MORE/collator.py at main · dunzeng/MORE · GitHub

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
import torch
from arguments import TrainingArguments
from transformers import PreTrainedTokenizer, PreTrainedModel
from typing import List, Dict
from copy import deepcopy
from torch.nn.utils.rnn import pad_sequence
from utils import print_rank_0
from datasets import Dataset
from random import sample

def _llm_tokenize(prompts: List[str], texts: List[str], tokenizer: PreTrainedTokenizer, args: TrainingArguments) -> Dict[str, torch.Tensor]:
    if prompts is None:
        input_ids = []
        labels = []
        for text in zip(texts):
            text_ids = tokenizer.encode(text, add_special_tokens=False)
            label = deepcopy(text_ids)

            input_ids.append(torch.tensor(text_ids[-args.max_length:]))
            labels.append(torch.tensor(label[-args.max_length:]))

        input_ids = pad_sequence(input_ids, batch_first=True, padding_value=tokenizer.pad_token_id)
        if args.pad_labels_with_ignore:
            labels = pad_sequence(labels, batch_first=True, padding_value=args.ignore_token_id)
        else:
            labels = pad_sequence(labels, batch_first=True, padding_value=tokenizer.pad_token_id)
        attention_mask = torch.ne(input_ids, tokenizer.pad_token_id)
    else:
        input_ids = []
        labels = []
        for prompt, text in zip(prompts, texts):
            prompt_ids = tokenizer.encode(prompt, add_special_tokens=False)
            text_ids = tokenizer.encode(text, add_special_tokens=False)
            label = deepcopy(text_ids)
            response_start_idx = len(prompt_ids)
            if prompt_ids != text_ids[:response_start_idx]:
                response_start_idx -= 1
                prompt_ids = text_ids[:response_start_idx]
            text_ids = [tokenizer.bos_token_id] + text_ids + [tokenizer.eos_token_id]
            label = [tokenizer.bos_token_id] + label + [tokenizer.eos_token_id]
            if args.only_predict_answer:
                label[:len(prompt_ids) + 1] = [args.ignore_token_id] * (len(prompt_ids) + 1)

            input_ids.append(torch.tensor(text_ids[-args.max_length:]))
            labels.append(torch.tensor(label[-args.max_length:]))

        input_ids = pad_sequence(input_ids, batch_first=True, padding_value=tokenizer.pad_token_id)
        if args.pad_labels_with_ignore:
            labels = pad_sequence(labels, batch_first=True, padding_value=args.ignore_token_id)
        else:
            labels = pad_sequence(labels, batch_first=True, padding_value=tokenizer.pad_token_id)
        attention_mask = torch.ne(input_ids, tokenizer.pad_token_id)

    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }


def classfication_data_collator(tokenizer: PreTrainedTokenizer, args: TrainingArguments):
    def collator(examples):
        texts = []
        labels = []
        for example in examples:
            texts.append(example['text'])
            labels.append(args.label2id[example['label']])

        encodings = tokenizer(texts, padding=True, truncation=True, add_special_tokens=True)

        return {
            "input_ids": torch.tensor(encodings['input_ids']),
            "attention_mask": torch.tensor(encodings['attention_mask']),
            "labels": torch.tensor(labels)
        }

    return collator


def reward_data_collator(tokenizer: PreTrainedTokenizer, args: TrainingArguments):
    def collator(examples: List[Dict[str, List]]):
        batch_size = len(examples)
        num_sample = max([len(example['texts']) for example in examples])
        all_texts = []
        all_scores = []
        for example in examples:
            if len(example['texts']) < num_sample:
                example['texts'].extend([' ']*(num_sample - len(example['texts'])))
                example['scores'].extend([-100]*(num_sample - len(example['scores'])))
            all_texts.extend(example['texts'])
            all_scores.extend(example['scores'])

        encodings = tokenizer(all_texts, padding=True, truncation=True, add_special_tokens=True)

        return {
            "input_ids": torch.tensor(encodings['input_ids']).reshape(batch_size, num_sample, -1),
            "attention_mask": torch.tensor(encodings['attention_mask']).reshape(batch_size, num_sample, -1),
            "scores": torch.tensor(all_scores).reshape(batch_size, -1)
        }

    return collator


def sft_data_collator(tokenizer: PreTrainedTokenizer, args: TrainingArguments):
    def collator(examples):
        texts = []
        prompts = []
        for example in examples:
            text = example['prompt'] + example['answer']
            texts.append(text)
            prompts.append(example['prompt'])

        return _llm_tokenize(prompts, texts, tokenizer, args)

    return collator

def replace_last_occurrence(text):
    # 找到最后一次出现'Assistant:'的位置
    last_occurrence = text.rfind('Assistant:')
    if last_occurrence != -1:
        # 将最后一次出现的'Assistant:'替换为'Assistant:<sep>'
        text = text[:last_occurrence] + 'Assistant:<sep>' + text[last_occurrence + len('Assistant:'):]
    return text

def rjs_data_collator(tokenizer: PreTrainedTokenizer, args: TrainingArguments):
    def collator(examples):
        best_texts: List[str] = []
        for example in examples:
            texts = example['text']
            scores = example['score']
            best_text = texts[torch.argmax(torch.tensor(scores))]
            best_texts.append(best_text)

        if not args.only_predict_answer:
            prompts = None
            texts = best_texts
        else:
            prompts = []
            texts = []
            for text in best_texts:
                if "<sep>" not in text:
                    text = replace_last_occurrence(text)
                prompt, answer = text.split(args.sep_token)
                prompts.append(prompt)
                texts.append(prompt + answer)

        return _llm_tokenize(prompts, texts, tokenizer, args)

    return collator

def multi_reward_data_collator(batch):
    num_task = len(batch[0])
    batch_list = []
    for task_id in range(num_task):
        scores = []
        input_ids = []
        attention_mask = []
        for item in batch:
            scores.append(item[task_id]["score"])
            input_ids.append(item[task_id]["tokens"]["input_ids"])
            attention_mask.append(item[task_id]["tokens"]["attention_mask"])
        task_batch = {
            "score": torch.Tensor(scores).float(),
            "input_ids": torch.Tensor(input_ids).long(),
            "attention_mask": torch.Tensor(attention_mask).float(),
        }
        batch_list.append(task_batch)
    return batch_list