Teaching NVIDIA Nemotron 3.5 Lightning to route code reviews

CodeRabbit, NVIDIA, and Baseten tested whether the compact NVIDIA Nemotron 3.5 Lightning model could be post-trained to make code-review routing decisions, using supervised fine-tuning on distilled examples followed by reinforcement learning with verifiable rewards scored against CodeRabbit’s routing policy. Training drew from 39,566 public-repository examples, with a filtered 9,996-example fine-tuning set and a repository-separated 1,000-task evaluation set to avoid leakage. Supervised fine-tuning improved exact route agreement from 75.8% for a GPT-class baseline to 80.4%, while adding RLVR raised it marginally to 80.7% and improved Cohen’s kappa output agreement from 0.461 to 0.544. The experiment took under three hours and cost less than $100, with the resulting model served on a single A100 GPU at measured throughput of about 315 aggregate output tokens per second across eight concurrent requests. Compared with the baseline’s estimated $2.34 inference cost for the evaluation workload, the tuned model cost an estimated $1.16 at peak throughput, generated 63.4% fewer tokens, and offered a projected 50.4% cost reduction, suggesting it could be useful for other high-volume, narrowly scoped CodeRabbit tasks.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.