A model made with curated synthetic data and then KTO'd on a small curated set. Total time to train was 4 H100 hours.
I quite like the results this gave despite the dataset sizes involved. Its also a lot cheaper to iterate.
I plan to hand review the human data I've been using and slowly work that back into the datamix.
Additionally, planning to make a focused instruction following KTO set to improve system prompt adherance and steerability.
Use chatML and minP.
