Quickstart

Use these commands after installation to validate the main AReno paths. Full training and agentic rollout require a CUDA-capable NVIDIA GPU.

Training smoke test

Run the smallest official training task to verify that the CLI can load a model, build batches, execute the training loop, and write outputs locally.

areno train \
  --ckpt Qwen/Qwen3-0.6B \
  --dataset-path gsm8k:main \
  --dataset-loader-fn examples/math/dataset_loader.py \
  --reward-fn-path examples/math/math_verify_reward.py \
  --algo gspo \
  --tp-size 1 \
  --world-size 1 \
  --batch-size 1

RLVR path

RLVR connects a dataset, model rollout, reward function, and policy loss. The math example is the fastest way to see that path end to end.

areno train \
  --ckpt Qwen/Qwen3-0.6B \
  --dataset-path gsm8k:main \
  --dataset-loader-fn examples/math/dataset_loader.py \
  --reward-fn-path examples/math/math_verify_reward.py \
  --algo gspo \
  --tp-size 1 \
  --world-size 1

Read Training Loop for the mental model and Math RLVR recipe for the runnable recipe shape.

Agentic rollout path

Agentic rollout is for tasks where the model interacts with tools, games, services, or an environment before AReno scores the trajectory.

areno train \
  --agent-fn examples/agentic/tictactoe/run_agent.py \
  --reward-fn-path examples/agentic/tictactoe/reward.py \
  --algo gspo

Read Agentic rollout API for the agentic rollout boundary and TicTacToe agentic RL for the first recipe.