Pith. sign in

REVIEW 4 major objections 4 minor 31 references

Can Post-Training Quantization Benefit from an Additional QLoRA Integration?

T0 review · 4 major / 4 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 This paper claims that adding a QLoRA adapter to a 4-bit post-training-quantized LLM restores enough accuracy to match or beat 16-bit full fine-tuning on the business tasks tested.

desk verdict A practical, honest empirical study of PTQ+QLoRA with a real result, but the missing unquantized-QLoRA control leaves the mechanism unproven. read the letter →

arxiv 2502.10202 v1 pith:DN4ZJ44J submitted 2025-02-14 cs.CL

Xiliang Zhu , Elena Khasanova , Cheng Chen This is my paper · ORCID
classification cs.CL
keywords post-trainingquantizationQLoRALo4-bitparameter-efficientfine-tuninglargelanguagemodelsintentclassificationdialoguesummarization
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 asks whether a large language model that has been compressed to 4-bit weights can be brought back to full-quality performance by training a small LoRA adapter on top of it. Across three 7B models, two quantization methods, and both proprietary business transcripts and public benchmarks, the authors report that this quantize-then-adapter pipeline matches or beats standard full-parameter fine-tuning at 16-bit precision on most tasks. The practical stake is that, if true, a 4-bit model plus a small adapter can be deployed instead of a much heavier 16-bit fine-tuned model, at a fraction of the memory and inference cost. The authors also report that the QLoRA step consistently improves plain 4-bit quantization, and that factual consistency of generated summaries is not significantly harmed.

What carries the argument

The engine of the argument is the PTQ-QLoRA pipeline: first full-parameter supervised fine-tuning (SFT) on a 16-bit base model, then 4-bit post-training quantization of the fine-tuned weights, then a QLoRA pass that trains a low-rank adapter on the frozen 4-bit model using the same task data. The QLoRA adapter is the part that carries the recovery: it injects a small number of trainable parameters into the frozen quantized backbone, so the deployed artifact is 4-bit weights plus a compact adapter rather than a full-precision model.

What would settle it

A concrete falsifying test: run the same two-stage fine-tuning recipe on the unquantized 16-bit model, with full SFT followed by a second LoRA or full SFT pass using the same task data but skipping quantization. If that model matches PTQ-QLoRA, the reported gains come from extra training, not from quantization recovery.

Watch

Extended reading notes

Core claim

The central claim is that the PTQ-QLoRA integration—full-parameter supervised fine-tuning, then 4-bit post-training quantization, then a QLoRA fine-tuning pass on the frozen quantized model—delivers task performance that matches or surpasses 16-bit full fine-tuning. In the authors' experiments across three 7B decoder-only base models (LLaMA-2, Qwen2, Mistral-7B-v0.3) and two quantization methods (bitsandbytes and GPTQ), the combined pipeline is either the best or tied for best accuracy on nearly every task, and it consistently improves over plain 4-bit PTQ. The authors interpret this as evidence that the integration is a viable deployment solution for resource-constrained environments without sacrificing task accuracy.

Load-bearing premise

The load-bearing premise is that the benefit of adding QLoRA comes from its interaction with the 4-bit quantized weights, rather than from the extra supervised fine-tuning pass, because the same task data is used for both the initial SFT and the QLoRA step and no control does a second fine-tuning pass on the unquantized model.

Editorial extensions

If this is right

  • If the claim holds, a 4-bit quantized model with a LoRA adapter can replace a 16-bit fully fine-tuned model for the tested business tasks, cutting memory and inference cost.
  • The improvement from PTQ-QLoRA over plain PTQ is statistically significant for both classification and generation metrics, so the adapter step is not a negligible patch.
  • Factual consistency, measured by AlignScore, is not significantly different between PTQ-QLoRA and 16-bit SFT, so the quality gain is not obtained by trading away faithfulness.
  • Neither bitsandbytes nor GPTQ is consistently better, so the choice of quantizer can be made on infrastructure grounds.
  • The pipeline is reproducible on public datasets such as DialogSum, banking77, and bitext, so deployment teams can validate it without proprietary data.

Reading between the lines

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

  • Editorial inference: the experimental design does not isolate quantization recovery from extra training; a control that runs a second fine-tuning pass (LoRA or full) on the unquantized 16-bit model with the same data would tell whether the gain is specific to the quantized setting.
  • Editorial inference: the same pipeline could be tested at other precisions (8-bit, 3-bit) and on instruction-tuned variants; the paper only studies 4-bit and base models, leaving those natural extensions open.
  • Editorial inference: because the LoRA adapter is small and separate from the 4-bit base, one could serve many task-specific adapters on a single shared quantized backbone, turning the method into a low-cost multi-task deployment strategy.
  • Editorial inference: the findings suggest that quantization loss is largely recoverable by gradient updates through the quantized weights, at least at 4-bit; a direct comparison of adapter quality on quantized versus unquantized backbones would test whether the frozen low-precision weights themselves constrain the achievable accuracy.
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 / 4 minor

Summary. The paper proposes a pipeline that applies full-parameter supervised fine-tuning (SFT) to a 16-bit base model, then applies 4-bit post-training quantization (PTQ), and finally runs QLoRA on the quantized model. The authors evaluate this 'PTQ-QLoRA' integration on three 7B models (LLaMA2, Qwen2, Mistral), two quantization methods (bitsandbytes and GPTQ), and a mix of proprietary business tasks and public datasets (DialogSum, banking77, Bitext). The central claims are that PTQ-QLoRA outperforms plain PTQ in nearly all tasks and can match or exceed 16-bit full-parameter fine-tuning in some cases. The paper includes statistical significance tests (Wilcoxon signed-rank) and AlignScore for factual consistency.

Significance. If the causal attribution were established, the paper would offer a practical recipe: a 4-bit quantized model plus a small LoRA adapter can recover quantization-induced accuracy loss and, in some settings, match a full-precision fine-tuned model, which is valuable for deployment in resource-constrained environments. The experimental sweep is broad: three base models, two quantization algorithms, seven tasks, and both generation and classification metrics, with significance testing and a factual consistency check. The main caveat is that the missing control for an additional fine-tuning pass on the unquantized model undermines the specific claim about the quantization–LoRA interaction, although the aggregate empirical pattern is a useful engineering data point.

major comments (4)
  1. [§3.1 and Figure 1] The pipeline applies the same fine-tuning datasets twice: once for full-parameter SFT on the 16-bit model and again for QLoRA on the 4-bit quantized model. The experimental design does not include a control condition in which QLoRA (or a second SFT pass) is applied to the unquantized 16-bit SFT model. Without this control, the gains of PTQ-QLoRA over plain PTQ shown in Tables 2 and 3 cannot be attributed to the interaction between quantization and LoRA; they could arise simply from an additional supervised fine-tuning pass on the same data. This control is required to answer the title question affirmatively.
  2. [§4.3, Tables 2 and 3] The aggregate Wilcoxon tests are presented as evidence that PTQ-QLoRA outperforms PTQ, but the tables contain multiple counterexamples (e.g., Llama2-7b + PTQ-BNB-4bit + QLoRA on DialogSum ROUGE-1 is 0.5737 vs 0.5739 for PTQ-BNB-4bit; Mistral-7b + PTQ-BNB-4bit + QLoRA on DialogSum ROUGE-1 is 0.5716 vs 0.5789 for PTQ-BNB-4bit). Moreover, each reported number is from a single run with no variance or confidence intervals; treating each task-model pair as one observation in a paired test is a fragile basis for the strong claim that the integration 'enhances the accuracy of PTQ' in nearly all tasks.
  3. [Table 4] The QLoRA step uses substantially different learning rates from the SFT step (e.g., Llama2-7B SFT uses 6e-6 while QLoRA uses 2e-4 or 5e-4; Mistral-7B SFT uses 6e-6 while QLoRA uses 5e-4). Therefore the additional training pass is not a controlled 'more of the same' operation but a different optimization regime. Even with a 16-bit QLoRA control, the hyperparameters would need to be matched to isolate the effect of quantization; as reported, the comparison conflates quantization-aware adaptation with simply a different fine-tuning schedule.
  4. [§1, Contribution 1] The contribution statement 'delivering task performance that matches or surpasses 16-bit full fine-tuning' is stronger than the evidence in Tables 2 and 3. Several PTQ-QLoRA rows fall below the corresponding SFT-16bit row (e.g., Llama2-7b + PTQ-BNB-4bit + QLoRA on DialogSum ROUGE-1 is 0.5737 vs SFT-16bit 0.5816; Llama2-7b + PTQ-GPTQ-4bit + QLoRA on Call Outcome F1-micro is 0.8493 vs SFT-16bit 0.8554). The abstract's 'in some cases' is a more accurate phrasing; the contribution should be revised to match the data.
minor comments (4)
  1. [Table 3] The header 'bitext_custcomer_support' contains a typo; it should be 'bitext_customer_support'.
  2. [§6 Limitations] 'Autometrics' appears to be a typo for 'automatic metrics'; please correct it.
  3. [§A.3 / Table 4] The paper does not report LoRA rank, LoRA alpha, or other QLoRA-specific hyperparameters (e.g., dropout, target modules), which are needed for reproducibility.
  4. [§2 Background] The sentence 'the comparison of QLoRA and 16-bit tuning on other larger language models has not been studied to our best knowledge' is inaccurate given the extensive QLoRA literature since 2023; please rephrase to 'not studied in this particular pipeline' or cite additional comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's PTQ-QLoRA claim rests on direct experimental comparisons, not on a derivation that reduces to its own inputs.

full rationale

This is an empirical systems paper with no fitted law, no analytic derivation, and no parameter calibrated to the reported outcome. The central pipeline (Section 3.1) is: full-parameter SFT on a 16-bit base, 4-bit PTQ of the SFT model, then QLoRA fine-tuning on the quantized model, with evaluation on held-out task splits (Tables 2-3). The claimed benefit of PTQ-QLoRA over plain PTQ is read directly from measured ROUGE/F1/AlignScore differences, and the statistical tests compare those measured values. There is no equation in which a predicted quantity is defined in terms of the target metric, and no fitted parameter is renamed as a prediction. The authors cite their own work? They do not: no reference is authored by Zhu, Khasanova, or Chen, and the paper does not invoke a self-authored theorem to force its choice. The strongest concern is a missing control, stated in the Limitations section: the authors never run QLoRA on the unquantized 16-bit SFT model, so the observed gain could in principle be attributed to a second supervised fine-tuning pass rather than to the quantization-LoRA interaction. That is an experimental confound and an attribution weakness, but it is not circularity under the defined patterns: the input data and the measured outcome are distinct, and the conclusion is not equivalent to an assumption by construction. Accordingly, no circular step is flagged and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No theoretical constructs or entities are introduced; the claim is purely empirical. The free parameters are standard training choices, not fitted constants. The main unexamined burden is the missing control for a second training pass, which sits outside the ledger and is captured in red_flags.

free parameters (3)
  • Per-model learning rates for SFT and QLoRA = 3e-5, 6e-6, 2e-4, 5e-4 (Table 4)
    Hyperparameters chosen by hand for each model and dataset; the paper does not derive them or show sensitivity.
  • LoRA rank and alpha = not reported
    The LoRA configuration is never specified, leaving adapter capacity as an unstated tunable that affects the QLoRA results.
  • Number of training epochs = 2 for both SFT and QLoRA
    Fixed at two epochs without ablations, so the extra training amount is arbitrary.
assumptions (4)
  • domain assumption The quantization tools bitsandbytes and GPTQ behave as documented and allow LoRA training on frozen 4-bit base models.
    Invoked in Section 3.3 without verifying the internals; the paper excludes AWQ due to a claimed CUDA compatibility issue.
  • domain assumption GPT-4 generated labels and summaries are sufficiently accurate ground truth for evaluating task performance.
    Used in Sections 4.1.1 and 4.1.2; internal labels are manually reviewed, but no inter-annotator agreement is reported, and the DialogSum references are GPT-4 outputs.
  • domain assumption ROUGE, F1, and AlignScore adequately measure quality for generation and classification without human evaluation.
    The paper states in Section 6 that a human review could reveal meaningful differences.
  • domain assumption Wilcoxon signed-rank tests are computed over valid paired samples with sufficient power.
    Section 4.3 reports p-values but not the number of paired observations or how tasks and methods are pooled, which is needed to interpret the claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Post-Training Quantization Benefit from an Additional QLoRA Integration?." pith.science (2026). https://pith.science/paper/DN4ZJ44J

@misc{pith2026250210202,
  author       = {Pith},
  title        = {Pith review of: Can Post-Training Quantization Benefit from an Additional QLoRA Integration?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DN4ZJ44J}},
  note         = {Machine review of arXiv:2502.10202}
}
read the original abstract

Large language models (LLMs) have transformed natural language processing but pose significant challenges for real-world deployment. These models necessitate considerable computing resources, which can be costly and frequently unavailable. Model compression techniques such as quantization are often leveraged to alleviate resource demand, but they may have a negative impact on the generation quality. In this study, we explore the integration of 4-bit Post-training Quantization (PTQ) with QLoRA to address these issues. We demonstrate through extensive experiments that this integration outperforms standard PTQ, and in some cases even 16-bit full-parameter fine-tuning on LLMs, validated across proprietary and public datasets with different quantization algorithms. The results demonstrate the efficacy of PTQ-QLoRA integration, offering a viable solution for deploying powerful LLMs in resource-constrained environments without compromising on performance.

Figures

Figures reproduced from arXiv: 2502.10202 by the authors.

Figure 1
Figure 1. Diagram of the PTQ-QLoRA integration. Note that we apply the same fine-tuning datasets twice during full-parameter SFT and QLoRA fine-tuning respectively. • LLaMA2-7B2 : The LLaMA2 series of LLM models (Touvron et al., 2023) developed by Meta. • Qwen2-7B3 . The Qwen2 series LLMs (Bai et al., 2023; Yang et al., 2024) from Alibaba, supporting long context lengths with strong performance on various benchmarks. • Mistra… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 9 canonical work pages

  1. [1]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966

    arXiv 2023

  2. [2]

    I \ n igo Casanueva, Tadas Temcinas, Daniela Gerz, Matthew Henderson, and Ivan Vulic. 2020. https://arxiv.org/abs/2003.04807 Efficient intent detection with dual sentence encoders . In Proceedings of the 2nd Workshop on NLP for ConvAI - ACL 2020. Data available at https://github.com/PolyAI-LDN/task-specific-datasets

    arXiv 2020

  3. [3]

    Yulong Chen, Yang Liu, Liang Chen, and Yue Zhang. 2021. https://doi.org/10.18653/v1/2021.findings-acl.449 D ialog S um: A real-life scenario dialogue summarization dataset . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 5062--5074, Online. Association for Computational Linguistics

    doi:10.18653/v1/2021.findings-acl.449 2021

  4. [4]

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2024. Llm.int8(): 8-bit matrix multiplication for transformers at scale. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc

    work page 2024

  5. [5]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html Qlora: Efficient finetuning of quantized llms . In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, Neur...

    2023

  6. [6]

    Tim Dettmers and Luke Zettlemoyer. 2023. The case for 4-bit precision: k-bit inference scaling laws. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org

    work page 2023

  7. [7]

    Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. https://doi.org/10.18653/v1/P18-1128 The hitchhiker ' s guide to testing statistical significance in natural language processing . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383--1392, Melbourne, Australia. Associ...

    doi:10.18653/v1/p18-1128 2018

  8. [8]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. GPTQ : Accurate post-training compression for generative pretrained transformers. arXiv preprint arXiv:2210.17323

    arXiv 2022

Show all 31 references
  1. [9]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. https://openreview.net/forum?id=tcbBPnfwxS OPTQ: accurate quantization for generative pre-trained transformers . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwa...

    2023

  2. [10]

    Mahoney, and Kurt Keutzer

    Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, and Kurt Keutzer. 2021. https://api.semanticscholar.org/CorpusID:232352683 A survey of quantization methods for efficient neural network inference . ArXiv, abs/2103.13630

    2021 arXiv

  3. [11]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://api.semanticscholar.org/CorpusID:59599816 Parameter-efficient transfer learning for nlp . ArXiv, abs/1902.00751

    2019 arXiv

  4. [12]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022 a . https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  5. [13]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022 b . https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Represen...

    2022

  6. [14]

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. 2018. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vis...

    2018

  7. [15]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  8. [16]

    Jemin Lee, Sihyeong Park, Jinse Kwon, Jihun Oh, and Yongin Kwon. 2024. https://arxiv.org/abs/2409.11055 A comprehensive evaluation of quantized instruction-tuned large language models: An experimental analysis up to 405b . Preprint, arXiv:2409.11055

    2024 arXiv

  9. [17]

    Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. 2021. https://api.semanticscholar.org/CorpusID:231861390 Brecq: Pushing the limit of post-training quantization by block reconstruction . ArXiv, abs/2102.05426

    2021 arXiv

  10. [18]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024 a . https://proceedings.mlsys.org/paper_files/paper/2024/file/42a452cbafa9dd64e9ba4aa95cc1ef21-Paper-Conference.pdf Awq: Activation-aware w...

    2024

  11. [19]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei - Ming Chen, Wei - Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024 b . https://proceedings.mlsys.org/paper\_files/paper/2024/hash/42a452cbafa9dd64e9ba4aa95cc1ef21-Abstract-Conference.html AWQ: activatio...

    2024

  12. [20]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  13. [21]

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. 2020. Up or down? adaptive rounding for post-training quantization. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org

    2020

  14. [22]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. 2021. https://api.semanticscholar.org/CorpusID:235435934 A white paper on neural network quantization . ArXiv, abs/2106.08295

    2021 arXiv

  15. [23]

    Jonas Pfeiffer, Ivan Vuli \'c , Iryna Gurevych, and Sebastian Ruder. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.617 MAD-X : A n A dapter- B ased F ramework for M ulti- T ask C ross- L ingual T ransfer . In Proceedings of the 2020 Conference on Empirical Methods in Natur...

    2020 doi

  16. [24]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2023. https://arxiv.org/abs/1910.10683 Exploring the limits of transfer learning with a unified text-to-text transformer . Preprint, arXiv:1910.10683

    2023 arXiv

  17. [25]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  18. [26]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.acl-long.754 Self-instruct: Aligning language models with self-generated instructions . In Proceedings of the 61st Annual Mee...

    2023 doi

  19. [27]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical met...

    2020

  20. [28]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng...

    2024 arXiv

  21. [29]

    Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. https://aclanthology.org/2023.acl-long.634 A lign S core: Evaluating factual consistency with a unified alignment function . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vo...

    2023

  22. [30]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  23. [31]

    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 gl...

Pith tools

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