We finetuned GPT-J for title and teaser snippet generation.

We evaluated the following approaches for model tuning:

• Classic model fine tuning
• Multitask fine tuning
• Prompt tuning

Prompt tuning did not prove to be successfull.
Fine tuning and multitask fine tuning delivered promising results.

We finetuned GPT-J-6B-8bit for title and teaser generation with multitask finetuning.
We finetuned on two datasets with different sizes.