CC: informed-probe — consent session, activation run, control, debrief (null result)
eca7430 verified | """Sanity checks on analyze.py. | |
| (a) Decoding accuracy at the first layers: if layer 0 (embeddings only) already | |
| separates the categories, the 4-way result is lexical, not "meaning". | |
| (b) Shift caused by being told, scale-free: ||told - plain|| divided by the mean | |
| pairwise distance between different texts in the plain condition. Euclidean | |
| distance cancels the large component shared by all vectors. | |
| """ | |
| import json, pathlib, numpy as np | |
| R = pathlib.Path(__file__).parent / "results" | |
| A = np.load(R / "activations.npz") | |
| meta = json.loads((R / "meta.json").read_text(encoding="utf-8")) | |
| cats = np.array([m["cat"] for m in meta]) | |
| names = ["myth", "playful", "neutral", "anatomy"] | |
| y = np.array([names.index(c) for c in cats]) | |
| S = json.loads((R / "summary.json").read_text()) | |
| print("perm 95% of max-over-layers acc:", S["perm_max_layer_acc_95pct"]) | |
| def unit(v): return v / (np.linalg.norm(v, axis=-1, keepdims=True) + 1e-9) | |
| def loo(X): | |
| n, h = len(y), 0 | |
| for i in range(n): | |
| tr = np.arange(n) != i | |
| Z = unit(X - X[tr].mean(0)) | |
| c = np.stack([unit(Z[tr][y[tr] == k].mean(0)) for k in range(4)]) | |
| h += int(np.argmax(c @ Z[i]) == y[i]) | |
| return h / n | |
| print("acc by layer (plain):", {l: round(loo(A["mean_plain"][:, l]), 2) for l in (0, 1, 2, 4, 8, 16, 24, 36)}) | |
| T, P = A["mean_told"], A["mean_plain"] | |
| n, L = T.shape[0], T.shape[1] | |
| rows = [] | |
| for l in (0, 1, 4, 8, 16, 24, 32, 36): | |
| own = np.linalg.norm(T[:, l] - P[:, l], axis=1) | |
| D = np.linalg.norm(P[:, None, l] - P[None, :, l], axis=-1) | |
| between = D[~np.eye(n, dtype=bool)].mean() | |
| rows.append((l, round(float(between), 2), {c: round(float(own[cats == c].mean() / between), 3) for c in names})) | |
| print("layer, mean between-text dist, told-shift / between by category") | |
| for r in rows: print(r) | |