Instruct tuning for lora/finetune?

Please correct me if I'm wrong, but it looks like the current examples for lora training all build a loss function around completion, which lines up with the lora example of using only the 'text' field from the jsonl dataset.

Are there any forks or plans to allow for instruct tuning, where the input is an input prompt, and the loss function is targeting the input/output pair?

Or did I miss something?

Thanks!

edit: example below:

{
"prompt": "[INST] Your input prompt here[/INST]",
"text": "The expected output result here"
}

Whereas it looks like the current lora process is:
{
"text": Predict what comes [next]
}

Start by reviewing the repository's current LoRA training examples and how they load JSONL records and construct the completion loss. Compare that flow with the prompt/text instruct-tuning example in this issue. Done would require a decided scope and documented or implemented support for prompt/input-output training, but no files or tests are named.