daVinci-kernel

RL framework from SII's GAIR lab for GPU kernel optimization: co-trains Skill Selection, Policy, and Skill Summary agents on a single backbone, admitting new skills into the library only after execution-verified speedups. daVinci-kernel-14B reaches 37.2% / 70.6% / 32.2% on KernelBench Levels 1/2/3 under the Fast₁ threshold, outperforming Dr.Kernel-14B by up to 46% on Level 3. Ships 8B/14B SFT and RL checkpoints plus the SFT dataset. Extends SII's daVinci line into systems-level code optimization.