Replace HAKARI-Bench results for hotchpotch/bekko-static-embedding-multilingual-v1

Replace all 563 existing task files for this model with a fresh full evaluation of
Hugging Face revision 49887c8b23ecbeb9026b9cd6e86656ce9b14836e.
The official model now stores its embedding table in FP16; the previous result
revision was 48e94648fd2e28dfb9ac57e8759c559fafa836d6 (FP32 storage).
Compute remains FP32, matching the author's storage-precision comparison.

The change modifies the same 563 existing .json.xz paths under
hakari-results/hotchpotch__bekko-static-embedding-multilingual-v1/.
There are no added or removed task paths and no changes to other models.
All 563 new compressed payload hashes differ from the existing main files.

Default Overall micro nDCG@10 x 100 is 45.883531401985564 over 550 canonical
tasks, matching the official FP16-storage result rounded to 45.88353.
Overall candidate-reranking micro nDCG@10 x 100 is 47.52946363047525.

Coverage was audited in the result JSON and rebuilt DuckDB: 563 tasks x 30
variants = 16,890 task-variant rows, with 550 Overall tasks for every variant.
The 30 variants include base, five truncation widths, full-dimensional int8 and
binary search/rescore, and all truncation x quantization/rescore combinations.
All scores are finite; every variant retains its retrieval and candidate-rerank
top-100 artifacts and query coverage.

Model card: https://huggingface.co/hotchpotch/bekko-static-embedding-multilingual-v1/blob/49887c8b23ecbeb9026b9cd6e86656ce9b14836e/README.md

Summary

Per-component nDCG@10

The template groups MNanoBEIR by task name before aggregation (381 grouped
units from 550 raw tasks). Its grouped summary differs from the default viewer
Overall micro score reported above. Component scores below use the template grouping.

Reproducibility

Command

Equivalent complete --all reproduction command (use a fresh result directory):

OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 TOKENIZERS_PARALLELISM=false \
uv run --with sentence-transformers==5.6.0 --with transformers==5.14.1 \
  hakari-bench evaluate dense \
  --model hotchpotch/bekko-static-embedding-multilingual-v1 \
  --model-revision 49887c8b23ecbeb9026b9cd6e86656ce9b14836e \
  --all \
  --dtype fp32 --device cuda:0 --retrieval-score-device cpu \
  --batch-size 4096 \
  --embedding-variant truncate:512,256,128,64,32 \
  --results-dir output/bekko-static-fp16-refresh-20260914

The actual run first validated all 13 NanoBEIR-en tasks, then partitioned the
remaining exact standard task manifest by dataset/split across two RTX 5090 GPUs
(cuda:0 and cuda:1). Every shard used the options above, replacing --all
with its explicit --dataset and --split selection. The union was checked
against all 563 standard targets, including only NanoDAPFAM's standard splits.

Submitter Notes

• The downloaded embedding.weight tensor is F16 with shape [200000, 1024].
  Its SHA-256 is 586e165a868f306822174e8110020fa8759577cc62a537a702941747732858f8,
  matching the official card. The loader converts it to FP32 for computation.
  These results do not measure FP16 inference performance.
• Query/document prompts remain empty; cosine similarity, the default candidate
  ranking reranking_hybrid, all candidate rows, and native input limits are preserved.
  No trust_remote_code, attention override, or sequence-length override was used.
• All documented MRL dimensions were checked before evaluation. Native 1024 is
  base; only 512, 256, 128, 64, and 32 are derived truncations.
• Initial orchestration used four workers; it was resumed with twelve workers
  to increase CPU scoring/compression throughput. Complete results were retained,
  and one interrupted output was rerun. Thread counts were 4 in the initial
  validation, 2 in the first four-worker pass, and 1 in the twelve-worker pass.
  Batch size remained 4096. There are no failed or omitted final tasks.
• Dataset revisions are recorded in every task JSON and are consistent within
  each dataset. All tasks use the same pinned new model revision and package versions.
• The rebuilt viewer rendered the model successfully in Overall, Overall with
  variants, and MNanoBEIR via in-process HTTP checks.

Checklist

• Result files are committed under hakari-results/hotchpotch__bekko-static-embedding-multilingual-v1/.
• 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.