NVIDIA-NeMo/Skills

A project to improve skills of large language models

What it solves

Nemo-Skills provides a unified set of pipelines to improve the specific capabilities ("skills") of large language models. It streamlines the entire LLM development lifecycle, including synthetic data generation, model training, and comprehensive evaluation, while allowing developers to scale their workflows from a single local workstation to massive Slurm clusters with minimal configuration changes.

How it works

The project operates as a collection of pipelines that integrate various tools and processes:

  • Inference: It supports flexible LLM inference by switching between API providers, local servers, and large-scale cluster jobs. It integrates with high-performance engines like TensorRT-LLM, vLLM, sglang, and Megatron.
  • Evaluation: It provides a framework to run models against a wide array of benchmarks covering math (natural and formal), code, scientific knowledge, instruction following, long-context, tool-calling, multilingual capabilities, and vision-language tasks.
  • Training: It enables model training through integrations with NeMo-RL and verl.
  • Data Generation: It facilitates synthetic data generation (SDG) that can be scaled across thousands of GPUs.

Who it’s for

It is designed for AI researchers and developers who are building, tuning, and evaluating high-performance LLMs, particularly those working in environments that require scaling from local prototyping to large-scale GPU clusters.

Highlights

  • Scalable Infrastructure: Move from a local machine to a Slurm cluster with a one-line change.
  • Broad Benchmark Support: Built-in evaluations for diverse domains including formal math, coding, and scientific reasoning.
  • Flexible Inference: Compatibility with multiple industry-standard inference engines (vLLM, TensorRT-LLM, etc.).
  • Research-Driven: Used to create and release SOTA open-weight models and datasets like OpenReasoning and OpenMathReasoning.

Related

  • Project
    liruiw/GenSim

    GenSim is an LLM-powered pipeline that automatically generates robotic simulation tasks and environments to augment training data for multitask robotic policies.

  • Project
    vllm-project/guidellm

    An SLO-aware benchmarking and evaluation platform that simulates real-world workloads to optimize LLM inference performance, resource needs, and operational limits.

  • Project
    ServiceNow/PipelineRL

    A scalable asynchronous reinforcement learning framework for LLM agents that uses in-flight weight updates to maximize GPU utilization while maintaining on-policy data freshness.

  • Project
    PaddlePaddle/PaddleNLP

    A comprehensive LLM development suite based on PaddlePaddle that provides tools for efficient training, lossless compression, and high-performance inference across multiple hardware platforms.

  • Project
    modelscope/ms-swift

    A large model and multimodal fine-tuning and deployment framework that supports over 1,000 models and provides a full pipeline from training to deployment.