Data Cartography for Detecting Memorization Hotspots and Guiding Data Interventions in Generative Models
Fuente:
arXiv
Guardado en:
| Autores principales: | , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866916925884334080 |
|---|---|
| author | Patel, Laksh Shanbhag, Neel |
| author_facet | Patel, Laksh Shanbhag, Neel |
| contents | Modern generative models risk overfitting and unintentionally memorizing rare training examples, which can be extracted by adversaries or inflate benchmark performance. We propose Generative Data Cartography (GenDataCarto), a data-centric framework that assigns each pretraining sample a difficulty score (early-epoch loss) and a memorization score (frequency of ``forget events''), then partitions examples into four quadrants to guide targeted pruning and up-/down-weighting. We prove that our memorization score lower-bounds classical influence under smoothness assumptions and that down-weighting high-memorization hotspots provably decreases the generalization gap via uniform stability bounds. Empirically, GenDataCarto reduces synthetic canary extraction success by over 40\% at just 10\% data pruning, while increasing validation perplexity by less than 0.5\%. These results demonstrate that principled data interventions can dramatically mitigate leakage with minimal cost to generative performance. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_00083 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Data Cartography for Detecting Memorization Hotspots and Guiding Data Interventions in Generative Models Patel, Laksh Shanbhag, Neel Machine Learning Artificial Intelligence Computation and Language Modern generative models risk overfitting and unintentionally memorizing rare training examples, which can be extracted by adversaries or inflate benchmark performance. We propose Generative Data Cartography (GenDataCarto), a data-centric framework that assigns each pretraining sample a difficulty score (early-epoch loss) and a memorization score (frequency of ``forget events''), then partitions examples into four quadrants to guide targeted pruning and up-/down-weighting. We prove that our memorization score lower-bounds classical influence under smoothness assumptions and that down-weighting high-memorization hotspots provably decreases the generalization gap via uniform stability bounds. Empirically, GenDataCarto reduces synthetic canary extraction success by over 40\% at just 10\% data pruning, while increasing validation perplexity by less than 0.5\%. These results demonstrate that principled data interventions can dramatically mitigate leakage with minimal cost to generative performance. |
| title | Data Cartography for Detecting Memorization Hotspots and Guiding Data Interventions in Generative Models |
| topic | Machine Learning Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2509.00083 |