seeds: 5, 17, 29
profile: smoke
iterations: 40
group_size: 8
learning_rate: 0.005
ppo_epochs: 3
policy: TinyAutoregressivePolicy (embedding + GRU + linear head)
prompts: 3 classes
max_response_tokens: 4
reward: deterministic verifier, maximum 1.2
evaluation_samples_per_prompt_per_iteration: 24
