I tried fine-tuning qwen3-14b on stock data a few times.
I tuned this model with textbook-based data + data I collected...
My agent quantized it and threw away the learned data.
I was planning to fine-tune it step by step and build up learning, but... I found out today that it was all deleted.
ㅜㅜ
So, in a hurry, I brought back the original model of the existing dataset and the dataset I'll be working on today and set it up again. It takes 33 hours.
The previous one took 27 hours...
dgx spark is good for fine-tuning, but it's slow.
The problem is that the previous one was lucky enough to be well fine-tuned, but there's no guarantee that I'll be lucky this time.
I'm backing up the quantized model and using it, but I have to start fine-tuning from scratch because of some randomness in the results. I just have to wait with a patient mind. ㅠㅠ
