Evaluating Expert Specialization in Mixture-of-Experts Antibody Language Models

Datasets for Evaluating Expert Specialization in Mixture-of-Experts Antibody Language Models, published in the ICLR 2026 FM4Science Workshop.

Files:

classification-data.tar.gz: Two classification datasets used to train classification models in Figure 4. The datasets are: HD-CoV and HD-CoV-Flu. CoV-specific sequences were obtained from CoV-AbDab, Flu-specific sequences were obtained from Wang et al., and healthy donor sequences were obtained from Neyestanak et al.

paired-TTE.tar.gz: Paired datasets used to train, test, and evaluate mixed models.

unpaired-TE.tar.gz: Unpaired datasets used to test and evaluate mixed models.

unpaired-train-shards.tar.gz: Unpaired dataset used to train models, sharded into 128 parquet files

Code: All code used for model training and evaluation is available under the MIT license on GitHub.

Model: Model weights for BALM-MoE are available on Hugging Face.