Trained checkpoints for the paper "From Interpretability to Performance: Optimizing Retrieval Heads for Long-Context Language Models"