new up
This commit is contained in:
weixin_43297441
parent
c806a931d1
commit
940134f4c9
204
hal_det_llama.py
204
hal_det_llama.py
|
|
@ -58,7 +58,7 @@ def main():
|
|||
parser.add_argument('--num_gene', type=int, default=1)
|
||||
parser.add_argument('--use_rouge', type=bool, default= False)
|
||||
parser.add_argument('--weighted_svd', type=int, default=0)
|
||||
parser.add_argument('--feat_loc_svd', type=int, default=0)
|
||||
parser.add_argument('--feat_loc_svd', type=int, default=1)
|
||||
parser.add_argument('--wild_ratio', type=float, default=0.75)
|
||||
parser.add_argument('--thres_gt', type=float, default=0.5)
|
||||
parser.add_argument('--most_likely', type=bool, default=True)
|
||||
|
|
@ -171,10 +171,7 @@ def main():
|
|||
embed_generated = []
|
||||
embed_generated_h =[]
|
||||
embed_generated_t=[]
|
||||
embed_generated_t_loc2 = []
|
||||
embed_generated_t_loc1 = []
|
||||
embed_generated_h_loc2 = []
|
||||
embed_generated_h_loc1 = []
|
||||
|
||||
|
||||
if args.dataset_name == 'tydiqa':
|
||||
length = len(used_indices)
|
||||
|
|
@ -191,10 +188,6 @@ def main():
|
|||
info = 'batch_generations_'
|
||||
answers = np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
|
||||
truths= np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/truths/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_truths_index_{i}.npy')
|
||||
hallucinations= np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/hallucinations/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_hallucinations_index_{i}.npy')
|
||||
for anw in answers:
|
||||
|
||||
if args.dataset_name == 'tydiqa':
|
||||
|
|
@ -216,63 +209,10 @@ def main():
|
|||
embed_generated = np.asarray(np.stack(embed_generated), dtype=np.float32)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_layer_wise.npy', embed_generated)
|
||||
|
||||
for tru in truths:
|
||||
|
||||
if args.dataset_name == 'tydiqa':
|
||||
prompt = tokenizer(
|
||||
"Concisely answer the following question based on the information in the given passage: \n" + \
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:" + tru,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
elif args.dataset_name == 'coqa':
|
||||
prompt = tokenizer(dataset[i]['prompt'] + tru, return_tensors='pt').input_ids.cuda()
|
||||
else:
|
||||
prompt = tokenizer(
|
||||
f"Answer the question concisely. Q: {question}" + " A:" + tru,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
with torch.no_grad():
|
||||
with TraceDict(model, HEADS + MLPS) as ret:
|
||||
output = model(prompt, output_hidden_states=True)
|
||||
head_wise_hidden_states = [ret[head].output.squeeze().detach().cpu() for head in HEADS]
|
||||
head_wise_hidden_states = torch.stack(head_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
mlp_wise_hidden_states = [ret[mlp].output.squeeze().detach().cpu() for mlp in MLPS]
|
||||
mlp_wise_hidden_states = torch.stack(mlp_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
|
||||
embed_generated_t_loc2.append(mlp_wise_hidden_states[:, -1, :])
|
||||
embed_generated_t_loc1.append(head_wise_hidden_states[:, -1, :])
|
||||
embed_generated_t_loc2 = np.asarray(np.stack(embed_generated_t_loc2), dtype=np.float32)
|
||||
embed_generated_t_loc1 = np.asarray(np.stack(embed_generated_t_loc1), dtype=np.float32)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_head_wise.npy', embed_generated_t_loc2)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_mlp_wise.npy', embed_generated_t_loc1)
|
||||
|
||||
for hal in hallucinations:
|
||||
|
||||
if args.dataset_name == 'tydiqa':
|
||||
prompt = tokenizer(
|
||||
"Concisely answer the following question based on the information in the given passage: \n" + \
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:" + hal,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
elif args.dataset_name == 'coqa':
|
||||
prompt = tokenizer(dataset[i]['prompt'] + hal, return_tensors='pt').input_ids.cuda()
|
||||
else:
|
||||
prompt = tokenizer(
|
||||
f"Answer the question concisely. Q: {question}" + " A:" + hal,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
with torch.no_grad():
|
||||
with TraceDict(model, HEADS + MLPS) as ret:
|
||||
output = model(prompt, output_hidden_states=True)
|
||||
head_wise_hidden_states = [ret[head].output.squeeze().detach().cpu() for head in HEADS]
|
||||
head_wise_hidden_states = torch.stack(head_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
mlp_wise_hidden_states = [ret[mlp].output.squeeze().detach().cpu() for mlp in MLPS]
|
||||
mlp_wise_hidden_states = torch.stack(mlp_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
|
||||
embed_generated_h_loc2.append(mlp_wise_hidden_states[:, -1, :])
|
||||
embed_generated_h_loc1.append(head_wise_hidden_states[:, -1, :])
|
||||
embed_generated_h_loc2 = np.asarray(np.stack(embed_generated_t_loc2), dtype=np.float32)
|
||||
embed_generated_h_loc1 = np.asarray(np.stack(embed_generated_t_loc1), dtype=np.float32)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_head_wise.npy', embed_generated_h_loc2)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_mlp_wise.npy', embed_generated_h_loc1)
|
||||
|
||||
|
||||
embed_generated_t_loc2 = []
|
||||
embed_generated_t_loc1 = []
|
||||
embed_generated_h_loc2 = []
|
||||
embed_generated_h_loc1 = []
|
||||
embed_generated_loc2 = []
|
||||
embed_generated_loc1 = []
|
||||
for i in tqdm(range(length)):
|
||||
|
|
@ -284,11 +224,15 @@ def main():
|
|||
|
||||
answers = np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
|
||||
truths= np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/truths/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_truths_index_{i}.npy')
|
||||
hallucinations= np.load(
|
||||
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/hallucinations/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_hallucinations_index_{i}.npy')
|
||||
for anw in answers:
|
||||
if args.dataset_name == 'tydiqa':
|
||||
prompt = tokenizer(
|
||||
"Concisely answer the following question based on the information in the given passage: \n" + \
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:",
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:" + anw,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
elif args.dataset_name == 'coqa':
|
||||
prompt = tokenizer(dataset[i]['prompt'] + anw, return_tensors='pt').input_ids.cuda()
|
||||
|
|
@ -307,13 +251,66 @@ def main():
|
|||
|
||||
embed_generated_loc2.append(mlp_wise_hidden_states[:, -1, :])
|
||||
embed_generated_loc1.append(head_wise_hidden_states[:, -1, :])
|
||||
for hal in hallucinations:
|
||||
if args.dataset_name == 'tydiqa':
|
||||
prompt = tokenizer(
|
||||
"Concisely answer the following question based on the information in the given passage: \n" + \
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:" + hal,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
elif args.dataset_name == 'coqa':
|
||||
prompt = tokenizer(dataset[i]['prompt'] + hal, return_tensors='pt').input_ids.cuda()
|
||||
else:
|
||||
prompt = tokenizer(
|
||||
f"Answer the question concisely. Q: {question}" + " A:" + hal,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
|
||||
with torch.no_grad():
|
||||
with TraceDict(model, HEADS + MLPS) as ret:
|
||||
output = model(prompt, output_hidden_states=True)
|
||||
head_wise_hidden_states = [ret[head].output.squeeze().detach().cpu() for head in HEADS]
|
||||
head_wise_hidden_states = torch.stack(head_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
mlp_wise_hidden_states = [ret[mlp].output.squeeze().detach().cpu() for mlp in MLPS]
|
||||
mlp_wise_hidden_states = torch.stack(mlp_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
|
||||
embed_generated_h_loc2.append(mlp_wise_hidden_states[:, -1, :])
|
||||
embed_generated_h_loc1.append(head_wise_hidden_states[:, -1, :])
|
||||
for tru in truths:
|
||||
if args.dataset_name == 'tydiqa':
|
||||
prompt = tokenizer(
|
||||
"Concisely answer the following question based on the information in the given passage: \n" + \
|
||||
" Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:" + tru,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
elif args.dataset_name == 'coqa':
|
||||
prompt = tokenizer(dataset[i]['prompt'] + tru, return_tensors='pt').input_ids.cuda()
|
||||
else:
|
||||
prompt = tokenizer(
|
||||
f"Answer the question concisely. Q: {question}" + " A:" + tru,
|
||||
return_tensors='pt').input_ids.cuda()
|
||||
|
||||
with torch.no_grad():
|
||||
with TraceDict(model, HEADS + MLPS) as ret:
|
||||
output = model(prompt, output_hidden_states=True)
|
||||
head_wise_hidden_states = [ret[head].output.squeeze().detach().cpu() for head in HEADS]
|
||||
head_wise_hidden_states = torch.stack(head_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
mlp_wise_hidden_states = [ret[mlp].output.squeeze().detach().cpu() for mlp in MLPS]
|
||||
mlp_wise_hidden_states = torch.stack(mlp_wise_hidden_states, dim=0).squeeze().numpy()
|
||||
|
||||
embed_generated_t_loc2.append(mlp_wise_hidden_states[:, -1, :])
|
||||
embed_generated_t_loc1.append(head_wise_hidden_states[:, -1, :])
|
||||
|
||||
embed_generated_loc2 = np.asarray(np.stack(embed_generated_loc2), dtype=np.float32)
|
||||
embed_generated_loc1 = np.asarray(np.stack(embed_generated_loc1), dtype=np.float32)
|
||||
embed_generated_h_loc2 = np.asarray(np.stack(embed_generated_h_loc2), dtype=np.float32)
|
||||
embed_generated_h_loc1 = np.asarray(np.stack(embed_generated_h_loc1), dtype=np.float32)
|
||||
embed_generated_t_loc2 = np.asarray(np.stack(embed_generated_t_loc2), dtype=np.float32)
|
||||
embed_generated_t_loc1 = np.asarray(np.stack(embed_generated_t_loc1), dtype=np.float32)
|
||||
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_head_wise.npy', embed_generated_loc1)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_embeddings_mlp_wise.npy', embed_generated_loc2)
|
||||
|
||||
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_head_wise.npy', embed_generated_h_loc2)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_mlp_wise.npy', embed_generated_h_loc1)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_head_wise.npy', embed_generated_t_loc2)
|
||||
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_mlp_wise.npy', embed_generated_t_loc1)
|
||||
|
||||
# get the split and label (true or false) of the unlabeled data and the test data.
|
||||
if args.use_rouge:
|
||||
|
|
@ -355,7 +352,7 @@ def main():
|
|||
|
||||
|
||||
|
||||
def svd_embed_score(embed_generated_wild, gt_label,embed_generated_h,embed_generated_t, begin_k, k_span, mean=1, svd=1, weight=0):
|
||||
def svd_embed_score(embed_generated_wild, gt_label,embed_generated_h,embed_generated_t, begin_k, k_span, mean=1, svd=10, epsilon=1e-20):
|
||||
embed_generated = embed_generated_wild
|
||||
# embed_hallucination= embed_generated_h
|
||||
best_auroc_over_k = 0
|
||||
|
|
@ -369,37 +366,68 @@ def main():
|
|||
mean_recorded = None
|
||||
# best_projection = None
|
||||
for layer in range(len(embed_generated_wild[0])):
|
||||
# print(len(embed_generated_wild[0]))
|
||||
if mean:
|
||||
mean_recorded = embed_generated[:, layer, :].mean(0)
|
||||
centered = embed_generated[:, layer, :] - mean_recorded
|
||||
mean_h=embed_generated_h[:, layer, :].mean(0)
|
||||
centered_h=embed_generated_h[:, layer, :]-mean_h
|
||||
mean_t=embed_generated_t[:, layer, :].mean(0)
|
||||
centered_t=embed_generated_t[:, layer, :]-mean_t
|
||||
else:
|
||||
centered = embed_generated[:, layer, :]
|
||||
|
||||
mean_h=embed_generated_h[:, layer, :].mean(0)
|
||||
centered_h=embed_generated_h[:, layer, :]-mean_h
|
||||
|
||||
mean_t=embed_generated_t[:, layer, :].mean(0)
|
||||
centered_t=embed_generated_t[:, layer, :].mean(0)-mean_t
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
# if not svd:
|
||||
# assert "Not implemented!"
|
||||
# else:
|
||||
_, sin_value, V_p = torch.linalg.svd(torch.from_numpy(centered).cuda())
|
||||
C=(1 / centered.shape[0])* V_p.T.cpu().data.numpy() @ np.diag(sin_value.cpu().data.numpy() ** 2)
|
||||
_, sin_value_h, V_p_h = torch.linalg.svd(torch.from_numpy(centered_h).cuda())
|
||||
C_h=(1 / centered_h.shape[0])* V_p_h.T.cpu().data.numpy() @ np.diag(sin_value_h.cpu().data.numpy() ** 2)
|
||||
_, sin_value_t, V_p_t = torch.linalg.svd(torch.from_numpy(centered_t).cuda())
|
||||
C_t=(1 / centered_t.shape[0])* V_p_t.T.cpu().data.numpy() @ np.diag(sin_value_t.cpu().data.numpy() ** 2)
|
||||
scores= (centered*np.invert(C)*centered.T) ** 0.5 - ((embed_generated[:, layer, :]-mean_t)*np.invert(C_t)*(embed_generated[:, layer, :]-mean_t).T) ** 0.5
|
||||
+ ((embed_generated[:, layer, :]-mean_h)*np.invert(C_h)*(embed_generated[:, layer, :]-mean_h).T) ** 0.5
|
||||
centered=torch.from_numpy(centered).cuda()
|
||||
centered_h=torch.from_numpy(centered_h).cuda()
|
||||
centered_t=torch.from_numpy(centered_t).cuda()
|
||||
_, sin_value, V_p = torch.linalg.svd(centered, full_matrices=False)
|
||||
sin_value_squared = torch.diag(sin_value[:svd]) ** 2
|
||||
V_p = V_p[:svd, :]
|
||||
C=(1 / centered.shape[0])* V_p.T @ sin_value_squared @ V_p
|
||||
|
||||
# clf = Perceptron(tol=1e-3, random_state=0)
|
||||
|
||||
_, 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
|
||||
V_p_h = V_p_h[:svd, :]
|
||||
C_h=(1 / centered_h.shape[0])* V_p_h.T @ sin_value_h_squared @ V_p_h
|
||||
|
||||
# print(centered_t.shape)
|
||||
_, 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
|
||||
V_p_t = V_p_t[:svd, :]
|
||||
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)
|
||||
+ torch.sqrt(torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda() @ inv_C_h @ torch.from_numpy(embed_generated[:, layer, :]-mean_h).cuda().T)
|
||||
print(torch.isnan(scores).any())
|
||||
scores = torch.mean(scores, -1, keepdim=True)
|
||||
scores = torch.sqrt(torch.sum(torch.square(scores), dim=1))
|
||||
|
||||
projection=V_p[:k, :].T
|
||||
scores1 = torch.mean(centered @ projection, -1, keepdim=True)
|
||||
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,
|
||||
# so we test both. similar practices are in the representation engineering paper
|
||||
# https://arxiv.org/abs/2310.01405
|
||||
scores=scores.data.cpu().numpy()
|
||||
# scores1=scores1.data.cpu().numpy()
|
||||
measures1 = get_measures(scores[gt_label == 1],
|
||||
scores[gt_label == 0], plot=False)
|
||||
measures2 = get_measures(-scores[gt_label == 1],
|
||||
|
|
@ -416,8 +444,7 @@ def main():
|
|||
best_auroc = measures[0]
|
||||
best_result = [100 * measures[2], 100 * measures[0]]
|
||||
best_layer = layer
|
||||
# best_scores = sign_layer * scores
|
||||
# best_projection = projection
|
||||
|
||||
best_mean = mean_recorded
|
||||
best_sign = sign_layer
|
||||
print('k: ', k, 'best result: ', best_result, 'layer: ', best_layer,
|
||||
|
|
@ -494,7 +521,10 @@ def main():
|
|||
else:
|
||||
embed_generated_wild = embed_generated[feat_indices_wild]
|
||||
embed_generated_eval = embed_generated[feat_indices_eval]
|
||||
embed_generated_hal,embed_generated_tru=embed_generated_h[feat_indices_wild], embed_generated_t[feat_indices_wild]
|
||||
# print(embed_generated.shape)
|
||||
# print(embed_generated_h.shape)
|
||||
# print(embed_generated_t.shape)
|
||||
embed_generated_hal,embed_generated_tru=embed_generated_h[feat_indices_eval], embed_generated_t[feat_indices_eval]
|
||||
|
||||
|
||||
|
||||
|
|
@ -504,7 +534,7 @@ def main():
|
|||
# 1, 11, mean=0, svd=0, weight=args.weighted_svd)
|
||||
# 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,
|
||||
1, 11, mean=1, svd=1, wei1ght=args.weighted_svd)
|
||||
1, 11, mean=1, svd=10)
|
||||
|
||||
pca_model = PCA(n_components=returned_results['k'], whiten=False).fit(embed_generated_wild[:,returned_results['best_layer'],:])
|
||||
projection = pca_model.components_.T
|
||||
|
|
|
|||
Loading…
Reference in New Issue