CORAL Benchmarking Datasets

Benchmarking datasets used to train/evaluate Coral - **Positive pairs** were compiled from RNAInter, RPI488, RPI369, RPI2241, and RPI1807. We retained RNAInter interactions tagged as strong experimental evidence; from the RPI datasets we kept only positives. Proteins longer than 1,250 aa and RNAs longer than 6,000 nt were excluded, yielding 24,012 high-confidence positive interactions. - **Splits** come in five folds per strategy: redundant, pairwise at 80/60% and 50/30% RNA/protein identity, component-wise at the same two thresholds (25 folds total). Negative pairs are random RNA-protein re-pairings within the same split, balanced 1:1 with positives.SHA-256: def4ca6d99abcc6177d02f72237f3381b3a2833976a0fe30da04dcfa7dc75234