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

z

This commit is contained in:
weixin_43297441 2025-02-07 16:35:55 +08:00
parent fe1613a481
commit 0f40e33a1d
1 changed files with 57 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

96
hal_generate.py
View File

@ -18,8 +18,14 @@ from baukit import Trace, TraceDict
from metric_utils import get_measures, print_measures from metric_utils import get_measures, print_measures
import re import re
from torch.autograd import Variable from torch.autograd import Variable
from openai import OpenAI
import openai
API={
'gpt-3.5-turbo':{'base_url':"https://api.agicto.cn/v1",'key':''},
'deepseek-chat':{'base_url':"https://api.agicto.cn/v1",'key':''},
'qwen-turbo':{'base_url':"https://api.agicto.cn/v1",'key':''},
}
def seed_everything(seed: int): def seed_everything(seed: int):
import random, os import random, os
@ -53,7 +59,7 @@ def main():
parser.add_argument('--dataset_name', type=str, default='triviaqa') 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=False) parser.add_argument('--use_api', type=bool, default=False)
parser.add_argument('--most_likely', type=bool, default=False) parser.add_argument('--sample', 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')
args = parser.parse_args() args = parser.parse_args()
@ -152,9 +158,7 @@ def main():
raise ValueError("Invalid dataset name") raise ValueError("Invalid dataset name")
if not args.use_api: if not args.use_api:
tokenizer = llama_iti.LlamaTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = llama_iti.LlamaForCausalLM.from_pretrained(MODEL, low_cpu_mem_usage=True, torch_dtype=torch.float16,
device_map="auto").cuda()
begin_index = 0 begin_index = 0
if args.dataset_name == 'tydiqa': if args.dataset_name == 'tydiqa':
@ -162,50 +166,64 @@ def main():
else: else:
end_index = len(dataset) end_index = len(dataset)
if not os.path.exists(f'./save_for_eval/{args.dataset_name}_hal_det/'): if not os.path.exists(f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/'):
os.mkdir(f'./save_for_eval/{args.dataset_name}_hal_det/') os.mkdir(f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/')
if not os.path.exists(f'./save_for_eval/{args.dataset_name}_hal_det/answers'): if not os.path.exists(f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers'):
os.mkdir(f'./save_for_eval/{args.dataset_name}_hal_det/answers') os.mkdir(f'./save_for_eval/{args.dataset_name}/{args.model_name}_hal_det/answers')
period_token_id = [tokenizer(_)['input_ids'][-1] for _ in ['\n']] # period_token_id = [tokenizer(_)['input_ids'][-1] for _ in ['\n']]
period_token_id += [tokenizer.eos_token_id] # period_token_id += [tokenizer.eos_token_id]
for i in range(begin_index, end_index): for i in range(begin_index, end_index):
answers = [None] * args.num_gene answers = [None] * args.num_gene
if args.dataset_name == 'tydiqa': if args.dataset_name == 'tydiqa':
question = dataset[int(used_indices[i])]['question'] question = dataset[int(used_indices[i])]['question']
prompt = tokenizer( prompt = "Concisely answer the following question based on the information in the given passage: \n" + \
"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:", # prompt = tokenizer(
return_tensors='pt').input_ids.cuda() # "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:",
# return_tensors='pt').input_ids.cuda()
elif args.dataset_name == 'coqa': elif args.dataset_name == 'coqa':
prompt = tokenizer( prompt = dataset[i]['prompt']
dataset[i]['prompt'], return_tensors='pt').input_ids.cuda() # prompt = tokenizer(
# dataset[i]['prompt'], return_tensors='pt').input_ids.cuda()
else: else:
question = dataset[i]['question'] question = dataset[i]['question']
prompt = tokenizer(f"Answer the question concisely. Q: {question}" + " A:", return_tensors='pt').input_ids.cuda() prompt = f"Answer the question concisely. Q: {question}" + " A:"
# prompt = tokenizer(f"Answer the question concisely. Q: {question}" + " A:", return_tensors='pt').input_ids.cuda()
for gen_iter in range(args.num_gene): for gen_iter in range(args.num_gene):
if args.most_likely: if args.sample:
generated = model.generate(prompt, response = openai.Completion.create(engine=args.model_name,
num_beams=5, prompt=prompt,
num_return_sequences=1, max_tokens=50,
do_sample=False, n=1,
max_new_tokens=64, stop=None,
) top_p=1,
temperature=0.9,)
decoded=response.choices[0].text
# generated = model.generate(prompt,
# num_beams=5,
# num_return_sequences=1,
# do_sample=False,
# max_new_tokens=64,
# )
else: else:
generated = model.generate(prompt, response = openai.Completion.create(engine=args.model_name,
do_sample=True, prompt=prompt,
num_return_sequences=1, max_tokens=50,
num_beams=1, n=5,
max_new_tokens=64, best_of=1,
temperature=0.5, stop=None,
top_p=1.0) top_p=0.5,
temperature=0.5,)
decoded=response.choices[0].text
# decoded = tokenizer.decode(generated[0, prompt.shape[-1]:],
decoded = tokenizer.decode(generated[0, prompt.shape[-1]:], # skip_special_tokens=True)
skip_special_tokens=True)
if args.dataset_name == 'tqa' or args.dataset_name == 'triviaqa': if args.dataset_name == 'tqa' or args.dataset_name == 'triviaqa':
# corner case. # corner case.
if 'Answer the question concisely' in decoded: if 'Answer the question concisely' in decoded:
@ -232,10 +250,10 @@ def main():
np.save(f'./save_for_eval/{args.dataset_name}_hal_det/answers/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy', np.save(f'./save_for_eval/{args.dataset_name}_hal_det/answers/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy',
answers) answers)
else: else:
tokenizer = llama_iti.LlamaTokenizer.from_pretrained(MODEL, trust_remote_code=True) client = OpenAI(
model = llama_iti.LlamaForCausalLM.from_pretrained(MODEL, low_cpu_mem_usage=True, api_key=API[args.model_name]['key'],
torch_dtype=torch.float16, base_url =API[args.model_name]['base_url']
device_map="auto").cuda() )
# firstly get the embeddings of the generated question and answers. # firstly get the embeddings of the generated question and answers.
embed_generated = [] embed_generated = []