NeMo Gym

NeMo Gym is a framework for building reinforcement learning (RL) training environments for large language models (LLMs). It provides infrastructure to develop environments, scale rollout collection, and integrate seamlessly with your preferred training framework.

NeMo Gym is a component of the NVIDIA NeMo Framework, NVIDIA’s GPU-accelerated platform for building and training generative AI models.

🏆 Why NeMo Gym?

• Scaffolding and patterns to accelerate environment development: multi-step, multi-turn, and user modeling scenarios
• Contribute environments without expert knowledge of the entire RL training loop
• Test environment and throughput end-to-end independent of the RL training loop
• Interoperable with existing environments, systems and RL training frameworks
• Growing collection of training environments and datasets to enable Reinforcement Learning from Verifiable Reward (RLVR)

Important

NeMo Gym is currently in early development. You should expect evolving APIs, incomplete documentation, and occasional bugs. We welcome contributions and feedback - for any changes, please open an issue first to kick off discussion!

📋 Requirements

Hardware Requirements

NeMo Gym is designed to run on standard development machines:

• GPU: Not required for NeMo Gym framework operation
  • GPU may be needed for specific resource servers or model inference (see individual server documentation)
• CPU: Any modern x86_64 or ARM64 processor (e.g., Intel, AMD, Apple Silicon)
• RAM: Minimum 8 GB (16 GB+ recommended for larger environments)
• Storage: Minimum 5 GB free disk space for installation and basic usage

Software Requirements

• Operating System:
  • Linux (Ubuntu 20.04+, or equivalent)
  • macOS (11.0+ for x86_64, 12.0+ for Apple Silicon)
  • Windows (via WSL2)
• Python: 3.12 or higher
• Git: For cloning the repository
• Internet Connection: Required for downloading dependencies and API access

Additional Requirements

• API Keys: OpenAI API key with available credits (for the quickstart examples)
  • Other model providers supported (Azure OpenAI, self-hosted models via vLLM)
• Ray: Automatically installed as a dependency (no separate setup required)

🚀 Quick Start

Setup

# Clone the repository
git clone git@github.com:NVIDIA-NeMo/Gym.git
cd Gym

# Install UV (Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

# Create virtual environment
uv venv --python 3.12
source .venv/bin/activate

# Install NeMo Gym
uv sync --extra dev --group docs

Configure Your API Key

Create an env.yaml file that contains your OpenAI API key and the policy model you want to use. Replace your-openai-api-key with your actual key. This file helps keep your secrets out of version control while still making them available to NeMo Gym.

echo "policy_base_url: https://api.openai.com/v1
policy_api_key: your-openai-api-key
policy_model_name: gpt-4.1-2025-04-14" > env.yaml

Note

We use GPT-4.1 in this quickstart because it provides low latency (no reasoning step) and works reliably out-of-the-box. NeMo Gym is not limited to OpenAI models—you can use self-hosted models via vLLM or any OpenAI-compatible inference server. See the documentation for details.

Start Servers

Terminal 1 (start servers):

# Start servers (this will keep running)
config_paths="resources_servers/example_simple_weather/configs/simple_weather.yaml,\
responses_api_models/openai_model/configs/openai_model.yaml"
ng_run "+config_paths=[${config_paths}]"

Terminal 2 (interact with agent):

# In a NEW terminal, activate environment
source .venv/bin/activate

# Interact with your agent
python responses_api_agents/simple_agent/client.py

Collect Rollouts

Terminal 2 (keep servers running in Terminal 1):

# Create a simple dataset with one query
echo '{"responses_create_params":{"input":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"What is the weather in Seattle?"}]}}' > weather_query.jsonl

# Collect verified rollouts
ng_collect_rollouts \
    +agent_name=simple_weather_simple_agent \
    +input_jsonl_fpath=weather_query.jsonl \
    +output_jsonl_fpath=weather_rollouts.jsonl

# View the result
cat weather_rollouts.jsonl | python -m json.tool

This generates training data with verification scores!

Clean Up Servers

Terminal 1 with the running servers: Ctrl+C to stop the ng_run process.

What's Next?

Now that you can generate rollouts, choose your path:

• Use an existing training environment — Browse the Available Resource Servers below to find a training-ready environment that matches your goals.

• Build a custom training environment — Implement or integrate existing tools and define task verification logic. Get started with the Creating a Resource Server tutorial.

📖 Documentation

• Documentation - Technical reference docs
• Tutorials - Hands-on tutorials and practical examples

🤝 Community & Support

We'd love your contributions! Here's how to get involved:

• Report Issues - Bug reports and feature requests
• Contributing Guide - How to contribute code, docs, or new environments

📚 Citations

If you use NeMo Gym in your research, please cite it using the following BibTeX entry:

@misc{nemo-gym,
  title = {NeMo Gym: An Open Source Framework for Scaling Reinforcement Learning Environments for LLM},
  howpublished = {\url{https://github.com/NVIDIA-NeMo/Gym}},
  author={NVIDIA},
  year = {2025},
  note = {GitHub repository},
}

📦 Available Resource Servers

NeMo Gym includes a curated collection of resource servers for training and evaluation across multiple domains:

Table 1: Example Resource Servers

Purpose: Demonstrate NeMo Gym patterns and concepts.

Name	Demonstrates	Config	README
Multi Step	Instruction_Following example	example_multi_step.yaml	README
Simple Weather	Basic single-step tool calling	simple_weather.yaml	README
Stateful Counter	Session state management (in-memory)	stateful_counter.yaml	README

Table 2: Resource Servers for Training

Purpose: Training-ready environments with curated datasets.

Tip

Each resource server includes example data, configuration files, and tests. See each server's README for details.

Resource Server	Domain	Dataset	Description	Value	Config	Train	Validation	License
Google Search	agent	Nemotron-RL-knowledge-web_search-mcqa	Multi-choice question answering problems with search tools integrated	Improve knowledge-related benchmarks with search tools	config	✓	-	Apache 2.0
Math Advanced Calculations	agent	Nemotron-RL-math-advanced_calculations	An instruction following math environment with counter-intuitive calculators	Improve instruction following capabilities in specific math environments	config	✓	-	Apache 2.0
Workplace Assistant	agent	Nemotron-RL-agent-workplace_assistant	Workplace assistant multi-step tool-using environment	Improve multi-step tool use capability	config	✓	✓	Apache 2.0
Mini Swe Agent	coding	SWE-bench_Verified	A software development with mini-swe-agent orchestration	Improve software development capabilities, like SWE-bench	config	✓	✓	MIT
Instruction Following	instruction_following	Nemotron-RL-instruction_following	Instruction following datasets targeting IFEval and IFBench style instruction following capabilities	Improve IFEval and IFBench	config	✓	-	Apache 2.0
Structured Outputs	instruction_following	Nemotron-RL-instruction_following-structured_outputs	Check if responses are following structured output requirements in prompts	Improve instruction following capabilities	config	✓	✓	Apache 2.0
Equivalence Llm Judge	knowledge	Nemotron-RL-knowledge-openQA	Short answer questions with LLM-as-a-judge	Improve knowledge-related benchmarks like GPQA / HLE	config	✓	-	Apache 2.0
Mcqa	knowledge	Nemotron-RL-knowledge-mcqa	Multi-choice question answering problems	Improve benchmarks like MMLU / GPQA / HLE	config	✓	-	Apache 2.0
Math With Judge	math	Nemotron-RL-math-OpenMathReasoning	Math dataset with math-verify and LLM-as-a-judge	Improve math capabilities including AIME 24 / 25	config	✓	✓	Creative Commons Attribution 4.0 International