I am new to huggingface and the task of text generation.
I’d like to to create my own train-eval loop to finetune text generation model based on the following checkpoint: dbmdz/german-gpt2 · Hugging Face
I found a very good tutorial on how to do that using the Trainer class (https://www.philschmid.de/fine-tune-a-non-english-gpt-2-model-with-huggingface).
The blog lead to some questions:
How exactly do we train text generation models under the boilerplate, if we want to use our own loop for full control instead of the fancy huggingface trainer?
How does the Trainer class decide how to train the model?
Where can I find out which training strategy is used by the Trainer?
Thank you for your help! =)