Add NanoSSRB results for 80 entries and NanoMTEB-BR BM25

#30
HAKARI-Bench org
edited 3 days ago

Add NanoSSRB results for 80 entries and NanoMTEB-BR BM25

Summary

This submission adds hakari-bench/NanoSSRB results for every model currently
registered in HAKARI-Bench: 76 public-weight models, two OpenAI embedding
models, Gemini Embedding 2, and the BM25 baseline. It also fills the six missing
hakari-bench/NanoMTEB-BR BM25 results; BM25 was accidentally omitted from the
earlier model-type inventory used by PR #29.

  • Entries: 80 (79 models and one BM25 baseline)
  • Tasks per entry: 6
  • Result files: 486 .json.xz (480 NanoSSRB and 6 NanoMTEB-BR BM25)
  • NanoSSRB dataset revision: 80dc6df1b0aa641950cf503842b6e7ef3be79d4e
  • NanoMTEB-BR dataset revision: 00541a0fce4048057fb7ddec30d37155a5c23d95
  • Submission root: hakari-results/{model_dir}/hakari-bench__NanoSSRB/
  • Candidate ranking: reranking_hybrid for every non-BM25 entry; bm25 for
    the BM25 baseline

NanoSSRB is an explicit-only benchmark over JSON-serialized semi-structured
objects. Its six domains are Academic, FinanceAndEconomics,
HumanResources, LLMAgentAndTool, ProductSearch, and ResumeSearch.

Evaluation and reproducibility

  • Public-weight models were evaluated on one physical RTX 5090 each. Separate
    model processes ran in parallel across two GPUs; no model used distributed or
    multi-GPU inference.
  • Model revision, dtype, attention implementation, maximum sequence length,
    prompts/task adapters, and embedding variants follow the reviewed model card
    and existing result metadata. Exact runtime and package versions are recorded
    in every result payload. The principal environment used Python 3.12, torch
    2.9.0, CUDA 12.8, Transformers 5.x, Sentence Transformers 5.x, and datasets
    4.8.4; isolated historical environments were used where reproduction
    required them.
  • Dense results retain the default int8/binary and rescore variants plus each
    model card's requested truncation variants. Provider models retain their
    requested dimension grids.
  • OpenAI text-embedding-3-small and text-embedding-3-large were evaluated
    through the Batch API. Large outputs were sharded after the provider's
    unsharded file repeatedly timed out; all corrected shards completed without
    failed requests before materialization.
  • google/gemini-embedding-2 was evaluated through Vertex batch inference.
    Transient socket failures were retried, and all 563 missing embeddings were
    recovered before the six complete task results were materialized.
  • Hosted inputs over the provider limit are truncated locally with the
    provider-compatible tokenizer. Gemini uses its official retrieval prompts.
  • hotchpotch/japanese-splade-v2 uses the reviewed compatibility loader and
    truncates raw text to the first 4000 characters before tokenization.
  • hotchpotch/japanese-reranker-xsmall-v2 uses the reproduced FP32 + SDPA
    runtime rather than the non-reproducing BF16 + FlashAttention path.

The accepted local-weight runtimes were checked on
NanoBEIR-en/NanoArguAna. The nine models with the largest negative NanoSSRB
Borda shifts all remain within 0.005 absolute nDCG@10 of their existing
NanoArguAna result; five reproduce exactly and the maximum difference is
0.003861334 for the mMARCO reranker. This indicates that the NanoSSRB shifts
are benchmark-specific rather than a general model-loading regression.

Result plausibility

  • Mean base nDCG@10 spans 0.0859291033 to 0.5627851267.
  • Reference means: Gemini Embedding 2 0.4413860634, OpenAI small
    0.3256673212, OpenAI large 0.3345185477, BM25 0.2821417014, and
    Harrier OSS 270m 0.3597489962.
  • Mean-score ordering correlates with existing results at Spearman 0.731329
    versus NanoBEIR-en (79 shared models), 0.756522 versus NanoMTEB-BR (78),
    and 0.630112 versus NanoMIRACL (79).
  • Using the viewer's normalized Borda formula on the same 78-model complete
    population, NanoSSRB versus the fixed 544-task Overall manifest has Spearman
    0.770894 and Kendall 0.571476. The result is stable against all 557
    standard tasks (Spearman 0.773558).
  • NanoSSRB favors several lexical sparse models and lowers several otherwise
    strong multilingual dense/reranker models. This is consistent with its mix
    of exact filters and semantic conditions over serialized objects. The large
    negative shifts were retained as task-family diagnostics after the
    NanoArguAna reproduction check.
  • NanoMTEB-BR BM25 mean nDCG@10 is 0.5178022169; all six scores come from
    the fixed dataset bm25 candidate subset rather than a local recomputation.

Validation

  • Final inventory audit passed 80/80 entries with no missing or extra model
    directories.
  • Confirmed exactly six NanoSSRB files per entry and 480 NanoSSRB files in
    total, plus the six NanoMTEB-BR BM25 files.
  • All files are non-empty valid XZ payloads with finite aggregate scores.
  • Every payload records dataset revision
    80dc6df1b0aa641950cf503842b6e7ef3be79d4e and the expected logical model ID.
  • Every model preserves the variant set from its existing NanoArguAna result.
  • Every non-BM25 payload records reranking_hybrid; BM25 records bm25.
  • The staged submission contains only the intended .json.xz result files.

Included entries

The submission contains the complete 80-entry inventory recorded by the audit:

  • 76 Hugging Face public-weight dense, sparse, reranker, and late-interaction
    models already registered in HAKARI-Bench
  • openai/text-embedding-3-small
  • openai/text-embedding-3-large
  • google/gemini-embedding-2
  • bm25

Notable newly registered coverage relative to the previous NanoMTEB-BR wave
includes nvidia/Nemotron-3-Embed-8B-BF16; no registered model was omitted.

hotchpotch changed pull request title from Add NanoSSRB results for 80 entries to Add NanoSSRB results for 80 entries and NanoMTEB-BR BM25
hotchpotch changed pull request status to merged

Sign up or log in to comment