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

Compare commits

merge into: liwenyun:main
Branches Tags
liwenyun:main
liwenyun:jailbreak
liwenyun:3-hidden-state
...
pull from: liwenyun:3-hidden-state
Branches Tags
liwenyun:jailbreak
liwenyun:main
liwenyun:3-hidden-state

1 Commits
main ... 3-hidden-s

Author SHA1 Message Date
weixin_43297441 be4e5789d9 save branch 2025-03-05 16:56:01 +08:00
3 changed files with 16 additions and 4 deletions
Show Stats Expand all files Collapse all files

16
hal_det_llama.py
View File

@ -165,6 +165,8 @@ def main():
model = llama_iti.LlamaForCausalLM.from_pretrained(MODEL, low_cpu_mem_usage=True,
torch_dtype=torch.float16,
device_map="auto").cuda()
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)]
# firstly get the embeddings of the generated question and answers.
embed_generated = []
embed_generated_h =[]
@ -252,8 +254,7 @@ def main():
embed_generated_h = np.asarray(np.stack(embed_generated_h), 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_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)]
embed_generated_loc2 = []
embed_generated_loc1 = []
for i in tqdm(range(length)):
@ -435,6 +436,10 @@ def main():
if feat_loc == 3:
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)
embed_generated_t = np.load(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_t_layer_wise.npy',
allow_pickle=True)
embed_generated_h = np.load(f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_h_layer_wise.npy',
allow_pickle=True)
elif feat_loc == 2:
embed_generated = np.load(
f'save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/' + info + f'{args.model_name}_gene_embeddings_mlp_wise.npy',
@ -458,12 +463,19 @@ def main():
elif i in wild_q_indices2:
feat_indices_eval.extend(np.arange(i, i + 1).tolist())
if feat_loc == 3:
# print(embed_generated.shape)
# print(embed_generated_h.shape)
# print(embed_generated_t.shape)
embed_generated_wild = embed_generated[feat_indices_wild][:,1:,:]
embed_generated_eval = embed_generated[feat_indices_eval][:, 1:, :]
embed_generated_hal,embed_generated_tru=embed_generated_h[feat_indices_wild][:,1:,:], embed_generated_t[feat_indices_wild][:,1:,:]
else:
embed_generated_wild = embed_generated[feat_indices_wild]
embed_generated_eval = embed_generated[feat_indices_eval]
# print(embed_generated.shape)
# print(embed_generated_h.shape)
# print(embed_generated_t.shape)
# print(feat_indices_wild)
embed_generated_hal,embed_generated_tru=embed_generated_h[feat_indices_wild], embed_generated_t[feat_indices_wild]

2
hal_generate.py
View File

@ -65,7 +65,7 @@ def main():
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)
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("--instruction", type=str, default='/home/liwenyun/code/haloscope/generation/qa/qa_one-turn_instruction.txt', help='local directory of instruction file.')
args = parser.parse_args()

2
hal_gt.py
View File

@ -54,7 +54,7 @@ def main():
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=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("--instruction", type=str, default=None, help='local directory of instruction file.')
parser.add_argument('--use_rouge', type=bool, default=False)