Pith. sign in

REVIEW 5 major objections 5 minor 18 references

EnTao-GPM: DNA Foundation Model for Predicting the Germline Pathogenic Mutations

T0 review · 5 major / 5 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 A DNA foundation model pre-trained on 27 mammal genomes reaches AUC 0.963 on germline variant classification after fine-tuning on ClinVar and HGMD, and its zero-shot backbone matches a 40-times-larger model.

desk verdict A standard fine-tuned DNA foundation model with plausible but poorly supported performance claims: the zero-shot comparison has an internal inconsistency, fine-tuned models are compared only to zero-shot baselines, and no external validation is provided. read the letter →

arxiv 2507.21706 v1 pith:J7DRUFQZ submitted 2025-07-29 q-bio.GN cs.AI

Zekai Lin , Haoran Sun , Yucheng Guo , Yujie Yang , Yanwen Wang , Bozhen Hu , Chonghang Ye , Qirong Yang
show 3 more authors
Fan Zhong Xiaoming Zhang Lei Liu
This is my paper
classification q-bio.GNcs.AI
keywords DNAfoundationmodelgermlinepathogenicitypredictionClinVarHGMDzero-shotvarianteffectcross-speciespre-trainingindelclassificationwhole-genomemutationscan
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 a recipe for germline pathogenicity prediction: take a DNA language model pre-trained on disease-relevant mammals (27 species including human, pig, and mouse), fine-tune it on curated clinical labels from ClinVar and HGMD, and it can classify point mutations and small insertions or deletions with high accuracy. The authors report that the resulting model reaches an AUC of 0.963 for single-nucleotide variants and 0.933 for a mixed set that also includes indels, and that the pre-trained backbone, before any clinical fine-tuning, matches a model 40 times larger in zero-shot performance. If the claims hold, such a model would give clinical labs an efficient whole-genome scanning tool and better coverage of the noncoding and non-SNV variants that rule-based and older data-driven tools struggle with.

What carries the argument

The machinery is TrinityDNA's bio-inspired transformer backbone, which combines three components: a Groove Fusion Module (parallel convolution filters of kernel sizes 3, 5, and 7 with GELU activations), Sliding Multi-Window Attention (per-head variable window sizes to capture both short regulatory motifs and long-range interactions), and a Gated Reverse Complement mechanism (a shared encoder processes both strand orientations and fuses them with a learned gate). This backbone is post-trained on 27 mammalian genomes with masked language modeling at a 100,000-base context window, yielding TrinityDNA-LabFauna. Fine-tuning then uses LoRA low-rank updates to attach a feedforward head that maps sequence representations to a pathogenicity probability, keeping the trainable parameter count small while adapting to ClinVar and HGMD labels.

What would settle it

Take EnTao-GPMFast and EnTao-GPMPro and score variants that were added to ClinVar after the April 2025 release used here, or variants from an independent clinical sequencing cohort; if the AUPRC falls sharply relative to the reported 0.974 and 0.966, the database-split evaluation was optimistic and the generalizability claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that cross-species pre-training on 27 mammalian genomes captures sequence features of pathogenic variation that transfer to human disease: the pre-trained backbone TrinityDNA-LabFauna reaches zero-shot AUC 0.943 and AUPRC 0.961 on the ClinVar SNV test, beating the same-size EVO2-1B and matching EVO2-40B. After LoRA fine-tuning on ClinVar and HGMD, EnTao-GPMFast (which reads only the reference sequence) reaches AUC 0.963 and AUPRC 0.974 on the SNV test, while EnTao-GPMPro (which reads both reference and variant sequence) reaches AUC 0.933 and AUPRC 0.966 on a mixed SNV/non-SNV test that also samples from HGMD. The paper also reports that the fine-tuned Fast model supports whole-genome mutation scanning, with precomputed GRCh38 scan results made publicly available.

Load-bearing premise

The load-bearing premise is that a highest-confidence subset of ClinVar (5,406 variants) held out from the same database used for training is a valid test of real-world clinical generalizability; if the model has learned database-specific artifacts rather than general biological features, the reported AUCs will not transfer to newly discovered variants.

Editorial extensions

If this is right

  • A single fine-tuned model can score both SNVs and indels, whereas many previous predictors handle only point mutations.
  • The reference-only Fast design enables whole-genome pathogenicity scans; the authors have published a GRCh38 scan for rapid look-up.
  • Pre-training on lab-mammal genomes achieves zero-shot performance on par with a 40-billion-parameter model, suggesting efficient cross-species transfer.
  • Because fine-tuning uses LoRA, the specialization step is cheap to rerun as ClinVar and HGMD are updated.

Reading between the lines

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

  • If the held-out test set is not separated chronologically from the training data, the reported AUCs may partly reflect database conventions rather than true generalization; a temporal or external-cohort validation would be the decisive check.
  • The paper's stated LLM-based statistical explanations are not described or evaluated in the methods or experiments, so the interpretability claim should be treated as a design goal rather than a validated result.
  • An ablation that varies the number of pre-training species, such as human-only versus all 27 mammals, would show directly how much of the zero-shot gain comes from cross-species conservation rather than from model size or data volume.
  • The zero-shot perplexity-difference method reaches only AUC 0.704 on the mixed SNV/non-SNV test, while fine-tuned EnTao-GPMPro reaches 0.933; this gap implies that most non-SNV ability is learned from the clinical labels rather than inherited from mammal pre-training.
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

5 major / 5 minor

Summary. The manuscript presents EnTao-GPM, a DNA foundation model for germline pathogenicity prediction. It post-trains the TrinityDNA architecture on 27 mammalian genomes to obtain TrinityDNA-LabFauna, then applies LoRA fine-tuning on ClinVar and HGMD to produce EnTao-GPMFast for SNVs and EnTao-GPMPro for SNVs and non-SNVs. The authors report zero-shot AUC 0.943 and AUPRC 0.961 for TrinityDNA-LabFauna, fine-tuned AUC 0.963 for EnTao-GPMFast on an SNV test set, and AUC 0.933 for EnTao-GPMPro on a mixed SNV/non-SNV test set, and they claim superior accuracy in mutation classification plus an interpretable clinical framework combining sequence embeddings with LLM-based explanations.

Significance. If substantiated, a mammalian-pretrained DNA foundation model that handles both SNVs and non-SNVs with a fast whole-genome screening mode would be practically valuable for clinical variant interpretation. The paper's strengths include the targeted multi-species pretraining idea, the use of parameter-efficient LoRA fine-tuning, and the two-model Fast/Pro design for different clinical workflows. However, the reported results do not currently establish the central claims: the zero-shot comparison table is internally inconsistent, the fine-tuned models are benchmarked only against zero-shot baselines, and the evaluation uses held-out splits of the same ClinVar and HGMD releases used for training without documented variant-level deduplication or external validation. The significance of the work therefore cannot be assessed until these evaluation issues are addressed.

major comments (5)
  1. [§3.2.1, Table 4] The zero-shot comparison is internally inconsistent: the text states that TrinityDNA-LabFauna achieves AUC 0.943 and AUPRC 0.961 and outperforms EVO2-1B with AUC 0.950 and AUPRC 0.955, but Table 4 shows AUC 0.950 and AUPRC 0.955 for EVO2-40b and AUC 0.927 and AUPRC 0.946 for EVO2-1b-base, while the column with 0.943 and 0.961 is labeled FudanDNA-1B rather than TrinityDNA-LabFauna. The claimed zero-shot superiority over EVO2-1B and parity with EVO2-40B therefore needs a corrected table and an explicit statement of which model is which.
  2. [§2.2.4, Tables 3 and 5] The evaluation protocol does not support the clinical generalizability claim. Test variants come from the same April 2025 ClinVar release and HGMD 2025.2 used for fine-tuning, and the paper does not state that training and test sets were deduplicated at the chromosome-position-reference-alternate level. The ClinVar SNV test set is also the 5,406 'highest-confidence' variants, which is not a random sample and may be biased toward well-characterized loci. Without a temporal split, external validation set, or explicit deduplication, the AUC values 0.963 and 0.933 in Table 5 could reflect database-specific memorization rather than generalizable biological signal, so the central claim of superior mutation classification is not established.
  3. [Table 5, §3.2.2] The fine-tuned models are compared only against zero-shot TrinityDNA-LabFauna; no supervised state-of-the-art predictor (e.g., REVEL, ClinPred, or a fine-tuned DNABERT-2) is run under the same training and test protocol. As a result, the abstract's 'superior accuracy' claim is unsupported, because the experiments do not isolate the benefit of the proposed fine-tuning procedure or architecture. The authors should add supervised baselines and report confidence intervals or statistical tests for the AUC differences.
  4. [§1.5, §2.1.3, §5] The third claimed innovation, an interpretable clinical framework integrating DNA sequence embeddings with LLM-based statistical explanations, is not implemented or evaluated anywhere in the paper; Section 5 lists the fusion of sequence models with LLMs as future work. The abstract and introduction therefore overstate the current contribution, and the interpretability claim must either be removed from the claims or supported by a concrete method and results.
  5. [§2.2.3, Table 3] The HGMD integration is under-specified at a point that affects split integrity. The text reports 97.2% concordance between the DM category and ClinVar pathogenic labels and says overlapping variants were removed, but it does not describe the matching procedure, nor whether the 2,000 HGMD test variants were selected before or after overlap removal and merging with ClinVar. If any HGMD test variant overlaps a ClinVar training variant, the Dataset 2 test results reported in Table 5 would be inflated; the split construction must be stated precisely.
minor comments (5)
  1. [§2.2.1] The word 'Ensemble' should be 'Ensembl', and the species table has several missing GCA/GCF identifiers (e.g., Nothocricetulus migratorius, Abo bat, Sind bat) that should be completed for reproducibility.
  2. [§2.2.3, Table 4] Section 2.2.3 contains unresolved cross-references 'Table??', Figures 1 and 2 are not referenced in the text, and Table 4's column header 'FudanDNA-1B' is inconsistent with the name 'TrinityDNA-LabFauna' used in the text.
  3. [§3.1.1] Section 3.1.1 uses unresolved citation markers '[?]' for TrinityDNA and for the Megatron/DeepSpeed training framework; the corresponding references need to be completed.
  4. [§6, §7] Sections 6 and 7 promise a Python library, a web API, and publicly available whole-genome scan results but give no URL, license, or version information, which prevents independent verification.
  5. [§1.4, §3.2.2] The text uses inconsistent capitalization of 'Mutation' and 'Mutations' and introduces the acronym 'DMS' (Disease Mutation Scanning) without definition; these should be corrected.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity via self-citation: the base architecture and training paradigm are imported from the authors' own unpublished TrinityDNA preprint, while the headline AUCs themselves are measured rather than derived.

  1. self citation load bearing [Section 2.1.1 and Section 3.1.1; also Introduction, second paragraph (ref [7])]
    "TrinityDNA serves as the core foundational model, meticulously designed for DNA sequence modeling with bioinspired components that lay the groundwork for subsequent specialized training phases. Inspired by the TrinityDNA framework, we first learn generic sequence representations, then specialise them through post-training on human and the other 26 laboratory mammalian genomes to obtain TrinityDNA-LabFauna."

    The core architectural premise (GFM/SMWA/GRC components and the evolutionary pre-training strategy) is imported from reference [7], a preprint by the same group (Qirong Yang, Yucheng Guo, Yujie Yang, and Xiaoming Zhang appear in both author lists). The introduction's assertions that TrinityDNA 'outperform[s] EVO and Caduceus' and sets 'new benchmarks' are also citations to [7] rather than results established or re-derived here. Thus the claimed foundation of EnTao-GPM rests on a self-citation that is itself unverified. The impact is mitigated because Table 4 provides a direct zero-shot ClinVar benchmark of TrinityDNA-LabFauna against EVO2, and the final EnTao-GPM AUCs are empirical measurements rather than equations forced by the citation.

full rationale

The paper's new contributions (EnTao-GPMFast and EnTao-GPMPro) are empirical fine-tuned classifiers, and the reported AUCs (0.963 and 0.933) are measured on held-out ClinVar/HGMD test splits. No equation in the paper defines the prediction score in terms of the test labels, so there is no construction-level circularity in the headline numbers. The main circularity-adjacent issue is load-bearing self-citation: TrinityDNA-LabFauna is obtained by post-training the authors' own TrinityDNA preprint (ref [7]), and the introduction's claims that TrinityDNA is state-of-the-art and outperforms EVO/Caduceus are supported only by that same preprint. The paper does provide a direct zero-shot ClinVar benchmark against EVO2 in Table 4, which gives some independent evidence for the derived model, but the choice and validity of the base architecture itself are not independently established here. The evaluation also uses a test split carved from the same ClinVar release used for fine-tuning, which is not circular by itself but limits the strength of the generalizability claim. Overall, the central claim is not forced by definition; the self-citation is minor and partially mitigated by external benchmarks, so a score of 2 is appropriate.

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

The central claims rest on the reliability of ClinVar and HGMD labels as ground truth, on the transfer of MLM pre-training from 27 mammals to human pathogenicity, and on the representativeness of the internal test splits. No free parameters are disclosed, since the authors do not report hyperparameters for fine-tuning. The models EnTao-GPM and TrinityDNA-LabFauna are introduced without released weights or APIs, so they are entities with no independent evidence.

assumptions (3)
  • domain assumption ClinVar and HGMD labels accurately reflect true pathogenicity
    Used as training and test ground truth in Sections 2.2.2 and 2.2.3; label noise would bias both training and evaluation.
  • domain assumption Pre-training on 27 mammalian genomes improves human pathogenicity prediction
    Motivates TrinityDNA-LabFauna in Section 2.1.2; no direct evidence that these 27 species are optimal beyond a general conservation argument.
  • domain assumption The held-out ClinVar split (highest confidence variants) is representative of future clinical variants
    The test set is drawn from the same database and time period as training (Section 2.2.4); no external or prospective validation is provided.
invented entities (2)
  • EnTao-GPM (Fast and Pro)
    purpose: Fine-tuned DNA model for germline pathogenic mutation prediction
    Model weights and inference code are not released; the paper provides only the architecture description and AUC numbers, so the entity cannot be independently tested.
  • TrinityDNA-LabFauna
    purpose: Mammal-adapted DNA foundation model
    Built from the authors' own TrinityDNA; no checkpoint or API is provided, so its zero-shot performance cannot be independently reproduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EnTao-GPM: DNA Foundation Model for Predicting the Germline Pathogenic Mutations." pith.science (2026). https://pith.science/paper/J7DRUFQZ

@misc{pith2026250721706,
  author       = {Pith},
  title        = {Pith review of: EnTao-GPM: DNA Foundation Model for Predicting the Germline Pathogenic Mutations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7DRUFQZ}},
  note         = {Machine review of arXiv:2507.21706}
}
read the original abstract

Distinguishing pathogenic mutations from benign polymorphisms remains a critical challenge in precision medicine. EnTao-GPM, developed by Fudan University and BioMap, addresses this through three innovations: (1) Cross-species targeted pre-training on disease-relevant mammalian genomes (human, pig, mouse), leveraging evolutionary conservation to enhance interpretation of pathogenic motifs, particularly in non-coding regions; (2) Germline mutation specialization via fine-tuning on ClinVar and HGMD, improving accuracy for both SNVs and non-SNVs; (3) Interpretable clinical framework integrating DNA sequence embeddings with LLM-based statistical explanations to provide actionable insights. Validated against ClinVar, EnTao-GPM demonstrates superior accuracy in mutation classification. It revolutionizes genetic testing by enabling faster, more accurate, and accessible interpretation for clinical diagnostics (e.g., variant assessment, risk identification, personalized treatment) and research, advancing personalized medicine.

Figures

Figures reproduced from arXiv: 2507.21706 by the authors.

Figure 1
Figure 1. Overview of the curated ClinVar dataset used for fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Overview of the HGMD dataset used for fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. EnTao-GPM model architecture for mutation prediction. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 9 canonical work pages

  1. [1]

    The expanding scope of dna sequencing.Nature biotech- nology, 30(11):1084–1094, 2012

    Jay Shendure and Erez Lieberman Aiden. The expanding scope of dna sequencing.Nature biotech- nology, 30(11):1084–1094, 2012

    work page 2012

  2. [2]

    A method and server for predicting damaging missense mutations.Nature methods, 7(4):248–249, 2010

    Ivan A Adzhubei, Steffen Schmidt, Leonid Peshkin, Vasily E Ramensky, Anna Gerasimova, Peer Bork, Alexey S Kondrashov, and Shamil R Sunyaev. A method and server for predicting damaging missense mutations.Nature methods, 7(4):248–249, 2010

    work page 2010

  3. [3]

    Clinvar: updates to support classifications of both germline and somatic variants

    MJ Landrum, S Chitipiralla, K Kaur, G Brown, C Chen, J Hart, D Hoffman, W Jang, C Liu, Z Maddipatla, R Maiti, J Mitchell, T Rezaie, G Riley, G Song, J Yang, L Ziyabari, A Russette, and BL Kattman. Clinvar: updates to support classifications of both germline and somatic variants. Nucleic Acids Research, 53(D1):D1313–D1321, 2025

    work page 2025

  4. [4]

    Ioannidis, Joseph H

    Nilah M. Ioannidis, Joseph H. Rothstein, Vikas Pejaver, Sumit Middha, Shannon K. McDonnell, Saurabh Baheti, Anthony Musolf, Qing Li, Emily Holzinger, Danielle Karyadi, Lisa A. Cannon- Albright, Craig C. Teerlink, Janet L. Stanford, William B. Isaacs, Jianfeng Xu, Kathleen A. Cooney, Ethan M. Lange, Johanna Schleutker, John D. Carpten, Isaac J. Powell, Oli...

    work page 2016

  5. [5]

    Genome modeling and design across all domains of life with evo 2.BioRxiv, pages 2025–02, 2025

    Garyk Brixi, Matthew G Durrant, Jerome Ku, Michael Poli, Greg Brockman, Daniel Chang, Gabriel A Gonzalez, Samuel H King, David B Li, Aditi T Merchant, et al. Genome modeling and design across all domains of life with evo 2.BioRxiv, pages 2025–02, 2025

    2025

  6. [6]

    Dnabert- 2: Efficient foundation model and benchmark for multi-species genome.arXiv preprint arXiv:2306.15006, 2023

    Zhihan Zhou, Yanrong Ji, Weijian Li, Pratik Dutta, Ramana Davuluri, and Han Liu. Dnabert- 2: Efficient foundation model and benchmark for multi-species genome.arXiv preprint arXiv:2306.15006, 2023

    arXiv 2023

  7. [7]

    Qirong Yang, Yucheng Guo, Zicheng Liu, Yujie Yang, Qijin Yin, Siyuan Li, Shaomin Ji, Linlin Chao, Xiaoming Zhang, and Stan Z. Li. Trinitydna: A bio-inspired foundational model for efficient long-sequence dna modeling, 2025

    work page 2025

  8. [8]

    The human gene mutation database (hgmd®): optimizing its use in a clinical diagnostic or research setting.Human genetics, 139(10):1197–1207, 2020

    Peter D Stenson, Matthew Mort, Edward V Ball, Molly Chapman, Katy Evans, Luisa Azevedo, Matthew Hayden, Sally Heywood, David S Millar, Andrew D Phillips, et al. The human gene mutation database (hgmd®): optimizing its use in a clinical diagnostic or research setting.Human genetics, 139(10):1197–1207, 2020

    work page 2020

Show all 18 references
  1. [9]

    Dnabert: pre-trained bidirectional encoder representations from transformers model for dna-language in genome.Bioinformatics, 37(15):2112–2120, 2021

    Yanrong Ji, Zhihan Zhou, Han Liu, and Ramana V Davuluri. Dnabert: pre-trained bidirectional encoder representations from transformers model for dna-language in genome.Bioinformatics, 37(15):2112–2120, 2021

    2021

  2. [10]

    Sequence modeling and design from molecular to genome scale with evo.Science, 386(6723):eado9336, 2024

    Eric Nguyen, Michael Poli, Matthew G Durrant, Brian Kang, Dhruva Katrekar, David B Li, Liam J Bartie, Armin W Thomas, Samuel H King, Garyk Brixi, et al. Sequence modeling and design from molecular to genome scale with evo.Science, 386(6723):eado9336, 2024

    2024

  3. [11]

    Holmes, Jieping Ye, Jun Li, Yuelong Shu, Mang Shi, and Zhaorong Li

    Yong He, Pan Fang, Yongtao Shan, Yuanfei Pan, Yanhong Wei, Yichang Chen, Yihao Chen, Yi Liu, Zhenyu Zeng, Zhan Zhou, Feng Zhu, Edward C. Holmes, Jieping Ye, Jun Li, Yuelong Shu, Mang Shi, and Zhaorong Li. Lucaone: Generalized biological foundation model with unified nucleic ac...

    2024

  4. [12]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  5. [13]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining.Bioinformatics, 36(4):1234–1240, 2020

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining.Bioinformatics, 36(4):1234–1240, 2020

    2020

  6. [14]

    Toward expert-level medical question answering with large language models.Nature Medicine, 31(3):943–950, 2025

    Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, et al. Toward expert-level medical question answering with large language models.Nature Medicine, 31(3):943–950, 2025

    2025

  7. [15]

    Sue Richards, Nazneen Aziz, Sherri Bale, David Bick, Soma Das, Julie Gastier-Foster, Wayne W Grody, Madhuri Hegde, Elaine Lyon, Elaine Spector, et al. Standards and guidelines for the in- terpretation of sequence variants: a joint consensus recommendation of the american colle...

    2015

  8. [16]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  9. [17]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  10. [18]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langua...

    2019

Pith tools

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