refactor utils.data module for line count linter (#1476)
Browse files
src/axolotl/utils/data/__init__.py
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Data processing modules
|
3 |
+
"""
|
4 |
+
from axolotl.utils.data.dpo import load_prepare_dpo_datasets # noqa: F401
|
5 |
+
from axolotl.utils.data.pretraining import ( # noqa: F401
|
6 |
+
encode_pretraining,
|
7 |
+
wrap_pretraining_dataset,
|
8 |
+
)
|
9 |
+
from axolotl.utils.data.sft import ( # noqa: F401
|
10 |
+
get_dataset_wrapper,
|
11 |
+
load_prepare_datasets,
|
12 |
+
load_tokenized_prepared_datasets,
|
13 |
+
prepare_dataset,
|
14 |
+
)
|
15 |
+
from axolotl.utils.data.utils import md5 # noqa: F401
|
src/axolotl/utils/data/dpo.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""data handling specific to DPO"""
|
2 |
+
|
3 |
+
import logging
|
4 |
+
from pathlib import Path
|
5 |
+
from typing import Any, List
|
6 |
+
|
7 |
+
import yaml
|
8 |
+
from datasets import concatenate_datasets, load_dataset, load_from_disk
|
9 |
+
|
10 |
+
from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
|
11 |
+
from axolotl.prompt_strategies.dpo import load as load_dpo
|
12 |
+
from axolotl.utils.data.utils import md5
|
13 |
+
from axolotl.utils.dict import DictDefault
|
14 |
+
from axolotl.utils.distributed import is_main_process, zero_first
|
15 |
+
|
16 |
+
LOG = logging.getLogger("axolotl")
|
17 |
+
|
18 |
+
|
19 |
+
def _get_path(ds_hash, cfg):
|
20 |
+
prepared_ds_path = (
|
21 |
+
Path(cfg.dataset_prepared_path) / ds_hash
|
22 |
+
if cfg.dataset_prepared_path
|
23 |
+
else Path(DEFAULT_DATASET_PREPARED_PATH) / ds_hash
|
24 |
+
)
|
25 |
+
|
26 |
+
return prepared_ds_path
|
27 |
+
|
28 |
+
|
29 |
+
def _load_preprocessed_ds(cfg, sub_cfg):
|
30 |
+
ds_hash = md5(yaml.dump(sub_cfg, Dumper=yaml.Dumper))
|
31 |
+
prepared_ds_path = _get_path(ds_hash, cfg)
|
32 |
+
dataset = None
|
33 |
+
|
34 |
+
# pylint: disable=duplicate-code
|
35 |
+
if (
|
36 |
+
cfg.dataset_prepared_path
|
37 |
+
and any(prepared_ds_path.glob("*"))
|
38 |
+
and not cfg.is_preprocess
|
39 |
+
):
|
40 |
+
LOG.info(f"Loading prepared dataset from disk at {prepared_ds_path}...")
|
41 |
+
dataset = load_from_disk(str(prepared_ds_path))
|
42 |
+
|
43 |
+
return dataset
|
44 |
+
|
45 |
+
|
46 |
+
def _save_preprocessed_ds(cfg, sub_cfg, dataset):
|
47 |
+
ds_hash = md5(yaml.dump(sub_cfg, Dumper=yaml.Dumper))
|
48 |
+
prepared_ds_path = _get_path(ds_hash, cfg)
|
49 |
+
|
50 |
+
if cfg.is_preprocess and is_main_process():
|
51 |
+
LOG.info(f"Loading prepared dataset from disk at {prepared_ds_path}...")
|
52 |
+
dataset.save_to_disk(str(prepared_ds_path))
|
53 |
+
|
54 |
+
|
55 |
+
def load_prepare_dpo_datasets(cfg):
|
56 |
+
def load_split(dataset_cfgs, _cfg):
|
57 |
+
split_datasets: List[Any] = []
|
58 |
+
for i, ds_cfg in enumerate(dataset_cfgs):
|
59 |
+
if ds_cfg["ds_type"] == "json":
|
60 |
+
for data_file in ds_cfg["data_files"]:
|
61 |
+
data_files = {ds_cfg["split"]: data_file}
|
62 |
+
ds = load_dataset( # pylint: disable=invalid-name
|
63 |
+
"json",
|
64 |
+
data_files=data_files,
|
65 |
+
split=ds_cfg["split"],
|
66 |
+
)
|
67 |
+
split_datasets.insert(i, ds)
|
68 |
+
else:
|
69 |
+
ds = load_dataset( # pylint: disable=invalid-name
|
70 |
+
ds_cfg["path"],
|
71 |
+
split=ds_cfg["split"],
|
72 |
+
)
|
73 |
+
split_datasets.insert(i, ds)
|
74 |
+
|
75 |
+
for i, data_set in enumerate(split_datasets):
|
76 |
+
_type = dataset_cfgs[i]["type"]
|
77 |
+
if _type:
|
78 |
+
if isinstance(_type, DictDefault):
|
79 |
+
_type = "user_defined.default"
|
80 |
+
ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
|
81 |
+
split_datasets[i] = data_set.map(
|
82 |
+
ds_transform_fn,
|
83 |
+
desc="Mapping RL Dataset",
|
84 |
+
)
|
85 |
+
else:
|
86 |
+
# If no `type` is provided, assume the dataset is already in the expected format with
|
87 |
+
# "prompt", "chosen" and "rejected" already preprocessed
|
88 |
+
split_datasets[i] = data_set
|
89 |
+
|
90 |
+
return concatenate_datasets(split_datasets)
|
91 |
+
|
92 |
+
with zero_first(is_main_process()):
|
93 |
+
train_is_preprocessed = False
|
94 |
+
eval_is_preprocessed = False
|
95 |
+
if train_dataset := _load_preprocessed_ds(cfg, cfg.datasets):
|
96 |
+
train_is_preprocessed = True
|
97 |
+
else:
|
98 |
+
train_dataset = load_split(cfg.datasets, cfg)
|
99 |
+
|
100 |
+
eval_dataset = None
|
101 |
+
if cfg.test_datasets:
|
102 |
+
if eval_dataset := _load_preprocessed_ds(cfg, cfg.test_datasets):
|
103 |
+
eval_is_preprocessed = True
|
104 |
+
else:
|
105 |
+
eval_dataset = load_split(cfg.test_datasets, cfg)
|
106 |
+
if not eval_dataset:
|
107 |
+
eval_dataset = None
|
108 |
+
|
109 |
+
if not train_is_preprocessed:
|
110 |
+
_save_preprocessed_ds(cfg, cfg.datasets, train_dataset)
|
111 |
+
if eval_dataset and not eval_is_preprocessed:
|
112 |
+
_save_preprocessed_ds(cfg, cfg.test_datasets, eval_dataset)
|
113 |
+
|
114 |
+
return train_dataset, eval_dataset
|
src/axolotl/utils/data/pretraining.py
ADDED
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""data handling specific to pretraining"""
|
2 |
+
|
3 |
+
import functools
|
4 |
+
import logging
|
5 |
+
from collections import defaultdict
|
6 |
+
from typing import Callable, Dict, List, Optional
|
7 |
+
|
8 |
+
import torch
|
9 |
+
from datasets import Dataset
|
10 |
+
from torch.utils.data import RandomSampler
|
11 |
+
from transformers import PreTrainedTokenizerBase
|
12 |
+
|
13 |
+
from axolotl.utils.collators import PretrainingBatchSamplerDataCollatorForSeq2Seq
|
14 |
+
from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
|
15 |
+
from axolotl.utils.trainer import process_pretraining_datasets_for_packing
|
16 |
+
|
17 |
+
LOG = logging.getLogger("axolotl")
|
18 |
+
|
19 |
+
|
20 |
+
def encode_pretraining(
|
21 |
+
tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: List[str]
|
22 |
+
) -> Dict[str, List]:
|
23 |
+
res = tokenizer(
|
24 |
+
examples,
|
25 |
+
truncation=True,
|
26 |
+
max_length=max_tokens - 2,
|
27 |
+
add_special_tokens=True,
|
28 |
+
)
|
29 |
+
# Convert to PyTorch tensors
|
30 |
+
input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
|
31 |
+
attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
|
32 |
+
new_input_ids = []
|
33 |
+
new_attention_mask = []
|
34 |
+
# Append EOS and PAD tokens to input_ids, and correct attention_mask
|
35 |
+
for i, _ in enumerate(input_ids):
|
36 |
+
input_ids[i] = torch.cat(
|
37 |
+
(
|
38 |
+
input_ids[i],
|
39 |
+
torch.tensor([tokenizer.eos_token_id, tokenizer.pad_token_id]),
|
40 |
+
),
|
41 |
+
dim=0,
|
42 |
+
)
|
43 |
+
attention_mask[i] = torch.cat((attention_mask[i], torch.tensor([1, 0])), dim=0)
|
44 |
+
|
45 |
+
# Concatenate tokens so that their lengths are less than max_tokens
|
46 |
+
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
47 |
+
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
48 |
+
|
49 |
+
for ids, mask in zip(input_ids, attention_mask):
|
50 |
+
if buffer_input_ids.numel() == max_tokens:
|
51 |
+
new_input_ids.append(buffer_input_ids)
|
52 |
+
new_attention_mask.append(buffer_attention_mask)
|
53 |
+
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
54 |
+
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
55 |
+
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
56 |
+
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
57 |
+
elif buffer_input_ids.numel() + ids.numel() <= max_tokens:
|
58 |
+
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
59 |
+
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
60 |
+
else:
|
61 |
+
buffer_input_ids = torch.cat(
|
62 |
+
(
|
63 |
+
buffer_input_ids,
|
64 |
+
torch.full(
|
65 |
+
(max_tokens - buffer_input_ids.numel(),),
|
66 |
+
tokenizer.pad_token_id,
|
67 |
+
dtype=torch.long,
|
68 |
+
),
|
69 |
+
),
|
70 |
+
dim=0,
|
71 |
+
)
|
72 |
+
buffer_attention_mask = torch.cat(
|
73 |
+
(
|
74 |
+
buffer_attention_mask,
|
75 |
+
torch.full(
|
76 |
+
(max_tokens - buffer_attention_mask.numel(),),
|
77 |
+
0,
|
78 |
+
dtype=torch.long,
|
79 |
+
),
|
80 |
+
),
|
81 |
+
dim=0,
|
82 |
+
)
|
83 |
+
new_input_ids.append(buffer_input_ids)
|
84 |
+
new_attention_mask.append(buffer_attention_mask)
|
85 |
+
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
86 |
+
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
87 |
+
|
88 |
+
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
89 |
+
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
90 |
+
|
91 |
+
if buffer_input_ids.numel() > 0: # for any leftover tokens
|
92 |
+
while buffer_input_ids.numel() < max_tokens: # make all sequences equal in size
|
93 |
+
buffer_input_ids = torch.cat(
|
94 |
+
(
|
95 |
+
buffer_input_ids,
|
96 |
+
torch.full(
|
97 |
+
(max_tokens - buffer_input_ids.numel(),),
|
98 |
+
tokenizer.pad_token_id,
|
99 |
+
dtype=torch.long,
|
100 |
+
),
|
101 |
+
),
|
102 |
+
dim=0,
|
103 |
+
)
|
104 |
+
buffer_attention_mask = torch.cat(
|
105 |
+
(
|
106 |
+
buffer_attention_mask,
|
107 |
+
torch.full(
|
108 |
+
(max_tokens - buffer_attention_mask.numel(),),
|
109 |
+
0,
|
110 |
+
dtype=torch.long,
|
111 |
+
),
|
112 |
+
),
|
113 |
+
dim=0,
|
114 |
+
)
|
115 |
+
new_input_ids.append(buffer_input_ids)
|
116 |
+
new_attention_mask.append(buffer_attention_mask)
|
117 |
+
|
118 |
+
ret = {
|
119 |
+
"input_ids": [seq.tolist() for seq in new_input_ids],
|
120 |
+
"labels": [seq.tolist() for seq in new_input_ids],
|
121 |
+
"attention_mask": [seq.tolist() for seq in new_attention_mask],
|
122 |
+
}
|
123 |
+
|
124 |
+
LOG.debug(len(ret["input_ids"]))
|
125 |
+
return ret
|
126 |
+
|
127 |
+
|
128 |
+
def wrap_pretraining_dataset(
|
129 |
+
dataset,
|
130 |
+
tokenizer,
|
131 |
+
cfg,
|
132 |
+
ds_wrapper_fn,
|
133 |
+
max_tokens=2048,
|
134 |
+
batch_size=1,
|
135 |
+
seed=42,
|
136 |
+
buffer_size=10_000,
|
137 |
+
):
|
138 |
+
if cfg.sample_packing:
|
139 |
+
collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
|
140 |
+
tokenizer,
|
141 |
+
return_tensors="pt",
|
142 |
+
padding=True,
|
143 |
+
pad_to_multiple_of=max_tokens * batch_size,
|
144 |
+
multipack_attn=cfg.pretrain_multipack_attn,
|
145 |
+
)
|
146 |
+
encode = functools.partial(
|
147 |
+
encode_packed_pretraining,
|
148 |
+
collate_fn,
|
149 |
+
ds_wrapper_fn,
|
150 |
+
max_seq_length=max_tokens,
|
151 |
+
batch_size=batch_size,
|
152 |
+
multipack_attn=cfg.pretrain_multipack_attn,
|
153 |
+
)
|
154 |
+
# set this to 1 so downstream data_loader doesn't try to increase the batch again
|
155 |
+
cfg.micro_batch_size = 1
|
156 |
+
else:
|
157 |
+
encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
|
158 |
+
|
159 |
+
if cfg.shuffle_merged_datasets:
|
160 |
+
dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
|
161 |
+
else:
|
162 |
+
LOG.debug("NOT shuffling merged pretraining datasets")
|
163 |
+
|
164 |
+
# remove all the existing columns after mapping since they end up having
|
165 |
+
# a different length than the encoded/tokenized column
|
166 |
+
# this is empty during streaming/pretraining
|
167 |
+
remove_columns = []
|
168 |
+
if dataset.features is None:
|
169 |
+
for first_row in dataset:
|
170 |
+
remove_columns = first_row.keys()
|
171 |
+
break
|
172 |
+
else:
|
173 |
+
remove_columns = dataset.features.keys()
|
174 |
+
|
175 |
+
dataset = dataset.map(
|
176 |
+
encode,
|
177 |
+
batched=True,
|
178 |
+
batch_size=buffer_size,
|
179 |
+
# input_columns="text",
|
180 |
+
remove_columns=remove_columns,
|
181 |
+
)
|
182 |
+
return dataset
|
183 |
+
|
184 |
+
|
185 |
+
def encode_packed_pretraining(
|
186 |
+
collate_fn,
|
187 |
+
ds_wrapper: Callable,
|
188 |
+
examples: Dict[str, List],
|
189 |
+
max_seq_length: int = 2048,
|
190 |
+
batch_size: int = 4,
|
191 |
+
multipack_attn: Optional[bool] = False,
|
192 |
+
) -> Dict[str, List]:
|
193 |
+
# pylint: disable=duplicate-code
|
194 |
+
# tokenize all the examples
|
195 |
+
# rows get split with stride (overlap)
|
196 |
+
train_dataset = ds_wrapper(Dataset.from_dict(examples))[0]
|
197 |
+
|
198 |
+
train_dataset = process_pretraining_datasets_for_packing(
|
199 |
+
train_dataset,
|
200 |
+
max_seq_length,
|
201 |
+
skip_position_ids=not multipack_attn,
|
202 |
+
)
|
203 |
+
|
204 |
+
sampler = MultipackBatchSampler(
|
205 |
+
RandomSampler(train_dataset),
|
206 |
+
batch_size=1,
|
207 |
+
drop_last=True,
|
208 |
+
batch_max_len=batch_size * max_seq_length,
|
209 |
+
lengths=get_dataset_lengths(train_dataset),
|
210 |
+
)
|
211 |
+
|
212 |
+
chunked_data = defaultdict(list)
|
213 |
+
|
214 |
+
for batch in sampler:
|
215 |
+
for data in batch:
|
216 |
+
features = train_dataset[data]
|
217 |
+
if "num_truncated_tokens" in features:
|
218 |
+
del features["num_truncated_tokens"]
|
219 |
+
if "num_truncated_tokens" in features:
|
220 |
+
del features["num_truncated_tokens"]
|
221 |
+
if "overflow_to_sample_mapping" in features:
|
222 |
+
del features["overflow_to_sample_mapping"]
|
223 |
+
if "labels" not in features:
|
224 |
+
features["labels"] = features["input_ids"].copy()
|
225 |
+
collated_features = collate_fn(features)
|
226 |
+
|
227 |
+
for feature in features.keys():
|
228 |
+
if feature == "length":
|
229 |
+
continue
|
230 |
+
chunked_data[feature].append(collated_features[feature].squeeze(0))
|
231 |
+
|
232 |
+
return chunked_data
|
src/axolotl/utils/{data.py → data/sft.py}
RENAMED
@@ -1,14 +1,10 @@
|
|
1 |
-
"""
|
2 |
|
3 |
import functools
|
4 |
-
import hashlib
|
5 |
import logging
|
6 |
-
from collections import defaultdict
|
7 |
from pathlib import Path
|
8 |
-
from typing import
|
9 |
|
10 |
-
import torch
|
11 |
-
import yaml
|
12 |
from datasets import (
|
13 |
Dataset,
|
14 |
DatasetDict,
|
@@ -18,13 +14,11 @@ from datasets import (
|
|
18 |
)
|
19 |
from huggingface_hub import hf_hub_download
|
20 |
from huggingface_hub.utils import HFValidationError
|
21 |
-
from torch.utils.data import RandomSampler
|
22 |
from transformers import PreTrainedTokenizerBase
|
23 |
|
24 |
from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
|
25 |
from axolotl.datasets import TokenizedPromptDataset
|
26 |
from axolotl.prompt_strategies import load
|
27 |
-
from axolotl.prompt_strategies.dpo import load as load_dpo
|
28 |
from axolotl.prompt_tokenizers import (
|
29 |
AlpacaMultipleChoicePromptTokenizingStrategy,
|
30 |
AlpacaPromptTokenizingStrategy,
|
@@ -45,26 +39,18 @@ from axolotl.prompters import (
|
|
45 |
SummarizeTLDRPrompter,
|
46 |
UnsupportedPrompter,
|
47 |
)
|
48 |
-
from axolotl.utils.
|
|
|
49 |
from axolotl.utils.dict import DictDefault
|
50 |
from axolotl.utils.distributed import is_main_process, zero_first
|
51 |
-
from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
|
52 |
from axolotl.utils.trainer import (
|
53 |
calculate_total_num_steps,
|
54 |
process_datasets_for_packing,
|
55 |
-
process_pretraining_datasets_for_packing,
|
56 |
)
|
57 |
|
58 |
LOG = logging.getLogger("axolotl")
|
59 |
|
60 |
|
61 |
-
def md5(to_hash: str, encoding: str = "utf-8") -> str:
|
62 |
-
try:
|
63 |
-
return hashlib.md5(to_hash.encode(encoding), usedforsecurity=False).hexdigest()
|
64 |
-
except TypeError:
|
65 |
-
return hashlib.md5(to_hash.encode(encoding)).hexdigest() # nosec
|
66 |
-
|
67 |
-
|
68 |
def prepare_dataset(cfg, tokenizer):
|
69 |
prompters = []
|
70 |
if not cfg.pretraining_dataset:
|
@@ -182,6 +168,7 @@ def load_tokenized_prepared_datasets(
|
|
182 |
except Exception: # pylint: disable=broad-except # nosec
|
183 |
pass
|
184 |
|
|
|
185 |
if dataset:
|
186 |
...
|
187 |
elif (
|
@@ -691,315 +678,3 @@ def get_dataset_wrapper(
|
|
691 |
)
|
692 |
|
693 |
return dataset_wrapper, dataset_prompter
|
694 |
-
|
695 |
-
|
696 |
-
def encode_pretraining(
|
697 |
-
tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: List[str]
|
698 |
-
) -> Dict[str, List]:
|
699 |
-
res = tokenizer(
|
700 |
-
examples,
|
701 |
-
truncation=True,
|
702 |
-
max_length=max_tokens - 2,
|
703 |
-
add_special_tokens=True,
|
704 |
-
)
|
705 |
-
# Convert to PyTorch tensors
|
706 |
-
input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
|
707 |
-
attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
|
708 |
-
new_input_ids = []
|
709 |
-
new_attention_mask = []
|
710 |
-
# Append EOS and PAD tokens to input_ids, and correct attention_mask
|
711 |
-
for i, _ in enumerate(input_ids):
|
712 |
-
input_ids[i] = torch.cat(
|
713 |
-
(
|
714 |
-
input_ids[i],
|
715 |
-
torch.tensor([tokenizer.eos_token_id, tokenizer.pad_token_id]),
|
716 |
-
),
|
717 |
-
dim=0,
|
718 |
-
)
|
719 |
-
attention_mask[i] = torch.cat((attention_mask[i], torch.tensor([1, 0])), dim=0)
|
720 |
-
|
721 |
-
# Concatenate tokens so that their lengths are less than max_tokens
|
722 |
-
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
723 |
-
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
724 |
-
|
725 |
-
for ids, mask in zip(input_ids, attention_mask):
|
726 |
-
if buffer_input_ids.numel() == max_tokens:
|
727 |
-
new_input_ids.append(buffer_input_ids)
|
728 |
-
new_attention_mask.append(buffer_attention_mask)
|
729 |
-
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
730 |
-
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
731 |
-
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
732 |
-
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
733 |
-
elif buffer_input_ids.numel() + ids.numel() <= max_tokens:
|
734 |
-
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
735 |
-
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
736 |
-
else:
|
737 |
-
buffer_input_ids = torch.cat(
|
738 |
-
(
|
739 |
-
buffer_input_ids,
|
740 |
-
torch.full(
|
741 |
-
(max_tokens - buffer_input_ids.numel(),),
|
742 |
-
tokenizer.pad_token_id,
|
743 |
-
dtype=torch.long,
|
744 |
-
),
|
745 |
-
),
|
746 |
-
dim=0,
|
747 |
-
)
|
748 |
-
buffer_attention_mask = torch.cat(
|
749 |
-
(
|
750 |
-
buffer_attention_mask,
|
751 |
-
torch.full(
|
752 |
-
(max_tokens - buffer_attention_mask.numel(),),
|
753 |
-
0,
|
754 |
-
dtype=torch.long,
|
755 |
-
),
|
756 |
-
),
|
757 |
-
dim=0,
|
758 |
-
)
|
759 |
-
new_input_ids.append(buffer_input_ids)
|
760 |
-
new_attention_mask.append(buffer_attention_mask)
|
761 |
-
buffer_input_ids = torch.tensor([], dtype=torch.long)
|
762 |
-
buffer_attention_mask = torch.tensor([], dtype=torch.long)
|
763 |
-
|
764 |
-
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
|
765 |
-
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
|
766 |
-
|
767 |
-
if buffer_input_ids.numel() > 0: # for any leftover tokens
|
768 |
-
while buffer_input_ids.numel() < max_tokens: # make all sequences equal in size
|
769 |
-
buffer_input_ids = torch.cat(
|
770 |
-
(
|
771 |
-
buffer_input_ids,
|
772 |
-
torch.full(
|
773 |
-
(max_tokens - buffer_input_ids.numel(),),
|
774 |
-
tokenizer.pad_token_id,
|
775 |
-
dtype=torch.long,
|
776 |
-
),
|
777 |
-
),
|
778 |
-
dim=0,
|
779 |
-
)
|
780 |
-
buffer_attention_mask = torch.cat(
|
781 |
-
(
|
782 |
-
buffer_attention_mask,
|
783 |
-
torch.full(
|
784 |
-
(max_tokens - buffer_attention_mask.numel(),),
|
785 |
-
0,
|
786 |
-
dtype=torch.long,
|
787 |
-
),
|
788 |
-
),
|
789 |
-
dim=0,
|
790 |
-
)
|
791 |
-
new_input_ids.append(buffer_input_ids)
|
792 |
-
new_attention_mask.append(buffer_attention_mask)
|
793 |
-
|
794 |
-
ret = {
|
795 |
-
"input_ids": [seq.tolist() for seq in new_input_ids],
|
796 |
-
"labels": [seq.tolist() for seq in new_input_ids],
|
797 |
-
"attention_mask": [seq.tolist() for seq in new_attention_mask],
|
798 |
-
}
|
799 |
-
|
800 |
-
LOG.debug(len(ret["input_ids"]))
|
801 |
-
return ret
|
802 |
-
|
803 |
-
|
804 |
-
def wrap_pretraining_dataset(
|
805 |
-
dataset,
|
806 |
-
tokenizer,
|
807 |
-
cfg,
|
808 |
-
ds_wrapper_fn,
|
809 |
-
max_tokens=2048,
|
810 |
-
batch_size=1,
|
811 |
-
seed=42,
|
812 |
-
buffer_size=10_000,
|
813 |
-
):
|
814 |
-
if cfg.sample_packing:
|
815 |
-
collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
|
816 |
-
tokenizer,
|
817 |
-
return_tensors="pt",
|
818 |
-
padding=True,
|
819 |
-
pad_to_multiple_of=max_tokens * batch_size,
|
820 |
-
multipack_attn=cfg.pretrain_multipack_attn,
|
821 |
-
)
|
822 |
-
encode = functools.partial(
|
823 |
-
encode_packed_pretraining,
|
824 |
-
collate_fn,
|
825 |
-
ds_wrapper_fn,
|
826 |
-
max_seq_length=max_tokens,
|
827 |
-
batch_size=batch_size,
|
828 |
-
multipack_attn=cfg.pretrain_multipack_attn,
|
829 |
-
)
|
830 |
-
# set this to 1 so downstream data_loader doesn't try to increase the batch again
|
831 |
-
cfg.micro_batch_size = 1
|
832 |
-
else:
|
833 |
-
encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
|
834 |
-
|
835 |
-
if cfg.shuffle_merged_datasets:
|
836 |
-
dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
|
837 |
-
else:
|
838 |
-
LOG.debug("NOT shuffling merged pretraining datasets")
|
839 |
-
|
840 |
-
# remove all the existing columns after mapping since they end up having
|
841 |
-
# a different length than the encoded/tokenized column
|
842 |
-
# this is empty during streaming/pretraining
|
843 |
-
remove_columns = []
|
844 |
-
if dataset.features is None:
|
845 |
-
for first_row in dataset:
|
846 |
-
remove_columns = first_row.keys()
|
847 |
-
break
|
848 |
-
else:
|
849 |
-
remove_columns = dataset.features.keys()
|
850 |
-
|
851 |
-
dataset = dataset.map(
|
852 |
-
encode,
|
853 |
-
batched=True,
|
854 |
-
batch_size=buffer_size,
|
855 |
-
# input_columns="text",
|
856 |
-
remove_columns=remove_columns,
|
857 |
-
)
|
858 |
-
return dataset
|
859 |
-
|
860 |
-
|
861 |
-
def encode_packed_pretraining(
|
862 |
-
collate_fn,
|
863 |
-
ds_wrapper: Callable,
|
864 |
-
examples: Dict[str, List],
|
865 |
-
max_seq_length: int = 2048,
|
866 |
-
batch_size: int = 4,
|
867 |
-
multipack_attn: Optional[bool] = False,
|
868 |
-
) -> Dict[str, List]:
|
869 |
-
# pylint: disable=duplicate-code
|
870 |
-
# tokenize all the examples
|
871 |
-
# rows get split with stride (overlap)
|
872 |
-
train_dataset = ds_wrapper(Dataset.from_dict(examples))[0]
|
873 |
-
|
874 |
-
train_dataset = process_pretraining_datasets_for_packing(
|
875 |
-
train_dataset,
|
876 |
-
max_seq_length,
|
877 |
-
skip_position_ids=not multipack_attn,
|
878 |
-
)
|
879 |
-
|
880 |
-
sampler = MultipackBatchSampler(
|
881 |
-
RandomSampler(train_dataset),
|
882 |
-
batch_size=1,
|
883 |
-
drop_last=True,
|
884 |
-
batch_max_len=batch_size * max_seq_length,
|
885 |
-
lengths=get_dataset_lengths(train_dataset),
|
886 |
-
)
|
887 |
-
|
888 |
-
chunked_data = defaultdict(list)
|
889 |
-
|
890 |
-
for batch in sampler:
|
891 |
-
for data in batch:
|
892 |
-
features = train_dataset[data]
|
893 |
-
if "num_truncated_tokens" in features:
|
894 |
-
del features["num_truncated_tokens"]
|
895 |
-
if "num_truncated_tokens" in features:
|
896 |
-
del features["num_truncated_tokens"]
|
897 |
-
if "overflow_to_sample_mapping" in features:
|
898 |
-
del features["overflow_to_sample_mapping"]
|
899 |
-
if "labels" not in features:
|
900 |
-
features["labels"] = features["input_ids"].copy()
|
901 |
-
collated_features = collate_fn(features)
|
902 |
-
|
903 |
-
for feature in features.keys():
|
904 |
-
if feature == "length":
|
905 |
-
continue
|
906 |
-
chunked_data[feature].append(collated_features[feature].squeeze(0))
|
907 |
-
|
908 |
-
return chunked_data
|
909 |
-
|
910 |
-
|
911 |
-
def _get_path(ds_hash, cfg):
|
912 |
-
prepared_ds_path = (
|
913 |
-
Path(cfg.dataset_prepared_path) / ds_hash
|
914 |
-
if cfg.dataset_prepared_path
|
915 |
-
else Path(DEFAULT_DATASET_PREPARED_PATH) / ds_hash
|
916 |
-
)
|
917 |
-
|
918 |
-
return prepared_ds_path
|
919 |
-
|
920 |
-
|
921 |
-
def _load_preprocessed_ds(cfg, sub_cfg):
|
922 |
-
ds_hash = md5(yaml.dump(sub_cfg, Dumper=yaml.Dumper))
|
923 |
-
prepared_ds_path = _get_path(ds_hash, cfg)
|
924 |
-
dataset = None
|
925 |
-
|
926 |
-
if (
|
927 |
-
cfg.dataset_prepared_path
|
928 |
-
and any(prepared_ds_path.glob("*"))
|
929 |
-
and not cfg.is_preprocess
|
930 |
-
):
|
931 |
-
LOG.info(f"Loading prepared dataset from disk at {prepared_ds_path}...")
|
932 |
-
dataset = load_from_disk(str(prepared_ds_path))
|
933 |
-
|
934 |
-
return dataset
|
935 |
-
|
936 |
-
|
937 |
-
def _save_preprocessed_ds(cfg, sub_cfg, dataset):
|
938 |
-
ds_hash = md5(yaml.dump(sub_cfg, Dumper=yaml.Dumper))
|
939 |
-
prepared_ds_path = _get_path(ds_hash, cfg)
|
940 |
-
|
941 |
-
if cfg.is_preprocess and is_main_process():
|
942 |
-
LOG.info(f"Loading prepared dataset from disk at {prepared_ds_path}...")
|
943 |
-
dataset.save_to_disk(str(prepared_ds_path))
|
944 |
-
|
945 |
-
|
946 |
-
def load_prepare_dpo_datasets(cfg):
|
947 |
-
def load_split(dataset_cfgs, _cfg):
|
948 |
-
split_datasets: List[Any] = []
|
949 |
-
for i, ds_cfg in enumerate(dataset_cfgs):
|
950 |
-
if ds_cfg["ds_type"] == "json":
|
951 |
-
for data_file in ds_cfg["data_files"]:
|
952 |
-
data_files = {ds_cfg["split"]: data_file}
|
953 |
-
ds = load_dataset( # pylint: disable=invalid-name
|
954 |
-
"json",
|
955 |
-
data_files=data_files,
|
956 |
-
split=ds_cfg["split"],
|
957 |
-
)
|
958 |
-
split_datasets.insert(i, ds)
|
959 |
-
else:
|
960 |
-
ds = load_dataset( # pylint: disable=invalid-name
|
961 |
-
ds_cfg["path"],
|
962 |
-
split=ds_cfg["split"],
|
963 |
-
)
|
964 |
-
split_datasets.insert(i, ds)
|
965 |
-
|
966 |
-
for i, data_set in enumerate(split_datasets):
|
967 |
-
_type = dataset_cfgs[i]["type"]
|
968 |
-
if _type:
|
969 |
-
if isinstance(_type, DictDefault):
|
970 |
-
_type = "user_defined.default"
|
971 |
-
ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
|
972 |
-
split_datasets[i] = data_set.map(
|
973 |
-
ds_transform_fn,
|
974 |
-
desc="Mapping RL Dataset",
|
975 |
-
)
|
976 |
-
else:
|
977 |
-
# If no `type` is provided, assume the dataset is already in the expected format with
|
978 |
-
# "prompt", "chosen" and "rejected" already preprocessed
|
979 |
-
split_datasets[i] = data_set
|
980 |
-
|
981 |
-
return concatenate_datasets(split_datasets)
|
982 |
-
|
983 |
-
with zero_first(is_main_process()):
|
984 |
-
train_is_preprocessed = False
|
985 |
-
eval_is_preprocessed = False
|
986 |
-
if train_dataset := _load_preprocessed_ds(cfg, cfg.datasets):
|
987 |
-
train_is_preprocessed = True
|
988 |
-
else:
|
989 |
-
train_dataset = load_split(cfg.datasets, cfg)
|
990 |
-
|
991 |
-
eval_dataset = None
|
992 |
-
if cfg.test_datasets:
|
993 |
-
if eval_dataset := _load_preprocessed_ds(cfg, cfg.test_datasets):
|
994 |
-
eval_is_preprocessed = True
|
995 |
-
else:
|
996 |
-
eval_dataset = load_split(cfg.test_datasets, cfg)
|
997 |
-
if not eval_dataset:
|
998 |
-
eval_dataset = None
|
999 |
-
|
1000 |
-
if not train_is_preprocessed:
|
1001 |
-
_save_preprocessed_ds(cfg, cfg.datasets, train_dataset)
|
1002 |
-
if eval_dataset and not eval_is_preprocessed:
|
1003 |
-
_save_preprocessed_ds(cfg, cfg.test_datasets, eval_dataset)
|
1004 |
-
|
1005 |
-
return train_dataset, eval_dataset
|
|
|
1 |
+
"""data handling specific to SFT"""
|
2 |
|
3 |
import functools
|
|
|
4 |
import logging
|
|
|
5 |
from pathlib import Path
|
6 |
+
from typing import List, Optional, Tuple, Union
|
7 |
|
|
|
|
|
8 |
from datasets import (
|
9 |
Dataset,
|
10 |
DatasetDict,
|
|
|
14 |
)
|
15 |
from huggingface_hub import hf_hub_download
|
16 |
from huggingface_hub.utils import HFValidationError
|
|
|
17 |
from transformers import PreTrainedTokenizerBase
|
18 |
|
19 |
from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
|
20 |
from axolotl.datasets import TokenizedPromptDataset
|
21 |
from axolotl.prompt_strategies import load
|
|
|
22 |
from axolotl.prompt_tokenizers import (
|
23 |
AlpacaMultipleChoicePromptTokenizingStrategy,
|
24 |
AlpacaPromptTokenizingStrategy,
|
|
|
39 |
SummarizeTLDRPrompter,
|
40 |
UnsupportedPrompter,
|
41 |
)
|
42 |
+
from axolotl.utils.data.pretraining import wrap_pretraining_dataset
|
43 |
+
from axolotl.utils.data.utils import md5
|
44 |
from axolotl.utils.dict import DictDefault
|
45 |
from axolotl.utils.distributed import is_main_process, zero_first
|
|
|
46 |
from axolotl.utils.trainer import (
|
47 |
calculate_total_num_steps,
|
48 |
process_datasets_for_packing,
|
|
|
49 |
)
|
50 |
|
51 |
LOG = logging.getLogger("axolotl")
|
52 |
|
53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
def prepare_dataset(cfg, tokenizer):
|
55 |
prompters = []
|
56 |
if not cfg.pretraining_dataset:
|
|
|
168 |
except Exception: # pylint: disable=broad-except # nosec
|
169 |
pass
|
170 |
|
171 |
+
# pylint: disable=duplicate-code
|
172 |
if dataset:
|
173 |
...
|
174 |
elif (
|
|
|
678 |
)
|
679 |
|
680 |
return dataset_wrapper, dataset_prompter
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/axolotl/utils/data/utils.py
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""data handling helpers"""
|
2 |
+
|
3 |
+
import hashlib
|
4 |
+
|
5 |
+
|
6 |
+
def md5(to_hash: str, encoding: str = "utf-8") -> str:
|
7 |
+
try:
|
8 |
+
return hashlib.md5(to_hash.encode(encoding), usedforsecurity=False).hexdigest()
|
9 |
+
except TypeError:
|
10 |
+
return hashlib.md5(to_hash.encode(encoding)).hexdigest() # nosec
|