File size: 6,342 Bytes
5e56f2f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 |
import torch
import re
from configs.hyperparametric import Reward_config
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments
from model.logitsprocessor import OutputControlLogitsProcessor,RewardControlLogitsProcessor
from tree.asts import AST
#reward_config = Reward_config()
def find_subarray_positions(large_array, small_array):
n = len(small_array)
result = []
for i in range(len(large_array) - n + 1):
if large_array[i:i+n] == small_array:
result.append(i)
return result
class Warp():
def __init__(self,args,):
self.args = args
self.generate_tokenizer = None
self.generate_model = None
self.logits_processor = None
self.reward_tokenizer = None
self.reward_model = None
self.reward_processor = None
def load_generate_model(self):
args = self.args
self.generate_tokenizer = AutoTokenizer.from_pretrained(args.generate_model_path, trust_remote_code=True)
self.generate_model = AutoModelForCausalLM.from_pretrained(args.generate_model_path,trust_remote_code=True).half().cuda()
if args.logits_control:
syntax_tree = AST(path=args.control_file,tokenizer=self.generate_tokenizer)
self.logits_processor = OutputControlLogitsProcessor(ast=syntax_tree, tokenizer=self.generate_tokenizer)
if re.search('<think>',self.generate_tokenizer.chat_template):
self.generate_tokenizer.chat_template = re.sub('<think>','',self.generate_tokenizer.chat_template)
self.none_token = '<p>none</p>'
self.generate_bos = self.generate_tokenizer.bos_token
self.generate_eos = self.generate_tokenizer.eos_token
def load_reward_model(self,**kwargs):
args = self.args
if 'bnb_config' in kwargs.keys():
self.reward_model = AutoModelForCausalLM.from_pretrained(args.reward_model_path,
trust_remote_code=True,
device_map='auto',
quantization_config=kwargs['bnb_config'],
)
else:
self.reward_model = AutoModelForCausalLM.from_pretrained(args.reward_model_path,
trust_remote_code=True
).half().cuda()
self.reward_tokenizer = AutoTokenizer.from_pretrained(args.reward_model_path, trust_remote_code=True)
if args.logits_control:
self.reward_processor = RewardControlLogitsProcessor(tokenizer=self.generate_tokenizer)
if re.search('<think>',self.reward_tokenizer.chat_template):
self.reward_tokenizer.chat_template = re.sub('<think>','',self.reward_tokenizer.chat_template)
@staticmethod
def template_to_qwen(x,bos='<|im_start|>',eos='<|im_end|>'):
return """system\n{system}{eos}\n{bos}user\n{query}{eos}\n{bos}assistant\n""".format(
bos=bos,
eos=eos,
system="You are a helpful assistant.",
query=x
)
@staticmethod
def assistant_from_template_in_response(x,bos='<|im_start|>',eos='<|im_end|>'):
processed_string = x.split(bos)[3].strip()
processed_string = processed_string.split(eos)[0].strip()
processed_string = re.sub('^assistant','',processed_string).strip()
return processed_string
@staticmethod
def step_from_response(x):
step = '</none_response>'
if '</think>' in x:
x = x.split('</think>')[-1].strip()
if '<p>' in x:
x_ = re.search('<p>.+</p>',x)
step = x_.group() if x_ else x
if '<e>' in x:
x_ = re.search('<e>.+</e>',x)
step = x_.group() if x_ else x
return step
@staticmethod
def value_from_response(x):
if '<v>' in x:
x_ = re.search('<v>\w+</v>',x)
return x_.group()[3:-4] if x_ else ''
#return float(x_.group()[3:-4])/100 if x_ else 0
return ''
class WarpLJP(Warp):
def __init__(self,args,mode='p'):
super().__init__(args)
self.mode = mode
def processing_single(self,x,mode=''):
self.none_token = '<p>无</p>'
p = x['Procuratorate']
a = []
if 'd' in mode:
for d in x['Defence'].split('。'):
if len(d.strip()) > 0:
a.append(d)
if 'f' in mode:
for f in x['Fact'].split('。'):
if len(f.strip()) > 0:
a.append(f)
crime = [c['charge'] for c in x['Annotations'][0]['annotation']]
penalty = [c['penalty'] for c in x['Annotations'][0]['annotation']]
imprisonment = [c['imprisonment'] for c in x['Annotations'][0]['annotation']]
label = {'crime':';'.join(['<e>{c}罪</e>'.format(c=c) for c in crime]),
'penalty':';'.join(['<e>{c}</e>'.format(c=c) for c in penalty]),
'imprisonment':';'.join(['<e>{c}</e>'.format(c=c) for c in imprisonment])}
if a != []:
a = [_ for _ in map(lambda x_:'<p>{i}</p>'.format(i=x_) if not x_.startswith('<p>') else x_,a)]
return p,a,label
else:
return p,['<p>无</p>'],label
@staticmethod
def prompt_to_value(x,a,bos='<|im_start|>',eos='<|im_end|>'):
pmt = '根据案情描述对给出的已有推理步骤选择接受或拒绝，并在<v></v>中给出选择，例如<v>接受</v>\n案情描述:{x}\n已有推理步骤：\n{a}\n：'.format(x=x,a=a)
return Warp.template_to_qwen(pmt,bos=bos,eos=eos)
@staticmethod
def prompt_to_crime(x,a,bos='<|im_start|>',eos='<|im_end|>'):
pmt = '根据案情描述和已有步骤仅给出一个推理。如果是结论则直接输出<e></e>，例如<e>盗窃罪</e>。如果是步骤则直接输出<p></p>，例如<p>步骤1：…</p>\n案情描述:{x}\n已有推理步骤：\n{a}\n：'.format(x=x,a=a)
return Warp.template_to_qwen(pmt,bos=bos,eos=eos)
|