Pith. sign in

REVIEW 2 major objections 6 minor 42 references

HABIB_TAZ at SemEval-2026 Task 11: Disentangling Formal Logic from Content via Synthetic Training and Multi-Objective Optimization

T0 review · 2 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Training on rule-generated syllogisms across all 256 Aristotelian forms, with a multi-objective loss that penalizes plausibility bias, drives content effects to zero on three of four benchmark subtasks.

desk verdict The abstract's ST4 result is the English-only ST2 model on a ChatGPT translation of the test set, not the multilingual system described in §4.1 — a reporting flaw that undercuts the paper's headline claim. read the letter →

arxiv 2607.14349 v1 pith:M3GKJRDB submitted 2026-07-15 cs.CL cs.AI

Abdullah Shaikh , Zain Naqi , Taha Zahid , Sandesh Kumar , Abdul Samad This is my paper
classification cs.CLcs.AI
keywords contenteffectssyllogisticreasoninglogicalvaliditysyntheticdatadistributionallyrobustoptimizationbiaspenaltymultilingualconsistencyregularization
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 seeks to show that formal logical validity can be fully disentangled from real-world plausibility in a trained classifier, without relying on LLM-generated training data. The researchers build a rule-based generator that exhaustively covers all 256 Aristotelian syllogistic moods and figures across 12 languages, then fine-tune a multilingual transformer with a loss that combines distributionally robust optimization over logic-plausibility groups, a differentiable penalty on confidence gaps between plausible and implausible items, and KL-consistency between simple and complex phrasings. They report perfect ranking scores (100.0) with 0.00% content bias on English, noisy English, and clean multilingual subtasks, and 89.06% accuracy with 2.89% bias on the noisy multilingual subtask. The significance is that content effects—the tendency of models to favor real-world plausibility over logical form—can be removed by training on formal structure and directly optimizing the bias metric.

What carries the argument

The key machinery is the pairing of a rule-based generator covering all 256 Aristotelian syllogistic moods and figures with a multi-objective loss L = L_DRO + λ_bias L_bias + γ L_KL. The generator creates syllogisms from hypernym/hyponym chains for plausible cases, cross-chain or random nouns for implausible cases, symbolic gibberish variables, and language-specific templates, avoiding LLM-injected noise. The loss combines adaptive Group DRO over validity×plausibility subgroups (with per-language subgroups for multilingual runs), a differentiable penalty derived from mean confidence differences across quadrants, and KL divergence between output distributions of simple and complex rephrasings

What would settle it

Take a trained ST1 or ST3 model and evaluate it on a set of syllogisms that preserve logical form but are paraphrased by humans or a generative model (different surface wording, noun choices, and sentence order). If accuracy falls or Total Content Effect rises above zero, the perfect scores are an artifact of matching the synthetic template distribution rather than proof of content-independent reasoning.

Watch

Extended reading notes

Core claim

The central claim is that a model can judge syllogistic validity independently of whether its content is plausible, provided it is trained on a synthetic dataset covering the complete space of logical forms and optimized with a loss that explicitly rewards uniformity across plausibility groups. On the English, noisy-English, and clean multilingual subtasks the system achieves perfect ranking scores with zero measured content bias; on the hardest noisy-multilingual subtask it reaches 89.06% accuracy with 2.89% bias. The authors attribute the residual gap to cross-lingual variation and to reduced training data (≈4% of the full dataset) caused by compute limits, and they show in ablation that K

Load-bearing premise

The load-bearing premise is that the synthetic rule-based generator—covering the 256 syllogistic schemes, hypernym-based plausibility, language-specific templates, and injected distractors—produces training data that matches the distribution of the official test data in all 12 languages and both noisy conditions; if the test data came from a different scheme, the perfect scores and zero bias would reflect template overlap rather than genuine disentangling.

Editorial extensions

If this is right

  • If the zero-bias results hold up, content effects in syllogistic reasoning can be eliminated without LLM-generated data, replacing costly LLM augmentation with scripted templates.
  • The clean-multilingual perfect score implies the formal reasoning ability transfers across the 12 languages when the generator supplies language-specific templates and translated nouns.
  • The noisy subtasks demonstrate that the same model can jointly select relevant premises and classify validity, a capability needed for multi-premise reasoning beyond syllogisms.
  • The scheduling of the bias penalty and consistency anchor is load-bearing: ablations show DRO plus bias alone degrades performance, so debiasing works only when paired with a regularizer and warm-up.
  • The ST4 cross-lingual analysis indicates KL-consistency makes predictions robust to language shift by anchoring on simple English syllogisms, offering a recipe for multilingual reasoning.

Reading between the lines

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

  • We infer that part of the perfect scores on ST1–ST3 may reflect overlap between the synthetic template distribution and the official test-generation distribution; the paper's own translation experiment shows a sharp performance drop when the input distribution shifts, so zero TCE is not yet demonstrated for naturally written arguments.
  • A natural next probe is to apply the same debiasing loss to an instruction-tuned large language model with the same synthetic examples; the current system is a fine-tuned encoder, so the recipe's transfer to generative models remains open.
  • The binary plausibility labeling (hypernym vs cross-chain nouns) is a coarse proxy for real-world plausibility; extending the generator to graded plausibility or to argument forms other than syllogisms would test whether the disentangling generalizes.
  • Given that ST4 used only ≈4% of the generated data due to compute limits, the paper implies—though the authors do not test it—that scaling the same pipeline to the full dataset may close the remaining cross-lingual gap; that is a cheap, testable prediction.
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

2 major / 6 minor

Summary. This paper presents HABIB_TAZ, the system submitted to SemEval-2026 Task 11 on disentangling formal logic from content in syllogistic reasoning. The authors combine mDeBERTa-v3 backbones (with a DeBERTa-v3-large variant for English noisy subtasks), a rule-based synthetic data generator covering 256 Aristotelian moods/figures, WordNet-based plausibility manipulation, DeepL-based multilingual expansion, HANS-inspired syntactic robustness data, and a multi-objective loss consisting of group DRO, a scheduled differentiable bias penalty, and KL-consistency regularization. They report perfect Ranking Scores on Subtasks 1-3 and a 6th-place result on Subtask 4, with a detailed ablation table and public code/data release.

Significance. If the reported results are accurate and the ST4 attribution is clarified, this is a strong empirical contribution: it indicates that synthetic rule-based training plus bias-aware optimization can produce near-zero content effects on an external multilingual syllogism benchmark. The paper's concrete strengths include official shared-task evaluation, a detailed ablation study, and the release of the generation engine and codebase. The bias penalty is a differentiable surrogate aligned with the TCE metric, so the optimization is loss/metric alignment rather than circular reasoning. The main weakness is a serious inconsistency in the ST4 result reporting, which currently prevents verification of the central multilingual claim.

major comments (2)
  1. [Abstract; §5.1; Table 2] The abstract reports ST4 as: 89.06% Accuracy/F1, 2.89% Bias, 37.78 Ranking Score, 6th rank. The only row in Table 2 with these numbers is 'ST2 Best Model (ET)' (89.06 / 89.06 / 2.89 / 37.78), placed under the ST4 section. §4.1 says the ST4 architecture uses microsoft/mdeberta-v3-base, whereas the ST2 large system uses DeBERTa-v3-large-mnli-fever-anli-ling-wanli; Appendix C.4 likewise specifies ST4 as mDeBERTa-v3-base. The paper never states that the official ST4 submission was evaluated on an English translation or that the ST2 large model was applied to ST4. If the official submission used the English-translated test, the 'multilingual disentangling' claim for ST4 is unsupported. If the table is mislabeled, the true official ST4 configuration and scores must be reported. This is a load-bearing discrepancy: the central claim about the most complex subtask cannot be verified.
  2. [§5.1] Even setting aside the row misplacement, the text says 'the Bias configuration achieved a Ranking Score of 37.78 during official evaluation' for ST4, but Table 2 lists Bias (MT)=25.98 and Bias (ET)=41.43. No listed configuration has 37.78 except the ST2 Best Model (ET) row. Please clarify the exact official submission: model checkpoint, test language (original multilingual vs English translation), and which row in Table 2 corresponds to it. Without this, the reported 6th rank and 2.89% bias cannot be attributed to the described multilingual system.
minor comments (6)
  1. [Table 2] The row 'ST2 Best Model (ET)' appears in the ST4 block; move it to a separate post-hoc table or rename it clearly as a diagnostic evaluation of an ST2 model on translated ST4 data.
  2. [§4.4] The sentence 'Stratification is performed across to ensure balanced...' is missing an object. Specify which variables (quadrant, language, noise condition) were used for stratification.
  3. [Appendix B] B.1 says 'As described in Section 3.2', but the data strategy is described in §4.2. Correct the cross-reference.
  4. [§4.6] The differential learning-rate formula 'LR_backbone = 0.1 × LR_head' is not reconciled with the hyperparameter tables, which list a single learning rate per subtask. State which learning rate is reported and how the differential rate is applied.
  5. [§4.3] Dynamic Scheduling is described qualitatively. Provide the exact schedule (e.g., linear ramp over epochs) for lambda_bias and the DRO step size so the experiments can be reproduced.
  6. [Table 2 / §5.2] No repeated-seed variance is reported. For the non-perfect numbers, especially the ST4 comparisons (e.g., 25.50 vs 40.35), single runs may not support the qualitative conclusions. Add standard deviations or state the number of seeds used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: predictions are externally benchmarked; the ST4 reporting inconsistency is a correctness issue, not a circular derivation.

full rationale

The paper's claimed derivation chain is not circular at the equation level. The differentiable bias penalty L_bias is a soft confidence analogue of the Total Content Effect metric, but TCE is computed from hard accuracies on the official test set and is reported as an outcome; optimizing a differentiable surrogate of one's evaluation metric is standard loss/metric alignment, not a definitional reduction. The ablations (e.g., DRO+Bias raising TCE to 4.26% in ST2 while Vanilla gives 2.13%) show the loss and metric are not trivially identical. The synthetic data generator shares the same syllogistic form family as the task (256 Aristotelian moods/figures, WordNet, DeepL), which creates distributional overlap but is an empirical generalization concern, not a reduction of the predictions to their inputs. Relevant-premise labeling and KL consistency are ordinary supervised and regularization objectives. There are no load-bearing self-citations: the cited syllogistic scheme, DRO, HANS, and NLI checkpoints are external, machine-checkable or independently published. The serious problem in the paper is internal consistency rather than circularity: the abstract's ST4 numbers (89.06% Acc/F1, 2.89% bias, 37.78 S_rank) exactly match the Table 2 row labeled 'ST2 Best Model (ET)' (89.06 / 89.06 / 2.89 / 37.78), not any ST4 mDeBERTa-v3-base row, and Section 5.1 says 'the Bias configuration achieved a Ranking Score of 37.78 during official evaluation' while Table 2 lists Bias(ET)=41.43 and DRO+Bias(ET)=40.35. This means the central 'multilingual ST4' result is unsupported as reported; that is a verification/correctness flaw, but it is not an equation-level circularity.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central engineering result rests on a large set of design choices. The most consequential are the loss-weight hyperparameters, the DRO schedule, and the domain assumptions about the synthetic generator matching the official data distribution. None of these is a derived constant; the paper provides no sensitivity analysis for the main hyperparameters.

free parameters (7)
  • lambda_bias_max = 2.0 for ST1/ST2/ST4, 2.5 for ST3
    Controls the strength of the differentiable bias penalty; scheduled from 0; chosen per subtask without sensitivity analysis.
  • gamma_KL = 0.5
    Weight on the KL-divergence consistency loss; chosen by hand; in ST3 the KL component was not in the best configuration.
  • DRO step-size eta
    Adaptive group-weight update step; initialized at 0 and increased after epoch 1, but the actual value/schedule is not reported.
  • DRO warmup = 1 epoch
    The model trains with standard loss for the first epoch before DRO and bias penalty are enabled.
  • learning rate = 2e-5 (ST1/ST3), 5e-6 (ST2/ST4)
    Differential LR with backbone at 0.1x head; chosen per subtask.
  • batch size / gradient accumulation = 32, 6+10, 144, 72 depending on subtask
    Effective batch sizes differ by subtask; chosen for GPU memory and stability.
  • epochs = 10 (ST1/ST2), 3 (ST3/ST4)
    Early stopping patience 2 with min delta 0.001 also used.
assumptions (6)
  • domain assumption The 256 Aristotelian moods and figures (Bertolazzi et al., 2024) cover the logical structures used in the official SemEval task.
    The synthetic generator is built on this scheme; if the official test uses other logical forms, the training distribution is mismatched.
  • domain assumption mDeBERTa-v3 / DeBERTa-v3-large NLI-pretrained checkpoints provide suitable representations for syllogistic validity classification.
    Architecture choice is empirical and not derived; performance depends on the pretrained representations.
  • domain assumption WordNet hypernym/hyponym chains and random cross-chain nouns correctly instantiate plausible and implausible content.
    Used to control plausibility in the synthetic data; errors in this mapping could leak plausibility cues.
  • domain assumption DeepL translation plus language-specific rule-based templates preserve logical quantifiers and validity across 12 languages.
    Multilingual expansion relies on this; translation artifacts or template errors would break cross-lingual validity.
  • domain assumption HANS samples improve syntactic robustness without biasing the logical training signal.
    HANS is added to prevent lexical-overlap heuristics; its distributional effect on the final model is not separately isolated from the main synthetic data.
  • domain assumption The Ranking Score / TCE metric as defined by the task organizers is a sound measure of content effect.
    The paper optimizes a surrogate of this metric and reports it as the measure of disentangling; the metric itself is not scrutinized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HABIB_TAZ at SemEval-2026 Task 11: Disentangling Formal Logic from Content via Synthetic Training and Multi-Objective Optimization." pith.science (2026). https://pith.science/paper/M3GKJRDB

@misc{pith2026260714349,
  author       = {Pith},
  title        = {Pith review of: HABIB_TAZ at SemEval-2026 Task 11: Disentangling Formal Logic from Content via Synthetic Training and Multi-Objective Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3GKJRDB}},
  note         = {Machine review of arXiv:2607.14349}
}
read the original abstract

While Large Language Models (LLMs) excel in many general NLP tasks, their formal reasoning capabilities are often compromised by content effects, demonstrating a measurable bias towards real-world plausibility. In this paper, we present our system for SemEval-2026 Task 11, which evaluates the ability of models to disentangle formal logic from content across 12 languages with and without distractor premises. We address this challenge using mDeBERTa-v3 networks fine-tuned on a synthetic, rule-based dataset of syllogistic schemes to avoid the semantic noise of LLM-augmented data. To explicitly decouple plausibility from logical structure, our training pipeline employs a multi-objective loss function combining Adaptive Group Distributionally Robust Optimization (DRO), a scheduled differentiable bias penalty, and KL-Divergence consistency regularization. Our system achieved #1 ranks and perfect Ranking Scores (100.0) with 0.00% bias and 100.0% accuracy on Subtask 1 (English), Subtask 2 (Noisy English), and Subtask 3 (Multilingual). On the highly complex Subtask 4 (Noisy Multilingual), the system achieved the 6th rank with 89.06% Accuracy and F1-score, alongside a limited 2.89% Bias and a 37.78 Ranking Score. Our dataset generation engine and codebase are publicly available to facilitate future work on robust logical reasoning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 1 linked inside Pith

  1. [1]

    Cheng and H

    F. Cheng and H. Li and F. Liu and R. van Rooij and K. Zhang and Z. Lin , title =. arXiv preprint , volume =

  2. [2]

    Quan and M

    X. Quan and M. Valentino and L. A. Dennis and A. Freitas , title =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =. 2024 , note =

    2024

  3. [3]

    Wan and Z

    W. Wan and Z. Yang and Y. Chen and C. Luo and R. Wang and K. Cai and N. Kang and L. Lin and K. Wang , title =. arXiv preprint , volume =

  4. [4]

    Valentino and G

    M. Valentino and G. Kim and D. Dalal and Z. Zhao and A. Freitas , title =. arXiv preprint , volume =

  5. [5]

    Wu and Y

    X. Wu and Y. Bu and others , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  6. [6]

    Kim and M

    G. Kim and M. Valentino and D. Dalal and Z. Zhao and A. Freitas , title =. arXiv preprint , volume =

  7. [7]

    Patel and M

    N. Patel and M. Kulkarni and M. Parmar and A. Budhiraja and M. Nakamura and N. Varshney and C. Baral , title =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    2024

  8. [8]

    Bertolazzi and A

    L. Bertolazzi and A. Gatt and A. M. Gliozzo and G. Greco and A. Madotto and M. Valentino , title =. arXiv preprint , volume =

Show all 42 references
  1. [9]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [10]

    Publications Manual , year = "1983", publisher =

    1983

  3. [11]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    1981

  4. [12]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of. 2007 , url=

    2007

  5. [13]

    Dan Gusfield , title =. 1997

    1997

  6. [14]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [15]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =. 2005 , url=

    2005

  8. [16]

    and Tukey, John W

    Cooley, James W. and Tukey, John W. , journal=. An algorithm for the machine calculation of complex. 1965 , url=

    1965

  9. [17]

    International Conference on Learning Representations , year=

    DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing , author=. International Conference on Learning Representations , year=

  10. [18]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    XNLI: Evaluating Cross-lingual Sentence Representations , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    2018

  11. [19]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  12. [20]

    arXiv preprint arXiv:2505.12189 , year =

    Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steering , author =. arXiv preprint arXiv:2505.12189 , year =

  13. [21]

    Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference

    Kim, Geonhee and Valentino, Marco and Freitas, Andre. Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.525

    2025 doi

  14. [22]

    SemEval-2026 Task 11: Disentangling Content and Formal Reasoning in Language Models

    Valentino, Marco and Ranaldi, Leonardo and Pucci, Giulia and Ranaldi, Federico and Freitas, Andr. SemEval-2026 Task 11: Disentangling Content and Formal Reasoning in Language Models. Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026). 2026

    2026

  15. [23]

    2024 , eprint=

    Language models show human-like content effects on reasoning tasks , author=. 2024 , eprint=

    2024

  16. [24]

    2024 , eprint=

    Evaluating the Deductive Competence of Large Language Models , author=. 2024 , eprint=

    2024

  17. [25]

    2024 , eprint=

    TrustLLM: Trustworthiness in Large Language Models , author=. 2024 , eprint=

    2024

  18. [26]

    2023 , eprint=

    DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing , author=. 2023 , eprint=

    2023

  19. [27]

    2018 , eprint=

    XNLI: Evaluating Cross-lingual Sentence Representations , author=. 2018 , eprint=

    2018

  20. [28]

    Less Annotating, More Classifying -- Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and

    Laurer, Moritz and van Atteveldt, Wouter and Casas, Andreu Salleras and Welbers, Kasper , journal=. Less Annotating, More Classifying -- Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and. 2022 , url=

    2022

  21. [29]

    2018 , eprint=

    A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference , author=. 2018 , eprint=

    2018

  22. [30]

    2015 , eprint=

    A large annotated corpus for learning natural language inference , author=. 2015 , eprint=

    2015

  23. [31]

    2020 , eprint=

    Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization , author=. 2020 , eprint=

    2020

  24. [32]

    The Annals of Mathematical Statistics , volume=

    On information and sufficiency , author=. The Annals of Mathematical Statistics , volume=. 1951 , publisher=

    1951

  25. [33]

    CoRR , volume =

    Chen Zhu and Yu Cheng and Zhe Gan and Siqi Sun and Tom Goldstein and Jingjing Liu , title =. CoRR , volume =. 2019 , url =. 1909.11764 , timestamp =

    2019 arXiv

  26. [34]

    , title =

    Miller, George A. , title =. Commun. ACM , month = nov, pages =. 1995 , issue_date =. doi:10.1145/219717.219748 , abstract =

    1995

  27. [35]

    2009 , publisher=

    Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit , author=. 2009 , publisher=

    2009

  28. [36]

    Translation Artifacts in Cross-lingual Transfer Learning

    Artetxe, Mikel and Labaka, Gorka and Agirre, Eneko. Translation Artifacts in Cross-lingual Transfer Learning. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.618

    2020 doi

  29. [37]

    Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference

    McCoy, Tom and Pavlick, Ellie and Linzen, Tal. Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019

    2019

  30. [38]

    2019 , eprint=

    Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference , author=. 2019 , eprint=

    2019

  31. [39]

    2023 , url =

    OpenAI , title =. 2023 , url =

    2023

  32. [40]

    2026 , howpublished =

    2026

  33. [41]

    Advances in Neural Information Processing Systems 32 , pages =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , author =. Advances in Neural Information Processing Systems 32 , pages =. 2019 , publisher =

    2019

  34. [42]

    Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu a...

    2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.