Fine-tuning features
Greetings! Thank you for your work!
Could you please share the details of the fine tuning apart from the datasets used? Different QLora settings and ways of preparing datasets can significantly affect the generation results, which is what I am experiencing in my case.
@Whitepaper Thanks for your comment.
I'll be sharing the axolotl settings today, but keep in mind that tinkering with settings isn't the most crucial aspect and the training time isn't brief!
The quality of the dataset is paramount.
My primary motivation for creating this was to construct a model with a PL dataset and then generate or translate a new dataset into a high-quality synthetic dataset, or a dataset focused on a specific task.
The next level idea is to build multiple lora/qlora adapters that can be loaded on without having to restart your backend, similar to a hot swap.
However, I'm working on this alone and in my spare time. Any collaboration would be helpful.
Remember, the journey of a thousand miles begins with a single step. Keep calm, stay focused, and let's make great things happen together.
Thank you for your time and consideration.