text to text implementation basing on https://github.com/salesforce/DocNLI
DatasetDict({ train: Dataset({ features: ['idx', 'inputs', 'targets'], num_rows: 942314 }) validation: Dataset({ features: ['idx', 'inputs', 'targets'], num_rows: 234258 }) test: Dataset({ features: ['idx', 'inputs', 'targets'], num_rows: 267086 }) })