Add HAKARI-Bench results for geevec-ai/geevec-embeddings-1.0-lite

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

# Equivalent CLI invocation for each task claimed by the two-GPU queue.
# GPU_ID is physical GPU 0 or 1. DATASET/TASK/DATASET_SHA come from the
# standard 563-task manifest and the pinned revision table below.
# BATCH_SIZE=1 when mean document characters >6000 or maximum >120000;
# otherwise BATCH_SIZE=16. Every task ran once; no OOM retries occurred.
CUDA_VISIBLE_DEVICES="$GPU_ID" TOKENIZERS_PARALLELISM=false \
PYTORCH_ALLOC_CONF=expandable_segments:True \
uv run --group tf4-fa2 hakari-bench evaluate dense \
  --model geevec-ai/geevec-embeddings-1.0-lite \
  --model-revision 3c6073ded5dbe6eda5e8caed080400c61d5476b3 \
  --dataset "$DATASET" --split "$TASK" --dataset-revision "$DATASET_SHA" \
  --dtype bf16 --trust-remote-code --attn-implementation flash_attention_2 \
  --model-max-seq-length 40960 --device cuda:0 --batch-size "$BATCH_SIZE" \
  --query-prompt $'Instruct: Given a question, retrieve passages that answer the question.\nQuery: ' \
  --embedding-variant truncate:2048,1024,512,256 \
  --results-dir "output/geevec-full-fa2/gpu${GPU_ID}"

Submitter Notes

• The general domain was selected for both queries and documents; no document prefix was used. BF16, Flash Attention 2, cosine similarity, and last-token pooling were preserved. The shipped 40960-token limit was retained despite the README advertising 32768. Native 4096 dimensions plus the documented 2048/1024/512/256 truncations each include float, int8, binary, int8 rescore, and binary rescore (25 variants).
• Two RTX 5090 GPUs used a shared queue ordered by estimated text volume, with separate output roots. Batch 16 completed 512 tasks; long-document batch 1 completed 51 tasks. There were no OOM retries, failed tasks, or sequence-length reductions. Each worker reused one stock SentenceTransformer instance in the upstream general domain, warmed with eight short texts before its first task. Dataset configurations and revisions were verified online before offline evaluation. A 30-minute monitor checked processes, memory, errors, and output coverage. Total evaluation wall time was 8194 seconds. The CLI invocation above is equivalent to each in-process cli.main call; the recorded package versions describe the isolated environment actually used.
• Standard full coverage: all 563 tasks, preserving the 13 overlapping NanoBEIR-en copies. Overall follows the configured 550-task/381-score-unit grouping. Every task has all 25 variants (14,075 rows), finite scores, pinned model/dataset revisions, and the intended prompt/runtime metadata. The standalone DuckDB audit found no duplicates or missing scores. These files are from the full Flash Attention 2 run; earlier SDPA and domain/prompt ablations are excluded.

Checklist

• Result files are committed under hakari-results/geevec-ai__geevec-embeddings-1.0-lite/.
• 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.

Pinned dataset revisions

Code and model metadata: https://github.com/hakari-bench/hakari-bench/pull/18