Nemo Skills
Nemo Skills is NVIDIA's collection of pipelines for improving large language models: synthetic data generation, model training, and evaluation across benchmarks in math, code, science and instruction following, scaling from a workstation to a Slurm cluster.
Nemo Skills is a collection of pipelines for improving the skills of large language models, covering synthetic data generation, model training, and evaluation across a wide range of benchmarks. A project can start on a local workstation and move to a large-scale Slurm cluster by changing one configuration line.
Supported inference backends include TensorRT-LLM, vLLM, SGLang and Megatron, hosted on one or many nodes, and benchmark categories include natural-language and formal mathematics, code, scientific knowledge, instruction following, long-context, tool-calling, multilingual tasks, speech and audio, and vision-language models. Each evaluation can be parallelized across Slurm jobs, self-host its own LLM judge, or use custom prompts, and model training runs through NeMo-RL or verl. Benchmark rollouts are moving to a companion project called NeMo-Gym.