Intent2QoS: Language Model-Driven Automation of Traffic Shaping Configurations

Burak Kantarci; Sudipta Acharya

REVIEW 3 major objections 4 minor 20 references

Intent2QoS: Language Model-Driven Automation of Traffic Shaping Configurations

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper claims the first end-to-end pipeline that converts high-level traffic-shaping intents into validated, deployable Linux traffic control command sets by combining a queue-simulation semantic model, a language model, and a determini

desk verdict A plausible intent-to-tc pipeline and a useful new benchmark, but the deployability claim is unverified because no generated config is ever executed and the metrics are text similarity to in-house references. read the letter →

arxiv 2601.18974 v1 pith:2T2XS646 submitted 2026-01-26 cs.NI cs.CL

Sudipta Acharya , Burak Kantarci This is my paper

classification cs.NIcs.CL

keywords intent-basednetworkingtrafficcontrolQoSenforcementlanguagemodelsactivequeuemanagementsemanticmodelqueuingtheoryLinuxtcconfiguration

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

The paper sets out to close the last manual step in intent-based networking: turning a high-level statement like "minimize voice delay during 8 PM–1 AM" into concrete Linux traffic control commands. It claims to be the first end-to-end pipeline that does this by first simulating a priority queue with active queue management to derive feasible delay, drop, and utilization numbers, then feeding those numbers plus a traffic profile to a language model that writes structured sub-intents and tc rules, and finally passing the raw output through a deterministic rule-based critic that repairs syntax and policy mistakes. On its own 100-intent benchmark, the best model with AQM-grounded two-shot prompting reaches 0.88 semantic similarity and 0.87 semantic unit coverage, a large margin over zero-shot or one-shot baselines. The significance is that, if the mapping from simulation to real queue behavior holds, operators can express QoS goals in plain language and receive deployable configuration with far less manual expertise.

What carries the argument

The carrying mechanism is the semantic model: a discrete-event simulation of a non-preemptive priority queue with bounded capacity and drop-tail AQM, calibrated offline with target utilization and per-class arrival rates to produce per-class average waiting times, drop rates, and actual utilization. These numbers are injected into the language-model prompt as grounding constraints—for example, avg_wait_high ≤ 0.13 s or drop_rate_low ≤ 5%—which is what turns an underspecified natural-language intent into quantifiable, enforceable sub-intents and then into tc commands. A second deterministic component, the rule-based critic, enforces structural and policy semantics by removing invalid argument

What would settle it

Deploy a critic-validated configuration on a Linux testbed under synthetic traffic matching the traffic profile, then compare measured per-class average delay and drop rate with the semantic model's thresholds (for example, the stated 0.142 s average delay). If measured latency or loss exceeds the threshold by more than the simulation's stated error margin, the simulation-to-device mapping fails and the central claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that an intent-to-configuration pipeline can be built where a queuing-theoretic semantic model, not just prompt examples, supplies the quantitative grounding that language models need to generate correct low-level network configuration. The pipeline has three stages: a digital twin of a non-preemptive priority queue with drop-tail AQM computes per-class delay, drop, and utilization values; a language model uses those values together with a traffic profile to produce declarative sub-intents and then raw tc rules; and a deterministic rule-based critic validates and repairs the raw rules before deployment. On 100 expert-curated intents, the AQM-grounded two-shot strategy ac

Load-bearing premise

The load-bearing premise is that the delay and drop parameters computed by an offline steady-state queue simulation correspond to what the generated tc commands actually produce on a real Linux device, so that a textually similar configuration truly meets the intent.

Editorial extensions

If this is right

Network operators could express QoS goals in plain language and receive validated tc rule sets without handwriting low-level configuration, lowering the expertise barrier.
The corrected sub-intents form a persistent, auditable semantic layer that lets operators verify what a configuration is supposed to enforce and re-synthesize rules for a different platform.
Because the rule-based critic deterministically repairs common language-model errors, the pipeline can tolerate imperfect raw generations and still emit deployable output.
AQM-grounded prompting provides a cheap, model-agnostic way to increase output stability; the threefold variability reduction makes repeated runs more trustworthy.
The 100-intent benchmark, while synthetic, provides a first testbed for intent-to-tc translation and a baseline for later work in this direction.

Reading between the lines

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

If the simulation-to-device mapping is validated on real hardware, the same semantic-model-with-critic design should extend beyond Linux tc to other QoS mechanisms such as RED, CoDel, or fq_codel by swapping the queue model and critic rules, because the pipeline structure is agnostic to the specific queuing discipline.
The evaluation metric is textual similarity to expert references, which is a proxy: two configurations that edit-match a reference could still behave differently under real traffic, so a production-ready claim would need a testbed comparing measured per-class delay and loss against the semantic model's predictions.
Time-sensitive intents are currently encoded as comments in the configuration rather than enforced dynamically; true temporal enforcement would require the pipeline to interoperate with a scheduler or controller that applies the configuration at the specified times.
The reported gains over zero-shot suggest that injecting quantitative grounding—numbers rather than just examples—is the active ingredient; a useful ablation would separate the effect of the semantic model's numbers from the effect of additional examples.

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.

Desk Editor's Note

read the letter

The genuinely new thing here is the assembly: a queuing-simulation semantic model feeding an LLM that generates Linux tc rules, with a rule-based critic for structural fixes, plus a 100-intent benchmark. No one has shipped that exact end-to-end package before, and the benchmark could be useful to the IBN community. The authors also deserve credit for stating their limitations clearly: the semantic model is a steady-state abstraction, production validation is deferred, and the Fig. 4 note concedes that the netem delay parameter is not a true representation of queue dynamics. That honesty makes the paper easier to trust than the abstract implies.

The soft spot is load-bearing, though. The paper claims the generated configurations are “valid,” “correct,” and “deployable,” but zero experiments actually deploy them. No tc command is executed on a Linux host, no traffic is generated, no delay or drop is measured. Every metric in Tables IV and V compares generated text against reference strings written by the same group. High SBERT similarity and semantic-unit coverage show that the LLM learned the authors' formatting conventions, not that the commands enforce the intended QoS. The critic performs only syntactic/structural repairs; it does not check whether a config produces the simulated delay or loss. So the central claim—that this pipeline closes the last manual step in intent-to-tc automation—is not supported by the evidence as written.

A related issue is circularity. The benchmark intents, reference sub-intents, reference configs, and semantic-unit definitions all come from the same team using aligned conventions. That doesn't make the benchmark worthless, but it means the scores partly measure self-consistency. Independent ground truth—for example, configs written by an operator who didn't see the paper, or actual network measurements—would be needed to break the cycle. The absence of released code, prompts, critic rules, and data makes independent reproduction impossible right now.

The sensitivity study is fine for what it is: it shows that AQM-grounded prompting reduces run-to-run variance. But low variance across ten LLM calls says little about whether the outputs are correct under real network conditions.

Who should read this? Researchers working on LLM-based network management, especially those who need a concrete example of benchmarking pitfalls. The pipeline design is worth discussing; the reported numbers are not evidence of deployability. I would send it to peer review—the idea and benchmark deserve referee time—but the authors should be pushed hard to either execute the configs and report network outcomes or scale back the claims. If they refuse to release artifacts, that's a major strike.

For my own work, I wouldn't cite it as a validated system, but I might cite the benchmark as a starting point.

Referee Report

3 major / 4 minor

Summary. The paper proposes Intent2QoS, a three-phase pipeline that translates high-level traffic-shaping intents into Linux tc configurations. Phase I builds an offline SimPy priority-queue digital twin with AQM to derive delay/drop/utilization semantics; Phase II uses an LLM to generate declarative sub-intents and then tc commands, optionally with zero-, one-, or two-shot AQM-grounded prompting; Phase III applies a deterministic rule-based critic for structural correction. The authors evaluate four open models on a 100-intent synthetic benchmark, reporting that LLaMA3 with AQM-guided two-shot prompting achieves 0.88 semantic similarity, 0.87 semantic-unit coverage, and NED 0.16, claiming the first end-to-end intent-to-tc pipeline producing deployable configurations.

Significance. If the central claim were fully supported, the paper would make a useful contribution to intent-based networking by linking declarative intents to concrete Linux queuing-discipline artifacts. The pipeline design is sensible, the deterministic critic is a clean idea, and the synthetic benchmark provides reasonable coverage of QoS objectives and traffic types. However, the current evaluation does not establish deployability: no generated configuration is executed or measured, and the correctness standard is internal to the authors' own references and critic. The strengths — a structured three-phase architecture, careful prompting-strategy comparison, and a reported sensitivity analysis — are real, but they support a claim about text-to-text translation quality, not a claim about validated, deployable configurations.

major comments (3)

[§IV-B, §IV-C, Fig. 4 note] The central claim of producing 'deployable Linux traffic control configuration sets' is not supported by the evaluation. No generated tc configuration is executed on a Linux host, subjected to `tc` parsing, or measured for delay, drop, or priority under load. The metrics in Tables IV and V are all textual or semantic similarities to author-curated reference strings (Table III); a generated config can match a reference lexically while failing to enforce the intended behavior. This concern is acknowledged in the paper itself: the Fig. 4 note states that the netem delay parameter 'represents an average queuing delay ... derived from buffer sizing' but 'in practice, delay arises from queue dynamics; future versions will adopt queue-length-adaptive AQM mechanisms.' The mapping from SimPy avg_wait/drop to `tc netem delay`/`loss` is thus asserted, not validated. I request a real testbed or netw
[§IV-A, Tables III–V] The evaluation is internally consistent but does not establish correctness against an independent standard. The 100-intent dataset, the reference sub-intents/configs, the semantic model, and the traffic profile are all created by the same group; the references are described as 'curated and validated by domain experts' (§IV-A) without inter-annotator agreement, without a precise expert-identification protocol, and without release of data or code. Because the references encode the authors' conventions, the high scores for AQM-grounded prompting partly measure agreement with those conventions. This circularity directly affects the primary outcome claim. I ask for public release of the benchmark, the raw/generated outputs, and the critic rules, plus either an independent validation set or a detailed account of how reference correctness was determined and how disagreements were resolved.
[§III-d, Algorithm 1 lines 12–13] Phase III is described as 'a rule-based critic ... checks and adjusts the rules for correctness and policy compliance,' but the corrections shown in Fig. 4 are purely structural and lexical: removing invalid `prio` arguments, normalizing thresholds/units, and fixing filters. The critic does not query the kernel, simulate the resulting configuration, or verify that a specific traffic class receives the intended delay/drop behavior. It therefore cannot detect semantically wrong but syntactically plausible artifacts, such as a mis-mapped `netem loss` affecting the wrong class or an HTB hierarchy whose rates contradict the stated bandwidth shares. The claim that the output is 'ready for deployment on the physical or virtual edge device' goes beyond what the critic can certify. The paper should either add an execution-based validation layer or weaken the certification language throughout.

minor comments (4)

[Abstract, §I] The claim of being 'the first end-to-end pipeline' is strong given that reference [4] is titled 'Towards end-to-end network intent management with large language models.' Please clarify the precise differentiator against [4] (e.g., generation of Linux tc commands versus higher-level orchestration).
[Table I] The 'Semantic model' and 'traffic profile' columns are vague. Clarify whether the LLM receives numeric thresholds, natural-language descriptions, or structured JSON; this matters for reproducibility of the prompting strategy.
[§IV-C, Tables IV and V] The 'over 30%' gain reported in the abstract and §IV-C is not tied to a specific baseline in Tables IV/V. State which metric and which model/prompt pair is the reference for this percentage.
[Algorithm 1] Line 12 passes `I_d` (the generated sub-intents for the current intent) to `fix_subs`, while the algorithm signature uses `I` for the full input set; clarify the naming to avoid confusion between full and per-intent variables.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation circularity: the pipeline's components are independently specified; internal-benchmark and deployment-validation gaps are correctness risks, not circular reasoning.

full rationale

The paper's claimed derivation chain is not circular in the sense defined by the review rules. Phase I builds a SimPy non-preemptive priority queue with drop-tail AQM and computes waiting times, drop rates, and utilization from stated inputs (target utilization, arrival rates, service rates). Phase II uses these semantic-model metrics plus a traffic profile as prompt context for the LM. Phase III applies a deterministic rule-based critic that repairs syntax and normalizes values. None of these steps defines the outputs in terms of the inputs by construction: the LM is not fitted to reference strings; the critic does not compute the evaluation scores; and the evaluation metrics compare generated text against independently curated expert references (Section IV-A: 'The reference sub-intents and tc configurations used for evaluation were curated and validated by domain experts in network traffic engineering'). No load-bearing self-citation appears: the novelty claim is supported by the cited literature, not by an imported uniqueness theorem from the same authors. The limitations the paper itself states (Section V: semantic model is 'an offline, steady-state abstraction ... does not capture transient dynamics'; Fig. 4 note that delay 'arises from queue dynamics'; future 'production-level validation') identify a real validity gap: scores measure text/semantic alignment with in-house references rather than actual deployment behavior. But a missing external benchmark or an unvalidated simulation-to-tc mapping is a correctness/evidence concern, not circularity, because the derivation does not reduce to its own fitted values or to a self-citation chain. Per the hard rules, an honest non-finding is appropriate.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on domain assumptions about queue-model fidelity, traffic classification, dataset representativeness, metric validity, and critic correctness. The only hand-chosen numeric inputs are the simulation parameters and threshold values that flow into both generation and the critic; no new physical or testable entity is introduced.

free parameters (3)

Queue simulation inputs (lambda_high, lambda_low, mu_high, mu_low, target utilization) = not reported
Phase I semantic model is calibrated by hand-set arrival/service rates and target utilization; these determine delay/drop values (e.g., 0.142 s, 2.8%) that later seed prompts and critic thresholds.
AQM drop thresholds and delay bounds used as semantic constraints = not reported; examples include avg_wait_high <= 0.13 s, drop_rate_low <= 5%
The semantic model outputs thresholds that constrain generated sub-intents; these are chosen, not derived from a stated requirement or real deployed QoS target.
Critic normalization values = 0.142 s, 0.4%, 2.8% in the case study
Section IV-B says the critic 'normalized thresholds/units to semantic model values'; this aligns outputs to the hand-set simulated values rather than to verified system behavior.

assumptions (6)

domain assumption The non-preemptive priority-queue model with drop-tail AQM is a valid semantic model for Linux tc traffic shaping.
Section III Phase I; the delay/drop values from this model are injected into prompts and used by the critic. The Fig. 4 note concedes the model is an approximation.
domain assumption Average simulated waiting time can be instantiated as a fixed netem delay in tc.
The note under Fig. 4 says 'In practice, delay arises from queue dynamics', so equating simulated delay with netem delay is not validated.
domain assumption Traffic classes in operator intents can be identified from IP subnet and port ranges in a synthetic traffic profile.
Section III Phase I uses RFC 1918 blocks and conventional ports; no real traffic data or application-layer inspection is used.
domain assumption The 100 handcrafted intents and expert reference outputs are representative of real-world network operator intents.
Section IV-A; the dataset is synthetic and in-house, with no public benchmark or independent corpus to check representativeness.
domain assumption Lexical/semantic similarity to reference text is a valid proxy for correctness and deployability of tc configurations.
Section IV-C Table III; no tc execution, kernel acceptance test, or network outcome is measured.
domain assumption The rule-based critic's rules are correct and complete enough to guarantee Linux tc syntax and policy compliance.
Section III Phase III; the critic rules are not listed or validated against an actual tc parser or kernel.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intent2QoS: Language Model-Driven Automation of Traffic Shaping Configurations." pith.science (2026). https://pith.science/paper/2T2XS646

@misc{pith2026260118974,
  author       = {Pith},
  title        = {Pith review of: Intent2QoS: Language Model-Driven Automation of Traffic Shaping Configurations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2T2XS646}},
  note         = {Machine review of arXiv:2601.18974}
}

read the original abstract

Traffic shaping and Quality of Service (QoS) enforcement are critical for managing bandwidth, latency, and fairness in networks. These tasks often rely on low-level traffic control settings, which require manual setup and technical expertise. This paper presents an automated framework that converts high-level traffic shaping intents in natural or declarative language into valid and correct traffic control rules. To the best of our knowledge, we present the first end-to-end pipeline that ties intent translation in a queuing-theoretic semantic model and, with a rule-based critic, yields deployable Linux traffic control configuration sets. The framework has three steps: (1) a queuing simulation with priority scheduling and Active Queue Management (AQM) builds a semantic model; (2) a language model, using this semantic model and a traffic profile, generates sub-intents and configuration rules; and (3) a rule-based critic checks and adjusts the rules for correctness and policy compliance. We evaluate multiple language models by generating traffic control commands from business intents that comply with relevant standards for traffic control protocols. Experimental results on 100 intents show significant gains, with LLaMA3 reaching 0.88 semantic similarity and 0.87 semantic coverage, outperforming other models by over 30\. A thorough sensitivity study demonstrates that AQM-guided prompting reduces variability threefold compared to zero-shot baselines.

Figures

Figures reproduced from arXiv: 2601.18974 by the authors.

**Figure 2.** Overview of the proposed edge-executed pipeline for [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Summary statistics for the 100-intent dataset [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: End-to-end case study flow: High-level Intent [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 4 linked inside Pith

[1]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2022

2022
[2]

Intent-based networking-concepts and definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent-based networking-concepts and definitions,” 2022

2022
[3]

Llm-based policy generation for intent-based management of applications,

K. Dzeparoska, J. Lin, A. Tizghadam, and A. Leon-Garcia, “Llm-based policy generation for intent-based management of applications,” in2023 19th International Conference on Network and Service Management (CNSM). IEEE, 2023, pp. 1–7

2023
[4]

Towards end-to-end network intent management with large language models,

L. Dinh, S. Cherrared, X. Huang, and F. Guillemin, “Towards end-to-end network intent management with large language models,”arXiv preprint arXiv:2504.13589, 2025

arXiv 2025
[5]

Netconfeval: Can llms facilitate network configuration?

C. Wang, M. Scazzariello, A. Farshin, S. Ferlin, D. Kosti ´c, and M. Chiesa, “Netconfeval: Can llms facilitate network configuration?” Proceedings of the ACM on Networking, vol. 2, no. CoNEXT2, pp. 1– 25, 2024

2024
[6]

Ferguson and G

P. Ferguson and G. Huston,Quality of service: delivering QoS on the Internet and in corporate networks. John Wiley & Sons, Inc., 1998

1998
[7]

Llama 3: Open foundation and instruction models,

H. Touvron, T. Lavril, G. Izacardet al., “Llama 3: Open foundation and instruction models,” https://ai.meta.com/llama/, 2024, meta AI

2024
[8]

Mistral 7b,

Y . Jiang and M. A. Team, “Mistral 7b,” https://mistral.ai/news/ announcing-mistral-7b, 2023, mistral AI

2023

Show all 20 references

[9]

Gemma: Open models by google deepmind,

DeepMind, “Gemma: Open models by google deepmind,” https:// deepmind.google/technologies/gemma/, 2024, google DeepMind

2024
[10]

Ieee 802.1qcr asyn- chronous traffic shaping with linux traffic control,

C. Pfefferle, F. Wiedner, and C. Schwarzenberg, “Ieee 802.1qcr asyn- chronous traffic shaping with linux traffic control,”Technical University of Munich Seminar Report, 2022

2022
[11]

Preconfig: A pretrained model for automating network configuration,

F. Li, H. Lang, J. Zhang, J. Shen, and X. Wang, “Preconfig: A pretrained model for automating network configuration,”arXiv preprint arXiv:2403.09369, 2024

2024 arXiv
[12]

Intent assurance using llms guided by intent drift,

K. Dzeparoska, A. Tizghadam, and A. Leon-Garcia, “Intent assurance using llms guided by intent drift,” inNOMS 2024-2024 IEEE Network Operations and Management Symposium. IEEE, 2024, pp. 1–7

2024
[13]

Address allocation for private internets,

Y . Rekhter, B. Moskowitz, D. Karrenberg, G. J. de Groot, and E. Lear, “Address allocation for private internets,” https://datatracker.ietf.org/doc/ html/rfc1918, 1996, rFC 1918

1996
[14]

Service name and transport protocol port number registry,

“Service name and transport protocol port number registry,” https://www.iana.org/assignments/service-names-port-numbers/ service-names-port-numbers.xhtml, iANA, Accessed July 2025

2025
[15]

Internet of things: A survey on enabling technologies, protocols, and applications,

A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, “Internet of things: A survey on enabling technologies, protocols, and applications,”IEEE Communications Surveys & Tutorials, vol. 17, no. 4, pp. 2347–2376, 2015

2015
[16]

Sentence-bert: Sentence embeddings using siamese bert-networks,

N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,”arXiv preprint arXiv:1908.10084, 2019

1908 arXiv
[17]

Rouge: A package for automatic evaluation of summaries,

C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” inText summarization branches out, 2004, pp. 74–81

2004
[18]

Sch ¨utze, C

H. Sch ¨utze, C. D. Manning, and P. Raghavan,Introduction to informa- tion retrieval. Cambridge University Press Cambridge, 2008, vol. 39

2008
[19]

Few- shot natural language generation for task-oriented dialog,

B. Peng, C. Zhu, C. Li, X. Li, J. Li, M. Zeng, and J. Gao, “Few- shot natural language generation for task-oriented dialog,”arXiv preprint arXiv:2002.12328, 2020

2002 arXiv
[20]

A normalized levenshtein distance metric,

L. Yujian and L. Bo, “A normalized levenshtein distance metric,”IEEE transactions on pattern analysis and machine intelligence, vol. 29, no. 6, pp. 1091–1095, 2007

2007

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2022

2022

[2] [2]

Intent-based networking-concepts and definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent-based networking-concepts and definitions,” 2022

2022

[3] [3]

Llm-based policy generation for intent-based management of applications,

K. Dzeparoska, J. Lin, A. Tizghadam, and A. Leon-Garcia, “Llm-based policy generation for intent-based management of applications,” in2023 19th International Conference on Network and Service Management (CNSM). IEEE, 2023, pp. 1–7

2023

[4] [4]

Towards end-to-end network intent management with large language models,

L. Dinh, S. Cherrared, X. Huang, and F. Guillemin, “Towards end-to-end network intent management with large language models,”arXiv preprint arXiv:2504.13589, 2025

arXiv 2025

[5] [5]

Netconfeval: Can llms facilitate network configuration?

C. Wang, M. Scazzariello, A. Farshin, S. Ferlin, D. Kosti ´c, and M. Chiesa, “Netconfeval: Can llms facilitate network configuration?” Proceedings of the ACM on Networking, vol. 2, no. CoNEXT2, pp. 1– 25, 2024

2024

[6] [6]

Ferguson and G

P. Ferguson and G. Huston,Quality of service: delivering QoS on the Internet and in corporate networks. John Wiley & Sons, Inc., 1998

1998

[7] [7]

Llama 3: Open foundation and instruction models,

H. Touvron, T. Lavril, G. Izacardet al., “Llama 3: Open foundation and instruction models,” https://ai.meta.com/llama/, 2024, meta AI

2024

[8] [8]

Mistral 7b,

Y . Jiang and M. A. Team, “Mistral 7b,” https://mistral.ai/news/ announcing-mistral-7b, 2023, mistral AI

2023