File size: 6,180 Bytes
2852136
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple

from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
from .processor_utils import get_paligemma_token_type_ids, get_pixel_values


if TYPE_CHECKING:
    from transformers import PreTrainedTokenizer, ProcessorMixin

    from ...hparams import DataArguments
    from ..template import Template


logger = get_logger(__name__)


def _encode_pairwise_example(
    prompt: Sequence[Dict[str, str]],
    response: Sequence[Dict[str, str]],
    system: Optional[str],
    tools: Optional[str],
    template: "Template",
    tokenizer: "PreTrainedTokenizer",
    processor: Optional["ProcessorMixin"],
    data_args: "DataArguments",
) -> Tuple[List[int], List[int], List[int], List[int]]:
    if processor is not None and not hasattr(processor, "image_seq_length"):  # llava-like models
        prompt[0]["content"] = template.image_token + prompt[0]["content"]

    chosen_messages = prompt + [response[0]]
    rejected_messages = prompt + [response[1]]
    prompt_ids, chosen_ids = template.encode_oneturn(
        tokenizer, chosen_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
    )
    _, rejected_ids = template.encode_oneturn(
        tokenizer, rejected_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
    )

    if template.efficient_eos:
        chosen_ids += [tokenizer.eos_token_id]
        rejected_ids += [tokenizer.eos_token_id]

    if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
        prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids

    chosen_input_ids = prompt_ids + chosen_ids
    chosen_labels = [IGNORE_INDEX] * len(prompt_ids) + chosen_ids
    rejected_input_ids = prompt_ids + rejected_ids
    rejected_labels = [IGNORE_INDEX] * len(prompt_ids) + rejected_ids

    return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels


def preprocess_pairwise_dataset(
    examples: Dict[str, List[Any]],
    template: "Template",
    tokenizer: "PreTrainedTokenizer",
    processor: Optional["ProcessorMixin"],
    data_args: "DataArguments",
) -> Dict[str, List[List[int]]]:
    # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
    model_inputs = {
        "chosen_input_ids": [],
        "chosen_attention_mask": [],
        "chosen_labels": [],
        "rejected_input_ids": [],
        "rejected_attention_mask": [],
        "rejected_labels": [],
    }
    if processor is not None:
        model_inputs["pixel_values"] = []
        if hasattr(processor, "image_seq_length"):  # paligemma models
            model_inputs["chosen_token_type_ids"] = []
            model_inputs["rejected_token_type_ids"] = []

    for i in range(len(examples["prompt"])):
        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
            continue

        chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels = _encode_pairwise_example(
            prompt=examples["prompt"][i],
            response=examples["response"][i],
            system=examples["system"][i],
            tools=examples["tools"][i],
            template=template,
            tokenizer=tokenizer,
            processor=processor,
            data_args=data_args,
        )
        model_inputs["chosen_input_ids"].append(chosen_input_ids)
        model_inputs["chosen_attention_mask"].append([1] * len(chosen_input_ids))
        model_inputs["chosen_labels"].append(chosen_labels)
        model_inputs["rejected_input_ids"].append(rejected_input_ids)
        model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids))
        model_inputs["rejected_labels"].append(rejected_labels)
        if processor is not None:
            model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
            if hasattr(processor, "image_seq_length"):  # paligemma models
                model_inputs["chosen_token_type_ids"].append(
                    get_paligemma_token_type_ids(len(chosen_input_ids), processor)
                )
                model_inputs["rejected_token_type_ids"].append(
                    get_paligemma_token_type_ids(len(rejected_input_ids), processor)
                )

    return model_inputs


def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
    valid_chosen_labels = list(filter(lambda x: x != IGNORE_INDEX, example["chosen_labels"]))
    valid_rejected_labels = list(filter(lambda x: x != IGNORE_INDEX, example["rejected_labels"]))
    print("chosen_input_ids:\n{}".format(example["chosen_input_ids"]))
    print("chosen_inputs:\n{}".format(tokenizer.decode(example["chosen_input_ids"], skip_special_tokens=False)))
    print("chosen_label_ids:\n{}".format(example["chosen_labels"]))
    print("chosen_labels:\n{}".format(tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)))
    print("rejected_input_ids:\n{}".format(example["rejected_input_ids"]))
    print("rejected_inputs:\n{}".format(tokenizer.decode(example["rejected_input_ids"], skip_special_tokens=False)))
    print("rejected_label_ids:\n{}".format(example["rejected_labels"]))
    print("rejected_labels:\n{}".format(tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)))