mlsanityData to 7 Oct 2026

Data Similarity is Not Enough to Explain Language Model Performance

arXiv 2311.09006 ↗First posted 15 Nov 2023cs.CL cs.LGcited by 9

Not on the map. Why?

Abstract (arXiv)

Large language models achieve high performance on many but not all downstream tasks. The interaction between pretraining data and task data is commonly assumed to determine this variance: a task with data that is more similar to a model's pretraining data is assumed to be easier for that model. We test whether distributional and example-specific similarity measures (embedding-, token- and model-based) correlate with language model performance through a large-scale comparison of the Pile and C4 pretraining datasets with downstream benchmarks. Similarity correlates with performance for multilingual datasets, but in other benchmarks, we surprisingly find that similarity metrics are not correlated with accuracy or even each other. This suggests that the relationship between pretraining data and downstream tasks is more complex than often assumed.

Key facts. Cited by 9 papers in the map (data through 7 Oct 2026). Builds on 40 of its 54 references that are in the map. First posted on arXiv 15 Nov 2023.

Builds on

40 of 54 references.

References · 40

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  4. 2018GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understandingcited by 4,854
  5. 2022Scaling Instruction-Finetuned Language Modelscited by 2,997
  6. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,346
  7. 2022Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language modelscited by 1,970
  8. 2022BLOOM: A 176B-Parameter Open-Access Multilingual Language Modelcited by 1,968
  9. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  10. 2020Don't Stop Pretraining: Adapt Language Models to Domains and Taskscited by 1,548
  11. 2020WILDS: A Benchmark of in-the-Wild Distribution Shiftscited by 1,366
  12. 2018XNLI: Evaluating Cross-lingual Sentence Representationscited by 913
28 more
  1. 2023The Flan Collection: Designing Data and Methods for Effective Instruction Tuningcited by 734
  2. 2021Deduplicating Training Data Makes Language Models Bettercited by 705
  3. 2020Estimating Training Data Influence by Tracing Gradient Descentcited by 565
  4. 2021Datasets: A Community Library for Natural Language Processingcited by 526
  5. 2021Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpuscited by 481
  6. 2021MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontierscited by 466
  7. 2021Sentence-T5: Scalable Sentence Encoders from Pre-trained Text-to-Text Modelscited by 462
  8. 2022Large Language Models Struggle to Learn Long-Tail Knowledgecited by 452
  9. 2020Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamicscited by 423
  10. 2023Data Selection for Language Models via Importance Resamplingcited by 322
  11. 2020Pretrained Transformers Improve Out-of-Distribution Robustnesscited by 293
  12. 2021Understanding Dataset Difficulty with V-Usable Informationcited by 288
  13. 2021Few-shot Learning with Multilingual Language Modelscited by 278
  14. 2022Data Contamination: From Memorization to Exploitationcited by 234
  15. 2022Demystifying Prompts in Language Models via Perplexity Estimationcited by 215
  16. 2022Out-of-Distribution Detection and Selective Generation for Conditional Language Modelscited by 193
  17. 2023A Pretrainer's Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, & Toxicitycited by 180
  18. 2020Adversarial Filters of Dataset Biasescited by 159
  19. 2020With Little Power Comes Great Responsibilitycited by 138
  20. 2022Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIPcited by 97
  21. 2023Revisiting Out-of-distribution Robustness in NLP: Benchmark, Analysis, and LLMs Evaluationscited by 96
  22. 2022Language Contamination Helps Explain the Cross-lingual Capabilities of English Pretrained Modelscited by 90
  23. 2020A Mathematical Exploration of Why Language Models Help Solve Downstream Taskscited by 71
  24. 2022ORCA: Interpreting Prompted Language Models via Locating Supporting Data Evidence in the Ocean of Pretraining Datacited by 34
  25. 2017Replicability Analysis for Natural Language Processing: Testing Significance with Multiple Datasetscited by 32
  26. 2021Rissanen Data Analysis: Examining Dataset Characteristics via Description Lengthcited by 22
  27. 2021Comparing Text Representations: A Theory-Driven Approachcited by 5
  28. 2023Gradient-Based Automated Iterative Recovery for Parameter-Efficient Tuningcited by 5

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 4; 2025: 3; 2026: 2.

9 papers.

  1. 2024No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performancecited by 75
  2. 2024Unleashing the Power of Data Tsunami: A Comprehensive Survey on Data Assessment and Selection for Instruction Tuning of Language Modelscited by 29
  3. 2024LMD3: Language Model Data Density Dependencecited by 7
  4. 2025Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (short paper)cited by 7
  5. 2024The Right Time Matters: Data Arrangement Affects Zero-Shot Generalization in Instruction Tuningcited by 4
  6. 2025Interrogating LLM design under a fair learning doctrinecited by 1
  7. 2025Rewriting History: A Recipe for Interventional Analyses to Study Data Effects on Model Behaviorcited by 0
  8. 2026The Role of Dataset Linguistic Structure in the Cultural Awareness of Large Language Modelscited by 0
  9. 2026Mind Your Tone: Does Tone Alter LLM Performance?cited by 0