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

new 6

This commit is contained in:
weixin_43297441 2025-03-03 11:48:07 +08:00
parent 3332163dd6
commit 24923b65f1
3 changed files with 45 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
hal_det_llama.py
View File

@ -52,11 +52,10 @@ def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name', type=str, default='llama2_chat_7B')
parser.add_argument('--dataset_name', type=str, default='triviaqa')
parser.add_argument('--model', type=str, default='llama2_chat_7B')
parser.add_argument('--model_name', type=str, default='step-1-8k')
parser.add_argument('--dataset_name', type=str, default='tqa')
parser.add_argument('--num_gene', type=int, default=1)
# parser.add_argument('--gene', type=int, default=0)
# parser.add_argument('--generate_gt', type=int, default=0)
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)
@ -67,7 +66,7 @@ def main():
parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data')
args = parser.parse_args()
MODEL = HF_NAMES[args.model_name] if not args.model_dir else args.model_dir
MODEL = HF_NAMES[args.model] if not args.model_dir else args.model_dir
@ -180,12 +179,16 @@ def main():
question = dataset[int(used_indices[i])]['question']
else:
question = dataset[i]['question']
if args.most_likely:
info = 'most_likely_'
else:
info = 'batch_generations_'
answers = np.load(
f'save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
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}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
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}_hal_det/hallucinations/most_likely_hal_det_{args.model_name}_{args.dataset_name}_hallucinations_index_{i}.npy')
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':
@ -205,7 +208,7 @@ def main():
hidden_states = hidden_states.detach().cpu().numpy()[:, -1, :]
embed_generated.append(hidden_states)
embed_generated = np.asarray(np.stack(embed_generated), dtype=np.float32)
np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_layer_wise.npy', embed_generated)
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:
@ -226,7 +229,7 @@ def main():
hidden_states = hidden_states.detach().cpu().numpy()[:, -1, :]
embed_generated_t.append(hidden_states)
embed_generated_t = np.asarray(np.stack(embed_generated_t), dtype=np.float32)
np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_t_layer_wise.npy', embed_generated_t)
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_layer_wise.npy', embed_generated_t)
for hal in hallucinations:
@ -247,7 +250,7 @@ def main():
hidden_states = hidden_states.detach().cpu().numpy()[:, -1, :]
embed_generated_h.append(hidden_states)
embed_generated_h = np.asarray(np.stack(embed_generated_h), dtype=np.float32)
np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_h_layer_wise.npy', embed_generated_h)
np.save(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_layer_wise.npy', embed_generated_h)
HEADS = [f"model.layers.{i}.self_attn.head_out" for i in range(model.config.num_hidden_layers)]
MLPS = [f"model.layers.{i}.mlp" for i in range(model.config.num_hidden_layers)]
@ -261,7 +264,7 @@ def main():
answers = np.load(
f'save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
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')
for anw in answers:
if args.dataset_name == 'tydiqa':
prompt = tokenizer(
@ -288,18 +291,18 @@ def main():
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)
np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_head_wise.npy', embed_generated_loc1)
np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{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_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)
# get the split and label (true or false) of the unlabeled data and the test data.
if args.use_rouge:
gts = np.load(f'./ml_{args.dataset_name}_rouge_score.npy')
gts_bg = np.load(f'./bg_{args.dataset_name}_rouge_score.npy')
gts = np.load(f'./ml_{args.dataset_name}_{args.model_name}_rouge_score.npy')
gts_bg = np.load(f'./bg_{args.dataset_name}_{args.model_name}_rouge_score.npy')
else:
gts = np.load(f'./ml_{args.dataset_name}_bleurt_score.npy')
gts_bg = np.load(f'./bg_{args.dataset_name}_bleurt_score.npy')
gts = np.load(f'./ml_{args.dataset_name}_{args.model_name}_bleurt_score.npy')
gts_bg = np.load(f'./bg_{args.dataset_name}_{args.model_name}_bleurt_score.npy')
thres = args.thres_gt
gt_label = np.asarray(gts> thres, dtype=np.int32)
gt_label_bg = np.asarray(gts_bg > thres, dtype=np.int32)
@ -430,15 +433,15 @@ def main():
if args.most_likely:
if feat_loc == 3:
embed_generated = np.load(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_layer_wise.npy',
embed_generated = np.load(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_layer_wise.npy',
allow_pickle=True)
elif feat_loc == 2:
embed_generated = np.load(
f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_mlp_wise.npy',
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_mlp_wise.npy',
allow_pickle=True)
else:
embed_generated = np.load(
f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_head_wise.npy',
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_head_wise.npy',
allow_pickle=True)
feat_indices_wild = []
feat_indices_eval = []

28
hal_generate.py
View File

@ -62,7 +62,7 @@ def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name', type=str, default='step-1-8k')
parser.add_argument('--dataset_name', type=str, default='triviaqa')
parser.add_argument('--dataset_name', type=str, default='tqa')
parser.add_argument('--num_gene', type=int, default=1)
parser.add_argument('--use_api', type=bool, default=True)
parser.add_argument('--most_likely', type=bool, default=True)
@ -226,14 +226,14 @@ def main():
response = client.chat.completions.create(
model = args.model_name,
messages = prompt,
# max_tokens=256,
max_tokens=256,
top_p=1,
temperature = 1,
)
hallucination_response = client.chat.completions.create(
model = args.model_name,
messages = hallucination_prompt,
# max_tokens=256,
max_tokens=256,
top_p=1,
temperature = 1,
)
@ -241,7 +241,7 @@ def main():
truth_response=client.chat.completions.create(
model = args.model_name,
messages = truth_prompt,
# max_tokens=256,
max_tokens=256,
top_p=1,
temperature=1
)
@ -252,7 +252,7 @@ def main():
response = client.chat.completions.create(
model = args.model_name,
messages = prompt,
# max_tokens=256,
max_tokens=256,
n=1,
# best_of=1,
top_p=0.5,
@ -300,22 +300,12 @@ def main():
truths[gen_iter]=truth_decoded
# if args.dataset_name == 'tydiqa':
# pass
# elif args.dataset_name == 'triviaqa':
# pass
if args.dataset_name == 'coqa':
truths[0]=dataset[i]['answer']
if args.num_gene >1 and dataset[i]['additional_answers']>= args.num_gene-1:
left_truth=dataset[i]['additional_answers'][:args.num_gene-1]
truths=truths+left_truth
truths=[dataset[i]['answer']]+dataset[i]['additional_answers']
truths=truths[:args.num_gene]
elif args.dataset_name == 'tqa':
truths[0]=dataset[i]['Best Answer']
if args.num_gene >1:
correct=dataset[i]['Correct Answers'].split(";")
if len(correct) >= args.num_gene-1:
left_truth=correct[:args.num_gene-1]
truths=truths+left_truth
truths=[dataset[i]['best_answer']]+dataset[i]['correct_answers']
truths=truths[:args.num_gene]
else:
assert 'Not supported dataset!'

23
hal_gt.py
View File

@ -49,14 +49,15 @@ def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name', type=str, default='llama2_chat_7B')
parser.add_argument('--dataset_name', type=str, default='triviaqa')
parser.add_argument('--model', type=str, default='llama2_chat_7B')
parser.add_argument('--model_name', type=str, default='step-1-8k')
parser.add_argument('--dataset_name', type=str, default='tqa')
parser.add_argument('--num_gene', type=int, default=1)
parser.add_argument('--use_api', type=bool, default=False)
parser.add_argument('--most_likely', type=bool, default=False)
parser.add_argument('--most_likely', type=bool, default=True)
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('--use_rouge', type=bool, default=True)
parser.add_argument('--use_rouge', type=bool, default=False)
parser.add_argument('--thres_gt', type=float, default=0.5)
# parser.add_argument('--model_name', type=str, default='llama2_chat_7B')
@ -74,7 +75,7 @@ def main():
# parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data')
args = parser.parse_args()
MODEL = HF_NAMES[args.model_name] if not args.model_dir else args.model_dir
MODEL = HF_NAMES[args.model] if not args.model_dir else args.model_dir
@ -198,10 +199,10 @@ def main():
if args.most_likely:
answers = np.load(
f'./save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
else:
answers = np.load(
f'./save_for_eval/{args.dataset_name}_hal_det/answers/batch_generations_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers/batch_generations_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy')
# get the gt.
if args.use_rouge:
@ -240,14 +241,14 @@ def main():
# breakpoint()
if args.most_likely:
if args.use_rouge:
np.save(f'./ml_{args.dataset_name}_rouge_score.npy', gts)
np.save(f'./ml_{args.dataset_name}_{args.model_name}_rouge_score.npy', gts)
else:
np.save(f'./ml_{args.dataset_name}_bleurt_score.npy', gts)
np.save(f'./ml_{args.dataset_name}_{args.model_name}_bleurt_score.npy', gts)
else:
if args.use_rouge:
np.save(f'./bg_{args.dataset_name}_rouge_score.npy', gts)
np.save(f'./bg_{args.dataset_name}_{args.model_name}_rouge_score.npy', gts)
else:
np.save(f'./bg_{args.dataset_name}_bleurt_score.npy', gts)
np.save(f'./bg_{args.dataset_name}_{args.model_name}_bleurt_score.npy', gts)