-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdata_processer.py
164 lines (147 loc) · 6.53 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
# @Time : 2023/3/25 18:36
# @Author : tk
import copy
import json
import random
import typing
from enum import Enum
import numpy as np
from deep_training.zoo.model_zoo.chatglm3.llm_model import ChatGLMTokenizer
class DataStrategy(Enum):
truncation = 1
siding = 2
class TokenIdsMaker:
def __init__(self,tokenizer: ChatGLMTokenizer, config):
self.tokenizer = tokenizer
self.config = config
self.bos_token_id = self.tokenizer.get_command("<bos>")
self.pad_token_id = self.tokenizer.get_command("<pad>")
self.eos_token_id = self.tokenizer.get_command("<eos>")
def build_single_message(self, role, metadata, message):
assert role in ["system", "user", "assistant", "observation"], role
role_tokens = [self.tokenizer.get_command(f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n")
message_tokens = self.tokenizer.encode(message)
tokens = role_tokens + message_tokens
return tokens
@classmethod
def final(cls, input_ids: typing.List, labels, max_seq_length, tokenizer):
input_ids = np.asarray(input_ids, dtype=np.int32)
labels = np.asarray(labels, dtype=np.int32)
seqlen = np.asarray(len(input_ids), dtype=np.int32)
pad_len = max_seq_length - seqlen
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,
}
return d
def build_chat_input(self, query, history=None, role="user"):
if history is None:
history = []
input_ids = []
for item in history:
content = item["content"]
input_ids.extend(self.build_single_message(item["role"], item.get("metadata", ""), content))
if query is not None:
input_ids.extend(self.build_single_message(role, "", query))
return self.tokenizer.encode(input_ids, is_split_into_words=True)
def parse_history_from_answers(self, output, history):
content = ""
metadata = ""
history = copy.deepcopy(history)
responses = output.split("<|assistant|>")
for response in responses:
#ensure 语料包含换行符 格式为{metadata}\n{content}
if len(responses) > 1:
metadata, content = response.split("\n", maxsplit=1)
else:
metadata = ""
content = response
if not metadata.strip():
content = content.strip()
history.append({"role": "assistant", "metadata": metadata, "content": content})
else:
history.append({"role": "assistant", "metadata": metadata, "content": content})
return metadata, content, history
def trunction(self, tokenizer: ChatGLMTokenizer,config, examples, max_seq_length,sup=True):
ds = []
history = []
for sid, (q_role,q,a) in enumerate(examples):
if q_role == "system":
prefix = {
"role": "system",
"content": q,
}
history += [prefix]
continue
if q_role == "function":
q_role = "observation"
if q_role != "observation":
q_role = "user"
history += [ {
"role": q_role,
"content": q,
} ]
a_ids = self.build_chat_input(query=None, history=history)
metadata, content, history = self.parse_history_from_answers(a,history)
b_ids = self.tokenizer.encode(self.tokenizer.encode(a),is_split_into_words=True)
role_tokens = [ self.tokenizer.get_command("<|assistant|>") ] + self.tokenizer.encode(f"{metadata}\n")
while len(a_ids) + len(b_ids) > max_seq_length - len(role_tokens) - 2:
if len(b_ids) > len(a_ids):
b_ids.pop(-1)
else:
a_ids.pop(0)
assert len(b_ids) > 0
b_ids += [ self.eos_token_id ]
a_ids = a_ids + role_tokens
input_ids = a_ids + b_ids
labels = copy.deepcopy(input_ids) if not sup else [ -100 ] * len(a_ids) + copy.deepcopy(b_ids)
input_ids = [self.bos_token_id] + input_ids
labels = [self.bos_token_id] + labels if not sup else [ -100 ] + labels
assert len(input_ids) <= max_seq_length
ds.append(self.final(input_ids, labels, max_seq_length, tokenizer))
return ds
# def slidding(cls, tokenizer: ChatGLMTokenizer,config, messages,
# max_seq_length,
# sliding_size = None,
# src_max_length=-1,
# dst_max_length=-1,
# sup=True):
#
#
# if sliding_size is None or sliding_size < 0:
# sliding_size = max_seq_length - 1
#
# assert sliding_size <= max_seq_length - 1
#
# ds = []
#
# for sid, (q, a) in enumerate(messages):
# a_ids = tokenizer.encode(text=build_template(q,prefix=prefix, history=examples[:sid]), add_special_tokens=False)
# b_ids = tokenizer.encode(text=a, add_special_tokens=False)
# 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 + b_ids
# labels_all = copy.deepcopy(input_ids_qa) if not sup else [-100] * len(a_ids) + b_ids
#
# pos = 0
# while pos < len(input_ids_qa):
# input_ids = input_ids_qa[pos:pos + max_seq_length - len(sptoken)]
# labels = labels_all[pos:pos + max_seq_length - len(sptoken)]
#
# pos += sliding_size
# if np.all(np.asarray(labels) == -100):
# continue
#
# input_ids = sptoken + input_ids
# labels = sptoken + labels if not sup else [-100] * len(sptoken) + labels
# ds.append(cls.final(input_ids,labels,max_seq_length,tokenizer))
# return ds