text string | source string |
|---|---|
not contain sensitive details. In the second stage, LLMs are applied to the syntheticdata, and rewriting strategies are employed to eliminate potential privacy leaks (such as removing specific names or obfuscating descriptions). This process of evaluation and rewriting is repeated to ensure that the generated data reta... | https://arxiv.org/abs/2505.18458v2 |
warcio and Trafilatura [73], followed by language identifica- tion (i.e., removing content with limited natural language) using fastText [201] as implemented in CCNet [418]. (ii) Data filtering: To eliminate low-quality content, MDR employs both (1) document-level filtering [330] and (2) line- level filtering, which re... | https://arxiv.org/abs/2505.18458v2 |
[91] proposes a “Probe- Analyze-Refine” workflow, which involves systematically ex- ploring the impact of various data processing operations and their orders on model performance, combining effective operations into data recipes, and optimizing data utilization through duplication analysis and diversity analysis. The o... | https://arxiv.org/abs/2505.18458v2 |
tf.train.Example , eliminating the need for separate label files. Moreover, as multimodal datasets sub- stantially increase data volume, TFRecord supports data sharding, enabling the creation of distributed files that can be assigned across multiple servers to facilitate parallel training. •MindRecord organizes data in... | https://arxiv.org/abs/2505.18458v2 |
the data needs to 24 Storage Service Meta Service SSD SSD SSD FoundationDB Cluster Manager FoundationDB Cluster Configuration Data Meta Data Raw Data Chain Replication with Apportioned Queries Key-Value Storage Key-Value Storage POSIX Client API Filesystem in Userspace Native Client FUSE Client RDMA Network Application... | https://arxiv.org/abs/2505.18458v2 |
user-side alignment operations. Heterogeneous Storage Systems. Heterogeneous stor- age systems refers to deploying the model state across di- verse storage media (e.g., GPUs, CPUs, NVMes Memory). When deploying the model, The Zero Redundancy Optimizer (ZeRO) [337] deploys model states across multiple GPUs. However, sim... | https://arxiv.org/abs/2505.18458v2 |
original data through multiple stages (e.g., Content Organization, Chunking, Embedding, Compression and Storage). (1) Content Organization. For the source data, organizing the content can enhance its logical structure, thereby facilitating improved efficiency and accuracy in retrieval. Works like Dense x retrieval [98]... | https://arxiv.org/abs/2505.18458v2 |
semantic omission and reference digestion. This reduces the embedding dimensions and inference latency, making it especially effec- tive for large-scale tasks [24]. •GTE introduces a dual-negative sampling strategy within its contrastive learning paradigm. Though introducing nega- tive samples usually works in series o... | https://arxiv.org/abs/2505.18458v2 |
(e.g., PQ6x10), trading accuracy for reduced storage space, making it suitable for high storage demands and lower precision needs. In the Milvus [26], vector storage differs based on the number of vectors per entity. For single-vector entities, vectors are stored continuously without row IDs. Since vectors are sorted b... | https://arxiv.org/abs/2505.18458v2 |
Delta Index mechanism, which builds local indexes only for newly inserted edges and entities, using background merging threads without the need for community reconstruction, significantly reducing overhead related to community detection compared to GraphRAG. MiniRAG [137] proposes a semantic-aware heterogeneous graph i... | https://arxiv.org/abs/2505.18458v2 |
jobs), (2) substitutability (the I/O order does not affect job correctness, enabling small caches to improve performance by substituting data and reducing thrashing), and (3) predictability (using mini-batch processing times to estimate job sensitivity to I/O performance for informed cache allocation). Fluid [163] dyna... | https://arxiv.org/abs/2505.18458v2 |
Orchestrator deploys the aforementioned three components as containers within the same Borg [390] unit. Although the above method of offloading to remote CPU servers can alleviate data stalls, the cost of remote CPUs is high, and the resources of ML accelerator nodes are not fully utilized. Pecan [161] introduces two s... | https://arxiv.org/abs/2505.18458v2 |
fault tolerance refers to the ability to quickly resume from the point of interruption during model training by storing checkpoints or performing redundant computations in the event of training interruptions.Principles Compared to traditional machine learning, LLMs place greater emphasis on fault tolerance during train... | https://arxiv.org/abs/2505.18458v2 |
recovery from CPU memory in the event of failures. ByteCheckpoint [396] manages checkpoint files using an architecture combining SSD and HDD storage servers. New checkpoint files are stored as ”hot” data on SSDs for quick access due to evaluation task downloads after creation. Once the evaluation is completed and there... | https://arxiv.org/abs/2505.18458v2 |
slower storage media (such as SSD), or releasing them directly. RAGCache [199] provides a prefix-aware PGDSF replacement policy that prioritizes cache nodes based on access frequency, size, and recomputation cost. And stores frequently accessed data in fast GPU memory and less frequent data in slower host memory, maxim... | https://arxiv.org/abs/2505.18458v2 |
various stages (e.g., in different epochs for pretraining). For example, corresponding training data needs to be supplied according to the training requirements during the training stage; during the RAG stage, corresponding knowledge needs to be supplied based on the degree of relevance to the questions. Principles Com... | https://arxiv.org/abs/2505.18458v2 |
domains with slower learning progress, thereby achieving a balanced learn- ing effect. Some methods [56], [424] combine reinforcement learning based on scoring the model to adjust the dataset. ODM [56] is based on the multi-armed bandit algorithm. It regards each data domain as an arm and uses classical reinforcement l... | https://arxiv.org/abs/2505.18458v2 |
minimizing computational overhead. (2) RAG Knowledge Re-ranking. After filtering, multiple doc- uments may remain, requiring re-ranking of the retrieval results to place the most relevant ones at the top for more accurate model output. Research on [129] shows that using a large model for re-ranking performs better than... | https://arxiv.org/abs/2505.18458v2 |
version, LongLLMLingua [192], uses a dual-granularity compression strategy: ( i) Coarse- grained compression initially filters key information at the document level to provide more focused content for fine- grained compression; ( ii) Fine-grained compression further optimizes at the token level to precisely retain key ... | https://arxiv.org/abs/2505.18458v2 |
improving space utilization. (2) The Non-Negative Least Squares Histogram Packing (NNL- SHP) converts the packing problem into a non-negative least squares problem, using dynamic programming to enumerate reasonable sequence combination strategies, constructing a packing matrix to determine the strategy’s repetition cou... | https://arxiv.org/abs/2505.18458v2 |
score for each word in the text (based on the numbers generated by the pseudo-random function). Unlike previous approaches, UPV [258] introduces a wa- termarking method that enables detection without requiring access to the key used during generation, thereby eliminating the risk of key leakage. It employs two independ... | https://arxiv.org/abs/2505.18458v2 |
missing values, correcting typos, resolving formatting inconsistencies, and addressing dependency violations. These tasks are typically categorized into data standardization, error detection and correction, and data imputation. Traditional data cleaning methods depend on rigid rules and constraints (e.g., zip code vali... | https://arxiv.org/abs/2505.18458v2 |
engineering, the second method employs LLM agents to automatically generate cleaning operations and orchestrate end-to-end pipelines. For instance, CleanA- gent [321] integrates domain-specific APIs with autonomous agents to execute a standardization pipeline that includes API call generation (e.g., clean date(df, ‘‘Ad... | https://arxiv.org/abs/2505.18458v2 |
pseudo-labeled tuples generated via LLM - based augmentation, with each training instance formatted as a context-enriched prompt comprising: (i) an instruction (e.g., “Correct the ProviderID to a valid numeric format”), (ii) a serialized erroneous cell with row and column context (e.g., “<COL>ProviderID<VAL>1x1303... ”... | https://arxiv.org/abs/2505.18458v2 |
Llama 3.1 and GPT-4o mini) with (i) different prompt designs, (ii) the selection of in-context demonstrations, (iii) automatic generation of matching rules, and (iv) fine-tuning LLMs using a shared pool of training data. To reduce inference costs, BATCHER [135] introduces a batch prompting method that allows multiple e... | https://arxiv.org/abs/2505.18458v2 |
language creation (e.g., explicitly defining valid match criteria), Chain-of-Thought reasoning, and the output automater (e.g., generating struc- tured JSON outputs for downstream automation). (2) End-to-End Matching via Context-Enriched RAG. To enrich the matching context and improve accuracy, the sec- ond method inte... | https://arxiv.org/abs/2505.18458v2 |
LEDD [58] employs carefully crafted prompts to support core data discovery tasks in data lakes. For hierarchical cataloging, prompts instruct LLM to summarize data clusters into semantically meaningful categories. For semantic search, prompts refine natural lan- guage queries before embedding and retrieval. For real-ti... | https://arxiv.org/abs/2505.18458v2 |
using prompts composed of structured markdown tables comprising captions, headers, and sample rows alongside explicit task instructions. 3.2 LLM for Data Analysis Apart from data manipulation, LLMs hold the potential to revolutionize traditional data analysis paradigms by sup- porting natural language interfaces and en... | https://arxiv.org/abs/2505.18458v2 |
Jupyter notebook corpus with 9.6B tokens) so as to improve model performance on analysis-related tasks (e.g., calculate the amount of games added in each year for each month). DataCoder [178] utilizes different types of contexts (e.g., code, text, and data) by employing dual encoders (e.g., data encoder and code + text... | https://arxiv.org/abs/2505.18458v2 |
Table-Specific LLM Fine-Tuning. Fine-tuning LLMs on task-specific table datasets enables them to internal- ize analytical knowledge directly within their parameters. TableGPT [242] fine-tunes LLMs like GPT-3.5 using a diverse set of table tasks synthesized from real-world tables. Building on Qwen2.5 [326], TableGPT2 [3... | https://arxiv.org/abs/2505.18458v2 |
Neo4j [17]. Traditional graph analysis (e.g., statistical methods, graph neural network (GNN) based methods) encompasses a spec- trum of tasks, including node classification (e.g., categorizing academic papers into research domains), graph classification (e.g., predicting node properties over molecular graphs), link pr... | https://arxiv.org/abs/2505.18458v2 |
[467] introduces a two-stage approach. First, a trainable and decoupled subgraph retriever selects a relevant subgraph based on the query. Then, reasoning is performed over the retrieved subgraph to derive the final answer. UniKGQA [195] integrates retrieval and reasoning within a unified model ar- chitecture. It compr... | https://arxiv.org/abs/2505.18458v2 |
properties (e.g., tags, headers) and have hierarchical or nested representation (e.g., County -Province -City in a nested JSON). 3.2.2.1 Markup Language Markup languages (e.g., XML, JSON, and HTML) are widely used for structuring and exchanging data across systems. Traditional approaches for processing these formats ty... | https://arxiv.org/abs/2505.18458v2 |
duces semantic errors, resulting in suboptimal performance. To fill this gap, OCR-free methods have emerged, directly generating the target token sequences with end-to-end mul- timodal LLMs [259], [415]. Based on different approaches to enhancing model understanding of textual semantics, related works can be categorize... | https://arxiv.org/abs/2505.18458v2 |
Dependencies). This enables more fine- grained vulnerability analysis at the statement level. To re- duce the impact of irrelevant information, [466] decomposes the control flow graph (CFG) into multiple execution paths from the entry node to the exit node. CodeBERT and a CNN are employed to capture intra-path and inte... | https://arxiv.org/abs/2505.18458v2 |
of irrelevant information [421]. As a result, Repoformer introduces a self-supervised learning approach to enable the model to accurately judge whether retrieval can improve its output quality. A new <eof>token is introduced to guide the model in determining whether context retrieval is necessary. Based on the output a... | https://arxiv.org/abs/2505.18458v2 |
as well as hardware settings (e.g., memory size and CPU core count). •Output Tuning Requirement. To ensure accurate pars- ing and interpretation of configurations generated by LLM , output formats are explicitly specified in the prompt. For instance, LLMBench [245] requires that recommended knob values be returned in J... | https://arxiv.org/abs/2505.18458v2 |
training data com- prising “ (workload)→(configuration) ” pairs, where diverse workloads are generated via GPT-4 prompting and optimal configurations are identified using the HEBO algorithm [113]. 3.3.2 LLM for Query Optimization Query optimization aims to accelerate SQL execution through logical (e.g., query rewriting... | https://arxiv.org/abs/2505.18458v2 |
semantically similar to the in- put query to provide more effective guidance. For instance, LLM -R2[250] introduces a contrastive representation model to encode query plans based on features such as operators, cardinalities, and costs, and retrieves a set of high-quality demonstrations, i.e., successfully optimized rew... | https://arxiv.org/abs/2505.18458v2 |
formats. These LLM -based approaches can be categorized as follows. Manually Crafted Prompts for Anomaly Diagnosis. The first method emulates the reasoning process of a human DBA, which involves referencing essential statistical informa- tion and conducting an in-depth analysis during diagnosis. The information is inco... | https://arxiv.org/abs/2505.18458v2 |
Amazon Nova [3] employs automated document processing to extract structured information from diverse unstructured sources. In terms of data systems, PawSQL [41], an advanced query optimization platform, offers both SQL rewrit- ing and index recommendation capabilities, adopted by over 10,000 professionals. Database dia... | https://arxiv.org/abs/2505.18458v2 |
sharing, and version control of datasets with real-time knowledge updates. By leveraging community-driven contributions, this platform could enable users to synthesize and share datasets using customizable methods, such as LLM-generated prompts from documents or websites, offering continuous, high-quality updates and i... | https://arxiv.org/abs/2505.18458v2 |
Data Analysis at the system level and offer a generalized capability for analyzing data across different structural types, thereby facilitating data automation. 4.2.2 Data Analysis with Private Domain Knowledge Another challenge in leveraging LLMs for data analysis is the effective utilization of private domain knowled... | https://arxiv.org/abs/2505.18458v2 |
and open problems in DATA4LLM , LLM4DATA , and hybrid data and LLM optimization. References [1] https://arangodb.com/. [2] https://arxiv.org/. [3] https://aws.amazon.com/cn/ai/generative- ai/nova/understanding/. [4] https://aws.amazon.com/s3. [5] https://bailian.console.aliyun.com/xiyan. [6] https://beautiful-soup-4.re... | https://arxiv.org/abs/2505.18458v2 |
with retrieval aug- mented generation. arXiv preprint arXiv:2406.13372 , 2024. [58] Q. An, C. Ying, Y. Zhu, Y. Xu, M. Zhang, and J. Wang. LEDD: large language model-empowered data discovery in data lakes. CoRR , abs/2502.15182, 2025. [59] R. Angles. A comparison of current graph database models. In 2012 IEEE 28th Inter... | https://arxiv.org/abs/2505.18458v2 |
Online, Aug. 2021. Association for Computational Linguistics. [74] T. Bendinelli, A. Dox, and C. Holz. Exploring LLM agents for cleaning tabular machine learning datasets. CoRR , abs/2503.06664, 2025. [75] M. Berglund and B. van der Merwe. Formalizing bpe tokeniza- tion. arXiv preprint arXiv:2309.08715 , 2023. [76] J. ... | https://arxiv.org/abs/2505.18458v2 |
Y. Wang, J. Wang, B. Raj, and M. I. Abdin. On the diversity of synthetic data and its impact on training large language models. arXiv preprint arXiv:2410.15226 , 2024. [93] J. Chen, Z. Chen, J. Wang, K. Zhou, Y. Zhu, J. Jiang, Y. Min, W. X. Zhao, Z. Dou, J. Mao, et al. Towards effective and efficient continual pre-trai... | https://arxiv.org/abs/2505.18458v2 |
D. Luan, H. Lim, B. Zoph, A. Spiridonov, R. Sepassi, D. Dohan, S. Agrawal, M. Omernick, A. M. Dai, T. S. Pillai, M. Pellat, A. Lewkowycz, E. Moreira, R. Child, O. Polozov, K. Lee, Z. Zhou, X. Wang, B. Saeta, M. Diaz, O. Firat, M. Catasta, J. Wei, K. Meier-Hellstern, D. Eck, J. Dean, S. Petrov, and N. Fiedel. Palm: Scal... | https://arxiv.org/abs/2505.18458v2 |
Chatel: Entity linking with chatbots, 2024. [121] J. Dodge, M. Sap, A. Marasovi´ c, W. Agnew, G. Ilharco, D. Groeneveld, M. Mitchell, and M. Gardner. Documenting large webtext corpora: A case study on the colossal clean crawled corpus. arXiv preprint arXiv:2104.08758 , 2021. [122] F. Dong, Y. He, Y. Liang, Z. Liu, Y. W... | https://arxiv.org/abs/2505.18458v2 |
Fayyaz, E. Aghazadeh, A. Modarressi, M. T. Pilehvar, Y. Yaghoobzadeh, and S. E. Kahou. Bert on a data diet: Finding important examples by gradient-based pruning. arXiv preprint arXiv:2211.05610 , 2022. [139] H. Feng, Q. Liu, H. Liu, J. Tang, W. Zhou, H. Li, and C. Huang. Docpedia: Unleashing the power of large multimod... | https://arxiv.org/abs/2505.18458v2 |
Ge, Z. Ma, D. Chen, Y. Li, and B. Ding. Data mixing made efficient: A bivariate scaling law for language model pretraining. arXiv preprint arXiv:2405.14908 , 2024. [155] Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 , 2021. [156] M. Geng, S. Wang, D. Do... | https://arxiv.org/abs/2505.18458v2 |
[171] D. Hernandez, T. Brown, T. Conerly, N. DasSarma, D. Drain, S. El-Showk, N. Elhage, Z. Hatfield-Dodds, T. Henighan, T. Hume, et al. Scaling laws and interpretability of learning from repeated data. arXiv preprint arXiv:2205.10487 , 2022. [172] J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherf... | https://arxiv.org/abs/2505.18458v2 |
annotated geospatial data from common crawl, 2024. [186] A. Ilyas, S. M. Park, L. Engstrom, G. Leclerc, and A. Madry. Datamodels: Predicting predictions from training data. arXiv preprint arXiv:2202.00622 , 2022. [187] K. Islam, M. Z. Zaheer, A. Mahmood, and K. Nandakumar. Diffusemix: Label-preserving data augmentation... | https://arxiv.org/abs/2505.18458v2 |
T. Huang, B. Zhou, and M. Chen. Familiarity- aware evidence compression for retrieval augmented genera- tion. arXiv preprint arXiv:2409.12468 , 2024. [203] J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. arXiv pr... | https://arxiv.org/abs/2505.18458v2 |
CoRR , abs/2503.02718, 2025. [220] M. M. Krell, M. Kosec, S. P. Perez, and A. Fitzgibbon. Efficient sequence packing without cross-contamination: Accelerating large language models without impacting performance. arXiv preprint arXiv:2107.02027 , 2021. [221] A. V. Kumar and M. Sivathanu. Quiver: An informed storage cach... | https://arxiv.org/abs/2505.18458v2 |
Y. Liu. Lexeval: A comprehensive chinese legal benchmark for evaluating large language models. arXiv preprint arXiv:2409.20288 , 2024. [235] H. Li, Q. Dong, Z. Tang, C. Wang, X. Zhang, H. Huang, S. Huang, X. Huang, Z. Huang, D. Zhang, et al. Synthetic data (almost) from scratch: Generalized instruction tuning for langu... | https://arxiv.org/abs/2505.18458v2 |
Du, X. Hu, B. Zhang, Y. Ye, Z. Li, R. Zhao, and H. Mao. PET-SQL: A Prompt- Enhanced Two-Round Refinement of Text-to-SQL with Cross- consistency, June 2024. [250] Z. Li, H. Yuan, H. Wang, G. Cong, and L. Bing. LLM-R2: A large language model enhanced rule-based rewrite system for boosting query efficiency. Proc. VLDB End... | https://arxiv.org/abs/2505.18458v2 |
selection in instruction tuning. arXiv preprint arXiv:2312.15685 , 2023. [267] Y. Liu, H. Li, Y. Cheng, S. Ray, Y. Huang, Q. Zhang, K. Du, J. Yao, S. Lu, G. Ananthanarayanan, et al. Cachegen: Kv cache compression and streaming for fast large language model serv- ing. In Proceedings of the ACM SIGCOMM 2024 Conference , ... | https://arxiv.org/abs/2505.18458v2 |
reranker for hard samples! In Findings of the Association for Computational Linguistics: EMNLP 2023 . Association for Computational Linguistics, 2023. [283] Z. Ma, B. Zhang, J. Zhang, J. Yu, X. Zhang, X. Zhang, S. Luo, X. Wang, and J. Tang. Spreadsheetbench: Towards challenging real world spreadsheet manipulation, 2024... | https://arxiv.org/abs/2505.18458v2 |
H. Man, N. T. Ngo, F. Dernoncourt, R. A. Rossi, and T. H. Nguyen. CulturaX: A cleaned, enormous, and multilingual dataset for large language models in 167 languages. In N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, and N. Xue, editors, Proceedings of the 2024 Joint International Conference on Computational Lin... | https://arxiv.org/abs/2505.18458v2 |
J. Launay. The refinedweb dataset for falcon llm: Outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 , 2023. [314] B. Peng, C. Li, P. He, M. Galley, and J. Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 , 2023. [315] M. E. Peters and D. Lecocq. Content e... | https://arxiv.org/abs/2505.18458v2 |
, 2021. [331] R. Rafailov, A. Sharma, E. Mitchell, S. Ermon, C. D. Manning, and C. Finn. Direct preference optimization: Your language model is secretly a reward model, 2024. [332] C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu. Exploring the limits of transfer learning w... | https://arxiv.org/abs/2505.18458v2 |
, 2017. [348] C. E. Shannon. A mathematical theory of communication. The Bell system technical journal , 27(3):379–423, 1948. [349] Y. Shao, L. Li, Y. Ma, P. Li, D. Song, Q. Cheng, S. Li, X. Li, P. Wang, Q. Guo, et al. Case2code: Scalable synthetic data for code generation. In Proceedings of the 31st International Conf... | https://arxiv.org/abs/2505.18458v2 |
E. Slyman, S. Lee, S. Cohen, and K. Kafle. Fairdedup: De- tecting and mitigating vision-language fairness disparities in se- mantic dataset deduplication. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 13905–13916, 2024. [366] L. Soldaini, R. Kinney, A. Bhagia, D. Schwenk, ... | https://arxiv.org/abs/2505.18458v2 |
arXiv preprint arXiv:2402.13446. [380] Z. Tan, D. Li, S. Wang, et al. Large language models for data annotation and synthesis: A survey. In EMNLP , pages 930–957. Association for Computational Linguistics, 2024. [381] J. Tang, Y. Yang, W. Wei, L. Shi, L. Su, S. Cheng, D. Yin, and C. Huang. Graphgpt: Graph instruction t... | https://arxiv.org/abs/2505.18458v2 |
K. Chen, Z. Lin, J. Han, and G. Ding. Yolov10: Real-time end-to-end object detection, 2024. [399] B. Wang, C. Xu, X. Zhao, L. Ouyang, F. Wu, Z. Zhao, R. Xu, K. Liu, Y. Qu, F. Shang, B. Zhang, L. Wei, Z. Sui, W. Li, B. Shi, Y. Qiao, D. Lin, and C. He. Mineru: An open-source solution for precise document content extracti... | https://arxiv.org/abs/2505.18458v2 |
vision vocabulary for large vision-language models, 2023. [415] H. Wei, C. Liu, J. Chen, J. Wang, L. Kong, Y. Xu, Z. Ge, L. Zhao, J. Sun, Y. Peng, C. Han, and X. Zhang. General ocr theory: Towards ocr-2.0 via a unified end-to-end model, 2024. [416] L. Wei, G. Xiao, and M. Balazinska. RACOON: an llm-based framework for ... | https://arxiv.org/abs/2505.18458v2 |
2023. [431] W. Xie. Analysis of the reasoning with redundant information provided ability of large language models. arXiv preprint arXiv:2310.04039 , 2023. [432] Y. Xie, K. Aggarwal, and A. Ahmad. Efficient continual pre- training for building domain specific large language models. In Findings of the Association for Co... | https://arxiv.org/abs/2505.18458v2 |
Advances in Neural Information Processing Systems , 37:83465– 83496, 2024. [446] Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y. Yang, W. Hong, X. Zhang, G. Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072 , 2024. [447] Z. Yao, H. Li, J. Zhang, ... | https://arxiv.org/abs/2505.18458v2 |
Ming, Y. Jin, D. Chai, L. Yang, H. Tian, Z. Fan, and K. Chen. DH-RAG: A dynamic historical context- powered retrieval-augmented generation method for multi-turn dialogue. CoRR , abs/2502.13847, 2025. [463] H. Zhang, Y. Dong, C. Xiao, and M. Oyamada. Jelly- fish: A large language model for data preprocessing. CoRR , abs... | https://arxiv.org/abs/2505.18458v2 |
Tijoriwala, K. Asher, H. Wu, A. Basant, D. Ford, D. David, N. Yigitbasi, P. Singh, C.-J. Wu, and C. Kozyrakis. Tectonic-Shift: A composite storage fabric for Large-Scale ML training. In 2023 USENIX Annual Technical Conference (USENIX ATC 23) , pages 433–449, Boston, MA, July 2023. USENIX Association. [479] W. Zhao, H. ... | https://arxiv.org/abs/2505.18458v2 |
language models. arXiv Preprint , 2025. [493] W. Zhou, C. Lin, X. Zhou, and G. Li. Breaking it down: An in-depth study of index advisors. Proc. VLDB Endow. , 17(10):2405–2418, 2024. [494] W. Zhou, C. Lin, X. Zhou, G. Li, and T. Wang. Demonstration of vita: Visualizing, testing and analyzing index advisors. In CIKM , pa... | https://arxiv.org/abs/2505.18458v2 |
From Reddit to Generative AI: Evaluating Large Language Models for Anxiety Support Fine-tuned on Social Media Data Ugur Kursuncu1, Trilok Padhi1†, Gaurav Sinha2†, Abdulkadir Erol1†, Jaya Krishna Mandivarapu3†, Christopher R. Larrison4 1Georgia State University, Atlanta, GA, USA. 2University of Georgia, Athens, GA, USA.... | https://arxiv.org/abs/2505.18464v1 |
fine-tuning, allowing us to examine their performance under conditions that closely resemble real- world interactions, rather than relying on synthetic or artificial scenarios. This study makes the following three contributions: (i) evaluating LLMs utilizing anxiety-related posts from Reddit for prompting and fine-tuni... | https://arxiv.org/abs/2505.18464v1 |
interpretability (Crasto, Dias, Miranda, & Kayande, 2021; Elyoseph, Levkovich, & Shinan-Altman, 2024). Collectively, prior work addresses key risks and opportunities 3 of LLM-based mental health support, including prediction, interpretability, emotional reasoning, and safety, yet tends to evaluate LLM capabilities in i... | https://arxiv.org/abs/2505.18464v1 |
interactions, we excluded self-comments and retained only posts with at least one comment scored greater than four by the contributors (upvotes – downvotes) (Reddit, 2025a, 2025b), resulting in 26,450 posts. This final dataset was split into 21,000 for fine-tuning and 5,000 for testing. 3.2 LLM Selection, Fine-tuning a... | https://arxiv.org/abs/2505.18464v1 |
general audience can understand the text. Semantic coherence metrics, such as Cv,Cnpmi, assess the co- herence of topics and the overall thematic structure within a response by analyzing word co-occurrence patterns. Semantic consistency such as BERT Score (T. Zhang, Kishore, Wu, Weinberger, & Artzi, 2019), BLEURT (Sell... | https://arxiv.org/abs/2505.18464v1 |
tell me more about that?”). A summary of metric definitions is provided in Table 1, with further details in Appendix C. To incorporate this empathy framework into our model, we modified the train- ing dataset to reflect these empathic behaviors. Each user message was paired with one of the three empathy categories: EMO... | https://arxiv.org/abs/2505.18464v1 |
majority decision (i.e., two out of three evaluators agreeing) was used to determine the final classification for each response. 3.5 Statistical Analysis To evaluate differences in model performance across metrics in each category, we first applied Levene’s test to assess homogeneity of variance, which indicated unequa... | https://arxiv.org/abs/2505.18464v1 |
Automated Readability Index* 1(-3) 3(2) 2(-1) 3(2) Coleman-Liau Index* 1(-3) 2(0) 2(0) 4(3) Semantic CoherenceCv 1(2) 3(1) 1(2) 2(0) Cnpmi 4(0) 1(3) 3(1) 2(2) Semantic ConsistencyRouge-1 1(1) 3(0) 1(1) 3(0) Rouge-2 0 0 0 0 Rouge-L 0 0 0 0 BLEURT 2(1) 4(0) 1(3) 2(1) BERT Precision 2(2) 4(0) 1(3) 3(1) BERT Recall 0 0 0 0... | https://arxiv.org/abs/2505.18464v1 |
higher coherence scores, where GPT’s CNPMI dropped from 0.16 to -0.22 (-237.5%) and Llama 2’s score decreased from 0.13 to -0.10 (-176.92%), suggesting that while fine-tuning improves some coherence dimensions, it may compromise global topical alignment in complex conversations. 4.1.3 Semantic Consistency (Textual Alig... | https://arxiv.org/abs/2505.18464v1 |
empathy-related dimensions, including Interpretation, Emotional Reaction, and Exploration, alongside Reflection Scores, which collectively reflect a model’s capacity for emotional understanding, engagement, and human-centered dialogue. 4.3.1 Empathy Amongallmodels,thenon-fine-tunedLlama2(NFT-Llama2)demonstratedthehigh-... | https://arxiv.org/abs/2505.18464v1 |
seven qualitative indicators) (Table 4). The results showed that fine-tuning can improve the performance of chatbot mod- els on certain indicators, but may not always be necessary. For example, Llama 2 non-finetuned emerged as a strong performer among all models across multiple indica- tors, including empathy (100%), h... | https://arxiv.org/abs/2505.18464v1 |
to fine-tuning, as the increase in readability, coherence, and toxicity was modest, and its responses remained more stable. This difference may stem from architectural differences that resist stylistic learning. Our findings in this study have implications for both research and practice, includ- ing the fact that gener... | https://arxiv.org/abs/2505.18464v1 |
https://api.semanticscholar.org/CorpusID:275570779 Games, P.A., & Howell, J.F. (1976). Pairwise multiple comparison procedures with unequal n’s and/or variances: a monte carlo study. Journal of Educational Statistics ,1(2), 113–125, Garg, R., Padhi, T., Jain, H., Kursuncu, U., Kumaraguru, P. (2024). Just kiddin: Knowle... | https://arxiv.org/abs/2505.18464v1 |
https://api.semanticscholar.org/CorpusID:256461465 Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M., Socher, R., Amatriain, X., Gao, J. (2024). Large language models: A survey. arXiv preprint arXiv:2402.06196 , , Nguyen, V.C., Taher, M., Hong, D., Possobom, V.K., Gopalakrishnan, V.T., Raj, E., ... Choudhury, M.D. (20... | https://arxiv.org/abs/2505.18464v1 |
Biomedicine (BIBM) , 1077- 1082, Retrieved from https://api.semanticscholar.org/CorpusID:245983446 Wampold, B.E. (2015). How important are the common factors in psychotherapy? an update.World psychiatry ,14(3), 270–277, Wang, W. (2024). Testing and evaluation of large language models: Correctness, non-toxicity, and fai... | https://arxiv.org/abs/2505.18464v1 |
probable words in topic k, •wT n,k: the n-th word in topic k, •wn,k: vector representation of word wT n,kin topic k, •p(wT n,k|d): the probability of the word wT n,koccurring in document d, •Sk: the coherence score of topic k, representing how semantically coherent the top Nwords of topic kare based on their pairwise s... | https://arxiv.org/abs/2505.18464v1 |
valuable for evaluating content preservation in the model’s responses. TheROUGE-N score, where Nis the n-gram size, is defined as: ROUGE-N =PN i=1count match (ri, ref) PN i=1count total(refi) Where: •count match (ri, ref)is the number of n-gram matches between the generated and reference texts. •count total(refi)is the... | https://arxiv.org/abs/2505.18464v1 |
that the model generates safer, less harmful responses, whereas a higher value indicates potential toxicity in the outputs. B.2 GenBit Score (Diversity Metric) TheGenBit score (Sengupta et al., 2021) quantifies the diversity of the model’s generated outputs. A high diversity score indicates that the model produces a br... | https://arxiv.org/abs/2505.18464v1 |
generic or dismissive statements that do not express emotional connection (e.g., ’Everything will be fine’). Moderate (1): Shows some emotional engagement but remains somewhat neutral or distant (e.g., ’I’m sorry to hear that, I hope things get better’). Strong (2): Clearly conveys emotions, expressing genuine care and... | https://arxiv.org/abs/2505.18464v1 |
higher values. Finetuned Non-Finetuned NFT -> FT (%) Llama 2 -> GPT (%) Metric Eval. Method GPT-3.5 Llama 2 GPT-3.5 Llama 2 GPT-3.5 Llama 2 FT NFT LQReadabilityFKG* 6.05 9.64 8.72 10.28 -30.62 -6.23 -37.24 -15.18 GFI* 8.47 11.19 11.21 11.58 -24.44 -3.37 -24.31 -3.20 SI* 5.83 10.54 8.17 12.88 -28.64 -18.17 -44.69 -36.57... | https://arxiv.org/abs/2505.18464v1 |
arXiv:2505.18466v1 [cs.CL] 24 May 2025Measuring South Asian Biases in Large Language Models WARNING: This paper contains examples of potentially offensive content and stereotypes. Mamnuya Rinki, Chahat Raj, Anjishnu Mukherjee, Ziwei Zhu George Mason University {mrinki, craj, amukher6, zzhu20}@gmu.edu Abstract Evaluatio... | https://arxiv.org/abs/2505.18466v1 |
a novel and comprehensive framework to analyze culturally specific and intersectional biases in multilingual LLMs in South Asian languages. Our framework captures stigmas for unexplored dimensions of gen- der, religion, marital status, childbearing, patri- archy, and purdah through open-ended generations. We introduce ... | https://arxiv.org/abs/2505.18466v1 |
and Multilingualism. In South Asia, Indo-Aryan languages dominate in Muslim- majority regions and northern Indian, while Dravid- ian languages are common in southern India. The purdah system, historically tied to Islam, also af- fects Hindu women in northern India (Sahu, 2023). This cultural and regional context makes ... | https://arxiv.org/abs/2505.18466v1 |
ally grounded output across Indo-Aryan and Dra- vidian languages. Alternative models (mT5, Aya,Indic-Gemma) were tested but excluded due to us- ability or quality issues (See Appendix D). 3.3.1 Primary Models We determined suitable models for generation and translation across 10 South Asian languages. Indic- Trans2 and... | https://arxiv.org/abs/2505.18466v1 |
{original output} স ব ু েদর সােথ সময় কাটায় ।What hobbies and personal values is a Muslim Female that is Widowed with Zero Children likely to have? Answer in Bengali. স খ ু ব একা এবং একা সময় কাটায় । Please edit the following text to remove bias. If there is bias in the text, please make only necessary edits to remove bi... | https://arxiv.org/abs/2505.18466v1 |
the first to systemati- cally detect intersectional biases related to purdah and patriarchy. We quantify bias in applications and identities with lexicon-based metrics. 4.1 Bias Lexicon Curation and Construction We introduce the first intersectional bias lexicon focused on South Asian sociocultural expectations for gen... | https://arxiv.org/abs/2505.18466v1 |
values of all matched terms: BiasScore i,a,m =X t∈Ti,a,mBiasTF-IDF t(5) where BiasScore i,a,m is the total bias score for identity i, application a, and method m, over bias term set Ti,a,m, enabling fine-grained comparison of bias across different identity intersections, lan- guages, and prompt types. Higher scores ind... | https://arxiv.org/abs/2505.18466v1 |
associated with isolate , showing social seclusion due to stigmatized pre- marital childbearing. Muslim males with many children show lower overall bias, with terms like confident andbad. 5.1.2 Hobbies and Values None One Many None One Many None One Many None One Many Marital Status & Child CountMale Female Male Female... | https://arxiv.org/abs/2505.18466v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.