Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

HuggingGraph: Understanding the Supply Chain of LLM Ecosystem

T0 review · 4 major / 6 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 This paper constructs a directed heterogeneous graph of 402,654 Hugging Face models and datasets to capture the LLM supply chain, then uses forward and backward analysis to trace provenance and structural vulnerability.

desk verdict Useful graph, but the model/dataset node typing is wrong in a load-bearing place, and the 'accurate capture' claim is unsubstantiated. read the letter →

arxiv 2507.14240 v3 pith:G3TMBTVK submitted 2025-07-17 cs.CL cs.AI

Mohammad Shahedur Rahman , Peng Gao , Yuede Ji This is my paper
classification cs.CLcs.AI
keywords LLMsupplychainHuggingFaceheterogeneousgraphmodelprovenancedatasetlineagedependencyanalysisfine-tuninganalytics
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 sets out to make the LLM supply chain visible: it collects metadata from Hugging Face for models and datasets and builds a directed heterogeneous graph with 402,654 nodes and 462,524 edges that record which models were fine-tuned, adapted, quantized, or merged from other models, and which datasets were used to train them. The paper claims that this graph accurately captures supply-chain provenance and supports forward and backward queries, so a practitioner can trace a model's full lineage back to its base model and training data. With that graph the authors report that the ecosystem is sparse and heavy-tailed, dominated by a single component of 61.4% of nodes, that base models like Llama-3.1-8B have spawned thousands of derivatives, and that the platform changes by thousands of models per day. If correct, the construction gives researchers and auditors a concrete tool for provenance checks, vulnerability propagation, and compliance, and it transfers to other platforms such as Kaggle.

What carries the argument

The load-bearing object is the directed heterogeneous supply-chain graph $G=(V,E)$, whose nodes are models (base, fine-tune, adapter, quantization, merge) and datasets, and whose edges are dependency relations among them. It is built from Hugging Face metadata fields ($\text{finetune}$, $\text{adapter}$, $\text{quantization}$, $\text{merge}$, and the dataset-side $\text{trained\_fine\_tune\_models}$ field), supplemented by cross-reference link scraping and NER-based textual pattern extraction when structured fields are absent. Analysis runs BFS forward from base models to count impacted derivatives and backward from task-specific models to reconstruct lineage, Louvain community detection for structure, and daily delta updates so the graph can be maintained as $G_{t+1}=G_t \cup \Delta_{t+1}$.

What would settle it

Sample, say, 100 edges from the graph at random and open the cited model or dataset card: if most supposed dependencies are not present in the card's structured fields or text, then the graph overstates supply-chain coverage. Additionally, check the Hugging Face repository type of 'Mistral-v0.1': if it is a model checkpoint rather than a dataset, the leading entry in the dataset ranking is a misclassification and the dataset-side results need rebuilding.

Watch

Extended reading notes

Core claim

The central claim is that the LLM supply chain—who built what from which base model and which datasets—can be captured at platform scale from Hugging Face's self-reported metadata. The resulting graph has six node types (base, fine-tune, adapter, quantization, and merge models, plus datasets) and edges for model-model, dataset-dataset, and model-dataset dependencies; missing metadata is patched by cross-referencing filtered listing pages and running NER over model and dataset cards. On this graph the paper reports structural findings: a sparse, heavy-tailed topology with a dominant weak component; Louvain communities with modularity 0.96 that align with tasks such as math and instruction tuning; base models with forward subgraph sizes in the thousands; datasets used by hundreds of models; and a dynamic view in which roughly 4,622 models and 1,538 datasets change daily. The same pipeline applied to Kaggle yields a smaller graph with similar structural properties, which the paper takes as evidence that the approach generalizes.

Load-bearing premise

The whole graph inherits the accuracy of Hugging Face's self-reported metadata: the fields 'finetune', 'adapter', 'quantization', 'merge', and 'trained_fine_tune_models', plus the text patterns extracted by NER, must faithfully and completely record every real dependency, and every node labeled a dataset must actually be one.

Editorial extensions

If this is right

  • A security patch or license change in one base model can be traced to every derivative: the graph shows Llama-3.1-8B feeding 7,544 downstream models across fine-tuning, adapter, quantization, and merge paths.
  • Dataset provenance becomes checkable at scale: top entries such as Mistral-v0.1 are recorded as training over a thousand models, and models like DeBERTa-ST-AllLayers-v3.1 draw on 116 datasets, so inherited bias or licensing risk can be attributed to specific upstream data.
  • The supply chain is not static: with roughly 4,622 models and 1,538 datasets changing daily, any snapshot is quickly stale, and delta-based updates are necessary for accurate tracking.
  • Task specialization has a signature: communities with modularity 0.96 split into math, instruction tuning, and other semantic groups, indicating that reuse clusters around shared tasks rather than random attachment.
  • The methodology is portable: applying the same collection and graph construction to Kaggle reproduces the heavy-tailed, fragmented structure despite far sparser metadata.

Reading between the lines

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

  • If Hugging Face metadata is noisier than assumed, the descendant counts and lineage depths reported here are better read as lower bounds on connectivity; a confidence score per edge would let downstream users distinguish well-documented provenance from inferred provenance.
  • The graph is a natural testbed for blast-radius simulation: seed a node with a hypothetical vulnerability and measure how many nodes become reachable through forward BFS—an experiment the paper describes but does not run.
  • The Table 5 entry 'Mistral-v0.1' is a model checkpoint by naming convention, not a dataset; if that is right, the top dataset ranking would change and the dataset-model edge construction should be re-examined for similar misclassifications.
  • Because dynamic deltas are captured daily, the pipeline could be used to detect coordinated upload events (fine-tuning waves, mass quantization releases) in near real time, turning the graph into an early-warning signal for ecosystem-level shifts.
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 / 6 minor

Summary. The paper introduces HuggingGraph, a directed heterogeneous graph of the Hugging Face LLM ecosystem containing 402,654 nodes and 462,524 edges as of June 30, 2025. Nodes represent models (base, fine-tuned, adapter, quantization, merge) and datasets; edges represent dependency relationships among models, among datasets, and between models and datasets. The graph is constructed from Hugging Face APIs, cross-referenced model listing webpages, and NER-based extraction from model/dataset cards. The authors then perform forward and backward BFS analysis, degree and connectivity analysis, Louvain community detection, dynamic update tracking, and a generalization study on Kaggle. They report six findings, including the sparsity and heavy-tailed degree distribution of the graph, the existence of a dominant weakly connected component, the identification of impactful base models (e.g., Llama-3.1-8B), top training datasets, and the high daily churn of models and datasets. A demonstration website and a GitHub repository are referenced without a commit hash or detailed release artifacts.

Significance. If the graph construction is accurate, HuggingGraph would be a valuable public resource for studying model provenance and dataset reuse in the LLM ecosystem, with potential applications to vulnerability propagation, bias auditing, and compliance. The paper's contribution is primarily empirical: it releases a large-scale provenance graph and conducts descriptive analyses. The credibility of the claims rests entirely on the correctness of the node typing and edge extraction pipeline, which is not validated in the manuscript. The paper does not provide machine-checked proofs or a reproducible code release with versioning, so the central artifact cannot be independently verified from the submission alone. The descriptive findings are plausible but several are near-tautological (e.g., a base model with a large forward subgraph is, by construction, one with many outgoing dependency edges).

major comments (4)
  1. [Section 3.2, Table 5] The node typing is not validated, and Table 5 contains entries that are clearly model checkpoints, not datasets: Mistral-v0.1, TinyLlama-1.1B-v1.0, open_llama_3b, Yarn-Mistral-7b-128k, and others are well-known Hugging Face model identifiers. If these are misclassified as datasets, then the 'dataset→model' edges in Table 5 are actually model→model edges, and downstream claims in RQ #4 (Finding #4) such as 'datasets like Mistral-v0.1 spawning hundreds of models' are artifacts of the type confusion. The authors report no validation set, no precision/recall for node or edge extraction, and no manual audit of a random sample. Since the paper's central claim is that the graph 'accurately capture[s] the LLM supply chain information' (Section 1.2), this issue is load-bearing and must be resolved with a concrete validation protocol and corrected tables.
  2. [Section 4.1, RQ #1] The graph is built from a highly biased sample: only 50,156 of 1.8 million models (2.79%) provide a model tree, and only 68,634 of 450K datasets (15.26%) provide dataset cards, yet the findings are phrased as ecosystem-wide properties (e.g., Finding #1: 'The LLM supply chain graph is medium-scale, sparse, and heavy-tailed distribution'). The paper acknowledges the metadata gap but does not quantify how the missing 97.21% of models and 84.74% of datasets could affect degree distributions, WCC size, community structure, or the rankings in Tables 3-6. At minimum, the authors must state that all findings are conditional on the subpopulation with available metadata, and ideally provide a sensitivity analysis or a comparison with a random sample of the full platform.
  3. [Section 3.1, Section 3.2] The edge extraction pipeline is described only qualitatively. The 'NER-based text extraction' is cited to spaCy but no details are given about the model, the keyword set (beyond 'fine-tuned', 'train', 'adapt'), the disambiguation rules, or the error rate. Similarly, the cross-reference link scraping step is said to be 'only employed when reverse dependency data is otherwise inaccessible', but there is no specification of the trigger condition or the deduplication logic when multiple sources (API fields, HTML links, NER) disagree. Without a precise pipeline specification or released code with a commit hash, the reader cannot distinguish systematic misclassification from isolated noise, which is essential for the graph's validity.
  4. [Table 4, Table 5] Several numerical items are inconsistent with the definitions given in the text. For Table 4, the 'Level' column is said to represent 'transformation levels in its backward lineage chain', but for command-r-1-layer the total is 40, fine-tune is 39, quantization is 0, and level is 39, suggesting the level is counted as the number of fine-tune ancestors, not the graph distance (which would include the base model). For Table 5, the sum of the listed categories for Mistral-v0.1 (300+300+193+300 = 1,093) matches the total, but for TinyLlama-1.1B-v1.0 the sum is 300+300+100+28 = 728, which is consistent; however, the text says 'featuring 300 fine-tuned variants and 300 adapters' while the table also lists 100 quantized and 28 merged, so the text is incomplete. The authors should clarify the exact definition of 'Level' and ensure that the textual description of each table fully accounts for all columns.
minor comments (6)
  1. [Abstract and Section 1.2] The phrase 'theLLM supply chain' appears without a space in several places (e.g., 'theLLM supply chain' in the abstract and 'theLLM supply chain' in Section 1.2); please fix the spacing.
  2. [Section 3.2, Figure 2] In the text describing Figure 2, the path '(Base model (Meta-llama))→ merged model (MistLlama)' is written with a double closing parenthesis 'Meta-llama))'; please correct the typo.
  3. [Section 3.3] The forward and backward analysis sections use BFS but do not state whether edge directions are respected (forward: outgoing edges; backward: incoming edges) or whether the graph is treated as unweighted; please state this explicitly.
  4. [Section 4.5, Figure 5] The y-axis label is 'Changed Nodes' but the caption and text refer to 'the number of changed datasets and models'; please clarify whether the figure shows node counts per day or cumulative changes, and specify the time zone used for daily boundaries.
  5. [Section 4.2, Table 2] The table lists community examples but does not report the number of communities found in total or the resolution parameter used for Louvain; adding these details would improve reproducibility.
  6. [References] Reference [43] is cited for a claim about command-r-1-layer's BF16 precision and its lineage, but the cited paper appears to be about zero-shot action evaluation in RL; please verify that the reference is correct or replace it with the model card or a more appropriate source.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports descriptive statistics of a provenance graph built directly from Hugging Face metadata; no fitted parameter is renamed as a prediction and no load-bearing self-citation is used.

full rationale

The paper's contribution is a data-collection pipeline and a directed heterogeneous graph whose edges are read from Hugging Face metadata fields ("finetune", "adapter", "quantization", "merge", and the dataset-side "trained_fine_tune_models") plus NER-augmented card text (Sections 3.1-3.2). Every finding (degree distributions, WCC/SCC sizes, Louvain modularity, forward/backward BFS counts, daily change counts) is a deterministic statistic of that constructed graph. There is no target quantity that is fitted and then called a prediction; "impact" in Table 3 is literally defined as forward-subgraph size, so the label is a restatement of the metric rather than a separately derived conclusion. The apparent listing of model checkpoints (Mistral-v0.1, TinyLlama-1.1B-v1.0) as "datasets" in Table 5 is a possible metadata/typing accuracy issue that affects the validity of the column, but it does not make the reasoning circular. Self-citations are limited to standard graph-algorithm references (e.g., [26-28] for BFS/WCC/SCC), which are external, implementable methods, not a uniqueness theorem or an ansatz smuggled in by the authors. No reduction of a claim to its own input is exhibited, so no circularity step is scored.

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

The paper's central artifact is a data collection, not a theoretical model, so it introduces no free parameters and no invented physical entities. It does rely on unvalidated assumptions about metadata quality and representativeness, listed above.

assumptions (3)
  • domain assumption Hugging Face metadata fields (finetune, adapter, quantization, merge, trained_fine_tune_models) encode actual dependency relationships.
    Invoked in Sections 3.1 and 3.2; if the metadata is missing, noisy, or misinterpreted, the graph edges are wrong.
  • domain assumption Named-entity recognition and cross-reference scraping recover dependency information from unstructured text with sufficient accuracy.
    Described in Section 3.1 under 'Handling missing information'; no precision, recall, or manual validation is provided.
  • ad hoc to paper The subset of models and datasets with available metadata is representative of the overall Hugging Face ecosystem.
    The paper acknowledges only 2.79% of models and 15.26% of datasets are included, yet still frames descriptive statistics as ecosystem-wide findings (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of HuggingGraph: Understanding the Supply Chain of LLM Ecosystem." pith.science (2026). https://pith.science/paper/G3TMBTVK

@misc{pith2026250714240,
  author       = {Pith},
  title        = {Pith review of: HuggingGraph: Understanding the Supply Chain of LLM Ecosystem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G3TMBTVK}},
  note         = {Machine review of arXiv:2507.14240}
}
read the original abstract

Large language models (LLMs) leverage deep learning architectures to process and predict sequences of words, enabling them to perform a wide range of natural language processing tasks, such as translation, summarization, question answering, and content generation. As existing LLMs are often built from base models or other pre-trained models and use external datasets, they can inevitably inherit vulnerabilities, biases, or malicious components that exist in previous models or datasets. Therefore, it is critical to understand these components' origin and development process to detect potential risks, improve model fairness, and ensure compliance with regulatory frameworks. Motivated by that, this project aims to study such relationships between models and datasets, which are the central parts of the LLM supply chain. First, we design a methodology to systematically collect LLMs' supply chain information. Then, we design a new graph to model the relationships between models and datasets, which is a directed heterogeneous graph, having 402,654 nodes and 462,524 edges. Lastly, we perform different types of analysis and make multiple interesting findings.

Figures

Figures reproduced from arXiv: 2507.14240 by the authors.

Figure 1
Figure 1. The number of AI models and datasets (in million [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An example subgraph centering on base model [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Cumulative distribution function (CDF) of WCC. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: The number of changed datasets and models on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. LLM DNA: Tracing Model Evolution via Functional Representations

    cs.LG 2025-09 unverdicted novelty 7.0 of 10

    LLM DNA is introduced as a low-dimensional bi-Lipschitz functional representation proven to satisfy inheritance and genetic determinism, with a training-free extraction pipeline tested on 305 models to reveal relation...

  2. ArtifactLinker: Linking Scientific Artifacts for Automatic State-of-the-Art Discovery

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    ArtifactLinker frames SOTA discovery as missing-link prediction on an artifact graph of models and datasets, with a two-stage ranking-plus-verification pipeline and a new benchmark of 14k artifacts.

Reference graph

Works this paper leans on

67 extracted references · 51 canonical work pages · cited by 2 Pith papers

  1. [1]

    Alexandre Agossah, Frédérique Krupa, Matthieu Perreira Da Silva, and Patrick Le Callet. 2023. Llm-based interaction for content generation: A case study on the perception of employees in an it department. InProceedings of the 2023 ACM International Conference on Interactive Media Experiences. 237–241

    work page 2023

  2. [2]

    Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. 2025. Evolu- tionary optimization of model merging recipes.Nature Machine Intelligence7, 2 (2025), 195–204

    work page 2025

  3. [3]

    Dean Allemang and Juan Sequeda. 2024. Increasing the LLM Accuracy for Question Answering: Ontologies to the Rescue!arXiv preprint arXiv:2405.11706 (2024)

    arXiv 2024

  4. [4]

    Anonymous. 2022. Review of Supply Chain Management in Manufacturing Organizations.ResearchGate(2022). https://www.researchgate.net/publication/ 377659033_Review_of_supply_chain_management_in_manufacturing_ organizations

    work page 2022

  5. [5]

    Mourad Bahani, Aziza El Ouaazizi, and Khalil Maalmi. 2023. The effectiveness of T5, GPT-2, and BERT on text-to-image generation task.Pattern Recognition Letters173 (2023), 57–63

    work page 2023

  6. [6]

    Casper Solheim Bojer and Jens Peder Meldgaard. 2021. Kaggle forecasting compe- titions: An overlooked learning opportunity.International Journal of Forecasting 37, 2 (2021), 587–603

    work page 2021

  7. [7]

    Euan Bonner, Ryan Lege, and Erin Frazier. 2023. Large Language Model-Based Artificial Intelligence in the Language Classroom: Practical Ideas for Teaching. Teaching English with Technology23, 1 (2023), 23–41

    work page 2023

  8. [8]

    Dalton A Brucker-Hahn, Wang Feng, Shanchao Li, Matthew Petillo, Alexan- dru G Bardas, Drew Davidson, and Yuede Ji. 2024. CloudCover: Enforcement of Multi-Hop Network Connections in Microservice Deployments. In2024 Annual Computer Security Applications Conference (ACSAC). IEEE, 1186–1202

    work page 2024

Show all 67 references
  1. [9]

    Feng Chen, Mario Valerio Giuffrida, and Sotirios A Tsaftaris. 2023. Adapting vision foundation models for plant phenotyping. InProceedings of the IEEE/CVF International Conference on Computer Vision. 604–613

    2023

  2. [10]

    M. C. Chou, H. Ye, X. M. Yuan, Y. N. Cheng, L. Chua, Y. Guan, S. E. Lee, and Y. C. Tay. 2006. Analysis of a Software-Focused Products and Service Supply Chain. IEEE Transactions on Industrial Informatics2, 4 (2006), 295–303. doi:10.1109/TII. 2006.884368

    2006

  3. [11]

    Lei Cui, Jiancong Cui, Yuede Ji, Zhiyu Hao, Lun Li, and Zhenquan Ding. 2023. API2Vec: Learning Representations of API Sequences for Malware Detection. In International Symposium on Software Testing and Analysis (ISSTA)

    2023

  4. [12]

    Jelena Cupać, Hendrik Schopmans, and İrem Tuncer-Ebetürk. 2024. Democ- ratization in the age of artificial intelligence: introduction to the special issue. 899–921 pages

    2024

  5. [13]

    Pasquale De Meo, Emilio Ferrara, Giacomo Fiumara, and Alessandro Provetti

  6. [14]

    Hugging Face. [n. d.]. Hugging Face – The AI community building the future. https://huggingface.co/

  7. [15]

    Wang Feng, Shiyang Chen, Hang Liu, and Yuede Ji. 2023. Peek: A Prune-Centric Approach for K Shortest Path Computation. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. 1–14

    2023

  8. [16]

    Qiang Fu, Yuede Ji, and H Howie Huang. 2022. TLPGNN: A lightweight two-level parallelism paradigm for graph neural network computation on GPU. InHPDC

    2022

  9. [17]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  10. [18]

    Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu, Yuan Yao, Ao Zhang, Liang Zhang, et al . 2021. Pre-trained models: Past, present and future.AI Open2 (2021), 225–250

    2021

  11. [19]

    Haojie He, Xingwei Lin, Ziang Weng, Ruijie Zhao, Shuitao Gan, Libo Chen, Yuede Ji, Jiashui Wang, and Zhi Xue. 2024. Code is not Natural Language: Unlock the Power of Semantics-Oriented Graph Representation for Binary Code Similarity Detection. InThe 33rd USENIX Security Sympos...

    2024

  12. [20]

    Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, et al. 2024. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526(2024)

    2024 arXiv

  13. [21]

    Runbang Hu, Chaoqun Li, Xiaojiang Du, and Yuede Ji. 2025. Adaptive Optimiza- tions for Parallel Single-Source Shortest Paths. InProceedings of the 1st FastCode Programming Challenge. 53–56

    2025

  14. [22]

    Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Ka-Wei Lee. 2023. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models.arXiv preprint arXiv:2304.01933(2023)

    2023 arXiv

  15. [23]

    Yuede Ji. 2025. High-Performance Computing for Graph AI: A Top-Down Per- spective. InProceedings of the 15th NSF/TCPP Workshop on Parallel and Distributed Computing Education (EduPar ’25)

    2025

  16. [24]

    Howie Huang

    Yuede Ji, Lei Cui, and H. Howie Huang. 2021. BugGraph: Differentiating Source- Binary Code Similarity with Graph Triplet-Loss Network. In16th ACM ASIA Conference on Computer and Communications Security (AsiaCCS)

    2021

  17. [25]

    Yuede Ji, Mohamed Elsabagh, Ryan Johnson, and Angelos Stavrou. 2021. DEFInit: An Analysis of Exposed Android Init Routines. In30th USENIX Security Sympo- sium (USENIX Security)

    2021

  18. [26]

    Howie Huang

    Yuede Ji and H. Howie Huang. 2020. Aquila: Adaptive Parallel Computation of Graph Connectivity Queries. InProceedings of the 29th International Symposium on High-Performance Parallel and Distributed Computing (HPDC)

    2020

  19. [27]

    Howie Huang

    Yuede Ji, Hang Liu, and H. Howie Huang. 2018. iSpan: Parallel Identification of Strongly Connected Components with Spanning Trees. InInternational Con- ference for High Performance Computing, Networking, Storage and Analysis (SC). IEEE, 731–742

    2018

  20. [28]

    Howie Huang

    Yuede Ji, Hang Liu, and H. Howie Huang. 2020. SwarmGraph: Analyzing Large- Scale In-Memory Graphs on GPUs. InInternational Conference on High Perfor- mance Computing and Communications (HPCC). IEEE

    2020

  21. [29]

    Wenxin Jiang, Nicholas Synovic, Matt Hyatt, Taylor R Schorlemmer, Rohan Sethi, Yung-Hsiang Lu, George K Thiruvathukal, and James C Davis. 2023. An empirical study of pre-trained model reuse in the hugging face deep learning model registry. In2023 IEEE/ACM 45th International Co...

    2023

  22. [30]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, Vol. 1. Minneapolis, Minnesota, 2

    2019

  23. [31]

    TN Kipf. 2016. Semi-Supervised Classification with Graph Convolutional Net- works.arXiv preprint arXiv:1609.02907(2016)

    2016 arXiv

  24. [32]

    Donald E. Knuth. 1974.The Art of Computer Programming, Volume 1: Fundamental Algorithms(2nd ed.). Addison-Wesley

    1974

  25. [33]

    Philippe Laban, Wojciech Kryściński, Divyansh Agarwal, Alexander Richard Fabbri, Caiming Xiong, Shafiq Joty, and Chien-Sheng Wu. 2023. SUMMEDITS: measuring LLM ability at factual reasoning through the lens of summarization. InProceedings of the 2023 Conference on Empirical Met...

    2023

  26. [34]

    Beibin Li, Konstantina Mellou, Bo Zhang, Jeevan Pathuri, and Ishai Menache

  27. [35]

    Chaoqun Li, Runbang Hu, Xiaojiang Du, and Yuede Ji. 2025. Optimized Parallel Breadth-First Search with Adaptive Strategies. InProceedings of the 1st FastCode Programming Challenge. 28–32

    2025

  28. [36]

    Jing Li, Aixin Sun, Jianglei Han, and Chenliang Li. 2020. A survey on deep learning for named entity recognition.IEEE transactions on knowledge and data engineering34, 1 (2020), 50–70

    2020

  29. [37]

    Yinquan Lu, Wenhao Zhu, Lei Li, Yu Qiao, and Fei Yuan. 2024. Llamax: Scal- ing linguistic horizons of llm by enhancing translation capabilities beyond 100 languages.arXiv preprint arXiv:2407.05975(2024)

    2024 arXiv

  30. [38]

    Felix Merrick, Maria Radcliffe, and Rupert Hensley. 2024. Upscaling a smaller llm to more parameters via manual regressive distillation. (2024)

    2024

  31. [39]

    Suvir Mirchandani, Fei Xia, Pete Florence, Brian Ichter, Danny Driess, Montser- rat Gonzalez Arenas, Kanishka Rao, Dorsa Sadigh, and Andy Zeng. 2023. Large language models as general pattern machines.arXiv preprint arXiv:2307.04721 (2023)

    2023 arXiv

  32. [40]

    Moran Mizrahi, Guy Kaplan, Dan Malkin, Rotem Dror, Dafna Shahaf, and Gabriel Stanovsky. 2024. State of what art? a call for multi-prompt llm evaluation. Transactions of the Association for Computational Linguistics12 (2024), 933–949

    2024

  33. [41]

    ONNX Community. 2025. ONNX Model Zoo: Pre-trained Models for ONNX. https://github.com/onnx/models

    2025

  34. [42]

    Minseop Park, Jaeseong You, Markus Nagel, and Simyung Chang. 2022. Quadapter: Adapter for gpt-2 quantization.arXiv preprint arXiv:2211.16912(2022)

    2022 arXiv

  35. [43]

    Eduardo Pignatelli, Johan Ferret, Tim Rockäschel, Edward Grefenstette, Davide Paglieri, Samuel Coward, and Laura Toni. 2024. Assessing the zero-shot capa- bilities of LLMs for action evaluation in RL.arXiv preprint arXiv:2409.12798 (2024)

    2024 arXiv

  36. [44]

    PyTorch Core Team. 2024. PyTorch Hub. https://pytorch.org/hub

    2024

  37. [45]

    Matteo Riva, Tommaso Lorenzo Parigi, Federica Ungaro, and Luca Massimino

  38. [46]

    Paul Röttger, Fabio Pernisi, Bertie Vidgen, and Dirk Hovy. 2024. Safetyprompts: a systematic review of open datasets for evaluating and improving large language model safety.arXiv preprint arXiv:2404.05399(2024)

    2024 arXiv

  39. [47]

    Zhiqiang Shen, Tianhua Tao, Liqun Ma, Willie Neiswanger, Zhengzhong Liu, Hongyi Wang, Bowen Tan, Joel Hestness, Natalia Vassilieva, Daria Soboleva, et al

  40. [48]

    Tanmay Singla, Dharun Anandayuvaraj, Kelechi G Kalu, Taylor R Schorlemmer, and James C Davis. 2023. An empirical study on using large language models to analyze software supply chain security failures. InProceedings of the 2023 Workshop on Software Supply Chain Offensive Resea...

    2023

  41. [49]

    2015.2015 State of the Software Supply Chain Report

    Sonatype. 2015.2015 State of the Software Supply Chain Report. Technical Report. Sonatype. https://www.sonatype.com/hubfs/White_Papers/2015_State_of_the_ Software_Supply_Chain_Report-.pdf

    2015

  42. [50]

    Xin Tan, Taichuan Li, Ruohe Chen, Fang Liu, and Li Zhang. 2024. Challenges of Using Pre-trained Models: the Practitioners’ Perspective.arXiv preprint arXiv:2404.14710(2024)

    2024 arXiv

  43. [51]

    Slimpajama-dc: Understanding data combinations for llm training.arXiv preprint arXiv:2309.10818(2023)

    2023 arXiv

  44. [52]

    2020.Natural language processing with Python and spaCy: A practical introduction

    Yuli Vasiliev. 2020.Natural language processing with Python and spaCy: A practical introduction. No Starch Press

    2020

  45. [53]

    Neelay Velingker, Jason Liu, Amish Sethi, William Dodds, Zhiqiu Xu, Saikat Dutta, Mayur Naik, and Eric Wong. [n. d.]. CLAM: Unifying Finetuning, Quantization, and Pruning by Chaining LLM Adapter Modules. InWorkshop on Efficient Systems for Foundation Models II@ ICML2024

  46. [54]

    Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn. 2024. Will we run out of data? Limits of LLM scaling based on human-generated data.arXiv preprint arXiv:2211.04325(2024), 13–29

    2024 arXiv

  47. [55]

    Hiren Thakkar and A Manimaran. 2023. Comprehensive examination of instruction-based language models: A comparative analysis of mistral-7b and llama-2-7b. In2023 International Conference on Emerging Research in Computa- tional Science (ICERCS). IEEE, 1–6

    2023

  48. [56]

    Pinhuan Wang, Chengying Huan, Zhibin Wang, Chen Tian, Yuede Ji, and Hang Liu. 2025. Bingo: Radix-based Bias Factorization for Random Walk on Dynamic Graphs. InProceedings of the Twentieth European Conference on Computer Systems (Rotterdam, Netherlands)(EuroSys ’25). Associatio...

    2025

  49. [57]

    Shenao Wang, Yanjie Zhao, Xinyi Hou, and Haoyu Wang. 2024. Large language model supply chain: A research agenda.ACM Transactions on Software Engineer- ing and Methodology(2024)

    2024

  50. [58]

    Mengwei Xu, Wangsong Yin, Dongqi Cai, Rongjie Yi, Daliang Xu, Qipeng Wang, Bingyang Wu, Yihao Zhao, Chen Yang, Shihe Wang, et al . 2024. A survey of resource-efficient llm and multimodal foundation models.arXiv preprint arXiv:2401.08092(2024)

    2024 arXiv

  51. [59]

    Kushala VM, Harikrishna Warrier, Yogesh Gupta, et al . 2024. Fine Tuning LLM for Enterprise: Practical Guidelines and Recommendations.arXiv preprint arXiv:2404.10779(2024)

    2024 arXiv

  52. [60]

    Gokul Yenduri, M Ramalingam, G Chemmalar Selvi, Y Supriya, Gautam Srivastava, Praveen Kumar Reddy Maddikunta, G Deepti Raj, Rutvij H Jhaveri, B Prabadevi, Weizheng Wang, et al. 2024. Gpt (generative pre-trained transformer)–a com- prehensive review on enabling technologies, po...

    2024

  53. [61]

    Wentao Zou, Qi Li, Jidong Ge, Chuanyi Li, Xiaoyu Shen, Liguo Huang, and Bin Luo. 2023. A Comprehensive Evaluation of Parameter-Efficient Fine-Tuning on Software Engineering Tasks.arXiv preprint arXiv:2312.15614(2023)

    2023 arXiv

  54. [62]

    Ajdin Čolaković, Aleksandar Ðorđević, Branislav Cvetić, Milan Danilović, and Dejan Vasiljević. 2021. Traditional vs Digital Supply Chains.ResearchGate (2021). https://www.researchgate.net/publication/381617842_Traditional_vs_ Digital_Supply_Chains 9

    2021

  55. [63]

    Zhuoyan Xu, Zhenmei Shi, Junyi Wei, Fangzhou Mu, Yin Li, and Yingyu Liang

  56. [64]

    Towards Few-Shot Adaptation of Foundation Models via Multitask Fine- tuning.arXiv preprint arXiv:2402.15017(2024)

    2024 arXiv

  57. [2011]

    In2011 11th international conference on intelligent systems design and applications

    Generalized louvain method for community detection in large networks. In2011 11th international conference on intelligent systems design and applications. IEEE, 88–93

  58. [2023]

    Large language models for supply chain optimization.arXiv preprint arXiv:2307.03875(2023)

    2023 arXiv

  59. [2024]

    Computational and Structural Biotechnology Reports(2024), 100003

    HuggingFace’s impact on medical applications of artificial intelligence. Computational and Structural Biotechnology Reports(2024), 100003. 8 HuggingGraph: Understanding the Supply Chain of LLM Ecosystem CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea

    2024

Pith tools

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