nemo-gym

What it is and what it does

NeMo Gym is a framework for building and running evaluation environments where language models and agents interact with tasks to solve problems. It abstracts the evaluation loop—task datasets, agent harnesses, verifiers, and execution state—into modular, extensible components. You define or reuse an environment, configure your model (openai, anthropic, local via vLLM, or hosted providers), and run agents against tasks at scale, collecting trajectories and metrics.

The library is designed for teams needing reproducible, stateful evaluation across shared environments, or transitioning from evaluation into agent optimization and training. It includes a hub of popular benchmarks and agent harnesses, integrates with training frameworks, and provides CLI tools for discovery, validation, and diagnostics. With 31 runtime dependencies (ray, mlflow, wandb, fastapi, pydantic, aiohttp, and others), it trades breadth of built-in capability for a heavier dependency footprint.

Use it for:

• Evaluate agents on standardized benchmarks with reproducible verifiers and shared task datasets across teams.
• Collect training data by running agents on tasks at scale, generating trajectories for supervised fine-tuning or reinforcement learning.
• Benchmark agent skill impact by running the same tasks with different skill sets to isolate which capabilities drive performance.
• Run tool-using agents in isolated sandboxes to safely test code execution and external tool interactions.
• Transition from evaluation to training by using the same environment and verifier definitions with training frameworks.
• Diagnose evaluation failures to identify which tasks failed, why, and the highest-impact fixes.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

NeMo Gym provides infrastructure for building, running, and scaling evaluation and training environments where agents interact with tasks, datasets, verifiers, and execution state to solve problems.

Yes, if you need to evaluate or train agents in stateful environments at scale with reproducible verifiers and shared benchmarks. Active maintenance, permissive license, low install friction, and integration with major providers make it solid for agentic workflows. However, the strict Python 3.13.14+ requirement and heavy dependency footprint may conflict with existing projects. Early development status and evolving APIs warrant caution in production.

Install

nemo-gym on PyPI

pip install nemo-gym

uv add nemo-gym

poetry add nemo-gym

Installing nemo-gym

Before you install

Low install friction with a pure Python wheel. Active maintenance with a release 7 days old and 1114 repository stars. However, requires Python 3.13.14 or higher, which is a strict version floor that may conflict with existing projects on older Python versions.

License in practice

Apache License 2.0 is permissive—you can use, modify, and distribute NeMo Gym freely in commercial and private projects, provided you include the license and notice of changes. No viral copyleft obligations.

Quickstart

# Install NeMo Gym (requires Python 3.13.14+)
git clone git@github.com:NVIDIA-NeMo/Gym.git
cd Gym
uv venv --python 3.13.14 && source .venv/bin/activate
uv sync

# Configure model (e.g., OpenAI)
echo 'policy_base_url: https://api.openai.com/v1' > env.yaml
echo 'policy_api_key: <your-key>' >> env.yaml
echo 'policy_model_name: gpt-4.1-2025-04-14' >> env.yaml

# Start servers
gym env start --resources-server mcqa --model-type openai_model

# Run evaluation
gym eval run --no-serve --agent mcqa_simple_agent --input data/example.jsonl --output results.jsonl --limit 5 --num-repeats 1

Requires Python 3.13.14 or higher; earlier Python versions are not supported. Installation uses uv package manager, not pip. Requires API keys for model providers (openai, anthropic, or self-hosted via vLLM).

Verify before relying

• Whether the 31 runtime dependencies are all required for basic usage or only for advanced features.
• Performance characteristics when scaling to thousands of concurrent environments—actual throughput and latency benchmarks.
• Stability guarantees given the 'early development' status and 'evolving APIs' warning.
• Whether Windows WSL2 support is fully tested or experimental.

Package facts

Evidence: nemo_gym-0.5.0-py3-none-any.whl

Keywords: reinforcement-learning, RL, agent, agentic, LLM, large-language-models, training-data, rollout, trajectory, verification, verifier, reward, tool-calling, function-calling, data-collection, NeMo, NVIDIA

Tags

More Libraries packages

Further reading