Pith. sign in

REVIEW 4 major objections 7 minor 29 references

HyReC: Exploring Hybrid-based Retriever for Chinese

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HyReC unifies dense, lexicon-based, and learned term-segmentation retrieval in a single Chinese retriever, reporting nDCG@10 of 70.54 on C-MTEB, above the 67.10 of the BGE-M3 hybrid baseline.

desk verdict A novel semantic-union retriever for Chinese whose empirical claims are undercut by training/eval overlap; the idea deserves a careful controlled replication. read the letter →

arxiv 2506.21913 v1 pith:GX5MGTSN submitted 2025-06-27 cs.IR cs.CL

Zunran Wang , Zheng Shenpeng , Wang Shenglan , Minghui Zhao , Zhonghua Li This is my paper
classification cs.IRcs.CL
keywords hybridretrievaldenselexicon-basedChineseinformationsemanticunionoftermsGlobal-Local-AwareEncoderNormalizationModuleC-MTEB
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 argues that hybrid retrieval for Chinese can be improved by letting the model itself learn word boundaries as a 'semantic union of terms' instead of relying on a separate frequency-based segmenter like Jieba. It introduces HyReC, a single model with a dense branch, a lexicon branch, and a learned term-union branch, plus two components: GLAE shares low-level semantics between branches while keeping their representations separate, and NM normalizes the two score types during training so that neither dominates. On the C-MTEB retrieval benchmark, HyReC reports an average nDCG@10 of 70.54 at base scale, beating the 67.10 of the strong BGE-M3 hybrid baseline and 69.53 of BGE-base-zh. If those numbers hold, the method suggests that end-to-end learning of word segmentation is a productive direction for Chinese hybrid retrieval.

What carries the argument

The central objects are (i) the semantic union of terms, a learned word-segmentation output produced by a union projector that assigns B/M/E/S tags to tokens and lets the bagging module merge token weights into word weights; (ii) the Global-Local-Aware Encoder (GLAE), a three-part transformer layout (shared semantic backbone, global-aware lexicon encoder, local-aware dense encoder) that lets the two retrieval branches share low-level semantics without coupling their representations; and (iii) the Normalization Module (NM), which L2-normalizes dense vectors and lexicon term-weight vectors so the two matching scores live on comparable scales during end-to-end training. The final hybrid score is a sum of the lexicon and dense matching scores.

What would settle it

A controlled experiment that fine-tunes BGE-M3-hybrid (or the strongest baseline) on the exact same training sets and recipe as HyReC, then compares on C-MTEB: if the nDCG@10 gap shrinks to under one point or reverses, the proposed components are not the cause. A second check is training HyReC without the semantic-union labeling (e.g., using only Jieba segments) on held-out domains not represented in the training data, to see whether the claimed generalization still holds.

Watch

Extended reading notes

Core claim

HyReC's central claim is that the weaknesses of Chinese hybrid retrieval stem from treating tokenizer terms as the atomic unit of lexicon matching, even though Chinese lacks word boundaries. The paper proposes to learn a semantic union of terms: a union projector labels each token as S/B/M/E so that adjacent tokens can be merged into meaningful words, and the bagging module aggregates token weights within each merged word. Combined with a Global-Local-Aware Encoder, which uses a shared backbone plus separate global-aware lexicon and local-aware dense encoders, and a Normalization Module that scales both matching scores to comparable ranges during training, the model is optimized end-to-end with contrastive losses for both retrieval branches plus a cross-entropy loss for term unions. The reported result is an average nDCG@10 of 70.54 for the base-hybrid model on C-MTEB retrieval, versus 67.10 for BGE-M3-hybrid, with ablations showing each proposed component contributes to the gain.

Load-bearing premise

The load-bearing premise is that the reported gains come from the proposed components rather than from the training data: HyReC is fine-tuned on T2Ranking, DuReader, mMARCO, and CMedQA-v2, which overlap with several C-MTEB retrieval tasks, and the baselines are not trained on the same data, so if most of the improvement comes from benchmark-domain fine-tuning, the central claim collapses.

Editorial extensions

If this is right

  • Chinese hybrid retrieval can be built as a single end-to-end model instead of coupling a separate word-segmentation tool with a retriever.
  • Learned semantic unions improve not only the lexicon branch but also the dense and hybrid scores, suggesting the two branches benefit from shared segmentation knowledge.
  • Training-time normalization of dense and lexicon scores makes hybrid training more stable and avoids arbitrary weights for combining scores.
  • The approach decouples the lexicon and dense encoders, so lexicon-based retrieval can be served with an inverted index and dense retrieval with dot-product operations independently.

Reading between the lines

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

  • The gains may partly reflect training on domains that overlap with the C-MTEB evaluation tasks; a controlled experiment with identical data for baselines would be needed to separate component effect from data effect.
  • The semantic-union idea could transfer to other languages with compounding morphology (e.g., Japanese, Korean, German compounds) if a label source for word boundaries is available.
  • The normalization module's training-time score alignment could be applied to other hybrid retrieval systems even without the semantic-union component.
  • A testable extension: evaluate HyReC on retrieval tasks entirely outside its training domains (e.g., legal or code retrieval) to test whether the learned segmentation generalizes.
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 / 7 minor

Summary. HyReC is an end-to-end hybrid retrieval model for Chinese that unifies three components in a single BERT-style architecture: a dense retriever using the [CLS] embedding, a lexicon-based retriever using per-token term weights, and a learned semantic union of terms obtained by labeling tokens as single/begin/middle/end of words. The model adds a Global-Local-Aware Encoder (GLAE) that shares low-level layers and then splits into separate dense and lexicon encoders, and a Normalization Module (NM) that L2-normalizes both score branches. The total loss is the sum of contrastive losses for dense and lexical retrieval plus a cross-entropy loss for the union labels, where labels come from a Jieba-plus-regex labelling tool. The paper evaluates on the C-MTEB retrieval benchmark and reports nDCG@10 of 70.54 for the base hybrid model versus 67.10 for BGE-M3-hybrid, with ablations attributing gains to NM, GLAE, and the semantic union.

Significance. If the empirical claims hold up under controlled comparison, HyReC is a useful engineering contribution to Chinese hybrid retrieval: it integrates sparse and dense retrieval with a learned term-union module in one model, and the paper provides a reasonably detailed training recipe and a component-level ablation. The GLAE idea of sharing low-level features while separating global and local encoders, and the NM idea of normalizing both branches during training, are sensible and could benefit practitioners. The paper is not a theoretical advance, and its current evidence is weakened by the overlap between the fine-tuning corpus and the evaluation benchmark, by the lack of size-matched or same-recipe baselines, and by the absence of released code or checkpoints. Those issues must be addressed before the central attribution claim can be accepted.

major comments (4)
  1. [Section 4.2 / Appendix A] The comparison with existing baselines is confounded by training/evaluation domain overlap. Appendix A states that high-quality fine-tuning uses T2-Ranking, DuReader, mMARCO, CMedQA-v2, and Multi-CPR; the C-MTEB retrieval evaluation in Table 1 includes T2Retrieval, MMarcoRetrieval, DuRetrieval, CmedqaRetrieval, CovidRetrieval, EcomRetrieval, MedicalRetrieval, and VideoRetrieval. Seven of the eight evaluation tasks therefore draw from the same source datasets as the fine-tuning corpus (Multi-CPR covers Ecom/Medical/Video). The only leakage remark, 'even without CovidRetrieval in the training data,' does not address the other seven tasks. Because the BGE and BGE-M3 baselines were not retrained on the same corpus under the same recipe, Table 1 may reflect in-domain fine-tuning rather than the proposed SU/GLAE/NM components. The ablations in Table 2 are measured inside the same setup and cannot resolve this attribution. Please add controlled same-data baselines, for example BGE-small/base fine-tuned with the HyReC recipe, or evaluate on a strictly held-out, non-overlapping task set.
  2. [Equation (10)] The definition of S_lex(q,p) in Eq. (10) is inconsistent with the text and with the notion of sparse lexical matching. As written, the double sum over all i and j multiplies the weight of every query term with the weight of every passage term, so non-overlapping terms also contribute. The preceding sentence says the score is 'weights multiplications of the common terms shared in the query and the passages,' which would require a sum over the intersection, and the standard sparse-retrieval score is a single sum over shared terms. Please correct the formula, specify whether non-overlapping terms are forced to zero, and state how sparsity is enforced if the summed product form is intended.
  3. [Section 3.1.4 / Eqs. (7)-(8)] The inference procedure for the semantic union is underspecified. The paper states that Uword_j may encompass multiple tokens when Uterm_i belongs to {B, M, E}, and then defines Uword_j and Wword_j by max pooling, but it never gives the decoding algorithm that converts the per-token four-class predictions into word boundaries. Since the bagging module is central to the claimed improvement, please provide the exact decoding and grouping procedure used at inference.
  4. [Section 4.2 / Table 1] The headline comparison against BGE-M3-hybrid is also confounded by model scale: HyReC-base has 153M parameters, while BGE-M3 is a 568M-parameter model, and the two use different training corpora and objectives. The +3.44% margin is therefore not an apples-to-apples comparison. The small-scale comparison similarly lacks a same-recipe baseline. Please include size-matched baselines trained under the identical recipe, or qualify the comparison explicitly as being against a larger, differently trained model.
minor comments (7)
  1. [Section 4.1.2] The heading 'Evalution Metrics' should be 'Evaluation Metrics'.
  2. [Appendix A] The dataset name 'DURreader' should be 'DuReader', and the number '118,944,5 paired texts' appears to be a formatting typo.
  3. [Equation (12)] The denominator contains a parenthesis/braces typo; the term e^{S*(q,p+)}/τ appears with an extra parenthesis.
  4. [Table 2] The ablation table lacks explicit row labels; please make clear which checkmarks indicate the full model and which component is removed in each row.
  5. [Section 4.3.1] The text about training exclusively on the lexicon-based or dense-vector retrieval task does not state whether the GLAE encoders are present in those runs; please clarify the configuration.
  6. [Section 3.3] Please clarify whether the L2 normalization of term-importance vectors is computed over the tokens in the current text or over the full vocabulary.
  7. [General] The paper uses both 'Bge' and 'BGE' inconsistently; please unify the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: HyReC's empirical claims rest on standard supervised training and external benchmarks; the benchmark-domain overlap is a data-attribution concern, not a circular reduction.

full rationale

HyReC is an empirical retrieval system, not a mathematical derivation, and its components are trained with standard objectives. The semantic union module is supervised by labels produced by Jieba plus regular expressions (Section 4.1.1), but those labels are auxiliary supervision for a segmentation-style task; the retrieval scores are computed from learned term weights and dense vectors via contrastive losses (Eqs. 10-14). No equation in the paper defines the reported nDCG@10 in terms of the training labels or fitted parameters by construction. The ablations in Tables 2 and 3 compare architectures and supervision signals, and the gains are empirical rather than forced. The paper contains no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in solely through citation. The main validity threat is that fine-tuning data listed in Appendix A (T2-Ranking, DuReader, mMARCO, CMedQA-v2, Multi-CPR) overlaps several C-MTEB retrieval evaluation tasks, which weakens attribution of the gains to the proposed components; however, this is a benchmark-contamination / data-attribution issue, not circularity. Because the central claims are not equivalent to the inputs by definition or by fit, the circularity score is 0.

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

The central claim rests mainly on the training data recipe and benchmark choice; the only introduced conceptual element is the semantic union of terms, which is supervised by an external Jieba-plus-regex labeling tool, not by a new unobservable entity.

free parameters (4)
  • Temperature tau = 0.05
    Used in contrastive losses for both retrieval branches (Eq 12); chosen by hand, not tuned.
  • Data filtering threshold = 0.43
    Threshold for filtering 160M web text pairs with Text2Vec-Chinese2 in preliminary fine-tuning (Appendix A).
  • Loss weights for Llex, Lden, Lunion = 1, 1, 1
    Total loss in Eq (14) sums the three losses without learned weighting; no tuning is reported.
  • GLAE layer split = small 1/1/1; base 5/7/7
    Number of layers in sharing backbone, global lexicon encoder, and local dense encoder; chosen via ablation in Fig. 3.
assumptions (4)
  • domain assumption The C-MTEB retrieval benchmark is a valid and representative measure of Chinese retrieval performance.
    The paper uses nDCG@10 on C-MTEB as the sole evaluation standard (Section 4.1.2).
  • ad hoc to paper Jieba plus handcrafted regexes provide adequate ground-truth labels for learning the semantic union of terms.
    The labeling tool (Section 4.1.1) defines supervision for the union projector; if these labels are poor, the semantic union component cannot help.
  • domain assumption The tokenizer's subword terms plus predicted merged words cover the matching vocabulary needed for lexicon retrieval.
    The lexicon representation is defined over tokenizer terms and semantic unions (Section 3.1.4); any query term outside this vocabulary receives zero weight.
  • ad hoc to paper Training on datasets that overlap in domain with the C-MTEB retrieval evaluation does not unfairly advantage HyReC over baselines.
    Appendix A lists T2Ranking, DuReader, mMARCO, and CMedQA-v2 as fine-tuning data; these share domains with C-MTEB retrieval tasks, so the comparison assumes no leakage-induced bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyReC: Exploring Hybrid-based Retriever for Chinese." pith.science (2026). https://pith.science/paper/GX5MGTSN

@misc{pith2026250621913,
  author       = {Pith},
  title        = {Pith review of: HyReC: Exploring Hybrid-based Retriever for Chinese},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GX5MGTSN}},
  note         = {Machine review of arXiv:2506.21913}
}
read the original abstract

Hybrid-based retrieval methods, which unify dense-vector and lexicon-based retrieval, have garnered considerable attention in the industry due to performance enhancement. However, despite their promising results, the application of these hybrid paradigms in Chinese retrieval contexts has remained largely underexplored. In this paper, we introduce HyReC, an innovative end-to-end optimization method tailored specifically for hybrid-based retrieval in Chinese. HyReC enhances performance by integrating the semantic union of terms into the representation model. Additionally, it features the Global-Local-Aware Encoder (GLAE) to promote consistent semantic sharing between lexicon-based and dense retrieval while minimizing the interference between them. To further refine alignment, we incorporate a Normalization Module (NM) that fosters mutual benefits between the retrieval approaches. Finally, we evaluate HyReC on the C-MTEB retrieval benchmark to demonstrate its effectiveness.

Figures

Figures reproduced from arXiv: 2506.21913 by the authors.

Figure 1
Figure 1. A example for lexicon-based retrieval. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The architecture of our HyReC. HyReC first utilizes a semantic sharing backbone to extract low-level [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ablation study on the performance of GLAE [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 16 canonical work pages

  1. [1]

    Yang Bai, Xiaoguang Li, Gang Wang, Chaoliang Zhang, Lifeng Shang, Jun Xu, Zhaowei Wang, Fangshan Wang, and Qun Liu. 2020. Sparterm: Learning term-based sparse representation for fast text retrieval. arXiv preprint arXiv:2010.00768

    arXiv 2020

  2. [2]

    Luiz Bonifacio, Vitor Jeronymo, Hugo Queiroz Abonizio, Israel Campiotti, Marzieh Fadaee, Roberto Lotufo, and Rodrigo Nogueira. 2022. https://arxiv.org/abs/2108.13897 mmarco: A multilingual version of the ms marco passage ranking dataset . Preprint, arXiv:2108.13897

    arXiv 2022

  3. [3]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. https://arxiv.org/abs/2402.03216 Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . Preprint, arXiv:2402.03216

    arXiv 2024

  4. [4]

    Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. arXiv preprint arXiv:1910.10687

    arXiv 2019

  5. [5]

    Jacob Devlin, Ming Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, vol. 6. long and short papers: Conference of the North American Chapter of the A...

    work page 2019

  6. [6]

    Thibault Formal, Benjamin Piwowarski, and St \'e phane Clinchant. 2021. Splade: Sparse lexical and expansion model for first stage ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2288--2292

    work page 2021

  7. [7]

    Michael Fuller, Marcin Kaszkiel, Sam Kimberley, Corinna Ng, Ross Wilkinson, Mingfang Wu, and Justin Zobel. 2008. 1 ad-hoc task 1.1 background

    work page 2008

  8. [8]

    Luyu Gao and Jamie Callan. 2021 a . Condenser: a pre-training architecture for dense retrieval. arXiv preprint arXiv:2104.08253

    arXiv 2021

Show all 29 references
  1. [9]

    Luyu Gao and Jamie Callan. 2021 b . Unsupervised corpus aware language model pre-training for dense passage retrieval. arXiv preprint arXiv:2108.05540

    2021 arXiv

  2. [10]

    Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. Coil: Revisit exact lexical match in information retrieval with contextualized inverted list. arXiv preprint arXiv:2104.07186

    2021 arXiv

  3. [11]

    Wei He, Kai Liu, Jing Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yuan Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, Tian Wu, and Haifeng Wang. 2018. https://arxiv.org/abs/1711.05073 Dureader: a chinese machine reading comprehension dataset from real-world applications . Pr...

    2018 arXiv

  4. [12]

    Kaiyu Huang, Degen Huang, Zhuang Liu, and Fengran Mo. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.318 A joint multiple criteria model in transfer learning for cross-domain C hinese word segmentation . pages 3873--3882, Online. Association for Computational Linguistics

    2020 doi

  5. [13]

    Lin, Xueguang Ma, Sheng Chieh Lin, Jheng Hong Yang, Ronak Pradeep, Rodrigo Nogueira, and D

    Jimmy J. Lin, Xueguang Ma, Sheng Chieh Lin, Jheng Hong Yang, Ronak Pradeep, Rodrigo Nogueira, and D. Cheriton. 2021. Pyserini: A python toolkit for reproducible information retrieval research with sparse and dense representations. Proceedings of the 44th International ACM SIGI...

    2021

  6. [14]

    Dingkun Long, Qiong Gao, Kuan Zou, Guangwei Xu, Pengjun Xie, Ruijie Guo, Jian Xu, Guanjun Jiang, Luxi Xing, and Ping Yang. 2022. Multi-cpr: A multi domain chinese dataset for passage retrieval

    2022

  7. [15]

    Mieradilijiang Maimaiti, Yang Liu, Yuanhang Zheng, Gang Chen, Kaiyu Huang, Ji Zhang, Huanbo Luan, and Maosong Sun. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.158 Segment, mask, and predict: Augmenting C hinese word segmentation with self-supervision . pages 2068--2077, ...

    2021 doi

  8. [16]

    Tao Shen, Xiubo Geng, Chongyang Tao, Can Xu, Guodong Long, Kai Zhang, and Daxin Jiang. 2023. Unifier: A unified retriever for large-scale retrieval. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4787--4799

    2023

  9. [17]

    Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Andreas Koukounas, Nan Wang, and Han Xiao. 2024. https://arxiv.org/abs/2409.10173 jina-embeddings-v3: Multilingual embeddings with task...

    2024 arXiv

  10. [18]

    Yuanhe Tian, Yan Song, Fei Xia, Tong Zhang, and Yonggang Wang. 2020. https://doi.org/10.18653/v1/2020.acl-main.734 Improving C hinese word segmentation with wordhood memory networks . pages 8274--8285, Online. Association for Computational Linguistics

    2020 doi

  11. [19]

    Shuai Wang, Shengyao Zhuang, and G. Zuccon. 2021. Bert-based dense retrievers require interpolation with bm25 for effective passage retrieval. Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval

    2021

  12. [20]

    Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. https://arxiv.org/abs/2205.12035 Retromae: Pre-training retrieval-oriented language models via masked auto-encoder . Preprint, arXiv:2205.12035

    2022 arXiv

  13. [21]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. https://arxiv.org/abs/2309.07597 C-pack: Packaged resources to advance general chinese embedding . Preprint, arXiv:2309.07597

    2023 arXiv

  14. [22]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808

    2020 arXiv

  15. [23]

    Bennett, Junaid Ahmed, and Arnold Overwikj

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwikj. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations

    2021

  16. [24]

    Sha Yuan, Hanyu Zhao, Zhengxiao Du, Ming Ding, and Jie Tang. 2021. Wudaocorpora: A super large-scale chinese corpora for pre-training language models. AI Open

    2021

  17. [25]

    Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1503--1512

    2021

  18. [26]

    Sheng Zhang, Xin Zhang, Hui Wang, Lixiang Guo, and Shanshan Liu. 2018. Multi-scale attentive interaction networks for chinese medical question answer selection. IEEE Access, pages 1--1

    2018

  19. [27]

    Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuccon. 2024. Promptreps: Prompting large language models to generate dense and sparse representations for zero-shot document retrieval

    2024

  20. [28]

    Shengyao Zhuang and Guido Zuccon. 2021 a . Fast passage re-ranking with contextualized exact term matching and efficient passage expansion. arXiv preprint arXiv:2108.08513

    2021 arXiv

  21. [29]

    Shengyao Zhuang and Guido Zuccon. 2021 b . Tilde: Term independent likelihood model for passage re-ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1483--1492

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.