Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Visualization for interactively adjusting the de-bias effect of word embedding

T0 review · 3 major / 5 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 Per-category debias strength preserves word categorization accuracy at 0.9325 versus 0.8548 for uniform Hard Debias while still reducing gender bias.

desk verdict A promising per-category debias idea undermined by an unbuilt interactive system and an in-sample evaluation that likely inflates the headline accuracy gain. read the letter →

arxiv 2506.02447 v1 pith:4PUKRKZC submitted 2025-06-03 cs.HC

Arisa Sugino , Takayuki Itoh This is my paper
classification cs.HC
keywords wordembeddinggenderbiasdebiasinginteractivevisualizationJapanesecategoryclassificationHardDebiasParetofront
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 addresses a known failure of word-embedding debiasing: removing gender bias uniformly also removes useful information, degrading downstream model performance. It proposes to make the debias strength a per-category dial, parameterizing Hard Debias with a ratio θ in [0,1], and to show users the resulting accuracy-versus-bias trade-off through confusion matrices and a line chart. On Japanese Wikipedia embeddings with words split into five categories, the paper reports that category-wise tuning reaches 0.9325 accuracy and 0.9325 weighted F1, compared with 0.8548 and 0.8574 for uniform Hard Debias, while still reducing gender bias. A sympathetic reader would care because this makes debiasing a tunable, inspectable choice rather than an all-or-nothing information loss.

What carries the argument

The load-bearing object is a parameterized projection removal: $v_{\text{debias}} = v - \theta\, \mathrm{proj}_G(v)$, where $G$ is the gender axis obtained by PCA over gender word pairs and $\theta \in [0,1]$ scales how much of the gender component is removed ($\theta=0$ no debias, $\theta=1$ Hard Debias). Around this sits a measurement loop: a BERT model fine-tuned on news headlines classifies words into five categories, HNSW approximate nearest-neighbor search links each post-debias vector to its pre-debias category, and the change in category labels before and after debiasing is treated as the semantic damage. A line chart of accuracy/F1 versus $\theta$ plus a confusion-matrix heatmap expose the trade-off, and a multi-objective optimization computes the Pareto front so users can choose debias-versus-performance priorities.

What would settle it

Take the same five-category setup but evaluate the debiased vectors on a human-labeled held-out set of Japanese nouns or on a standard word-similarity benchmark; if per-category $\theta$ does not preserve accuracy better than uniform $\theta=1$ on that independent benchmark, the paper's central accuracy advantage is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that the damage done by gender debiasing to a word embedding is not uniform across semantic categories, and that a category-specific debias ratio can exploit this unevenness. Concretely, it replaces the full projection removal of Hard Debias with $v_{\text{debias}} = v - \theta \cdot \mathrm{proj}_G(v)$, where $\theta \in [0,1]$ is chosen separately for politics, science, business, sports, and entertainment. Using a BERT-based word-category classifier with HNSW nearest-neighbor search to measure semantic drift, the paper finds that politics and science suffer most under full debias. Selecting $\theta$ per category from a visualized Pareto front yields accuracy and weighted F1 scores near 0.9325, versus 0.8548 and 0.8574 for uniform Hard Debias, and the paper reads this as evidence that flexible per-category debiasing suppresses performance degradation while still neutralizing gender bias.

Load-bearing premise

The method assumes that the accuracy of its word-category classifier, trained with labels produced by a large language model, truly reflects how much meaning is lost when a word is debiased.

Editorial extensions

If this is right

  • Users can tune debiasing per category instead of accepting a single global debias, so the choice between fairness and accuracy becomes explicit.
  • The reported numbers imply that category-aware debiasing can cut the accuracy loss of Hard Debias roughly in half, from about 9.4 points to about 1.6 points, while still reducing bias.
  • Visualizing the confusion-matrix delta and the accuracy/F1 curve makes the semantic damage of debiasing visible, so a user can choose a configuration that matches their objectives.
  • The Pareto-front mechanism offers multiple acceptable configurations rather than a single 'correct' debias, which matches the paper's premise that acceptable bias levels are subjective.

Reading between the lines

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

  • Not explored in the paper, but a direct extension: the same per-category $\theta$ dial could be applied to other protected attributes, such as race or religion, whenever a defining subspace and a semantic-drift measure exist.
  • Because the accuracy proxy is an LLM-labeled classifier, an independent human-labeled benchmark would be needed to confirm that the 93%-versus-85% gap reflects real downstream value rather than label bias; this is a testable check, not a claim the paper makes.
  • The paper's flat-accuracy region for moderate $\theta$ hints at a possible 'free lunch' zone where gender bias can be reduced with almost no semantic cost; automating the search for such regions across categories is a plausible next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a visualization-supported method for adjusting the strength of gender debiasing per word category in Japanese word embeddings. It parameterizes Hard Debias with a per-category ratio θ, visualizes accuracy and bias trade-offs, and reports that category-specific θ values preserve downstream categorization accuracy much better than uniform Hard Debias (Table 3: 0.93252 vs. 0.85481). The paper also describes an optimization-based framework for presenting Pareto-optimal θ choices. The experiment uses a Japanese Wikipedia word2vec model, a BERT/HNSW word-category classifier trained on news headlines and GPT-4o-generated labels, and five categories (entertainment, science, politics, sports, business). Main claims are that debiasing impact varies by category and that per-category adjustment reduces performance degradation.

Significance. If the empirical claim were properly validated, the idea of per-category debias strength with interactive visual adjustment would be a useful contribution to the debiasing literature: it addresses a real trade-off and provides a user-centered mechanism. The paper also deserves credit for explicitly parameterizing Hard Debias and for attempting to visualize a bias-performance Pareto front. However, the current evaluation does not establish the central claim: the reported accuracy advantage appears to be an in-sample selection artifact, the evaluation proxy is circular in a way that may favor the method by construction, and the interactive system promised in the title and abstract was not built. These issues are load-bearing, so the contribution as presented is not yet supported.

major comments (3)
  1. [Section 4.4.2 and Table 3] The reported 7.8-point accuracy advantage over Hard Debias is obtained by selecting θ per category after inspecting the accuracy and bias curves produced by the same BERT/HNSW classifier on the same evaluation data. Section 4.4.2 describes how θ was chosen from Figure 6 for each category, and Table 3 then reports accuracy on that same setup. No held-out split, cross-validation, or independent downstream task is described. The θ values are therefore fitted to the evaluation metric, so the 0.93252 vs. 0.85481 gap may reflect selection rather than a genuine preservation of word meaning. This is the central empirical claim and it is not supported by the reported evidence.
  2. [Section 3.3.1] The evaluation metric is circular in an important way. The BERT/HNSW classifier treats the pre-debiased categorization as ground truth and measures how much the debiased vectors change those labels. Any debiasing that moves vectors will appear as degradation, and the proposed method's advantage may partly consist of choosing θ to keep vectors close to their original positions on this particular nearest-neighbor metric. The paper provides no evidence that this classifier is a valid proxy for real downstream task performance, and no independent validation of the GPT-4o-generated labels beyond manual inspection. Without such validation, the accuracy-bias trade-off shown in Figure 6 and the comparisons in Tables 1 and 3 are not established.
  3. [Section 5] The paper's title and abstract promise an interactive visualization method in which 'the user can adjust the parameters based on the visualization results,' but Section 5 explicitly states that no interactive system was developed: 'It was difficult to archive a practical response time ... and therefore we have not develop such an interactive system.' The experiments thus demonstrate static, offline analysis only. The actual interactive-adjustment contribution, which is the paper's stated novelty, is not demonstrated. This is not a minor omission; it is a mismatch between the claimed contribution and the reported implementation.
minor comments (5)
  1. [Section 4.4] The sentence 'the 0.9 value in the top right corner indicates that 0.9 percent of the words in the category business were misclassified to sports' appears to mean 90 percent (or a proportion of 0.9), not 0.9 percent; this should be corrected for readability.
  2. [Section 3.3.2] There is a typo: 'perfonmed' should be 'performed'.
  3. [Section 4.2.1] The phrase 'the the k-means clusterig' contains a typo ('the the' and 'clusterig'); it should be 'k-means clustering'.
  4. [Section 4.2.2] The caption for Figure 1 does not explain which line corresponds to training loss versus validation loss; the text does so, but the figure itself should be self-contained.
  5. [General] The paper would benefit from a precise statement of the dataset sizes (number of words per category), the number of gender word pairs used, and the exact number of test words, as these are important for interpreting the confusion matrices and the reported scores.

Circularity Check

2 steps flagged · score 6.0 of 10

Table 3's 0.93252 accuracy is a selected optimum: the θ values in Table 2 are chosen from the same BERT/HNSW Accuracy/F1 curves that Table 3 reports, and that Accuracy metric is defined as preserving pre-debias labels, so the reported gain over Hard Debias is fitted rather than independently predicted.

  1. fitted input called prediction [Section 4.4.2 (θ selection) and Table 3]
    "We calculated θ by repeating the above steps for each category for which the model performance degradation and the bias effect are comparable. Table 2 compares the results for the “emphasis on debiasing”, the “emphasis on minimizing performance degradation”, and the “emphasis on both equally”."

    The 'above steps' vary θ for a category and read the Accuracy/F1 and bias curves in Figure 6 to pick a θ. Those curves are produced by the Section 3.3.1 BERT/HNSW word-category scorer. Table 3 then reports Accuracy and weighted F1 (0.93252 and 0.93254) from the exact same scorer on the same words after applying the selected θ values, with no held-out split or independent downstream task. The comparison to Hard Debias (0.85481/0.85743) is therefore an optimized point versus a fixed point (θ=1) of the same curve; the improvement is selected, not predicted.

  2. self definitional [Section 3.3.1 (evaluation metric) and Section 3.4 (optimization objective)]
    "By using the pre-debiased categorization results as correct data and comparing them to the categorization results using the post-debiased vectors, it is possible to evaluate the change in word meaning due to debiasing. ... Specifically, the Pareto front is obtained based on a problem setting that maximizes model performance (Accuracy and F1 Score) and minimizes gender bias at the same time."

    The paper's 'model performance' is defined as the fraction of words whose HNSW nearest-neighbor category under the pre-debias index is unchanged after debiasing. Since a debiased word is v - θ·proj_G(v), small θ keeps vectors near their original neighborhoods and thus keeps this accuracy high. Section 3.4 sets the multi-objective optimization to maximize this same Accuracy/F1, so the optimization objective, the evaluation metric, and the reported success criterion are the same quantity by construction. 'Preserving pre-debias labels' is the definition of success, not an externally verified measure of downstream capability.

full rationale

No load-bearing self-citation chain is present: the paper's own self-citations (Tochigi et al., Nakai et al.) appear only in related work, and the debiasing parameterization is original to the paper. The category-dependent effect is an empirical measurement rather than a tautology. However, the headline quantitative claim is partially circular. The θ values in Table 2 are selected by inspecting Accuracy/F1-versus-θ curves from the same BERT/HNSW categorization used for the final evaluation, and Table 3 reports Accuracy/F1 on that same metric with no held-out split or independent downstream task. In addition, the metric itself treats pre-debias HNSW labels as ground truth, so the method's goal (keep vectors close to their original neighborhood) is exactly what is measured. Section 5 also concedes that no interactive system was implemented, so the interactive-adjustment contribution is not demonstrated in the reported experiment. Because the category-specific curves and bias measurements retain some independent empirical content, the result is partially circular rather than fully reducible to the inputs.

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

The central claim rests on standard linear debiasing assumptions, on the BERT/HNSW categorization being a valid proxy for downstream model performance, on the cosine-similarity difference as a bias measure, and on labels generated by GPT-4o with manual filtering. The per-category debias ratios in Table 2 are free parameters fitted to the same data used for evaluation, and no new entities are introduced.

free parameters (3)
  • per-category debias ratio theta_c = politics: 0.7; science: 0.8; business: 0.7; sports: 0.9; entertainment: 0.9 (equal emphasis setting)
    Chosen by inspecting Accuracy/F1 and bias curves on the same evaluation data (Section 4.4.2, Table 2); the Table 3 accuracy depends directly on these choices.
  • BERT fine-tuning epochs = 2
    Selected from Fig. 1 by comparing training and validation loss; this affects the categorization model used as the performance proxy.
  • number of word categories = 5
    Fixed by the authors in Section 4.2.2; the trade-off curves and Table 2 depend on this grouping.
assumptions (4)
  • domain assumption Gender bias in Japanese word embeddings is captured by a linear gender axis computed from translated English gender word pairs, and removing a projection along this axis reduces bias.
    Invoked in Section 3.2.2 when defining G and v_debias; no evidence that the linear subspace assumption holds for Japanese nouns.
  • domain assumption Accuracy of the BERT/HNSW categorization on the pre-debiased labels is a valid proxy for downstream model performance.
    Section 3.3.1 treats pre-debiased categorization results as correct data; the whole trade-off analysis rests on this proxy.
  • domain assumption The difference in cosine similarity to male and female reference words for 'large bias' words in a category is a valid measure of gender bias.
    Used in Section 3.4 and Figure 6; selection of 'large bias' words is not specified, and no validation of this metric is given.
  • ad hoc to paper GPT-4o-generated and manually selected Japanese nouns provide reliable category labels for training and testing.
    Section 3.3.2 and 4.2.2; this data construction is specific to the paper and not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visualization for interactively adjusting the de-bias effect of word embedding." pith.science (2026). https://pith.science/paper/4PUKRKZC

@misc{pith2026250602447,
  author       = {Pith},
  title        = {Pith review of: Visualization for interactively adjusting the de-bias effect of word embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PUKRKZC}},
  note         = {Machine review of arXiv:2506.02447}
}
read the original abstract

Word embedding, which converts words into numerical values, is an important natural language processing technique and widely used. One of the serious problems of word embedding is that the bias will be learned and affect the model if the dataset used for pre-training contains bias. On the other hand, indiscriminate removal of bias from word embeddings may result in the loss of information, even if the bias is undesirable to us. As a result, a risk of model performance degradation due to bias removal will be another problem. As a solution to this problem, we focus on gender bias in Japanese and propose an interactive visualization method to adjust the degree of debias for each word category. Specifically, we visualize the accuracy in a category classification task after debiasing, and allow the user to adjust the parameters based on the visualization results, so that the debiasing can be adjusted according to the user's objectives. In addition, considering a trade-off between debiasing and preventing degradation of model performance, and that different people perceive gender bias differently, we developed a mechanism to present multiple choices of debiasing configurations applying an optimization scheme. This paper presents the results of an experiment in which we removed the gender bias for word embeddings learned from the Japanese version of Wikipedia. We classified words into five categories based on a news corpus, and observed that the degree of influence of debiasing differed greatly among the categories. We then adjusted the degree of debiasing for each category based on the visualization results.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    Yongsu Ahn, Yu-Ru Lin, Fairsight: Visual analytics for fa irness in decision making, IEEE transactions on visualization and computer graphics, Vol. 26, No. 1, pp. 1086– 1095, 2019

    work page 2019

  2. [2]

    5454–5476, 2020

    Su Lin Blodgett, Solon Barocas, Hal Daumé III and Hanna W al lach, Language (Tech- nology) is Power: A Critical Survey of “Bias” in NLP, Proceed ings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 5454–5476, 2020. 20 Arisa Sugino, Takayuki Itoh

    work page 2020

  3. [3]

    4356–4364, 2016

    Tolga Bolukbasi, Kai-W ei Chang, James Zou, Venkatesh Sal igrama and Adam Kalai, Man is to Computer Programmer as W oman is to Homemaker? Debia sing W ord Em- beddings, Proceedings of the 30th International Conferenc e on Neural Information Pro- cessing Systems, pp. 4356–4364, 2016

    work page 2016

  4. [4]

    Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgen- stern, Duen H

    Alexander A. Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgen- stern, Duen H. Chau, FairVis: Visual analytics for discover ing intersectional bias in ma- chine learning, IEEE Conference on Visual Analytics Scienc e and Technology (V AST), pp. 46–56, 2019,

    work page 2019

  5. [5]

    Florent Cabric, Margrét V. Bjarnadóttir, Meng Ling, Guðb jörg Linda Rafnsdóttir, Petra Isenberg, Eleven Years of Gender Data Visualization: A Step Towards More Inclusive Gender Representation, IEEE Transactions on Visualizatio n and Computer Graphics, Vol. 30, No. 1, pp. 316–326, 2023

    work page 2023

  6. [6]

    561–582, 2024

    Hannah Chen, Yangfeng Ji and David Evans, Addressing Both Statistical and Causal Gender Fairness in NLP Models, Findings of the Association f or Computational Lin- guistics: NAACL 2024, pp. 561–582, 2024

    work page 2024

  7. [7]

    Jacob Devlin, Ming-W ei Chang, Kenton Lee, Kristina Touta nova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understan ding, Proceedings of the 2019 Conference of the North American Chapter of the Associa tion for Computational Linguistics: Human Language Technologies, Vol. 1, pp. 4171 –4186, 2019

    work page 2019

  8. [8]

    67–73, 2 018

    Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain and Lu cy Vasserman, Measur- ing and Mitigating Unintended Bias in Text Classification, P roceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pp. 67–73, 2 018

    work page 2018

Show all 28 references
  1. [9]

    11–21, 2018

    Yanai Elazar and Yoav Goldberg, Adversarial Removal of De mographic Attributes from Text Data, Proceedings of the 2018 Conference on Empirical M ethods in Natural Lan- guage Processing, pp. 11–21, 2018

    2018

  2. [10]

    Enoch Opanin Gyamfi, Yunbo Rao, Miao Gou, Yanhua Shao, deb 2viz: Debiasing gender in word embedding data using subspace visualization, Eleve nth International Conference on Graphics and Image Processing, 2019

    2019

  3. [11]

    5491–5501, 2020

    Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton, Kellie W ebster, Yu Zhong and Stephen Denuyl, Social Biases in NLP Models as Barriers f or Persons with Dis- abilities, Proceedings of the 58th Annual Meeting of the Ass ociation for Computational Linguistics, pp. 5491–5501, 2020

    2020

  4. [12]

    14914–14921, 2024

    Bar Iluz, Yanai Elazar, Asaf Yehudai and Gabriel Stanovs ky, Applying Intrinsic Debi- asing on Downstream Tasks: Challenges and Considerations f or Machine Translation, Proceedings of the 2024 Conference on Empirical Methods in N atural Language Pro- cessing, pp. 14914–14921, 2024

    2024

  5. [13]

    Kazuki Kobayashi, Ken W akita, Performance Degradation of NLP Models Caused by Bias Removal, The 36th Annual Conference of the Japanese Soc iety for Artificial Intel- ligence, 2022

    2022

  6. [14]

    5502– 5515, 2020

    Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim , Ruslan Salakhutdinov and Louis-Philippe Morency, Towards Debiasing Sentence Re presentations, Proceedings of the 58th Annual Meeting of the Association for Computatio nal Linguistics, pp. 5502– 5515, 2020

    2020

  7. [15]

    189–202, 2019

    Kaiji Lu, Piotr Mardziel, Fangjing W u, Preetam Amanchar la and Anupam Datta, Gender Bias in Neural Natural Language Processing, Logic, L anguage, and Security, pp. 189–202, 2019

    2019

  8. [16]

    Malkov, Dmiyry A

    Yu A. Malkov, Dmiyry A. Yashunin, Efficient and robust appr oximate nearest neighbor search using Hierarchical Navigable Small W orld graphs, IE EE Transactions on Pattern Analysis and Machine Intelligence, Vol. 42, pp. 824–836, 20 16

  9. [17]

    3111–3119, 2013

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, J eff Dean, Distributed Rep- resentations of W ords and Phrases and their Compositionali ty, Advances in Neural Information Processing Systems 26, pp. 3111–3119, 2013

    2013

  10. [18]

    Tomas Mikolov, Kai Chen, Gref Corrado, Jeff Dean, Efficient Estimation of W ord Rep- resentations in Vector Space, The workshop papers accepted to the International Con- ference on Learning Representations (ICLR), 2013

    2013

  11. [19]

    178–183, 2023

    Yuki Nakai, Takayuki Itoh, Hidekazu Takahashi, Satoshi Nakashima, Tetsu Yamamoto, Hierarchical Data Visualization of Gender Difference: Appl ication to Feeling of Tem- perature, 27th International Conference Information Visu alisation, pp. 178–183, 2023. Visualization for inter...

    2023

  12. [20]

    Archit Rathore, Sunipa Dev, Jeff M. Phillips, Vivek Sriku mar, Yan Zheng, Chin-Chia Michael Yeh, Junpeng W ang, W ei Zhang, Bei W ang, VERB: Visual izing and Inter- preting Bias Mitigation Techniques for W ord Representatio ns, ACM Transactions on Interactive Intelligent System...

    2024

  13. [21]

    Shintaro Sakai and Yasuhiro Suzuki, Evaluation of Gende r Bias in Japanese W ord Em- bedding Model, The 37th Annual Conference of the Japanese So ciety for Artificial In- telligence, 2023

    2023

  14. [22]

    449–482, 2 023

    Nurullah Sevim, Furkan Şahinuç and Aykut Koç, Gender bia s in legal corpora and debiasing it, Natural Language Engineering, pp. 449–482, 2 023

  15. [23]

    Smith and Luke Zettlemoyer, E valuating Gender Bias in Machine Translation, Proceedings of the 57th Annual Meetin g of the Association for Computational Linguistics, pp

    Gabriel Stanovsky, Noah A. Smith and Luke Zettlemoyer, E valuating Gender Bias in Machine Translation, Proceedings of the 57th Annual Meetin g of the Association for Computational Linguistics, pp. 1679–1684, 2019

    2019

  16. [24]

    1630–1640, 2019

    Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElS herief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-W ei Chang and William Yang W a ng, Mitigating Gen- der Bias in Natural Language Processing: Literature Review , Proceedings of the 57th Annual Meeting of the Assoc...

    2019

  17. [25]

    Ami Tochigi, Takayuki Itoh, Xiting W ang, Visualization of Bias of Machine Learning for Content Recommendation, IEEE VIS, Posters, 2021

    2021

  18. [26]

    Natkamon Tovanich, Pierre Dragicevic, Petra Isenberg, Gender in 30 Years of IEEE Visualization, IEEE Transactions on Visualization and Com puter Graphics, Vol. 28, No. 1, pp. 497507, 2021

    2021

  19. [27]

    Qianwen W ang, Zhenhua Xu, Zhutian Chen, Yong W ang, Shixia Liu, Huamin Qu, Visual analysis of discrimination in machine learning, IEEE Trans actions on Visualization and Computer Graphics, Vol. 27, No. 2, pp. 1470–1480, 2020

    2020

  20. [28]

    5443–5453, 2020

    Tianlu W ang, Xi Victoria Lin, Nazneen Fatema Rajani, Bry an McCann, Vicente Or- donez and Caiming Xiong, Double-Hard Debias: Tailoring W ord Embeddings for Gender Bias Mitigation, Proceedings of the 58th Annual Meeting of t he Association for Com- putational Linguistics, pp. ...

    2020

Pith tools

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