Instruction fine tuning Flan-T5, non-english language

[holmbuar]

Are there any conventions, rules or templates when it comes to fine-tuning the Flan-T5 models in a non-English language?

How would the following example be formatted in e.g. German? Would it be a direct translation like

Q: Answer the following yes/no question by
reasoning step-by-step.
Could a dandelion suffer from hepatitis?
A: Hepatitis only affects organisms with livers.
Dandelions don’t have a liver. The answer is no.

to

F: Beantworten Sie die folgende Ja/Nein-Frage mit
Schritt für Schritt argumentieren.
Könnte ein Löwenzahn an Hepatitis leiden?
A: Hepatitis befällt nur Organismen mit Lebern.
Löwenzahn hat keine Leber. Die Antwort ist nein.

?

Similarly, in German, would a randomly picked FewShot pattern from Flan templates be:

 FewShotPattern(
      inputs="Ist \"{text1}\" in diesem Satz dasselbe wie \"{text2}\"?\n{context}\n\n{options_}",
      targets="{answer}",
      inputs_prefix="FRAGE: ",
      targets_prefix="ANT: ",
      x_y_delimiter="\n",
      example_separator="\n"),

?

(Pardon the Google translated texts)

[kushmuratoff]

I also need answer, @torlarse if you found answer, share with us