Add HAKARI-Bench results for google/embeddinggemma-2

Evaluation coverage and default Overall

Validated text-only FlexAttention evaluation: 569 tasks × 20 conditions = 11,380 task-variant rows.
The submission contains 563 standard tasks plus six extended NanoDAPFAM FullText tasks.
Extended FullText scores are reported separately and excluded from the standard Overall.
Each condition has 550 canonical Overall tasks; all 563 task artifacts are retained.
Default Overall micro nDCG@10 ×100: 58.60100107.
Base candidate-reranking micro nDCG@10 ×100: 59.56060656.

Extended FullText results (outside standard Overall)

Summary

DuckDB Nano-set Comparison

Computed from the submitted and published DuckDB task results using 381 grouped units from 550 canonical tasks. These grouped means are on a 0–1 scale; the default 550-task micro score on a 0–100 scale is reported above. Dataset revisions may differ across models. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Comparison database: hakari-bench/leaderboard_database revision 33d59d83d2bbe0eddf91846645e709e07cb945e7.

Grouped component nDCG@10

The template aggregates 381 grouped units; its grouped mean differs from the default Overall micro mean over 550 tasks.

Reproducibility

Command

# Run each dataset at the original pinned revision (see command manifest).
CUDA_VISIBLE_DEVICES=1 PYTHONPATH=tmp uv run --no-project \
  --python tmp/embeddinggemma2-venv/bin/python python -c \
  'from hakari_bench.cli import main; main()' evaluate dense \
  --model google/embeddinggemma-2 \
  --model-revision 914f7f89142e33e77833254d9c9b90c3cef7303b \
  --model-loader embeddinggemma2_flex_loader:load_model \
  --model-loader-kwargs-json '{"config_kwargs":{"vision_config":null,"audio_config":null},"input_modalities":["text"],"kernel_options":{"BLOCK_M":16,"BLOCK_N":32,"num_stages":1,"num_warps":4,"BACKEND":"TRITON"}}' \
  --dtype bf16 --attn-implementation flex_attention --model-max-seq-length 8192 \
  --query-prompt-name SearchQuery --document-prompt-name Document \
  --embedding-variant truncate:128,256,512 --device cuda:0 --batch-size 16 \
  --dataset DATASET_ID --dataset-revision ORIGINAL_DATASET_COMMIT --split PINNED_TASK_LIST

Pinned dataset revisions and explicit task lists

Submitter Notes

• All results use the text embedding encoder only, with official FP32 RoPE frequency buffers preserved. This complete FlexAttention submission replaces the original uploaded scores affected by a redundant bf16 cast; original local artifacts are retained separately.
• Runtime: Python 3.12, SentenceTransformers 6.1.0, torch 2.14.1, CUDA 13.0, Transformers 5.19.0.dev0 at a96730c8c97b8efbf35bbaf7f5da33ec99231a49. GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q (physical GPU 1). bf16, batch 16, 8192-token context, and SearchQuery/Document prompts. Repository dependency requirements and lockfile are unchanged.
• Attention: FlexAttention, Triton backend, kernel options BLOCK_M=16, BLOCK_N=32, num_stages=1, num_warps=4, BACKEND=TRITON. Smaller tiles avoid the GPU shared-memory limit.
• Torch Dynamo limits: recompile_limit=256, accumulated_recompile_limit=2048, fail_on_recompile_limit_hit=True. The initial English MLDR trial hit the default limit and fell back to an unfused implementation; it was stopped and discarded. The corrected complete run has no unfused fallback warnings.
• Coverage: 563 standard tasks plus all six NanoDAPFAM FullText tasks (18 NanoDAPFAM tasks total). All 569 task files have the same 20 embedding conditions, finite scores, and verified original pinned dataset revisions. The standard Overall uses 550 canonical tasks; extended FullText tasks are excluded from both micro and grouped comparisons.
• The completed English MLDR Flex trial was reused with identical settings; the remaining 568 tasks were evaluated in the separate Flex result root. The full wave ran from 2026-10-08 07:31 UTC to 13:15 UTC. All 20 conditions share one query/corpus encoding pass per task.
• English MLDR nDCG@10 ×100: corrected SDPA / torch 2.9.1 56.7163, Flex / torch 2.14.1 57.2295. Evaluation time including compilation: 1203.3s → 652.7s (~1.84× faster). Both attention and torch version differ; this is not an isolated attention-only comparison.
• Loading fix, model card and reproduction instructions: GitHub PR #7.

Text encoder parameter measurement

Counts were measured from the actual text-only loaded encoder. Active parameters follow HAKARI's convention: total minus input embedding parameters. Vision/audio encoders are excluded. Result payloads record the count scope and exact Transformers commit; experiment fingerprints were rebuilt for the submitted metadata.

Text-only loader and Flex wrapper

from sentence_transformers import SentenceTransformer
from hakari_bench.models import resolve_torch_dtype


def load_model(config):
    model = SentenceTransformer(
        config.model_name_or_path,
        revision=config.model_revision,
        device=config.device,
        config_kwargs={'vision_config': None, 'audio_config': None},
        model_kwargs={
            'torch_dtype': resolve_torch_dtype(config.dtype),
            'attn_implementation': config.attn_implementation or 'sdpa',
        },
    )
    if config.max_seq_length is not None:
        model.max_seq_length = config.max_seq_length
    # This IR benchmark passes text, including strings that look like media URLs.
    # Restrict the formatter's modality inference without changing input content.
    model[0].input_formatter.supported_modalities = ['text']
    return model

Save the following wrapper as tmp/embeddinggemma2_flex_loader.py:

from embeddinggemma2_text_loader import load_model as load_text_model

def load_model(config):
    import torch
    torch._dynamo.config.recompile_limit=256
    torch._dynamo.config.accumulated_recompile_limit=2048
    torch._dynamo.config.fail_on_recompile_limit_hit=True
    model=load_text_model(config)
    options={'BLOCK_M':16,'BLOCK_N':32,'num_stages':1,'num_warps':4,'BACKEND':'TRITON'}
    options.update(config.model_loader_kwargs.get('kernel_options',{}))
    original_forward=model[0].auto_model.forward
    def tuned_forward(*args,**kwargs):
        kwargs['kernel_options']=options
        return original_forward(*args,**kwargs)
    model[0].auto_model.forward=tuned_forward
    return model

Checklist

• Result files are committed under hakari-results/google__embeddinggemma-2/.
• 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.