Embedding Benchmarks
54 beliefs (54 IN, 0 OUT)
This topic centers on the Massive Text Embedding Benchmark (MTEB) and its role as a unified evaluation framework for text embedding models. MTEB, introduced by Muennighoff et al. (2022, arXiv 2210.07316v3) by researchers at Hugging Face and Cohere, covers 8 task types, 58 constituent datasets, 10 multilingual datasets, and 112 languages (mteb-scope-8-task-types-58-datasets-112-languages, mteb-112-languages). It fills a gap left by prior benchmarks such as SemEval/SentEval (STS and classifier evaluation only), USEB (reranking), and BEIR (zero-shot retrieval), none of which offered a single cross-task comparison (mteb-prior-benchmarks-narrow-scope). The benchmark's scope is bounded: it excludes code datasets and cross-modal evaluation, with CodeSearchNet, TyDi QA, XOR QA, and MIRACL noted as future extensions (mteb-no-code-or-multimodal-datasets). Across approximately 5,000 experiments on 30+ models, MTEB found that no single embedding method dominates all task categories (mteb-no-dominant-embedding-model), a finding consistent with the benchmark's broader interpretive claim.
The most consequential belief in this group is the interpretive principle that embedding benchmarks measure readout fidelity rather than representational completeness (evaluation-as-readout-fidelity-measure). Under this reading, the specific evaluation pipelines are telling: classification uses a logistic regression head with max_iter=100 trained on frozen embeddings (mteb-classification-head-params, mteb-classification-logistic-regression-100-iterations), clustering uses mini-batch k-means with batch size 32 scored by V-measure (mteb-clustering-params, mteb-clustering-vmeasure-kmeans-batch32), STS and summarization rely on Spearman correlation of cosine similarities (mteb-sts-main-metric, mteb-sts-evaluation-metric), bitext mining uses F1 over closest cosine pairs (mteb-bitext-mining-metric-f1), pair classification reports average precision via cosine (mteb-pair-class-main-metric), and reranking uses MAP (mteb-reranking-main-metric). All of these are linear or near-linear probes that extract a specific geometric direction from the embedding space. Combined with MTEB's dataset-count-weighted scoring (which biases toward retrieval, classification, and clustering simply because they have more datasets, mteb-average-score-dataset-count-weighted) and the cosine-to-Spearman pipeline, the leaderboard becomes a statement about calibration fidelity rather than about the semantic richness of the underlying space. The GloVe baseline, which applies a context-agnostic word-to-vector averaging layer and scores roughly 42 on average (mteb-glove-baseline-avg-score, mteb-glove-baseline-average-42, mteb-glove-averaging-layer), establishes a floor for learned embeddings, while ST5-XXL leads English scores at 59.51, followed by GTR-XXL (58.97) and SGPT-5.8B-msmarco (58.81) (st5-xxl-best-avg-score). In multilingual settings, LaBSE tops bitext mining (labse-top-bitext-mining) and MPNet leads classification and STS (mpnet-best-multilingual-cls-sts), reinforcing the no-dominant-model pattern.
The remaining beliefs in this group are largely premise-level: they fix concrete parameters, dataset characteristics, and practical infrastructure. These include the specific language counts per multilingual dataset (Tatoeba=112, Massive=51, STS22=18, BUCC=5, mteb-multilingual-dataset-language-counts, mteb-bucc-5-languages, mteb-massive-51-languages, sts22-language-pairs), dataset sizes (BIOSS/BIOSSES at exactly 100 pairs, SICK-R at roughly 10,000, STSBenchmark as monolingual English, mteb-bioss-100-sentence-pairs, mteb-biosses-100-biomedical-pairs, mteb-sickr-approximately-10000-pairs, stsbenchmark-monolingual-english), the input-type taxonomy of s2s, p2p, and s2p (mteb-input-type-taxonomy), embedding sizes ranging from 1.2 kB (GloVe) to 16.4 kB (SGPT-5.8B) (mteb-embedding-size-range), text lengths from ~32 to ~1981 characters (mteb-text-length-range), the summarization task's single-dataset imbalance (mteb-summarization-single-dataset), clustering and reranking's strong inter-task correlation versus summarization's outlier status (mteb-clustering-reranking-strongest-correlation), and the arXiv clustering's three split strategies (mteb-arxiv-clustering-three-splits). Practical details cover the code repository at github.com/embeddings-benchmark/mteb (mteb-code-repo, mteb-github-repository), the Hugging Face leaderboard and JSON results (mteb-leaderboard-url, mteb-results-hosted-huggingface, mteb-results-json-url), the ease of adding models (fewer than 10 lines) or datasets (a single config file) (mteb-add-model-10-lines-code, mteb-add-dataset-single-config-file), speed benchmarking on a single A100 80GB (mteb-speed-bench-hardware), and the Jean Zay cluster at IDRIS used for evaluation runs (mteb-hpc-jean-zay). STS scores range from 0 to 5 (mteb-sts-score-range-0-to-5), and 56 of the 58 datasets are English subsets in the main results table (mteb-total-datasets-count).
All beliefs in this topic are marked IN; none have been retracted. The interpretive belief about readout fidelity (evaluation-as-readout-fidelity-measure) functions as a derived or synthetic claim that draws on the methodological premises (linear probes, dataset-count weighting, cosine-based metrics) to reframe what the leaderboard numbers actually signify. The model-performance beliefs (GloVe baseline, ST5-XXL top score, LaBSE and MPNet multilingual leadership) are empirical observations that are consistent with the readout-fidelity reading, since different tasks probe different geometric directions and no single model excels at all of them simultaneously.
-
IN
evaluation-as-readout-fidelity-measure
MTEB's dataset-count-weighted bias, SBERT's task-specificity gap, and the cosine→Spearman pipeline consistency are all manifestations of a single principle: embedding benchmarks measure readout fidelity (how faithfully a linear probe extracts a specific geometric direction) rather than the completeness or quality of the underlying representation space, making leaderboard rankings a statement about calibration rather than about semantics. -
IN
glove-baseline-avg-score
GloVe word-embedding baseline achieves an average MTEB score of 41.97 -
IN
labse-top-bitext-mining
LaBSE is the top-performing model for multilingual bitext mining in MTEB evaluations -
IN
mpnet-best-multilingual-cls-sts
MPNet (multilingual) is the strongest model on multilingual classification and STS tasks in MTEB -
IN
mteb-112-languages
MTEB covers 112 languages across its multilingual datasets -
IN
mteb-add-dataset-single-config-file
Adding a new dataset for an existing MTEB task requires a single configuration file specifying the task name and a Hugging Face dataset identifier; a new task requires implementing a task interface plus an evaluator class. -
IN
mteb-add-model-10-lines-code
Adding a model to the MTEB benchmark requires fewer than 10 lines of code; the model must accept a list of texts and return a fixed-shape vector per item. -
IN
mteb-arxiv-2210-07316
The MTEB paper (Muennighoff et al., 2022) has arXiv identifier 2210.07316v3, authored by researchers at Hugging Face and Cohere. -
IN
mteb-arxiv-clustering-three-splits
arXiv clustering in MTEB uses three split strategies: main category (coarse), secondary category within main (fine-grained), and secondary category across all (multi-scale). -
IN
mteb-average-score-dataset-count-weighted
MTEB's overall average score is dataset-count-weighted rather than task-weighted, biasing results toward tasks with more datasets (retrieval, classification, clustering). -
IN
mteb-beir-embedding-benchmarks
MTEB (Massive Text Embedding Benchmark) covers cross-domain and cross-language embedding evaluation, while BEIR specifically benchmarks heterogeneous zero-shot information retrieval. -
IN
mteb-bioss-100-sentence-pairs
The BIOSS biomedical STS dataset in MTEB contains 100 sentence pairs. -
IN
mteb-biosses-100-biomedical-pairs
BIOSSES contains exactly 100 sentence pairs from the biomedical domain for domain-specific semantic similarity. -
IN
mteb-bitext-mining-metric-f1
MTEB bitext mining evaluation uses F1 as the primary metric, with closest pairs found via cosine similarity between embedded language sets. -
IN
mteb-bitext-sts-spearman-correlation-multilingual
In multilingual STS evaluation, MTEB reports Spearman correlation of cosine similarities as the scoring metric. -
IN
mteb-bucc-5-languages
BUCC in MTEB supports exactly 5 languages: EN, FR, RU, DE, ZH. -
IN
mteb-classification-head-params
MTEB classification evaluation uses a logistic regression head with max_iter=100 trained on embeddings (no end-to-end fine-tuning) -
IN
mteb-classification-logistic-regression-100-iterations
MTEB classification evaluation trains a logistic regression (max 100 iterations) on frozen embeddings; the embedding model is not fine-tuned. -
IN
mteb-clustering-params
MTEB clustering evaluation uses mini-batch k-means with batch_size=32 and k=number of labels, scored with V-measure -
IN
mteb-clustering-reranking-strongest-correlation
Among MTEB tasks, clustering and reranking show the strongest inter-task performance correlation, while summarization is an outlier with weaker correlation to other tasks. -
IN
mteb-clustering-vmeasure-kmeans-batch32
MTEB clustering evaluation uses mini-batch k-means with batch size 32 and V-measure as the primary metric. -
IN
mteb-code-repo
The MTEB codebase is hosted at https://github.com/embeddings-benchmark/mteb -
IN
mteb-embedding-size-range
MTEB embedding sizes range from 1.2 kB per example (GloVe/Komninos) to 16.4 kB per example (SGPT-5.8B) -
IN
mteb-experiment-count
MTEB conducted approximately 5,000 experiments across 30+ embedding models -
IN
mteb-github-repository
The MTEB benchmark source code is hosted at https://github.com/embeddings-benchmark/mteb and the public leaderboard at https://huggingface.co/spaces/mteb/leaderboard. -
IN
mteb-glove-averaging-layer
GloVe-style word embedding models are context-agnostic, using a word-to-vector layer followed by an averaging operation, making them invariant to input length. -
IN
mteb-glove-baseline-average-42
GloVe average-of-word-embeddings baseline scores approximately 42 average across MTEB tasks, establishing a floor for learned embeddings. -
IN
mteb-hpc-jean-zay
MTEB evaluations and data processing ran on the Jean Zay cluster at IDRIS (CNRS) under GENCI allocation 2021-A0101012475 -
IN
mteb-input-type-taxonomy
MTEB uses three input type notations: s2s (sentence-to-sentence), p2p (passage-to-passage), and s2p (sentence-to-passage). -
IN
mteb-leaderboard-url
The MTEB interactive leaderboard is hosted at https://huggingface.co/spaces/mteb/leaderboard -
IN
mteb-massive-51-languages
The Massive Intent/Scenario dataset in MTEB covers 51 typologically diverse languages. -
IN
mteb-multilingual-dataset-language-counts
MTEB multilingual dataset language counts are: Massive=51, Tatoeba=112, STS22=18, STS17=11, AmazonReviews=6, AmazonCounterfactual=4. -
IN
mteb-no-code-or-multimodal-datasets
MTEB includes no code datasets and no cross-modal (image/text) evaluation; CodeSearchNet, TyDi QA, XOR QA, and MIRACL are suggested as future extensions. -
IN
mteb-no-dominant-embedding-model
Across the 33 models evaluated in MTEB, no single text embedding method dominates all tasks; different models top different task categories. -
IN
mteb-pair-class-main-metric
The main metric for MTEB Pair Classification is average precision (AP) computed via cosine similarity -
IN
mteb-pair-classification-distance-metrics
MTEB pair classification searches over cosine, dot, Euclidean, and Manhattan distance metrics to select the best binary threshold, reporting average precision with cosine similarity. -
IN
mteb-prior-benchmarks-narrow-scope
SemEval/SentEval cover only STS and classifier-based evaluation, USEB focuses on reranking, and BEIR covers zero-shot retrieval; none provided a unified cross-task embedding evaluation prior to MTEB. -
IN
mteb-reranking-main-metric
The main metric for MTEB reranking evaluation is MAP (mean average precision) -
IN
mteb-results-hosted-huggingface
MTEB per-dataset results are published as JSON at https://huggingface.co/datasets/mteb/results and the leaderboard is at https://huggingface.co/spaces/mteb/leaderboard. -
IN
mteb-results-json-url
MTEB benchmark results are published as versioned JSON files at https://huggingface.co/datasets/mteb/results -
IN
mteb-scope-8-task-types-58-datasets-112-languages
MTEB covers 8 task types, 58 datasets, 10 multilingual datasets, and 112 languages total. -
IN
mteb-sickr-approximately-10000-pairs
SICK-R contains approximately 10,000 sentence pairs focused on compositional semantics (lexical, syntactic, and semantic variation). -
IN
mteb-speed-bench-hardware
MTEB speed benchmarks were run on 1× Nvidia A100 80GB with CUDA 11.6 using the STS15 dataset -
IN
mteb-sts-evaluation-metric
STS evaluation in MTEB correlates model-predicted similarity scores with human-annotated ratings using Spearman or Pearson correlation. -
IN
mteb-sts-main-metric
The main metric for MTEB STS and Summarization evaluation is Spearman correlation based on cosine similarity -
IN
mteb-sts-score-range-0-to-5
STS scores in MTEB range from 0 to 5 inclusive, where 5.0 indicates near-identical meaning. -
IN
mteb-summarization-single-dataset
MTEB includes only one summarization dataset (SummEval, evaluated on CNN/DailyMail), making summarization the most imbalanced task in the benchmark. -
IN
mteb-task-types-count
MTEB defines 8 task types: Bitext Mining, Classification, Clustering, Pair Classification, Reranking, Retrieval, Semantic Textual Similarity, and Summarization -
IN
mteb-tatoeba-112-languages
Tatoeba Bitext Mining in MTEB covers 112 languages. -
IN
mteb-text-length-range
MTEB dataset average text lengths range from approximately 32 characters (TwentyNewsgroups) to approximately 1981 characters (MedrxivClusteringP2P). -
IN
mteb-total-datasets-count
MTEB contains 58 constituent datasets, of which 56 are English subsets used in the main results table -
IN
st5-xxl-best-avg-score
ST5-XXL achieves the highest average MTEB English score of 59.51, followed by GTR-XXL (58.97) and SGPT-5.8B-msmarco (58.81) -
IN
sts22-language-pairs
STS22 in MTEB covers 18 crosslingual language pairs -
IN
stsbenchmark-monolingual-english
STSBenchmark in MTEB is a monolingual English Semantic Textual Similarity dataset