Countering incomplete sentences

[saikrishnarallabandi]

We have been working with T5 models and found them to be quite effective. However, we have encountered an issue where the models consistently generate incomplete sentences. This behavior persists even with instruction-finetuned versions like FLAN. We have tried various approaches to address this problem including prompt engineering, specific finetuning, etc but the issue persists.

We would appreciate any guidance or suggestions on how to counter this behavior and improve the sentence generation quality of the T5 models and FLAN.