Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

LGAI-EMBEDDING-Preview Technical Report

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A LoRA-tuned 7B model can rank among the best general-purpose text embedders on the new 41-task English benchmark, without any architectural change.

desk verdict A 7B LoRA-tuned embedder reportedly ranks near the top of MTEB English v2, but the paper's zero-shot framing is undercut by training on MTEB task splits and the soft-label loss never appears in the equation. read the letter →

arxiv 2506.07438 v2 pith:SJTERNCM submitted 2025-06-09 cs.CL

Jooyoung Choi , Hyun Kim , Hansol Jang , Changwook Jun , Kyunghoon Bae , Hyewon Choi , Stanley Jungkyu Choi , Honglak Lee
show 1 more author
Chulmin Yun
This is my paper
classification cs.CL
keywords textembeddingsin-contextlearningsoftlabelinghard-negativeminingLoRAfine-tuningMTEBEnglishv2retrievaldecoder-onlyLLM
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

The paper tries to show that a general-purpose embedding model can be built cheaply: no full fine-tuning, no bidirectional attention, no pooling surgery. It combines in-context learning with soft supervision and adaptive hard-negative mining on a Mistral-7B backbone, and reports that the resulting model ranks among the top systems on the 41-task MTEB(English, v2) benchmark by Borda count, beating several larger or fully fine-tuned baselines. If true, the practical consequence is that task-conditioned prompting plus high-quality teacher labels can substitute for expensive architectural modifications in embedding generation.

What carries the argument

Three mechanisms carry the argument. First, in-context prompting: each query is prefixed with a task instruction and few-shot query-passage demonstrations, and the embedding is read from the last token's hidden state, so the model adapts to task format without weight changes. Second, soft-label distillation: an in-house two-stage retriever-reranker pipeline (ANNA) scores query-document pairs using BM25, dense retrieval, and a cross-encoder; the three rankings are merged by reciprocal rank fusion, and the resulting continuous scores replace binary labels in an InfoNCE contrastive loss. Third, adaptive margin-based hard-negative mining: negatives whose teacher score exceeds 95% of the corresponding positive's score are excluded, and a random subset of the top-K remaining negatives is used, reducing false negatives and training noise.

What would settle it

Re-run the comparison after removing the single summarization task, since Borda rank is sensitive to one task; if the model no longer outranks Seed1.5-Embedding, the headline ranking rests on one task. Separately, a hard-label-only training run on the same data and negatives would show whether soft labels from ANNA are the source of the retrieval and STS gains or whether the prompting and mining alone account for them.

Watch

Extended reading notes

Core claim

The paper's central claim is that a decoder-only 7B model, LoRA fine-tuned for one epoch and never re-architected, can reach a top Borda rank on MTEB(English, v2) by conditioning embeddings on task instructions and few-shot examples, training on continuous relevance scores distilled from a retriever-reranker pipeline, and mining hard negatives with a 95% margin relative to each positive. Its reported results include second-best retrieval (66.18), second-best STS (86.69), second-best pair classification (88.67), and first-best summarization (38.93), with a mean task score of 74.12 that is slightly below Seed1.5-Embedding's 74.76 but a better overall Borda ranking.

Load-bearing premise

The results depend on the in-house ANNA teacher's relevance scores being unbiased, with ANNA never trained or tuned on the benchmark's 41 tasks; if those scores already encode the benchmark answers, the reported generalization is inherited from the teacher, and the 95% margin is assumed rather than derived.

Editorial extensions

If this is right

  • Because the recipe changes no architecture and only LoRA-adapts the backbone, the same instruction-plus-distillation pipeline should transfer to other decoder-only language models at modest training cost.
  • Soft labels replace binary positives and negatives in the contrastive loss, so the model can rank candidates by degree of relevance rather than only by membership in a positive set.
  • The adaptive 95% margin filter removes negatives that score nearly as high as the positive, which should reduce false-negative noise and stabilize contrastive training.
  • Borda aggregation across 41 tasks ranks this model above competitors with higher average scores, implying the model is more consistently good across task types rather than dominant on a few.
  • The top summarization score, despite the model's compact 7B size relative to commercial baselines, suggests the instruction-following prompt is especially effective for single-output language generation tasks.

Reading between the lines

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

  • This reader's inference: the decisive check would be an ablation that replaces ANNA soft labels with the original hard labels while keeping the same data and negatives; if retrieval and STS scores drop substantially, the teacher signal is the active ingredient, and if they do not, the prompting and mining are doing the work.
  • This reader's inference: the 95% margin is a single global proportion applied to every task's positive score; tuning it per task or making it adaptive during training could separate true hard negatives from near-duplicates more cleanly, and would be a natural next experiment.
  • This reader's inference: because the report discloses no details of ANNA's training data or benchmark overlap, the portability of the recipe to other settings depends on whether the teacher is itself a general retriever or has been exposed to the evaluation tasks; this is a testable condition, not a settled fact.
  • This reader's inference: the Borda-based claim implies that average-score leaderboards can be misleading for comparing generalist embedders; re-ranking the published per-task numbers under other aggregation rules would show how robust the top position is.
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

4 major / 6 minor

Summary. The manuscript describes LGAI-Embedding-Preview, a text embedding model that fine-tunes Mistral-7B with LoRA and a contrastive objective, augmented by in-context instructions, soft-label supervision from an in-house retrieval pipeline (ANNA), and adaptive margin-based hard-negative mining. The model is evaluated on the MTEB(English, v2) benchmark, and the authors report that it ranks among the top models by Borda score, with second place in retrieval and STS and first place in summarization, despite using only parameter-efficient fine-tuning and no task-specific per-task fine-tuning.

Significance. If the reported results are accurate and the methodology is reproducible, the paper would show that LoRA fine-tuning combined with in-context prompting and teacher-derived soft labels can match or beat much larger fully fine-tuned models across a broad set of embedding tasks. The use of the external MTEB(eng,v2) benchmark and the Borda count protocol is a genuine strength, and the training mixture is largely based on public datasets. However, the current manuscript does not disclose the overlap between its training data and the evaluation tasks, does not show how the soft labels enter the loss, and does not provide details of the ANNA teacher pipeline. These omissions make it impossible at present to separate genuine cross-task generalization from in-distribution supervision or potential leakage.

major comments (4)
  1. [Section 3 / Table 1 / Section 5.1] The paper does not report the overlap between its training datasets and the evaluation tasks of MTEB(eng,v2). Section 3 explicitly states that non-retrieval training data are drawn from "classification, clustering, reranking, and STS within the MTEB tasks," and Table 1 lists datasets such as Banking77Classification, EmotionClassification, STS12, STS22, STSBenchmark, and several Arxiv/Reddit/StackExchange clustering datasets. Section 5.1 says MTEB(eng,v2) "emphasizes zero-shot evaluation by excluding tasks commonly used during fine-tuning." These two statements are in tension. The authors must either provide a precise mapping showing that none of the Table 1 task families appear in the 41-task MTEB(eng,v2) evaluation set, or acknowledge that the model was fine-tuned on training splits of benchmark tasks, which would invalidate the zero-shot generalization claim made in the abstract and conclusion. Please report the complete task list of MTEB(eng,v2) and an explicit overlap analysis.
  2. [Section 4.1, Eq. (1)] The contrastive loss in Eq. (1) is the standard InfoNCE loss with positives and hard negatives, and the teacher's continuous relevance scores do not appear anywhere in the objective. The text states that soft targets serve as fine-grained supervision signals, but no distillation term, soft-label weighting, or alternative mechanism is shown. As written, the soft-label framework has no effect on the training dynamics. Please provide the actual loss function that incorporates the soft scores, or revise the description to clarify how the soft labels influence the training (for example, through a distillation loss or by adjusting the logits before the softmax). Without this, the central methodological novelty of the paper is unsupported.
  3. [Section 4.1] The ANNA IR pipeline, which provides both the soft labels and the hard-negative scores, is referenced as [32] but its training data, architecture, and any possible overlap with MTEB(eng,v2) are not disclosed. Because every soft label and every hard-negative margin in this work is derived from ANNA, the teacher's behavior fully determines the student's supervision. If ANNA or its training data were exposed to MTEB(eng,v2) tasks or test labels, the reported evaluation would partly measure leakage rather than generalization. Please provide a detailed description of ANNA's training data and an explicit statement that neither ANNA nor its training corpora include MTEB(eng,v2) evaluation examples or labels.
  4. [Section 5.2, Table 2] The ranking claims rest on a single evaluation run with no error bars, confidence intervals, or significance tests. In particular, the reported first-place summarization score (38.93 vs 38.28 for gemini-embedding-001) and the overall Borda rank advantage over Seed1.5-Embedding (whose mean score is higher, 74.76 vs 74.12) may be within run-to-run variation. The authors should report per-task standard deviations or, where feasible, results over multiple seeds, so that readers can assess whether the Borda ranking is stable and whether the stated 'top-performing' claims are statistically meaningful.
minor comments (6)
  1. [Section 4.3] The '95% margin' threshold for hard-negative filtering is introduced without justification or sensitivity analysis; please provide a rationale for this value and report how the results change as the margin is varied.
  2. [Section 4.2] The number of few-shot examples k used in the actual evaluations is never specified; please report the value per task or state that it was kept fixed across all tasks.
  3. [Abstract / Conclusion] The phrase 'without task-specific fine-tuning' is misleading given that the training mixture includes tasks from the same benchmark families as the evaluation. Consider phrasing such as 'without training on MTEB(eng,v2) test labels' or 'without per-task fine-tuning after the single multi-task training run.'
  4. [Section 3, Table 1] Some dataset names are incomplete or nonstandard (e.g., 'ArxivClustering2', 'BiorxivClustering', 'MedrxivClustering'); use the full versioned names and provide links or citations for each dataset.
  5. [Appendix B] The license text in Appendix B is not accompanied by a URL or release repository where the model weights and code can be downloaded; for a technical report that claims reproducible results, a public release link is necessary.
  6. [Section 4.1, RRF formula] The reciprocal rank fusion formula in Section 4.1 does not specify how the lexical, dense, and reranker scores are normalized before combining ranks; please clarify the preprocessing step or note if raw ranks are used directly.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: the reported MTEB(eng,v2) scores rest on external held-out test labels, though the self-cited ANNA teacher and the use of MTEB-suite training tasks create evidence gaps that are not circular reductions.

full rationale

The central claim is an empirical ranking on MTEB(eng,v2), computed from held-out test labels rather than from the paper's own equations. The training chain—public retrieval corpora plus MTEB-suite classification/clustering/reranking/STS training splits, LoRA fine-tuning with an InfoNCE-style contrastive loss, and soft labels/hard negatives from the in-house ANNA pipeline (ref [32], overlapping authors)—does not define any reported score in terms of itself. The paper explicitly states 'we use only the training data and omit test splits to maintain evaluation integrity,' so the benchmark numbers are not equal by construction to the training signals. The ANNA self-citation is used to describe the teacher and negative-mining source, not to prove the benchmark result, and no uniqueness/ansatz argument is imported from it. The training-set overlap with MTEB task families (Table 1 lists STS12/STS22/STSBenchmark and several clustering datasets) and the undisclosed training data of ANNA are validity/contamination concerns about the zero-shot framing, but they are not demonstrated circular reductions under the hard rules; no Eq. X = Eq. Y or fitted-parameter-renamed-as-prediction step can be exhibited. Score 2 reflects the presence of a self-cited in-house teacher without treating it as construction-level circularity.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on a set of training hyperparameters chosen without ablation, an undisclosed teacher pipeline, and a benchmark whose zero-shot status is complicated by training on its own train splits. None of these are fatal by themselves, but together they mean the report's results are not independently auditable from the text.

free parameters (9)
  • margin percentage for hard-negative filtering = 95%
    Used to set max_negative_score_threshold = positive_score x 0.95 in Section 4.3. Chosen by hand; no sensitivity analysis.
  • number of hard negatives per query = 7
    Reported in Section 5.1 without ablation.
  • LoRA rank = 64
    Chosen in Section 5.1; not justified.
  • LoRA alpha = 32
    Chosen in Section 5.1; not justified.
  • learning rate = 1e-4
    Reported in Section 5.1; not justified.
  • loss temperature tau = not reported
    Appears in the InfoNCE loss in Section 4.1, but its value is unspecified and affects training dynamics.
  • top-K for random hard-negative sampling = not reported
    Mentioned in Section 4.3 as 'top-K ranked negatives' without specifying K.
  • few-shot example count k = not reported
    The input format in Section 4.2 uses k support examples, but k is never specified.
  • RRF constant k = 60
    Fixed to 60 per the RRF reference [33]; standard but still a hyperparameter.
assumptions (4)
  • domain assumption MTEB(eng, v2) is a valid zero-shot benchmark for the claimed generalization.
    The evaluation section (Section 5.1) treats leaderboard scores as evidence of generalization, but training on MTEB train splits (Section 3) makes the zero-shot status ambiguous.
  • domain assumption The ANNA IR teacher provides unbiased relevance scores for soft supervision and hard-negative mining.
    Section 4.1 relies on an in-house pipeline with no disclosed training data; if it overlapped MTEB v2, the supervision signal would leak evaluation information.
  • standard math Soft-label supervision improves student generalization as claimed by Mandal et al. [8].
    The paper adopts the theoretical result as background without proving it; cited as reference [8] in Section 4.1.
  • standard math The InfoNCE loss with hard negatives is an appropriate objective for the embedding model.
    Used in Sections 4.1 and 5.1; a standard contrastive learning assumption with no comparison to alternatives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LGAI-EMBEDDING-Preview Technical Report." pith.science (2026). https://pith.science/paper/SJTERNCM

@misc{pith2026250607438,
  author       = {Pith},
  title        = {Pith review of: LGAI-EMBEDDING-Preview Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJTERNCM}},
  note         = {Machine review of arXiv:2506.07438}
}
read the original abstract

This report presents a unified instruction-based framework for learning generalized text embeddings optimized for both information retrieval (IR) and non-IR tasks. Built upon a decoder-only large language model (Mistral-7B), our approach combines in-context learning, soft supervision, and adaptive hard-negative mining to generate context-aware embeddings without task-specific fine-tuning. Structured instructions and few-shot examples are used to guide the model across diverse tasks, enabling strong performance on classification, semantic similarity, clustering, and reranking benchmarks. To improve semantic discrimination, we employ a soft labeling framework where continuous relevance scores, distilled from a high-performance dense retriever and reranker, serve as fine-grained supervision signals. In addition, we introduce adaptive margin-based hard-negative mining, which filters out semantically ambiguous negatives based on their similarity to positive examples, thereby enhancing training stability and retrieval robustness. Our model is evaluated on the newly introduced MTEB (English, v2) benchmark, covering 41 tasks across seven categories. Results show that our method achieves strong generalization and ranks among the top-performing models by Borda score, outperforming several larger or fully fine-tuned baselines. These findings highlight the effectiveness of combining in-context prompting, soft supervision, and adaptive sampling for scalable, high-quality embedding generation.

Figures

Figures reproduced from arXiv: 2506.07438 by the authors.

Figure 1
Figure 1. Flowchart illustrating the conversion process from Natural Language Inference (NLI) data to STS-style [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the ANNA IR pipeline. The system employs a two-stage retrieval framework. In the first stage, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. QZhou-Embedding Technical Report

    cs.CL 2025-08 conditional novelty 4.0 of 10

    QZhou-Embedding reports state-of-the-art average scores on MTEB and CMTEB as of August 27, 2025, using a two-stage multi-task pipeline with LLM-based data synthesis.

Reference graph

Works this paper leans on

38 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [32]

    Anna: Enhanced language representation for question answering.ACL 2022, page 121, 2022

    Changwook Jun, Hansol Jang, Myoseop Sim, Hyun Kim, Jooyoung Choi, Kyungkoo Min, and Kyunghoon Bae. Anna: Enhanced language representation for question answering.ACL 2022, page 121, 2022

    work page 2022

  2. [1]

    Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961, 2024

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961, 2024

    arXiv 2024

  3. [2]

    Improving text embeddings with large language models.arXiv preprint arXiv:2401.00368, 2023

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models.arXiv preprint arXiv:2401.00368, 2023

    arXiv 2023

  4. [3]

    Nv-embed: Improved techniques for training llms as generalist embedding models.arXiv preprint arXiv:2405.17428, 2024

    Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models.arXiv preprint arXiv:2405.17428, 2024

    arXiv 2024

  5. [4]

    Making text embedders few-shot learners.arXiv preprint arXiv:2409.15700, 2024

    Chaofan Li, MingHao Qin, Shitao Xiao, Jianlyu Chen, Kun Luo, Yingxia Shao, Defu Lian, and Zheng Liu. Making text embedders few-shot learners.arXiv preprint arXiv:2409.15700, 2024

    arXiv 2024

  6. [5]

    Mmteb: Massive multilingual text embedding benchmark.arXiv preprint arXiv:2502.13595, 2025

    Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, Márton Kardos, Ashwin Mathur, David Stap, Jay Gala, Wissam Siblini, Dominik Krzemi´nski, Genta Indra Winata, et al. Mmteb: Massive multilingual text embedding benchmark.arXiv preprint arXiv:2502.13595, 2025

    arXiv 2025

  7. [6]

    Linq-embed-mistral:elevating text retrieval with improved gpt data through task-specific control and quality refinement

    Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral:elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024

    work page 2024

  8. [7]

    Gecko: Versatile text embeddings distilled from large language models.arXiv preprint arXiv:2403.20327, 2024

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, et al. Gecko: Versatile text embeddings distilled from large language models.arXiv preprint arXiv:2403.20327, 2024

    arXiv 2024

Show all 38 references
  1. [8]

    A theoretical analysis of soft-label vs hard-label training in neural networks.arXiv preprint arXiv:2412.09579, 2024

    Saptarshi Mandal, Xiaojun Lin, and R Srikant. A theoretical analysis of soft-label vs hard-label training in neural networks.arXiv preprint arXiv:2412.09579, 2024

    2024 arXiv

  2. [9]

    Ms marco: A human-generated machine reading comprehension dataset

    Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016

    2016

  3. [10]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christo- pher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    2018 arXiv

  4. [11]

    Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research.Transactions of the Association for Computat...

    2019

  5. [12]

    Squad: 100,000+ questions for machine comprehension of text.arXiv preprint arXiv:1606.05250, 2016

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text.arXiv preprint arXiv:1606.05250, 2016

    2016 arXiv

  6. [13]

    Eli5: Long form question answering.arXiv preprint arXiv:1907.09190, 2019

    Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. Eli5: Long form question answering.arXiv preprint arXiv:1907.09190, 2019

    1907 arXiv

  7. [14]

    Retrieval of the best counterargument without prior topic knowledge

    Henning Wachsmuth, Shahbaz Syed, and Benno Stein. Retrieval of the best counterargument without prior topic knowledge. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 241–251, 2018. 8 APREPRINT- SEPTEMBER17, 2025

    2018

  8. [15]

    Www’18 open challenge: financial opinion mining and question answering

    Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. Www’18 open challenge: financial opinion mining and question answering. InCompanion proceedings of the the web conference 2018, pages 1941–1942, 2018

    2018

  9. [16]

    Fever: a large-scale dataset for fact extraction and verification.arXiv preprint arXiv:1803.05355, 2018

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. Fever: a large-scale dataset for fact extraction and verification.arXiv preprint arXiv:1803.05355, 2018

    2018 arXiv

  10. [17]

    Natural language understanding with the quora question pairs dataset.arXiv preprint arXiv:1907.01041, 2019

    Lakshay Sharma, Laura Graesser, Nikita Nangia, and Utku Evci. Natural language understanding with the quora question pairs dataset.arXiv preprint arXiv:1907.01041, 2019

    1907 arXiv

  11. [18]

    I wish i would have loved this one, but i didn’t–a multilingual dataset for counterfactual detection in product reviews.arXiv preprint arXiv:2104.06893, 2021

    James O’Neill, Polina Rozenshtein, Ryuichi Kiryo, Motoko Kubota, and Danushka Bollegala. I wish i would have loved this one, but i didn’t–a multilingual dataset for counterfactual detection in product reviews.arXiv preprint arXiv:2104.06893, 2021

    2021 arXiv

  12. [19]

    Hidden factors and hidden topics: understanding rating dimensions with review text

    Julian McAuley and Jure Leskovec. Hidden factors and hidden topics: understanding rating dimensions with review text. InProceedings of the 7th ACM conference on Recommender systems, pages 165–172, 2013

    2013

  13. [20]

    Efficient intent detection with dual sentence encoders.arXiv preprint arXiv:2003.04807, 2020

    Iñigo Casanueva, Tadas Temˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli´c. Efficient intent detection with dual sentence encoders.arXiv preprint arXiv:2003.04807, 2020

    2003 arXiv

  14. [21]

    Carer: Contextualized affect representations for emotion recognition

    Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. Carer: Contextualized affect representations for emotion recognition. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 3687–3697, 2018

    2018

  15. [22]

    Learning word vectors for sentiment analysis

    Andrew Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, pages 142–150, 2011

    2011

  16. [23]

    Mtop: A comprehen- sive multilingual task-oriented semantic parsing benchmark.arXiv preprint arXiv:2008.09335, 2020

    Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. Mtop: A comprehen- sive multilingual task-oriented semantic parsing benchmark.arXiv preprint arXiv:2008.09335, 2020

    2008 arXiv

  17. [24]

    Jigsaw unintended bias in toxicity classification

    cjadams, Daniel Borkan, inversion, Jeffrey Sorensen, Lucas Dixon, Lucy Vasserman, and nithum. Jigsaw unintended bias in toxicity classification. https://kaggle.com/competitions/ jigsaw-unintended-bias-in-toxicity-classification, 2019. Kaggle

    2019

  18. [25]

    Tweet sentiment extraction

    Maggie, Phil Culliton, and Wei Chen. Tweet sentiment extraction. https://kaggle.com/competitions/ tweet-sentiment-extraction, 2020. Kaggle

    2020

  19. [26]

    Newsweeder: Learning to filter netnews

    Ken Lang. Newsweeder: Learning to filter netnews. InMachine learning proceedings 1995, pages 331–339. Elsevier, 1995

    1995

  20. [27]

    Tweac: Transformer with extendable qa agent classifiers.arXiv preprint, abs/2104.07081, 2021

    Gregor Geigle, Nils Reimers, Andreas Rücklé, and Iryna Gurevych. Tweac: Transformer with extendable qa agent classifiers.arXiv preprint, abs/2104.07081, 2021

    2021 arXiv

  21. [28]

    Specter: Document-level representation learning using citation-informed transformers.arXiv preprint arXiv:2004.07180, 2020

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S Weld. Specter: Document-level representation learning using citation-informed transformers.arXiv preprint arXiv:2004.07180, 2020

    2004 arXiv

  22. [29]

    Linkso: a dataset for learning to retrieve similar question answer pairs on software development forums

    Xueqing Liu, Chi Wang, Yue Leng, and ChengXiang Zhai. Linkso: a dataset for learning to retrieve similar question answer pairs on software development forums. InProceedings of the 4th ACM SIGSOFT International Workshop on NLP for Software Engineering, pages 2–5, 2018

    2018

  23. [30]

    Semeval-2012 task 6: A pilot on semantic textual similarity.* sem 2012: The first joint conference on lexical and computational semantics—

    Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity.* sem 2012: The first joint conference on lexical and computational semantics—. InProceedings of the Sixth International Workshop on Semantic Evaluation ...

    2012

  24. [31]

    Hale, David Jurgens, and Mattia Samory

    Xi Chen, Ali Zeynali, Chico Camargo, Fabian Flöck, Devin Gaffney, Przemyslaw Grabowicz, Scott A. Hale, David Jurgens, and Mattia Samory. SemEval-2022 task 8: Multilingual news article similarity. In Guy Emerson, Natalie Schluter, Gabriel Stanovsky, Ritesh Kumar, Alexis Palmer,...

    2022

  25. [33]

    Reciprocal rank fusion outperforms condorcet and individual rank learning methods

    Gordon V Cormack, Charles LA Clarke, and Stefan Buettcher. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. InProceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 758–759, 2009

    2009

  26. [34]

    Unsupervised dense information retrieval with contrastive learning.arXiv preprint arXiv:2112.09118, 2021

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning.arXiv preprint arXiv:2112.09118, 2021. 9 APREPRINT- SEPTEMBER17, 2025

    2021 arXiv

  27. [35]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023

  28. [36]

    Nv-retriever: Improving text embedding models with effective hard-negative mining.arXiv preprint arXiv:2407.15831, 2024

    Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. Nv-retriever: Improving text embedding models with effective hard-negative mining.arXiv preprint arXiv:2407.15831, 2024

    2024 arXiv

  29. [37]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  30. [38]

    Agreement

    Pierre Colombo, Nathan Noiry, Ekhine Irurozki, and Stéphan Clémençon. What are the best systems? new perspectives on nlp benchmarking.Advances in neural information processing systems, 35:26915–26932, 2022. A Contributors All authors are listed in alphabetical order by last na...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.