EntropyLong: Effective Long-Context Training via Predictive Uncertainty