Pith. sign in

REVIEW 3 major objections 3 minor 56 references

A Simple Ensemble Strategy for LLM Inference: Towards More Stable Text Classification

T0 review · 3 major / 3 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A five-run ensemble of an 8-billion-parameter language model produced lower average error than a single 70-billion-parameter run on restaurant review sentiment, at nearly identical accuracy and less than half the processing time.

desk verdict A practical, re-implementable cost-performance comparison that is undermined by a single-run 70B baseline and missing error bars; still worth a careful referee. read the letter →

arxiv 2504.18884 v2 pith:NWGFEKVI submitted 2025-04-26 cs.CL cs.AI

Junichiro Niimi This is my paper · ORCID
classification cs.CLcs.AI
keywords largelanguagemodelsensembleinferencesentimentanalysismedianaggregationlabelstabilityRMSErestaurantreviewstextclassification
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 claims that running a mid-sized language model several times and combining the answers is a better use of compute than one pass with a much larger model. On a 1,000-review restaurant sentiment task, the five-run ensemble of an 8-billion-parameter model reached a root mean square error of 0.424 and accuracy of 0.778, while a single run of a 70-billion-parameter model reached RMSE 0.521 and accuracy 0.779—an 18.6 percent smaller average error at nearly identical accuracy, in less than half the processing time. The proposed method requires no fine-tuning, and it borrows the common human-annotation practice of aggregating several judgments.

What carries the argument

The load-bearing mechanism is a five-worker median ensemble. A single 8-billion-parameter instruction-tuned model is called five times with different random seeds, each call producing one star rating through a one-shot prompt, and the final prediction is the median of the five ratings. This turns the model's natural output variability into five virtual annotators and cancels single-run outliers. A 4-bit quantized implementation keeps the eight-billion-parameter runs cheap enough that five runs still cost less than one 70-billion-parameter run.

What would settle it

Run the 70-billion-parameter model on the same 1,000 reviews across, say, twenty random seeds, and compare the RMSE of every single run with the five-run 8B ensemble's 0.424. If most 70B seeds beat 0.424, the claimed 18.6 percent improvement is an artifact of one unlucky run.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that stochastic variation in LLM inference can be exploited as a cheap ensemble: five independent generations by the same 8-billion-parameter model, aggregated with the median of the predicted 1–5 star scores, produce more stable and on average less wrong predictions than a single attempt with a 70-billion-parameter model. The paper reports RMSE 0.424 versus 0.521, an 18.6 percent reduction, and accuracy 0.778 versus 0.779, with runtime 27.2 versus 64.9 seconds per review. The median aggregation is robust to the occasional invalid or extreme token, so the ensemble gains do not depend on a scoring scheme that rewards only exactly correct labels.

Load-bearing premise

The comparison treats one 70-billion-parameter run as a fair representative of that model; since LLM outputs vary with seed, a single run could be atypical, and the paper gives no distribution of repeated 70B runs.

Editorial extensions

If this is right

  • If the central comparison holds, practitioners can replace one expensive large-model inference with several cheap mid-model inferences and get smaller average error in less wall-clock time.
  • Because the ensemble treats the model as an off-the-shelf scorer, it can be added to existing LLM annotation pipelines without training or extensive prompt engineering.
  • The method turns a reproducibility nuisance—random seed sensitivity—into an accuracy lever, so it should transfer to other stochastic models and other ordinal or categorical labeling tasks.
  • The RMSE gain with equal accuracy means the ensemble is not merely moving correct labels around; it shrinks the severity of misclassifications, which matters for star ratings and other graded scales.
  • The strategy's benefit should grow when single-run outputs include out-of-scope tokens, because the median discards extreme votes.

Reading between the lines

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

  • A testable extension the paper leaves implicit: repeat the 70B run across several seeds to obtain a distribution, since the claimed 18.6 percent RMSE advantage is currently measured against one sample of the 70B output distribution.
  • The five-seed choice is arbitrary; plotting RMSE as a function of the number of workers would show whether gains saturate quickly and whether a 70B median ensemble would eventually pull ahead.
  • Because cost scales with parameters and with number of runs, the result suggests a compute-budget view of model choice: for a fixed wall-clock or dollar budget, repeated mid-size runs may dominate a single large run on ordinal classification, and checking this on other model families and tasks would test the generality.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes a simple ensemble strategy for LLM-based sentiment classification: run Llama-3-8B-Instruct five times with different random seeds (as 'virtual workers'), aggregate the ordinal 1–5 predictions by the median, and compare the resulting ensemble against a single Llama-3-70B-Instruct run. On a 1,000-review subsample of the Yelp Open Dataset, the paper reports that the 8B median ensemble achieves RMSE 0.424 and accuracy 0.778, while the single 70B run achieves RMSE 0.521 and accuracy 0.779, with the 8B ensemble also being faster (27.2 vs 64.9 seconds per review). The paper frames this as evidence that medium-sized LLMs with multiple inference attempts are more robust and efficient than a single large-model attempt, and it claims an 18.6% RMSE reduction over the 70B baseline.

Significance. If the headline comparison were statistically supported, the result would be practically useful: it would show that cheap repeated inference with a small model can match or beat a much larger model at a fraction of the compute cost, without tuning, fine-tuning, or prompt engineering. The manuscript is commendably concrete: the strategy is simple, the code is released, the evaluation dataset is public, and the number of workers and the median aggregation rule are fixed a priori rather than tuned on the test set. These properties make the result easily checkable and reusable. However, the central quantitative claim currently rests on a single 70B run, and the claimed accuracy improvement is actually a tie, so the magnitude of the headline effect is not yet established.

major comments (3)
  1. [Section 4.3, Table 3] The headline 18.6% RMSE reduction (0.521 vs 0.424) compares a five-run median ensemble against a single Llama-3-70B run. The paper provides no repeated-seed distribution, confidence interval, or significance test for the 70B baseline, even though LLM outputs are stochastic. A different seed or sampling temperature could plausibly move the 70B RMSE into the 0.45–0.55 range, which would shrink or erase the claimed gain. Please report the variance across repeated runs (for both the 70B single model and the 8B ensemble), state the exact seed values and sampling parameters, and run a paired significance test. This is load-bearing for the abstract's central claim.
  2. [Abstract and Section 4.3, Table 3] The abstract says the ensemble produces 'more robust and accurate results' than the large model, but the reported accuracy is 0.778 for the ensemble and 0.779 for the 70B single run: statistically and practically a tie. The evidence for improved accuracy is therefore absent; the supportable claim is about RMSE (and even that currently lacks uncertainty quantification). Please revise the wording to distinguish the supported RMSE claim from the unsupported accuracy claim.
  3. [Section 3.2, Eq. (1)] The paper says 'the model's reproducibility parameter is utilized' and that five seed values create five workers, but it does not report the sampling temperature, top-p, or other decoding parameters used during inference. Since seed alone may not fully determine the sampling distribution across different inference stacks, please state the exact decoding configuration in the methods so that the reported numbers are reproducible and the ensemble diversity mechanism is precisely defined.
minor comments (3)
  1. [Section 4.3] The text reads 'Llama 2 required more inference time than 80B model (Model 3)'; this should be '8B model'.
  2. [Table 3] The row 'Lift from the LLM baseline (8B)' reports -500.0% for time, which is confusing because a negative percentage could be read as a speedup rather than an increase; please clarify the sign convention in the table note.
  3. [Section 4.2] The evaluation uses only accuracy and RMSE; since the task is ordinal, reporting a confusion matrix or per-class error breakdown would help readers see where the median aggregation actually helps.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the ensemble result is an external benchmark against held-out Yelp labels with no fitted parameters and no load-bearing self-citation.

full rationale

The paper's central comparison is an empirical measurement against held-out Yelp ratings. The ensemble method—five seed-valued workers with median aggregation, defined in Eq. (2)—is fixed before evaluation and is not fitted to the test set; no parameter is estimated from the data and then renamed as a prediction. The reported 18.6% RMSE reduction is a direct comparison of two measurement vectors, not a quantity whose definition contains its own output. The only self-citation ([27], Niimi) appears in the related-work sentence noting that most prior work used single-run LLM annotations; it is background context, not a load-bearing premise, and no uniqueness theorem or ansatz is imported from it. The reader-flagged weakness that the single Llama-3-70B run is one stochastic draw without variance reporting is a statistical robustness limitation, not a circularity, because the method and evaluation are externally anchored to held-out labels. No derivation chain reduces to its inputs, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on no fitted numeric parameters. The main modeling choices are the ensemble size of 5, the specific seeds, and the median aggregation rule; these are fixed a priori but not validated by sensitivity analysis. The dataset-memorization assumption is untested. No new entities are introduced.

free parameters (2)
  • Number of workers (ensemble size) = 5
    The ensemble uses exactly five inference runs with seeds 1 through 5; no sensitivity analysis is provided, so the RMSE result depends on this hand-chosen number.
  • Seed values = 1, 2, 3, 4, 5
    The specific seeds are arbitrary. Different seeds could produce a different ensemble estimate, and the paper does not analyze seed sensitivity.
assumptions (3)
  • domain assumption The Yelp test reviews were not included in Llama pretraining
    Stated in Section 4.1: 'it is unlikely that an entire corpus was included for pre-training process of Llama family.' If any test reviews were memorized, one-shot accuracy would be inflated.
  • domain assumption Different seeds produce enough output diversity for a median ensemble to help
    The method relies on stochastic variation across runs (Section 3.2), but the paper does not measure the diversity of the five outputs or test whether five runs are sufficient.
  • domain assumption A single randomly selected one-shot example is representative for all 1,000 reviews
    The prompt includes one annotated example selected at random (Section 3.2). The authors acknowledge in Section 5 that LLM performance is prompt-sensitive, but they do not study prompt variation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Simple Ensemble Strategy for LLM Inference: Towards More Stable Text Classification." pith.science (2026). https://pith.science/paper/NWGFEKVI

@misc{pith2026250418884,
  author       = {Pith},
  title        = {Pith review of: A Simple Ensemble Strategy for LLM Inference: Towards More Stable Text Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NWGFEKVI}},
  note         = {Machine review of arXiv:2504.18884}
}
read the original abstract

With the advance of large language models (LLMs), LLMs have been utilized for the various tasks. However, the issues of variability and reproducibility of results from each trial of LLMs have been largely overlooked in existing literature while actual human annotation uses majority voting to resolve disagreements among annotators. Therefore, this study introduces the straightforward ensemble strategy to a sentiment analysis using LLMs. As the results, we demonstrate that the ensemble of multiple inference using medium-sized LLMs produces more robust and accurate results than using a large model with a single attempt with reducing RMSE by 18.6%.

Figures

Figures reproduced from arXiv: 2504.18884 by the authors.

Figure 1
Figure 1. Basic Prompt For the text-generation task, each base model predicts a single token that follows the prompt, corresponding to the underbar at the end of the prompt. To introduce prob￾abilistic diversity in the responses, the model’s repro￾ducibility parameter is utilized. In most machine learning methods, specifying the initial value of the random num￾ber generator (i.e., seed value), ensures reproducibility in train… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 45 canonical work pages

  1. [1]

    Whence consumer loyalty? Jour- nal of marketing, 63(4 suppl1):33–44, 1999

    Richard L Oliver. Whence consumer loyalty? Jour- nal of marketing, 63(4 suppl1):33–44, 1999

    work page 1999

  2. [2]

    The customer relationship management process: Its measurement and impact on perfor- mance

    Werner Reinartz, Manfred Krafft, and Wayne D Hoyer. The customer relationship management process: Its measurement and impact on perfor- mance. Journal of marketing research, 41(3):293– 305, 2004

    work page 2004

  3. [3]

    Natural lan- guage processing in marketing

    Jochen Hartmann and Oded Netzer. Natural lan- guage processing in marketing. In Artificial intel- ligence in marketing, pages 191–215. Emerald Pub- lishing Limited, 2023. 4 A Simple Ensemble Strategy for LLM Inference A PREPRINT Table 3: Comparison of the accuracy and the processing time Model Name Ensemble RMSE Acc. Time ( s) LLMs

    work page 2023

  4. [4]

    Llama-3-8B-Instruct adopted 0.424 0.778 27.215

    work page

  5. [5]

    Llama-3-70B-Instruct - 0.521 0.779 64.879

    work page

  6. [6]

    Llama-3-8B-Instruct - 0.562 0.749 5.443

    work page

  7. [7]

    Llama-2-7B-chat - 0.860 0.721 5.986 Reference models

    work page

  8. [8]

    BERT (Large) [40] - 0.941 0.639 -

    work page

Show all 56 references
  1. [9]

    Bi-LSTM [41] - 0.930 0.636 -

  2. [10]

    CNN [42] - 1.098 0.596 -

  3. [11]

    V ADER [14] - 1.111 - -

  4. [12]

    Linear SVM [46] - 1.134 0.627 -

  5. [13]

    Bold metric indicates that the model performs better than all references while cell shading represents the highest performance

    Random - 1.850 0.358 - Lift from the best single LLM (70B) 18.6% -0.1% 58.1% from the LLM baseline (8B) 32.7% 3.9% -500.0% from the best reference model 54.4% 17.9% - Note. Bold metric indicates that the model performs better than all references while cell shading represents t...

  6. [14]

    consistent 4 4 4 4 4 4

  7. [15]

    The effect of word of mouth on sales: Online book reviews

    Judith A Chevalier and Dina Mayzlin. The effect of word of mouth on sales: Online book reviews. Jour- nal of marketing research, 43(3):345–354, 2006

    2006

  8. [16]

    Managing marketing decision-making with senti- ment analysis: An evaluation of the main prod- uct features using text data mining

    Erick Kauffmann, Jes ´us Peral, David Gil, Anto- nio Ferr ´andez, Ricardo Sellers, and Higinio Mora. Managing marketing decision-making with senti- ment analysis: An evaluation of the main prod- uct features using text data mining. Sustainability, 11(15):4235, 2019

    2019

  9. [17]

    Ex- tra: Extracting prominent review aspects from cus- tomer feedback

    Zhiyi Luo, Shanshan Huang, Frank F Xu, Bill Yuchen Lin, Hanyuan Shi, and Kenny Zhu. Ex- tra: Extracting prominent review aspects from cus- tomer feedback. In Proceedings of the 2018 Con- ference on Empirical Methods in Natural Language Processing, pages 3477–3486, 2018

    2018

  10. [18]

    An empirical examination of the impact of ewom information on young con- sumers’ online purchase intention: Mediating role of ewom information adoption

    Affifa Sardar, Amir Manzoor, Khurram Adeel Shaikh, and Liaqat Ali. An empirical examination of the impact of ewom information on young con- sumers’ online purchase intention: Mediating role of ewom information adoption. Sage Open, 11(4), 2021

    2021

  11. [19]

    A BERT- Based Multi-Criteria Recommender System for Hotel Promotion Management

    Yuanyuan Zhuang and Jaekyeong Kim. A BERT- Based Multi-Criteria Recommender System for Hotel Promotion Management. Sustainability, 13(14):8039, 2021

    2021

  12. [20]

    An mturk crisis? shifts in data quality and the impact on study results

    Michael Chmielewski and Sarah C Kucker. An mturk crisis? shifts in data quality and the impact on study results. Social Psychological and Person- ality Science, 11(4):464–473, 2020

    2020

  13. [21]

    Data quality in online human-subjects research: Comparisons between mturk, prolific, cloudresearch, qualtrics, and sona

    Benjamin D Douglas, Patrick J Ewell, and Markus Brauer. Data quality in online human-subjects research: Comparisons between mturk, prolific, cloudresearch, qualtrics, and sona. Plos one , 18(3):e0279720, 2023

    2023

  14. [22]

    Is GPT-3 a good data annotator? In Proceedings of the 61st Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 11173–11195

    Bosheng Ding, Chengwei Qin, Linlin Liu, Yew Ken Chia, Boyang Li, Shafiq Joty, and Lidong Bing. Is GPT-3 a good data annotator? In Proceedings of the 61st Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 11173–11195. Association f...

    2023

  15. [23]

    Chatgpt outperforms crowd workers for text-annotation tasks

    Fabrizio Gilardi, Meysam Alizadeh, and Ma ¨el Kubli. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30):e2305016120, 2023. 5 A Simple Ensemble Strategy for LLM Inference A PREPRINT

    2023

  16. [24]

    Automated annotation with gener- ative ai requires validation

    Nicholas Pangakis, Samuel Wolken, and Neil Fasching. Automated annotation with gener- ative ai requires validation. arXiv preprint arXiv:2306.00176, 2023

    2023 arXiv

  17. [25]

    Vader: A parsi- monious rule-based model for sentiment analysis of social media text

    Clayton Hutto and Eric Gilbert. Vader: A parsi- monious rule-based model for sentiment analysis of social media text. In Proceedings of the interna- tional AAAI conference on web and social media , volume 8, pages 216–225, 2014

    2014

  18. [26]

    Seeing the wood for the trees: How machine learning can help firms in iden- tifying relevant electronic word-of-mouth in social media

    Susan AM Vermeer, Theo Araujo, Stefan F Bernrit- ter, and Guda van Noort. Seeing the wood for the trees: How machine learning can help firms in iden- tifying relevant electronic word-of-mouth in social media. International Journal of Research in Mar- keting, 36(3):492–508, 2019

    2019

  19. [27]

    Exploring ewom in online customer reviews: Sen- timent analysis at a fine-grained level

    Qing Sun, Jianwei Niu, Zhong Yao, and Hao Yan. Exploring ewom in online customer reviews: Sen- timent analysis at a fine-grained level. Engineer- ing Applications of Artificial Intelligence, 81:68–78, 2019

    2019

  20. [28]

    Deep learn- ing for sentiment analysis: A survey

    Lei Zhang, Shuai Wang, and Bing Liu. Deep learn- ing for sentiment analysis: A survey. Wiley Inter- disciplinary Reviews: Data Mining and Knowledge Discovery, 8(4):e1253, 2018

    2018

  21. [29]

    Deep Learning for Aspect-Based Senti- ment Analysis: A Comparative Review

    Hai Ha Do, PWC Prasad, Angelika Maag, and Abeer Alsadoon. Deep Learning for Aspect-Based Senti- ment Analysis: A Comparative Review. Expert Sys- tems with Applications, 118:272–299, 2019

    2019

  22. [30]

    Sentiment analysis in the age of generative ai

    Jan Ole Krugmann and Jochen Hartmann. Sentiment analysis in the age of generative ai. Customer Needs and Solutions, 11(1):3, 2024

    2024

  23. [32]

    Neural net- work ensembles

    Lars Kai Hansen and Peter Salamon. Neural net- work ensembles. IEEE transactions on pattern anal- ysis and machine intelligence , 12(10):993–1001, 1990

    1990

  24. [33]

    Ensemble methods in ma- chine learning

    Thomas G Dietterich. Ensemble methods in ma- chine learning. In International workshop on mul- tiple classifier systems, pages 1–15. Springer, 2000

    2000

  25. [34]

    Ensemble based systems in deci- sion making

    Robi Polikar. Ensemble based systems in deci- sion making. IEEE Circuits and systems magazine, 6(3):21–45, 2006

    2006

  26. [35]

    A survey on ensemble learning

    Xibin Dong, Zhiwen Yu, Wenming Cao, Yifan Shi, and Qianli Ma. A survey on ensemble learning. Frontiers of Computer Science, 14:241–258, 2020

    2020

  27. [36]

    A survey of hi- erarchical classification across different application domains

    Carlos N Silla and Alex A Freitas. A survey of hi- erarchical classification across different application domains. Data mining and knowledge discovery , 22:31–72, 2011

    2011

  28. [37]

    Ensemble methods: foundations and algorithms

    Zhi-Hua Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012

    2012

  29. [38]

    Dynamic sentiment analysis with local large language models using majority voting: A study on factors affecting restaurant evaluation

    Junichiro Niimi. Dynamic sentiment analysis with local large language models using majority voting: A study on factors affecting restaurant evaluation. arXiv preprint arXiv:2407.13069, 2024

    2024 arXiv

  30. [39]

    Designing heterogeneous llm agents for financial sentiment analysis

    Frank Xing. Designing heterogeneous llm agents for financial sentiment analysis. ACM Transactions on Management Information Systems, 2024

    2024

  31. [40]

    Ensemble learning for heterogeneous large language models with deep parallel collaboration

    Yichong Huang, Xiaocheng Feng, Baohang Li, Yang Xiang, Hui Wang, Ting Liu, and Bing Qin. Ensemble learning for heterogeneous large language models with deep parallel collaboration. In The Thirty-eighth Annual Conference on Neural Infor- mation Processing Systems, 2024

    2024

  32. [41]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  33. [42]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, An- drew M

    Jason Wei, Maarten Bosma, Vincent Y . Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, An- drew M. Dai, and Quoc V . Le. Finetuned language models are zero-shot learners. In The Tenth Inter- national Conference on Learning Representations, ICLR 2022, 2022

    2022

  34. [43]

    Deep re- inforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep re- inforcement learning from human preferences. Ad- vances in neural information processing systems, 30, 2017

    2017

  35. [44]

    Gerganov

    G. Gerganov. llama.cpp: LLM inference in C/C++. (https://github.com/ggerganov/llama.cpp, accessed May. 26th, 2024), 2022

    2024

  36. [45]

    Yelp Open Dataset, An all-purpose dataset for learning

    Yelp. Yelp Open Dataset, An all-purpose dataset for learning. (https://www.yelp.com/dataset, accessed Nov. 20th, 2023), 2022

    2023

  37. [46]

    Sentiment classification and aspect-based sentiment analysis on yelp reviews using deep learning and word embeddings

    Eman Saeed Alamoudi and Norah Saleh Alghamdi. Sentiment classification and aspect-based sentiment analysis on yelp reviews using deep learning and word embeddings. Journal of Decision Systems , 30(2-3):259–281, 2021

    2021

  38. [47]

    Yelp dataset challenge: Review rat- ing prediction

    Nabiha Asghar. Yelp dataset challenge: Review rat- ing prediction. arXiv preprint arXiv:1605.05362 , 2016

    2016 arXiv

  39. [48]

    Yelp review rating prediction: Machine learning and deep learning models

    Zefang Liu. Yelp review rating prediction: Machine learning and deep learning models. arXiv preprint arXiv:2012.06690, 2020

    2012 arXiv

  40. [49]

    Seriema: A framework to enhance clustering stability, compactness, and separation by fusing multimodal data

    Fillipe dos Santos Silva, Julio Cesar dos Reis, and Marcelo S Reis. Seriema: A framework to enhance clustering stability, compactness, and separation by fusing multimodal data. In International Conference on Applications of Natural Language to Information Systems, pages 394–40...

    2024

  41. [50]

    tiktoken: a fast BPE to- keniser for use with OpenAI’s models

    OpenAI. tiktoken: a fast BPE to- keniser for use with OpenAI’s models . (https://github.com/openai/tiktoken, accessed May. 26th, 2024), 2023. 6 A Simple Ensemble Strategy for LLM Inference A PREPRINT

    2024

  42. [51]

    BERT: Pre-training of deep bidirectional transformers for language understand- ing

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understand- ing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan...

    2019

  43. [52]

    Senti- ment analysis of text based on bidirectional lstm with multi-head attention

    Fei Long, Kai Zhou, and Weihua Ou. Senti- ment analysis of text based on bidirectional lstm with multi-head attention. IEEE Access, 7:141960– 141969, 2019

    2019

  44. [53]

    Impact of convolu- tional neural network and fasttext embedding on text classification

    Muhammad Umer, Zainab Imtiaz, Muhammad Ah- mad, Michele Nappi, Carlo Medaglia, Gyu Sang Choi, and Arif Mehmood. Impact of convolu- tional neural network and fasttext embedding on text classification. Multimedia Tools and Applications , 82(4):5569–5585, 2023

    2023

  45. [54]

    Efficient estimation of word rep- resentations in vector space

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word rep- resentations in vector space. arXiv preprint arXiv:1301.3781, 2013

    2013 arXiv

  46. [55]

    Bag of tricks for efficient text classification

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th Confer- ence of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427–431. Associat...

    2017

  47. [56]

    Codegen2: Lessons for training llms on programming and nat- ural languages

    Erik Nijkamp, Hiroaki Hayashi, Caiming Xiong, Silvio Savarese, and Yingbo Zhou. Codegen2: Lessons for training llms on programming and nat- ural languages. arXiv preprint arXiv:2305.02309 , 2023

    2023 arXiv

  48. [57]

    An im- proved text sentiment classification model using tf-idf and next word negation

    Bijoyan Das and Sarit Chakraborty. An im- proved text sentiment classification model using tf-idf and next word negation. arXiv preprint arXiv:1806.06407, 2018. 7

    2018 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.