Add HAKARI-Bench results for KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5

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

# Initial split across two GPUs, using the reviewed dense model card settings.
CUDA_VISIBLE_DEVICES=0 uv run --group tf4-fa2 hakari-bench evaluate dense \
  --model 'KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5' \
  --model-revision '52c687bbe81a62a223c924698b787ec05c9a978a' \
  --dataset 'NanoMTEB-Dutch,NanoMIRACL,NanoMMTEB-v2,NanoRARb,NanoRTEB,NanoBEIR-de,NanoBEIR-es,NanoBEIR-it,NanoBEIR-ko,NanoBEIR-pt,NanoBEIR-sv,NanoBEIR-vi,NanoDAPFAM,NanoIndicQA,NanoCoIR,NanoMedical,NanoMTEB-Thai,NanoMTEB-French,NanoIFIR,NanoMTEB-Scandinavian,NanoLongEmbed,NanoMTEB-Korean,NanoChemTEB,NanoRuMTEB' \
  --dtype bf16 \
  --device cuda:0 \
  --trust-remote-code \
  --flash-attn2 \
  --query-prompt-name query \
  --document-prompt-name document \
  --embedding-variant 'truncate:896,512,256,128,64' \
  --batch-size 32

CUDA_VISIBLE_DEVICES=1 uv run --group tf4-fa2 hakari-bench evaluate dense \
  --model 'KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5' \
  --model-revision '52c687bbe81a62a223c924698b787ec05c9a978a' \
  --dataset 'NanoVNMTEB,NanoBRIGHT,NanoFaMTEB-v2,NanoMTEB-Polish,NanoMuPLeR,NanoBEIR-ar,NanoBEIR-en,NanoBEIR-fr,NanoBEIR-ja,NanoBEIR-no,NanoBEIR-sr,NanoBEIR-th,NanoMLDR,NanoMTEB-Misc,NanoJMTEB-v2,NanoMTEB-v2,NanoCMTEB,NanoLaw,NanoR2MED,NanoMTEB-Spanish,NanoBIRCO,NanoMTEB-German,NanoCodeRAG,NanoBuiltBench' \
  --dtype bf16 \
  --device cuda:0 \
  --trust-remote-code \
  --flash-attn2 \
  --query-prompt-name query \
  --document-prompt-name document \
  --embedding-variant 'truncate:896,512,256,128,64' \
  --batch-size 32

# Remaining and retried tasks were resumed with the same options and
# PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, changing only batch size to 16.

Submitter Notes

• Dense SentenceTransformers evaluation using bf16, trust_remote_code, Transformers 4.57.6, flash_attention_2, and the tf4-fa2 dependency group. The pinned model revision is 52c687bbe81a62a223c924698b787ec05c9a978a.
• Prompt names were set explicitly from the model's SentenceTransformers config: query_prompt_name=query, document_prompt_name=document. The query prompt text is Instruct: Given a query, retrieve documents that answer the query \n Query: ; the document prompt is empty.
• MRL dimensions requested were 896, 512, 256, 128, 64. The 896-dimensional variant is the native embedding dimension and is skipped by the evaluator as a no-op duplicate of the base result; submitted variants include the base embedding plus 512/256/128/64 truncation variants and the default int8/binary/rescore combinations.
• The initial split run used --batch-size 32 on two GPUs. Some long tasks hit memory pressure, so incomplete tasks were resumed with --batch-size 16 and PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True. Model max sequence length stayed at 32768.
• These are standard full HAKARI-Bench dense results for the current model-card target set, split across GPUs for throughput rather than a narrowed benchmark subset.

Checklist

• Result files are committed under hakari-results/KaLM-Embedding__KaLM-embedding-multilingual-mini-instruct-v2.5/.
• 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.