1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
|
"""
Controlled frequency ablation: train model on doubled-q Shakespeare,
compare J-lens norms for 'q' vs control.
Usage (inside Docker on meru):
python3 src/freq_ablation.py
"""
import sys, os
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import torch, numpy as np, pickle
from model import GPT, GPTConfig
import jlens_v2
from jlens_v2 import compute_jlens_layer
device = 'cuda'
MODEL_ARGS = dict(n_layer=6, n_head=6, n_embd=384, block_size=128,
bias=False, vocab_size=65, dropout=0.2)
def get_batch(train_data, val_data, split, block_size=128, batch_size=32):
data = train_data if split == 'train' else val_data
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([torch.from_numpy(data[i:i+block_size].astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy(data[i+1:i+1+block_size].astype(np.int64)) for i in ix])
return x.to(device), y.to(device)
def train(data_dir, out_dir, max_iters=5000):
"""Train a model and return best checkpoint path."""
train_data = np.memmap(f'{data_dir}/train.bin', dtype=np.uint16, mode='r')
val_data = np.memmap(f'{data_dir}/val.bin', dtype=np.uint16, mode='r')
config = GPTConfig(**MODEL_ARGS)
model = GPT(config).to(device)
print(f" Params: {sum(p.numel() for p in model.parameters())/1e6:.2f}M")
optimizer = model.configure_optimizers(weight_decay=0.1, lr=1e-3,
betas=(0.9, 0.99), device_type='cuda')
os.makedirs(out_dir, exist_ok=True)
best_val = 1e9
for it in range(max_iters):
if it % 500 == 0:
model.eval()
losses = {}
for split in ['train', 'val']:
lv = []
for _ in range(50):
X, Y = get_batch(train_data, val_data, split)
_, loss = model(X, Y)
lv.append(loss.item())
losses[split] = np.mean(lv)
model.train()
print(f" step {it}: train={losses['train']:.4f}, val={losses['val']:.4f}")
if losses['val'] < best_val:
best_val = losses['val']
torch.save({'model': model.state_dict(), 'model_args': MODEL_ARGS,
'best_val_loss': best_val}, f'{out_dir}/ckpt.pt')
X, Y = get_batch(train_data, val_data, 'train')
logits, loss = model(X, Y)
loss.backward()
optimizer.step()
optimizer.zero_grad(set_to_none=True)
if it % 500 == 0:
print(f" iter {it}: loss={loss.item():.4f}")
print(f" Done. Best val: {best_val:.4f}")
return model
def run_jlens_on_model(model, data_dir, target_char, n_batches=10):
"""Compute J-lens and return norm for target_char."""
train_data = np.memmap(f'{data_dir}/train.bin', dtype=np.uint16, mode='r')
with open(f'{data_dir}/meta.pkl', 'rb') as f:
meta = pickle.load(f)
target_id = meta['stoi'][target_char]
batch_size = 16
block_size = MODEL_ARGS['block_size']
batches = []
for _ in range(n_batches):
ix = torch.randint(len(train_data) - block_size, (batch_size,))
x = torch.stack([torch.from_numpy(
train_data[i:i+block_size].astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy(
train_data[i+1:i+1+block_size].astype(np.int64)) for i in ix])
batches.append((x, y))
results = {}
for layer_idx in [2, 3, 4]:
jlens = compute_jlens_layer(model, layer_idx, batches, device)
results[layer_idx] = jlens[target_id].norm().item()
return results
# ── MAIN ───────────────────────────────────────────────
print("=" * 60)
print("CONTROLLED FREQUENCY ABLATION: 'q' DOUBLED")
print("=" * 60)
# 1. Train control model (or use existing)
print("\n[1/4] Using existing control model...")
ctrl_model, _ = jlens_v2.load_model('out-shakespeare-char/ckpt.pt', device)
ctrl_model.eval()
# 2. Train modified model
print("\n[2/4] Training modified model (doubled 'q')...")
mod_model = train('data/freq_experiment/doubled_q', 'out-freq-doubled-q')
mod_model.eval()
# 3. J-lens on both
print("\n[3/4] Computing J-lens...")
print(" Control model...")
ctrl_norms = run_jlens_on_model(ctrl_model, 'data/shakespeare_char', 'q')
print(" Modified model...")
mod_norms = run_jlens_on_model(mod_model, 'data/freq_experiment/doubled_q', 'q')
# 4. Compare
print("\n[4/4] RESULTS")
print("=" * 60)
print(f" Character: 'q'")
print(f" Original frequency: 0.055%")
print(f" Doubled frequency: 0.109%")
print()
print(f" {'Layer':<8} {'Control':>10} {'Modified':>10} {'Change':>10}")
print(f" {'-'*8} {'-'*10} {'-'*10} {'-'*10}")
for layer_idx in sorted(ctrl_norms.keys()):
c = ctrl_norms[layer_idx]
m = mod_norms[layer_idx]
pct = (m - c) / c * 100
direction = "↓" if pct < 0 else "↑"
print(f" L{layer_idx:<7} {c:>10.4f} {m:>10.4f} {direction}{abs(pct):>8.1f}%")
print()
avg_ctrl = np.mean(list(ctrl_norms.values()))
avg_mod = np.mean(list(mod_norms.values()))
avg_pct = (avg_mod - avg_ctrl) / avg_ctrl * 100
print(f" Average: {avg_ctrl:.4f} → {avg_mod:.4f} ({avg_pct:+.1f}%)")
print(f" HYPOTHESIS: doubling frequency should REDUCE J-lens norm")
print(f" RESULT: {'CONFIRMED' if avg_pct < -2 else 'INCONCLUSIVE' if abs(avg_pct) < 2 else 'CONTRADICTED'}")
|
