Replace Bekko small beta v8 with a25m results

#13

by hotchpotch - opened about 7 hours ago

base: refs/heads/main

←

from: refs/pr/13

Discussion Files changed

+1650

-1650

Replace Bekko small beta v8 with a25m resultsb46da8f5

hotchpotch

HAKARI-Bench org about 7 hours ago

•

edited about 7 hours ago

Replace Bekko small beta v8 with `hotchpotch/bekko-embedding-v1-a25m`

Replacement Scope

This PR removes the beta Bekko small embedding results and adds the official
hotchpotch/bekko-embedding-v1-a25m release results.

Action	Path	Files
Remove	`hakari-results/hotchpotch__bekko-embedding-small-beta-unir-v8/*/.json.xz`	551
Add	`hakari-results/hotchpotch__bekko-embedding-v1-a25m/*/.json.xz`	551

Summary

Field	Value
Model	`hotchpotch/bekko-embedding-v1-a25m`
Result directory	`hotchpotch__bekko-embedding-v1-a25m`
Target path	`hakari-results/hotchpotch__bekko-embedding-v1-a25m`
Result files	551 total, 551 `.json.xz`
Evaluation method	dense
Overall nDCG@10	0.5855
Overall score units	369 grouped units from 538 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	hotchpotch/bekko-embedding-v1-a25m	hotchpotch/bekko-embedding-small-beta-unir-v8 (384 dims)	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	0.5855	0.5829	0.5978	0.6322	0.5858	0.5190	0.4832
NanoMMTEB-v2	0.5065	0.5224	0.5581	0.5590	0.4846	0.4455	0.4550
NanoRTEB	0.5901	0.5910	0.6713	0.7005	0.5365	0.4711	0.3553
MNanoBEIR	0.5551	0.5489	0.5490	0.6052	0.5552	0.5099	0.4646
NanoBIRCO	0.2969	0.2873	0.3070	0.3526	0.2617	0.1613	0.2693
NanoMLDR	0.6053	0.6089	0.6239	0.5384	0.6621	0.3920	0.7396
NanoLongEmbed	0.7077	0.7178	0.7232	0.6680	0.6527	0.5014	0.8217
NanoDAPFAM	0.2831	0.2801	0.3018	0.3179	0.2406	0.2380	0.2400
NanoCoIR	0.7797	0.7838	0.8601	0.8777	0.6924	0.6915	0.5436
NanoIFIR	0.3233	0.3285	0.3364	0.3893	0.2391	0.2152	0.2761
NanoLaw	0.5650	0.5740	0.6075	0.6370	0.5597	0.4790	0.6854
NanoMedical	0.4942	0.5111	0.5694	0.5803	0.5371	0.5055	0.4145
NanoRARb	0.2297	0.2418	0.2689	0.2889	0.2343	0.2240	0.1359
NanoBRIGHT	0.3376	0.3282	0.3885	0.4284	0.2941	0.1758	0.2790
NanoCodeRAG	0.8252	0.8209	0.8712	0.9139	0.7155	0.7464	0.5823
NanoChemTEB	0.7643	0.7845	0.8035	0.7980	0.7777	0.8081	0.7012
NanoR2MED	0.2761	0.2870	0.3180	0.3630	0.2088	0.1099	0.2094
NanoBuiltBench	0.4876	0.4758	0.5129	0.5277	0.4248	0.4291	0.3958
NanoCMTEB	0.7022	0.7085	0.7982	0.8052	0.7591	0.6999	0.6003
NanoIndicQA	0.7362	0.7205	0.6413	0.7056	0.7586	0.7009	0.5653
NanoMuPLeR	0.8411	0.8025	0.7122	0.8388	0.8912	0.7837	0.7994
NanoMTEB-v2	0.5810	0.5751	0.6372	0.6450	0.5726	0.5348	0.5028
NanoMTEB-Dutch	0.5809	0.5810	0.5686	0.6213	0.5863	0.5287	0.4673
NanoMTEB-French	0.5505	0.5539	0.5771	0.6377	0.5527	0.4702	0.4261
NanoMTEB-German	0.6350	0.6189	0.6298	0.6536	0.6189	0.5711	0.5522
NanoJMTEB-v2	0.7863	0.7711	0.7732	0.8008	0.7906	0.7165	0.7465
NanoMTEB-Korean	0.7853	0.7624	0.7792	0.8246	0.8183	0.7668	0.6743
NanoFaMTEB-v2	0.6562	0.6545	0.6338	0.6882	0.6652	0.6135	0.5651
NanoMTEB-Polish	0.4697	0.4655	0.4738	0.5316	0.4999	0.4365	0.3424
NanoRuMTEB	0.8822	0.8823	0.8622	0.9121	0.9169	0.8643	0.7089
NanoMTEB-Scandinavian	0.7641	0.7580	0.6981	0.7596	0.7740	0.7029	0.6091
NanoMTEB-Spanish	0.5508	0.5558	0.5662	0.6292	0.5624	0.4848	0.3679
NanoMTEB-Thai	0.7319	0.7329	0.7455	0.7670	0.7672	0.7107	0.5216
NanoVNMTEB	0.5509	0.5438	0.5717	0.6066	0.5616	0.5197	0.4571
NanoMTEB-Misc	0.7542	0.7531	0.7629	0.8011	0.7766	0.6423	0.4939
NanoMIRACL	0.7910	0.7799	0.7879	0.8351	0.8475	0.7871	0.5715

Overall nDCG@10

Overall component	nDCG@10	Score units	Raw task results
NanoMMTEB-v2	0.5065	18	18
NanoRTEB	0.5901	14	14
MNanoBEIR	0.5551	13	182
NanoBIRCO	0.2969	5	5
NanoMLDR	0.6053	13	13
NanoLongEmbed	0.7077	6	6
NanoDAPFAM	0.2831	12	12
NanoCoIR	0.7797	10	10
NanoIFIR	0.3233	4	4
NanoLaw	0.5650	4	4
NanoMedical	0.4942	7	7
NanoRARb	0.2297	14	14
NanoBRIGHT	0.3376	20	20
NanoCodeRAG	0.8252	4	4
NanoChemTEB	0.7643	3	3
NanoR2MED	0.2761	8	8
NanoBuiltBench	0.4876	2	2
NanoCMTEB	0.7022	8	8
NanoIndicQA	0.7362	11	11
NanoMuPLeR	0.8411	14	14
NanoMTEB-v2	0.5810	10	10
NanoMTEB-Dutch	0.5809	27	27
NanoMTEB-French	0.5505	8	8
NanoMTEB-German	0.6350	5	5
NanoJMTEB-v2	0.7863	11	11
NanoMTEB-Korean	0.7853	5	5
NanoFaMTEB-v2	0.6562	17	17
NanoMTEB-Polish	0.4697	14	14
NanoRuMTEB	0.8822	3	3
NanoMTEB-Scandinavian	0.7641	7	7
NanoMTEB-Spanish	0.5508	7	7
NanoMTEB-Thai	0.7319	9	9
NanoVNMTEB	0.5509	26	26
NanoMTEB-Misc	0.7542	12	12
NanoMIRACL	0.7910	18	18

Reproducibility

Field	Value
Model source	`hotchpotch/bekko-embedding-v1-a25m`
Model revision	`f2e4e3555fd9452984578ce112e72ee64e25ddd8`
Dataset revision(s)	`01736efbaa96f020c2a4d996efdacc18071e2fcb`, `017849a95097eea984680cbab35972f8d3812376`, `0f3a6f43b8a26a9b8c8d5f31b09bd60dc4cd572d`, `1726763179e1e114ad9ffcdc7262923471e8ecc8`, `175ff423246cdbca9c3a992c4d68d312701b3f2a`, ... (48 total)
Evaluated at UTC	`2026-06-25T06:05:33.305496+00:00` to `2026-06-25T07:46:48.013635+00:00`
Generated at UTC	`2026-06-25T06:05:33.479256+00:00` to `2026-06-25T07:46:48.013649+00:00`
dtype	bf16
device	cuda:0
batch size	8
attention implementation	sdpa
trust remote code	False
max sequence length	8192
candidate ranking	reranking_hybrid
rerank top-k	not recorded
query prompt name	query
document prompt name	passage
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	`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

uv run --group tf4-fa2 hakari-bench evaluate from-model-card \
  --model-card config/model_cards/hotchpotch__bekko-embedding-v1-a25m.yaml \
  --output-dir output/hakari-results/hotchpotch__bekko-embedding-v1-a25m

Submitter Notes

Evaluated as a dense SentenceTransformer model from the local model card.
Runtime used bf16, sdpa, batch size 8, max sequence length 8192, and trust_remote_code=False.
Prompt names are query for queries and passage for documents.
Dense truncation variants were included for 64, 128, and 256 dimensions.
Full expected result coverage was audited after the run: 551 present, 0 missing, 0 unreadable, 0 metadata mismatches.
This PR intentionally replaces hotchpotch/bekko-embedding-small-beta-unir-v8 with the official hotchpotch/bekko-embedding-v1-a25m release.

Checklist

Result files are committed under hakari-results/hotchpotch__bekko-embedding-v1-a25m/.
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 merged about 7 hours ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment