Ecosystem

View as Markdown

NeMo Gym is one of several specialized libraries in the NVIDIA NeMo platform. Alongside NeMo RL for post-training, NeMo Gym focuses on the environments agents act in — pairing a dataset, agent harness, and verifier so the same environment can drive both evaluation and training.

For what NeMo Gym does and when to use it, see the Overview and What NeMo Gym Provides. For how environments are implemented as composable servers, see Architecture.

NeMo Gym is also integrated with the broader agentic ecosystem. We would love your contribution! Open a PR to add an integration, or file an issue to share what would be valuable for you.

NeMo Gym Within NVIDIA NeMo

NVIDIA NeMo is a GPU-accelerated platform for training generative AI models and optimizing AI agents. It includes specialized libraries that span the model lifecycle end to end:

  • NeMo Curator: Prepares, filters, and deduplicates training data at scale.
  • NeMo Gym: Defines the environments — dataset, agent harness, and verifier — that agents are evaluated and trained in (this library).
  • NeMo RL: Runs scalable post-training (GRPO, DPO, and SFT) against those environments.

NeMo Gym’s role: NeMo Gym is where a task becomes an interactive environment. It standardizes how an agent interacts with the world (the agent harness), how task completion is scored (the verifier), and how per-task state is isolated — producing the rollouts and reward signals that training frameworks like NeMo RL consume, and the consistent scoring that evaluation depends on.

Environment Libraries

Seamlessly combine environments and benchmarks from other libraries alongside NeMo Gym environments, with access to over 1,000 community-contributed environments across integrated libraries.

Training Framework Libraries

Use environments for SFT and RL training. If you’re interested in integrating another training framework, see the Training Framework Integration Guide.

Agent Harnesses

Agent harnesses are available out of the box for evaluation and training. Some examples:

  • OpenHands - software engineering agent harness
  • Mini SWE Agent - software engineering agent harness
  • LangGraph - agent patterns built with LangGraph (reflection, orchestration, parallel thinking)

Depending on your workflow, you may also find these NeMo libraries useful across data preparation, evaluation and training, and deployment.

Data

LibraryPurpose
NeMo CuratorScalable data preprocessing and curation
NeMo Data DesignerGenerate synthetic training data from scratch or seed examples
NeMo Safe SynthesizerGenerate privacy-safe synthetic copies of sensitive datasets
NeMo AnonymizerDetect and replace sensitive data

Evaluation & Training

LibraryPurpose
NeMo GymEvaluate and improve models and agents using environments (this project)
NeMo EvaluatorModel evaluation and benchmarking
NeMo RLScalable post-training with GRPO, DPO, and SFT
NeMo Megatron-BridgePretraining and fine-tuning with Megatron-Core
NeMo AutoModelPyTorch native training for Hugging Face models

Deployment

LibraryPurpose
NeMo Agent ToolkitConnect and optimize teams of AI agents
NeMo GuardrailsEnforce safety and policy rules
NeMo RetrieverDocument extraction and retrieval for RAG pipelines