EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

TARA: Task-Adaptive Rank Allocation for Efficient Large Language Model Fine-Tuning in Geo-Information Text Classification

Canhui Wang, Juntao Shen, Yicong Feng, Jin Huang, Yanwu Jing, Weiwei Chen, Wanqiang Zhang, Min Wang

ISPRS International Journal of Geo-Information · 18.08.2026 · 15 (8) · 372

DOI: 10.3390/ijgi15080372 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ISPRS International Journal of Geo-Information, online erschienen am 18.08.2026. Verfasst von Canhui Wang und 7 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Geo-information texts, including geospatial data-use regulations and Earth observation metadata, are central to data governance and compliance auditing in remote sensing ecosystems. Full fine-tuning of large pre-trained language models is often computationally impractical, while standard LoRA reduces cost but assigns a fixed rank to all adapted modules, ignoring differences across layers and projection types. This paper proposes TARA, a task-adaptive rank allocation method for LoRA-based fine-tuning. TARA assigns learnable importance scores to rank dimensions and uses Gumbel–Sigmoid sampling with the Straight-Through Estimator to learn discrete rank masks under a global sparsity constraint. We further construct RSRegulation, a geospatial regulatory compliance benchmark containing 4032 English-language samples derived from 168 clauses across seven regulatory and policy sources with clause-level data isolation. Across five random seeds, TARA achieves 95.30 ± 0.10% accuracy and 95.44 ± 0.10% F1 with a maximum trainable adapter budget of 1.57 M parameters. The learned soft allocation corresponds to approximately 0.38 M effective adapter parameters and 75.8% soft rank compression. Physical hard pruning reduces the deployed adapter to 0.086 M parameters while retaining 95.12 ± 0.11% accuracy and 95.21 ± 0.10% F1. Layer-wise analysis shows that value projections retain higher ranks than query projections under the current task and backbone, revealing a task-dependent non-uniform allocation pattern.

Zitieren

APA

Wang, C., Shen, J., Feng, Y., Huang, J., Jing, Y., Chen, W., Zhang, W., & Wang, M. (2026). TARA: Task-Adaptive Rank Allocation for Efficient Large Language Model Fine-Tuning in Geo-Information Text Classification. ISPRS International Journal of Geo-Information, 15(8), 372. https://doi.org/10.3390/ijgi15080372

Vancouver

Wang C, Shen J, Feng Y, Huang J, Jing Y, Chen W, et al. TARA: Task-Adaptive Rank Allocation for Efficient Large Language Model Fine-Tuning in Geo-Information Text Classification. ISPRS International Journal of Geo-Information. 2026;15(8):372. doi:10.3390/ijgi15080372

BibTeX
@article{Wang2026,
  author = {Wang, Canhui and Shen, Juntao and Feng, Yicong and Huang, Jin and Jing, Yanwu and Chen, Weiwei and Zhang, Wanqiang and Wang, Min},
  title = {TARA: Task-Adaptive Rank Allocation for Efficient Large Language Model Fine-Tuning in Geo-Information Text Classification},
  journal = {ISPRS International Journal of Geo-Information},
  year = {2026},
  volume = {15},
  number = {8},
  pages = {372},
  doi = {10.3390/ijgi15080372}
}
Kostenlos starten →