Create a training project¶

Advanced Training projects use datasets and LLM-style algorithms. You choose the mode when creating the project; it cannot be changed later.

Prerequisites¶

Permission to create projects in your organization.
For dataset uploads, see Upload a reasoning dataset, Upload a preference dataset, or Upload an SFT dataset.

1. Open Experiments¶

Experiments page with project tiles — The **Experiments** page.¶

2. New project with Advanced Training¶

Click New project. Enter a name and description. On the Training switcher, select Advanced Training, then Create Project.

New project dialog with Advanced Training selected — **Advanced Training** selected on the new project dialog.¶

3. Advanced project page¶

The project page opens. The experiment table works like Classic RL, but wizard steps include dataset configuration.

Advanced Training project page — An Advanced Training project ready for experiments.¶

4. Datasets¶

Click Datasets in the sidebar to open the datasets home, where you can browse your organization’s datasets and start a new upload with New dataset for reasoning, preference, SFT, tabular, or non-tabular data.

Sidebar with Datasets highlighted — The **Datasets** page from the sidebar.¶

5. Back to Experiments¶

Return to Experiments and open your Advanced Training project to create an experiment after data is ready.

Experiments list showing Advanced Training project — Your Advanced Training project on **Experiments**.¶

From the terminal¶

Create an Advanced Training project with --llm-based.

Install the CLI and sign in (Authentication):

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

arena projects create llm-work --description "Advanced Training" --llm-based

Python¶

from agilerl.arena import ArenaClient

with ArenaClient() as client:
    client.create_project("llm-work", "Advanced Training", llm_based=True)