Add HAKARI-Bench results for minishlab/potion-multilingual-128M

Evaluation coverage and default Overall

Full standard evaluation: 563 tasks x 5 conditions = 2,815 task-variant rows.
Each condition has 550 canonical Overall tasks; all 563 task artifacts are retained.
Default Overall micro nDCG@10 x 100: 34.99437094.
Base candidate-reranking micro nDCG@10 x 100: 38.01505988.

Summary

DuckDB Nano-set Comparison

Computed from the latest published DuckDB task_results using the template grouping (381 grouped units from 550 canonical tasks). The Overall row below is the grouped-unit mean on a 0–1 scale, not the default 550-task micro score of 34.99437094 on a 0–100 scale reported above. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Comparison database: hakari-bench/leaderboard_database revision bc869f95f6979026963fa7093d1b05fd899a15b2.

Grouped component nDCG@10

The template aggregates 381 grouped units; its grouped mean differs from the default Overall micro mean over 550 tasks.

Reproducibility

Command

# Equivalent full-target command; actual jobs partitioned the standard target
# by dataset across 12 workers on cuda:0 and cuda:1.
uv run hakari-bench evaluate dense \
  --model minishlab/potion-multilingual-128M \
  --model-revision 73908c3438cf03b6a01bcb9611d62b23d0726f08 \
  --all --dtype bf16 --device cuda:0 --retrieval-score-device cpu \
  --batch-size 4096 --results-dir output/potion-multilingual-128M-20260917

Submitter Notes

• The official checkpoint stores FP32 weights. Actual inference parameters and encoded output were verified as BF16; this is not an FP32 inference run.
• Native output is 256 dimensions. The official model card and checkpoint configuration do not document Matryoshka/prefix-truncation support or recommended lower dimensions, so no truncate variants were added. All four default quantization/rescore variants are included.
• Standard SentenceTransformers StaticEmbedding + Normalize backend, empty query/document prompts, no context-length override, no remote code, and no attention implementation (not applicable to this architecture).
• Post-encoding retrieval scoring uses the benchmark's CPU path. Candidate ranking is the default reranking_hybrid; all supplied candidates are reranked.
• One NanoArguAna smoke task was retained after validation; the remaining 562 tasks were partitioned by dataset. Every worker succeeded; no failures, OOM retries, or batch-size changes occurred. Thread counts were limited to one per worker.
• All 50 dataset revisions were pinned for the partitioned full run; the smoke task's resolved revision was verified to match. Exact revisions are recorded in each result JSON and listed below.
• Coverage, finite scores, model/dataset revisions, dtype, all five conditions, and retrieval/candidate ranking artifacts were audited. The local DuckDB independently confirms 563 tasks per condition and 550 Overall tasks.
• Model metadata was prepared and validated locally (40 model-card tests passed). This Dataset PR contains only result files; no GitHub PR is part of this submission.

{
  "hakari-bench/NanoBEIR-ar": "5c2e7952832d26a760796a2fcbdd5d9486f9ef7e",
  "hakari-bench/NanoBEIR-de": "175ff423246cdbca9c3a992c4d68d312701b3f2a",
  "hakari-bench/NanoBEIR-en": "d3962aa8efe48ed79044c5e155b848982667b4ba",
  "hakari-bench/NanoBEIR-es": "c05f7c8fa2578e3e8424e5f334e15b31640a5603",
  "hakari-bench/NanoBEIR-fr": "f57775ea2de74d275fdba0079ff8e8d3258d7afb",
  "hakari-bench/NanoBEIR-it": "f98eb62d9bbd2acd6099c99b18a1ddb6a821944b",
  "hakari-bench/NanoBEIR-ja": "5c1d5564643f9ca7a8c275688acf09fd940aa5f2",
  "hakari-bench/NanoBEIR-ko": "f7163a578dd87b9095c845371cdd7f4cc86af2bf",
  "hakari-bench/NanoBEIR-no": "a2ed691465a7b647c65b14db0547b5e6109ea3ed",
  "hakari-bench/NanoBEIR-pt": "0f3a6f43b8a26a9b8c8d5f31b09bd60dc4cd572d",
  "hakari-bench/NanoBEIR-sr": "b78aac98559de02f3b61de2c632b10ffad0d49e9",
  "hakari-bench/NanoBEIR-sv": "225ee695b2078612900dc77c175892cc3ead3c28",
  "hakari-bench/NanoBEIR-th": "d43a3c685818a0883a995162c7231b1f6e4355d7",
  "hakari-bench/NanoBEIR-vi": "ae6c498faed84ece5e10e82702cc26f006fec0ef",
  "hakari-bench/NanoBIRCO": "2796bcbef7f68779f597b5e213d84008aa27409b",
  "hakari-bench/NanoBRIGHT": "c192bca3345d3b1bb8736ce76c107cd5a55e616d",
  "hakari-bench/NanoBuiltBench": "4ab96ccaf9ad073982bee6d044a34eddd8340daf",
  "hakari-bench/NanoCMTEB": "d4de21e475c1aefa2799a56dc88e6a9916bde1e7",
  "hakari-bench/NanoChemTEB": "954f4b70773eb1d737b29e367384aa366a444608",
  "hakari-bench/NanoCoIR": "f6fa3c5c630d51870e260d00ebccdb32abb0afb6",
  "hakari-bench/NanoCodeRAG": "e37a708bc79ff069031849cbeb8f2623366b89da",
  "hakari-bench/NanoDAPFAM": "d444cdd6bdc676a5e2d702755ce9c8ed9fe795c5",
  "hakari-bench/NanoFaMTEB-v2": "e93f78c6b4bbbf08803f464600586820606296cb",
  "hakari-bench/NanoIFIR": "d17cefbebbe3ce26b9a4ed78d75c28c3d666d31c",
  "hakari-bench/NanoIndicQA": "6ad861f7d4cc2c3316e725978aaadbb53224e277",
  "hakari-bench/NanoJMTEB-v2": "322a48e91659b48c78583bb37367c62917d89776",
  "hakari-bench/NanoLaw": "47dce6a36e9ae44fab47d735c5dec8c16c946495",
  "hakari-bench/NanoLongEmbed": "ec27cccf70e6ac6afef6678b7645f0d2d02d3e1e",
  "hakari-bench/NanoMIRACL": "ea97b8cca4e0e0e19910c43e7f3ee9a60fef3a74",
  "hakari-bench/NanoMLDR": "c632cfe56f83c2361a4754a788de39ab890f786b",
  "hakari-bench/NanoMMTEB-v2": "01736efbaa96f020c2a4d996efdacc18071e2fcb",
  "hakari-bench/NanoMTEB-BR": "00541a0fce4048057fb7ddec30d37155a5c23d95",
  "hakari-bench/NanoMTEB-Dutch": "232f47f59d39a34dcaa69f1cb024d4a2edac8044",
  "hakari-bench/NanoMTEB-French": "fb12ec58d36aed39e568327f155199fbce38817e",
  "hakari-bench/NanoMTEB-German": "9dca935e210750bcf7dcda9b97869934238571fb",
  "hakari-bench/NanoMTEB-Korean": "7cba9121a3231a8f60b41975560f27d105c2ee8c",
  "hakari-bench/NanoMTEB-Misc": "f138caeb0ad78a7a1e17152f604c855995774127",
  "hakari-bench/NanoMTEB-Polish": "4a5000eae0ae7a51bf48f57e1c933a5199d8fe7b",
  "hakari-bench/NanoMTEB-Scandinavian": "3479ad3d391780b5f7f737df9faa853d1c9972b3",
  "hakari-bench/NanoMTEB-Spanish": "31e0395fb3c003b985676209e4902cdcfb173470",
  "hakari-bench/NanoMTEB-Thai": "c1fc94233015a5c7dbb2aea4099d67900de612f9",
  "hakari-bench/NanoMTEB-v2": "ae70aef0f08a3506066ae483ee3c31c536fdd79b",
  "hakari-bench/NanoMedical": "2403e84e42e745b70e913a3c97fb5315478dc9e5",
  "hakari-bench/NanoMuPLeR": "b99e4b047a1f7e29231ff527b555c160dd74afef",
  "hakari-bench/NanoR2MED": "4f46f9e6ca0283efad0c1251b6da7f05ebde33e0",
  "hakari-bench/NanoRARb": "017849a95097eea984680cbab35972f8d3812376",
  "hakari-bench/NanoRTEB": "3a002b35576f30f114f94bb417026139e70fb9db",
  "hakari-bench/NanoRuMTEB": "434fe43bb55a85954758fdf319993365591db1b1",
  "hakari-bench/NanoSSRB": "80dc6df1b0aa641950cf503842b6e7ef3be79d4e",
  "hakari-bench/NanoVNMTEB": "1726763179e1e114ad9ffcdc7262923471e8ecc8"
}

Checklist

• Result files are committed under hakari-results/minishlab__potion-multilingual-128M/.
• 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.