Work for my project in Texas Luminescence, Chrom.io (text-to-color palette generator). This project did get scrapped because the final full stack product was pretty basic, but I learned a lot experimenting on the ML end!
The main thing I did was design and execute an RLVR fine-tuning protocol to improve color palette generation in small LLMs. Used deterministic reward metrics for aesthetic output, and serverless LoRA fine-tuning on the Fireworks.ai platform
Improved Qwen3-8B palette similarity scores by 13% over baseline, beating gpt-5-mini and gpt-4o!
Need to do a write-up of this project eventually because it was really interesting, but basically we used this paper + fine tuning to get SOTA performance.
Github Repository! (mainly look at /ml and /experiments)
(score_R is generated pallete similarity to ground truth palettes)
