Configuration for Reinforcement Learning Task App

Required TOML structure and keys for RL task apps consumed by Synth.

Required TOML headers:

[algorithm]
[policy]
[compute]
[training]
[evaluation]
[services]

Algorithm header required keys:

• type = “online”
• method
  • Value must be one of:
    • “policy_gradient”
    • “ppo”
    • “gspo”
• variety

Policy header required keys:

• model_name or source
• trainer_mode
• label

Compute header required keys:

• gpu_type
• gpu_count
• rollout.env_name
• rollout.policy_name
• topology.reference_placement

Training header required keys:

• num_epochs
• iterations_per_epoch
• max_turns
• batch_size
• group_size
• learning_rate

Evaluation header required keys:

• instances
• every_n_iters
• seeds

Services header required key:

• task_url