Pith. sign in

REVIEW 5 major objections 6 minor 40 references

Meta-aware Learning in text-to-SQL Large Language Model

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Combining schema, reasoning, domain knowledge, and tokenized prompts during fine-tuning produces more accurate text-to-SQL on complex business databases.

desk verdict Plausible industrial text-to-SQL fine-tuning recipe undermined by a missing train/test disjointness guarantee on template-generated data. read the letter →

arxiv 2505.18929 v1 pith:KXLPV2U6 submitted 2025-05-25 cs.AI cs.CL

classification cs.AIcs.CL
keywords text-to-SQLLLMfine-tuningchain-of-thoughtschema-basedlearningdomainknowledgemetadatatokenizationcatastrophicforgettingexecutionaccuracy
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 the best way to adapt an LLM to a specific business database is to fine-tune it not just on question-SQL pairs but on a meta-aware training signal that includes four components: schema-based learning, chain-of-thought reasoning, domain-knowledge subtasks, and tokenization of schema and prompt tags. In two experiments on proprietary business data, the strongest meta-aware configurations outperform schema-only fine-tuning by large margins, with execution accuracy reaching 0.928 to 0.940 depending on the test setting. The paper also reports that tokenized prompt structures delay overfitting at small training sizes and that combining all four strategies avoids the catastrophic forgetting seen when the same knowledge is taught sequentially. The contribution is a practical fine-tuning framework that ties database metadata into the model rather than relying on few-shot prompting, and the evidence is generated from template-based question-SQL pairs over the same business tables.

What carries the argument

The central object is the meta-aware learning framework, an integrated fine-tuning setup with four sub-modules: schema-based learning, which feeds the exact tables and columns used in each query; chain-of-thought learning, which inserts step-by-step reasoning about tables, columns, and joins before the SQL; domain-knowledge enhancement learning, which trains the model on seven metadata subtasks such as identifying columns, recognizing tables, and understanding JOIN relationships; and key information tokenization, which adds database table and column names plus prompt structure tags like <question> and <answer> directly into the model's tokenizer. The machinery works by making the fine-tuning objective carry metadata relationships explicitly, so the model does not have to derive them from question-SQL pairs alone. Prompt structure tokenization is the component the paper shows delays overfitting and improves small-sample performance, while the combination of all four components is what yields the highest accuracy.

What would settle it

Take the best meta-aware model and test it on a set of business questions that are not producible from the paper's templates, such as questions written by analysts about the same tables using unseen phrasings and filter combinations, and check whether execution accuracy stays near 0.93 or falls toward the 0.12 to 0.7 range of the baselines.

Watch

Extended reading notes

Core claim

The paper's central claim is that a meta-aware learning framework—one that simultaneously trains an LLM on schema, chain-of-thought reasoning, domain-knowledge relationships, and tokenized key information—produces substantially more accurate SQL generation for complex business databases than schema-only fine-tuning. The proposed method treats these not as separate tricks but as one integrated fine-tuning objective: the model learns to map business questions to SQL while also learning which columns answer which questions, how tables relate through JOINs, what column descriptions and data types mean, and how the schema should be tokenized. Evidence is provided in two scenarios: on a single-table dataset, tokenized prompt structures beat a base prompt at every training size, including a 27.7% accuracy improvement at 250 samples; on a five-table dataset, Schema-CoT-OPT reaches 0.928 (full schema) and 0.930 (dynamic schema) while Schema-CoT-Kn reaches 0.940 on chain-of-thought tasks. Schema-only baselines stay below 0.7 on the same full and dynamic schema tests, and sequential learning of the same components collapses to below 0.1, which the paper attributes to catastrophic forgetting.

Load-bearing premise

The training and test question-SQL pairs are generated from the same template set over the same business tables, so the measured accuracies may reflect pattern memorization rather than the ability to handle genuinely new business questions.

Editorial extensions

If this is right

  • If the framework is right, teams fine-tuning small open-source LLMs on proprietary databases should combine schema, chain-of-thought, and metadata subtasks rather than relying on schema-only prompts.
  • Tokenized prompt tags are a low-cost way to delay overfitting when training data is small: with 250 training examples the tokenized prompt reaches 0.914 execution accuracy, and it outperforms the base prompt at every training size tested.
  • Progressive learning that teaches schema and knowledge sequentially drops all text-to-SQL task accuracies below 0.1 in the Schema@Kn order, whereas the integrated meta-aware objective keeps accuracy near 0.9, so the framework avoids the failure mode of staged fine-tuning.
  • Chain-of-thought instructions alone lift schema-only models above 0.9 on CoT test cases, while the meta-aware Schema-CoT-Kn reaches 0.940, indicating that reasoning traces can substitute for some schema familiarity at test time.

Reading between the lines

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

  • A direct extension the paper does not run is to test the same meta-aware training on a public cross-domain benchmark where train and test questions come from different sources, which would separate template memorization from genuine schema linking.
  • The template-based generation means the reported accuracy gap likely upper-bounds real-world gains; a fair estimate of production benefit would require testing on analyst-written questions outside the template space, using unseen phrasings and filter combinations.
  • The tokenization component is a tokenizer-level intervention whose benefit at very small n (250 samples) suggests it acts like an embedding prior or regularizer; this could be studied by measuring generalization to schema elements never seen during training.
  • The paper's comparison to progressive learning hints that catastrophic forgetting in text-to-SQL is not solved by curriculum order; a multi-task unified objective appears to be the more robust path, which would be a testable design principle for enterprise fine-tuning pipelines.
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. This paper proposes a meta-aware learning framework for fine-tuning text-to-SQL large language models on business databases. The framework combines schema-based learning, chain-of-thought reasoning, domain knowledge enhancement, and key information tokenization. The authors evaluate the framework in two scenarios using template-generated, private Walmart data, comparing execution accuracy against several baselines and progressive learning variants. They report that meta-aware approaches, particularly Schema-CoT-OPT and Schema-CoT-Kn, achieve the highest execution accuracy on full and dynamic schemas and on chain-of-thought tasks. The paper also examines overfitting behavior and catastrophic forgetting.

Significance. If the reported results are reliable, the proposed framework is a potentially practical recipe for adapting LLMs to proprietary business SQL environments, and the comparison of learning strategies would be useful to practitioners. The paper's strengths include a clear decomposition of four complementary learning strategies and an effort to ablate them. However, the evaluation has several serious gaps: the training and test sets are drawn from the same template-generated pool without an explicit disjoint split, the base model is unnamed, and no statistical significance is reported. These issues currently prevent the results from supporting the paper's central claim.

major comments (5)
  1. [Section IV.A (Scenario I and Scenario II); Tables II and III] The manuscript states that the test set is sampled from the same full dataset that is used to create training subsets, but it never states that the training and test samples are disjoint. In Scenario I, the 500 test points are drawn from the same ≈30,000 pool used for training subsets up to n=10,000; in Scenario II, both the 500 test points and the 5,000 training points are sampled from the same ≈1,000,000 pool. Without an explicit disjointness guarantee, the reported execution accuracies (e.g., 0.928 and 0.930 in Table III) could be inflated by memorization of question-SQL pairs that appear in both sets, so the claimed superiority of meta-aware learning over baselines is not established.
  2. [Section IV.E, second paragraph] The model was trained using 'the exact schema, containing only the tables and columns referenced in each ground truth SQL query.' This oracle-like schema condition may favor methods that do not need to perform schema linking; because the test conditions (full and dynamic schemas) contain additional tables and columns, the model is never trained to select the relevant subset from a larger schema. The authors should either train all methods without the oracle schema or explicitly justify this design choice and explain how it interacts with the claimed superiority.
  3. [Section IV.B; Tables II and III] No error bars, confidence intervals, or statistical significance tests are reported, and inference uses non-deterministic sampling (do_sample=True). With a test set of only 500 examples, differences such as 0.928 vs. 0.923 in Table III are likely within noise, so the claim that one method is superior to another is not statistically supported. The authors should report variance over multiple fine-tuning runs or at least provide a significance test for the key comparisons.
  4. [Section IV.D, first sentence] The base LLM is not named. The sentence 'Our fine-tuning experiments utilize the open-source LLM model (models are not open to public due to Walmart privacy requirements)' is self-contradictory: an open-source model should be identifiable, and the parenthetical conflates the base model with the fine-tuned models. Naming the base model is essential for reproducibility and for separating the effect of the proposed method from base-model choice.
  5. [Section IV.A and Algorithm 1] Both training and test data are generated from the same template-based process over the same metrics and filters, so the test measures pattern recall on near-duplicates rather than generalization to naturally occurring business questions that might not be covered by the templates. This limits the external validity of the reported accuracies and should be acknowledged explicitly; an evaluation on a held-out set of human-written or out-of-template questions would materially strengthen the paper's claims about real-world business databases.
minor comments (6)
  1. [Section IV.E, paragraph on progressive learning] The sentence 'As shown in Table II, introducing Kn after Schema-based learning masked prior text-to-SQL knowledge' should refer to Table III, since the results described are in the Scenario II comparison table.
  2. [Section III, first paragraph] The word 'meata-aware' should be 'meta-aware'.
  3. [Section IV.E, first paragraph] The phrase 'based line' should be 'baseline'.
  4. [Section IV.A, Scenario I] The sentence 'We sampled 500 data points in the testing dataset' is grammatically awkward; 'from the dataset' would be clearer.
  5. [Algorithm 1, step 14] The notation 'Pj = {S, Ij, Qj, Aj|T } ←Ij' is unclear; the assignment of the sampled instruction and the replacement of the static instruction should be described more explicitly.
  6. [Section II, paragraph on LLM approaches] The typo 'd.g., SQLCoder 2' should be 'e.g., SQLCoder 2'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the paper's claims are empirical comparisons of fine-tuning strategies, and no prediction reduces to a fitted input or self-citation.

full rationale

The paper contains no mathematical derivation chain in which an output reduces to an input by construction. The central claims are comparative execution accuracies from fine-tuned models; the methods (schema, CoT, knowledge, tokenization) are not defined in terms of the outcome, and no fitted parameter is relabeled as a prediction. The reference list contains no prior work by the author that is invoked as load-bearing, and no uniqueness theorem or ansatz is imported from self-citations. The reported evaluation on a private, template-generated dataset sampled from the same pool as training raises external-validity and potential-leakage concerns, but those are empirical weaknesses rather than circular reasoning, so under the hard rules no circularity step is recorded.

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

The framework adds no physical or conceptual entities beyond existing schema metadata and prompt tags. The empirical claims rest on assumptions about representativeness of template-generated data, the sufficiency of execution accuracy, and the suitability of an unnamed base model. Hand-chosen hyperparameters are listed because the reported gains have no sensitivity analysis.

free parameters (4)
  • Learning rate = 2e-5
    Hand-chosen in Section IV.D; no grid search or sensitivity analysis reported, yet results may depend on it.
  • LoRA rank = 32
    Hand-chosen in Section IV.D 'to ensure a reasonable scale'; no sensitivity analysis.
  • LoRA scaling factor alpha = 64
    Hand-chosen in Section IV.D; no sensitivity analysis.
  • Training sample size for best single-table result = 1000
    Section IV.E Scenario I states n=1,000 was sufficient to achieve the best accuracy; the value was selected after observing results across n=250 to 10,000.
assumptions (4)
  • domain assumption Template-generated question-SQL pairs from metrics and filters represent the distribution of real business queries.
    Section IV.A relies on templates for both training and test; if the template space differs from natural queries, accuracies are inflated.
  • domain assumption A balanced sample of 500 test examples from 5 tables is representative of production query mix.
    Scenario II samples 'averagely' from each table, which can overrepresent rare tables and underrepresent common ones.
  • domain assumption Execution accuracy against ground truth SQL outputs is a complete measure of SQL generation quality.
    Section IV.B defines the metric; no human evaluation, syntax validity checks, or robustness checks are reported.
  • domain assumption An unnamed open-source LLM is a suitable base model for the comparisons.
    Section IV.D says 'We use the open-source LLM model' without naming it, so the starting point cannot be assessed or reproduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Meta-aware Learning in text-to-SQL Large Language Model." pith.science (2026). https://pith.science/paper/KXLPV2U6

@misc{pith2026250518929,
  author       = {Pith},
  title        = {Pith review of: Meta-aware Learning in text-to-SQL Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXLPV2U6}},
  note         = {Machine review of arXiv:2505.18929}
}
read the original abstract

The advancements of Large language models (LLMs) have provided great opportunities to text-to-SQL tasks to overcome the main challenges to understand complex domain information and complex database structures in business applications. In this paper, we propose a meta-aware learning framework to integrate domain knowledge, database schema, chain-of-thought reasoning processes, and metadata relationships to improve the SQL generation quality. The proposed framework includes four learning strategies: schema-based learning, Chain-of-Thought (CoT) learning, knowledge-enhanced learning, and key information tokenization. This approach provides a comprehensive understanding of database structure and metadata information towards LLM through fine-tuning to improve its performance on SQL generation within business domains. Through two experimental studies, we have demonstrated the superiority of the proposed methods in execution accuracy, multi-task SQL generation capability, and reduction of catastrophic forgetting.

Figures

Figures reproduced from arXiv: 2505.18929 by the authors.

Figure 1
Figure 1. The comparison in SQL complexity from public database and the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of meta-aware learning with sub-modules: (a) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison plots between tokenized-prompt training [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 9 canonical work pages

  1. [1]

    Next-generation database interfaces: A survey of llm-based text-to-sql,

    Z. Hong, Z. Yuan, Q. Zhang, H. Chen, J. Dong, F. Huang, and X. Huang, “Next-generation database interfaces: A survey of llm-based text-to-sql,” arXiv preprint arXiv:2406.08426 , 2024

  2. [2]

    Din-sql: Decomposed in-context learning of text-to-sql with self-correction,

    M. Pourreza and D. Rafiei, “Din-sql: Decomposed in-context learning of text-to-sql with self-correction,” Advances in Neural Information Processing Systems, vol. 36, 2024

  3. [3]

    Codes: Towards building open-source language models for text-to-sql,

    H. Li, J. Zhang, H. Liu, J. Fan, X. Zhang, J. Zhu, R. Wei, H. Pan, C. Li, and H. Chen, “Codes: Towards building open-source language models for text-to-sql,” Proceedings of the ACM on Management of Data, vol. 2, no. 3, pp. 1–28, 2024

  4. [4]

    Constructing an interactive natural language interface for relational databases,

    F. Li and H. V . Jagadish, “Constructing an interactive natural language interface for relational databases,” Proceedings of the VLDB Endow- ment, vol. 8, no. 1, pp. 73–84, 2014

  5. [5]

    A rule based approach for nlp based query processing,

    T. Mahmud, K. A. Hasan, M. Ahmed, and T. H. C. Chak, “A rule based approach for nlp based query processing,” in 2015 2nd international conference on electrical information and communication technologies (EICT). IEEE, 2015, pp. 78–82

  6. [6]

    Sequence to sequence learning with neural networks,

    I. Sutskever, “Sequence to sequence learning with neural networks,” arXiv preprint arXiv:1409.3215 , 2014

  7. [7]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Informa- tion Processing Systems , 2017

  8. [8]

    Evaluating the text-to-sql ca- pabilities of large language models,

    N. Rajkumar, R. Li, and D. Bahdanau, “Evaluating the text-to-sql ca- pabilities of large language models,” arXiv preprint arXiv:2204.00498, 2022

Show all 40 references
  1. [9]

    Text- to-sql empowered by large language models: A benchmark evaluation,

    D. Gao, H. Wang, Y . Li, X. Sun, Y . Qian, B. Ding, and J. Zhou, “Text- to-sql empowered by large language models: A benchmark evaluation,” arXiv preprint arXiv:2308.15363 , 2023

  2. [10]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al. , “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 , 2023

  3. [11]

    Siren’s song in the ai ocean: a survey on hallucination in large language models,

    Y . Zhang, Y . Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y . Zhang, Y . Chen et al. , “Siren’s song in the ai ocean: a survey on hallucination in large language models,” arXiv preprint arXiv:2309.01219, 2023

  4. [12]

    Sql-palm: Improved large language model adaptation for text-to-sql (extended),

    R. Sun, S. ¨O. Arik, A. Muzio, L. Miculicich, S. Gundabathula, P. Yin, H. Dai, H. Nakhost, R. Sinha, Z. Wang et al. , “Sql-palm: Improved large language model adaptation for text-to-sql (extended),” arXiv preprint arXiv:2306.00739, 2023

  5. [13]

    Benchmarking the text-to-sql capability of large language models: A comprehensive evaluation,

    B. Zhang, Y . Ye, G. Du, X. Hu, Z. Li, S. Yang, C. H. Liu, R. Zhao, Z. Li, and H. Mao, “Benchmarking the text-to-sql capability of large language models: A comprehensive evaluation,” arXiv preprint arXiv:2403.02951, 2024

  6. [14]

    Towards complex text-to-sql in cross-domain database with interme- diate representation,

    J. Guo, Z. Zhan, Y . Gao, Y . Xiao, J.-G. Lou, T. Liu, and D. Zhang, “Towards complex text-to-sql in cross-domain database with interme- diate representation,” arXiv preprint arXiv:1905.08205 , 2019

  7. [15]

    Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql,

    H. Li, J. Zhang, C. Li, and H. Chen, “Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 11, 2023, pp. 13 067–13 075

  8. [16]

    An introduction to convolutional neural networks,

    K. O’Shea, “An introduction to convolutional neural networks,” arXiv preprint arXiv:1511.08458, 2015

  9. [17]

    A survey on in-context learning,

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, T. Liu et al., “A survey on in-context learning,” arXiv preprint arXiv:2301.00234, 2022

  10. [18]

    A systematic survey of prompt engineering in large language models: Techniques and applications,

    P. Sahoo, A. K. Singh, S. Saha, V . Jain, S. Mondal, and A. Chadha, “A systematic survey of prompt engineering in large language models: Techniques and applications,” arXiv preprint arXiv:2402.07927, 2024

  11. [19]

    Dynamic self-attention: Computing attention over words dynamically for sentence embedding,

    D. Yoon, D. Lee, and S. Lee, “Dynamic self-attention: Computing attention over words dynamically for sentence embedding,” arXiv preprint arXiv:1808.07383, 2018

  12. [20]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al. , “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

  13. [21]

    Act-sql: In- context learning for text-to-sql with automatically-generated chain-of- thought,

    H. Zhang, R. Cao, L. Chen, H. Xu, and K. Yu, “Act-sql: In- context learning for text-to-sql with automatically-generated chain-of- thought,” arXiv preprint arXiv:2310.17342 , 2023

  14. [22]

    Explor- ing chain-of-thought style prompting for text-to-sql,

    C.-Y . Tai, Z. Chen, T. Zhang, X. Deng, and H. Sun, “Explor- ing chain-of-thought style prompting for text-to-sql,” arXiv preprint arXiv:2305.14215, 2023

  15. [23]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” Advances in Neural Information Processing Systems , vol. 33, pp. 9459–9474, 2020

  16. [24]

    Dubo-sql: Diverse retrieval-augmented generation and fine tuning for text-to-sql,

    D. G. Thorpe, A. J. Duberstein, and I. A. Kinsey, “Dubo-sql: Diverse retrieval-augmented generation and fine tuning for text-to-sql,” arXiv preprint arXiv:2404.12560, 2024

  17. [25]

    A survey on employing large language models for text-to-sql tasks,

    L. Shi, Z. Tang, and Z. Yang, “A survey on employing large language models for text-to-sql tasks,” arXiv preprint arXiv:2407.15186 , 2024

  18. [26]

    Mitigating forgetting in llm supervised fine-tuning and preference learning,

    H. Fernando, H. Shen, P. Ram, Y . Zhou, H. Samulowitz, N. Baracaldo, and T. Chen, “Mitigating forgetting in llm supervised fine-tuning and preference learning,” arXiv preprint arXiv:2410.15483 , 2024

  19. [27]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

  20. [28]

    Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks,

    R. K. Mahabadi, S. Ruder, M. Dehghani, and J. Henderson, “Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks,” arXiv preprint arXiv:2106.04489 , 2021

  21. [29]

    Seq2sql: Generating structured queries from natural language using reinforcement learning,

    V . Zhong, C. Xiong, and R. Socher, “Seq2sql: Generating structured queries from natural language using reinforcement learning,” arXiv preprint arXiv:1709.00103, 2017

  22. [30]

    Universal language model fine-tuning with subword tokenization for polish,

    P. Czapla, J. Howard, and M. Kardas, “Universal language model fine-tuning with subword tokenization for polish,” arXiv preprint arXiv:1810.10222, 2018

  23. [31]

    Typesql: Knowledge-based type-aware neural text-to-sql generation,

    T. Yu, Z. Li, Z. Zhang, R. Zhang, and D. Radev, “Typesql: Knowledge-based type-aware neural text-to-sql generation,” arXiv preprint arXiv:1804.09769, 2018

  24. [32]

    Grappa: Grammar-augmented pre-training for table semantic parsing,

    T. Yu, C.-S. Wu, X. V . Lin, B. Wang, Y . C. Tan, X. Yang, D. Radev, R. Socher, and C. Xiong, “Grappa: Grammar-augmented pre-training for table semantic parsing,” arXiv preprint arXiv:2009.13845 , 2020

  25. [33]

    Data augmentation instead of explicit regularization,

    A. Hern ´andez-Garc´ıa and P. K ¨onig, “Data augmentation instead of explicit regularization,” arXiv preprint arXiv:1806.03852 , 2018

  26. [34]

    Ada-instruct: Adapting instruction generators for complex reasoning,

    W. Cui and Q. Wang, “Ada-instruct: Adapting instruction generators for complex reasoning,” arXiv preprint arXiv:2310.04484 , 2023

  27. [35]

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task,

    T. Yu, R. Zhang, K. Yang, M. Yasunaga, D. Wang, Z. Li, J. Ma, I. Li, Q. Yao, S. Roman et al., “Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task,” arXiv preprint arXiv:1809.08887 , 2018

  28. [36]

    Code llama: Open foundation models for code,

    B. Roziere, J. Gehring, F. Gloeckle, S. Sootla, I. Gat, X. E. Tan, Y . Adi, J. Liu, R. Sauvestre, T. Remez et al., “Code llama: Open foundation models for code,” arXiv preprint arXiv:2308.12950 , 2023

  29. [37]

    Pae: Llm-based product attribute extraction for e-commerce fashion trends,

    A. Sinha and E. Gujral, “Pae: Llm-based product attribute extraction for e-commerce fashion trends,” arXiv preprint arXiv:2405.17533 , 2024

  30. [38]

    Increasing model capacity for free: A simple strategy for parameter efficient fine-tuning,

    H. Song, H. Zhao, S. Majumder, and T. Lin, “Increasing model capacity for free: A simple strategy for parameter efficient fine-tuning,” arXiv preprint arXiv:2407.01320 , 2024

  31. [39]

    Progressive learning: A deep learning framework for continual learning,

    H. M. Fayek, L. Cavedon, and H. R. Wu, “Progressive learning: A deep learning framework for continual learning,” Neural Networks , vol. 128, pp. 345–357, 2020

  32. [40]

    Investi- gating the catastrophic forgetting in multimodal large language model fine-tuning,

    Y . Zhai, S. Tong, X. Li, M. Cai, Q. Qu, Y . J. Lee, and Y . Ma, “Investi- gating the catastrophic forgetting in multimodal large language model fine-tuning,” in Conference on Parsimony and Learning . PMLR, 2024, pp. 202–227

Pith tools

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