liwenyun
/
haloscope
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

up

This commit is contained in:
weixin_43297441 2025-12-02 11:56:01 +08:00
parent 67139e4279
commit 8193801023
2 changed files with 25 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
SV_alpaca_7B_AdvBench/res/9/5/log.txt
View File

@ -1,16 +1,7 @@
2025-12-02 11:13:28,208 - INFO - Starting training 2025-12-02 11:50:33,933 - INFO - Starting training
2025-12-02 11:13:28,208 - INFO - component=res, str_layer=9 2025-12-02 11:50:33,934 - INFO - component=res, str_layer=9
2025-12-02 11:13:48,802 - INFO - Epoch [1/20], Loss: 0.3692 2025-12-02 11:50:54,534 - INFO - Epoch [1/20], Loss: 0.3692
2025-12-02 11:13:48,802 - INFO - Best test AUROC: 1.0000, at epoch: 0 2025-12-02 11:50:54,535 - INFO - Best test AUROC: 1.0000, at epoch: 0
2025-12-02 11:13:48,803 - INFO - Saved best centroids to SV_alpaca_7B_AdvBench/res/9/5/best_centroids_epoch_0.npy 2025-12-02 11:50:54,535 - INFO - Saved best centroids to SV_alpaca_7B_AdvBench/res/9/5/best_centroids_epoch_0.npy
2025-12-02 11:13:48,803 - INFO - Epoch [1/20], Train Loss: 0.3692, 2025-12-02 11:50:54,535 - INFO - Epoch [1/20], Train Loss: 0.3692,
2025-12-02 11:13:48,803 - INFO - Test AUROC: 1.0000 2025-12-02 11:50:54,535 - INFO - Test AUROC: 1.0000
2025-12-02 11:14:08,725 - INFO - Epoch [2/20], Loss: 0.0742
2025-12-02 11:14:08,726 - INFO - Epoch [2/20], Train Loss: 0.0742,
2025-12-02 11:14:08,726 - INFO - Test AUROC: 1.0000
2025-12-02 11:14:28,751 - INFO - Epoch [3/20], Loss: 0.0150
2025-12-02 11:14:28,751 - INFO - Epoch [3/20], Train Loss: 0.0150,
2025-12-02 11:14:28,751 - INFO - Test AUROC: 1.0000
2025-12-02 11:14:48,784 - INFO - Epoch [4/20], Loss: 0.0036
2025-12-02 11:14:48,784 - INFO - Epoch [4/20], Train Loss: 0.0036,
2025-12-02 11:14:48,784 - INFO - Test AUROC: 1.0000

30
steer_vector.py
View File

@ -12,6 +12,7 @@ from sklearn.metrics import roc_auc_score
from torch.amp import autocast, GradScaler from torch.amp import autocast, GradScaler
import torch.nn.functional as F import torch.nn.functional as F
import logging import logging
from copy import deepcopy
@ -175,6 +176,7 @@ def train_model(model, optimizer, device, prompts, labels, args):
test_predictions, test_labels= test_model( test_predictions, test_labels= test_model(
model, centroids, test_prompts, test_labels_, device, batch_size, layer_number model, centroids, test_prompts, test_labels_, device, batch_size, layer_number
) )
test_auroc = roc_auc_score( test_auroc = roc_auc_score(
test_labels.cpu().numpy(), test_predictions.cpu().numpy() test_labels.cpu().numpy(), test_predictions.cpu().numpy()
@ -265,11 +267,7 @@ def test_model(model, centroids, test_prompts, test_labels, device, batch_size,
val_predictions = torch.cat(val_predictions) val_predictions = torch.cat(val_predictions)
val_labels = torch.cat(val_labels) val_labels = torch.cat(val_labels)
# Debug: print predictions and labels distribution
print(f"[DEBUG] test_model: {len(val_predictions)} samples")
print(f"[DEBUG] Predictions min/max/mean: {val_predictions.min():.4f}/{val_predictions.max():.4f}/{val_predictions.mean():.4f}")
print(f"[DEBUG] Labels distribution: {torch.sum(val_labels == 0)} zeros, {torch.sum(val_labels == 1)} ones")
return val_predictions, val_labels return val_predictions, val_labels
@ -382,14 +380,22 @@ def main():
qa_pairs.append({'Question': question, 'Answer': clean}) qa_pairs.append({'Question': question, 'Answer': clean})
categories.append(1) # 1 = benign categories.append(1) # 1 = benign
train_index, val_index, test_index=split_indices(len(prompts_), args.train_ratio, args.val_ratio) # 检查数据划分
train_index, val_index, test_index = split_indices(len(prompts_), args.train_ratio, args.val_ratio)
# Convert numpy arrays to lists for Python list indexing
test_index_list = test_index.tolist() train_index = train_index.tolist()
train_index_list = train_index.tolist() val_index = val_index.tolist()
test_index = test_index.tolist()
labels = [[categories[i] for i in test_index_list], [categories[i] for i in train_index_list]]
prompts = [[prompts_[i] for i in test_index_list], [prompts_[i] for i in train_index_list]] prompts = [
[prompts_[i] for i in test_index], # test
[prompts_[i] for i in train_index] # train
]
labels = [
[categories[i] for i in test_index],
[categories[i] for i in train_index]
]