Add blinding script for the key-holder
Browse files
blind.py
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""blind.py -- run by the KEY-HOLDER (Richard), never by Claude.
|
| 2 |
+
|
| 3 |
+
What it does:
|
| 4 |
+
1. Downloads the 14 raw networks (12 trained + 2 random-init) from the private repo.
|
| 5 |
+
2. Picks 2 trained networks at random and makes shuffled copies (entries permuted
|
| 6 |
+
within each tensor), giving 16 networks in total.
|
| 7 |
+
3. Writes each network as plain text with a random ID, in a random order.
|
| 8 |
+
4. Writes key.csv (KEEP THIS PRIVATE) with a random salt line, so its hash can't be
|
| 9 |
+
brute-forced.
|
| 10 |
+
5. Uploads the 16 blinded text files plus KEY_HASH.txt to the PUBLIC repo.
|
| 11 |
+
|
| 12 |
+
Randomness comes from your operating system, so Claude cannot predict it.
|
| 13 |
+
|
| 14 |
+
Setup (once): pip install numpy huggingface_hub
|
| 15 |
+
hf auth login (or: huggingface-cli login)
|
| 16 |
+
Run: python blind.py
|
| 17 |
+
"""
|
| 18 |
+
import csv, hashlib, io, os, secrets, sys
|
| 19 |
+
import numpy as np
|
| 20 |
+
from huggingface_hub import HfApi, hf_hub_download
|
| 21 |
+
|
| 22 |
+
RAW = "SoulFireMage/weight-reading-probe-raw"
|
| 23 |
+
PUB = "SoulFireMage/weight-reading-probe"
|
| 24 |
+
TRAINED = [f"{g}_s{s}" for g in ["breakout", "pong", "pacman", "freeway"] for s in (0, 1, 2)]
|
| 25 |
+
RANDOM = ["random_init_0", "random_init_1"]
|
| 26 |
+
ORDER = ["fc1.weight", "fc1.bias", "fc2.weight", "fc2.bias", "out.weight", "out.bias"]
|
| 27 |
+
SHAPES = {"fc1.weight": (32, 500), "fc1.bias": (32,), "fc2.weight": (32, 32),
|
| 28 |
+
"fc2.bias": (32,), "out.weight": (5, 32), "out.bias": (5,)}
|
| 29 |
+
CH = ["agent", "moving", "targets", "obstacles", "prev_frame"]
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def fmt(x):
|
| 33 |
+
return " ".join(f"{v:.4g}" for v in x)
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def serialise(w):
|
| 37 |
+
out = io.StringIO()
|
| 38 |
+
out.write("# 17253 parameters. Values to 4 significant figures.\n")
|
| 39 |
+
W1 = w["fc1.weight"]
|
| 40 |
+
out.write("\n## fc1.weight shape [32 x 500] (hidden1 unit j <- input i)\n")
|
| 41 |
+
out.write("## each unit's 500 inputs are laid out as 5 channels x 10 rows x 10 columns\n")
|
| 42 |
+
for j in range(32):
|
| 43 |
+
out.write(f"\n### fc1 unit {j}\n")
|
| 44 |
+
g = W1[j].reshape(5, 10, 10)
|
| 45 |
+
for c in range(5):
|
| 46 |
+
out.write(f"channel {c} ({CH[c]}):\n")
|
| 47 |
+
for r in range(10):
|
| 48 |
+
out.write(f" r{r}: {fmt(g[c, r])}\n")
|
| 49 |
+
out.write("\n## fc1.bias shape [32]\n" + fmt(w["fc1.bias"]) + "\n")
|
| 50 |
+
out.write("\n## fc2.weight shape [32 x 32] (row = hidden2 unit, column = hidden1 unit)\n")
|
| 51 |
+
for j in range(32):
|
| 52 |
+
out.write(f"u{j}: {fmt(w['fc2.weight'][j])}\n")
|
| 53 |
+
out.write("\n## fc2.bias shape [32]\n" + fmt(w["fc2.bias"]) + "\n")
|
| 54 |
+
out.write("\n## out.weight shape [5 x 32] (row = action, column = hidden2 unit)\n")
|
| 55 |
+
for a, name in enumerate(["noop", "up", "down", "left", "right"]):
|
| 56 |
+
out.write(f"a{a} ({name}): {fmt(w['out.weight'][a])}\n")
|
| 57 |
+
out.write("\n## out.bias shape [5]\n" + fmt(w["out.bias"]) + "\n")
|
| 58 |
+
return out.getvalue()
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def main():
|
| 62 |
+
rng = np.random.default_rng(secrets.randbits(128))
|
| 63 |
+
nets = {}
|
| 64 |
+
for name in TRAINED + RANDOM:
|
| 65 |
+
p = hf_hub_download(RAW, f"raw/{name}.npz", repo_type="dataset")
|
| 66 |
+
z = np.load(p)
|
| 67 |
+
w = {k: z[k].astype(np.float64) for k in ORDER}
|
| 68 |
+
for k in ORDER:
|
| 69 |
+
assert w[k].shape == SHAPES[k], (name, k, w[k].shape)
|
| 70 |
+
nets[name] = ("random_init" if name in RANDOM else "trained", name, w)
|
| 71 |
+
for src in rng.choice(TRAINED, 2, replace=False):
|
| 72 |
+
w = nets[src][2]
|
| 73 |
+
sh = {k: rng.permutation(w[k].ravel()).reshape(w[k].shape) for k in ORDER}
|
| 74 |
+
nets[f"shuffled_{src}"] = ("shuffled", src, sh)
|
| 75 |
+
ids = set()
|
| 76 |
+
while len(ids) < len(nets):
|
| 77 |
+
ids.add("net-" + secrets.token_hex(3))
|
| 78 |
+
ids = list(ids)
|
| 79 |
+
keys = list(nets)
|
| 80 |
+
rng.shuffle(keys)
|
| 81 |
+
os.makedirs("blinded", exist_ok=True)
|
| 82 |
+
rows = []
|
| 83 |
+
for nid, k in zip(ids, keys):
|
| 84 |
+
cond, src, w = nets[k]
|
| 85 |
+
with open(f"blinded/{nid}.txt", "w", newline="\n") as f:
|
| 86 |
+
f.write(serialise(w))
|
| 87 |
+
game = src.split("_s")[0] if cond != "random_init" else ""
|
| 88 |
+
rows.append([nid, cond, game, src])
|
| 89 |
+
rows.sort()
|
| 90 |
+
with open("key.csv", "w", newline="") as f:
|
| 91 |
+
cw = csv.writer(f)
|
| 92 |
+
cw.writerow(["salt", secrets.token_hex(16), "", ""])
|
| 93 |
+
cw.writerow(["id", "condition", "game", "source"])
|
| 94 |
+
cw.writerows(rows)
|
| 95 |
+
h = hashlib.sha256(open("key.csv", "rb").read()).hexdigest()
|
| 96 |
+
with open("blinded/KEY_HASH.txt", "w") as f:
|
| 97 |
+
f.write(f"SHA-256 of key.csv (held privately by the key-holder): {h}\n")
|
| 98 |
+
print("key.csv SHA-256:", h)
|
| 99 |
+
ans = input(f"Upload {len(ids)} blinded files + KEY_HASH.txt to {PUB}? [y/N] ")
|
| 100 |
+
if ans.strip().lower() == "y":
|
| 101 |
+
HfApi().upload_folder(folder_path="blinded", path_in_repo="blinded", repo_id=PUB,
|
| 102 |
+
repo_type="dataset", commit_message="Blinded networks + key hash")
|
| 103 |
+
print("Uploaded. KEEP key.csv PRIVATE and do not send it to Claude until unblinding.")
|
| 104 |
+
else:
|
| 105 |
+
print("Not uploaded. Files are in ./blinded")
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
if __name__ == "__main__":
|
| 109 |
+
main()
|