-
Notifications
You must be signed in to change notification settings - Fork 4
/
data_processer.py
151 lines (135 loc) · 6.27 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
# @Time : 2023/3/25 18:36
# @Author : tk
import copy
import random
import typing
from enum import Enum
import numpy as np
from transformers import PreTrainedTokenizer
class DataStrategy(Enum):
sup = 1
unsup = 2
sub_rounds = 3
mos_rounds = 4
class TokenIdsFinal:
@classmethod
def process(cls,tokenizer,input_ids,labels,max_seq_length):
seqlen = np.asarray(len(input_ids), dtype=np.int32)
pad_len = max_seq_length - seqlen
input_ids = np.asarray(input_ids, dtype=np.int32)
attention_mask = np.asarray([1] * len(input_ids), dtype=np.int32)
labels = np.asarray(labels, dtype=np.int32)
if pad_len:
pad_val = tokenizer.eos_token_id
input_ids = np.pad(input_ids, (0, pad_len), 'constant', constant_values=(pad_val, pad_val))
attention_mask = np.pad(attention_mask, (0, pad_len), 'constant', constant_values=(0, 0))
labels = np.pad(labels, (0, pad_len), 'constant', constant_values=(-100, -100))
d = {
'input_ids': input_ids,
'attention_mask': attention_mask,
'labels': labels,
'seqlen': seqlen
}
return d
class TokenUnSupervision:
@classmethod
def process(cls, tokenizer: PreTrainedTokenizer,config,stride, max_seq_length, examples):
input_ids_all = []
for idx, session in enumerate(examples):
question, answer = session['q'], session['a']
if isinstance(answer, list):
answer = '\n'.join(answer)
text = question + answer
ids = tokenizer.encode(text=text)
if len(ids) <= 3:
continue
input_ids_all += ids
# decoder_start_token_id = self.config.decoder_start_token_id
decoder_start_token_id = config.bos_token_id
pos = 0
ds = []
while pos < len(input_ids_all):
input_ids = [decoder_start_token_id] + input_ids_all[pos: pos + max_seq_length - 1]
pos += stride
if len(input_ids) <= 5:
continue
d = TokenIdsFinal.process(tokenizer,input_ids,copy.deepcopy(input_ids),max_seq_length)
ds.append(d)
return ds
class TokenSupervision:
@classmethod
def process(cls, tokenizer: PreTrainedTokenizer,config,stride, max_seq_length, examples):
ds = []
for idx, session in enumerate(examples):
question, answer = session['q'], session['a']
if isinstance(answer, list):
answer = '\n'.join(answer)
a_ids = tokenizer.encode(text=question,add_special_tokens=False)[:max_seq_length-2]
b_ids = tokenizer.encode(text=answer, add_special_tokens=False)
assert len(b_ids)
input_ids_all = a_ids + b_ids + [config.eos_token_id]
labels_all = [-100] * len(a_ids) + b_ids + [config.eos_token_id]
pos = 0
while pos < len(input_ids_all):
input_ids = [config.bos_token_id] + input_ids_all[pos: pos + max_seq_length - 1]
labels = [config.bos_token_id] + labels_all[pos: pos + max_seq_length - 1]
pos += stride
d = TokenIdsFinal.process(tokenizer, input_ids, labels, max_seq_length)
ds.append(d)
return ds
class TokenSupervisionRounds:
@classmethod
def process(cls, tokenizer: PreTrainedTokenizer,config,stride, max_seq_length, examples):
ds = []
prompt_text = ''
for idx, session in enumerate(examples):
question, answer = session['q'], session['a']
if isinstance(answer, list):
answer = '\n'.join(answer)
if idx == 0:
a_text = question
else:
a_text = prompt_text + "[Round {}]\n问：{}\n答：".format(idx, question)
prompt_text += "[Round {}]\n问：{}\n答：{}".format(idx, question, answer)
a_ids = tokenizer.encode(text=a_text,add_special_tokens=False)[:max_seq_length-2]
b_ids = tokenizer.encode(text=answer, add_special_tokens=False)
assert len(b_ids)
input_ids_all = a_ids + b_ids + [config.eos_token_id]
labels_all = [-100] * len(a_ids) + b_ids + [config.eos_token_id]
pos = 0
while pos < len(input_ids_all):
input_ids = [config.bos_token_id] + input_ids_all[pos: pos + max_seq_length - 1]
labels = [config.bos_token_id] + labels_all[pos: pos + max_seq_length - 1]
pos += stride
d = TokenIdsFinal.process(tokenizer, input_ids, labels, max_seq_length)
ds.append(d)
return ds
class TokenRoundsForMoss:
@classmethod
def process(cls, tokenizer: PreTrainedTokenizer,config,max_seq_length, examples):
meta_instruction = examples[0]
instruction_ids = tokenizer.encode(meta_instruction)
assert isinstance(instruction_ids, list) and len(instruction_ids) > 0
input_ids = copy.deepcopy(instruction_ids)
no_loss_spans = [(0, len(instruction_ids))]
for idx, session in enumerate(examples[1]):
cur_turn_ids = []
cur_no_loss_spans = []
for key, value in session.items():
cur_ids = tokenizer.encode(value)
if key == 'Tool Responses':
# The format tokens (<|Results|>:...<eor>\n) should have losses.
cur_no_loss_spans.append(
(len(input_ids + cur_turn_ids) + 5, len(input_ids + cur_turn_ids + cur_ids) - 2))
assert isinstance(cur_ids, list) and len(cur_ids) > 0
cur_turn_ids.extend(cur_ids)
if len(input_ids + cur_turn_ids) > max_seq_length - 1:
break
input_ids.extend(cur_turn_ids)
no_loss_spans.extend(cur_no_loss_spans)
input_ids.append(config.eos_token_id)
labels = np.asarray(copy.deepcopy(input_ids),dtype=np.int32)
for no_loss_span in no_loss_spans:
labels[no_loss_span[0]: no_loss_span[1]] = -100
d = TokenIdsFinal.process(tokenizer, input_ids, labels, max_seq_length)
return [d]