← Rakesh Kariya
Writing
What a 16-image LoRA taught us about post-training an image model
September 24, 2026
A reproducible FLUX.2 Klein LoRA experiment: what style learning, checkpoint selection, serving, and a failed high-five prompt taught us.
Three GRPO runs to give a 1.5B model a thinking-budget dial. I got a 3.2x compression instead.
August 28, 2026
Three GRPO runs and 31 hours of A100 time trying to reproduce L1/LCPO on a 1.5B model. The budget dial never formed, but MATH-500 reasoning compressed 3.2x. A writeup of all three runs, including the two that failed.