Skills

Why we included this project

NeMo Skills bundles the full LLM development loop into one set of pipelines: synthetic data generation, training through NeMo-RL or verl, and evaluation across a broad set of benchmarks covering math, code, instruction following, long-context, tool calling, and multilingual tasks. Teams that need to take a raw checkpoint to a benchmarked, production-ready model get that in one place. It also scales cleanly: start a job on a single local GPU, then move the same workload to a large Slurm cluster with a one-line config change, which suits small teams that prototype locally before scaling up. The evaluation layer is solid too, with parallelized jobs, self-hosted LLM judges, and reproducible recipes for models like Nemotron and OpenReasoning. If you are building or fine-tuning reasoning-focused models and want reproducible data, training, and eval workflows rather than a single-purpose tool, this is a strong fit.