CRAFTS-FRT: A Pixel-Level Annotated Dataset for Fast Radio Transient Segmentation and Detection

To facilitate the development of FRTSearch, we constructed CRAFTS-FRT, a pixel-level annotated dataset for Fast Radio Transients (FRTs) derived from the Commensal Radio Astronomy FAST Survey. By leveraging the 19-beam receiver of the FAST telescope, we captured natural morphological variations across diverse viewing geometries. The dataset comprises 2,392 annotated instances, encompassing 2,115 pulses from 43 pulsars, 15 pulses from 3 Rotating Radio Transients (RRATs), and 262 bursts from two repeating Fast Radio Burst (FRB) sources (FRB 20121102 and FRB 20201124A). The raw data has been pre-processed into single-channel matrices with dimensions of $256 \times 8,192$ (frequency $\times$ time), covering a broad parameter space with Dispersion Measures (DM) ranging from 14.9 to 780 pc cm$^{-3}$ and Signal-to-Noise Ratios (S/N) from 5.2 to 102.9. For model development, the dataset is partitioned into a training set containing 1,805 instances (from 400 observation files) and a validation set containing 587 instances (from 100 observation files). Ultimately, CRAFTS-FRT establishes a standardized benchmark for instance segmentation tasks, supporting deep learning applications in radio transient astronomy.