NEED help fast!!

[sorobedio]

Hello everyone,

I’m looking for guidance on fine-tuning the LLaMA Instruct 3.2-1B model on my dataset to generate specific sequences. During training, the setup requires the user to input a sequence prompt (e.g., "generate sequence xx") and provide the associated target output ("xx"). At test time, however, the user should only need to provide the prompt or request, without needing to supply the target sequence.

Could you provide insights on setting up this training process and suggest a template or structure for it? This is quite urgent—thanks in advance for your help!

[rasbt]

Hi there,

During training, the setup requires the user to input a sequence prompt (e.g., "generate sequence xx") and provide the associated target output ("xx")

LitGPT doesn't allow interaction during training, but I think what you have in mind can be achieved by preparing such an instruction dataset in advance. For information on how to prepare the dataset, I recommend the "Preparing Custom Datasets From a JSON File" section in https://github.com/Lightning-AI/litgpt/blob/main/tutorials/prepare_dataset.md#preparing-custom-datasets-from-a-json-file

At test time, however, the user should only need to provide the prompt or request, without needing to supply the target sequence.

This should already be supported by

litgpt chat ...

and

litgpt generate ...

after training.

I am closing this issue as potentially addressed to clear out the issue tracker a bit, but please feel free to reopen if there are follow-up questions.