Pith. sign in

REVIEW 4 major objections 4 minor 3 cited by

Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Fine-tuned AI models on Hugging Face form measurable family trees: models in the same lineage share more traits, siblings resemble each other more than parents, and the ecosystem drifts toward permissive licenses, English-only coverage, and

desk verdict Not the paper we were asked to review—the full text is an unrelated report—and the abstract alone cannot support the claimed findings. read the letter →

arxiv 2508.06811 v1 pith:QJ5TON3Q submitted 2025-08-09 cs.SI cs.AIcs.CYcs.LG

Benjamin Laufer , Hamidah Oderinwale , Jon Kleinberg This is my paper · ORCID
classification cs.SIcs.AIcs.CYcs.LG
keywords HuggingFacemodelfine-tuningfamilytreesevolutionaryanalogyopen-sourceAIecosystemlicensedriftmetadatatraitinheritance
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 tries to establish that the sprawling practice of fine-tuning open machine-learning models can be studied as an evolutionary ecosystem. Analyzing 1.86 million models on Hugging Face, it builds family trees that link fine-tuned models to their base models and then measures how traits such as license, language coverage, and model-card text change along these lineages. The central claims are that models display family resemblance, that mutations are fast and directed so siblings look more like each other than like their parents, and that the ecosystem drifts directionally: licenses migrate toward permissive or copyleft terms, often against upstream terms, multilingual models shift toward English-only, and model cards grow shorter and more templated. If correct, this gives researchers and regulators an empirical, map-like view of how open AI actually evolves, not just how it is designed.

What carries the argument

The central object is the model family tree: a network connecting each fine-tuned model to its declared base or parent model, constructed from Hugging Face metadata. Traits are read from model metadata and model cards, treated as genetic markers, and similarity between models is measured as overlap in these markers. The sibling-versus-parent/child comparison is the mechanism that distinguishes this ecosystem's mutation pattern from classic asexual-reproduction models: fast, directed mutation makes sibling similarity the dominant signal.

What would settle it

Reconstruct family trees directly from model weights instead of declared parent metadata, and compare siblings to parent-child pairs while controlling for card-template usage. If declared siblings show no greater weight-based similarity than unrelated models with identical card templates, the family-resemblance and sibling-similarity claims would collapse. Conversely, checking a random sample of declared parent-child pairs against their upstream licenses would settle whether the license-violation drift is real or a metadata artifact.

Watch

Extended reading notes

Core claim

Using model metadata and model cards as genetic markers, the paper finds that fine-tuned models inherit and transmit traits in a way that resembles asexual reproduction, but with a distinctive twist: mutations are fast and directed. Sibling models, meaning two models fine-tuned from the same parent, tend to be more similar to each other than a child is to its own parent, a pattern that departs from standard evolutionary expectations. The same trait-tracking lens reveals directional ecosystem drift: licenses tend to move from restrictive commercial terms toward permissive or copyleft terms, frequently in violation of the upstream license; models tend to lose multilingual capability and conver

Load-bearing premise

The entire analysis relies on Hugging Face users truthfully declaring which base model each fine-tune descends from, and accurately filling in traits such as license, language, and model-card text; if uploaders misstate parentage or copy cards and license fields automatically, family membership, sibling structure, and trait-drift measurements all inherit that noise.

Editorial extensions

If this is right

  • Family trees provide a tractable unit of analysis for the open-source ML ecosystem, letting researchers ask who descends from whom and how traits propagate.
  • License drift toward permissive and copyleft terms, often violating upstream licenses, implies that a substantial share of fine-tuned derivatives may carry compliance risk.
  • The shift toward English-only models suggests that the open ecosystem, despite its diversity, may be concentrating on a single language rather than broadening coverage.
  • Shorter, templated, auto-generated model cards imply that the metadata used to discover and evaluate models is becoming less informative over time.
  • If sibling models are more similar than parent-child pairs, then lineage breadth, not just depth, determines how traits cluster and spread in the ecosystem.

Reading between the lines

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

  • A direct test of the family-resemblance claim would reconstruct lineages from model weights rather than declared parent fields; if sibling similarity disappears once card-templating artifacts are controlled, the trait signal may be partly an artifact of how cards are written.
  • The paper's evolutionary framing suggests a measurable extension: track whether traits such as permissive-license adoption or English-only behavior spread through copying of model-card text versus actual weight-level inheritance, which would separate social mimicry from technical descent.
  • The license-violation finding, if robust, has a concrete policy corollary: platform-level audits of fine-tuned derivatives could compare declared licenses against upstream terms, turning an ecological observation into a compliance tool.
  • The attached full text is a different technical report on Q-programming-language fine-tuning; the ecosystem findings rest on the abstract's description, since the full text supplied here does not contain the family-tree analysis itself.
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 submission as received consists of the abstract of arXiv:2508.06811, 'Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face,' followed by a full text that is not that paper. The full text is arXiv:2508.06813v2, 'Technical Report: Full-Stack Fine-Tuning for the Q Programming Language' by Hogan et al. The abstract claims an empirical study of 1.86 million Hugging Face models, construction of model family trees, measurement of 'genetic similarity' from metadata and model cards, and findings of family resemblance, sibling-over-parent similarity, license drift, language narrowing, and template-based card generation. None of the corresponding methods, data descriptions, analyses, figures, or validation appear in the supplied full text. The claimed study therefore cannot be verified, reproduced, or even formally assessed from the submitted material.

Significance. If the claimed study were fully documented, it would be a timely and potentially valuable contribution: a large-scale, quantitative map of fine-tuning lineages on Hugging Face, with concrete findings on inheritance, mutation, and directional drift in licenses, language coverage, and documentation practices. The evolutionary-biology framing and the use of family trees are appealing and could generate new hypotheses about open ML ecosystems. However, the submitted artifact contains no evidence for these findings. The paper's contribution is therefore entirely conditional on a manuscript that is not present. There is also a real methodological threat, identified in the abstract itself: because traits are extracted from model cards and metadata, and because model cards are increasingly templated and auto-generated, the central family-resemblance result could be an artifact of card-generation tooling rather than fine-tuning relatedness.

major comments (4)
  1. [Full Text (first page)] The supplied full text is not the manuscript described in the abstract. It is 'Technical Report: Full-Stack Fine-Tuning for the Q Programming Language' (arXiv:2508.06813v2). It contains no methods for constructing family trees, no definition of genetic markers or traits, no data description for the 1.86M models, no statistical tests, and no results on family resemblance, sibling similarity, license drift, language narrowing, or card templating. The central claims of the abstract are thus entirely unsupported in the submitted material. This is a load-bearing deficiency that cannot be repaired by local revision.
  2. [Abstract] The abstract asserts specific quantitative findings: family resemblance, greater sibling similarity than parent/child similarity, directional license drift 'often in violation of upstream license terms,' a shift from multilingual to English-only coverage, and a trend toward shorter, templated model cards. No measurement procedures, error bars, baselines, null models, or validation are provided in the available text. For a study whose novelty depends on these empirical patterns, the absence of any methodological detail makes the claims untestable as submitted.
  3. [Abstract (threat to validity)] The stress-test concern about templated model cards lands. The abstract states that traits are measured from 'model metadata and model cards' and also reports that model cards are 'reduced in length and standardize by turning, more often, to templates and automatically generated text.' If sibling models inherit a parent's card template or are created with card-generation tools, then trait overlap between siblings is expected independently of fine-tuning relatedness. Without controls such as removing boilerplate, stratifying on manual versus templated cards, or demonstrating that card similarity is not explained by template sharing, the central 'siblings more similar than parents' result could be an artifact. No such controls are visible in the submitted material.
  4. [Abstract (license analysis)] The claim that licenses drift 'often in violation of upstream license's terms' requires a precise classification rule: which license combinations count as violations, how upstream license terms are determined, and how self-reported metadata is validated against actual license files or model weights. The supplied text gives no such rule. Because the abstract itself relies on self-reported metadata, the license-violation counts inherit any noise or strategic misreporting in those fields. This issue is secondary to the missing manuscript, but it would need to be addressed in any resubmission.
minor comments (4)
  1. [Title vs. Abstract] The title states '2 Million Models' while the abstract analyzes '1.86 million models.' Please reconcile the numbers.
  2. [Abstract (definitions)] The terms 'genetic markers,' 'traits,' 'family resemblance,' and 'mutation' are not operationally defined. Even in an abstract, a sentence indicating what fields constitute markers and what distance/similarity measure is used would help.
  3. [Abstract (data snapshot)] No data collection date, snapshot version, or deduplication/filtering criteria are mentioned. Hugging Face metadata changes over time, so a temporal snapshot is needed for reproducibility.
  4. [Abstract (statistical claims)] Statements such as 'mutations are fast and directed' and 'sibling models tend to exhibit more similarity than parent/child pairs' require a null model or baseline (e.g., comparing to random pairs or reshuffled trees) and confidence intervals. The abstract reports none.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity is demonstrable from the available evidence; the supplied full text is a different paper, so the claimed derivation chain cannot be inspected, and the abstract alone does not exhibit a reduction.

full rationale

The claimed paper (arXiv:2508.06811) is represented only by its abstract; the supplied full text is arXiv:2508.06813v2, a technical report on Q-language fine-tuning by Hogan et al., which does not contain the family-tree methods, equations, or analyses described in the abstract. I therefore cannot quote a specific reduction of any prediction to its inputs. Based on the abstract alone, no circular step is exhibited: the family trees are described as networks built from metadata linking fine-tuned models to base/parent models, while traits (licenses, languages, model-card text) are measured from metadata and cards. Family resemblance is an empirical joint-distribution claim that overlap in traits is higher within families, not a claim that family membership is defined by trait similarity. The sibling-versus-parent finding is also stated as an empirical regularity rather than a construction. The abstract's own observation that model cards are increasingly templated and auto-generated is a real validity risk for trait measurement—card similarity could track template provenance rather than fine-tuning relatedness—but that is a confound or measurement-error concern, not circularity, because the paper does not fit a parameter to the outcome and then rename it as a prediction. No load-bearing self-citation is visible in the abstract, and the unrelated full text provides no derivation chain to audit. Under the hard rule that circularity requires quoting the paper and exhibiting the specific reduction, the honest finding is no significant circularity: score 0.

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

The review package contains only the abstract for the claimed paper; the full text is a different manuscript. The ledger therefore records the load-bearing analytical choices visible in the abstract: trait definitions, the license-violation rule, the assumption that metadata parentage is accurate, and the assumption that model-card traits are heritable signals. Self-reported data plus hand-chosen trait definitions mean the reader is paying for the definitions and metadata assumptions upstream. No fitted numeric parameters are visible at abstract level. No new physical or conceptual entities are introduced; "genetic markers" and "family trees" are metaphors for metadata relations.

free parameters (2)
  • Trait/genetic-marker definitions on model cards
    The abstract measures family resemblance via "genetic markers and traits" drawn from model metadata and model cards, but does not specify which fields count as traits or how overlap is computed. These choices determine every similarity finding and function as hand-selected analytical parameters.
  • License-violation classification rule
    The claim that license drift is "often in violation of upstream license's terms" requires a rule mapping declared licenses to upstream obligations. That rule is a modeling choice not described in the abstract.
assumptions (3)
  • domain assumption Hugging Face metadata records true fine-tuning parentage (base_model fields are accurate)
    Family trees connect fine-tuned models to their base or parent using self-reported metadata and model cards. If uploaders misstate or omit bases, the trees encode platform convention, not training history.
  • domain assumption Model-card traits are heritable and meaningful similarity signals
    The "genetic similarity" claim treats model card content and metadata as heritable traits of a model lineage. The abstract itself notes model cards are increasingly templated and auto-generated, which means trait signals may reflect card-generation tooling rather than model behavior.
  • domain assumption An evolutionary analogy (asexual reproduction, mutation) maps onto fine-tuning
    The interpretation of similarities and differences as "fast, directed mutations" assumes fine-tuning is an asexual inheritance process with heritable traits. This is a framing assumption, not established by the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face." pith.science (2026). https://pith.science/paper/QJ5TON3Q

@misc{pith2026250806811,
  author       = {Pith},
  title        = {Pith review of: Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJ5TON3Q}},
  note         = {Machine review of arXiv:2508.06811}
}
read the original abstract

Many have observed that the development and deployment of generative machine learning (ML) and artificial intelligence (AI) models follow a distinctive pattern in which pre-trained models are adapted and fine-tuned for specific downstream tasks. However, there is limited empirical work that examines the structure of these interactions. This paper analyzes 1.86 million models on Hugging Face, a leading peer production platform for model development. Our study of model family trees -- networks that connect fine-tuned models to their base or parent -- reveals sprawling fine-tuning lineages that vary widely in size and structure. Using an evolutionary biology lens to study ML models, we use model metadata and model cards to measure the genetic similarity and mutation of traits over model families. We find that models tend to exhibit a family resemblance, meaning their genetic markers and traits exhibit more overlap when they belong to the same model family. However, these similarities depart in certain ways from standard models of asexual reproduction, because mutations are fast and directed, such that two `sibling' models tend to exhibit more similarity than parent/child pairs. Further analysis of the directional drifts of these mutations reveals qualitative insights about the open machine learning ecosystem: Licenses counter-intuitively drift from restrictive, commercial licenses towards permissive or copyleft licenses, often in violation of upstream license's terms; models evolve from multi-lingual compatibility towards english-only compatibility; and model cards reduce in length and standardize by turning, more often, to templates and automatically generated text. Overall, this work takes a step toward an empirically grounded understanding of model fine-tuning and suggests that ecological models and methods can yield novel scientific insights.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When Model Release Meets Model Reuse: Producer-Consumer Misalignment in Hugging Face

    cs.SE 2026-07 conditional novelty 7.0 of 10

    Producers and consumers of pre-trained models rely on the same documentation but systematically disagree on where metadata belongs, why lineage is traced, and which governance mechanisms help.

  2. Controlling tail risk in two-slope ski rental

    cs.DS 2025-08 unverdicted novelty 6.0 of 10

    Characterizes optimal tail-risk strategies for two-slope ski rental and gives a near-optimal greedy algorithm and an exact linear-programming algorithm.

  3. When Models Meet Users: An Empirical Study of Perceptions of General LLMs and Multimodal LLMs on Hugging Face

    cs.SE 2026-04 unverdicted novelty 5.0 of 10

    Hugging Face discussions show that access barriers, output quality, and setup complexity are the main user concerns for both general and multimodal LLMs.

Reference graph

Works this paper leans on

25 extracted references · 16 canonical work pages · cited by 3 Pith papers

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025

    2025

  2. [2]

    sanity check

    from the start, and agreeing on consistent logging and formatting conventions, proved essential for making sense of results, facilitating collaboration, and ensuring reproducibility. A clean, searchable experiment log is as important as the code or models themselves. 8.�������� �� ��� ���� ����� �� ���� ���� ��� ���� ��� �������With well- designed dataset...

    work page

  3. [3]

    Experiment tracking with weights and biases, 2020

    Lukas Biewald. Experiment tracking with weights and biases, 2020. Software available from wandb.com

    2020

  4. [4]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    work page 2021

  5. [5]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    work page 2025

  6. [6]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, As- ton Zhang, Aurelien Rodriguez, Austen Gregerson, Ava...

    work page 2024

  7. [7]

    Accelerate: Training and inference at scale made simple, efficient and adaptable.������������������� ����������������������, 2022

    Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, and Benjamin Bossan. Accelerate: Training and inference at scale made simple, efficient and adaptable.������������������� ����������������������, 2022

    work page 2022

  8. [8]

    Suchin Gururangan, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. Don’t stop pretraining: Adapt language models to domains and tasks, 2020

    work page 2020

Show all 25 references
  1. [9]

    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. Lora: Low-rank adaptation of large language models, 2021

    2021

  2. [10]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

    2017

  3. [11]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In����������� �� ��� ��� ������ ���� ��������� �� ��������� ������� ...

    2023

  4. [12]

    Documentation for kdb�and q.����������������������, 2025

    Kx Systems, Inc. Documentation for kdb�and q.����������������������, 2025. Accessed 27 July 2025

    2025

  5. [13]

    kdb�: Time-series database and the q programming language

    Kx Systems, Inc. kdb�: Time-series database and the q programming language. ����������������������������, 2025. Accessed 27 July 2025

    2025

  6. [14]

    Understanding r1-zero-like training: A critical perspective, 2025

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective, 2025

    2025

  7. [15]

    OpenAI, :, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El- Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexander Prokofiev, Alexander Wei, Allison Tam, Ally Benne...

    2024

  8. [16]

    Qwen2.5 technical report, 2025

    Qwen, :, AnYang, BaosongYang, BeichenZhang, BinyuanHui, BoZheng, BowenYu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu,...

    2025

  9. [17]

    Zero: Memory optimizations toward training trillion parameter models, 2020

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models, 2020

    2020

  10. [18]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024

    2024

  11. [19]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, ...

    2025

  12. [20]

    Qwq-32b: Embracing the power of reinforcement learning, March 2025

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, March 2025

    2025

  13. [21]

    Trl: Transformer reinforcement learning.����������������������������������, 2020

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. Trl: Transformer reinforcement learning.����������������������������������, 2020

    2020

  14. [22]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  15. [23]

    Leetcodedataset: A temporal dataset for robust evaluation and efficient training of code llms, 2025

    Yunhui Xia, Wei Shen, Yan Wang, Jason Klein Liu, Huifeng Sun, Siyue Wu, Jian Hu, and Xiaolong Xu. Leetcodedataset: A temporal dataset for robust evaluation and efficient training of code llms, 2025

    2025

  16. [24]

    Qwen3 technical report, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    2025

  17. [25]

    reasoning

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen...

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.