NeMo Gym

Requirements • Quick Start • Available Environments • Documentation & Resources • Community & Support • Citations

NeMo Gym is a library 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.

🏆 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 environments 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 for 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!

🔗 Ecosystem

NeMo Gym is part of NVIDIA NeMo, NVIDIA's GPU-accelerated platform for building and training generative AI models. NeMo Gym integrates with a growing number of RL training frameworks and environment libraries; see the Ecosystem page for full details and tutorials.

Training Frameworks:NeMo RL • OpenRLHF • Unsloth • more →

Environment Libraries:Reasoning Gym • Aviary • more →

📋 Requirements

NeMo Gym is designed to run on standard development machines:

Hardware Requirements	Software Requirements
GPU: Not required for NeMo Gym library operation • GPU may be needed for specific resources servers or model inference (see individual server documentation)	Operating System: • Linux (Ubuntu 20.04+, or equivalent) • macOS (11.0+ for x86_64, 12.0+ for Apple Silicon) • Windows (via WSL2)
CPU: Any modern x86_64 or ARM64 processor (e.g., Intel, AMD, Apple Silicon)	Python: 3.12 or higher
RAM: Minimum 8 GB (16 GB+ recommended for larger environments)	Git: For cloning the repository
Storage: Minimum 5 GB free disk space for installation and basic usage	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

Install NeMo Gym, start the servers, and collect your first verified rollouts for RL training.

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/v1policy_api_key: your-openai-api-keypolicy_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_single_tool_call/configs/example_single_tool_call.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 environmentsource .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 queryecho'{"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=example_single_tool_call_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.

Next Steps

Now that you can generate rollouts, choose your path:

• Start training — Train models using NeMo Gym with your preferred RL framework. See the Training Tutorials.

• Use an existing environment — Browse the Available Environments below to find an environment that matches your goals.

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

📦 Available Environments

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

Example Environment Patterns

Purpose: Demonstrate NeMo Gym patterns and concepts.

Name	Demonstrates	Config	README
Multi Step	Multi-step tool calling	example_multi_step.yaml	README
Session State Mgmt	Session state management (in-memory)	example_session_state_mgmt.yaml	README
Single Tool Call	Basic single-step tool calling	example_single_tool_call.yaml	README

Environments for Training & Evaluation

Purpose: Training-ready environments with curated datasets.

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

The Dataset column links to publicly available datasets (e.g., on HuggingFace). A - means the train/validation data has not been publicly released yet, or that it is procedurally generated using a provided script. If no data is released yet, new data can be generated, or the environment can be used as a reference. Each server includes 5 example tasks in data/example.jsonl.

Resources Server	Domain	Description	Value	Train	Validation	License	Config	Dataset
Aalcr	other	-	-	-	-	-	aalcr.yaml	-
Abstention	rlhf	Train models to abstain when unsure using three-tier reward on HotPotQA with LLM judge	Improve calibration by rewarding abstention over incorrect answers	✓	✓	Creative Commons Attribution-ShareAlike 4.0 International	abstention.yaml	-
Arc Agi	knowledge	Solve puzzles designed to test intelligence. See https://arcprize.org/arc-agi.	Improve puzzle-solving capabilities.	-	✓	-	arc_agi.yaml	-
Aviary	agent	Multi-hop question answering on the HotPotQA dataset with Wikipedia search	Improve knowledge and agentic capability	✓	✓	Apache 2.0	hotpotqa_aviary.yaml	-
Aviary	math	GSM8k benchmark with calculator tool	Test math and agentic capability	✓	✓	Apache 2.0	gsm8k_aviary.yaml	-
Calendar	agent	Multi-turn calendar scheduling dataset. User states events and constraints in natural language; model schedules events to satisfy all constraints.	Improve multi-turn instruction following capabilities	✓	✓	Apache 2.0	calendar.yaml	Nemotron-RL-agent-calendar_scheduling
Calendar	agent	Multi-turn calendar scheduling dataset. User states events and constraints in natural language; model schedules events to satisfy all constraints.	Improve multi-turn instruction following capabilities	✓	✓	Creative Commons Attribution 4.0 International	calendar_v2.yaml	Nemotron-RL-Instruction-Following-Calendar-v2
Circle Click	other	Click on circles in images	-	-	-	-	circle_click.yaml	-
Code Gen	coding	Model must submit the right code to solve a problem	Improve competitive coding capabilities	✓	✓	Apache 2.0	code_gen.yaml	nemotron-RL-coding-competitive_coding
Equivalence Llm Judge	agent	Short bash command generation questions with LLM-as-a-judge	Improve foundational bash and IF capabilities	✓	✓	GNU General Public License v3.0	nl2bash-equivalency.yaml	-
Equivalence Llm Judge	knowledge	Short answer questions with LLM-as-a-judge	Improve knowledge-related benchmarks like GPQA / HLE	-	-	-	equivalence_llm_judge.yaml	-
Ether0	knowledge	ether0 chemistry benchmark verifiers	Evalutate chemistry knowledge and reasoning with ether0 benchmark	-	✓	-	ether0.yaml	-
Finance Sec Search	agent	SEC EDGAR filing search for financial analysis questions	Enable LLMs to search and analyze SEC filings	-	-	-	finance_sec_search.yaml	-
Genrm Compare	rlhf	GenRM pairwise comparison for RLHF training	Compare multiple candidate responses using GenRM model	-	-	-	genrm_compare.yaml	-
Google Search	agent	Multi-choice question answering problems with search tools integrated	Improve knowledge-related benchmarks with search tools	✓	-	Apache 2.0	google_search.yaml	Nemotron-RL-knowledge-web_search-mcqa
Gpqa Diamond	knowledge	GPQA Diamond multiple-choice question answering problems	Evaluate graduate-level scientific reasoning via MCQ verification	✓	-	MIT	gpqa_diamond.yaml	-
Instruction Following	instruction_following	Instruction following datasets targeting IFEval and IFBench style instruction following capabilities	Improve IFEval and IFBench	✓	-	Apache 2.0	instruction_following.yaml	Nemotron-RL-instruction_following
Jailbreak Detection	safety	Jailbreak detection with Nemotron judge + combined reward	-	-	✓	-	jailbreak_detection_nemotron_combined_reward_tp8.yaml	-
Math Advanced Calculations	agent	An instruction following math environment with counter-intuitive calculators	Improve instruction following capabilities in specific math environments	✓	-	Apache 2.0	math_advanced_calculations.yaml	Nemotron-RL-math-advanced_calculations
Math Formal Lean	math	Lean4 formal proof verification environment	Improve formal theorem proving capabilities	✓	-	Apache 2.0	nemotron_clean_easy.yaml	-
Math Formal Lean	math	Lean4 formal proof verification environment	Improve formal theorem proving capabilities	✓	-	Apache 2.0	nemotron_first_try_hard.yaml	-
Math Formal Lean	math	Lean4 formal proof verification environment	Improve formal theorem proving capabilities	✓	-	Apache 2.0	nemotron_medium_500.yaml	-
Math Formal Lean	math	Lean4 formal proof verification environment	Improve formal theorem proving capabilities	✓	-	Apache 2.0	nemotron_very_easy.yaml	-
Math Formal Lean	math	Lean4 formal proof verification environment	Improve formal theorem proving capabilities	✓	-	MIT	math_formal_lean.yaml	-
Math Formal Lean	math	Lean4 formal proof verification environment with multi-turn self-correction	Improve formal theorem proving capabilities	✓	-	MIT	math_formal_lean_multi_turn.yaml	-
Math With Code	math	Model solves competitive math problems using simple calculator tools	Improve math and simple tool use capabilities	✓	-	Apache 2.0	math_with_code.yaml	-
Math With Judge	math	DAPO17k math dataset with math-verify	Improve math capabilities including AIME 24 / 25	✓	✓	Apache 2.0	dapo17k.yaml	-
Math With Judge	math	MathStackOverflow math dataset with math-verify	Improve math capabilities including AIME 24 / 25	✓	✓	Creative Commons Attribution-ShareAlike 4.0 International	math_stack_overflow.yaml	Nemotron-RL-math-stack_overflow
Math With Judge	math	OpenMathReasoning math dataset with math-verify and LLM-as-a-judge	Improve math capabilities including AIME 24 / 25	✓	✓	Creative Commons Attribution 4.0 International	math_with_judge.yaml	Nemotron-RL-math-OpenMathReasoning
Mcqa	knowledge	Multi-choice question answering problems	Improve benchmarks like MMLU / GPQA / HLE	✓	✓	Apache 2.0	mcqa.yaml	Nemotron-RL-knowledge-mcqa
Mini Swe Agent	coding	A software development with mini-swe-agent orchestration	Improve software development capabilities, like SWE-bench	✓	✓	MIT	mini_swe_agent.yaml	SWE-Gym
Multichallenge	knowledge	Targets inference memory, instruction retention, version editing, and self-coherence.	Improve complex multi-turn conversational capability	✓	-	Creative Commons Attribution 4.0 International	multichallenge_nrl.yaml	Nemotron-RL-Instruction-Following-MultiTurnChat-v1
Newton Bench	math	Scientific law discovery tasks through agentic experimentation across 12 physics domains	Improve science, reasoning, and tool use capabilities	✓	-	Apache 2.0	newton_bench.yaml	-
Ns Tools	agent	NeMo Skills tool execution with math verification	-	-	-	-	ns_tools.yaml	-
Nvarc	knowledge	ARC-AGI inductive mode: model outputs Python code with transform()	Improve ARC-AGI puzzle-solving by inducing executable transformation programs	✓	✓	Apache 2.0	inductive.yaml	-
Nvarc	knowledge	ARC-AGI transductive mode: model outputs grid directly	Improve ARC-AGI puzzle-solving by directly predicting transformed grids	✓	✓	Apache 2.0	transductive.yaml	-
Openenv	agent	Echo environment via OpenEnv (MCP). Echoes messages back with length-based rewards.	-	-	-	-	openenv_echo.yaml	-
Openenv	coding	Python code execution environment via OpenEnv. Executes code and returns stdout/stderr.	-	-	-	-	openenv_coding.yaml	-
Openenv	games	Maze navigation environment via OpenEnv. Agent navigates an 8x8 grid to find the exit.	-	-	-	-	openenv_maze.yaml	-
Over Refusal Detection		-	-	✓	-	TBD	over_refusal_detection.yaml	-
Proof Genselect	math	Pairwise proof selection with binary correctness reward	-	-	-	-	proof_genselect.yaml	-
Proof Judge	math	Theorem proving with verifier + meta-verifier judge (combined env)	-	-	-	-	proof_judge.yaml	-
Proof Verification	math	Proof verification scored against ground truth and meta-verifier agreement	-	-	-	-	proof_verification.yaml	-
Rdkit Chemistry	knowledge	Molecular chemistry question answering: calculate properties of SMILES. Includes a mix of tool-use (python + rdkit) and no-tool-use questions.	Improve molecular reasoning and SMILES parsing.	✓	-	TBD	rdkit_chemistry.yaml	-
Reasoning Gym	knowledge	LangGraph orchestrator agent compatible with resource servers that do not use tools; enables diverse agent training data and test time scaling vs a simple agent, extensible to use tools or other agent architectures	Iterative test time scaling for improved performance in reasoning tasks	✓	-	Apache 2.0	orchestrator_agent.yaml	-
Reasoning Gym	knowledge	LangGraph parallel thinking agent compatible with resource servers that do not use tools; enables diverse agent training data and test time scaling vs a simple agent, extensible to use tools or other agent architectures	Iterative test time scaling for improved performance in reasoning tasks	✓	-	Apache 2.0	parallel_thinking_agent.yaml	-
Reasoning Gym	knowledge	LangGraph reflection agent compatible with resource servers that do not use tools; provides iterative reflection for diverse agent training data and test time scaling, extensible to use tools or other agent architectures	Iterative test time scaling for improved performance in reasoning tasks	✓	-	Apache 2.0	reflection_agent.yaml	-
Reasoning Gym	knowledge	LangGraph ReWOO agent compatible with resource servers that do not use tools; enables diverse agent training data and test time scaling vs a simple agent, extensible to use tools or other agent architectures	Iterative test time scaling for improved performance in reasoning tasks	✓	-	Apache 2.0	rewoo_agent.yaml	-
Reasoning Gym	knowledge	Over 100 tasks including algebra, arithmetic, computation, cognition, geometry, graph theory, logic, and many common games.	Improve robustness, generalization, broad knowledge and reasoning	✓	-	Creative Commons Attribution 4.0 International	reasoning_gym.yaml	Nemotron-RL-ReasoningGym-v1
Ruler	other	-	-	-	-	-	ruler.yaml	-
Single Step Tool Use With Argument Comparison	agent	Conversational tool-use RL from expert trajectories; behavior cloning per step across auth, lookup, and servicing domains.	-	✓	✓	Creative Commons Attribution 4.0 International	single_step_tool_use_with_argument_comparison.yaml	Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1
Single Step Tool Use With Argument Comparison	agent	General function-calling RL dataset using expert trajectories; behavior cloning to match expert tool calls per step.	-	✓	✓	Creative Commons Attribution 4.0 International	toolcall_schema_single_step_tool_use_with_argument_comparison.yaml	Nemotron-RL-Agentic-Function-Calling-Pivot-v1
Single Step Tool Use With Argument Comparison	agent	GitHub-issue dataset for software-engineering agents; refactored from SWE-Gym and SWE-Bench-Verified for NeMo Gym.	-	✓	✓	Creative Commons Attribution 4.0 International	swe_pivot_single_step_tool_use_with_argument_comparison.yaml	Nemotron-RL-Agentic-SWE-Pivot-v1
Single Step Tool Use With Argument Comparison	agent	The model must output the next correct call in a given trajectory involving search tools.	Improve agentic search capability.	✓	✓	Apache 2.0	search_pivot_single_step_tool_use_with_argument_comparison.yaml	-
Spider2 Lite	coding	Text-to-SQL with execution-based evaluation on Spider 2.0-Lite (135 SQLite tasks). Binary reward based on result-set equivalence.	Improve text-to-SQL capabilities for real-world enterprise queries using execution-based binary reward without an LLM judge.	-	✓	-	spider2_lite.yaml	-
Structured Outputs	instruction_following	Check if responses are following structured output requirements in prompts	Improve instruction following capabilities	✓	✓	Apache 2.0	structured_outputs_json.yaml	Nemotron-RL-instruction_following-structured_outputs
Structured Outputs	instruction_following	Check if responses are following structured output requirements in prompts	Improve instruction following capabilities	✓	✓	Apache 2.0	structured_outputs_json_yaml_xml_v1.yaml	-
Swerl Gen	coding	Running sandboxed evaluation for SWE-style tasks (either patch generation or reproduction test generation)	Improve SWE capabilities useful for benchmarks like SWE-bench	✓	✓	Apache 2.0	swerl_gen.yaml	-
Swerl Llm Judge	coding	SWE-style multiple-choice LLM-judge tasks scored via ... choice.	Improve SWE capabilities useful for benchmarks like SWE-bench	✓	✓	MIT	swerl_llm_judge.yaml	-
Tavily Search	agent	Model uses search tools to satisfy a user query.	Measure agentic search capability	✓	✓	Apache 2.0	tavily_search_judge_vllm_model.yaml	-
Terminus Judge	agent	single-step terminal based task (rubrics v4 judge prompt)	Improve on terminal-style tasks	✓	✓	Apache 2.0	terminus_judge.yaml	-
Terminus Judge	agent	single-step terminal based task (simple judge prompt)	Improve on terminal-style tasks	✓	✓	Apache 2.0	terminus_judge_simple.yaml	-
Terminus Judge	agent	single-step terminal based task (string similarity only)	Improve on terminal-style tasks	✓	✓	Apache 2.0	terminus_judge_string_only.yaml	-
Text To Sql	coding	Text-to-SQL generation with LLM-as-a-judge equivalence checking	Improve text-to-SQL capabilities across multiple dialects	-	-	-	text_to_sql.yaml	-
Vlm Eval Kit	other	-	Measure VLM capabilities	-	✓	-	MMBench_DEV_EN_V11.yaml	-
Vlm Eval Kit	other	-	Measure VLM capabilities	-	✓	-	OCRBench.yaml	-
Vlm Eval Kit	other	Run all supported VLMEvalKit benchmarks.	Measure VLM capabilities	-	✓	-	vlm_eval_kit.yaml	-
Workplace Assistant	agent	Workplace assistant multi-step tool-using environment	Improve multi-step tool use capability	✓	✓	Apache 2.0	workplace_assistant.yaml	Nemotron-RL-agent-workplace_assistant
Xlam Fc	agent	Salesforce xlam-function-calling-60k tool calling tasks	Improve tool-calling capabilities	✓	✓	Apache 2.0	xlam_fc.yaml	-
Xstest	safety	XSTest safety benchmark - exaggerated safety (over-refusal) evaluation	Evaluate model safety calibration between helpfulness and harmlessness	-	-	-	xstest.yaml	-
Xstest	safety	XSTest safety benchmark - exaggerated safety (over-refusal) evaluation	Evaluate model safety calibration between helpfulness and harmlessness	-	-	-	xstest_string_match.yaml	-

📖 Documentation & Resources

• Documentation - Technical reference docs
• Training Tutorials - Train with NeMo Gym environments
• API Reference - Complete class and function reference

🤝 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, new environments, or training framework integrations

📚 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 Library for Scaling Reinforcement Learning Environments for LLM},
howpublished = {\url{https://github.com/NVIDIA-NeMo/Gym}},
author={NVIDIA},
year = {2025},
note = {GitHub repository},
}