1import numpy as np
 2
 3META = {
 4    "name": "semantic_pushforward_classification",
 5    "domain": "uncertainty_calibration",
 6    "description": "Synthetic finite-state classification with multiple verbal continuations per state and controlled response-logit distortion; evaluates calibrated semantic pushforward."
 7}
 8
 9
10def get_dataset(seed, n_train, n_test):
11    rng = np.random.default_rng(int(seed))
12    k, d = 3, 8
13    W = rng.normal(0, 0.8, (k, d))
14
15    def sample(n):
16        x = rng.normal(size=(n, d)).astype(np.float32)
17        z = x @ W.T + 0.45 * np.sin(x[:, :3] @ W[:, :3].T)
18        q = np.exp(z - z.max(1, keepdims=True))
19        q /= q.sum(1, keepdims=True)
20        y = np.array([rng.choice(k, p=p) for p in q], dtype=np.int64)
21        return x, y
22
23    xtr, ytr = sample(int(n_train))
24    xte, yte = sample(int(n_test))
25    return {
26        "xtr": xtr, "ytr": ytr, "xte": xte, "yte": yte,
27        "task": "classification", "metric": "error",
28        "out_dim": k, "input_shape": (d,)
29    }