Add NanoMTEB-BR results for 78 models

Summary

This submission adds hakari-bench/NanoMTEB-BR results for 75 existing
HAKARI-Bench public-weight models and three hosted embedding models (two OpenAI
models and Gemini Embedding 2). The
public-weight models were initially gated by reproducing the
existing hakari-bench/NanoBEIR-en / NanoArguAna and NanoMIRACL/en
results before running the six NanoMTEB-BR tasks. Base nDCG@10 differences
below 0.005 were treated as small runtime/numerical differences.

• Models: 78 (75 public-weight, 3 hosted API)
• Tasks per model: 6
• Result files: 468 .json.xz
• Dataset revision recorded by the target results:
  00541a0fce4048057fb7ddec30d37155a5c23d95
• Submission root: hakari-results/{model_dir}/hakari-bench__NanoMTEB-BR/

Included models

• Alibaba-NLP/gte-multilingual-base
• answerdotai/answerai-colbert-small-v1
• BAAI/bge-m3
• BAAI/bge-small-en-v1.5
• HIT-TMG/KaLM-embedding-multilingual-mini-v1
• KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5
• Lajavaness/bilingual-embedding-base
• Lajavaness/bilingual-embedding-small
• LiquidAI/LFM2.5-ColBERT-350M
• Qwen/Qwen3-Embedding-0.6B
• Snowflake/snowflake-arctic-embed-l-v2.0
• cl-nagoya/ruri-v3-30m
• cl-nagoya/ruri-v3-310m
• cl-nagoya/ruri-v3-reranker-310m
• codefuse-ai/F2LLM-v2-160M
• codefuse-ai/F2LLM-v2-330M
• codefuse-ai/F2LLM-v2-80M
• colbert-ir/colbertv2.0
• google/embeddinggemma-300m
• google/gemini-embedding-2
• hotchpotch/bekko-embedding-v1-a25m
• hotchpotch/bekko-embedding-v1-a8m
• hotchpotch/japanese-splade-v2
• ibm-granite/granite-embedding-107m-multilingual
• ibm-granite/granite-embedding-278m-multilingual
• ibm-granite/granite-embedding-311m-multilingual-r2
• ibm-granite/granite-embedding-97m-multilingual-r2
• intfloat/e5-base-v2
• intfloat/e5-large-v2
• intfloat/e5-small-v2
• intfloat/multilingual-e5-base
• intfloat/multilingual-e5-large
• intfloat/multilingual-e5-small
• jinaai/jina-embeddings-v5-text-nano
• jinaai/jina-embeddings-v5-text-small
• lightonai/ColBERT-Zero
• lightonai/GTE-ModernColBERT-v1
• microsoft/harrier-oss-v1-0.6b
• microsoft/harrier-oss-v1-270m
• mixedbread-ai/mxbai-edge-colbert-v0-17m
• mixedbread-ai/mxbai-edge-colbert-v0-32m
• mixedbread-ai/mxbai-embed-xsmall-v1
• mixedbread-ai/mxbai-rerank-base-v2
• nomic-ai/nomic-embed-text-v1.5
• nomic-ai/nomic-embed-text-v2-moe
• perplexity-ai/pplx-embed-v1-0.6b
• sbintuitions/sarashina-embedding-v2-1b
• sentence-transformers/LaBSE
• sentence-transformers/all-MiniLM-L12-v2
• sentence-transformers/all-MiniLM-L6-v2
• sentence-transformers/all-mpnet-base-v2
• sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
• sentence-transformers/paraphrase-multilingual-mpnet-base-v2
• sentence-transformers/static-similarity-mrl-multilingual-v1
• Alibaba-NLP/gte-multilingual-reranker-base
• BAAI/bge-reranker-v2-m3
• LiquidAI/LFM2.5-Embedding-350M
• Qwen/Qwen3-Embedding-4B
• Qwen/Qwen3-Embedding-8B
• Qwen/Qwen3-Reranker-0.6B
• cross-encoder/ettin-reranker-150m-v1
• cross-encoder/ettin-reranker-17m-v1
• cross-encoder/ettin-reranker-32m-v1
• cross-encoder/ettin-reranker-400m-v1
• cross-encoder/ettin-reranker-68m-v1
• cross-encoder/mmarco-mMiniLMv2-L12-H384-v1
• hotchpotch/japanese-reranker-xsmall-v2
• ibm-granite/granite-embedding-30m-sparse
• jinaai/jina-embeddings-v3
• jinaai/jina-reranker-v2-base-multilingual
• naver/splade-v3
• nvidia/Nemotron-3-Embed-1B-BF16
• opensearch-project/opensearch-neural-sparse-encoding-multilingual-v1
• openai/text-embedding-3-large
• openai/text-embedding-3-small
• perplexity-ai/pplx-embed-v1-4B
• prithivida/Splade_PP_en_v2
• voyageai/voyage-4-nano

Evaluation and reproducibility

• Every model evaluation used one physical GPU. Separate models were run in
  parallel across two RTX 5090 GPUs; no evaluation used multi-GPU model or data
  parallelism.
• The NanoArguAna gate replayed the existing result's model and dataset
  revisions, batch size, dtype, attention implementation, Transformers version,
  prompt behavior, candidate subset/top-k, reranker document limit, and
  retrieval score device.
• Dense results retain the model card's requested truncation variants and the
  default int8/binary plus rescore variants. Late-interaction results retain
  their reviewed model-card settings.
• ibm-granite/granite-embedding-97m-multilingual-r2 passed the gate at its
  original batch size, but the NanoMTEB-BR run required batch size 1 after an
  illegal CUDA memory access at the original target batch size. The result JSON
  records batch size 1.
• hotchpotch/japanese-splade-v2 uses a tokenizer compatibility loader and
  truncates each raw input to its first 4000 characters before MeCab/fugashi.
  This prevents the BRTaxQAR native-tokenizer segfault; NanoArguAna still
  reproduces exactly, and the guard is recorded in result metadata.
• microsoft/harrier-oss-v1-270m reproduces its existing NanoArguAna result to
  floating-point last-bit precision. Its NanoMTEB-BR mean is 59.03 versus the
  dataset README's stored candidate-generation mean of 60.90; investigation
  found Harrier bf16 batch-size sensitivity. The benchmark run retains the
  batch size 2 required by the existing-result reproduction gate.

Three models did not pass the strict reproduction gate, but were evaluated at
the user's request and retained after a separate plausibility audit. Across all
75 submitted models, NanoMTEB-BR means correlate with matching existing Overall
scores at Spearman 0.970982 and Pearson 0.965766. AnswerAI, Ruri, and
Mixedbread rank 51/75, 36/75, and 11/75 on NanoMTEB-BR versus 55/75, 41/75, and
16/75 by their existing Overall scores. All 18 task scores lie inside the
observed per-task ranges. Their original reproduction differences remain
documented rather than being presented as exact reproductions.

The hosted batch models locally truncate over-limit inputs to 8100 tokens
before submission: OpenAI uses tiktoken, while Gemini uses the Gemma2
SentencePiece tokenizer. Gemini uses its official retrieval prompts. A
provider-side transient socket failure affected one Quati and one JurisTCU
document embedding; incomplete outputs were rejected, those tasks were retried
as complete task batches, and only the complete retry outputs were retained.
Gemini's six-task mean is 0.731944 (rank 1/78), compared with rank 5/78 by its
existing Overall score. Across all 78 models, NanoMTEB-BR and existing Overall
remain strongly correlated (Spearman 0.968588, Pearson 0.965764).

Validation

• Confirmed exactly six files for every included model:
  Quati, JurisTCU, BRTaxQAR, FaQuADIR, MedPTRetrieval, and
  FaqBacenRetrieval.
• All 468 files pass xz -t.
• Parsed every payload and confirmed target.dataset_id is
  hakari-bench/NanoMTEB-BR and the six-split set is complete.
• Submission output contains no aggregate JSON, logs, caches, Markdown, or
  partial model directories.

NanoMTEB-BR is not yet part of the configured 551-task collection/viewer, so
the current generic DuckDB/PR helper does not produce an Overall comparison for
these tasks. The per-task result files are retained in the canonical results
layout for the future collection update.

NanoMTEB-BR vs. MTEB-BR Retrieval Borda-rank validation

This comparison uses the official MTEB-BR leaderboard results available on 2026-08-04 and the NanoMTEB-BR results in this PR.

We matched the 29 models that have complete results in both benchmarks. For each benchmark, we ranked these same 29 models independently on each of the six corresponding Retrieval tasks—Quati, JurisTCU, BRTaxQAR, FaQuADIR, MedPTRetrieval, and FaqBacenRetrieval—using nDCG@10. Ties receive the average rank. The six task ranks are summed, and the lowest sum receives the best Borda rank.

As reported for Nano-set rank preservation in the HAKARI-Bench paper, the expected difference is generally within ±2 ranks. 28 of the 29 matched models fall within ±2 Borda ranks. The only exception is sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2, with a difference of +5. The two Borda rankings have Spearman ρ = 0.9860 and Kendall τ = 0.9320, showing that NanoMTEB-BR preserves the original MTEB-BR Retrieval ordering very closely.

Δ rank = MTEB-BR rank − NanoMTEB-BR rank; positive values mean the model ranks higher on NanoMTEB-BR.