rLLM CLI

Evaluate, train, and monitor agents from the command line and web dashboard

rLLM provides two interfaces for working with agents: a CLI for running evaluations and training from the terminal, and a web UI for monitoring runs and exploring results.

CLI

The rLLM CLI help screen

The rllm CLI is the primary way to evaluate and train agents. Datasets are auto-pulled from HuggingFace, agents and evaluators are resolved from the built-in catalog, and a local LiteLLM proxy handles API routing.

First-time setup

Configure your model provider before running evaluations or training:

rllm model setup

This walks you through selecting a provider (OpenAI, Anthropic, etc.), entering your API key, and choosing a default model. Configuration is saved to ~/.rllm/config.json.

Core commands

rllm eval gsm8k                          # Evaluate on a benchmark
rllm train gsm8k --model Qwen/Qwen3-8B  # Train with RL
rllm init my-agent                       # Scaffold a new agent project
rllm dataset list --all                  # Browse available datasets
rllm model setup                         # Configure model providers

Typical workflow

1. Configure your model

   rllm model setup
   
2. Explore available datasets

   rllm dataset list --all
   
3. Evaluate on a benchmark

   rllm eval gsm8k
   
4. Train with RL

   rllm train gsm8k --model Qwen/Qwen3-8B
   

Global behavior

• Auto-pull: Datasets are automatically downloaded from HuggingFace when first referenced by eval or train.
• LiteLLM proxy: When no --base-url is provided, eval and train start a local LiteLLM proxy automatically, routing requests to your configured provider.
• Lazy loading: Commands are loaded on-demand, so rllm --help starts instantly regardless of installed extras.

Web UI

The rLLM web UI at ui.rllm-project.com provides a dashboard for monitoring training runs and exploring evaluation results.

Logging in

Authenticate with the hosted UI from the CLI:

rllm login

This opens a browser for authentication and stores your API key locally in ~/.rllm/config.json. You can also set the RLLM_API_KEY environment variable directly.

Enabling UI logging

Add the --ui flag to eval or train to stream live data to the dashboard:

rllm eval gsm8k --ui
rllm train gsm8k --model Qwen/Qwen3-8B --ui

When enabled, the CLI sends metrics, episode data, and step-by-step execution traces to the UI backend in real time.

What you can see

The UI provides:

• Episode explorer: Browse individual episodes with full trajectory and step-level detail
• Metrics dashboard: Track rewards, success rates, and custom signals across training runs
• Trajectory viewer: Inspect agent reasoning chains and tool calls step by step

To use a self-hosted UI instance, set the RLLM_UI_URL environment variable to your instance URL.