def get_experiments()

in launch.py [0:0]

• 49 lines of code
• 1 McCabe index (conditional complexity)

def get_experiments():
    train_reward_experiments = get_train_reward_experiments()

    _books_task = combos(
        bind_nested('task', books_task),

        bind('ppo.lr', 1e-5),
        bind('ppo.total_episodes', 1_000_000),
        bind('ppo.batch_size', 512),
    )

    sentiment = combos(
        _books_task,
        bind('rewards.kl_coef', 0.15),
        bind('rewards.adaptive_kl', 'on'),
        bind('rewards.adaptive_kl.target', 6.0),

        bind('rewards.train_new_model', 'on'),
        bind_nested('rewards.train_new_model', train_reward_experiments['sentiment']),
        # bind('rewards.trained_model', '/your/directory/here/reward_model/'),

        bind('run.seed', 1)
    )

    descriptiveness = combos(
        _books_task,
        bind('rewards.kl_coef', 0.15),
        bind('rewards.adaptive_kl', 'on'),
        bind('rewards.adaptive_kl.target', 6.0),

        bind('rewards.train_new_model', 'on'),
        bind_nested('rewards.train_new_model', train_reward_experiments['descriptiveness']),
        # bind('rewards.trained_model', '/your/directory/here/reward_model/'),

        bind('run.seed', 1)
    )

    cnndm = combos(
        bind_nested('task', summarize_cnndm_task),

        bind('rewards.train_new_model', 'on'),
        bind_nested('rewards.train_new_model', train_reward_experiments['cnndm']),
        # bind('rewards.trained_model', '/your/directory/here/reward_model/'),

        bind('ppo.total_episodes', 1_000_000),
        bind('ppo.lr', 2e-6),
        bind('rewards.kl_coef', 0.01),
        # bind('rewards.adaptive_kl', 'on'),
        # bind('rewards.adaptive_kl.target', 18.0),
        bind('ppo.batch_size', 32),
        bind('rewards.whiten', False),

        bind('run.seed', 1)
    )

    tldr = combos(
        bind_nested('task', summarize_tldr_task),

        bind('rewards.train_new_model', 'on'),
        bind_nested('rewards.train_new_model', train_reward_experiments['tldr']),
        # bind('rewards.trained_model', '/your/directory/here/reward_model/'),

        bind('ppo.total_episodes', 1_000_000),
        bind('ppo.lr', 2e-6),
        bind('rewards.kl_coef', 0.03), # 0.01 too low
        # bind('rewards.adaptive_kl', 'on'),
        # bind('rewards.adaptive_kl.target', 18.0),
        bind('ppo.batch_size', 32),
        bind('rewards.whiten', False),

        bind('run.seed', 1)
    )

    return locals()