Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Multilevel Analysis of Cryptocurrency News using RAG Approach with Fine-Tuned Mistral Large Language Model

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A fine-tuned Mistral 7B model, trained on GPT-4.1 instructions with 4-bit LoRA, claims that representing cryptocurrency news as a knowledge graph essentially eliminates LLM hallucinations, and demonstrates a multilevel pipeline that stacks

desk verdict A reproducible fine-tuning recipe for crypto news that overreaches: no retrieval, no grounding, no evaluation, and claims that contradict its own examples. read the letter →

arxiv 2509.03527 v1 pith:CS6A3NPR submitted 2025-08-25 cs.CL cs.AI

Bohdan M. Pavlyshenko This is my paper
classification cs.CLcs.AI
keywords cryptocurrencynewsknowledgegraphsentimentscoringMistral7BLoRAfine-tuningmultilevelsummarizationstackingsummariesanalytics
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 argues that a single small open-weight model—Mistral 7B, fine-tuned with 4-bit LoRA on instruction pairs produced by GPT-4.1—can carry a complete multilevel cryptocurrency news-analytics pipeline. At the first level the model converts an article into two complementary artifacts: a knowledge graph of entity–relationship–sentiment triples (sentiment on a -10..+10 scale) and a narrative text summary, each also emitted as JSON. At higher levels, the same model stacks batches of these summaries into consolidated reports that separate upward, downward, and contradictory trends, and finally merges the graph-level and text-level reports into a single verdict. The paper claims this graph representation essentially eliminates LLM hallucinations because facts are anchored as checkable triples, and that the sentiment scores are informative enough to serve as predictive features. The evidence shown is qualitative: printed example outputs across several July 2025 news items, not a scored benchmark.

What carries the argument

The load-bearing mechanism is the prompt-tuned multilevel summarization hierarchy executed by the fine-tuned Mistral 7B. First-level prompts such as "Generate a knowledge graph from cryptocurrency news" and "Generate summaries of cryptocurrency news and detect sentiment signals" force every article into a fixed set of entity types and relation types, closing with has_sentiment_signal on a -10..10 scale; a second prompt serializes the result to JSON. Higher-level prompts such as "Summarize the following list of cryptocurrency news summaries" and "Generate a single summary from the two provided summaries and output the result in JSON format" treat prior outputs as text input, so the same model

What would settle it

Annotate 50 recent crypto news articles with human-verified entity-relation-sentiment triples; into 10 of those articles, insert deliberately false but plausible claims such as "Bitcoin was acquired by Visa." Run the fine-tuned model and count how many graph triples reproduce the false claims or otherwise have no support in the source text. If any fabricated triple appears—especially with a confident sentiment score—the claim that knowledge-graph representation essentially eliminates hallucinations is falsified. A quantitative complement: compare the model's -10..10 sentiment scores against ma

Watch

Extended reading notes

Core claim

Stated at full strength in the abstract and conclusion, the paper's claim is that representing cryptocurrency news as a knowledge graph can essentially eliminate the hallucination problem in LLM-based news analysis, and that a fine-tuned Mistral 7B is enough to realize this. The model is trained with parameter-efficient fine-tuning (LoRA) on 4-bit quantized weights over a mixture of tasks: graph summary generation, text summary generation with sentiment detection, JSON serialization of both, list-level stacking of graph/text summaries, and a final stacking of the two stacked summaries into one JSON report. Outputs follow fixed schemas: graph summaries use triples of the form Entity | relatio

Load-bearing premise

The entire reliability chain rests on an unverified premise from Section 4.2: the GPT-4.1-generated instruction outputs used as training targets are accurate, hallucination-free ground truth. If the teacher outputs contain errors or style bias, the fine-tuned Mistral simply inherits them, since no human verification or retrieval step is described.

Editorial extensions

If this is right

  • Sentiment scores emitted as JSON can be dropped directly into regression or classification models for Bitcoin and altcoin price movement, as the paper states.
  • The same fine-tuned model can compress arbitrarily large news sets by hierarchical stacking: five summaries become one, and summaries of summaries continue upward, so a whole day's news collapses into a single readable verdict.
  • Because graph and text summaries are generated independently, disagreements between them become visible; the stack level is designed to produce an optimized score that weighs both views.
  • The pipeline runs on modest hardware thanks to a 4-bit LoRA adapter on a 7B model, which would let a newsroom or trading desk run this analytics locally instead of calling a proprietary API.
  • If the hallucination-control claim holds, the procedure becomes usable as an input layer for decision support, since each trend claim in the final report traces back to named entities and relations.

Reading between the lines

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

  • The paper never measures hallucination rates; the "essentially eliminate" claim is supported only by example outputs. Testing against human-annotated triples would show whether the gain comes from graph structure or from the fine-tuning itself.
  • The title's RAG label does not match the implementation: no retrieval happens—full news text is inserted directly into the prompt in Sections 5.1 and A.1. A direct extension would insert a genuine retrieval step over a news archive or knowledge base and compare graph quality with and without it.
  • Because teacher outputs from GPT-4.1 are unverified, the student model's ceiling is the teacher's accuracy; a human or rule-based consistency filter on training triples (e.g., checking each triple against the source sentence) would likely improve student reliability more than more training data.
  • In the examples, graph and text summaries genuinely disagree for the same entity (e.g., Circle at 8 in the graph but +6 in text; Solana at 0 in the graph but -2 in text). The stacking step could be turned into an uncertainty signal: disagreement between graph and text for the same entity marks the news as ambiguous, which is itself a useful feature for risk modeling.
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 a multilevel, multitask pipeline for cryptocurrency news analytics using a fine-tuned Mistral 7B model. At the first level, the model generates knowledge-graph summaries, text summaries, sentiment scores, and JSON representations of both. Higher levels perform hierarchical 'stacking' that consolidates lists of graph/text summaries into combined reports. The model is fine-tuned with 4-bit quantization and PEFT/LoRA on instruction pairs generated by GPT-4.1, and the authors test it on a handful of news articles from July 2025. The abstract and conclusion make strong claims: the approach is described as RAG-based, and graph representation is said to 'essentially eliminate' LLM hallucinations, with results showing 'informative qualitative and quantitative analytics.'

Significance. If the claims were validated, the contribution would be significant: a compact 7B open-weights model that produces structured, sentiment-scored graph/text summaries of crypto news at low cost, with a stacking mechanism for multi-article synthesis. The paper is strong on reproducibility artifacts: it provides the PEFT adapter on HuggingFace, training hyperparameters, prompt templates, and many full output examples, which is useful for replicating the pipeline. However, the scientific evidence is almost entirely missing. There is no quantitative evaluation, no baseline comparison, no human agreement study, and no external validation of sentiment scores. The RAG characterization is unsupported because no retrieval occurs, and the hallucination-elimination claim is contradicted by the paper's own examples and by its concluding caveat. As a peer-reviewed contribution, the current manuscript is closer to a technical report describing an untested prototype.

major comments (4)
  1. [Section 5.1, Appendix A.1] The title and abstract claim a 'retrieval-augmented generation (RAG)' approach, but no retrieval is implemented. In the test prompt and the code in Appendix A.1, the full news text is inserted directly into the prompt (`text=""" Cryptocurriency news text from 1Kb to 10Kb """`), and there is no query to a vector database, graph database, or any external knowledge source. The knowledge graph appears only as an output format, not as an input constraint or grounding mechanism. Hence the central 'RAG' contribution is not realized by the described system.
  2. [Abstract; Sections 4.2, 6] The assertion that representing news as a knowledge graph 'can essentially eliminate problems with large language model hallucinations' is not demonstrated. No factual-consistency metric, error analysis, or comparison to a non-graph baseline is provided. The paper's own outputs contradict the claim: Section 5.2.3's text summary says Circle was 'acquired by French media company Meteor,' while Section 5.2.1's graph lists Circle as 'involved_in IPO'; Solana receives sentiment 0 in the graph (Section 5.2.1) but -2 in the text summary (Section 5.2.3). Section 6 admits 'some inaccuracy can appear,' and Section 7 disclaims real-world use. The hallucination-elimination claim is therefore both unsupported and internally inconsistent with the reported examples.
  3. [Section 5 ('Testing')] The paper reports no evaluation whatsoever. Section 5 consists solely of hand-picked examples from an unspecified set of 'arbitrary' July 2025 news. There are no metrics, error bars, baselines, human-annotated gold standards, or statistical tests, and the sentiment scores are never validated against market data or human judgment. Consequently, the abstract's statement that the results 'demonstrate that the use of fine-tuned Mistral 7B LLM models ... can conduct informative qualitative and quantitative analytics' has no quantitative support. This is a load-bearing omission because the paper's main claimed contribution is the usable output of the analytics pipeline.
  4. [Section 4.2] The fine-tuning data are generated by GPT-4.1 prompts, with no described human verification or quality filtering. The subsequent 'results' are the author's inspection of outputs from a model trained to imitate GPT-4.1. This creates a closed validation loop: there is no external ground truth for the graph structures, summaries, or sentiment scores. Any quality estimate would therefore only reflect how well the student mimics the teacher, not how accurate the analytics are in an absolute sense. The paper should either provide external validation or substantially weaken the claims.
minor comments (5)
  1. [Section 4.1] Typo: 'Summarazing' should be 'Summarizing'. Similar typos appear throughout (e.g., 'crypotcurrency', 'cruptocurrencies', 'datasests').
  2. [Section 4.2] No details are given about the size of the GPT-4.1-generated training set, the number of prompts, prompt diversity, or any filtering/curation of the generated labels. This makes the fine-tuning step hard to reproduce or assess.
  3. [Section 5.2.1] The entity 'Crypto in general' appears as a catch-all aggregate but is not defined in the entity ontology of Section 4.2. If it is an intentional invention, it should be defined and its intended semantics stated.
  4. [Section 7] The disclaimer that the results are 'for academic purposes only and are not intended to inform real-world conclusions or recommendations' sits in tension with the abstract's claim that the pipeline provides 'important insights' and usable sentiment scores for predictive models. Please clarify the intended scope of the contribution.
  5. [Section 2] Reference [30] is listed twice in the sentence 'The use of knowledge graphs can reduce hallucinations in LLMs [9, 30, 30].' Also, the graph examples in Section 5.2 and the appendix are plain-text lists rather than actual graphs; visualizations would help the reader understand the proposed graph representation.

Circularity Check

2 steps flagged · score 6.0 of 10

Validation loop closes on itself: the model is fine-tuned on GPT-4.1 instruction outputs and those same kinds of outputs are then presented as evidence of informative analytics; the knowledge-graph hallucination claim is self-referential because the graph is generated by the LLM and never used as an external grounding input.

  1. fitted input called prediction [Section 4.2, Section 5.2, Conclusion]
    "We used the prompts for GPT-4.1 model to generate training datasets of instructions. ... The obtained results demonstrate that the use of fine-tuned Mistral 7B LLM models for cryptocurrency news analysis can give informative qualitative and quantitative analytics, providing important insights."

    The fine-tuned Mistral model is trained to reproduce GPT-4.1's instruction responses (graph summaries, text summaries, sentiment scores, stacking outputs). The paper's evidence that the approach is 'informative' is the author's qualitative inspection of outputs from that same model on test news (Section 5.2), with no external ground-truth labels, no human-annotation study, no factual-consistency metric, and no sentiment-calibration check. Since the model was optimized to imitate the teacher, exhibiting its outputs cannot independently demonstrate informativeness; the 'results' are by construction a sample from the teacher-imitating distribution. This is a fitted-input-called-prediction validation loop.

  2. self definitional [Abstract, Section 2, Section 5.1, Appendix A.1]
    "The representation of cryptocurrency news as knowledge graph can essentially eliminate problems with large language model hallucinations. ... text=""" Cryptocurriency news text from 1Kb to 10Kb """"

    The knowledge graph is not an external constraint; it is an output of the same fine-tuned LLM that is being trusted. In the testing prompt (Section 5.1) and Appendix A.1, the only input is the news text inserted directly into the prompt; no graph is retrieved, and no previously generated graph is used to ground or verify the response. Therefore the claimed hallucination-elimination mechanism is self-referential: the LLM's own structured output is presented as the ground truth that would eliminate the LLM's hallucinations. The graph cannot validate the LLM because it is produced by the LLM, so the claim holds only by definition, not by any implemented mechanism.

full rationale

The paper's central derivation is not a formal mathematical chain but an empirical pipeline. The clearest circularity is the teacher-student validation loop: training instructions are generated by GPT-4.1 (Section 4.2), the Mistral model is fine-tuned to imitate them, and the paper's evidence that the approach gives 'informative' analytics is the author's inspection of outputs from that same model (Section 5.2, Conclusion). No external benchmark, human evaluation, factual-consistency check, or sentiment-calibration test anchors the claimed results. This is a partial circularity: the outputs are a fitted imitation of the teacher, yet they are presented as independent evidence of quality. A second circularity affects the strongest claim: the abstract asserts that knowledge-graph representation 'can essentially eliminate' LLM hallucinations, but in the implementation the graph is generated by the LLM itself and never used as a retrieval source or verification constraint (Section 5.1, Appendix A.1). The graph thus cannot serve as an external ground truth; the claimed grounding is the model's own output, making the hallucination claim self-referential by construction. The paper also labels the pipeline 'RAG' although no retrieval step occurs, but that is an unsupported-labeling/correctness issue rather than a circularity. Self-citations ([1,2,3,13,33,34]) are background and are not load-bearing. The paper's own limitations ('In the test results, some inaccuracy can appear') and disclaimer (Section 7) further undercut the strong claims. Overall, the central validation loop closes on itself, so a score of 6 is warranted.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on an unverified training signal (GPT-4.1 outputs), a hand-built ontology, and an unmeasured assertion that graph structure removes hallucinations. No external ground truth is used anywhere in the pipeline.

free parameters (5)
  • Fine-tuning hyperparameters (learning rate, epochs, batch) = lr=3e-5, epochs=7, per_device_train_batch_size=2, gradient_accumulation_steps=4, fp16, 4-bit QLoRA
    Chosen by hand in Section 4 without ablations; directly shape what the adapter learns.
  • GPT-4.1-generated instruction training set = unspecified size and content; no samples shown
    Section 4.2 states training instructions were generated by GPT-4.1 with no described verification, size, coin coverage, or quality filter; all results inherit its biases.
  • Entity/relation ontology and sentiment scale = 17 entity types, 15 relation types, sentiment -10..10
    Hand-defined in Section 4.2; constrains the output space and hence the graph summaries.
  • Number of summaries per stacking prompt = 5
    Section 4.1: 'we used 5 arbitrary chosen summaries for one prompt'; an arbitrary choice affecting stacking outputs.
  • LoRA rank/alpha and target modules = not reported
    The paper gives no LoRA configuration (rank, alpha, target modules), a core free choice in PEFT/LoRA fine-tuning.
assumptions (5)
  • standard math LoRA/QLoRA fine-tuning preserves or improves base-model capability for instruction following
    Invoked in Section 4 as the basis for expecting the tuned Mistral to follow the trained tasks; relies on Hu et al. 2021 and Dettmers et al. 2023.
  • ad hoc to paper GPT-4.1-generated instruction/output pairs are an adequate ground truth for the analytics tasks
    Section 4.2: 'We used the prompts for GPT-4.1 model to generate training datasets of instructions.' No verification step is described; if labels contain errors or style bias, the student model inherits them.
  • domain assumption Knowledge-graph representation essentially eliminates LLM hallucinations
    Asserted in the abstract and Sections 2 and 4.2, citing surveys [9, 30]; the paper adds no measurement and its own examples show output inconsistencies.
  • domain assumption Sentiment scores from graph and text branches can be reconciled into an optimized score on the stacking level
    Section 3 and the conclusion claim stacking yields an 'optimized score'; no mechanism or validation is given, and Section 6 admits graph and text scores differ for the same news.
  • domain assumption The tested news texts (arbitrary July 2025 web articles) are representative of cryptocurrency news
    Section 5: 'we loaded arbitrary cryptocurrency news from the Web for July 2025'; no sampling protocol, so selection is uncontrolled.
invented entities (1)
  • 'Crypto in general' aggregate entity
    purpose: Catch-all bucket in graph and text summaries for market-wide statements not tied to a single coin
    The model emits this entity in nearly every example (5.2.1 through the appendix), but it is not in the entity-type list defined in Section 4.2. It is an output artifact with no external referent or verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multilevel Analysis of Cryptocurrency News using RAG Approach with Fine-Tuned Mistral Large Language Model." pith.science (2026). https://pith.science/paper/CS6A3NPR

@misc{pith2026250903527,
  author       = {Pith},
  title        = {Pith review of: Multilevel Analysis of Cryptocurrency News using RAG Approach with Fine-Tuned Mistral Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CS6A3NPR}},
  note         = {Machine review of arXiv:2509.03527}
}
read the original abstract

In the paper, we consider multilevel multitask analysis of cryptocurrency news using a fine-tuned Mistral 7B large language model with retrieval-augmented generation (RAG). On the first level of analytics, the fine-tuned model generates graph and text summaries with sentiment scores as well as JSON representations of summaries. Higher levels perform hierarchical stacking that consolidates sets of graph-based and text-based summaries as well as summaries of summaries into comprehensive reports. The combination of graph and text summaries provides complementary views of cryptocurrency news. The model is fine-tuned with 4-bit quantization using the PEFT/LoRA approach. The representation of cryptocurrency news as knowledge graph can essentially eliminate problems with large language model hallucinations. The obtained results demonstrate that the use of fine-tuned Mistral 7B LLM models for multilevel cryptocurrency news analysis can conduct informative qualitative and quantitative analytics, providing important insights.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    AI Approaches to Qualitative and Quantitative News Analytics on NATO Unity

    Bohdan M. Pavlyshenko. AI Approaches to Qualitative and Quantitative News Analytics on NATO Unity. arXiv preprint arXiv:2505.06313 , 2025. 14

    work page Pith review arXiv 2025

  2. [2]

    Financial News Analytics Using Fine-Tuned Llama 2 GPT Model

    Bohdan Pavlyshenko. Financial News Analytics Using Fine-Tuned Llama 2 GPT Model. arXiv preprint arXiv:2308.13032 , 2023

    work page Pith review arXiv 2023

  3. [3]

    Pavlyshenko

    Bohdan M. Pavlyshenko. Analysis of Disinformation and Fake News Detection Using Fine- Tuned Large Language Model. arXiv preprint arXiv:2309.04704 , 2023

    arXiv 2023

  4. [4]

    Forming Predictive Features of Tweets for Decision-Making Support

    Bohdan M. Pavlyshenko. Forming Predictive Features of Tweets for Decision-Making Sup- port. In Lecture Notes in Computational Intelligence and Decision Making: 2021 In- ternational Scientific Conference” Intellectual Systems of Decision-making and Problems of Computational Intelligence”, Proceedings , pages 479–490. Springer, Download PDF: https://arxiv.o...

    work page Pith review arXiv 2021

  5. [5]

    Methods of Informational Trends Analytics and Fake News Detection on Twitter

    Bohdan M. Pavlyshenko. Methods of informational trends analytics and fake news detection on twitter. arXiv preprint arXiv:2204.04891 , 2022

    work page Pith review arXiv 2022

  6. [6]

    FinGPT: Open-Source Fi- nancial Large Language Models

    Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. FinGPT: Open-Source Fi- nancial Large Language Models. FinLLM Symposium at IJCAI 2023, arXiv preprint arXiv:2306.06031, 2023

    arXiv 2023

  7. [7]

    ChatGPT-Informed Graph Neural Network for Stock Movement Prediction

    Zihan Chen, Lei Nico Zheng, Cheng Lu, Jialu Yuan, and Di Zhu. ChatGPT-Informed Graph Neural Network for Stock Movement Prediction. arXiv preprint arXiv:2306.03763 , 2023

    arXiv 2023

  8. [8]

    A survey on aug- menting knowledge graphs (KGs) with large language models (LLMs): models, evaluation metrics, benchmarks, and challenges

    Nourhan Ibrahim, Samar Aboulela, Ahmed Ibrahim, and Rasha Kashef. A survey on aug- menting knowledge graphs (KGs) with large language models (LLMs): models, evaluation metrics, benchmarks, and challenges. Discover Artificial Intelligence , 4(1):76, 2024

    work page 2024

Show all 46 references
  1. [9]

    Can knowledge graphs reduce hallucinations in LLMs?: A survey

    Garima Agrawal, Tharindu Kumarage, Zeyad Alghamdi, and Huan Liu. Can knowledge graphs reduce hallucinations in LLMs?: A survey. arXiv preprint arXiv:2311.07914 , 2023

    2023 arXiv

  2. [11]

    Relation extraction with fine-tuned large language models in retrieval augmented generation frameworks

    Sefika Efeoglu and Adrian Paschke. Relation extraction with fine-tuned large language models in retrieval augmented generation frameworks. arXiv preprint arXiv:2406.14745 , 2024

    2024 arXiv

  3. [12]

    Pavlyshenko

    B.M. Pavlyshenko. bpavlsh/Mistral-Fake-News-Detection: Fine-tuned Mistral-7B model for detecting and analyzing fake news, propaganda and offensive language in news articles. https://huggingface.co/bpavlsh/Mistral-Fake-News-Detection , 2025. Model card on Hugging Face

    2025

  4. [13]

    Pavlyshenko

    B.M. Pavlyshenko. bpavlsh/bart-crypto-summary: Fine-tuned Seq2Seq model is devel- oped for analysing and summarization of cryptocurrency news. https://huggingface. co/bpavlsh/bart-crypto-summary, 2024. Model card on Hugging Face

    2024

  5. [14]

    Clustering of authors’ texts of english fiction in the vector space of semantic fields

    Bohdan Pavlyshenko. Clustering of authors’ texts of english fiction in the vector space of semantic fields. Cybernetics and Information Technologies, 14(3):25–36, 2014

    2014

  6. [15]

    Pavlyshenko

    Bohdan M. Pavlyshenko. Bitcoin price predictive modeling using expert correction. In 2019 XIth International Scientific and Practical Conference on Electronics and Information Technologies (ELIT), pages 163–167. IEEE, 2019. 15

    2019

  7. [16]

    From Local to Global: A Graph RAG Approach to Query-Focused Summarization, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From Local to Global: A Graph RAG Approach to Query-Focused Summarization, 2024

    2024

  8. [17]

    Graph retrieval-augmented generation for large language models: A survey

    Tyler Thomas Procko and Omar Ochoa. Graph retrieval-augmented generation for large language models: A survey. In 2024 Conference on AI, Science, Engineering, and Technol- ogy (AIxSET), pages 166–169. IEEE, 2024

    2024

  9. [18]

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨ uttler, Mike Lewis, Wen tau Yih, Tim Rockt¨ aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Ne...

    2020

  10. [19]

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. Advances in Neural Information Processing Sys- tems, 37:132876–132907, 2024

    2024

  11. [20]

    Graph retrieval-augmented generation: A survey

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921, 2024

    2024 arXiv

  12. [21]

    Graph chain-of-thought: Augmenting large language models by reasoning on graphs

    Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103 , 2024

    2024 arXiv

  13. [22]

    A survey of graph retrieval-augmented generation for customized large language models

    Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, Junnan Dong, Hao Chen, Yi Chang, and Xiao Huang. A survey of graph retrieval-augmented generation for customized large language models. arXiv preprint arXiv:2501.13958, 2025

    2025

  14. [23]

    Lightrag: Simple and fast retrieval-augmented generation

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779 , 2024

    2024 arXiv

  15. [24]

    Retrieval-augmented generation with knowledge graphs for customer service question answering

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xi- aofeng Wang, and Zheng Li. Retrieval-augmented generation with knowledge graphs for customer service question answering. In Proceedings of the 47th international ACM SIGIR conference on research and...

    2024

  16. [25]

    Graph enhanced representation learning for news recommendation

    Suyu Ge, Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. Graph enhanced representation learning for news recommendation. In Proceedings of the web conference 2020, pages 2863–2869, 2020

    2020

  17. [26]

    A survey of large language models on generative graph analytics: Query, learning, and applications

    Wenbo Shang and Xin Huang. A survey of large language models on generative graph analytics: Query, learning, and applications. arXiv preprint arXiv:2404.14809 , 2024

    2024 arXiv

  18. [27]

    Advancing graph representation learning with large language models: A comprehensive survey of techniques

    Qiheng Mao, Zemin Liu, Chenghao Liu, Zhuo Li, and Jianling Sun. Advancing graph representation learning with large language models: A comprehensive survey of techniques. arXiv preprint arXiv:2402.05952 , 2024

    2024 arXiv

  19. [28]

    Large language models on graphs: A comprehensive survey

    Bowen Jin, Gang Liu, Chi Han, Meng Jiang, Heng Ji, and Jiawei Han. Large language models on graphs: A comprehensive survey. IEEE Transactions on Knowledge and Data Engineering, 2024. 16

    2024

  20. [29]

    Large language model enhanced knowledge representation learning: A survey

    Xin Wang, Zirui Chen, Haofen Wang, Leong Hou U, Zhao Li, and Wenbin Guo. Large language model enhanced knowledge representation learning: A survey. Data Science and Engineering, pages 1–24, 2025

    2025

  21. [30]

    Knowledge graphs, large language models, and hallucinations: An nlp perspective

    Ernests Lavrinovics, Russa Biswas, Johannes Bjerva, and Katja Hose. Knowledge graphs, large language models, and hallucinations: An nlp perspective. Journal of Web Semantics , 85:100844, 2025

    2025

  22. [31]

    HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models

    Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models. volume 37, pages 59532–59569, 2024

    2024

  23. [32]

    RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christo- pher D Manning. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. In The Twelfth International Conference on Learning Representations , 2024

    2024

  24. [33]

    Pavlyshenko

    Bohdan M. Pavlyshenko. Machine-learning models for sales time series forecasting. Data, 4(1):15, 2019

    2019

  25. [34]

    Pavlyshenko

    Bohdan M. Pavlyshenko. Using bayesian regression for stacking time series predictive mod- els. In 2020 IEEE Third International Conference on Data Stream Mining & Processing (DSMP), pages 305–309. IEEE, 2020

    2020

  26. [35]

    PEFT: State-of-the-art Parameter-Efficient Fine-Tuning methods

    Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. PEFT: State-of-the-art Parameter-Efficient Fine-Tuning methods. https://github.com/huggingface/peft, 2022

    2022

  27. [36]

    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. https://arxiv.org/abs/2106.09685, 2021

    2021 arXiv

  28. [37]

    StackLLaMA: A hands-on guide to train LLaMA with RLHF, 2023

    Edward Beeching, Younes Belkada, Kashif Rasul, Lewis Tunstall, Leandro von Werra, Nazneen Rajani, and Nathan Lambert. StackLLaMA: A hands-on guide to train LLaMA with RLHF, 2023

    2023

  29. [38]

    QLoRA: Efficient Fine-Tuning of Quantized LLMs

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient Fine-Tuning of Quantized LLMs. https://arxiv.org/abs/2305.14314, 2023

    2023 arXiv

  30. [39]

    8-bit Optimizers via Block-wise Quantization

    Tim Dettmers. 8-bit Optimizers via Block-wise Quantization. https://arxiv.org/abs/ 1909.05858, 2022

    1909 arXiv

  31. [40]

    QLoRA GitHub Repository

    Artidoro Pagnoni. QLoRA GitHub Repository. https://github.com/artidoro/qlora, 2023

    2023

  32. [41]

    BitsAndBytes GitHub Repository

    Tim Dettmers. BitsAndBytes GitHub Repository. https://github.com/TimDettmers/ bitsandbytes, 2023

    2023

  33. [42]

    Transformers GitHub Repository

    Hugging Face. Transformers GitHub Repository. https://github.com/huggingface/ transformers, 2023

    2023

  34. [43]

    Mistral-7B-v0.1 on Hugging Face

    Mistral AI. Mistral-7B-v0.1 on Hugging Face. https://huggingface.co/mistralai/ Mistral-7B-v0.1, 2023

    2023

  35. [44]

    Mixtral-8x7B-Instruct-v0.1 on Hugging Face

    Mistral AI. Mixtral-8x7B-Instruct-v0.1 on Hugging Face. https://huggingface.co/ mistralai/Mixtral-8x7B-Instruct-v0.1 , 2023. 17

    2023

  36. [45]

    Introducing Mistral 7B

    Mistral AI. Introducing Mistral 7B. https://mistral.ai/news/announcing-mistral-7b, 2023

    2023

  37. [46]

    TRL: Transformer Reinforcement Learning

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, and Shengyi Huang. TRL: Transformer Reinforcement Learning. https: //github.com/huggingface/trl, 2020

    2020

  38. [47]

    Huggingface access token

    B.M. Pavlyshenko. bpavlsh/Mistral-crypto-news: Fine-tuned Mistral-7B model for cryptocurrency news multilevel analytics. https://huggingface.co/bpavlsh/ Mistral-crypto-news, note = Model card on Hugging Face, 2025. 18 A Appendix A.1 Python Code for Testing Fine-Tuned Mistral 7...

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.