A user on Hugging Face discusses their attempt to fine-tune the Llama-3.1-8B model using QLoRA and a custom automation script called ChetGPT. The goal is to create a creative writing model rather than an instructional one, using Harry Potter novels as training data.
The user reports that following advice from GPT resulted in disastrous outcomes, with scenes being randomly cut or rearranged incorrectly. They share their Python code for `train_auto.py`, which includes features like automatic data analysis, learning amount determination based on data size, and generation tests to verify style learning.
The post serves as a request for feedback on whether this approach and the associated code are correct for achieving the desired creative writing capabilities.