Add HAKARI-Bench results for dmis-lab/UltraBERT-Embed-150M

Scope

This submission intentionally covers English and code search only (--all --english, equivalent to --all --lang en,code), as requested for this English embedding model. It is not a complete multilingual --all run.

All 165 task files are included, including the 13 overlapping NanoBEIR-en copies that remain required result artifacts. Applying the viewer's duplicate exclusions yields 152 English/code score units, with mean nDCG@10 0.525777 (52.5777/100). This is a scope-specific average, not the 550-task multilingual Overall score. English natural-language tasks alone: 127 tasks, 0.482152 (48.2152/100); the other 25 score units are code-search tasks.

Each of base, int8, binary, int8 rescore, and binary rescore has 165/165 task coverage (825 model-task-variant measurements in total). No undocumented truncation variants were added.

Summary

English/code nDCG@10 by Overall component

Reproducibility

Command

# Original entry point, using the reviewed UltraBERT model card:
OMP_NUM_THREADS=8 uv run --no-project --python tmp/ultrabert/venv/bin/python \
  python -m hakari_bench.cli evaluate from-model-card \
  --model-card config/model_cards/dmis-lab__UltraBERT-Embed-150M.yaml \
  --all --english --device cuda:0 --batch-size 32

# Equivalent single-GPU reproduction without a local model-card dependency:
OMP_NUM_THREADS=8 uv run --no-project --python /path/to/ultrabert-venv/bin/python \
  python -m hakari_bench.cli evaluate dense \
  --model dmis-lab/UltraBERT-Embed-150M \
  --model-revision b06b36724c5417a1ec1b7863dd4fa542b74e4899 \
  --all --english --dtype bf16 --attn-implementation flash_attention_2 \
  --model-max-seq-length 32768 \
  --query-prompt-name query --document-prompt-name document \
  --device cuda:0 --batch-size 32

Submitter Notes

• Native loading with Transformers 5.20.1 / SentenceTransformers 6.1.0; no remote code. FlashAttention 2.8.3, bf16 weights, and the model's native 32,768-token maximum were retained. No sequence-length reduction or batch-size change.
• CLS pooling, 768-dimensional embeddings. Query prompt: Instruct: Given a question, retrieve passages that answer the question\nQuery: ; document prompt: empty. Prompt names query and document preserve the checkpoint configuration.
• Total parameters: 151,032,576; input embedding parameters: 37,748,736; active/non-embedding parameters: 113,283,840.
• The candidate_ranking=reranking_hybrid field is the stored CLI default; these are exhaustive dense retrieval results, not candidate-limited cross-encoder reranking.
• Model reference: https://huggingface.co/dmis-lab/UltraBERT-Embed-150M/tree/b06b36724c5417a1ec1b7863dd4fa542b74e4899 . Code checkout: 4c2cd7401017ac9899c47d12b63ea3a24e622e50, with a local model card and model-specific documentation additions only; no benchmark implementation changes.
• An initial NanoArguAna smoke evaluation passed. After the first 33 task results, the single-GPU process was intentionally interrupted; the remaining 132 tasks were partitioned into two disjoint sets of 66 on cuda:0 and cuda:1 with identical runtime settings. Both partitions completed successfully. Results from the second output root were consolidated without changing payloads. No failed or missing task remains.
• Coverage and runtime metadata were audited across all submitted files. All files are compressed .json.xz; no aggregate result, model card, report, DuckDB, or scratch file is part of this dataset PR.
• Intentional subset: standard --all --english target only; all 165 resolved tasks and all five default dense representations are complete.

Checklist

• Result files are committed under hakari-results/dmis-lab__UltraBERT-Embed-150M/.
• 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.

Exact dataset revisions