donfu
commited on
Commit
•
007df59
1
Parent(s):
aae6ba6
Add hacky support for huge files
Browse files- process.py +154 -63
process.py
CHANGED
@@ -10,70 +10,80 @@ import sys
|
|
10 |
import re
|
11 |
from html2text import html2text
|
12 |
from datasets import load_dataset
|
|
|
|
|
|
|
|
|
13 |
|
14 |
-
|
|
|
15 |
SOURCE = "stackexchange-{0}"
|
16 |
MAX_ANSWERS = 10
|
17 |
QUESTION_SCORE_TRESHOLD = 0
|
18 |
ANSWER_SCORE_TRESHOLD = 0
|
19 |
HF_DATASET = "donfu/oa-stackexchange"
|
20 |
-
|
21 |
-
xml_format_map = {
|
22 |
-
"Id": int,
|
23 |
-
"PostTypeId": int,
|
24 |
-
"CreationDate": str,
|
25 |
-
"Score": int,
|
26 |
-
"ViewCount": int,
|
27 |
-
"Body": str,
|
28 |
-
"AnswerCount": int,
|
29 |
-
"CommentCount": int,
|
30 |
-
"ContentLicense": str,
|
31 |
-
"AcceptedAnswerId": int,
|
32 |
-
"ParentId": int,
|
33 |
-
}
|
34 |
|
35 |
|
36 |
def main():
|
37 |
datasets = sys.argv[1:] if len(sys.argv) > 1 else list_cached_datasets()
|
|
|
|
|
38 |
for dataset in datasets:
|
39 |
process_dataset(dataset)
|
40 |
|
41 |
|
42 |
def list_cached_datasets():
|
43 |
-
xml_files = glob.glob(f"{
|
44 |
datasets = [os.path.splitext(os.path.basename(file))[0] for file in xml_files]
|
45 |
datasets.sort()
|
46 |
return datasets
|
47 |
|
48 |
|
49 |
def process_dataset(dataset):
|
50 |
-
xml_file = f"{
|
|
|
51 |
source = SOURCE.format(dataset)
|
52 |
-
if os.path.exists(xml_file):
|
53 |
-
df =
|
54 |
-
oa = convert_to_oa(df)
|
55 |
save_parquet(oa, dataset)
|
56 |
# upload_hf(dataset)
|
57 |
else:
|
58 |
print(f"XML file {xml_file} not found, please download first. Skipping...")
|
59 |
|
60 |
|
61 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
"""
|
63 |
Convert dataframe to Open Assistant format with INSTRUCTION, RESPONSE, SOURCE, METADATA columns
|
64 |
|
65 |
Only include questions with an AcceptedAnswerId
|
66 |
"""
|
67 |
-
|
68 |
-
"
|
69 |
-
.replace("-", " ")
|
70 |
.replace("><", ", ")
|
71 |
.replace("<", "")
|
72 |
.replace(">", "")
|
73 |
-
|
74 |
-
|
75 |
-
"
|
76 |
-
"
|
|
|
77 |
}
|
78 |
questions = all[all["AcceptedAnswerId"] != 0]
|
79 |
merged = pd.merge(
|
@@ -84,11 +94,13 @@ def convert_to_oa(all):
|
|
84 |
right_on="Id",
|
85 |
suffixes=("_q", "_a"),
|
86 |
)
|
|
|
|
|
87 |
merged["INSTRUCTION"] = (
|
88 |
merged["Title_q"] + "\n" + merged["Body_q"].apply(to_markdown)
|
89 |
)
|
90 |
merged["RESPONSE"] = merged["Body_a"].apply(to_markdown)
|
91 |
-
merged["SOURCE"] =
|
92 |
merged["METADATA"] = merged.apply(create_metadata, axis=1)
|
93 |
|
94 |
return merged[["INSTRUCTION", "RESPONSE", "SOURCE", "METADATA"]]
|
@@ -99,63 +111,142 @@ def save_parquet(df, dataset):
|
|
99 |
Save Dataframe to Parquet. See here for specs:
|
100 |
https://projects.laion.ai/Open-Assistant/docs/data/datasets#creating-a-dataset-on-hugging-face
|
101 |
"""
|
102 |
-
parquet_file =
|
103 |
df.to_parquet(parquet_file, row_group_size=100, engine="pyarrow", index=False)
|
104 |
-
print("Converted
|
105 |
|
106 |
|
107 |
def upload_hf(dataset):
|
108 |
"""
|
109 |
Upload to Hugging Face
|
110 |
"""
|
111 |
-
parquet_file =
|
112 |
dataset = load_dataset("parquet", data_files=parquet_file, name=dataset)
|
113 |
dataset.push_to_hub(HF_DATASET, max_shard_size="500MB")
|
114 |
print("Uploaded to Hugging Face: " + HF_DATASET)
|
115 |
|
116 |
|
117 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
"""
|
119 |
-
|
120 |
-
|
121 |
-
pd.read_xml() errors when XML trees are too large, this is just a hack to
|
122 |
-
download a XML file and parse into a Dataframe. **Not Tested on huge XML files**
|
123 |
-
|
124 |
-
Parameters:
|
125 |
-
response (Requests.Response): Requests response object with the XML data
|
126 |
-
|
127 |
-
Returns:
|
128 |
-
df (DataFrame): A Dataframe from the XML file
|
129 |
"""
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
|
145 |
|
146 |
remove_markdown_links_pattern = r"\[([^\]]+)\]\(([^\)]+)\)"
|
147 |
remove_remaining_links = r"https?:\/\/[^\s]+"
|
148 |
|
149 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
# Replace HTML content to markdown but remove links
|
151 |
def to_markdown(text):
|
152 |
-
|
|
|
|
|
|
|
|
|
153 |
text = re.sub(remove_markdown_links_pattern, r"\1", text)
|
154 |
-
text =
|
155 |
-
|
156 |
-
if "http" in text:
|
157 |
-
raise "Found http in markdown: " + text
|
158 |
-
return text
|
159 |
|
160 |
|
161 |
if __name__ == "__main__":
|
|
|
10 |
import re
|
11 |
from html2text import html2text
|
12 |
from datasets import load_dataset
|
13 |
+
from lxml import etree
|
14 |
+
from tqdm import tqdm
|
15 |
+
import subprocess
|
16 |
+
from merge_parquets import merge_parquet_dir
|
17 |
|
18 |
+
|
19 |
+
XML_DIR = "./xml"
|
20 |
SOURCE = "stackexchange-{0}"
|
21 |
MAX_ANSWERS = 10
|
22 |
QUESTION_SCORE_TRESHOLD = 0
|
23 |
ANSWER_SCORE_TRESHOLD = 0
|
24 |
HF_DATASET = "donfu/oa-stackexchange"
|
25 |
+
PARQUET_FILE = "{0}.parquet"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
|
28 |
def main():
|
29 |
datasets = sys.argv[1:] if len(sys.argv) > 1 else list_cached_datasets()
|
30 |
+
if "temp" in datasets:
|
31 |
+
process_temp_datasets()
|
32 |
for dataset in datasets:
|
33 |
process_dataset(dataset)
|
34 |
|
35 |
|
36 |
def list_cached_datasets():
|
37 |
+
xml_files = glob.glob(f"{XML_DIR}/*.xml")
|
38 |
datasets = [os.path.splitext(os.path.basename(file))[0] for file in xml_files]
|
39 |
datasets.sort()
|
40 |
return datasets
|
41 |
|
42 |
|
43 |
def process_dataset(dataset):
|
44 |
+
xml_file = f"{XML_DIR}/{dataset}.xml"
|
45 |
+
parquet_file = PARQUET_FILE.format(dataset)
|
46 |
source = SOURCE.format(dataset)
|
47 |
+
if os.path.exists(xml_file) and not os.path.exists(parquet_file):
|
48 |
+
df = parse_xml(xml_file)
|
49 |
+
oa = convert_to_oa(df, source)
|
50 |
save_parquet(oa, dataset)
|
51 |
# upload_hf(dataset)
|
52 |
else:
|
53 |
print(f"XML file {xml_file} not found, please download first. Skipping...")
|
54 |
|
55 |
|
56 |
+
def process_temp_datasets():
|
57 |
+
parquet_files = glob.glob(f"temp/?.parquet")
|
58 |
+
for file in parquet_files:
|
59 |
+
print("Reading parquet file: ", file)
|
60 |
+
df = pd.read_parquet(file)
|
61 |
+
print("Converting to Open Assistant format...")
|
62 |
+
oa = convert_to_oa(df, SOURCE.format("stackoverflow"))
|
63 |
+
num = re.search(r"\d", file)[0]
|
64 |
+
parquet_file = f"so/stackoverflow-{num}.parquet"
|
65 |
+
df.to_parquet(parquet_file, row_group_size=100, engine="pyarrow", index=False)
|
66 |
+
print("Wrote parquet file: ", parquet_file)
|
67 |
+
|
68 |
+
merge_parquet_dir("so", "stackoverflow.parquet")
|
69 |
+
|
70 |
+
|
71 |
+
def convert_to_oa(all, source):
|
72 |
"""
|
73 |
Convert dataframe to Open Assistant format with INSTRUCTION, RESPONSE, SOURCE, METADATA columns
|
74 |
|
75 |
Only include questions with an AcceptedAnswerId
|
76 |
"""
|
77 |
+
convert_tags = (
|
78 |
+
lambda raw: raw.replace("-", " ")
|
|
|
79 |
.replace("><", ", ")
|
80 |
.replace("<", "")
|
81 |
.replace(">", "")
|
82 |
+
)
|
83 |
+
create_metadata = lambda row: {
|
84 |
+
"tags": convert_tags(row["Tags_q"]),
|
85 |
+
"question_score": row["Score_q"],
|
86 |
+
"answer_score": row["Score_a"],
|
87 |
}
|
88 |
questions = all[all["AcceptedAnswerId"] != 0]
|
89 |
merged = pd.merge(
|
|
|
94 |
right_on="Id",
|
95 |
suffixes=("_q", "_a"),
|
96 |
)
|
97 |
+
del all
|
98 |
+
|
99 |
merged["INSTRUCTION"] = (
|
100 |
merged["Title_q"] + "\n" + merged["Body_q"].apply(to_markdown)
|
101 |
)
|
102 |
merged["RESPONSE"] = merged["Body_a"].apply(to_markdown)
|
103 |
+
merged["SOURCE"] = source
|
104 |
merged["METADATA"] = merged.apply(create_metadata, axis=1)
|
105 |
|
106 |
return merged[["INSTRUCTION", "RESPONSE", "SOURCE", "METADATA"]]
|
|
|
111 |
Save Dataframe to Parquet. See here for specs:
|
112 |
https://projects.laion.ai/Open-Assistant/docs/data/datasets#creating-a-dataset-on-hugging-face
|
113 |
"""
|
114 |
+
parquet_file = PARQUET_FILE.format(dataset)
|
115 |
df.to_parquet(parquet_file, row_group_size=100, engine="pyarrow", index=False)
|
116 |
+
print(f"Converted {len(df)} instructions into {parquet_file}")
|
117 |
|
118 |
|
119 |
def upload_hf(dataset):
|
120 |
"""
|
121 |
Upload to Hugging Face
|
122 |
"""
|
123 |
+
parquet_file = PARQUET_FILE.format(dataset)
|
124 |
dataset = load_dataset("parquet", data_files=parquet_file, name=dataset)
|
125 |
dataset.push_to_hub(HF_DATASET, max_shard_size="500MB")
|
126 |
print("Uploaded to Hugging Face: " + HF_DATASET)
|
127 |
|
128 |
|
129 |
+
# Define a custom SAX ContentHandler to extract data from the XML file
|
130 |
+
class StackExchangeHandler:
|
131 |
+
def __init__(self, total_rows):
|
132 |
+
self.total_rows = total_rows
|
133 |
+
self.progress_bar = tqdm(total=self.total_rows)
|
134 |
+
self.df = pd.DataFrame(
|
135 |
+
columns=[
|
136 |
+
"Id",
|
137 |
+
"PostTypeId",
|
138 |
+
"Body",
|
139 |
+
"Title",
|
140 |
+
"Tags",
|
141 |
+
"Score",
|
142 |
+
"AcceptedAnswerId",
|
143 |
+
"ParentId",
|
144 |
+
]
|
145 |
+
)
|
146 |
+
|
147 |
+
def startElement(self, name, attrs):
|
148 |
+
if name == "row":
|
149 |
+
row = {}
|
150 |
+
row["Id"] = int(attrs.getValue("Id"))
|
151 |
+
row["PostTypeId"] = int(attrs.getValue("PostTypeId"))
|
152 |
+
row["Body"] = str(attrs.getValue("Body"))
|
153 |
+
row["Title"] = str(attrs.get("Title", ""))
|
154 |
+
row["Tags"] = str(attrs.get("Tags", ""))
|
155 |
+
row["Score"] = int(attrs.get("Score", 0))
|
156 |
+
row["ParentId"] = int(attrs.get("ParentId", 0))
|
157 |
+
row["AcceptedAnswerId"] = int(attrs.get("AcceptedAnswerId", 0))
|
158 |
+
|
159 |
+
self.df = pd.concat(
|
160 |
+
[self.df, pd.DataFrame([row], columns=self.df.columns)],
|
161 |
+
ignore_index=True,
|
162 |
+
)
|
163 |
+
self.progress_bar.update(1)
|
164 |
+
|
165 |
+
|
166 |
+
def parse_xml(path: str):
|
167 |
"""
|
168 |
+
Parse (very large) XML files with sax parser and load it into a pandas Dataframe
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
"""
|
170 |
+
total_rows = int(subprocess.getoutput(f"grep -c '<row' {path}"))
|
171 |
+
print(f"Parsing {total_rows} rows from {path}...")
|
172 |
+
columns = "Id PostTypeId Body Title Tags Score AcceptedAnswerId ParentId"
|
173 |
+
rows = []
|
174 |
+
if total_rows > 50000000:
|
175 |
+
huge_file = True
|
176 |
+
temp_file = 1
|
177 |
+
os.makedirs("temp", exist_ok=True)
|
178 |
+
|
179 |
+
context = etree.iterparse(path, events=("start", "end"))
|
180 |
+
|
181 |
+
for event, element in tqdm(
|
182 |
+
context, total=total_rows * 2
|
183 |
+
): # somehow it does not work just with start event, hence *2
|
184 |
+
if event == "start" and element.tag == "row":
|
185 |
+
row = [
|
186 |
+
int(element.get("Id")),
|
187 |
+
int(element.get("PostTypeId")),
|
188 |
+
element.get("Body"),
|
189 |
+
element.get("Title", ""),
|
190 |
+
element.get("Tags", ""),
|
191 |
+
int(element.get("Score", 0)),
|
192 |
+
int(element.get("AcceptedAnswerId", 0)),
|
193 |
+
int(element.get("ParentId", 0)),
|
194 |
+
]
|
195 |
+
rows.append(row)
|
196 |
+
if huge_file and len(rows) >= 10000000:
|
197 |
+
df = pd.DataFrame(rows, columns=columns.split())
|
198 |
+
df.to_parquet(
|
199 |
+
f"temp/{temp_file}.parquet", engine="pyarrow", index=False
|
200 |
+
)
|
201 |
+
print(f"Wrote temp/{temp_file}.parquet file")
|
202 |
+
rows = []
|
203 |
+
temp_file += 1
|
204 |
+
del df
|
205 |
+
element.clear()
|
206 |
+
element.getparent().remove(element)
|
207 |
+
|
208 |
+
df = pd.DataFrame(rows, columns=columns.split())
|
209 |
+
if huge_file:
|
210 |
+
df.to_parquet(f"temp/{temp_file}.parquet", engine="pyarrow", index=False)
|
211 |
+
del rows
|
212 |
+
del df
|
213 |
+
print("Merging all temp files...")
|
214 |
+
merge_parquet_dir("temp", "temp/merged.parquet")
|
215 |
+
df = pd.read_parquet("temp/merged.parquet")
|
216 |
+
print(f"Loaded full dataset with {len(df)} rows")
|
217 |
+
|
218 |
+
return df
|
219 |
|
220 |
|
221 |
remove_markdown_links_pattern = r"\[([^\]]+)\]\(([^\)]+)\)"
|
222 |
remove_remaining_links = r"https?:\/\/[^\s]+"
|
223 |
|
224 |
|
225 |
+
def remove_emojis(string):
|
226 |
+
emoji_pattern = re.compile(
|
227 |
+
"["
|
228 |
+
"\U0001F600-\U0001F64F" # emoticons
|
229 |
+
"\U0001F300-\U0001F5FF" # symbols & pictographs
|
230 |
+
"\U0001F680-\U0001F6FF" # transport & map symbols
|
231 |
+
"\U0001F1E0-\U0001F1FF" # flags (iOS)
|
232 |
+
"\U00002702-\U000027B0"
|
233 |
+
"\U000024C2-\U0001F251"
|
234 |
+
"]+",
|
235 |
+
flags=re.UNICODE,
|
236 |
+
)
|
237 |
+
return emoji_pattern.sub(r"", string)
|
238 |
+
|
239 |
+
|
240 |
# Replace HTML content to markdown but remove links
|
241 |
def to_markdown(text):
|
242 |
+
try:
|
243 |
+
text = html2text(text, bodywidth=0).strip()
|
244 |
+
except Exception as e:
|
245 |
+
print(e)
|
246 |
+
text = re.sub(r"<[^>]*>", "", str(text))
|
247 |
text = re.sub(remove_markdown_links_pattern, r"\1", text)
|
248 |
+
text = remove_emojis(text)
|
249 |
+
return re.sub(remove_remaining_links, "", text)
|
|
|
|
|
|
|
250 |
|
251 |
|
252 |
if __name__ == "__main__":
|