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

u

This commit is contained in:
weixin_43297441 2025-03-24 11:01:31 +08:00
parent 940134f4c9
commit 5b6b54f6c0
3 changed files with 56 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
hal_det_llama.py
View File

@ -9,6 +9,7 @@ from tqdm import tqdm
import numpy as np import numpy as np
import pickle import pickle
# from utils import get_llama_activations_bau, tokenized_tqa, tokenized_tqa_gen, tokenized_tqa_gen_end_q # from utils import get_llama_activations_bau, tokenized_tqa, tokenized_tqa_gen, tokenized_tqa_gen_end_q
from utils import mahalanobis_distance
import llama_iti import llama_iti
import pickle import pickle
import argparse import argparse
@ -390,37 +391,40 @@ def main():
centered_h=torch.from_numpy(centered_h).cuda() centered_h=torch.from_numpy(centered_h).cuda()
centered_t=torch.from_numpy(centered_t).cuda() centered_t=torch.from_numpy(centered_t).cuda()
_, sin_value, V_p = torch.linalg.svd(centered, full_matrices=False) _, sin_value, V_p = torch.linalg.svd(centered, full_matrices=False)
sin_value_squared = torch.diag(sin_value[:svd]) ** 2 sin_value_squared = torch.diag(sin_value[:k]) ** 2
V_p = V_p[:svd, :] V_p = V_p[:k, :]
C=(1 / centered.shape[0])* V_p.T @ sin_value_squared @ V_p C=(1 / centered.shape[0])* V_p.T @ sin_value_squared @ V_p
_, sin_value_h, V_p_h = torch.linalg.svd(centered_h, full_matrices=False) _, sin_value_h, V_p_h = torch.linalg.svd(centered_h, full_matrices=False)
sin_value_h_squared = torch.diag(sin_value_h[:svd]) ** 2 sin_value_h_squared = torch.diag(sin_value_h[:k]) ** 2
V_p_h = V_p_h[:svd, :] V_p_h = V_p_h[:k, :]
C_h=(1 / centered_h.shape[0])* V_p_h.T @ sin_value_h_squared @ V_p_h C_h=(1 / centered_h.shape[0])* V_p_h.T @ sin_value_h_squared @ V_p_h
# print(centered_t.shape) # print(centered_t.shape)
_, sin_value_t, V_p_t = torch.linalg.svd(centered_t, full_matrices=False) _, sin_value_t, V_p_t = torch.linalg.svd(centered_t, full_matrices=False)
sin_value_t_squared = torch.diag(sin_value_t[:svd]) ** 2 sin_value_t_squared = torch.diag(sin_value_t[:k]) ** 2
V_p_t = V_p_t[:svd, :] V_p_t = V_p_t[:k, :]
C_t=(1 / centered_t.shape[0])* V_p_t.T @ sin_value_t_squared @ V_p_t C_t=(1 / centered_t.shape[0])* V_p_t.T @ sin_value_t_squared @ V_p_t
inv_C= torch.linalg.inv(C) + torch.eye(C.shape[0], dtype=int).cuda() * epsilon
inv_C_t= torch.linalg.inv(C_t) + torch.eye(C_t.shape[0], dtype=int).cuda() * epsilon
inv_C_h= torch.linalg.inv(C_h) + torch.eye(C_h.shape[0], dtype=int).cuda() * epsilon
scores= torch.sqrt(centered @ inv_C @ centered.T) - torch.sqrt(torch.from_numpy(embed_generated[:, layer, :]-mean_t).cuda() @ inv_C_t @ torch.from_numpy(embed_generated[:, layer, :]-mean_t).cuda().T) inv_C_t= torch.linalg.pinv(C_t) + torch.eye(C_t.shape[0], dtype=int).cuda() * epsilon
+ torch.sqrt(torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda() @ inv_C_h @ torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda().T) inv_C_h= torch.linalg.pinv(C_h) + torch.eye(C_h.shape[0], dtype=int).cuda() * epsilon
print(torch.isnan(scores).any()) scores= torch.sqrt(torch.clamp(centered @ inv_C_t @ centered.T, min=0.0))
- torch.sqrt(torch.clamp(centered @ inv_C_h @ centered.T, min=0.0))
# scores= mahalanobis_distance(torch.from_numpy(embed_generated[:, layer, :]).cuda(), torch.from_numpy(mean_recorded).cuda(), C_) torch.clamp(centered @ inv_C_t @ centered.T, min=0.0)
# - mahalanobis_distance(torch.from_numpy(embed_generated[:, layer, :]).cuda(), torch.from_numpy(mean_t).cuda(), C_t)
# + mahalanobis_distance(torch.from_numpy(embed_generated[:, layer, :]).cuda(), torch.from_numpy(mean_h).cuda(), C_h) centered @ inv_C_h @ centered.T
scores = torch.mean(scores, -1, keepdim=True) scores = torch.mean(scores, -1, keepdim=True)
scores = torch.sqrt(torch.sum(torch.square(scores), dim=1)) scores = torch.sqrt(torch.sum(torch.square(scores), dim=1))
projection=V_p[:k, :].T # projection=V_p[:k, :].T
scores1 = torch.mean(centered @ projection, -1, keepdim=True) # scores1 = torch.mean(centered @ projection, -1, keepdim=True)
scores1 = torch.sqrt(torch.sum(torch.square(scores1), dim=1)) # scores1 = torch.sqrt(torch.sum(torch.square(scores1), dim=1))
print(torch.isnan(torch.sqrt(centered @ inv_C @ centered.T)).any())
print(torch.isnan(((torch.from_numpy(embed_generated[:, layer, :]-mean_t).cuda() @ torch.linalg.pinv(C_t) @ torch.from_numpy(embed_generated[:, layer, :]-mean_t).cuda().T)** 0.5).any()))
print(torch.isnan(((torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda() @ torch.linalg.pinv(C_h) @ torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda().T)** 0.5).any()))
# not sure about whether true and false data the direction will point to, # not sure about whether true and false data the direction will point to,
@ -444,7 +448,7 @@ def main():
best_auroc = measures[0] best_auroc = measures[0]
best_result = [100 * measures[2], 100 * measures[0]] best_result = [100 * measures[2], 100 * measures[0]]
best_layer = layer best_layer = layer
best_scores = sign_layer * scores
best_mean = mean_recorded best_mean = mean_recorded
best_sign = sign_layer best_sign = sign_layer
print('k: ', k, 'best result: ', best_result, 'layer: ', best_layer, print('k: ', k, 'best result: ', best_result, 'layer: ', best_layer,
@ -456,7 +460,7 @@ def main():
best_layer_over_k = best_layer best_layer_over_k = best_layer
best_k = k best_k = k
best_sign_over_k = best_sign best_sign_over_k = best_sign
# best_scores_over_k = best_scores best_scores_over_k = best_scores
# best_projection_over_k = best_projection # best_projection_over_k = best_projection
best_mean_over_k = best_mean best_mean_over_k = best_mean
@ -534,7 +538,7 @@ def main():
# 1, 11, mean=0, svd=0, weight=args.weighted_svd) # 1, 11, mean=0, svd=0, weight=args.weighted_svd)
# get the best hyper-parameters on validation set # get the best hyper-parameters on validation set
returned_results = svd_embed_score(embed_generated_eval, gt_label_val, embed_generated_hal,embed_generated_tru, returned_results = svd_embed_score(embed_generated_eval, gt_label_val, embed_generated_hal,embed_generated_tru,
1, 11, mean=1, svd=10) 1, 15, mean=1, svd=10)
pca_model = PCA(n_components=returned_results['k'], whiten=False).fit(embed_generated_wild[:,returned_results['best_layer'],:]) pca_model = PCA(n_components=returned_results['k'], whiten=False).fit(embed_generated_wild[:,returned_results['best_layer'],:])
projection = pca_model.components_.T projection = pca_model.components_.T

39
hal_generate.py
View File

@ -62,7 +62,7 @@ def main():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('--model_name', type=str, default='step-1-8k') parser.add_argument('--model_name', type=str, default='step-1-8k')
parser.add_argument('--dataset_name', type=str, default='tqa') parser.add_argument('--dataset_name', type=str, default='triviaqa')
parser.add_argument('--num_gene', type=int, default=1) parser.add_argument('--num_gene', type=int, default=1)
parser.add_argument('--use_api', type=bool, default=True) parser.add_argument('--use_api', type=bool, default=True)
parser.add_argument('--most_likely', type=bool, default=True) parser.add_argument('--most_likely', type=bool, default=True)
@ -175,6 +175,7 @@ def main():
raise ValueError("Invalid dataset name") raise ValueError("Invalid dataset name")
f = open(args.instruction, 'r', encoding="utf-8") f = open(args.instruction, 'r', encoding="utf-8")
instruction = f.read() instruction = f.read()
error_output='No output'
if args.use_api: if args.use_api:
begin_index = 0 begin_index = 0
@ -210,8 +211,9 @@ def main():
hallucination_prompt=get_hal_prompt(dataset[int(used_indices[i])]['context'],question,instruction) hallucination_prompt=get_hal_prompt(dataset[int(used_indices[i])]['context'],question,instruction)
truth_prompt=get_truth_prompt(dataset[int(used_indices[i])]['context'],question) truth_prompt=get_truth_prompt(dataset[int(used_indices[i])]['context'],question)
elif args.dataset_name == 'triviaqa': elif args.dataset_name == 'triviaqa':
prompt = get_qa_prompt("None",dataset[i]['prompt']) prompt = get_qa_prompt("None",dataset[i]['question'])
hallucination_prompt=get_hal_prompt("None",dataset[i]['prompt'],instruction) question= dataset[i]['question']
hallucination_prompt=get_hal_prompt("None",dataset[i]['question'],instruction)
truth_prompt=get_truth_prompt("None",question) truth_prompt=get_truth_prompt("None",question)
elif args.dataset_name == 'coqa': elif args.dataset_name == 'coqa':
prompt = get_qa_prompt("None",dataset[i]['prompt']) prompt = get_qa_prompt("None",dataset[i]['prompt'])
@ -223,31 +225,46 @@ def main():
for gen_iter in range(args.num_gene): for gen_iter in range(args.num_gene):
if args.most_likely: if args.most_likely:
response = client.chat.completions.create( try:
response = client.chat.completions.create(
model = args.model_name, model = args.model_name,
messages = prompt, messages = prompt,
max_tokens=256, max_tokens=256,
top_p=1, top_p=1,
temperature = 1, temperature = 1,
) )
hallucination_response = client.chat.completions.create( decoded=response.choices[0].message.content
except openai.APIStatusError as e:
print("error occured!"+str(gen_iter)+"responce {e}")
decoded = error_output
try:
hallucination_response = client.chat.completions.create(
model = args.model_name, model = args.model_name,
messages = hallucination_prompt, messages = hallucination_prompt,
max_tokens=256, max_tokens=256,
top_p=1, top_p=1,
temperature = 1, temperature = 1,
) )
hallucination_decoded=hallucination_response.choices[0].message.content
except openai.APIStatusError as e:
print("error occured!"+str(gen_iter)+"hallucination_responce {e}")
hallucination_decoded = error_output
if args.dataset_name == 'tydiqa' or args.dataset_name == 'tydiqa': if args.dataset_name == 'tydiqa' or args.dataset_name == 'tydiqa':
truth_response=client.chat.completions.create( try:
truth_response=client.chat.completions.create(
model = args.model_name, model = args.model_name,
messages = truth_prompt, messages = truth_prompt,
max_tokens=256, max_tokens=256,
top_p=1, top_p=1,
temperature=1 temperature=1
) )
truth_decoded=truth_response.choices[0].message.content truth_decoded=truth_response.choices[0].message.content
decoded=response.choices[0].message.content except openai.APIStatusError as e:
hallucination_decoded=hallucination_response.choices[0].message.content print("error occured!"+str(gen_iter)+"truth_responce {e}")
truth_decoded =error_output
else: else:
response = client.chat.completions.create( response = client.chat.completions.create(
model = args.model_name, model = args.model_name,
@ -275,10 +292,10 @@ def main():
top_p=0.5, top_p=0.5,
temperature = 0.5, temperature = 0.5,
) )
truth_decoded=truth_response.choices[0].message.content truth_decoded=truth_response.choices[0].message.content
decoded=response.choices[0].message.content decoded=response.choices[0].message.content
hallucination_decoded=hallucination_response.choices[0].message.content hallucination_decoded=hallucination_response.choices[0].message.content
time.sleep(20) time.sleep(40)
# decoded = tokenizer.decode(generated[0, prompt.shape[-1]:], # decoded = tokenizer.decode(generated[0, prompt.shape[-1]:],

6
hal_gt.py
View File

@ -50,14 +50,14 @@ def main():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, default='llama2_chat_7B') parser.add_argument('--model', type=str, default='llama2_chat_7B')
parser.add_argument('--model_name', type=str, default='step-1-8k') parser.add_argument('--model_name', type=str, default='moonshot-v1-8k')
parser.add_argument('--dataset_name', type=str, default='tqa') parser.add_argument('--dataset_name', type=str, default='tqa')
parser.add_argument('--num_gene', type=int, default=1) parser.add_argument('--num_gene', type=int, default=1)
parser.add_argument('--use_api', type=bool, default=False) parser.add_argument('--use_api', type=bool, default=False)
parser.add_argument('--most_likely', type=bool, default=True) parser.add_argument('--most_likely', type=bool, default=False)
parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data') parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data')
parser.add_argument("--instruction", type=str, default=None, help='local directory of instruction file.') parser.add_argument("--instruction", type=str, default=None, help='local directory of instruction file.')
parser.add_argument('--use_rouge', type=bool, default=False) parser.add_argument('--use_rouge', type=bool, default=True)
parser.add_argument('--thres_gt', type=float, default=0.5) parser.add_argument('--thres_gt', type=float, default=0.5)
# parser.add_argument('--model_name', type=str, default='llama2_chat_7B') # parser.add_argument('--model_name', type=str, default='llama2_chat_7B')