Pith. sign in

REVIEW 4 major objections 5 minor 44 references

FedChip: Federated LLM for Artificial Intelligence Accelerator Chip Design

T0 review · 4 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 Federated fine-tuning among multiple chip-design parties improves LLM-generated accelerator designs by more than 77% over high-end general-purpose LLMs, while keeping proprietary data private.

desk verdict The dataset and the federated recipe are real; the 77% number is a placeholder until Chip@k is defined and the experiments are visible. read the letter →

arxiv 2508.13162 v1 pith:KPPDUSZE submitted 2025-07-23 cs.AR cs.LG

Mahmoud Nazzal , Khoa Nguyen , Deepak Vungarala , Ramtin Zand , Shaahin Angizi , Hai Phan , Abdallah Khreishah This is my paper
classification cs.ARcs.LG
keywords FederatedlearningLargelanguagemodelsHardwaredesignautomationAIacceleratorPPAoptimizationDataprivacyChip@kmetricAPTPU-Gendataset
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

FedChip is a federated fine-tuning system that lets several chip-design companies jointly improve a shared large language model for generating AI accelerator designs without ever exposing their proprietary data. The paper claims that after a few rounds of local fine-tuning on emulated company-specific datasets, the federated model produces designs that score more than 77% higher on a new quality metric, Chip@k, than high-end general-purpose LLMs such as GPT-4o and Claude 3.5 Sonnet. To make this concrete, the authors release APTPU-Gen, a public dataset of 30,000 TPU design variations with power, performance, and area (PPA) values, and propose Chip@k, which statistically measures how often generated designs meet predefined acceptance criteria. The central idea is that federated learning can overcome the two obstacles holding back LLM-based hardware design: the shortage of domain-specific data caused by IP protections, and the difficulty of balancing competing PPA objectives.

What carries the argument

The load-bearing components are the FedChip federated fine-tuning pipeline, which combines local QLoRA-style parameter-efficient updates with federated aggregation; APTPU-Gen, a 30k-sample dataset of TPU design variations generated from APTPU and synthesized with OpenROAD to obtain PPA labels; and Chip@k, a statistical evaluation metric that counts how many of $k$ generated designs satisfy predefined PPA acceptance criteria. The system also uses a simulated multi-party setup in which K-means clustering partitions APTPU-Gen into three PPA-distinct clusters and a 20% Dirichlet reassignment introduces non-IID noise, emulating competing companies with different design priorities. These pieces work together to show that federated fine-tuning improves design quality on a realistic, privacy-constrained benchmark.

What would settle it

Run FedChip on real proprietary accelerator design data from at least three independently operated design teams, or on a public industrial-strength hardware dataset not derived from APTPU, and compare the Chip@k scores against the paper's synthetic-cluster baseline; if the federated model does not beat each party's locally fine-tuned model and the frontier LLM by a meaningful margin, the 77% claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that federated fine-tuning across multiple parties with non-identically distributed design data can turn a generic pre-trained LLM into a competitive AI-accelerator design generator, improving design quality by over 77% relative to frontier LLMs while keeping raw design data on each party's premises. The authors demonstrate this by partitioning the APTPU-Gen dataset into three clusters that emulate three design companies with different PPA priorities, fine-tuning a shared LLM locally at each party via parameter-efficient methods, and aggregating only the model updates. They evaluate generated designs with the proposed Chip@k metric, which accepts a design only when it meets user-defined thresholds across PPA metrics. The result is framed as evidence that privacy-preserving collaboration, rather than larger generic models, is the practical path to domain-competent LLMs in hardware design.

Load-bearing premise

The claim depends on the assumption that three K-means clusters of APTPU-Gen with 20% of samples randomly reassigned via Dirichlet sampling faithfully emulate real proprietary chip-design datasets from competing companies; if actual industrial design data are substantially more heterogeneous or differently distributed, the measured 77% improvement may not generalize to federated deployment.

Editorial extensions

If this is right

  • If FedChip's claim holds, chip companies can collectively advance LLM-based design automation without handing over their architectural IP, since only model updates are shared.
  • The Chip@k metric gives hardware teams a concrete, statistical way to judge whether a generated design is actually useful, rather than merely syntactically valid code.
  • The public APTPU-Gen dataset and the methodology for simulating non-IID client data provide a reproducible testbed for future federated LLM fine-tuning work in hardware design.
  • Because the federated model outperforms even strong proprietary LLMs, the result suggests that domain-specific tuning, not model scale, is the dominant factor in automated accelerator design quality.

Reading between the lines

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

  • A direct next test is to run the same federated fine-tuning protocol on genuinely industrial multi-company design data; the 77% gain might shrink, grow, or vanish depending on how much the real data resemble the synthetic clusters.
  • The privacy guarantee as stated means no raw data leaves each party, but it is not a formal differential-privacy guarantee; an adversary with access to model updates might still infer properties of the local datasets.
  • The recipe could plausibly transfer to other hardware description languages or accelerator families beyond TPU-style APTPU cores, though the PPA trade-offs would differ and the improvement would need re-measurement.
  • Chip@k only counts statistical acceptance; pairing it with an actual synthesis-and-verification pass on the generated Verilog would make the quality claim more robust to designs that happen to hit thresholds by chance.
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 / 5 minor

Summary. The paper proposes FedChip, a federated fine-tuning framework that allows multiple chip-design parties to collaboratively improve a shared LLM for AI accelerator hardware generation without sharing proprietary data. The authors create and release APTPU-Gen, a 30k-sample dataset of accelerator design variations with PPA metrics, and introduce a new evaluation metric, Chip@k, intended to statistically assess generated designs against acceptance criteria. To emulate multi-party deployment, the dataset is split into three clusters via K-means with 20% Dirichlet-based reassignment, and the non-IID-ness of the resulting sub-datasets is analyzed using KL and Jensen-Shannon divergences. The central claim, stated in the Abstract, is that FedChip improves design quality by more than 77% over high-end LLMs while maintaining data privacy. The provided manuscript text, however, does not include the experimental results section that would support this headline claim.

Significance. If substantiated, this work would address a real practical barrier in LLM-based chip design: the inability to fine-tune on proprietary data due to IP and privacy concerns. The idea of federated fine-tuning for hardware design is timely, and the release of APTPU-Gen provides a useful public benchmark for future work. The non-IID analysis using divergence measures is a reasonable step toward justifying federated learning and is clearly presented. However, the central 77% improvement claim is not currently verifiable because the manuscript lacks the precise definition of Chip@k, the experimental setup, baseline comparisons, and any external validation such as physical synthesis or timing closure. The synthetic party split is a plausible modeling choice but requires sensitivity analysis to be convincing. Strengths include the creation and planned release of the dataset and a clear articulation of the federated training concept.

major comments (4)
  1. [Abstract / Experimental Results] The central claim that FedChip improves design quality by more than 77% over high-end LLMs is not supported in the visible text: there is no experimental results section, no tables comparing against baseline LLMs, no error bars or multiple runs, and no concrete Chip@k values. Please add the full experimental setup, the exact definition and thresholds of Chip@k, and the raw results with statistical variability.
  2. [Abstract / Section III.A] The Chip@k metric is introduced in the Abstract but never formally defined in the provided text. Its 'predefined acceptance criteria' are not specified, and the metric is not calibrated against physical synthesis or an independent expert evaluation. Without this, the reported 77% margin is an internal, self-referential score that may reward matching the APTPU-Gen distribution rather than generating synthesizable, correct hardware. Please provide the formal metric definition, the threshold selection procedure, and a validation against physical design metrics.
  3. [Section III.A] The synthetic party-split procedure uses K-means on the APTPU-Gen dataset followed by 20% Dirichlet-based reassignment to emulate three companies. This assumes that real proprietary design data from competing companies resemble such clusters; no evidence is given for this assumption, and the subsequent non-IID analysis (Fig. 5) is partly circular because the clusters are constructed to be different. Please justify this modeling assumption and provide a sensitivity analysis over K, the reassignment fraction, and clustering initialization.
  4. [Overall experimental evaluation] The manuscript does not report the number of runs, standard deviations, or statistical significance for the claimed 77% improvement. Given that the paper introduces a new evaluation metric and uses several free parameters (LoRA rank, learning rate, aggregation rounds, Dirichlet reassignment fraction), please include a hyperparameter sensitivity analysis and error bars so that the central result is not a single-point estimate.
minor comments (5)
  1. [Section III.A] The text says 'the subscripts in the cluster pairs denote the pairwise comparisons' for Fig. 5, but the figure caption is not shown in the provided text; please ensure the figure labels are clear.
  2. [References] References [28] and [39] are duplicate LoRA citations; consolidate them.
  3. [Section III.A] The phrase 'coefficient sharing' is ambiguous; it likely means sharing of model parameters or updates. Please rephrase.
  4. [Abstract / Introduction] The paper mentions 'hallucinated variables' and 'inability to handle complex AI accelerators' as failure modes of general LLMs, citing [12]; a more specific citation to hardware-design LLM evaluations would strengthen this claim.
  5. [Section III.A] The mapping of clusters to parties (energy-efficient, high-complexity, timing-critical) is introduced after the distributions are shown; it would be clearer to state this mapping before the cluster descriptions.

Circularity Check

0 steps flagged · score 1.0 of 10

No demonstrated circularity: the 77% claim is benchmarked against external LLMs on an original released dataset; the undefined Chip@k metric is a validation gap, not a self-referential reduction.

full rationale

The available text does not exhibit any step in which a prediction equals an input by construction. FedChip is a parameter-efficient federated fine-tuning pipeline applied to a newly released dataset, APTPU-Gen, and the headline improvement is measured against external high-end LLMs (GPT-4o, Claude 3.5 Sonnet), not against a quantity fitted from those models. The synthetic three-party split is a K-means clustering of APTPU-Gen with 20% Dirichlet reassignment; this weakens the external validity of the privacy story, but it does not make the improvement score definitionally identical to the training objective. The one substantive gap is Chip@k: the abstract states that it 'statistically evaluates the quality of generated designs against predefined acceptance criteria,' but those criteria and any validation against physical synthesis or timing closure are not present in the supplied text. The reference list hints that the metric may rely on the three-sigma rule, yet no equation or threshold definition is given, so one cannot exhibit a specific reduction from Chip@k to the APTPU-Gen training distribution. Absence of specification is an evidence and correctness concern, not a demonstrated circular step. No self-citation is load-bearing: APTPU [32] and SA-DS [9] are prior-work dataset sources, and the core federated-learning ideas are cited to the general literature [17]-[19]. The derivation chain is therefore not circular on the available evidence, though the central claim is under-audited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim depends on the synthetic three-party split, the authors' own evaluation metric and thresholds, and the assumption that FL protects privacy. These are modeling choices, not derived results.

free parameters (4)
  • Number of parties/clusters (K) = 3
    K-means is configured to produce three clusters to emulate three companies; the headline result is tied to this setup.
  • Dirichlet reassignment fraction = 0.2
    20% of data points are reassigned to clusters to create non-IID variability; this controls how different the synthetic parties are.
  • Chip@k acceptance thresholds = not specified in provided text
    The 77% improvement is measured with Chip@k, which requires predefined PPA acceptance criteria; the thresholds are not shown in the excerpt.
  • Fine-tuning hyperparameters (LoRA rank, learning rate, aggregation rounds) = not reported in provided text
    Standard choices in LLM fine-tuning that affect the final model quality; not visible in the provided sections.
assumptions (4)
  • domain assumption Three K-means clusters with Dirichlet perturbation approximate real proprietary chip-design datasets.
    All multi-party experiments use these synthetic clusters (Section III.A); no real proprietary data or external validation appears.
  • domain assumption Chip@k with the chosen acceptance criteria is a valid measure of generated design quality.
    The headline result is defined through Chip@k (Abstract, Section I); its thresholds are not visible in the provided text.
  • domain assumption Federated model averaging preserves data privacy.
    The abstract claims privacy without presenting a privacy analysis or differential privacy mechanism.
  • domain assumption APTPU-Gen PPA labels from the APTPU/OpenROAD flow are accurate for training and evaluation.
    The dataset is the training and evaluation substrate; correctness of power, performance, and area values is taken for granted.
invented entities (2)
  • APTPU-Gen dataset independent evidence
    purpose: 30k generated accelerator design variations with PPA labels for federated training and evaluation.
    The dataset is released on GitHub, so it can be independently inspected and reused.
  • Chip@k metric
    purpose: Statistical evaluation of generated designs against preset acceptance criteria.
    Defined in this paper; no external validation or prior benchmark establishes its thresholds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedChip: Federated LLM for Artificial Intelligence Accelerator Chip Design." pith.science (2026). https://pith.science/paper/KPPDUSZE

@misc{pith2026250813162,
  author       = {Pith},
  title        = {Pith review of: FedChip: Federated LLM for Artificial Intelligence Accelerator Chip Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPPDUSZE}},
  note         = {Machine review of arXiv:2508.13162}
}
read the original abstract

AI hardware design is advancing rapidly, driven by the promise of design automation to make chip development faster, more efficient, and more accessible to a wide range of users. Amongst automation tools, Large Language Models (LLMs) offer a promising solution by automating and streamlining parts of the design process. However, their potential is hindered by data privacy concerns and the lack of domain-specific training. To address this, we introduce FedChip, a Federated fine-tuning approach that enables multiple Chip design parties to collaboratively enhance a shared LLM dedicated for automated hardware design generation while protecting proprietary data. FedChip enables parties to train the model on proprietary local data and improve the shared LLM's performance. To exemplify FedChip's deployment, we create and release APTPU-Gen, a dataset of 30k design variations spanning various performance metric values such as power, performance, and area (PPA). To encourage the LLM to generate designs that achieve a balance across multiple quality metrics, we propose a new design evaluation metric, Chip@k, which statistically evaluates the quality of generated designs against predefined acceptance criteria. Experimental results show that FedChip improves design quality by more than 77% over high-end LLMs while maintaining data privacy

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 17 canonical work pages

  1. [1]

    Artificial intelligence (ai) hardware market to exceed usd 84.9 billion by 2031: Skyquest technology,

    SkyQuest Technology, “Artificial intelligence (ai) hardware market to exceed usd 84.9 billion by 2031: Skyquest technology,” 2023, accessed: 2024-10-31. [Online]. Available: https://www.prnewswire.com

    work page 2023

  2. [2]

    A survey on collaborative dnn inference for edge intelligence,

    W.-Q. Renet al., “A survey on collaborative dnn inference for edge intelligence,”Machine Intelligence Research, vol. 20, no. 3, pp. 370– 395, 2023

    work page 2023

  3. [3]

    Gemmini: Enabling systematic deep-learning architec- ture evaluation via full-stack integration,

    H. Gencet al., “Gemmini: Enabling systematic deep-learning architec- ture evaluation via full-stack integration,” inDAC. IEEE, 2021, pp. 769–774

    work page 2021

  4. [4]

    Verigen: A large language model for verilog code generation,

    S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri, and S. Garg, “Verigen: A large language model for verilog code generation,” ACM Transactions on Design Automation of Electronic Systems, vol. 29, no. 3, pp. 1–31, 2024

    2024

  5. [5]

    Chateda: A large language model powered autonomous agent for eda,

    H. Wu, Z. He, X. Zhang, X. Yao, S. Zheng, H. Zheng, and B. Yu, “Chateda: A large language model powered autonomous agent for eda,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024

    work page 2024

  6. [6]

    Chipgpt: How far are we from natural language hardware design,

    K. Chang, Y . Wang, H. Ren, M. Wang, S. Liang, Y . Han, H. Li, and X. Li, “Chipgpt: How far are we from natural language hardware design,”arXiv preprint arXiv:2305.14019, 2023

    arXiv 2023

  7. [7]

    Autochip: Automating hdl generation using llm feedback,

    S. Thakuret al., “Autochip: Automating hdl generation using llm feedback,”arXiv preprint arXiv:2311.04887, 2023

    arXiv 2023

  8. [8]

    Gpt4aigchip: Towards next-generation ai accelerator design automation via large language models,

    Y . Fu, Y . Zhang, Z. Yu, S. Li, Z. Ye, C. Li, C. Wan, and Y . C. Lin, “Gpt4aigchip: Towards next-generation ai accelerator design automation via large language models,” in2023 IEEE/ACM International Confer- ence on Computer Aided Design (ICCAD). IEEE, 2023, pp. 1–9

    2023

Show all 44 references
  1. [9]

    Sa-ds: A dataset for large language model- driven ai accelerator design generation,

    D. Vungarala, M. Nazzal, M. Morsali, C. Zhang, A. Ghosh, A. Khreishah, and S. Angizi, “Sa-ds: A dataset for large language model- driven ai accelerator design generation,”arXiv e-prints, pp. arXiv–2404, 2024

    2024

  2. [10]

    Gpt-4o: Openai’s advanced generative language model,

    OpenAI, “Gpt-4o: Openai’s advanced generative language model,” Available online: https://openai.com, 2024, accessed: 2024-11-18

    2024

  3. [11]

    Claude 3.5 sonnet: Anthropic’s advanced language model,

    Anthropic, “Claude 3.5 sonnet: Anthropic’s advanced language model,” Available online: https://www.anthropic.com, 2024, accessed: 2024-11- 18

    2024

  4. [12]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskillet al., “On the opportunities and risks of foundation models,”arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  5. [13]

    Llm-aided efficient hardware design automation,

    K. Xu, R. Qiu, Z. Zhao, G. L. Zhang, U. Schlichtmann, and B. Li, “Llm-aided efficient hardware design automation,”arXiv preprint arXiv:2410.18582, 2024

    2024 arXiv

  6. [14]

    The dawn of ai-native eda: Op- portunities and challenges of large circuit models,

    L. Chen, Y . Chen, Z. Chu, W. Fang, T.-Y . Ho, R. Huang, Y . Huang, S. Khan, M. Li, X. Liet al., “The dawn of ai-native eda: Op- portunities and challenges of large circuit models,”arXiv preprint arXiv:2403.07257, 2024

    2024 arXiv

  7. [15]

    Cybercriminals who breached nvidia issue one of the most unusual demands ever,

    D. Goodin, “Cybercriminals who breached nvidia issue one of the most unusual demands ever,”Ars Technica, March 2022, accessed: 2024-10-31. [Online]. Available: https://arstechnica.com

    2022

  8. [16]

    Fast and accurate ppa modeling with transfer learning,

    W. R. Davis, P. Franzon, L. Francisco, B. Huggins, and R. Jain, “Fast and accurate ppa modeling with transfer learning,” in2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD). IEEE, 2021, pp. 1–8

    2021

  9. [17]

    Federated machine learning: Concept and applications,

    Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,”ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 2, pp. 1–19, 2019

    2019

  10. [18]

    Towards federated learning at scale: Syste m design,

    K. Bonawitz, “Towards federated learning at scale: Syste m design,” arXiv preprint arXiv:1902.01046, 2019

    1902 arXiv

  11. [19]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. PMLR, 2017, pp. 1273– 1282

    2017

  12. [20]

    Federated fine-tuning of large language models under heterogeneous language tasks and client resources,

    J. Bai, D. Chen, B. Qian, L. Yao, and Y . Li, “Federated fine-tuning of large language models under heterogeneous language tasks and client resources,”arXiv preprint arXiv:2402.11505, 2024

    2024 arXiv

  13. [21]

    Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning,

    W. Kuang, B. Qian, Z. Li, D. Chen, D. Gao, X. Pan, Y . Xie, Y . Li, B. Ding, and J. Zhou, “Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning,” inProceed- ings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Dat...

    2024

  14. [22]

    Federated lora with sparse communication,

    K. Kuo, A. Raje, K. Rajesh, and V . Smith, “Federated lora with sparse communication,”arXiv preprint arXiv:2406.05233, 2024

    2024 arXiv

  15. [23]

    Learned hardware/software co-design of neural accelerators,

    Z. Shi, C. Sakhuja, M. Hashemi, K. Swersky, and C. Lin, “Learned hardware/software co-design of neural accelerators,”arXiv preprint arXiv:2010.02075, 2020

    2010 arXiv

  16. [24]

    Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers,

    D. Dai, Y . Sun, L. Dong, Y . Hao, S. Ma, Z. Sui, and F. Wei, “Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers,”arXiv preprint arXiv:2212.10559, 2022

    2022 arXiv

  17. [25]

    Language models are few-shot learners,

    T. B. Brown, “Language models are few-shot learners,”arXiv preprint arXiv:2005.14165, 2020

    2005 arXiv

  18. [26]

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

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2019. [Online]. Available: https://arxiv.org/abs/1810.04805

    2019 arXiv

  19. [27]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, “Roberta: A robustly optimized bert pretraining approach,”arXiv preprint arXiv:1907.11692, vol. 364, 2019

    1907 arXiv

  20. [29]

    Qlora: Efficient finetuning of quantized llms,

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “Qlora: Efficient finetuning of quantized llms,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  21. [30]

    Opt: Open pre-trained transformer language models,

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V . Linet al., “Opt: Open pre-trained transformer language models,”arXiv preprint arXiv:2205.01068, 2022

    2022 arXiv

  22. [31]

    Motivation for and evaluation of the first tensor processing unit,

    N. Jouppi, C. Young, N. Patil, and D. Patterson, “Motivation for and evaluation of the first tensor processing unit,”IEEE Micro, vol. 38, no. 3, pp. 10–19, 2018

    2018

  23. [32]

    Aptpu: Approximate computing based tensor processing unit,

    M. E. Elbtity, P. S. Chandarana, B. Reidy, J. K. Eshraghian, and R. Zand, “Aptpu: Approximate computing based tensor processing unit,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 69, no. 12, pp. 5135–5146, 2022

    2022

  24. [33]

    [Online]

    (2018) Openroad. [Online]. Available: https://github.com/ The-OpenROAD-Project/OpenROAD

    2018

  25. [34]

    Least squares quantization in pcm,

    S. Lloyd, “Least squares quantization in pcm,”IEEE transactions on information theory, vol. 28, no. 2, pp. 129–137, 1982

    1982

  26. [35]

    Estimating a dirichlet distribution,

    T. Minka, “Estimating a dirichlet distribution,” 2000

    2000

  27. [36]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,”The annals of mathematical statistics, vol. 22, no. 1, pp. 79–86, 1951

    1951

  28. [37]

    The three sigma rule,

    F. Pukelsheim, “The three sigma rule,”The American Statistician, vol. 48, no. 2, pp. 88–91, 1994

    1994

  29. [38]

    Openfedllm: Training large language models on decentralized private data via federated learning,

    R. Ye, W. Wang, J. Chai, D. Li, Z. Li, Y . Xu, Y . Du, Y . Wang, and S. Chen, “Openfedllm: Training large language models on decentralized private data via federated learning,” 2024. [Online]. Available: https://arxiv.org/abs/2402.06954

    2024 arXiv

  30. [39]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,”

  31. [40]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,”

  32. [41]

    Stanford alpaca: An instruction-following llama model,

    R. Taori, I. Gulrajani, T. Zhang, Y . Dubois, X. Li, C. Guestrin, P. Liang, and T. B. Hashimoto, “Stanford alpaca: An instruction-following llama model,” https://github.com/tatsulab/stanford alpaca, 2023

    2023

  33. [42]

    Gpt-o1: Specialized generative pre-trained transformer for do- main applications,

    OpenAI, “Gpt-o1: Specialized generative pre-trained transformer for do- main applications,” Available online: https://openai.com, 2024, accessed: 2024-11-18

    2024

  34. [43]

    Gemini advanced: High-performance large language model,

    Gemini AI Lab, “Gemini advanced: High-performance large language model,” Available online: https://www.gemini-ai.com, 2024, accessed: 2024-11-18

    2024

  35. [2019]

    Available: https://arxiv.org/abs/1711.05101

    [Online]. Available: https://arxiv.org/abs/1711.05101

    arXiv

  36. [2021]

    Available: https://arxiv.org/abs/2106.09685

    [Online]. Available: https://arxiv.org/abs/2106.09685

    arXiv

Pith tools

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