Support custom prompt loss for text generation task/instruction tuning

[viethoangtranduong]

In the context of instruction tuning, it's common in many research papers to set the "prompt_loss" to 0, as demonstrated in the Alpaca implementation here. OpenAI also provides similar capabilities.

The objective here is to prevent the model from learning the prompt and instead focus solely on the generation aspect. I would greatly appreciate it if the autotrain-advanced feature could:

• Set the "prompt_loss" to 0
• Allow customization of the "prompt_loss"

As we purely train on text for generative tasks (--text-column), we might need to decode some special token to separate prompt and response.

If you believe this feature would be relevant and can guide me to the pertinent files or resources, I'd be willing to take an initial attempt at implementing it. Thank you!

[viethoangtranduong]

@abhishekkrthakur Thank you!