-
Notifications
You must be signed in to change notification settings - Fork 176
/
data_processer.py
179 lines (147 loc) · 6.34 KB
/
data_processer.py
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
# @Time : 2023/3/25 18:36
# @Author : tk
import copy
import random
import typing
from enum import Enum
import numpy as np
from deep_training.zoo.model_zoo.chatglm.llm_model import ChatGLMTokenizer
class DataStrategy(Enum):
truncation = 1
sliding = 2
def build_template_chatglm(query, answer = None,prefix=None, history=None):
prompt = prefix or ''
sid = 0
if history is not None:
for q, a in history:
prompt += "[Round {}]\n问：{}\n答：{}".format(sid,q, a)
sid += 1
prompt += query if sid == 0 else "[Round {}]\n问：{}\n答：".format(sid, query)
if answer is not None:
prompt += answer
return prompt
def build_template_chatglm2(query, answer = None,prefix=None, history=None):
prompt = prefix or ''
sid = 1
if history is not None:
for q, a in history:
prompt += "[Round {}]\n问：{}\n答：{}".format(sid,q, a)
sid += 1
prompt += "[Round {}]\n问：{}\n答：".format(sid, query)
if answer is not None:
prompt += answer
return prompt
def build_template_default(query, answer = None,prefix=None, history=None):
prompt = prefix or ''
if history is not None:
for q,a in history:
prompt += "User: {}\nAssistant:{}".format(q,a)
prompt += "User: {}\nAssistant:".format(query)
if answer is not None:
prompt += answer
return prompt
def build_template_tiger(query,answer = None,prefix=None, history=None):
prompt = prefix or ''
tok_ins = "\n\n### Instruction:\n"
tok_res = "\n\n### Response:\n"
if history is not None:
for q,a in history:
prompt += "{}{}{}{}".format(tok_ins,q,tok_res,a)
prompt += "{}{}{}".format(tok_ins, query, tok_res)
if answer is not None:
prompt += answer
return prompt
# 切换模版
build_template = build_template_chatglm
#对prompt 截断
class TokenIdsMaker:
@classmethod
def final(cls, input_ids: typing.List, sptoken, max_seq_length, tokenizer):
ctxlen = input_ids.index(sptoken[-1])
mask_position = ctxlen - 1
labels = [-100] * ctxlen + input_ids[mask_position + 1:]
seqlen = np.asarray(len(input_ids), dtype=np.int32)
pad_len = max_seq_length - seqlen
input_ids = np.asarray(input_ids, dtype=np.int32)
labels = np.asarray(labels, dtype=np.int32)
ctxlen = np.asarray(ctxlen, dtype=np.int32)
if pad_len:
pad_val = tokenizer.pad_token_id
input_ids = np.pad(input_ids, (0, pad_len), 'constant', constant_values=(pad_val, pad_val))
labels = np.pad(labels, (0, pad_len), 'constant', constant_values=(-100, -100))
d = {
'input_ids': input_ids,
'labels': labels,
'seqlen': seqlen,
'ctxlen': ctxlen
}
return d
@classmethod
def tunction(cls, tokenizer: ChatGLMTokenizer,config, examples, max_seq_length, sptoken: typing.List):
ds = []
prefix = None
history = [ ]
for sid, (q_role, q, a) in enumerate(examples):
if q_role == "system":
prefix = q
continue
history += [ (q, a) ]
a_ids = tokenizer.encode(text=build_template(q,prefix=prefix, history=history[:-1]), add_special_tokens=False)
b_ids = tokenizer.encode(text=a, add_special_tokens=False)
while len(a_ids) + len(b_ids) > max_seq_length - len(sptoken) - 1:
if len(b_ids) > len(a_ids):
b_ids.pop(-1)
else:
a_ids.pop(0)
b_ids += [config.eos_token_id]
input_ids = a_ids + sptoken + b_ids
assert len(input_ids) <= max_seq_length
ds.append(cls.final(input_ids, sptoken, max_seq_length, tokenizer))
return ds
@classmethod
def slidding(cls, tokenizer: ChatGLMTokenizer,config, examples, max_seq_length, sptoken: typing.List,
sliding_size=None,
src_max_length=-1,
dst_max_length=-1,p=1):
if sliding_size is None or sliding_size < 0:
sliding_size = max_seq_length - len(sptoken)
assert sliding_size <= max_seq_length - len(sptoken)
ds = []
prefix = None
history = [ ]
for sid, (q_role, q, a) in enumerate(examples):
if q_role == "system":
prefix = q
continue
history += [ (q, a) ]
a_ids = tokenizer.encode(text=build_template(q, prefix=prefix,history=history[:-1]), add_special_tokens=False)
b_ids = tokenizer.encode(text=a, add_special_tokens=False) + [config.eos_token_id]
if src_max_length and src_max_length > 0:
a_ids = a_ids[:src_max_length]
if dst_max_length and dst_max_length > 0:
b_ids = b_ids[:dst_max_length]
b_ids += [config.eos_token_id]
input_ids_qa = a_ids + sptoken + b_ids
a_length = len(a_ids)
pos = 0
while pos < len(input_ids_qa):
if pos + max_seq_length <= a_length:
input_ids = input_ids_qa[pos:pos + max_seq_length - 2]
if p > 0:
input_ids = input_ids[0:-p] + sptoken + input_ids[-p:]
else:
p = random.randint(0, max_seq_length - 2)
input_ids = input_ids[0:p] + sptoken + input_ids[p:]
elif sptoken[0] in input_ids_qa[pos:pos + max_seq_length]:
val = input_ids_qa[pos:pos + max_seq_length][-1]
if val == sptoken[-1]:
input_ids = input_ids_qa[pos + 1:pos + max_seq_length + 1]
elif val == sptoken[0]:
input_ids = input_ids_qa[pos + 2:pos + max_seq_length + 2]
else:
input_ids = input_ids_qa[pos:pos + max_seq_length]
else:
input_ids = sptoken + input_ids_qa[pos:pos + max_seq_length - 2]
pos += sliding_size
ds.append(cls.final(input_ids, sptoken, max_seq_length, tokenizer))
return ds