Upload a language-based environment¶

A language-based environment is custom gym code that drives an LLM through reset and step. Turn on Language-based in the create modal, then upload, pick an entrypoint, and save.

Prerequisites¶

An Advanced Training project. Classic RL projects can store the gym, but LLM algorithms such as LLMPPO run in Advanced Training.
A Python module or archive ready to upload. See Custom environments.

1. Open Environments¶

Click Environments in the sidebar.

Sidebar with Environments highlighted — **Environments** in the sidebar.¶

2. Language-based setup¶

Click New environment. Name it, then set Environment category to Language-based. That switch is the language-based control in the UI.

Create environment modal with Language-based selected — Setup with **Language-based** turned on.¶

Language-based gyms are single-agent.

3. Upload code¶

Upload the environment archive (file or folder). Include requirements.txt at the top of a package if you have extra dependencies.

Upload step for language-based environment archive — The upload step with a file selected.¶

4. Entrypoint¶

Pick the module and class Arena should instantiate, for example guess_the_number:GuessTheNumberEnv. Finish the stepper to create the first version.

Entrypoint selection step — Choosing an entrypoint class.¶

5. Environment detail¶

The environment page lists versions and validation status. Wait until the version is Validated before you launch an experiment.

Language-based environment detail page — The saved language-based environment.¶

6. View code¶

Open the environment version to inspect the gym source. You should see Guess The Number (guess_the_number:GuessTheNumberEnv).

Language-based environment source in the editor — Viewing Guess The Number source in the editor.¶

7. Validation results¶

Click Validate and open the checklist.

Language-based environment validation checklist — Validation checklist for the working set.¶

In the experiment wizard this gym appears as an RL Environment row. For multi-turn training, continue with Run a Multi-turn LLM experiment.

From the terminal¶

Upload and validate.

Install the CLI and sign in (Authentication):

pip install "agilerl[arena]"
arena login
# or: export ARENA_API_KEY=arena_pat_…

arena env validate ticket-sim --version v1 --source ./ticket_sim.tar.gz --language-based

List discovered entrypoints after validation finishes:

arena env entrypoints ticket-sim --version v1

Python¶

Validate the gym, then point a run spec at it:

from agilerl.arena import ArenaClient

with ArenaClient() as client:
    client.validate_environment(
        name="ticket-sim",
        version="v1",
        source="./ticket_sim.tar.gz",
        language_based=True,
    )
    spec = client.get_default_run_spec(
        algorithm="LLMPPO",
        gym_env="ticket-sim",
    )