Pith. sign in

REVIEW 4 major objections 5 minor 30 references

ClimateChat: Designing Data and Methods for Instruction Tuning LLMs to Answer Climate Change Queries

T0 review · 4 major / 5 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 The paper shows that an automated pipeline for generating climate-change instruction data, combined with a geoscience-pretrained base model, lifts accuracy on climate QA from 86% to 93%.

desk verdict A controlled, modestly novel study whose central claim survives the source-overlap worry thanks to the human evaluation in Appendix A; the objective eval and missing artifacts are the real weak points. read the letter →

arxiv 2506.13796 v1 pith:2PAYR7LP submitted 2025-06-12 cs.CL cs.AI

Zhou Chen , Xiao Wang , Yuanhong Liao , Ming Lin , Yuqi Bai This is my paper · ORCID
classification cs.CLcs.AI
keywords instructiontuningclimatechangelargelanguagemodelsquestionansweringdomain-specificLLMdataconstructionretrieval-augmentedgenerationgeoscience
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 study proposes a fully automated way to build climate-change instruction data and shows that instruction-tuning with it improves an LLM's performance on climate question answering. The authors construct ClimateChat-Corpus with three strategies—generating questions and answers from reliable documents, scraping high-quality StackExchange QA pairs, and self-instructing from seed questions—and fine-tune the geoscience model JiuZhou with it. On objective multiple-choice questions, the resulting ClimateChat scores 93.0% against 86.0% for the same base model tuned on general instruction data only and 75.5% for Mistral-7B tuned on the same climate data. The paper argues the gain comes both from the specialized instruction data and from starting with a base model that already carries climate knowledge; tuning a less knowledgeable base model on the same data leads to more hallucinations. If right, this lowers the expert burden of building domain instruction sets and clarifies when instruction tuning helps versus hurts.

What carries the argument

The load-bearing object is ClimateChat-Corpus, an instruction dataset assembled by three automated strategies. Self-QA segments reliable documents (Wikipedia, IPCC reports, scientific papers) and uses GPT-4 to generate questions from each paragraph's facts and then answers; Web Scraping pulls highly recommended question-answer pairs about climate, earth science, and remote sensing from StackExchange; Self-Instruct starts from seed questions and uses GPT-4 in a few-shot loop to generate additional instructions, filtering new ones back into the task pool. The dataset is the independent variable that is removed in the JiuZhouChat ablation, and it carries the paper's argument that specialized instruction data improves climate QA. Training itself uses LoRA low-rank adaptation, and the base model is JiuZhou, a Mistral-7B model continued-pretrained on geoscience literature.

What would settle it

Give climate experts a set of newly written climate questions that are not drawn from Wikipedia or GeoBench, and have human raters (or a different LLM) grade open-ended responses. If ClimateChat's advantage over JiuZhouChat shrinks to near zero or reverses, then the reported improvement reflects distribution matching to the training/evaluation pipeline rather than better climate knowledge.

Watch

Extended reading notes

Core claim

The central claim is that climate-change-specific instruction data, produced automatically rather than handwritten by experts, enables an LLM to make better use of the climate knowledge it already has. The paper demonstrates this by comparing ClimateChat, which is JiuZhou instruction-tuned on ClimateChat-Corpus plus general data, with JiuZhouChat, which is the same base model tuned only on general data, and with ClimateMistral, which is Mistral-7B tuned on the same climate data. ClimateChat clearly outperforms both on objective and subjective climate QA, and the paper attributes the gains to two factors: the instruction data activates domain knowledge, and the domain-appropriate base model supplies that knowledge in the first place. It also reports that retrieval-augmented generation cannot replace instruction tuning on subjective tasks, though combining the two gives the best results, and that an instruction-tuned model can assist with literature retrieval and hypothesis formulation.

Load-bearing premise

The paper's measured gains assume that the evaluation does not systematically reward answers that resemble the training data: objective questions were drawn from Wikipedia, the same source used to generate the training QA pairs, with only 5-gram deduplication, and subjective scores were assigned by GPT-4, the very model that generated the training instructions.

Editorial extensions

If this is right

  • An instruction corpus built automatically from documents, web QA, and self-instructed seeds improves climate QA accuracy by about 7 percentage points over general instruction tuning alone.
  • Base model choice dominates the outcome: the same climate instruction data that helps JiuZhou produces unsupported, hallucinated figures when used to tune Mistral-7B.
  • RAG improves objective climate questions but does not close the subjective gap, so it cannot replace instruction tuning; instruction tuning plus RAG performs best.
  • The three-part data construction recipe can generate large, specialized instruction sets without expert manual writing, which should transfer to other scientific domains.
  • An instruction-tuned climate LLM can support scientific literature retrieval and hypothesis formulation, not just direct question answering.

Reading between the lines

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

  • Because the objective eval questions come from the same Wikipedia source as the Self-QA training data, part of ClimateChat's accuracy edge may be memorization of near-duplicate QA pairs rather than better reasoning; a held-out expert-written test set would separate these.
  • The subjective scores come from GPT-4, the same model that generated the training data, so the scores may partly reward stylistic similarity to GPT-4's own writing; the human evaluation in the appendix, which agrees, is a more trustworthy check.
  • The paper does not ablate the three data strategies against each other; a natural follow-up would measure the marginal contribution of Self-QA, Web Scraping, and Self-Instruct to the final accuracy.
  • If the base-model finding generalizes, then the optimal path to a domain LLM is continued pretraining on domain literature first and instruction tuning second—the instruction data alone cannot supply knowledge the base model lacks.
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 introduces ClimateChat-Corpus, an automatically constructed instruction dataset for climate change QA built from three strategies: Self-QA (GPT-4-generated questions and answers from Wikipedia, IPCC reports, and papers), Web Scraping (StackExchange QA pairs), and Self-Instruct (seed tasks expanded with GPT-4). The authors fine-tune the geoscience model JiuZhou (and Mistral-7B as a control) with LoRA, producing ClimateChat, and compare it against two baselines: JiuZhouChat (general instruction data only) and ClimateMistral (Mistral-7B with the same climate data as ClimateChat). Evaluation consists of 105 objective multiple-choice questions drawn from Wikipedia and 45 subjective questions from GeoBench scored by GPT-4, with a supplementary human evaluation in Appendix A. The central claim is that climate-specific instruction tuning improves LLMs' ability to answer climate queries using internal knowledge, and that the choice of base model matters. The paper also reports a RAG comparison showing that RAG plus instruction tuning outperforms either alone.

Significance. If the central claim holds, the paper provides a useful recipe for building domain-specific instruction data without heavy expert annotation, and the comparison of base models is a meaningful contribution. The design has notable strengths: the authors control for instruction data (JiuZhouChat), for base model (ClimateMistral), and they include a human evaluation that qualitatively supports the ranking ClimateChat > JiuZhouChat > ClimateMistral. The public model checkpoint is also a practical asset. However, the objective evaluation shares its source with the training data, the subjective scores come from the same model family that generated the training instructions, and no statistical significance testing is reported; these issues currently undermine the strength of the central claim and must be addressed before the results can be taken as conclusive.

major comments (4)
  1. [Section 3, first paragraph; Section 2.1 (Self-QA)] The objective evaluation set is collected from Wikipedia, which is the same source used to generate the Self-QA training data. The 5-gram deduplication described in Section 3 cannot remove semantic or paraphrase overlap: an eval question and a training paragraph can state the same fact with no shared 5-gram. Because ClimateChat was fine-tuned on GPT-4-generated answers derived from those very Wikipedia paragraphs, the observed ClimateChat-vs-JiuZhouChat gap on objective tasks may reflect memorization of the eval source rather than improved internal climate knowledge. The authors should construct a held-out objective set from a source not used in training, or at minimum release both the eval items and the ClimateChat-Corpus so that contamination can be checked, and report results stratified by topical overlap with training paragraphs.
  2. [Section 3, Table 2; Section 2.1] The subjective scores are assigned by GPT-4 using a six-criterion rubric, and GPT-4 is also the model used to generate the training instructions and answers in Self-QA and Self-Instruct. This creates a risk of evaluator bias favoring outputs that resemble GPT-4 style, independent of factual correctness. Although the human evaluation in Appendix A supports the qualitative ranking, it is reported only as aggregate counts across 45 questions, with no inter-annotator agreement or statistical test. The authors should either provide a human evaluation of the objective set, or report agreement between GPT-4-as-judge and human raters, and consider a rubric-based blind human evaluation for the subjective tasks.
  3. [Section 3, Table 1] The abstract and findings state that ClimateChat 'significantly improves' performance, but no statistical significance testing is reported. The meaning of the '±' values in Table 1 is not defined, and it is unclear whether they are standard deviations across runs, standard errors, or something else. Given that the objective set has only 18–26 questions per discipline, a difference of a few questions could drive the aggregate gap. The authors should specify the number of evaluation runs, define the error bars, and apply a per-question paired test (e.g., McNemar's test) to compare ClimateChat with JiuZhouChat and ClimateMistral.
  4. [Section 3, Findings; Appendix A] The claim that 'climate change-related instruction data can enhance the ability of LLMs to leverage internal knowledge' is inferred primarily from the objective-task comparison, whose contamination issue is described above. The subjective-task evidence, while supported by human evaluation, is limited to 45 GeoBench questions and a single case study; the Appendix A case analysis is anecdotal and does not quantify the error categories with uncertainty or inter-rater reliability. The authors should state explicitly which evaluation evidence (objective vs. subjective) supports each stated conclusion and provide per-item analysis for the 45 subjective questions.
minor comments (5)
  1. [Section 5 heading] The heading 'DISCCUSION' is misspelled and should be 'DISCUSSION'.
  2. [Appendix B heading] The appendix title refers to 'CHATCLIMATE' while the model is named ClimateChat; the naming should be made consistent throughout the manuscript.
  3. [Section 2.1, Web Scraping] The text says 'only the highly recommended answers' but then states that answers with fewer than three recommendations were excluded; the criterion for 'highly recommended' should be clarified.
  4. [Table 1 caption] The caption reports accuracy in percent but does not explain the '±' terms; defining these in the caption would improve interpretability.
  5. [Appendix C.2] The RAG implementation is described as using an embedding model and 'top five chunks', but the identity of the embedding model, the chunk size, and the similarity measure are not specified; these details are needed for reproducibility.

Circularity Check

2 steps flagged · score 4.0 of 10

Objective evaluation shares its Wikipedia source with the Self-QA training data, and GPT-4 both generates the training instructions and scores the subjective responses, so part of the reported gain reflects distribution matching; external human evaluation and GeoBench tasks keep the paper from being fully circular.

  1. other [Section 2.1 (Self-QA) and Section 3 (Objective Evaluation, Table 1)]
    "Initially, extensive materials relating to climate change and earth sciences were gathered, including Wikipedia pages, reports (e.g., The Sixth Assessment Report of the IPCC), and academic papers. We then segmented these texts into smaller paragraphs and used GPT-4 to generate questions based on the facts contained within these paragraphs. ... A set of questions and answers was collected from climate change-related disciplines on Wikipedia, and the answer options were expanded using GPT-4 to construct the objective tasks. We used 5-grams to remove data related to the ClimateChat-Corpus."

    The objective evaluation set is drawn from Wikipedia, the same source used to construct the Self-QA portion of ClimateChat-Corpus, and the only decontamination is 5-gram removal. Exact n-gram deduplication does not remove paraphrased or topically overlapping items: an evaluation question asking for a fact contained in a training paragraph may share no 5-gram with that paragraph. ClimateChat was fine-tuned on GPT-4-generated answers derived from those Wikipedia paragraphs, whereas JiuZhouChat was not. A higher objective score for ClimateChat can therefore be explained by memorization of the evaluation source rather than by the paper's conclusion that instruction data enhances the model's ability to leverage internal climate knowledge.

  2. other [Section 2.1 (Self-QA and Self-Instruct) and Section 3 (Subjective Evaluation, Table 2)]
    "We then segmented these texts into smaller paragraphs and used GPT-4 to generate questions based on the facts contained within these paragraphs. Finally, we input both the questions and their corresponding paragraphs into GPT-4 to generate the answers. ... Responses were evaluated using GPT-4 based on six criteria: helpfulness, relevance, accuracy, depth, creativity, and detail, with each criterion scored on a scale from 1 to 3."

    The same model, GPT-4, is used both to generate the ClimateChat-Corpus instruction data and to score the subjective responses in Table 2. Because ClimateChat was instruction-tuned on GPT-4-generated instructions and answers, its output style and content are likely to resemble GPT-4's own preferences more closely than JiuZhouChat's. A GPT-4 evaluator may therefore systematically favor ClimateChat for stylistic and distributional reasons, not solely for better climate knowledge. This is not a formal derivation, but it makes the subjective-score gap partly a reflection of training-data and evaluator overlap. The human evaluation in Appendix A provides some external grounding, which is why this issue contributes moderate, not total, circularity.

full rationale

The paper's central empirical claim is that ClimateChat-Corpus instruction tuning improves climate-change QA performance. The strongest evidence for this claim is the ClimateChat-versus-JiuZhouChat comparison in Section 3. That comparison has two overlapping-input problems: the objective questions come from Wikipedia, the same source used for Self-QA training data, with only 5-gram deduplication; and the subjective scores are assigned by GPT-4, the same model that generated the training instructions and answers. Both issues mean that a non-trivial part of the measured improvement could reflect distribution matching rather than independent verification of enhanced internal climate knowledge. However, the paper is not wholly circular: the GeoBench subjective tasks are an external benchmark, and Appendix A reports human evaluation counts in which ClimateChat also performs best (34 accurate responses vs. 29 for JiuZhouChat and 6 for ClimateMistral). The RAG comparison in Appendix C also provides a somewhat independent behavioral check. The self-citation of the JiuZhou base model is not load-bearing in a circular sense because the base-model comparison is an ablation, not an imported theorem. On balance, the contamination and evaluator overlap are real but partial, so the circularity score is moderate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No physical constants or fitted equations are involved. The assumptions concern data quality, evaluation validity, and base model knowledge, all of which are load-bearing for the empirical claim.

free parameters (3)
  • StackExchange answer recommendation threshold = 3 recommendations
    Answers with fewer than three recommendations were excluded from the corpus (Section 2.1, Web Scraping). This hand-set threshold affects data quality and quantity.
  • RAG chunk count k = 5
    The top five retrieved chunks were used in RAG experiments (Appendix C.2). This influences the RAG results but not the central instruction-tuning claim.
  • N-gram deduplication length = 5
    5-grams were used to remove evaluation items related to ClimateChat-Corpus (Section 3). The choice affects how much overlap remains between training and evaluation data.
assumptions (4)
  • domain assumption GPT-4-generated instructions from documents are accurate enough, and LLM filtering ensures trustworthiness.
    This is central to the data construction pipeline. No human validation of the corpus is reported. Stated in Sections 2.1 and 5.
  • domain assumption GPT-4 evaluates subjective answer quality reliably on six criteria.
    Subjective results in Table 2 rely on GPT-4 scoring, and the model outputs may be favored if they match GPT-4 style. Appendix A uses a different categorical human evaluation, not the same rubric.
  • domain assumption Objective evaluation questions from Wikipedia are a valid measure of climate QA ability.
    The eval set is drawn from Wikipedia, the same source as Self-QA training data, and only n-gram dedup separates them (Section 3). This independence is assumed rather than demonstrated.
  • domain assumption JiuZhou contains more climate and geoscience knowledge than Mistral-7B due to continued pretraining.
    Used to explain why ClimateChat outperforms ClimateMistral (Section 3). It is asserted based on the prior JiuZhou paper, not independently measured in this study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ClimateChat: Designing Data and Methods for Instruction Tuning LLMs to Answer Climate Change Queries." pith.science (2026). https://pith.science/paper/2PAYR7LP

@misc{pith2026250613796,
  author       = {Pith},
  title        = {Pith review of: ClimateChat: Designing Data and Methods for Instruction Tuning LLMs to Answer Climate Change Queries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PAYR7LP}},
  note         = {Machine review of arXiv:2506.13796}
}
read the original abstract

As the issue of global climate change becomes increasingly severe, the demand for research in climate science continues to grow. Natural language processing technologies, represented by Large Language Models (LLMs), have been widely applied to climate change-specific research, providing essential information support for decision-makers and the public. Some studies have improved model performance on relevant tasks by constructing climate change-related instruction data and instruction-tuning LLMs. However, current research remains inadequate in efficiently producing large volumes of high-precision instruction data for climate change, which limits further development of climate change LLMs. This study introduces an automated method for constructing instruction data. The method generates instructions using facts and background knowledge from documents and enhances the diversity of the instruction data through web scraping and the collection of seed instructions. Using this method, we constructed a climate change instruction dataset, named ClimateChat-Corpus, which was used to fine-tune open-source LLMs, resulting in an LLM named ClimateChat. Evaluation results show that ClimateChat significantly improves performance on climate change question-and-answer tasks. Additionally, we evaluated the impact of different base models and instruction data on LLM performance and demonstrated its capability to adapt to a wide range of climate change scientific discovery tasks, emphasizing the importance of selecting an appropriate base model for instruction tuning. This research provides valuable references and empirical support for constructing climate change instruction data and training climate change-specific LLMs.

Figures

Figures reproduced from arXiv: 2506.13796 by the authors.

Figure 1
Figure 1. Construction of the ClimateChat-Corpus. Self-QA Initially, extensive materials relating to climate change and earth sciences were gathered, including Wikipedia pages, reports (e.g., The Sixth Assessment Report of the IPCC), and academic papers. We then segmented these texts into smaller paragraphs and used GPT-4 to generate questions based on the facts contained within these paragraphs. Finally, we input both the qu… view at source ↗
Figure 2
Figure 2. Responses of ClimateMistral, JiuZhouChat, and ClimateChat to the climate change ques [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Examples of ClimateChat facilitating scientific discovery for scientists. The left panel [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance of ClimateChat with and without the RAG method versus GPT-4 in answer [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 13 canonical work pages

  1. [1]

    Climate change as a global amplifier of human--wildlife conflict

    Briana Abrahms, Neil H Carter, TJ Clark-Wolf, Kaitlyn M Gaynor, Erik Johansson, Alex McInturff, Anna C Nisi, Kasim Rafiq, and Leigh West. Climate change as a global amplifier of human--wildlife conflict. Nature Climate Change, 13 0 (3): 0 224--234, 2023. doi:10.1038/s41558-023-01608-5

    work page doi:10.1038/s41558-023-01608-5 2023

  2. [2]

    Cheap talk in corporate climate commitments: The role of active institutional ownership, signaling, materiality, and sentiment

    Julia Anna Bingler, Mathias Kraus, Markus Leippold, and Nicolas Webersinke. Cheap talk in corporate climate commitments: The role of active institutional ownership, signaling, materiality, and sentiment. Technical report, Swiss Finance Institute, 2022. URL https://EconPapers.repec.org/RePEc:chf:rpseri:rp2201

    work page 2022

  3. [3]

    Detecting deception using natural language processing and machine learning in datasets on covid-19 and climate change

    Barbara Brzic, Ivica Boticki, and Marina Bagic Babac. Detecting deception using natural language processing and machine learning in datasets on covid-19 and climate change. Algorithms, 16 0 (5): 0 221, 2023. doi:10.1016/j.nlp.2024.100057

    arXiv 2023

  4. [4]

    Machine-learning-based evidence and attribution mapping of 100,000 climate impact studies

    Max Callaghan, Carl-Friedrich Schleussner, Shruti Nath, Quentin Lejeune, Thomas R Knutson, Markus Reichstein, Gerrit Hansen, Emily Theokritoff, Marina Andrijevic, Robert J Brecha, et al. Machine-learning-based evidence and attribution mapping of 100,000 climate impact studies. Nature climate change, 11 0 (11): 0 966--972, 2021. doi:10.1038/s41558-021-01168-6

    work page doi:10.1038/s41558-021-01168-6 2021

  5. [5]

    Preparedllm: Effective pre-pretraining framework for domain-specific large language models

    Zhou Chen, Ming Lin, Zimeng Wang, Mingrun Zang, and Yuqi Bai. Preparedllm: Effective pre-pretraining framework for domain-specific large language models. Big Earth Data, 8 0 (4): 0 649--672, 2024. doi:10.1080/20964471.2024.2396159

    arXiv 2024

  6. [6]

    Jiuzhou: Open foundation language models and effective pre-training framework for geoscience

    Zhou Chen, Ming Lin, Mingrun Zang, Zimeng Wang, and Yuqi Bai. Jiuzhou: Open foundation language models and effective pre-training framework for geoscience. International Journal of Digital Earth, 18 0 (1): 0 1--30, 2025. doi:10.1080/17538947.2025.2449708

    arXiv 2025

  7. [7]

    Cody, Andrew J

    Emily M. Cody, Andrew J. Reagan, Lewis Mitchell, Peter Sheridan Dodds, and Christopher M. Danforth. Climate change sentiment on twitter: An unsolicited public opinion poll. PLoS ONE, 10, 2015. URL https://api.semanticscholar.org/CorpusID:872939

    work page 2015

  8. [8]

    Free dolly: Introducing the world's first truly open instruction-tuned llm, 2023

    Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. Free dolly: Introducing the world's first truly open instruction-tuned llm, 2023. URL https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm

    2023

Show all 30 references
  1. [9]

    K2: A foundation language model for geoscience knowledge understanding and utilization

    Cheng Deng, Tianhang Zhang, Zhongmou He, Qiyuan Chen, Yuanyuan Shi, Yi Xu, Luoyi Fu, Weinan Zhang, Xinbing Wang, Chenghu Zhou, Zhouhan Lin, and Junxian He. K2: A foundation language model for geoscience knowledge understanding and utilization. WSDM '24, pp.\ 161–170, New York,...

    2024

  2. [10]

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

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pp.\ 1--16, 2019. doi:10.18653/v1/N19-1423

    2019 doi

  3. [11]

    Does fine-tuning llms on new knowledge encourage hallucinations?, 2024

    Zorik Gekhman, Gal Yona, Roee Aharoni, Matan Eyal, Amir Feder, Roi Reichart, and Jonathan Herzig. Does fine-tuning llms on new knowledge encourage hallucinations?, 2024. URL https://arxiv.org/abs/2405.05904

    2024 arXiv

  4. [12]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Eleventh International Conference on Learning Representations, pp.\ 1--26, 2022. doi:10.48550/arXiv.2106.09685

    2022 doi arXiv

  5. [13]

    Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases, 2023

    Yunjie Ji, Yong Deng, Yan Gong, Yiping Peng, Qiang Niu, Lei Zhang, Baochang Ma, and Xiangang Li. Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases, 2023. URL https://arxiv.org/abs/2303.14742

    2023 arXiv

  6. [14]

    \# instag: Instruction tagging for analyzing supervised fine-tuning of large language models

    Keming Lu, Hongyi Yuan, Zheng Yuan, Runji Lin, Junyang Lin, Chuanqi Tan, Chang Zhou, and Jingren Zhou. \# instag: Instruction tagging for analyzing supervised fine-tuning of large language models. In The Twelfth International Conference on Learning Representations, 2023. doi:1...

    2023 doi arXiv

  7. [15]

    Understanding the impact of climate change on critical infrastructure through nlp analysis of scientific literature

    Tanwi Mallick, Joshua David Bergerson, Duane R Verner, John K Hutchison, Leslie-Anne Levy, and Prasanna Balaprakash. Understanding the impact of climate change on critical infrastructure through nlp analysis of scientific literature. Sustainable and Resilient Infrastructure, p...

    2024

  8. [16]

    Analyzing regional impacts of climate change using natural language processing techniques

    Tanwi Mallick, John Murphy, Joshua David Bergerson, Duane R Verner, John K Hutchison, and Leslie-Anne Levy. Analyzing regional impacts of climate change using natural language processing techniques. arXiv preprint arXiv:2401.06817, 2024 b . doi:10.48550/arXiv.2401.06817

    2024 doi arXiv

  9. [17]

    A rabic mini- C limate GPT : A climate change and sustainability tailored A rabic LLM

    Sahal Mullappilly, Abdelrahman Shaker, Omkar Thawakar, Hisham Cholakkal, Rao Anwer, Salman Khan, and Fahad Khan. A rabic mini- C limate GPT : A climate change and sustainability tailored A rabic LLM . In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Associa...

    2023 doi

  10. [18]

    Instruction tuning with gpt-4, 2023

    Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4, 2023. URL https://arxiv.org/abs/2304.03277

    2023 arXiv

  11. [19]

    Climate bot: A machine reading comprehension system for climate change question answering

    Md Rashad Al Hasan Rony, Ying Zuo, Liubov Kovriguina, Roman Teucher, and Jens Lehmann. Climate bot: A machine reading comprehension system for climate change question answering. In IJCAI, pp.\ 5249--5252, 2022. doi:10.24963/ijcai.2022/729

    2022 doi

  12. [20]

    Harnessing the potential of nature-based solutions for mitigating and adapting to climate change

    Nathalie Seddon. Harnessing the potential of nature-based solutions for mitigating and adapting to climate change. Science, 376 0 (6600): 0 1410--1416, 2022. doi:10.1126/science.abn9668

    2022 doi

  13. [21]

    Analyzing the dynamics of climate change discourse on twitter: A new annotated corpus and multi-aspect classification

    Shuvam Shiwakoti, Surendrabikram Thapa, Kritesh Rauniyar, Akshyat Shah, Aashish Bhandari, and Usman Naseem. Analyzing the dynamics of climate change discourse on twitter: A new annotated corpus and multi-aspect classification. In Proceedings of the 2024 Joint International Con...

    2024

  14. [22]

    Environmental claim detection

    Dominik Stammbach, Nicolas Webersinke, Julia Bingler, Mathias Kraus, and Markus Leippold. Environmental claim detection. Available at SSRN 4207369, 2022. doi:10.18653/v1/2023.acl-short.91

    2022 doi

  15. [23]

    Climategpt: Towards ai synthesizing interdisciplinary research on climate change

    David Thulke, Yingbo Gao, Petrus Pelser, Rein Brune, Rricha Jalota, Floris Fok, Michael Ramos, Ian van Wyk, Abdallah Nasir, Hayden Goldstein, et al. Climategpt: Towards ai synthesizing interdisciplinary research on climate change. arXiv preprint arXiv:2401.09646, 2024. doi:10....

    2024 doi arXiv

  16. [24]

    Deep climate change: A dataset and adaptive domain pre-trained language models for climate change related tasks

    Saeid Vaghefi, Veruska Muccione, Christian Huggel, Hamed Khashehchi, and Markus Leippold. Deep climate change: A dataset and adaptive domain pre-trained language models for climate change related tasks. In NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning,...

    2022

  17. [25]

    Chatclimate: Grounding conversational ai in climate science

    Saeid Ashraf Vaghefi, Dominik Stammbach, Veruska Muccione, Julia Bingler, Jingwei Ni, Mathias Kraus, Simon Allen, Chiara Colesanti-Senni, Tobias Wekhof, Tobias Schimanski, et al. Chatclimate: Grounding conversational ai in climate science. Communications Earth & Environment, 4...

    2023 doi

  18. [26]

    Climatebert: A pretrained language model for climate-related text

    Nicolas Webersinke, Mathias Kraus, Julia Anna Bingler, and Markus Leippold. Climatebert: A pretrained language model for climate-related text. arXiv preprint arXiv:2110.12010, 2021. doi:10.48550/arXiv.2110.12010

    2021 doi arXiv

  19. [27]

    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 gl...

  20. [28]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  21. [29]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  22. [30]

    Tackling Climate Change with Machine Learning

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2025

Pith tools

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