SciDiagBench
A benchmark dataset and evaluation method for assessing the quality of scientific diagram generation. It systematically divides scientific diagrams into 11 types, covering 10 domains with a total of 165 diagrams, and employs a human-preference-guided calibration method to generate type-specific evaluation criteria.
Timeline (last 90 days)
SciDiagBench is a new benchmark for evaluating the quality of scientific figure generation, whose core is to systematically categorize scientific figures and adopt a human-preference-guided calibration method to generate type-specific evaluation criteria.
The SciDiagBench dataset divides scientific figures into 11 types, covering 10 broad domains and containing 165 scientific figures in total.
SciDiagBench's figure types include graphical abstracts, technical roadmaps, and illustrative examples.
SciDiagBench currently targets the evaluation of scientific figure generation quality and does not yet cover figure understanding or interactive scenarios.
The specific basis for dividing SciDiagBench's 11 types, the concrete implementation of preference calibration, and the distribution of the 165 figures across the 10 domains are not elaborated in the original abstract.
SciDiagBench could be used by academic publishing platforms or preprint services for automated figure quality screening, but its practical deployment effectiveness still awaits further experimental validation.
All Facts (6)
SciDiagBench is a new benchmark for evaluating the quality of scientific figure generation, whose core is to systematically categorize scientific figures and adopt a human-preference-guided calibration method to generate type-specific evaluation criteria.
50%UnverifiedThe SciDiagBench dataset divides scientific figures into 11 types, covering 10 broad domains and containing 165 scientific figures in total.
50%UnverifiedSciDiagBench's figure types include graphical abstracts, technical roadmaps, and illustrative examples.
50%UnverifiedSciDiagBench currently targets the evaluation of scientific figure generation quality and does not yet cover figure understanding or interactive scenarios.
50%UnverifiedThe specific basis for dividing SciDiagBench's 11 types, the concrete implementation of preference calibration, and the distribution of the 165 figures across the 10 domains are not elaborated in the original abstract.
50%UnverifiedSciDiagBench could be used by academic publishing platforms or preprint services for automated figure quality screening, but its practical deployment effectiveness still awaits further experimental validation.
50%