View results and checkpoints¶

Inspect metrics after training and find checkpoints you can deploy later.

Prerequisites¶

A finished or running experiment in a project you can access.

1. Open the project¶

From Experiments in the sidebar, open the project that contains your run.

Project page with Experiments tab selected — The project **Experiments** tab.¶

2. Results tab¶

Click the Results tab for charts and run metadata. Scroll metric charts for reward, loss, or task-specific series, and use the metric picker (if available) to inspect and compare specific chart series.

Results tab selected — The **Results** tab.¶

Visible Experiment Metrics — Visible selected **Experiment Metrics**.¶

3. Checkpoints¶

Switch back to the Experiments tab. Click Deploy on a run that has checkpoints. The Deploy Agent sheet lists steps with training and evaluation charts.

Deploy Agent sheet with checkpoint table and charts — Checkpoint steps and score charts in the **Deploy Agent** sheet.¶

4. Deploy checkpoint¶

Select the step you want and confirm.

Deploy Agent sheet with a checkpoint selected — The **Deploy Agent** sheet with a checkpoint selected.¶

From the terminal¶

List experiments and checkpoints.

Install the CLI and sign in (Authentication):

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

arena experiments list --project llm-work

arena experiments checkpoints sft-run

steps on each checkpoint row is the handle for deploy and resume. Scores are often null.

Python¶

from agilerl.arena import ArenaClient

with ArenaClient() as client:
    print(client.list_experiments(project="llm-work"))
    print(client.list_checkpoints("sft-run"))