4, 8
attention implementation sdpa
trust remote code False
max sequence length 32768
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.12 (main, Dec 9 2025, 19:02:36) [Clang 21.1.4 ]
Platform Linux-6.8.0-107-generic-x86_64-with-glibc2.39
torch 2.9.1
transformers 5.10.2
sentence-transformers 5.7.0.dev0
datasets 5.0.0
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

Checklist

  • Result files are committed under hakari-results/mixedbread-ai__mxbai-rerank-base-v2/.
  • 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.
hotchpotch changed pull request status to open
hotchpotch changed pull request status to merged

Sign up or log in to comment