Comparative Evaluation of Prompting and Fine-Tuning for Applying Large Language Models to Grid-Structured Geospatial Data

Akash Dhruv; Jordan Branham; Tanwi Mallick; Yangxinyu Xie

REVIEW 4 major objections 4 minor 20 references

Comparative Evaluation of Prompting and Fine-Tuning for Applying Large Language Models to Grid-Structured Geospatial Data

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning an 8B large language model on about 100 examples lifts accuracy on climate grid queries from 0.2889 to 1.0, making it essential for precise extraction of values, units, and scenarios.

desk verdict A small, honest pilot on LLMs for gridded climate data, but the central claim that fine-tuning is essential rests on a 12-example test set and needs a real leakage check before it can be believed. read the letter →

arxiv 2505.17116 v1 pith:AHXHNF23 submitted 2025-05-21 cs.CL cs.ET

Akash Dhruv , Yangxinyu Xie , Jordan Branham , Tanwi Mallick This is my paper

classification cs.CLcs.ET

keywords largelanguagemodelsgeospatialdatagrid-structuredfine-tuningpromptingClimRRclimateprojectionsLoRA

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 tests whether a small open-weight language model can interpret grid-structured climate data from the ClimRR dataset when the data are presented as structured JSON prompts, and whether fine-tuning on a small set of examples changes the outcome. The authors report that a LLaMA 3.1 8B model guided by structured prompts alone reaches a semantic similarity of 0.8335 but an accuracy score of only 0.2889, with frequent mistakes in numeric values, physical units, and emissions-scenario selection. After LoRA fine-tuning on roughly 100 user–input–assistant examples, the same model reaches 0.8954 similarity and a 1.0 accuracy score on the 10% test subset. The paper takes this as evidence that fine-tuning is essential for high-precision, domain-aligned answers on grid-structured geospatial data, and that an 8B model fine-tuned on about a hundred examples can be a practical component of lightweight agentic workflows for weather resilience planning. If correct, the result matters because it shows that small models can become reliable readers of structured scientific data without scaling up to much larger systems.

What carries the argument

The load-bearing mechanism is a curated dataset of roughly 120 structured JSON records built from ClimRR, a WRF-generated dataset of U.S. climate projections under RCP 4.5 and 8.5. Each record pairs a grid-cell identifier such as R073C493 with atmospheric variables for historical (1971–2000), mid-century (2041–2070), and end-century (2071–2100) periods, along with regional aggregates and a template-generated reference answer that GPT-4 later rephrased to add linguistic variety. These records are converted into user–input–assistant examples; about 100 are used for training and a 10% slice for testing. Fine-tuning applies Low-Rank Adaptation (LoRA) to the LLaMA 3.1 8B model with 8-bit quantization and bfloat16 mixed precision, and evaluation uses cosine similarity of sentence embeddings (all-MiniLM-L6-v2) plus a regex-based scorer that credits exact, partial, or no matches for grid cell, variable, units, RCP scenario, and values. This dataset and the two metrics together carry the comparison that yields the paper's central claim.

What would settle it

Build a held-out set of several hundred ClimRR queries for grid cells and time windows never used in training, score the fine-tuned model with the paper's regex protocol, and compare against a prompt-only baseline given the same structured inputs; the claim that fine-tuning is essential for high-precision answers is falsified if accuracy does not stay near 1.0 or if prompting alone matches it.

Watch

Extended reading notes

Core claim

In the paper's own terms, the central discovery is that fine-tuning an 8B LLaMA 3.1 model with LoRA on roughly 100 structured ClimRR question–answer pairs produces precise and consistent extraction of grid-cell values, physical units, and RCP 4.5/8.5 scenario references, raising the accuracy score from 0.2889 to 1.0 and cosine similarity from 0.8335 to 0.8954 relative to reference answers. The prompt-only base model could retrieve basic values and compare scenarios but made rounding or interpolation errors, ignored units, and fell back to an unspecified default scenario when the query did not name one. The fine-tuned model, by contrast, aligned with reference responses across value, unit, and scenario components, which the authors illustrate with side-by-side response comparisons and interpret as demonstrating that structured prompting alone is insufficient for scientific precision while fine-tuning yields domain-aligned answers.

Load-bearing premise

The whole result rests on the 10% test subset (about 12 examples) being representative of the task and truly independent of the roughly 100 training examples; if those test examples overlap with training examples in grid cells, templates, or quoted values, the perfect accuracy score is not a valid measure of generalization.

Editorial extensions

If this is right

A small 8B open-weight model fine-tuned on roughly a hundred examples can serve as a reliable interpreter of structured ClimRR queries, making lightweight agentic deployment for weather and hazard response feasible.
Structured prompting alone, at least for this data representation, should not be trusted for high-precision extraction; tasks that require exact values, units, and scenario labels should budget for fine-tuning.
The dataset-and-evaluation recipe can be reused to adapt LLMs to other gridded scientific datasets without retraining on raw multidimensional arrays.
The perfect accuracy on the test subset implies that, in an agentic workflow, the bottleneck may shift from value extraction to the external systems that select and supply the grid context.
The reported gains motivate expanding the dataset to more grid cells and query types so the approach can be stress-tested beyond its current scale.

Reading between the lines

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

If the fine-tuned model's accuracy survives an independently generated set of several hundred ClimRR queries for unseen grid cells, the same recipe could generalize to other structured scientific data, turning fine-tuning into a standard adapter step for data-to-text tasks.
Because the test queries come from the same templating pipeline that produced training queries, an evaluation with naturally typed, free-form user questions would separate genuine reasoning from pattern replication.
The paper does not report confidence or calibration, so a natural next experiment is to check whether the model's certainty tracks correctness on out-of-distribution grid cells, which matters for hazard alerts.
A direct comparison against a prompt-only baseline given explicit instructions to always list both RCP scenarios and include units would clarify whether the observed gap comes from missing knowledge or from prompting that does not enforce it.

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

Quick take: this is a small, honest, applied paper that makes a conventional point—fine-tuning beats prompting—on a new data format. The evidence, however, is too thin to support the strength of the claim. The paper introduces a ~120-example QA dataset over ClimRR grid cells and compares LLaMA 3.1 8B with structured prompting against a LoRA fine-tuned version. The direction is plausible, the authors are transparent about their setup, and the GitHub repo with training scripts and data is a genuine asset.

What is actually new: the dataset itself, and the direct head-to-head comparison on this particular gridded-data task. The qualitative finding that fine-tuning improves unit handling and scenario awareness is not surprising, but it's useful confirmation for practitioners building low-latency climate agents with small open-weight models.

The soft spot is the evaluation. The test set is 10% of ~120 examples, so about 12 items. Fine-tuned accuracy is 1.0; base accuracy is 0.2889. Those numbers have no error bars, no repeated runs, and no stratification. A 12/12 success rate only puts the true accuracy above roughly 0.74 at 95% confidence. More importantly, the dataset is generated by the same template pipeline for both training and test, with GPT-4 paraphrasing as the main diversity mechanism. There is no leakage analysis, so the perfect score could reflect near-duplicate templates rather than generalization. The regex-based accuracy script is also not shown, which is minor but easy to fix. The comparison itself is not circular—the method comparison is legitimate—but the conclusion "fine-tuning was essential" outruns the evidence. Presenting this as a pilot study with a call for larger evaluation would be more honest.

Who gets value: anyone working on LLMs for geospatial or climate data, and anyone interested in evaluation pitfalls with tiny test sets. It deserves a serious referee because it ships a new benchmark and code, but the referee should demand a larger, stratified held-out set and a leakage check before accepting the central claim.

Referee Report

4 major / 4 minor

Summary. The paper compares zero-shot structured prompting with LoRA fine-tuning of a LLaMA 3.1 8B model on a self-created dataset of ClimRR grid-cell queries. Using cosine similarity against a reference answer and a regex-based accuracy score, it reports that fine-tuning raises accuracy from 0.2889 to 1.0 on a 10% held-out subset and concludes that fine-tuning is essential for high-precision, domain-aligned answers.

Significance. If the conclusion holds, the paper offers a practical recipe for adapting small open-source LLMs to grid-structured climate data, with a public code repository and a clear task decomposition. However, the empirical support is currently too thin to establish the claim: the test set is roughly 12 examples, there is no leakage analysis, no repeated runs, and no alternative baselines. These are not fatal to the research direction, but they make the central comparative claim premature.

major comments (4)

[Section 3, Table 1] The fine-tuned model's reported accuracy of 1.0 in Table 1 is computed on the 10% test subset described in Section 3, i.e., roughly 12 examples of the ~120-example dataset. Even under ideal independence, 12/12 correct yields a one-sided 95% lower confidence bound of only about 0.74, so the claim that fine-tuning yields 'high-precision' answers is not statistically supported. Report error bars (e.g., binomial CIs), use a larger held-out set, and provide repeated runs.
[Section 2] The dataset is generated by a semi-automated template pipeline with GPT-4 paraphrasing (Section 2). Without a leakage analysis showing that the ~12 test records do not share grid cell, variable, period, and RCP combination with the ~100 fine-tuning examples, the 1.0 accuracy may reflect memorization of near-duplicate templates rather than generalization. Report overlap statistics and ideally split by grid cell or query template.
[Section 5] The accuracy score is computed by a 'Python script that used regex' (Section 5), but neither the script nor the exact scoring rubric is included, making reproducibility impossible and leaving room for subjective assignment of partial credits. Provide the script or regex patterns and report component-level accuracy (value, units, RCP, grid cell) in addition to the composite score.
[Section 5, Abstract] The central conclusion that 'fine-tuning was essential' is based on a single model (LLaMA 3.1 8B) evaluated on a single self-contained dataset with no alternative baselines. The comparison does not rule out that similar gains could be obtained by few-shot prompting of the base model or by a different base model. Add at least one prompting baseline (e.g., few-shot) and ideally another base model to support the comparative claim.

minor comments (4)

[Section 3] Report the exact number of examples in the test subset and the splitting procedure (e.g., random, stratified by grid cell or task type).
[Section 5] The sentence 'A comparison of select results in shown in Figure 2' contains a typo; it should be 'is shown'.
[Section 4] Clarify the relationship between the 'approximately 100' fine-tuning examples and the 'approximately 120' total dataset examples, and state explicitly whether the test subset was held out from the fine-tuning set.
[Figure 2] Since the highlights are described as red, ensure the figure is readable in grayscale or use additional markers for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the comparison is a standard held-out evaluation; limitations concern test size and leakage, not circular reasoning.

full rationale

The paper's central claim is an empirical comparison between a base LLM and a fine-tuned LLM on a held-out test subset. Section 3 states that model performance was evaluated on a test subset comprising 10% of the dataset, while Section 4 describes fine-tuning on approximately 100 user-assistant examples from the same curated dataset. This is a conventional train/test split, not a construction in which the test outcome is defined by the training input. No equation in the paper defines the reported similarity or accuracy scores as algebraic consequences of the fine-tuning examples, and no fitted parameter is renamed as a prediction. The self-citations to WildfireGPT (Xie et al. 2025a,b) and to the authors' own repository are contextual rather than load-bearing; they do not supply a uniqueness theorem or an ansatz that forces the results. The in-distribution nature of the evaluation, the small test set (~12 examples), and the lack of an explicit leakage analysis are legitimate correctness and generalization concerns, but they are not circularity under the definitions used here. Because the paper does not reduce its conclusions to its own inputs by construction or by self-citation chain, no circular step can be quoted with a specific reduction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No fitted parameters or invented entities underpin a derivation; the paper is an empirical comparison. The main unstated support comes from assuming the self-built dataset and the small test split are reliable evaluation instruments.

free parameters (4)

LoRA rank = 8
Hand-selected rank for parameter-efficient fine-tuning; affects model capacity but is not fitted to the evaluation metric.
LoRA scaling factor = 16
Chosen together with rank; influences the magnitude of fine-tuning updates.
Effective batch size = 8
Small batches with gradient accumulation; a training choice not fit to data.
Maximum input tokens = 2048
Context truncation limit for inputs; required by the model but arbitrary.

assumptions (4)

domain assumption ClimRR values used to build reference answers are correct ground truth.
Section 2 says the dataset was generated by querying ClimRR, but the paper does not independently validate the extracted values.
domain assumption The 10% test subset is representative and free from leakage with the fine-tuning set.
Sections 3 and 5 define a 10% test subset but report no stratification or overlap analysis.
domain assumption Regex-based component scoring with equal weights measures response correctness.
Section 5 describes the script but provides no validation against human judgment or alternative metrics.
domain assumption Cosine similarity from all-MiniLM-L6-v2 reflects semantic quality in this domain.
Section 5 uses sentence embeddings, but no evidence links higher similarity to better geospatial reasoning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Comparative Evaluation of Prompting and Fine-Tuning for Applying Large Language Models to Grid-Structured Geospatial Data." pith.science (2026). https://pith.science/paper/AHXHNF23

@misc{pith2026250517116,
  author       = {Pith},
  title        = {Pith review of: Comparative Evaluation of Prompting and Fine-Tuning for Applying Large Language Models to Grid-Structured Geospatial Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHXHNF23}},
  note         = {Machine review of arXiv:2505.17116}
}

read the original abstract

This paper presents a comparative study of large language models (LLMs) in interpreting grid-structured geospatial data. We evaluate the performance of a base model through structured prompting and contrast it with a fine-tuned variant trained on a dataset of user-assistant interactions. Our results highlight the strengths and limitations of zero-shot prompting and demonstrate the benefits of fine-tuning for structured geospatial and temporal reasoning.

Figures

Figures reproduced from arXiv: 2505.17116 by the authors.

**Figure 1.** Schematic showing the mapping of gridded ClimRR data over United States. Each grid cell is assigned an alphanumeric tag (e.g., R073C493) and contains atmospheric variable values in tabular form. These values can be transformed into a user–input–assistant format suitable for prompting and fine-tuning the language model. pretability, capabilities essential for real-world applications in weather resilience and adaptati… view at source ↗

**Figure 2.** Comparison of reference outputs with responses from the base and fine-tuned models to user queries, highlighting differences (shown in red) in accuracy and reasoning. These examples illustrate common ambiguities in base model responses to geospatial climate queries, including challenges in referencing correct RCP scenarios, handling measurement units, and making accurate regional comparisons. In contrast, fine-tune… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 3 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Daniel Han, M. H. and team, U. Unsloth, 2023. URL http://github.com/unslothai/unsloth

2023
[3]

8-bit optimizers via block-wise quantization, 2022

Dettmers, T., Lewis, M., Shleifer, S., and Zettlemoyer, L. 8-bit optimizers via block-wise quantization, 2022. URL https://arxiv.org/abs/2110.02861

arXiv 2022
[4]

and Dubey, A

Dhruv, A. and Dubey, A. Leveraging large language models for code translation and software development in scientific computing, 2025. URL https://arxiv.org/abs/2410.24119

arXiv 2025
[5]

The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes

Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J., Harrison, L., Hoell, A., and Michaelsen, J. The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes. Scientific Data, 2 0 (1): 0 150066, 2015. ISSN 2052-4463. doi:10.1038/sdata.2015.66. URL https://doi.or...

work page doi:10.1038/sdata.2015.66 2015
[6]

J., Todling, R., Molod, A., Takacs, L., Randles, C

Gelaro, R., McCarty, W., Suárez, M. J., Todling, R., Molod, A., Takacs, L., Randles, C. A., Darmenov, A., Bosilovich, M. G., Reichle, R., Wargan, K., Coy, L., Cullather, R., Draper, C., Akella, S., Buchard, V., Conaty, A., da Silva, A. M., Gu, W., Kim, G.-K., Koster, R., Lucchesi, R., Merkova, D., Nielsen, J. E., Partyka, G., Pawson, S., Putman, W., Riene...

2017
[7]

Geospatial reasoning: Unlocking insights with generative ai and multiple foundation models

Google Research . Geospatial reasoning: Unlocking insights with generative ai and multiple foundation models. https://research.google/blog/geospatial-reasoning-unlocking-insights-with-generative-ai-and-multiple-foundation-models/, 2025

work page 2025
[8]

The llama 3 herd of models

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

arXiv 2024

Show all 20 references

[9]

J., Hólm, E., Janisková, M., Keeley, S., Laloyaux, P., Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F., Villaume, S., and Thépaut, J.-N

Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G., Dahlgren, P....

1999
[10]

J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

2021 arXiv
[11]

and Hutter, F

Loshchilov, I. and Hutter, F. Sgdr: Stochastic gradient descent with warm restarts, 2017. URL https://arxiv.org/abs/1608.03983

2017 arXiv
[12]

and Hutter, F

Loshchilov, I. and Hutter, F. Decoupled weight decay regularization, 2019. URL https://arxiv.org/abs/1711.05101

2019 arXiv
[13]

Mixed precision training, 2018

Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., Ginsburg, B., Houston, M., Kuchaiev, O., Venkatesh, G., and Wu, H. Mixed precision training, 2018. URL https://arxiv.org/abs/1710.03740

2018 arXiv
[14]

Scaling neural machine translation, 2018

Ott, M., Edunov, S., Grangier, D., and Auli, M. Scaling neural machine translation, 2018. URL https://arxiv.org/abs/1806.00187

2018 arXiv
[15]

Llms for climate data analytics

Symufolk. Llms for climate data analytics. https://symufolk.com/llm-for-climate-data-analytics/, 2025

2025
[16]

Llama: Open and efficient foundation language models, 2023

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., and Lample, G. Llama: Open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971

2023 arXiv
[17]

N., Kaiser, L., and Polosukhin, I

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762

2023 arXiv
[18]

D., Hutchison, J

Xie, Y., Jiang, B., Mallick, T., Bergerson, J. D., Hutchison, J. K., Verner, D. R., Branham, J., Alexander, M. R., Ross, R. B., Feng, Y., Levy, L.-A., Su, W., and Taylor, C. J. Wildfiregpt: Tailored large language model for wildfire analysis, 2025 a . URL https://arxiv.org/abs...

2025 arXiv
[19]

D., Hutchison, J

Xie, Y., Jiang, B., Mallick, T., Bergerson, J. D., Hutchison, J. K., Verner, D. R., Branham, J., Alexander, M. R., Ross, R. B., Feng, Y., et al. A rag-based multi-agent llm system for natural hazard resilience and adaptation. arXiv preprint arXiv:2504.17200, 2025 b

2025 arXiv
[20]

Aphrodite: Constructing a long-term daily gridded precipitation dataset for asia based on a dense network of rain gauges

Yatagai, A., Kamiguchi, K., Arakawa, O., Hamada, A., Yasutomi, N., and Kitoh, A. Aphrodite: Constructing a long-term daily gridded precipitation dataset for asia based on a dense network of rain gauges. Bulletin of the American Meteorological Society, 93 0 (9): 0 1401--1415, 2...

2012 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Daniel Han, M. H. and team, U. Unsloth, 2023. URL http://github.com/unslothai/unsloth

2023

[3] [3]

8-bit optimizers via block-wise quantization, 2022

Dettmers, T., Lewis, M., Shleifer, S., and Zettlemoyer, L. 8-bit optimizers via block-wise quantization, 2022. URL https://arxiv.org/abs/2110.02861

arXiv 2022

[4] [4]

and Dubey, A

Dhruv, A. and Dubey, A. Leveraging large language models for code translation and software development in scientific computing, 2025. URL https://arxiv.org/abs/2410.24119

arXiv 2025

[5] [5]

The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes

Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J., Harrison, L., Hoell, A., and Michaelsen, J. The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes. Scientific Data, 2 0 (1): 0 150066, 2015. ISSN 2052-4463. doi:10.1038/sdata.2015.66. URL https://doi.or...

work page doi:10.1038/sdata.2015.66 2015

[6] [6]

J., Todling, R., Molod, A., Takacs, L., Randles, C

Gelaro, R., McCarty, W., Suárez, M. J., Todling, R., Molod, A., Takacs, L., Randles, C. A., Darmenov, A., Bosilovich, M. G., Reichle, R., Wargan, K., Coy, L., Cullather, R., Draper, C., Akella, S., Buchard, V., Conaty, A., da Silva, A. M., Gu, W., Kim, G.-K., Koster, R., Lucchesi, R., Merkova, D., Nielsen, J. E., Partyka, G., Pawson, S., Putman, W., Riene...

2017

[7] [7]

Geospatial reasoning: Unlocking insights with generative ai and multiple foundation models

Google Research . Geospatial reasoning: Unlocking insights with generative ai and multiple foundation models. https://research.google/blog/geospatial-reasoning-unlocking-insights-with-generative-ai-and-multiple-foundation-models/, 2025

work page 2025

[8] [8]

The llama 3 herd of models

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

arXiv 2024