A synthetic benchmark for evaluating information-theoretic metrics in multimodal text–time series forecasting. The signal is a sine wave with randomly inserted constant-value segments, paired with three annotation categories per time point: correct , incorrect , and irrelevant (describes global signal properties). Ground-truth information content is known exactly by construction, enabling principled evaluation of mutual information estimators without human annotation quality judgement. Temporally split into train, validation, and test sets. Associated with the paper: When Does Text Inform? Benchmarking Information-Theoretic Metrics for Multimodal Time-Series Forecasting[Under Review].