Add HAKARI-Bench results for lightonai/mDenseOn

Summary

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 nDCG@10

Reproducibility

Command

# NanoBEIR-en validation (13 tasks), on physical GPU 0.
CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 \
  --with transformers==5.6.2 \
  --with sentence-transformers==5.4.1 \
  hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__mDenseOn.yaml \
  --dataset hakari-bench/NanoBEIR-en \
  --device cuda:0 --batch-size 32 --show-progress

# The remaining standard tasks were split into two disjoint dataset groups.
# Each process could see exactly one physical GPU; both used logical cuda:0.
CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 \
  --with transformers==5.6.2 \
  --with sentence-transformers==5.4.1 \
  hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__mDenseOn.yaml \
  --dataset NanoMTEB-Dutch,NanoMIRACL,NanoMMTEB-v2,NanoRARb,NanoRTEB,NanoBEIR-de,NanoBEIR-es,NanoBEIR-it,NanoBEIR-ko,NanoBEIR-pt,NanoBEIR-sv,NanoBEIR-vi,NanoMTEB-Misc,NanoJMTEB-v2,NanoMedical,NanoMTEB-Thai,NanoLaw,NanoR2MED,NanoMTEB-Scandinavian,NanoLongEmbed,NanoMTEB-BR,NanoBIRCO,NanoMTEB-Korean,NanoRuMTEB \
  --device cuda:0 --batch-size 32

CUDA_VISIBLE_DEVICES=1 uv run --group tf4-fa2 \
  --with transformers==5.6.2 \
  --with sentence-transformers==5.4.1 \
  hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__mDenseOn.yaml \
  --dataset NanoVNMTEB,NanoBRIGHT,NanoFaMTEB-v2,NanoMTEB-Polish,NanoMuPLeR,NanoBEIR-ar,NanoBEIR-en,NanoBEIR-fr,NanoBEIR-ja,NanoBEIR-no,NanoBEIR-sr,NanoBEIR-th,NanoMLDR,NanoIndicQA,NanoCoIR,NanoMTEB-v2,NanoCMTEB,NanoMTEB-French,NanoIFIR,NanoMTEB-Spanish,NanoSSRB,NanoMTEB-German,NanoCodeRAG,NanoChemTEB,NanoBuiltBench \
  --device cuda:0 --batch-size 32

# NanoDAPFAM was run separately with its 12 standard splits so that the six
# extended ToFullText tasks were not accidentally included.
CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 \
  --with transformers==5.6.2 \
  --with sentence-transformers==5.4.1 \
  hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__mDenseOn.yaml \
  --dataset NanoDAPFAM \
  --split NanoDAPFAMAllTitlAbsClmToTitlAbs,NanoDAPFAMAllTitlAbsClmToTitlAbsClm,NanoDAPFAMAllTitlAbsToTitlAbs,NanoDAPFAMAllTitlAbsToTitlAbsClm,NanoDAPFAMInTitlAbsClmToTitlAbs,NanoDAPFAMInTitlAbsClmToTitlAbsClm,NanoDAPFAMInTitlAbsToTitlAbs,NanoDAPFAMInTitlAbsToTitlAbsClm,NanoDAPFAMOutTitlAbsClmToTitlAbs,NanoDAPFAMOutTitlAbsClmToTitlAbsClm,NanoDAPFAMOutTitlAbsToTitlAbs,NanoDAPFAMOutTitlAbsToTitlAbsClm \
  --device cuda:0 --batch-size 32

Submitter Notes

• The pinned checkpoint uses the model's built-in query and document prompts, bf16, Flash Attention 2, an 8,192-token maximum sequence length, and batch size 32. The model does not require trust_remote_code and does not advertise Matryoshka truncation dimensions.
• The current checkpoint metadata requires SentenceTransformers 5.4.1 and Transformers 5.6.2. The older framework versions stated in the model-card prose could not load the current checkpoint format.
• NanoBEIR-en was validated first (mean nDCG@10 0.6698). An initial single-GPU full run was stopped after saved task boundaries and resumed through the disjoint two-GPU dataset partition above. No result was overwritten, no task used a different batch size, and no OOM or failed task occurred.
• These are the complete standard --all-equivalent results: 563 task files. Each evaluation process used one RTX 5090 only; the two physical GPUs processed disjoint datasets independently.

Checklist

• Result files are committed under hakari-results/lightonai__mDenseOn/.
• 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.