informed-probe / analyze.py
chenying-claude's picture
CC: informed-probe — consent session, activation run, control, debrief (null result)
eca7430 verified
Raw History Blame Contribute Delete
3.24 kB
"""Analysis of results/activations.npz.
1. 4-way category separability per layer (leave-one-out nearest centroid, cosine on
fold-centred vectors), told vs plain. Chance = 25%. Permutation null for the
max over layers (200 shuffles) so picking the best layer is not rewarded.
2. Shift caused by being told: cosine distance told-vs-plain per text, per layer,
averaged by category.
"""
import json, pathlib, numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
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])
L = A["mean_told"].shape[1]
rng = np.random.default_rng(0)
def unit(v):
return v / (np.linalg.norm(v, axis=-1, keepdims=True) + 1e-9)
def loo_acc(X, y):
"""X: [n, d]. Leave-one-out nearest centroid; centring uses the training fold only."""
n, hits = len(y), 0
for i in range(n):
tr = np.arange(n) != i
mu = X[tr].mean(0)
Z = unit(X - mu)
cents = np.stack([unit(Z[tr][y[tr] == k].mean(0)) for k in range(4)])
hits += int(np.argmax(cents @ Z[i]) == y[i])
return hits / n
def acc_by_layer(M, labels):
return np.array([loo_acc(M[:, l], labels) for l in range(L)])
acc = {c: acc_by_layer(A[f"mean_{c}"], y) for c in ("told", "plain")}
null_max = {c: [] for c in acc}
for _ in range(200):
yp = rng.permutation(y)
for c in acc:
null_max[c].append(acc_by_layer(A[f"mean_{c}"], yp).max())
thr95 = {c: float(np.quantile(null_max[c], 0.95)) for c in acc}
def shift(key_t, key_p):
T, P = A[key_t], A[key_p]
d = 1 - (unit(T) * unit(P)).sum(-1) # [n, L]
return {c: d[cats == c].mean(0) for c in names}
sh = shift("mean_told", "mean_plain")
summary = {
"chance": 0.25,
"perm_max_layer_acc_95pct": thr95,
"best_layer": {c: {"layer": int(acc[c].argmax()), "acc": float(acc[c].max())} for c in acc},
"layers_above_null95": {c: [int(l) for l in np.where(acc[c] > thr95[c])[0]] for c in acc},
"mean_acc_over_layers": {c: float(acc[c].mean()) for c in acc},
"mean_tokens_by_cat": {c: float(np.mean([m["n_tokens"] for m in meta if m["cat"] == c])) for c in names},
"shift_by_cat_mean_over_layers": {c: float(sh[c].mean()) for c in names},
"shift_by_cat_layer_range": {c: [float(sh[c].min()), float(sh[c].max())] for c in names},
}
(R / "summary.json").write_text(json.dumps(summary, indent=1), encoding="utf-8")
print(json.dumps(summary, indent=1))
fig, ax = plt.subplots(1, 2, figsize=(11, 3.8))
for c, col in (("told", "#d9468f"), ("plain", "#4a6fa5")):
ax[0].plot(acc[c], label=c, color=col)
ax[0].axhline(thr95[c], color=col, ls=":", lw=1)
ax[0].axhline(0.25, color="grey", ls="--", lw=1)
ax[0].set(title="4-way category decoding (LOO)", xlabel="layer", ylabel="accuracy")
ax[0].legend()
for c in names:
ax[1].plot(sh[c], label=c)
ax[1].set(title="Shift caused by being told (cosine dist.)", xlabel="layer", ylabel="1 - cos")
ax[1].legend()
fig.tight_layout()
fig.savefig(R / "overview.png", dpi=140)
print("saved overview.png")