File size: 6,856 Bytes
236083b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
180
181
182
183
184
185
186
187
188
# Copyright Lightning AI. Licensed under the Apache License 2.0, see LICENSE file.

import json
from dataclasses import dataclass, field
from pathlib import Path

import torch
from torch.utils.data import DataLoader

from litgpt.data import DataModule, SFTDataset, get_sft_collate_fn
from litgpt.data.alpaca import download_if_missing
from litgpt.prompts import PromptStyle
from litgpt.tokenizer import Tokenizer

_URL = "https://huggingface.co/datasets/Muennighoff/flan/resolve/main"


# TODO: Including all subsets, FLAN is too large to be loaded in memory. Switch the implementation to cache
#   on disk or use Lightning Data
@dataclass
class FLAN(DataModule):
    """FLAN data module for supervised finetuning."""

    mask_prompt: bool = False
    """Whether to mask the prompt section from the label (with ``ignore_index``)."""
    prompt_style: str | PromptStyle = "flan"
    """The style to apply to instruction prompts. See `litgpt.prompts` for a list of available styles."""
    ignore_index: int = -100
    """The index to use for elements to be ignored in the label."""
    seed: int = 42
    """The random seed for shuffling the dataset."""
    num_workers: int = 4
    """How many DataLoader processes to use for loading."""
    download_dir: Path = Path("./data/flan")
    """The directory in which the downloaded dataset gets saved."""
    url: str = _URL
    """The URL from where to download the dataset."""
    subsets: str | None = None
    """A comma separated list of subsets to use. If None, all subsets are used."""

    tokenizer: Tokenizer | None = field(default=None, init=False, repr=False)
    batch_size: int = field(default=1, init=False, repr=False)
    max_seq_length: int = field(default=-1, init=False, repr=False)
    train_dataset: SFTDataset | None = field(default=None, init=False, repr=False)
    test_dataset: SFTDataset | None = field(default=None, init=False, repr=False)

    def __post_init__(self):
        super().__init__()
        if isinstance(self.prompt_style, str):
            self.prompt_style = PromptStyle.from_name(self.prompt_style)

        supported_subsets = _supported_subsets()
        if self.subsets is not None:
            self.subsets = self.subsets.split(",")
            for subset in self.subsets:
                if subset not in supported_subsets:
                    raise ValueError(f"{subset} not in {supported_subsets}")
        else:
            self.subsets = list(supported_subsets)

    def connect(
        self, tokenizer: Tokenizer | None = None, batch_size: int = 1, max_seq_length: int | None = None
    ) -> None:
        self.tokenizer = tokenizer
        self.batch_size = batch_size
        self.max_seq_length = -1 if max_seq_length is None else max_seq_length

    def prepare_data(self) -> None:
        self.download_dir.mkdir(parents=True, exist_ok=True)
        for subset in self.subsets:
            for split in ("train", "test"):
                data_file_path = self.download_dir / f"{subset}_{split}.jsonl"
                data_file_url = f"{self.url}/{split}/{subset}_{split}.jsonl"
                download_if_missing(data_file_path, data_file_url)

    def train_dataloader(self):
        return self._dataloader("train")

    def val_dataloader(self):
        return self._dataloader("test")

    def _dataloader(self, split: str) -> DataLoader:
        data = []
        for subset in self.subsets:
            data_file_path = self.download_dir / f"{subset}_{split}.jsonl"
            data.extend(load_jsonl(data_file_path))

        dataset = SFTDataset(
            data=data,
            tokenizer=self.tokenizer,
            prompt_style=self.prompt_style,
            max_seq_length=self.max_seq_length,
            mask_prompt=self.mask_prompt,
            ignore_index=self.ignore_index,
            transform=_transform,
        )
        return DataLoader(
            dataset=dataset,
            batch_size=self.batch_size,
            shuffle=(split == "train"),
            generator=torch.Generator().manual_seed(self.seed),
            num_workers=self.num_workers,
            collate_fn=get_sft_collate_fn(max_seq_length=self.max_seq_length, ignore_index=self.ignore_index),
        )


def load_jsonl(filename: Path) -> list[dict[str, str]]:
    data = []
    with open(filename, encoding="utf-8") as f:
        for line in f:
            data.append(json.loads(line))
    return data


def _transform(item: dict) -> dict:
    return {"instruction": item["inputs"], "input": "", "output": item["targets"]}


def _supported_subsets() -> set[str]:
    return {
        "aeslc_10templates",
        "ag_news_subset_10templates",
        "anli_r1_10templates",
        "anli_r2_10templates",
        "anli_r3_10templates",
        "arc_challenge_10templates",
        "arc_easy_10templates",
        "bool_q_10templates",
        "cb_10templates",
        "cnn_dailymail_10templates",
        "cola_10templates",
        "common_gen_10templates",
        "copa_10templates",
        "coqa_10templates",
        "cosmos_qa_10templates",
        "dart_10templates",
        "definite_pronoun_resolution_10templates",
        "drop_10templates",
        "e2e_nlg_10templates",
        "fix_punct_10templates",
        "gigaword_10templates",
        "glue_mrpc_10templates",
        "glue_qqp_10templates",
        "hellaswag_10templates",
        "imdb_reviews_10templates",
        "math_dataset_10templates",
        "mnli_matched_10templates",
        "mnli_mismatched_10templates",
        "multi_news_10templates",
        "multirc_10templates",
        "natural_questions_10templates",
        "openbookqa_10templates",
        "opinion_abstracts_idebate_10templates",
        "opinion_abstracts_rotten_tomatoes_10templates",
        "para_crawl_enes_10templates",
        "paws_wiki_10templates",
        "piqa_10templates",
        "qnli_10templates",
        "quac_10templates",
        "record_10templates",
        "rte_10templates",
        "samsum_10templates",
        "sentiment140_10templates",
        "snli_10templates",
        "squad_v1_10templates",
        "squad_v2_10templates",
        "sst2_10templates",
        "story_cloze_10templates",
        "stsb_10templates",
        "trec_10templates",
        "trivia_qa_10templates",
        "true_case_10templates",
        "web_nlg_en_10templates",
        "wic_10templates",
        "wiki_lingua_english_en_10templates",
        "wmt14_enfr_10templates",
        "wmt16_translate_csen_10templates",
        "wmt16_translate_deen_10templates",
        "wmt16_translate_fien_10templates",
        "wmt16_translate_roen_10templates",
        "wmt16_translate_ruen_10templates",
        "wmt16_translate_tren_10templates",
        "wnli_10templates",
        "word_segment_10templates",
        "wsc_10templates",
        "yelp_polarity_reviews_10templates",
    }