LLama 3.1 8B fine-tuning output repeats input

[dreamy dome]

was inference done in unsloth or externally?

make sure to use the exact same chat template for both training and inference: https://docs.unsloth.ai/basics/errors#running-in-unsloth-works-well-but-after-exporting-and-running-on-other-platforms-the-results-are-poo

[viscid pawn]

yeah, I had the same problem: https://discord.com/channels/1179035537009545276/1341946525575811112