Curated Paired Antibody Dataset for Masked Language Model Training (AbCDR-MLM)

This dataset contains processed and curated paired heavy-light antibody sequences derived from the Observed Antibody Space (OAS) for masked language model training and evaluation, as as used in the paper "Preferential CDR masking in paired antibody language models improves binding affinity prediction". Contents: Clustered training set: 1,617,948 paired sequences (MMseqs2, 95% sequence identity, 80% coverage) Validation set: 20,225 paired sequences Test set: 20,225 paired sequences IMGT-labeled CDR annotations for all sequences Citation If you use this dataset, please cite both the original OAS publication and our published article: Talaei, M., Walker, K. C., Hao, B. et al. (2026). Preferential CDR masking in paired antibody language models improves binding affinity prediction. Communications AI & Computing, 1, 7. https://doi.org/10.1038/s44488-026-00010-2