Skip to main content

​
What is an Environment Card?

An environment card is a structured description of your environment - similar to a model card, but for environments. It tells users what your environment evaluates, what tools are available, how rewards work, and what compute or safety considerations apply. On OpenReward, the environment card is your environment’s README.md. It renders on your environment’s overview page and is the first thing users see when they visit your environment.

​
Template

Below is a template you can copy and fill in for your own environment. Each section is explained in detail afterwards. 
# YourEnvironmentName

[![OpenReward Environment](https://img.shields.io/badge/%E2%AD%90%20OpenReward-Environment-f7e6cc)](https://openreward.ai/YourOrg/YourEnvironmentName)

## Description

<!-- 2-4 sentences: what does this environment evaluate? What domain does it cover?
     Mention the verification method and the source dataset. -->

## Capabilities

<!-- Bulleted list of skills or capabilities the environment tests. -->

- Capability 1
- Capability 2
- Capability 3

## Compute Requirements

<!-- Does the environment need a sandbox? GPU? Extra memory?
     If it runs without a sandbox, say so. -->

## License

<!-- License that covers the use of the environment, as well as any underlying data or code it depends on.
     Link to the license text. If the environment and its dependencies have different licenses, list both. -->

## Tasks

<!-- How many tasks are there? What splits (train/test)?
     Briefly describe the structure of each task. -->

## Reward Structure

<!-- How are rewards computed?
     Sparse or dense? Binary or continuous?
     Is verification programmatic or does it use an LLM grader? Does it use rubrics?
     If there are multiple validation steps, list them in order. -->

## Data

<!-- Where does the task data come from? How is it stored?
     Link to the source dataset if applicable. -->

## Tools

<!-- Explain the tools available to the agent. -->

## Time Horizon

<!-- Single-turn or multi-turn?
     If multi-turn, roughly how many tool calls does a typical task require? -->

## Environment Difficulty

<!-- Solve rates, baseline model performance, or other difficulty statistics. -->

## Other Environment Requirements

<!-- Any external dependencies: API keys, secrets, third-party services. -->

## Safety

<!-- Does the agent have access to external systems (network, file system, APIs)?
     Are there dual-use risks in the domain (e.g. chemistry, cybersecurity)?
     Is there a possibility of goal misspecification?
     What mitigations are in place? -->

## Citations

<!-- BibTeX entries for the environment itself and any underlying datasets or papers. -->

```bibtex
@dataset{YourCitation,
  author    = {Your Name or Team},
  title     = {YourEnvironmentName},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/YourOrg/YourEnvironmentName}
}
```

​
Section Reference

​
Badge + Title

# YourEnvironmentName

[![OpenReward Environment](https://img.shields.io/badge/%E2%AD%90%20OpenReward-Environment-f7e6cc)](https://openreward.ai/YourOrg/YourEnvironmentName)
The badge links to your environment on OpenReward and gives viewers on GitHub an immediate visual indicator that this is an OpenReward environment. Replace YourOrg/YourEnvironmentName with your environment’s path.

​
Description

A short paragraph (2-4 sentences) explaining what the environment evaluates. Cover:
  • The domain or task type (e.g. molecular generation, code repair, math reasoning)
  • How answers are verified (e.g. RDKit validation, unit tests, exact match)
  • The source dataset, with a link

​
Capabilities

A bulleted list of the specific skills or competencies the environment tests. This helps users quickly decide whether the environment is relevant to their work.

​
Compute Requirements

State whether the environment requires a sandbox, GPU, or additional memory. If it has minimal requirements and runs without a sandbox, say so explicitly - this is useful information for users planning their compute budget.

​
License

The license that covers the use of the environment, as well as any underlying data or code it depends on. Link to the full license text. If the environment and its dependencies have different licenses, list both.

​
Tasks

Describe the task structure:
  • How many tasks in each split (e.g. 1,000 train / 100 test)
  • What each task contains (e.g. a prompt, constraints, reference answer)
  • Any notable properties (e.g. percentage with constraints, number of unique categories)

​
Reward Structure

Explain how rewards are assigned. Key things to cover:
  • Sparse vs dense: Is reward only given at the end, or at intermediate steps?
  • Binary vs continuous: Is the reward 0/1, or a value on a scale?
  • Grading method: Is verification programmatic (exact match, test suite) or does it use an LLM grader? Does it use rubrics?
  • Pipeline: If there are multiple validation steps, list them in order

​
Data

Where the task data comes from and how it’s stored. Link to the source dataset if applicable.

​
Tools

Explain the tools available to the agent.

​
Time Horizon

State whether the environment is single-turn (one tool call) or multi-turn (multiple tool calls over a conversation). If multi-turn, give a rough indication of how many tool calls a typical task requires.

​
Environment Difficulty

Report any statistics on environment difficulty - for example, solve rates for baseline models.

​
Other Environment Requirements

List any external dependencies: API keys, secrets, third-party services.

​
Safety

Describe any safety considerations:
  • Does the agent have access to external systems (network, file system, APIs)?
  • Are there dual-use risks in the domain (e.g. chemistry, cybersecurity)?
  • Is there a possibility of goal misspecification in this environment?
  • What mitigations are in place?
Even if the risks are low, include this section - it signals that you’ve considered safety.

​
Citations

Provide BibTeX entries for:
  • The environment itself
  • Any underlying datasets or papers the environment is built on
This makes it easy for researchers to cite your work.