Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

FinLoRA: Benchmarking LoRA Methods for Fine-Tuning LLMs on Financial Datasets

T0 review · 5 major / 6 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 claims that LoRA fine-tuning raises general-purpose LLM accuracy by 36% on average across 19 financial datasets, with vanilla LoRA on Llama 3.1 8B reaching 74.74 versus the base model's 37.05.

desk verdict The four new XBRL datasets are the real asset here; the headline 36% gain is arithmetic that doesn't hold up. read the letter →

arxiv 2505.19819 v1 pith:X6W245LL submitted 2025-05-26 cs.CE cs.AI

Dannong Wang , Jaisal Patel , Daochen Zha , Steve Y. Yang , Xiao-Yang Liu This is my paper
classification cs.CEcs.AI
keywords low-rankadaptationparameter-efficientfine-tuningfinancialLLMbenchmarkXBRLanalysisLoRAvariantsexactmatchevaluation
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 sets out to show that parameter-efficient low-rank adaptation can turn general-purpose LLMs into credible financial specialists without the multi-million-dollar cost of training from scratch. It introduces FinLoRA, a benchmark of nineteen financial datasets, four of them newly built from XBRL-formatted annual reports, and compares five LoRA variants across five base models. The headline result is a 36% average accuracy gain over untuned base models, with vanilla LoRA on Llama 3.1 8B reaching 74.74 average score versus 37.05 for the base model. The paper also reports costs of roughly 14 to 16 hours on four A5000 GPUs, about $15, compared with BloombergGPT's estimated $2.7 million. If the claim holds, specialized financial capability no longer requires training a large model from scratch.

What carries the argument

The load-bearing object is the low-rank adapter: instead of updating the frozen weights $W_0$, LoRA learns $A \in \mathbb{R}^{r \times k}$ and $B \in \mathbb{R}^{d \times r}$ so that the forward pass is $y = W_0x + \gamma_r BAx$, with vanilla LoRA using $\gamma_r = \alpha/r$ and rsLoRA using $\alpha/\sqrt{r}$. Around this mechanism, the paper builds a benchmark: 19 datasets, including four new XBRL analysis datasets constructed by applying five templates to facts from Dow Jones 30 companies' 10-K filings and filtering the relevant XBRL text segment as context. The adapters isolate what fine-tuning contributes, letting the authors compare accuracy, F1, BERTScore, fine-tuning time, GPU memory, inference latency, federated training, and out-of-domain forgetting.

What would settle it

Take a random sample of a few hundred questions from the four new XBRL datasets, have independent financial analysts write ground-truth answers without seeing the template, and compare against the template strings under the same parser. If analysts regularly produce semantically equivalent tags or formulas that the parser counts as wrong, the reported accuracies, including the Formula Calculation jump from 27.27 to 98.68, are inflated.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that LoRA methods produce substantial performance gains of 36% on average over base models across nineteen financial datasets, and that a vanilla LoRA adapter (8-bit, rank 8) on Llama 3.1 8B Instruct reaches an overall average score of 74.74, a 37.69% increase over the base model's 37.05. The largest gains appear in the new XBRL analysis tasks, where the base model's Formula Calculation accuracy of 27.27 rises to 98.68. The paper also finds that rsLoRA leads in financial analysis, reporting, and statement analysis categories, that DoRA underperforms when the magnitude vector shares the low-rank learning rate, and that federated LoRA improves over the base model while trailing centralized LoRA. No catastrophic forgetting is observed on MMLU and GSM8K at the tested ranks.

Load-bearing premise

The headline gains on the new XBRL tasks assume exact-match scoring is valid: each templated question has exactly one correct tag, value, or formula, and the output parser maps model responses to those answers without error.

Editorial extensions

If this is right

  • A LoRA-tuned 8B model can beat or match far larger base models and a proprietary fine-tuned Gemini model on most of the 19 financial tasks, so specialized financial capability no longer requires training a large model from scratch.
  • XBRL-based question answering is a particularly high-return target for fine-tuning: formula calculation accuracy jumps from 27.27 to 98.68 after LoRA adaptation.
  • Fine-tuning with LoRA on the reported setup costs about $15 and 14 to 16 hours on four A5000 GPUs, versus an estimated $2.7 million for BloombergGPT, making domain adaptation accessible to small organizations.
  • At ranks 4 and 8 with the tested learning-rate ratios, financial LoRA adapters leave MMLU accuracy unchanged and slightly improve GSM8K, suggesting domain fine-tuning did not erase general knowledge.
  • Federated LoRA with four nodes improves sentiment analysis over the base model but underperforms centralized LoRA, defining a privacy-accuracy tradeoff for collaborative financial training.

Reading between the lines

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

  • Because the new XBRL datasets are template-generated with exact-match answers, a natural check is whether LoRA generalizes to filings outside the Dow 30 or to years outside 2019-2023; if scores collapse there, the adapters may have learned template patterns rather than XBRL semantics.
  • The contrast between large XBRL gains and small FinanceBench gains suggests that input standardization, not task difficulty, drives LoRA's benefit; a testable extension would be fine-tuning on other highly structured document formats such as regulatory filings, legal contracts, or clinical trial records.
  • The cost comparison depends on the stated $0.26 per GPU-hour rate for A5000s; at higher cloud rates the qualitative conclusion would survive, but the dollar figures would not, so the benchmark's cost claims are best read as an order-of-magnitude argument.
  • The absence of catastrophic forgetting was measured only on MMLU and GSM8K zero-shot; broader safety and reasoning benchmarks would be needed before claiming financial adapters are generally non-destructive.
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

5 major / 6 minor

Summary. The paper introduces FinLoRA, an open benchmark for LoRA-style parameter-efficient fine-tuning of LLMs on financial tasks. It curates 19 datasets (15 public and 4 newly constructed XBRL analysis datasets derived from Dow Jones 30 company filings), evaluates five base models and five fine-tuning approaches (LoRA, QLoRA, DoRA, rsLoRA, and a Gemini 2.0 Flash Lite fine-tune), and reports accuracy, F1, BERTScore, fine-tuning/inference cost, federated-LoRA results, and catastrophic-forgetting checks. The headline claim is that LoRA fine-tuned models achieve a 36% average improvement over base models, with vanilla LoRA (8-bit, rank 8) reaching an overall average of 74.74 versus 37.05 for the Llama 3.1 8B base model.

Significance. If the results hold, the paper would be a useful community resource: it provides a broad financial benchmark, four new XBRL question-answering datasets, open adapters/code, and cost measurements showing that local LoRA fine-tuning is substantially cheaper than closed API fine-tuning. The strengths are the scale of the evaluation (19 datasets, 46 fine-tuning rounds, 194 evaluation rounds), the open release of datasets and adapters, and the direct comparison of multiple LoRA variants in a high-stakes domain. However, the central quantitative claim is currently not reproducible from the reported data, and several evaluation details need correction before the benchmark results can be trusted.

major comments (5)
  1. [Abstract and Section 4.1] The headline improvement statistics are arithmetically inconsistent. Section 4.1 states that vanilla LoRA's 74.74 is a '37.69% increase' over the Llama 3.1 8B base model's 37.05, but (74.74 - 37.05) / 37.05 is approximately 1.017, i.e., a 101.7% relative increase; 37.69 is the absolute percentage-point gain, not a percent increase. The abstract's '36% on average over base models' is also not derivable from Table 4 under any stated aggregation: using the four Llama fine-tuned models against the Llama base gives an increase of (73.10 - 37.05) / 37.05 ≈ 97.3%, while comparing the four tuned models to the average of all five base models (53.79) gives about 35.9%, a baseline that mixes 8B open models with GPT-4o and DeepSeek V3. Because the central claim is precisely a quantitative improvement, the paper must define the baseline and aggregation and correct all reported percentages.
  2. [Section 3.1 and Table 4] The 'Overall Average' row in Table 4 is never defined. No formula is given for aggregating across datasets that use different metrics (accuracy, F1, and BERTScore multiplied by 100) or across datasets of very different sizes; it is unclear whether each dataset is weighted equally, whether F1 and accuracy are averaged together, and how the BERTScore values (e.g., 0.574) are converted into the reported average. Without this definition, the ranking of LoRA variants and the headline numbers cannot be reproduced. In addition, all reported results appear to be single runs, with no error bars, variance, or significance testing; because fine-tuning is stochastic, differences of a few tenths of a point between LoRA and QLoRA may not be meaningful.
  3. [Section 3.1 and Table 4 (XBRL datasets)] The construction of the four new XBRL analysis datasets and the Exact Match scoring are underspecified. The pipeline uses five templates to generate questions from XBRL facts, but the paper does not report (i) whether multiple XBRL tags can satisfy a natural-language description, (ii) how the output parser maps model text to tags, values, or formulas, (iii) whether a formula is judged semantically or as a string, and (iv) whether train/test splits are made at the filing level to avoid template leakage. These details are load-bearing because the largest gains, such as Formula Calculation jumping from 27.27 to 98.68, rest entirely on Exact Match. If the template construction admits multiple valid tags, semantically equivalent formulas, or parser artifacts, the reported XBRL improvements would be inflated. Please add human-validation statistics, the exact parsing procedure, and the split procedure.
  4. [Section 4.4 and Table 6] The catastrophic-forgetting evaluation reports MMLU accuracy of 0.229 and GSM8K accuracy of 0.011 for the Llama 3.1 8B Instruct base model. These values are far below published results for this model (typically around 0.68 on MMLU and substantially higher on GSM8K), indicating a systematic issue in the evaluation harness, prompt format, or answer parsing. The conclusion that 'LoRA does not exhibit catastrophic forgetting' relies entirely on this table, so with an invalid baseline the comparison is uninformative. Please report the exact evaluation setup, the number of examples used, and verify the harness against known published numbers for the base model.
  5. [Section 3.1 and Table 2] The FNXL dataset is listed with no training split ('-' under #Train), yet Table 4 reports fine-tuned model performance on it. Please clarify whether FNXL was used only for zero-shot evaluation, and if so, whether the corresponding adapter was trained on other financial-reporting data (e.g., FiNER or XBRL Term) or evaluated directly without task-specific fine-tuning. This ambiguity affects the interpretation of the financial-reporting results and the reported count of fine-tuning rounds.
minor comments (6)
  1. [Section 5.1] The phrase 'The-50-billion parameter model' contains a typographical artifact; it should read 'the 50-billion-parameter model'.
  2. [Table 4 caption] The caption says that accuracy, F1, and BERTScore are shown in blue, gray, and green, but the text version does not show colors; please label the columns or add a legend so the table is readable in monochrome.
  3. [Section 4.1 (DoRA discussion)] The claim that DoRA would benefit from separate learning rates for the magnitude vector and the direction matrix is speculative, since no experiment with separate learning rates is reported; please either add such an experiment or clearly label this as a hypothesis.
  4. [Section 3.2 and Section 4.4] The 'LoRA with Federated Learning' method is listed among the five methods, but Section 4.4 only says that FedAvg was used with four nodes; please specify the number of communication rounds, local epochs, client partitioning, and whether all clients share the same LoRA hyperparameters.
  5. [Section 4.3 and Table 5] The estimated fine-tuning cost of $0.26 per GPU hour should be justified, since A5000 cloud pricing varies; the GPT-4o-mini cost estimate should also state the assumed number of tokens and epochs.
  6. [Section 2 and Table 1] The case study in Table 1 is illustrative but not quantitative; consider adding a note about how representative the two examples are of overall base-model errors.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central LoRA-versus-base comparison is an empirical benchmark measured on held-out test splits; existing self-citations are contextual and not load-bearing.

full rationale

FinLoRA's headline claim is an experimental measurement, not a derivation. Table 4 compares five base models and five fine-tuned models on 19 datasets, and the fine-tuned models are evaluated on test splits after training on separate training splits; no fitted parameter is renamed as a prediction, and no equation defines the reported accuracy in terms of the LoRA construction itself. The four new XBRL datasets are template-generated from SEC filings, which raises benchmark-validity questions about answer uniqueness, but template construction is not a circular reduction of the evaluation to the method's output. The paper cites prior work by overlapping authors (FinGPT [22,23] and XBRL Agent [10]) for datasets and context, but those citations do not carry the load-bearing claim that LoRA improves over base models; that claim is supported by the paper's own held-out numbers and by public external benchmarks such as FPB, FiQA, TFNS, NER, Headline, CFA, FiNER, FNXL, FinanceBench, Financial Math, MMLU, and GSM8K. The '36%' and '37.69%' figures are arithmetically mislabeled percentage-point gains rather than relative percent increases, and the abstract's 36% mixes different base-model families, but these are reporting or definitional errors, not circularity. No uniqueness theorem, ansatz-by-citation, or definitional equivalence is invoked, so no circular step can be substantiated.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The benchmark depends on hand-chosen hyperparameters and on the validity of the newly constructed XBRL datasets and their exact-match scoring. No new physical or mathematical entities are introduced. The main unverifiable inputs are the dataset construction pipeline and the evaluation harness, which together determine whether the reported gains are real.

free parameters (5)
  • LoRA rank r = 8 for most variants, 4 for QLoRA
    Hand-chosen rank; the paper notes rsLoRA benefits at higher ranks, so the comparative results are specific to these ranks and may not generalize.
  • LoRA scaling ratio alpha:r = 8:1 at rank 8, 4:1 at rank 4
    Standard defaults chosen by hand; no sweep is performed, and DoRA's performance is known to be sensitive to this ratio.
  • Learning rate = 1e-4
    A single learning rate is used for all methods; the authors state DoRA may need a separate, higher learning rate for its magnitude vector, so this choice may disadvantage DoRA.
  • Batch size = 2 to 8 depending on prompt length
    Chosen per dataset based on prompt length; batch size affects optimization and was not varied systematically.
  • Aggregate improvement percentage = Claimed 36% but not reproducible from the table
    The 36% average improvement is stated in the abstract and Section 1, but the paper does not define how it is computed. Table 4 overall averages suggest a much larger relative increase, so the aggregation recipe is an unstated, hand-chosen statistic.
assumptions (4)
  • domain assumption Each of the four new XBRL datasets has a unique ground-truth answer such that Exact Match is a valid evaluation metric.
    Section 3.1 describes template-based question generation on XBRL segments; if multiple tags or semantically equivalent formulas are correct, EM understates true performance and the reported gains are artifacts.
  • domain assumption The public datasets are used with their original labels and splits, and no data leakage occurs between the XBRL dataset construction and the training sets.
    Section 3.1 and Table 2 list train/test sizes but do not describe the split procedure for the new datasets; leakage across companies or years would inflate accuracy.
  • domain assumption The LoRA variants (LoRA, QLoRA, DoRA, rsLoRA) were implemented correctly and comparably using standard libraries.
    Section 3.2 gives mathematical definitions, but the paper does not provide implementation details or verify correctness against reference implementations.
  • domain assumption The evaluation harness, including greedy decoding and output parsing, does not introduce systematic bias favoring any LoRA variant.
    Section 4 specifies temperature 0.0 and basic decoding settings, but parsing of free-form model outputs into tags, values, and formulas is not validated on a sample.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FinLoRA: Benchmarking LoRA Methods for Fine-Tuning LLMs on Financial Datasets." pith.science (2026). https://pith.science/paper/X6W245LL

@misc{pith2026250519819,
  author       = {Pith},
  title        = {Pith review of: FinLoRA: Benchmarking LoRA Methods for Fine-Tuning LLMs on Financial Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X6W245LL}},
  note         = {Machine review of arXiv:2505.19819}
}
read the original abstract

Low-rank adaptation (LoRA) methods show great potential for scaling pre-trained general-purpose Large Language Models (LLMs) to hundreds or thousands of use scenarios. However, their efficacy in high-stakes domains like finance is rarely explored, e.g., passing CFA exams and analyzing SEC filings. In this paper, we present the open-source FinLoRA project that benchmarks LoRA methods on both general and highly professional financial tasks. First, we curated 19 datasets covering diverse financial applications; in particular, we created four novel XBRL analysis datasets based on 150 SEC filings. Second, we evaluated five LoRA methods and five base LLMs. Finally, we provide extensive experimental results in terms of accuracy, F1, and BERTScore and report computational cost in terms of time and GPU memory during fine-tuning and inference stages. We find that LoRA methods achieved substantial performance gains of 36\% on average over base models. Our FinLoRA project provides an affordable and scalable approach to democratize financial intelligence to the general public. Datasets, LoRA adapters, code, and documentation are available at https://github.com/Open-Finance-Lab/FinLoRA

Figures

Figures reproduced from arXiv: 2505.19819 by the authors.

Figure 1
Figure 1. Average performance of base models and LoRA models. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Task suitability. 4.1 Angle I: LoRA Methods Performance on Financial Datasets Comparative Performance of LoRA Variants [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Average inference time of LoRA fine-tuned Llama 3.1 8B and LoRA fine￾tuned Gemini 2.0 FL across tasks FiNER FNXL Sentiment Analysis NER Headline Analysis FinancialMath XBRL Analysis FinanceBench 1 2 3 4 5 Inference Time (seconds) Llama LoRA Fine-tuned Gemini Fine-tuned [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MiniCache: Reusable Program Caching with Small Model Interfaces for Efficient LLM Inference

    cs.AI 2026-07 conditional novelty 6.0 of 10

    MiniCache caches the shared computation of similar LLM requests as parameterized programs, using a small model to fill in changing variables, cutting latency up to 3.1x and raising throughput about 2.8x in benchmarks.

Reference graph

Works this paper leans on

51 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    FinTral: A family of GPT-4 level multimodal financial large language models

    Gagan Bhatia, El Moatez Billah Nagoudi, Hasan Cavusoglu, and Muhammad Abdul-Mageed. FinTral: A family of GPT-4 level multimodal financial large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, pages 13064–13087, August 2024

    work page 2024

  2. [2]

    Can GPT models be financial analysts? an evaluation of ChatGPT and GPT-4 on mock CFA exams

    Ethan Callanan, Amarachi Mbakwe, Antony Papadimitriou, Yulong Pei, Mathieu Sibue, Xiao- dan Zhu, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. Can GPT models be financial analysts? an evaluation of ChatGPT and GPT-4 on mock CFA exams. In Proceedings of the Eighth Financial Technology and Natural Language Processing and the 1st Agent AI for Scenario Planning,...

    work page 2024

  3. [3]

    Data-driven detection of subtype-specific differentially expressed genes

    Lulu Chen, Yingzhou Lu, Chiung-Ting Wu, Robert Clarke, Guoqiang Yu, Jennifer E Van Eyk, David M Herrington, and Yue Wang. Data-driven detection of subtype-specific differentially expressed genes. Scientific Reports, 11(1):332, 2021

    work page 2021

  4. [4]

    Adaptformer: Adapting vision transformers for scalable visual recognition

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664–16678, 2022

    2022

  5. [5]

    Uncertainty quantification and interpretability for clinical trial approval prediction

    Tianyi Chen, Nan Hao, Capucine Van Rechem, Jintai Chen, and Tianfan Fu. Uncertainty quantification and interpretability for clinical trial approval prediction. Health Data Science, 4:0126, 2024

    work page 2024

  6. [6]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

    arXiv

  7. [7]

    QLoRA: Efficient finetuning of quantized LLMs

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient finetuning of quantized LLMs. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    work page 2023

  8. [8]

    The Llama 3 herd of models, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, and et al. The Llama 3 herd of models, 2024

    work page 2024

Show all 51 references
  1. [9]

    The Pile: An 800GB dataset of diverse text for language modeling

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. The Pile: An 800GB dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020

    2020 arXiv

  2. [10]

    XBRL Agent: Lever- aging large language models for financial report analysis

    Shijie Han, Haoqiang Kang, Bo Jin, Xiao-Yang Liu, and Steve Y Yang. XBRL Agent: Lever- aging large language models for financial report analysis. In Proceedings of the 5th ACM International Conference on AI in Finance, ICAIF ’24, page 856–864, 2024

    2024

  3. [11]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020

    2009 arXiv

  4. [12]

    LoRA: Low-rank adaptation of large language models

    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 International Conference on Learning Representations, 2022

    2022

  5. [13]

    Fine-tuning transformers efficiently: A survey on LoRA and its impact

    Muchen Huan and Jianhong Shun. Fine-tuning transformers efficiently: A survey on LoRA and its impact. Preprints, February 2025

    2025

  6. [14]

    GPT-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. GPT-4o system card. arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  7. [15]

    FinanceBench: A new benchmark for financial question answering, 2023

    Pranab Islam, Anand Kannappan, Douwe Kiela, Rebecca Qian, Nino Scherrer, and Bertie Vidgen. FinanceBench: A new benchmark for financial question answering, 2023

    2023

  8. [16]

    A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA, 2023

    Damjan Kalajdzievski. A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA, 2023. 10

    2023

  9. [17]

    Large language models in finance (finllms)

    Jean Lee, Nicholas Stevens, and Soyeon Caren Han. Large language models in finance (finllms). Neural Computing and Applications, January 2025

    2025

  10. [18]

    DeepSeek-V3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. DeepSeek-V3 technical report. arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  11. [19]

    Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning

    Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950–1965, 2022

    1950

  12. [20]

    SocraticLM: Exploring socratic personalized teaching with large language models

    Jiayu Liu, Zhenya Huang, Tong Xiao, Jing Sha, Jinze Wu, Qi Liu, Shijin Wang, and Enhong Chen. SocraticLM: Exploring socratic personalized teaching with large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  13. [21]

    DoRA: Weight-decomposed low-rank adaptation

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In Forty-first International Conference on Machine Learning, 2024

    2024

  14. [22]

    Data-centric FinGPT: De- mocratizing internet-scale data for financial large language models

    Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, and Daochen Zha. Data-centric FinGPT: De- mocratizing internet-scale data for financial large language models. In Workshop on Instruction Tuning and Instruction Following, NeurIPS, 2023

    2023

  15. [23]

    Efficient Pretraining and Finetuning of Quantized LLMs with Low-Rank Structure

    Xiao-Yang Liu, Jie Zhang, Guoxuan Wang, Weiqin Tong, and Anwar Walid. Efficient Pretraining and Finetuning of Quantized LLMs with Low-Rank Structure . In IEEE 44th International Conference on Distributed Computing Systems (ICDCS), pages 300–311, July 2024

    2024

  16. [24]

    Zhu, Daochen Zha, J

    Xiao-Yang Liu, R. Zhu, Daochen Zha, J. Gao, S. Zhong, Matt White, and Meikang Qiu. Differentially private low-rank adaptation of large language model using federated learning. ACM Transactions on Management Information Systems, 2024

    2024

  17. [25]

    FiNER: Financial numeric entity recognition for XBRL tagging

    Lefteris Loukas, Manos Fergadiotis, Ilias Chalkidis, Eirini Spyropoulou, Prodromos Malakasio- tis, Ion Androutsopoulos, and Georgios Paliouras. FiNER: Financial numeric entity recognition for XBRL tagging. In Proceedings of the 60th Annual Meeting of the Association for Comput...

    2022

  18. [26]

    COT: an effi- cient and accurate method for detecting marker genes among many subtypes

    Yingzhou Lu, Chiung-Ting Wu, Sarah J Parker, Zuolin Cheng, Georgia Saylor, Jennifer E Van Eyk, Guoqiang Yu, Robert Clarke, David M Herrington, and Yue Wang. COT: an effi- cient and accurate method for detecting marker genes among many subtypes. Bioinformatics Advances, 2(1), 2022

    2022

  19. [27]

    Www’18 open challenge: Financial opinion mining and question answering

    Macedo Maia, Siegfried Handschuh, Andre Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. Www’18 open challenge: Financial opinion mining and question answering. pages 1941–1942, 04 2018

    1941

  20. [28]

    Good debt or bad debt: Detecting semantic orientations in economic texts, 2013

    Pekka Malo, Ankur Sinha, Pyry Takala, Pekka Korhonen, and Jyrki Wallenius. Good debt or bad debt: Detecting semantic orientations in economic texts, 2013

    2013

  21. [29]

    A survey on lora of large language models

    Yuren Mao, Yuhang Ge, Yijiang Fan, Wenyi Xu, Yu Mi, Zhonghao Hu, and Yunjun Gao. A survey on lora of large language models. Frontiers of Computer Science, 19(7), December 2024

    2024

  22. [30]

    Communication-Efficient Learning of Deep Networks from Decentralized Data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence a...

    2017

  23. [31]

    Pissa: Principal singular values and singular vectors adaptation of large language models

    Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. Advances in Neural Information Processing Systems, 37:121038–121072, 2024

    2024

  24. [32]

    OpenAI API pricing

    OpenAI. OpenAI API pricing. https://platform.openai.com/docs/pricing, May 2025. Accessed: 2025-05-14. 11

    2025

  25. [33]

    Abdur Rahman

    Md. Abdur Rahman. Twitter financial news sentiment. http://precog.iiitd.edu.in/ people/anupama, 2022

    2022

  26. [34]

    An introduction to XBRL

    Ali Saeedi, Jim Richards, and Barry Smith. An introduction to XBRL. In British Accounting Association’s Annual Conference, 2007

    2007

  27. [35]

    Domain adaption of named entity recognition to support credit risk assessment

    Julio Cesar Salinas Alvarado, Karin Verspoor, and Timothy Baldwin. Domain adaption of named entity recognition to support credit risk assessment. In Ben Hachey and Kellie Webster, editors, Proceedings of the Australasian Language Technology Association Workshop 2015, pages 84–...

    2015

  28. [36]

    Financial numeric extreme labelling: A dataset and bench- marking

    Soumya Sharma, Subhendu Khatuya, Manjunath Hegde, Afreen Shaikh, Koustuv Dasgupta, Pawan Goyal, and Niloy Ganguly. Financial numeric extreme labelling: A dataset and bench- marking. In Findings of the Association for Computational Linguistics: ACL 2023 , pages 3550–3561, July 2023

    2023

  29. [37]

    Impact of news on the commodity market: Dataset and results, 2020

    Ankur Sinha and Tanmay Khandait. Impact of news on the commodity market: Dataset and results, 2020

    2020

  30. [38]

    Improving loRA in privacy-preserving federated learning

    Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. Improving loRA in privacy-preserving federated learning. In The Twelfth International Conference on Learning Representations, 2024

    2024

  31. [39]

    Gemini: A family of highly capable multimodal models, 2024

    Gemini Team, Rohan Anil, Sebastian Borgeaud, et al. Gemini: A family of highly capable multimodal models, 2024

    2024

  32. [40]

    PrivateLoRA for efficient privacy preserving LLM, 2023

    Yiming Wang, Yu Lin, Xiaodong Zeng, and Guannan Zhang. PrivateLoRA for efficient privacy preserving LLM, 2023

    2023

  33. [41]

    TWIN-GPT: Digital twins for clinical trials via large language model

    Yue Wang, Tianfan Fu, Yinlong Xu, Zihan Ma, Hongxia Xu, Bang Du, Yingzhou Lu, Honghao Gao, Jian Wu, and Jintai Chen. TWIN-GPT: Digital twins for clinical trials via large language model. ACM Trans. Multimedia Comput. Commun. Appl., July 2024. Just Accepted

    2024

  34. [42]

    BloombergGPT: A large language model for finance

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. BloombergGPT: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023

    2023 arXiv

  35. [43]

    Knowledge-infused legal wisdom: Navigating llm consultation through the lens of diagnostics and positive-unlabeled reinforcement learning

    Yang Wu, Chenghao Wang, Ece Gumusel, and Xiaozhong Liu. Knowledge-infused legal wisdom: Navigating llm consultation through the lens of diagnostics and positive-unlabeled reinforcement learning. In ACL (Findings), pages 15542–15555, 2024

    2024

  36. [44]

    FinBen: An holistic financial benchmark for large language models

    Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, Yijing Xu, Haoqiang Kang, Ziyan Kuang, Chenhan Yuan, Kailai Yang, Zheheng Luo, Tianlin Zhang, Zhiwei Liu, Guojun Xiong, Zhiyang Deng, Yuechen Jiang, Zhiy...

    2024

  37. [45]

    PIXIU: A comprehensive benchmark, instruction dataset and large language model for finance

    Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. PIXIU: A comprehensive benchmark, instruction dataset and large language model for finance. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and ...

    2023

  38. [46]

    Low-rank adaptation for foundation models: A comprehensive review, 2024

    Menglin Yang, Jialin Chen, Yifei Zhang, Jiahong Liu, Jiasheng Zhang, Qiyao Ma, Harshit Verma, Qianru Zhang, Min Zhou, Irwin King, and Rex Ying. Low-rank adaptation for foundation models: A comprehensive review, 2024

    2024

  39. [47]

    Enhancing financial sentiment analysis via retrieval augmented large language models

    Boyu Zhang, Hongyang Yang, Tianyu Zhou, Muhammad Ali Babar, and Xiao-Yang Liu. Enhancing financial sentiment analysis via retrieval augmented large language models. In ACM International Conference on AI in Finance, pages 349–356, 2023. 12

    2023

  40. [48]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020

    2020

  41. [49]

    A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023

    2023 arXiv

  42. [50]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595–46623, 2023

    2023

  43. [51]

    Large language models for disease diagnosis: A scoping review

    Shuang Zhou, Zidu Xu, Mian Zhang, Chunpu Xu, Yawen Guo, Zaifu Zhan, Sirui Ding, Jiashuo Wang, Kaishuai Xu, Yi Fang, et al. Large language models for disease diagnosis: A scoping review. arXiv preprint arXiv:2409.00097, 2024. 13

    2024 arXiv

Pith tools

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