Can I seed a DataCollatorForLanguageModeling to control what word will be masked?
DataCollatorForLanguageModeling