The present training pipeline supports training the Unified Encoders with Instance level unfrozen. Having the option to unfreeze the Instance Level Encoders and jointly optimise both would be useful for more extensive evaluation of CrossOver's capabilities.