danbraunai-apollo
commited on
Commit
•
6be2a8b
1
Parent(s):
a6177e5
Add upload script
Browse files- upload_script.py +385 -0
upload_script.py
ADDED
@@ -0,0 +1,385 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Taken and adapated from Alan Cooney's
|
3 |
+
https://github.com/ai-safety-foundation/sparse_autoencoder/tree/main/sparse_autoencoder.
|
4 |
+
"""
|
5 |
+
|
6 |
+
import subprocess
|
7 |
+
from collections.abc import Mapping, Sequence
|
8 |
+
from dataclasses import dataclass
|
9 |
+
from typing import TypedDict
|
10 |
+
|
11 |
+
from datasets import (
|
12 |
+
Dataset,
|
13 |
+
DatasetDict,
|
14 |
+
VerificationMode,
|
15 |
+
load_dataset,
|
16 |
+
)
|
17 |
+
from huggingface_hub import HfApi
|
18 |
+
from jaxtyping import Int
|
19 |
+
from pydantic import PositiveInt, validate_call
|
20 |
+
from torch import Tensor
|
21 |
+
from transformers import AutoTokenizer, PreTrainedTokenizerBase
|
22 |
+
|
23 |
+
|
24 |
+
class GenericTextDataBatch(TypedDict):
|
25 |
+
"""Generic Text Dataset Batch.
|
26 |
+
|
27 |
+
Assumes the dataset provides a 'text' field with a list of strings.
|
28 |
+
"""
|
29 |
+
|
30 |
+
text: list[str]
|
31 |
+
meta: list[dict[str, dict[str, str]]] # Optional, depending on the dataset structure.
|
32 |
+
|
33 |
+
|
34 |
+
TokenizedPrompt = list[int]
|
35 |
+
"""A tokenized prompt."""
|
36 |
+
|
37 |
+
|
38 |
+
class TokenizedPrompts(TypedDict):
|
39 |
+
"""Tokenized prompts."""
|
40 |
+
|
41 |
+
input_ids: list[TokenizedPrompt]
|
42 |
+
|
43 |
+
|
44 |
+
class TorchTokenizedPrompts(TypedDict):
|
45 |
+
"""Tokenized prompts prepared for PyTorch."""
|
46 |
+
|
47 |
+
input_ids: Int[Tensor, "batch pos vocab"]
|
48 |
+
|
49 |
+
|
50 |
+
class TextDataset:
|
51 |
+
"""Generic Text Dataset for any text-based dataset from Hugging Face."""
|
52 |
+
|
53 |
+
tokenizer: PreTrainedTokenizerBase
|
54 |
+
|
55 |
+
def preprocess(
|
56 |
+
self,
|
57 |
+
source_batch: GenericTextDataBatch,
|
58 |
+
*,
|
59 |
+
context_size: int,
|
60 |
+
) -> TokenizedPrompts:
|
61 |
+
"""Preprocess a batch of prompts.
|
62 |
+
|
63 |
+
Tokenizes a batch of text data and packs into context_size samples. An eos token is added
|
64 |
+
to the end of each document after tokenization.
|
65 |
+
|
66 |
+
Args:
|
67 |
+
source_batch: A batch of source data, including 'text' with a list of strings.
|
68 |
+
context_size: Context size for tokenized prompts.
|
69 |
+
|
70 |
+
Returns:
|
71 |
+
Tokenized prompts.
|
72 |
+
"""
|
73 |
+
prompts: list[str] = source_batch["text"]
|
74 |
+
|
75 |
+
tokenized_prompts = self.tokenizer(prompts, truncation=False, padding=False)
|
76 |
+
|
77 |
+
all_tokens = []
|
78 |
+
for document_tokens in tokenized_prompts[self._dataset_column_name]: # type: ignore
|
79 |
+
all_tokens.extend(document_tokens + [self.tokenizer.eos_token_id])
|
80 |
+
# Ignore incomplete chunks
|
81 |
+
chunks = [
|
82 |
+
all_tokens[i : i + context_size]
|
83 |
+
for i in range(0, len(all_tokens), context_size)
|
84 |
+
if len(all_tokens[i : i + context_size]) == context_size
|
85 |
+
]
|
86 |
+
|
87 |
+
return {"input_ids": chunks}
|
88 |
+
|
89 |
+
@validate_call(config={"arbitrary_types_allowed": True})
|
90 |
+
def __init__(
|
91 |
+
self,
|
92 |
+
dataset_path: str,
|
93 |
+
tokenizer: PreTrainedTokenizerBase,
|
94 |
+
context_size: PositiveInt = 256,
|
95 |
+
load_revision: str = "main",
|
96 |
+
dataset_dir: str | None = None,
|
97 |
+
dataset_files: str | Sequence[str] | Mapping[str, str | Sequence[str]] | None = None,
|
98 |
+
dataset_split: str | None = None,
|
99 |
+
dataset_column_name: str = "input_ids",
|
100 |
+
n_processes_preprocessing: PositiveInt | None = None,
|
101 |
+
preprocess_batch_size: PositiveInt = 1000,
|
102 |
+
):
|
103 |
+
"""Initialize a generic text dataset from Hugging Face.
|
104 |
+
|
105 |
+
Args:
|
106 |
+
dataset_path: Path to the dataset on Hugging Face (e.g. `'monology/pile-uncopyright'`).
|
107 |
+
tokenizer: Tokenizer to process text data.
|
108 |
+
context_size: The context size to use when returning a list of tokenized prompts.
|
109 |
+
*Towards Monosemanticity: Decomposing Language Models With Dictionary Learning* used
|
110 |
+
a context size of 250.
|
111 |
+
load_revision: The commit hash or branch name to download from the source dataset.
|
112 |
+
dataset_dir: Defining the `data_dir` of the dataset configuration.
|
113 |
+
dataset_files: Path(s) to source data file(s).
|
114 |
+
dataset_split: Dataset split (e.g., 'train'). If None, process all splits.
|
115 |
+
dataset_column_name: The column name for the prompts.
|
116 |
+
n_processes_preprocessing: Number of processes to use for preprocessing.
|
117 |
+
preprocess_batch_size: Batch size for preprocessing (tokenizing prompts).
|
118 |
+
"""
|
119 |
+
self.tokenizer = tokenizer
|
120 |
+
|
121 |
+
self.context_size = context_size
|
122 |
+
self._dataset_column_name = dataset_column_name
|
123 |
+
|
124 |
+
# Load the dataset
|
125 |
+
dataset = load_dataset(
|
126 |
+
dataset_path,
|
127 |
+
revision=load_revision,
|
128 |
+
streaming=False, # We need to pre-download the dataset to upload it to the hub.
|
129 |
+
split=dataset_split,
|
130 |
+
data_dir=dataset_dir,
|
131 |
+
data_files=dataset_files,
|
132 |
+
verification_mode=VerificationMode.NO_CHECKS, # As it fails when data_files is set
|
133 |
+
)
|
134 |
+
# If split is not None, will return a Dataset instance. Convert to DatasetDict.
|
135 |
+
if isinstance(dataset, Dataset):
|
136 |
+
assert dataset_split is not None
|
137 |
+
dataset = DatasetDict({dataset_split: dataset})
|
138 |
+
assert isinstance(dataset, DatasetDict)
|
139 |
+
|
140 |
+
for split in dataset:
|
141 |
+
print(f"Processing split: {split}")
|
142 |
+
# Setup preprocessing (we remove all columns except for input ids)
|
143 |
+
remove_columns: list[str] = list(next(iter(dataset[split])).keys()) # type: ignore
|
144 |
+
if "input_ids" in remove_columns:
|
145 |
+
remove_columns.remove("input_ids")
|
146 |
+
|
147 |
+
# Tokenize and chunk the prompts
|
148 |
+
mapped_dataset = dataset[split].map(
|
149 |
+
self.preprocess,
|
150 |
+
batched=True,
|
151 |
+
batch_size=preprocess_batch_size,
|
152 |
+
fn_kwargs={"context_size": context_size},
|
153 |
+
remove_columns=remove_columns,
|
154 |
+
num_proc=n_processes_preprocessing,
|
155 |
+
)
|
156 |
+
dataset[split] = mapped_dataset.shuffle()
|
157 |
+
|
158 |
+
self.dataset = dataset
|
159 |
+
|
160 |
+
@validate_call
|
161 |
+
def push_to_hugging_face_hub(
|
162 |
+
self,
|
163 |
+
repo_id: str,
|
164 |
+
commit_message: str = "Upload preprocessed dataset using sparse_autoencoder.",
|
165 |
+
max_shard_size: str = "500MB",
|
166 |
+
revision: str = "main",
|
167 |
+
*,
|
168 |
+
private: bool = False,
|
169 |
+
) -> None:
|
170 |
+
"""Share preprocessed dataset to Hugging Face hub.
|
171 |
+
|
172 |
+
Motivation:
|
173 |
+
Pre-processing a dataset can be time-consuming, so it is useful to be able to share the
|
174 |
+
pre-processed dataset with others. This function allows you to do that by pushing the
|
175 |
+
pre-processed dataset to the Hugging Face hub.
|
176 |
+
|
177 |
+
Warning:
|
178 |
+
You must be logged into HuggingFace (e.g with `huggingface-cli login` from the terminal)
|
179 |
+
to use this.
|
180 |
+
|
181 |
+
Warning:
|
182 |
+
This will only work if the dataset is not streamed (i.e. if `pre_download=True` when
|
183 |
+
initializing the dataset).
|
184 |
+
|
185 |
+
Args:
|
186 |
+
repo_id: Hugging Face repo ID to save the dataset to (e.g. `username/dataset_name`).
|
187 |
+
commit_message: Commit message.
|
188 |
+
max_shard_size: Maximum shard size (e.g. `'500MB'`).
|
189 |
+
revision: Branch to push to.
|
190 |
+
private: Whether to save the dataset privately.
|
191 |
+
"""
|
192 |
+
self.dataset.push_to_hub(
|
193 |
+
repo_id=repo_id,
|
194 |
+
commit_message=commit_message,
|
195 |
+
max_shard_size=max_shard_size,
|
196 |
+
private=private,
|
197 |
+
revision=revision,
|
198 |
+
)
|
199 |
+
|
200 |
+
|
201 |
+
@dataclass
|
202 |
+
class DatasetToPreprocess:
|
203 |
+
"""Dataset to preprocess info."""
|
204 |
+
|
205 |
+
source_path: str
|
206 |
+
"""Source path from HF (e.g. `roneneldan/TinyStories`)."""
|
207 |
+
|
208 |
+
tokenizer_name: str
|
209 |
+
"""HF tokenizer name (e.g. `gpt2`)."""
|
210 |
+
|
211 |
+
load_revision: str = "main"
|
212 |
+
"""Commit hash or branch name to download from the source dataset."""
|
213 |
+
|
214 |
+
data_dir: str | None = None
|
215 |
+
"""Data directory to download from the source dataset."""
|
216 |
+
|
217 |
+
data_files: list[str] | None = None
|
218 |
+
"""Data files to download from the source dataset."""
|
219 |
+
|
220 |
+
hugging_face_username: str = "apollo-research"
|
221 |
+
"""HF username for the upload."""
|
222 |
+
|
223 |
+
private: bool = False
|
224 |
+
"""Whether the HF dataset should be private or public."""
|
225 |
+
|
226 |
+
context_size: int = 2048
|
227 |
+
"""Number of tokens in a single sample. gpt2 uses 1024, pythia uses 2048."""
|
228 |
+
|
229 |
+
split: str | None = None
|
230 |
+
"""Dataset split to download from the source dataset. If None, process all splits."""
|
231 |
+
|
232 |
+
@property
|
233 |
+
def source_alias(self) -> str:
|
234 |
+
"""Create a source alias for the destination dataset name.
|
235 |
+
|
236 |
+
Returns:
|
237 |
+
The modified source path as source alias.
|
238 |
+
"""
|
239 |
+
return self.source_path.replace("/", "-")
|
240 |
+
|
241 |
+
@property
|
242 |
+
def tokenizer_alias(self) -> str:
|
243 |
+
"""Create a tokenizer alias for the destination dataset name.
|
244 |
+
|
245 |
+
Returns:
|
246 |
+
The modified tokenizer name as tokenizer alias.
|
247 |
+
"""
|
248 |
+
return self.tokenizer_name.replace("/", "-")
|
249 |
+
|
250 |
+
@property
|
251 |
+
def destination_repo_name(self) -> str:
|
252 |
+
"""Destination repo name.
|
253 |
+
|
254 |
+
Returns:
|
255 |
+
The destination repo name.
|
256 |
+
"""
|
257 |
+
split_str = f"{self.split}-" if self.split else ""
|
258 |
+
return f"{self.source_alias}-{split_str}tokenizer-{self.tokenizer_alias}"
|
259 |
+
|
260 |
+
@property
|
261 |
+
def destination_repo_id(self) -> str:
|
262 |
+
"""Destination repo ID.
|
263 |
+
|
264 |
+
Returns:
|
265 |
+
The destination repo ID.
|
266 |
+
"""
|
267 |
+
return f"{self.hugging_face_username}/{self.destination_repo_name}"
|
268 |
+
|
269 |
+
|
270 |
+
def upload_datasets(datasets_to_preprocess: list[DatasetToPreprocess]) -> None:
|
271 |
+
"""Upload datasets to HF.
|
272 |
+
|
273 |
+
Warning:
|
274 |
+
Assumes you have already created the corresponding repos on HF.
|
275 |
+
|
276 |
+
Args:
|
277 |
+
datasets_to_preprocess: List of datasets to preprocess.
|
278 |
+
|
279 |
+
Raises:
|
280 |
+
ValueError: If the repo doesn't exist.
|
281 |
+
"""
|
282 |
+
repositories_updating = [dataset.destination_repo_id for dataset in datasets_to_preprocess]
|
283 |
+
print("Updating repositories:\n" "\n".join(repositories_updating))
|
284 |
+
|
285 |
+
for dataset in datasets_to_preprocess:
|
286 |
+
print("Processing dataset: ", dataset.source_path)
|
287 |
+
|
288 |
+
# Preprocess
|
289 |
+
tokenizer = AutoTokenizer.from_pretrained(dataset.tokenizer_name)
|
290 |
+
text_dataset = TextDataset(
|
291 |
+
dataset_path=dataset.source_path,
|
292 |
+
tokenizer=tokenizer,
|
293 |
+
dataset_files=dataset.data_files,
|
294 |
+
dataset_dir=dataset.data_dir,
|
295 |
+
dataset_split=dataset.split,
|
296 |
+
context_size=dataset.context_size,
|
297 |
+
load_revision=dataset.load_revision,
|
298 |
+
)
|
299 |
+
# size_in_bytes and info gives info about the whole dataset regardless of the split index,
|
300 |
+
# so we just get the first split.
|
301 |
+
split = next(iter(text_dataset.dataset))
|
302 |
+
print("Dataset info:")
|
303 |
+
print(f"Size: {text_dataset.dataset[split].size_in_bytes / 1e9:.2f} GB") # type: ignore
|
304 |
+
print("Info: ", text_dataset.dataset[split].info)
|
305 |
+
|
306 |
+
# Upload
|
307 |
+
text_dataset.push_to_hugging_face_hub(
|
308 |
+
repo_id=dataset.destination_repo_id, private=dataset.private
|
309 |
+
)
|
310 |
+
# Also upload the current file to the repo for reproducibility and transparency
|
311 |
+
api = HfApi()
|
312 |
+
api.upload_file(
|
313 |
+
path_or_fileobj=__file__,
|
314 |
+
path_in_repo="upload_script.py",
|
315 |
+
repo_id=dataset.destination_repo_id,
|
316 |
+
repo_type="dataset",
|
317 |
+
commit_message="Add upload script",
|
318 |
+
)
|
319 |
+
|
320 |
+
|
321 |
+
if __name__ == "__main__":
|
322 |
+
# Check that the user is signed in to huggingface-cli
|
323 |
+
try:
|
324 |
+
result = subprocess.run(
|
325 |
+
["huggingface-cli", "whoami"], check=True, capture_output=True, text=True
|
326 |
+
)
|
327 |
+
if "Not logged in" in result.stdout:
|
328 |
+
print("Please sign in to huggingface-cli using `huggingface-cli login`.")
|
329 |
+
raise Exception("You are not logged in to huggingface-cli.")
|
330 |
+
except subprocess.CalledProcessError:
|
331 |
+
print("An error occurred while checking the login status.")
|
332 |
+
raise
|
333 |
+
|
334 |
+
datasets: list[DatasetToPreprocess] = [
|
335 |
+
DatasetToPreprocess(
|
336 |
+
source_path="roneneldan/TinyStories",
|
337 |
+
# Paper says gpt-neo tokenizer, and e.g. EleutherAI/gpt-neo-125M uses the same tokenizer
|
338 |
+
# as gpt2. They also suggest using gpt2 in (https://github.com/EleutherAI/gpt-neo).
|
339 |
+
tokenizer_name="gpt2",
|
340 |
+
hugging_face_username="apollo-research",
|
341 |
+
context_size=512,
|
342 |
+
),
|
343 |
+
DatasetToPreprocess(
|
344 |
+
source_path="Skylion007/openwebtext",
|
345 |
+
tokenizer_name="gpt2",
|
346 |
+
hugging_face_username="apollo-research",
|
347 |
+
context_size=1024,
|
348 |
+
),
|
349 |
+
DatasetToPreprocess(
|
350 |
+
source_path="Skylion007/openwebtext",
|
351 |
+
tokenizer_name="EleutherAI/gpt-neox-20b",
|
352 |
+
hugging_face_username="apollo-research",
|
353 |
+
context_size=2048,
|
354 |
+
),
|
355 |
+
DatasetToPreprocess(
|
356 |
+
source_path="monology/pile-uncopyrighted",
|
357 |
+
tokenizer_name="gpt2",
|
358 |
+
hugging_face_username="apollo-research",
|
359 |
+
context_size=1024,
|
360 |
+
# Get just the first few (each file is 11GB so this should be enough for a large dataset)
|
361 |
+
data_files=[
|
362 |
+
"train/00.jsonl.zst",
|
363 |
+
"train/01.jsonl.zst",
|
364 |
+
"train/02.jsonl.zst",
|
365 |
+
"train/03.jsonl.zst",
|
366 |
+
"train/04.jsonl.zst",
|
367 |
+
],
|
368 |
+
),
|
369 |
+
DatasetToPreprocess(
|
370 |
+
source_path="monology/pile-uncopyrighted",
|
371 |
+
tokenizer_name="EleutherAI/gpt-neox-20b",
|
372 |
+
hugging_face_username="apollo-research",
|
373 |
+
private=False,
|
374 |
+
context_size=2048,
|
375 |
+
data_files=[
|
376 |
+
"train/00.jsonl.zst",
|
377 |
+
"train/01.jsonl.zst",
|
378 |
+
"train/02.jsonl.zst",
|
379 |
+
"train/03.jsonl.zst",
|
380 |
+
"train/04.jsonl.zst",
|
381 |
+
],
|
382 |
+
),
|
383 |
+
]
|
384 |
+
|
385 |
+
upload_datasets(datasets)
|