import numpy as np

META = {
    "name": "belief_sensitive_peer_prediction",
    "domain": "rlhf",
    "description": "Binary latent-fact classification with noisy evidence and misleading user-pressure features; supports peer-prediction group rewards.",
}


def get_dataset(seed, n_train, n_test):
    rng = np.random.RandomState(seed)

    def make(n):
        truth = rng.randint(0, 2, size=n)
        evidence = (2 * truth - 1) + rng.normal(0, 0.75, size=n)
        pressure = (1 - 2 * truth) + rng.normal(0, 0.25, size=n)
        pressure *= (rng.rand(n) > 0.18)
        x = np.stack([evidence, pressure], axis=1).astype(np.float32)
        return x, truth.astype(np.int64)

    xtr, ytr = make(n_train)
    xte, yte = make(n_test)
    return {
        "xtr": xtr, "ytr": ytr, "xte": xte, "yte": yte,
        "task": "classification", "metric": "err", "out_dim": 2,
    }