Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Maintaining MTEB: Towards Long Term Usability and Reproducibility of Embedding Benchmarks

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that MTEB's long-term usefulness as a benchmark rests on engineering that makes every leaderboard result exactly reproducible, and it introduces a zero-shot score to reveal how much models were trained on benchmark data.

desk verdict An honest engineering report from the MTEB maintainers that deserves referee time, but its 'identical results' claim is asserted, not shown. read the letter →

arxiv 2506.21182 v1 pith:VLAP3GB3 submitted 2025-06-26 cs.CL cs.AIcs.SE

Isaac Chung , Imene Kerboua , Marton Kardos , Roman Solomatin , Kenneth Enevoldsen This is my paper
classification cs.CLcs.AIcs.SE
keywords MTEBtextembeddingbenchmarkreproducibilitymaintenancezero-shotevaluationdatacontaminationversioningcontinuousintegration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a benchmark's long-term value depends on the engineering that keeps it reproducible, usable, and extensible, and that MTEB has been scaled into a multilingual, multimodal ecosystem by getting that engineering right. Its central claim is that exact reproducibility is achieved by versioning every component separately—tasks, datasets, models, and the evaluation code—so that re-running a specified dataset revision against a specified model checkpoint with a specified MTEB version yields identical results on any machine. The paper also introduces a zero-shot score, $z = 1 - n_{\mathrm{train}}/n_{\mathrm{total}}$, that estimates how much of a model's training data overlaps with benchmark tasks, and uses it to argue that many top-ranked models on the legacy English MTEB score highly by training on benchmark data rather than by generalizing. Two case studies show how these mechanisms were developed in response to real failures: the first on measuring out-of-distribution behavior, the second on reproducing self-reported results that required supporting model-specific prefixes, prompts, normalization choices, and extra encoding stages.

What carries the argument

The load-bearing mechanism is a multi-level versioning system with four independent layers: tasks carry their own evaluation protocols and versions; datasets are referenced by a specific revision from their source repository so identical data is loaded every time; models are referenced by a specific checkpoint or API version; and the MTEB package itself uses semantic versioning with compatibility boundaries. Around this core sits a continuous-integration pipeline that validates dataset format and metadata on every contribution, runs unit and integration tests across platforms and Python versions, and routes leaderboard submissions through community review. The zero-shot score $z = 1 - n_{\mathrm{train}}/n_{\mathrm{total}}$ is a separate, deliberately simple mechanism that makes training overlap visible, where $n_{\mathrm{train}}$ counts the benchmark datasets a model was trained on and $n_{\mathrm{total}}$ the total number of benchmark datasets.

What would settle it

Find a model whose training data is known to include some MTEB benchmark splits but whose creators do not disclose that, and show that its zero-shot score is 100% while an independent contamination probe (for example kernel divergence) flags the overlap; that would demonstrate the transparency score fails exactly where it is meant to inform. A second, simpler falsifier is to re-run any published MTEB result under the same pinned versions on different hardware and find a task whose scores diverge, which would contradict the claim of identical results across machines and environments.

Watch

Extended reading notes

Core claim

The paper's central claim is that MTEB's leaderboard results are exactly reproducible because each evaluation is pinned to a task version, a dataset revision from its source repository, a model revision or API version, and a specific MTEB package version, and because deterministic seeds are used for procedures like K-means clustering and linear probing. In Section 3.3 the authors state directly that anyone can reproduce a benchmark result by running the specified dataset version against the specified model version using a specific MTEB version, yielding identical results across different machines and environments. The paper further claims that a zero-shot score $z = 1 - n_{\mathrm{train}}/n_{\mathrm{total}}$, computed from model creators' self-reported training datasets, is a usable transparency signal, and it presents evidence that some of the highest-ranked models on the legacy English MTEB reach their scores by training on benchmark tasks while lower-scoring models may generalize better out of distribution.

Load-bearing premise

The load-bearing premise is that model contributors disclose their training datasets accurately and completely, so the zero-shot score $z = 1 - n_{\mathrm{train}}/n_{\mathrm{total}}$ overstates out-of-distribution generalization whenever training data is omitted, misreported, or unknown to the contributors.

Editorial extensions

If this is right

  • Any result on the MTEB leaderboard can be replayed exactly by re-running the recorded dataset revision, model checkpoint, and MTEB version, so comparisons between models remain stable across time, machines, and environments.
  • New tasks, languages, and modalities can be added without invalidating existing results, because each addition is versioned and gated by automated validation rather than by a single monolithic release.
  • The zero-shot score converts training-data disclosure into a visible per-model number, letting users distinguish high scores earned by training on benchmark tasks from high scores earned by genuine out-of-distribution generalization.
  • Moving from scraped, self-reported model-card results to a centralized, peer-reviewed results repository means leaderboard rankings reflect verified runs rather than unverifiable claims.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reproducibility guarantee as stated in the paper presumes that API-served proprietary models remain unchanged over time; the paper itself flags this trust assumption, and a stricter guarantee would require pinning API versions or snapshots.
  • The zero-shot score would become a stronger transparency signal if it were cross-checked against an independent contamination detector that does not rely on self-reporting, a natural testable extension of the paper's approach.
  • Because the score treats all benchmark datasets equally, it may understate overlap when a model trains on only a few easy or highly similar tasks; weighting datasets by size or difficulty would be a refinement the paper does not explore.
  • The engineering pattern generalizes beyond embeddings: any benchmark whose datasets and evaluation code evolve independently could use the same multi-level versioning plus automated validation to keep older results comparable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper is an experience report from the maintainers of the Massive Text Embedding Benchmark (MTEB). It describes the engineering infrastructure built around the benchmark: multi-level versioning of tasks, datasets, models, and code; seeded execution for stochastic components; a CI pipeline with linting, tests, and dataset/metadata validation; a community-submission workflow with peer review; and a leaderboard that tracks version-specific results. Two case studies are presented: a zero-shot score intended to quantify distributional overlap between training data and benchmark tasks, and a set of fixes (prefix support, per-task prompts, LoRA parameters, embedding normalization, multi-stage encoders) that were needed to reproduce reported model results. The paper argues that these practices let MTEB scale while preserving reproducibility, usability, and extensibility.

Significance. If the reproducibility claims are accurate, this is a useful and timely contribution: benchmarks are increasingly central to model evaluation, and their long-term maintenance is rarely documented. The paper provides concrete, checkable artifacts: links to pull requests and discussions, a detailed description of the CI pipeline, and a structured metadata schema in Table 1. It also gives credit to the community process and is candid about limitations, which is welcome. The main weakness is that the central reproducibility claim is asserted rather than demonstrated: no cross-machine or cross-version reproduction experiment is reported, and the paper itself admits that versioning discipline has been imperfect. The zero-shot score is an interesting transparency tool but rests on self-reported training data with no independent validation. These issues are fixable within the scope of the paper, but they need to be addressed before the central claims can be taken at face value.

major comments (3)
  1. [§3.3, §6, §A.3.2] The strong claim in §3.3 that running a specified dataset version, model version, and MTEB version yields "identical results across different machines and environments" is not supported by any experiment reported in the paper. Section 3.2 mentions seeds for stochastic components, but no cross-machine, cross-OS, or cross-GPU bitwise comparison is shown; the CI tests described in §A.3.2 use mock tasks and synthetic models, not real-task output equivalence. Furthermore, §6 explicitly states that breaking changes were "at times introduced in minor/patch releases," which means that a version string does not currently guarantee a stable evaluation protocol. The authors should either provide evidence that identical results are achieved in practice (e.g., a small reproducibility study on a few tasks across platforms), or substantially weaken the §3.3 claim to say that versioning and seeding reduce variability rather than guarantee identity.
  2. [§5.1] The zero-shot score z = 1 - n_train / n_total is load-bearing for the contamination case study and for the leaderboard's transparency feature, but it depends entirely on self-reported training data in ModelMeta.trainingdatasets. The paper does not validate this self-report against any independent contamination detection method, and it does not discuss what happens when contributors omit or misreport training data or mark training data as "unknown." The formula also counts datasets rather than training examples or distributional similarity, so calling it a measure of "distributional overlap" overstates what it captures. The authors should either add a validation caveat and sensitivity analysis for misreporting, or reframe the metric as a transparency heuristic rather than a contamination estimate.
  3. [§5.2 and overall framing] The paper's central contribution is framed as improving reproducibility, but Section 5.2 provides only process-level evidence: it lists pull requests that added prefix support, prompt handling, LoRA parameters, and other features, and states that "the improved reproducibility increased community confidence." No quantitative before/after comparison is given, such as how many model cards' reported scores were successfully reproduced, how many submissions were rejected or corrected, or what fraction of MTEB results can currently be reproduced. Without such evidence, the reader cannot assess whether the infrastructure changes achieved their stated goal. A small reproduction audit, even on a handful of models and tasks, would materially strengthen the paper.
minor comments (5)
  1. [§A.2] The list of automated checks contains a typo: "4) and dataset availability on Hugging Face" should be "4) dataset availability on Hugging Face."
  2. [§6] The sentence "These refactoring effort did benefit the community" should be "These refactoring efforts did benefit the community," and "We have since then put more emphasis" reads awkwardly; consider "We have since placed more emphasis."
  3. [Figure 2 and §5.1] Figure 2's caption says "the legacy English MTEB," while the text in Section 5.1 refers to "MTEB (English, v2)"; please align the terminology so readers know which benchmark version is plotted.
  4. [§5.1] The variables in the zero-shot formula are not fully defined in the text: n_train presumably counts the number of benchmark datasets (or dataset splits) the model was trained on, and n_total is the total number of datasets in the benchmark. Please state this explicitly and note that the score treats all datasets equally regardless of size.
  5. [Table 1] The table reports parameters "as of version 1.38.4"; since the paper emphasizes versioning, it would be helpful to state the MTEB version at which the table was captured and to note that the schema may have changed since.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an engineering report; its central reproducibility and zero-shot claims do not reduce to fitted inputs or self-citation chains.

full rationale

This paper is a software-engineering and community-maintenance report rather than a derivation-driven scientific claim. The reproducibility claims in Section 3 are supported by versioning, seeds, CI checks, and community review; they are not obtained by fitting a parameter and then renaming the fit as a prediction. The zero-shot score z = 1 - n_train / n_total in Section 5.1 is explicitly a transparency metric based on self-reported training data; it is defined, not derived, and its dependence on contributor honesty is a trust limitation rather than a circular step. The claim in Section 3.3 that pinned versions yield identical results across machines is an unverified assertion, and Section 6 partially undermines it with the admission of semver-breaking changes in minor/patch releases, but an overclaim is not circularity. Self-citations to prior MTEB, MMTEB, and related benchmark papers appear mostly as contextual descriptions of the ecosystem and do not carry the load-bearing engineering argument. No equation in the paper is equivalent to its own input by construction, no fitted parameter is relabeled as a prediction, and no uniqueness theorem is imported from the authors' previous work. The reader's concern about self-reported training data is a validity and trust issue, not a circularity issue. Therefore the paper is self-contained with respect to circularity, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no fitted parameters and no new physical or theoretical entities. The zero-shot score is a derived metric from self-reported counts, not a free parameter. The listed axioms are unverified domain assumptions on which the reproducibility and contamination claims rest.

assumptions (3)
  • domain assumption Model contributors disclose training datasets completely and accurately.
    The zero-shot score in Section 5.1 counts n_train from self-reported training datasets; incomplete disclosure inflates the score.
  • domain assumption Community peer review of pull requests catches methodology errors and unusual results.
    Section 3.3 states submissions undergo community review, but no audit of review effectiveness is provided.
  • domain assumption Automated CI checks meaningfully validate dataset and model metadata.
    Appendix A describes checks based on Pydantic schemas, but there is no evidence connecting these checks to real error prevention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Maintaining MTEB: Towards Long Term Usability and Reproducibility of Embedding Benchmarks." pith.science (2026). https://pith.science/paper/VLAP3GB3

@misc{pith2026250621182,
  author       = {Pith},
  title        = {Pith review of: Maintaining MTEB: Towards Long Term Usability and Reproducibility of Embedding Benchmarks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLAP3GB3}},
  note         = {Machine review of arXiv:2506.21182}
}
read the original abstract

The Massive Text Embedding Benchmark (MTEB) has become a standard evaluation platform for text embedding models. While previous work has established the core benchmark methodology, this paper focuses on the engineering aspects that ensure MTEB's continued reproducibility and extensibility. We present our approach to maintaining robust continuous integration pipelines that validate dataset integrity, automate test execution, and assess benchmark results' generalizability. We detail the design choices that collectively enhance reproducibility and usability. Furthermore, we discuss our strategies for handling community contributions and extending the benchmark with new tasks and datasets. These engineering practices have been instrumental in scaling MTEB to become more comprehensive while maintaining quality and, ultimately, relevance to the field. Our experiences offer valuable insights for benchmark maintainers facing similar challenges in ensuring reproducibility and usability in machine learning evaluation frameworks. The MTEB repository is available at: https://github.com/embeddings-benchmark/mteb

Figures

Figures reproduced from arXiv: 2506.21182 by the authors.

Figure 1
Figure 1. MTEB is designed with a modular architecture that separates concerns into distinct components. Users can select which model, task, or benchmark to run via the Python API or command line interface. The produced results can then be submitted to the results repository, which is read by the leaderboard. MTEB version, yielding identical results across different machines and environments. These reproducibility measures re… view at source ↗
Figure 2
Figure 2. Models’ mean performance against their zero-shot score on the legacy English MTEB. The highest ranking models achieve their scores by training on benchmark tasks, even though models with lower scores might generalize better to out-of-distribution environments. 5.1. Case Study 1: Assessing zero-shot levels. A significant challenge in embedding benchmarks is de￾termining whether models genuinely demonstrate out-of￾dis… view at source ↗
Figure 3
Figure 3. The MTEB Leaderboard offers an expandable collection of embedding benchmarks, with its default set as Multilingual (MMTEB, (Enevoldsen et al., 2025)). For every benchmark, a model performance per model size graph is shown, with an option to show model performance per task type in a separate tab. Models are ranked by Borda Count by default. The MTEB leaderboard also offers options to customize a benchmark by filterin… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Position: Text Embeddings Should Capture Implicit Semantics, Not Just Surface Meaning

    cs.CL 2025-06 conditional novelty 5.0 of 10

    State-of-the-art text embeddings lag far behind on tasks requiring pragmatic inference, stance detection, and social meaning, relative to their strong performance on surface semantic benchmarks.

  2. Llama Nemoretriever Colembed: Top-Performing Text-Image Retrieval Model

    cs.CV 2025-07 conditional novelty 4.0 of 10

    A new family of text-image retrieval models, built from Eagle2 with bidirectional attention and ColBERT-style late interaction, reports state-of-the-art NDCG@5 scores on ViDoRe V1 (91.0) and V2 (63.5).

Reference graph

Works this paper leans on

24 extracted references · 6 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Recent advances in text embedding: A comprehensive review of top-performing methods on the mteb benchmark, 2024

    Cao, H. Recent advances in text embedding: A comprehensive review of top-performing methods on the mteb benchmark, 2024. URL https://arxiv.org/abs/2406.01607

    arXiv 2024

  3. [3]

    Writing style matters: An examination of bias and fairness in information retrieval systems

    Cao, H. Writing style matters: An examination of bias and fairness in information retrieval systems. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining, WSDM ’25, pp.\ 336–344. ACM, March 2025. doi:10.1145/3701551.3703514. URL http://dx.doi.org/10.1145/3701551.3703514

    arXiv 2025

  4. [4]

    K., Khanov, M., Wei, H., and Li, Y

    Choi, H. K., Khanov, M., Wei, H., and Li, Y. How contaminated is your benchmark? quantifying dataset leakage in large language models with kernel divergence, 2025. URL https://arxiv.org/abs/2502.00678

    arXiv 2025

  5. [5]

    Mteb-french: Resources for french sentence embedding evaluation and analysis, 2024

    Ciancone, M., Kerboua, I., Schaeffer, M., and Siblini, W. Mteb-french: Resources for french sentence embedding evaluation and analysis, 2024. URL https://arxiv.org/abs/2405.20468

    arXiv 2024

  6. [6]

    The scandinavian embedding benchmarks: Comprehensive assessment of multilingual and monolingual text embedding

    Enevoldsen, K., Kardos, M., Muennighoff, N., and Nielbo, K. The scandinavian embedding benchmarks: Comprehensive assessment of multilingual and monolingual text embedding. In Advances in Neural Information Processing Systems, 2024. URL https://nips.cc/virtual/2024/poster/97869

    work page 2024

  7. [7]

    O . V., Kundu, A., Bernstorff, M., Xiao, S., Sukhlecha, A., Pahwa, B., Po \'s wiata, R., GV, K. K., Ashraf, S., Auras, D., Pl \

    Enevoldsen, K., Chung, I., Kerboua, I., Kardos, M., Mathur, A., Stap, D., Gala, J., Siblini, W., Krzemi \'n ski, D., Winata, G. I., Sturua, S., Utpala, S., Ciancone, M., Schaeffer, M., Misra, D., Dhakal, S., Rystr m, J., Solomatin, R., C a g atan, \"O . V., Kundu, A., Bernstorff, M., Xiao, S., Sukhlecha, A., Pahwa, B., Po \'s wiata, R., GV, K. K., Ashraf,...

    work page 2025

  8. [8]

    S., Khodadad, M., Saloot, M

    Kasmaee, A. S., Khodadad, M., Saloot, M. A., Sherck, N., Dokas, S., Mahyar, H., and Samiee, S. Chemteb: Chemical text embedding benchmark, an overview of embedding models performance & efficiency on a specific domain. In Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop, PMLR 262:512-531, 2024. URL https://doi.org/10...

    arXiv doi:10.48550/arxiv.2412.00532 2024

Show all 24 references
  1. [9]

    CLIP \_benchmark: CLIP -like model evaluation

    LAION-AI. CLIP \_benchmark: CLIP -like model evaluation. https://github.com/LAION-AI/CLIP_benchmark, 2025. Accessed: 2025-05-03

    2025

  2. [10]

    Q., Xia, W., Yin, Y., Zhang, H., Liu, Y., Wang, Y., and Tang, R

    Li, X., Dong, K., Lee, Y. Q., Xia, W., Yin, Y., Zhang, H., Liu, Y., Wang, Y., and Tang, R. Coir: A comprehensive benchmark for code information retrieval models, 2024. URL https://arxiv.org/abs/2407.02883

    2024 arXiv

  3. [11]

    and Schwartz, R

    Magar, I. and Schwartz, R. Data contamination: From memorization to exploitation. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp.\ 157--165, Dublin, I...

    2022 doi

  4. [12]

    Morris, J. X. and Rush, A. M. Contextual document embeddings, 2024. URL https://arxiv.org/abs/2410.02525

    2024 arXiv

  5. [13]

    Mteb: Massive text embedding benchmark

    Muennighoff, N., Tazi, N., Magne, L., and Reimers, N. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022. doi:10.48550/ARXIV.2210.07316. URL https://arxiv.org/abs/2210.07316

    2022 doi arXiv

  6. [14]

    S., Enevoldsen, K., and Schneider-Kamp, P

    Nielsen, D. S., Enevoldsen, K., and Schneider-Kamp, P. Encoder vs decoder: Comparative analysis of encoder and decoder language models on multilingual nlu tasks. arXiv preprint arXiv:2406.13469, 2024

    2024 arXiv

  7. [15]

    Bias in text embedding models, 2024

    Rakivnenko, V., Maslej, N., Cervi, J., and Zhukov, V. Bias in text embedding models, 2024. URL https://arxiv.org/abs/2406.12138

    2024 arXiv

  8. [16]

    F., Mahendiran, A., Ko, W.-Y., Shandilya, H., Patel, J., Mataciunas, D., OMahony, L., Zhang, M., Hettiarachchi, R., Wilson, J., Machado, M., Moura, L

    Singh, S., Vargus, F., Dsouza, D., Karlsson, B. F., Mahendiran, A., Ko, W.-Y., Shandilya, H., Patel, J., Mataciunas, D., OMahony, L., Zhang, M., Hettiarachchi, R., Wilson, J., Machado, M., Moura, L. S., Krzemiński, D., Fadaei, H., Ergün, I., Okoh, I., Alaagib, A., Mudannayake,...

    2024 arXiv

  9. [17]

    K., Günther, M., Wang, B., Krimmel, M., Wang, F., Mastrapas, G., Koukounas, A., Koukounas, A., Wang, N., and Xiao, H

    Sturua, S., Mohr, I., Akram, M. K., Günther, M., Wang, B., Krimmel, M., Wang, F., Mastrapas, G., Koukounas, A., Koukounas, A., Wang, N., and Xiao, H. jina-embeddings-v3: Multilingual embeddings with task lora, 2024. URL https://arxiv.org/abs/2409.10173

    2024 arXiv

  10. [18]

    BEIR : A heterogeneous benchmark for zero-shot evaluation of information retrieval models

    Thakur, N., Reimers, N., R \"u ckl \'e , A., Srivastava, A., and Gurevych, I. BEIR : A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2),...

    2021

  11. [19]

    Improving text embeddings with large language models

    Wang, L., Yang, N., Huang, X., Yang, L., Majumder, R., and Wei, F. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023

    2023 arXiv

  12. [20]

    Improving text embeddings with large language models

    Wang, L., Yang, N., Huang, X., Yang, L., Majumder, R., and Wei, F. Improving text embeddings with large language models. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long ...

    2024 doi

  13. [21]

    Multilingual e5 text embeddings: A technical report, 2024 b

    Wang, L., Yang, N., Huang, X., Yang, L., Majumder, R., and Wei, F. Multilingual e5 text embeddings: A technical report, 2024 b . URL https://arxiv.org/abs/2402.05672

    2024 arXiv

  14. [22]

    German text embedding clustering benchmark

    Wehrli, S., Arnrich, B., and Irrgang, C. German text embedding clustering benchmark. arXiv preprint arXiv:2401.02709, 2024. URL https://arxiv.org/abs/2401.02709

    2024 arXiv

  15. [23]

    A., Enevoldsen, K., and Muennighoff, N

    Xiao, C., Chung, I., Kerboua, I., Stirling, J., Zhang, X., Kardos, M., Solomatin, R., Moubayed, N. A., Enevoldsen, K., and Muennighoff, N. Mieb: Massive image embedding benchmark. arXiv preprint arXiv:2504.10471, 2025. doi:10.48550/ARXIV.2504.10471. URL https://arxiv.org/abs/2...

    2025 doi arXiv

  16. [24]

    C-pack: Packed resources for general chinese embeddings

    Xiao, S., Liu, Z., Zhang, P., Muennighoff, N., Lian, D., and Nie, J.-Y. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, pp.\ 641–649, New Yor...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.