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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
|
"""
Loss-reweighting ablation (GPT-5.6-Terra's design, adapted to faithful J-lens).
Tests whether increasing a token's EFFECTIVE frequency/importance reduces its
faithful J-lens norm — without the confounds of the old random-insertion
ablation (which shifted positions, destroyed n-grams, and used an unmatched
control run).
Design per seed (identical init + identical minibatch order for all three):
q-upweight: cross-entropy terms whose target is 'q' are weighted x2.
control: ordinary loss.
ctrl_random: same-total-loss control: weight x2 on the SAME NUMBER of
randomly chosen non-'q' target positions (deterministic per
batch index, so all models share the same control positions).
If increased effective frequency causally reduces the faithful J-lens norm of
'q', the q-upweight model must show a lower norm than BOTH controls.
Steps:
python3 src/loss_reweight.py --step train --mode q --seed 0 [--max_iters 3000]
python3 src/loss_reweight.py --step jlens --mode q --seed 0 [--layers 2,3,4]
python3 src/loss_reweight.py --step summary
"""
import sys, os, argparse, subprocess, pickle
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from typing import Any
import numpy as np
import torch
import torch.nn.functional as F
DATA_DIR = 'data/shakespeare_char'
OUT_ROOT = 'out-loss-reweight'
JLENS_OUT = 'outputs/loss_reweight'
TARGET = 'q'
WEIGHT = 2.0
MODEL_ARGS: dict[str, Any] = dict(n_layer=6, n_head=6, n_embd=384, block_size=128,
bias=False, dropout=0.2)
LOSS_MODES = ('q', 'control', 'ctrl_random')
def _batch(data, blk, bs, g, device):
ix = torch.randint(len(data) - blk, (bs,), generator=g)
x = torch.stack([torch.from_numpy(data[i:i+blk].astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy(data[i+1:i+1+blk].astype(np.int64)) for i in ix])
return x.to(device), y.to(device)
def _weighted_loss(logits, y, mode, q_id, batch_k, V, device):
"""Per-token weighted CE. Returns scalar loss."""
logp = F.log_softmax(logits.view(-1, V), dim=-1)
nll = -logp.gather(1, y.view(-1, 1)).squeeze(1) # (B*T,)
w = torch.ones_like(nll)
if mode == 'q':
w[y.view(-1) == q_id] = WEIGHT
elif mode == 'ctrl_random':
g = torch.Generator().manual_seed(1000 + batch_k) # CPU generator (randperm)
n_q = int((y == q_id).sum().item())
flat = torch.arange(y.numel(), device=device)
non_q = flat[y.view(-1) != q_id]
if len(non_q) > 0 and n_q > 0:
pick = non_q[torch.randperm(len(non_q), generator=g)[:min(n_q, len(non_q))]]
w[pick] = WEIGHT
return (nll * w).mean()
def train(mode, seed, max_iters, batch_size):
sys.path.insert(0, '.')
from model import GPT, GPTConfig
torch.manual_seed(seed)
np.random.seed(seed)
device = 'cuda'
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')
with open(f'{DATA_DIR}/meta.pkl', 'rb') as f:
meta = pickle.load(f)
q_id = meta['stoi'][TARGET]
args = dict(
n_layer=int(MODEL_ARGS['n_layer']), n_head=int(MODEL_ARGS['n_head']),
n_embd=int(MODEL_ARGS['n_embd']), block_size=int(MODEL_ARGS['block_size']),
bias=bool(MODEL_ARGS['bias']), dropout=float(MODEL_ARGS['dropout']),
vocab_size=int(meta['vocab_size']),
)
model = GPT(GPTConfig(**args)).to(device)
print(f"[{mode}] seed {seed}: params={sum(p.numel() for p in model.parameters())/1e6:.2f}M")
opt = model.configure_optimizers(weight_decay=0.1, learning_rate=1e-3,
betas=(0.9, 0.99), device_type='cuda')
bs = batch_size
blk = args['block_size']
V = args['vocab_size']
# identical minibatch order for every model: per-seed CPU generator, fixed start
# (torch.randint does not accept CUDA generators on torch 2.4)
g = torch.Generator().manual_seed(20260731 + seed)
gval = torch.Generator().manual_seed(777 + seed)
def get_batch(split):
d = train_data if split == 'train' else val_data
gg = g if split == 'train' else gval
return _batch(d, blk, bs, gg, device)
best_val = 1e9
out_dir = f'{OUT_ROOT}/{mode}/seed{seed}'
os.makedirs(out_dir, exist_ok=True)
for it in range(max_iters):
if it % 500 == 0:
model.eval()
lv = []
for _ in range(50):
X, Y = get_batch('val')
with torch.no_grad():
_, loss = model(X, Y)
lv.append(loss.item())
v = np.mean(lv)
model.train()
if v < best_val:
best_val = v
torch.save({'model': model.state_dict(), 'model_args': args,
'best_val_loss': best_val}, f'{out_dir}/ckpt.pt')
if it % 1000 == 0:
print(f" iter {it}: val={v:.4f}")
X, Y = get_batch('train')
logits = model(X, Y)[0] # pass Y: targets=None would give last-position logits only
loss = _weighted_loss(logits, Y, mode, q_id, it, V, device)
loss.backward()
opt.step()
opt.zero_grad(set_to_none=True)
print(f"[{mode}] seed {seed} done. best_val={best_val:.4f}")
def jlens(mode, seed, layers, n_prompts):
out_dir = f'{JLENS_OUT}/{mode}/seed{seed}'
cmd = ["python3", "-u", "src/jlens_v3.py",
"--checkpoint", f'{OUT_ROOT}/{mode}/seed{seed}/ckpt.pt',
"--data_dir", DATA_DIR,
"--n_prompts", str(n_prompts),
"--layers", layers,
"--chunk", "16",
"--output_dir", out_dir]
print("running:", " ".join(cmd))
r = subprocess.run(cmd, cwd='/workspace/code')
assert r.returncode == 0, "jlens_v3 failed"
def summary(layers):
with open(f'{DATA_DIR}/meta.pkl', 'rb') as f:
meta = pickle.load(f)
q_id = meta['stoi'][TARGET]
seeds = sorted(set(
d.split('seed')[1] for mode in LOSS_MODES
for d in os.listdir(f'{JLENS_OUT}/{mode}')
if d.startswith('seed')))
print(f"\n{'='*78}")
print(f"LOSS-REWEIGHTING: faithful J-lens norm of '{TARGET}' "
f"({WEIGHT}x CE) vs controls")
print(f"{'='*78}")
print(f"{'seed':<5}{'layer':<6}" + "".join(f"{m:>14}" for m in LOSS_MODES))
for s in seeds:
for l in map(int, layers.split(',')):
row = [s, str(l)]
for m in LOSS_MODES:
d = torch.load(f'{JLENS_OUT}/{m}/seed{s}/layer{l}.pt',
map_location='cpu')
row.append(f"{d['faithful_norms'][q_id]:.4f}")
print(f"{row[0]:<5}{row[1]:<6}" + "".join(f"{v:>14}" for v in row[2:]))
# mean over middle layers per mode
mids = [l for l in map(int, layers.split(','))]
means = {}
for m in LOSS_MODES:
vals = []
for l in mids:
d = torch.load(f'{JLENS_OUT}/{m}/seed{s}/layer{l}.pt',
map_location='cpu')
vals.append(d['faithful_norms'][q_id])
means[m] = np.mean(vals)
print(f" -> mean over layers: q={means['q']:.4f} "
f"control={means['control']:.4f} ctrl_random={means['ctrl_random']:.4f}")
print(f" -> q/control = {means['q']/max(means['control'],1e-9):.3f} "
f"q/ctrl_random = {means['q']/max(means['ctrl_random'],1e-9):.3f}")
if __name__ == '__main__':
ap = argparse.ArgumentParser()
ap.add_argument('--step', required=True, choices=['train', 'jlens', 'summary'])
ap.add_argument('--mode', choices=LOSS_MODES)
ap.add_argument('--seed', type=int, default=0)
ap.add_argument('--max_iters', type=int, default=3000)
ap.add_argument('--batch_size', type=int, default=16)
ap.add_argument('--layers', default='2,3,4')
ap.add_argument('--n_prompts', type=int, default=10)
a = ap.parse_args()
if a.step == 'train':
assert a.mode, "need --mode"
train(a.mode, a.seed, a.max_iters, a.batch_size)
elif a.step == 'jlens':
assert a.mode, "need --mode"
jlens(a.mode, a.seed, a.layers, a.n_prompts)
else:
summary(a.layers)
|
