Datasets:

OxAISH-AL-LLM
/

wiki_toxic

	import re
	import math

	import pandas as pd

	from tqdm import tqdm

	seed = 7497

	TOXIC_COLUMNS = [
	"toxic",
	"severe_toxic",
	"obscene",
	"threat",
	"insult",
	"identity_hate",
	]

	# Time and date regexes
	TIME = r"([0-9]{1,2}:[0-9]{2}( (am\|AM\|pm\|PM))?)"
	DAY = r"([23]?(1(st)?\|2(nd)?\|3(rd)?\|[4-9](th)?)\|1[0-9](th)?)"
	MONTH = r"(January\|February\|March\|April\|May\|June\|July\|August\|September\|October\|November\|December\|Jan\|Feb\|Mar\|Apr\|May\|Jun\|Jul\|Aug\|Sep\|Nov\|Dec)"
	YEAR = r"('?[0-9]{2}\|[0-9]{4})"
	DATE = rf"(({DAY} {MONTH}\|{MONTH} {DAY})(,? {YEAR})?)"
	TIMESTAMP = rf"((({TIME},? ($UTC$ )?)?{DATE}\|({DATE},? )?{TIME})(\s+$UTC$)?)"

	# The 'talk' part at the end of a signature
	TALK = r"((\\|\s\|\(\s)?[tT]alk((\s[-\|•, ]\s\|\s+)[cC]ontribs)?(\s*[-\|)])?)"

	# IP addresses
	IP = r"([0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3})"

	# Username and the username part of a the signature
	USERNAME = r"([^#<>[\]\|{}/@\s]+)"
	USER_SIG = rf"((((?:\s)[-–—]\s)?($(User:)?{USERNAME}$\|User:{USERNAME})\|(?:\s)[-–—]\s{USERNAME})(\s+{TALK})?)"

	# A full signature
	SIGNATURE = rf"(((([-–—]\s)?{IP}(\s+{USER_SIG})?\|(?:\s)[-–—]\s[uU]nsigned\|{TALK}\|{USER_SIG})(\s+{TIMESTAMP})?)\|{TIMESTAMP}(\s+{TALK})?)"

	# List of the patterns to remove
	REGEX_REMOVE = [
	r"^(\"+\|'+)", # Initial quotation marks
	r"(\"+\|'+)$", # Final quotation marks
	r"^REDIRECT.*$", # The whole comment is a redirect
	rf"^\s*{SIGNATURE}", # Initial signature
	rf"{SIGNATURE}\s*$", # Final signature
	r" \[[0-9]+\]\|\[[0-9]+\] ", # Citations
	r"‖\s+[tT]alk - [-a-zA-Z0-9._()\s]+‖",
	r"==[^=]+==",
	r"^::+",
	r"^\s*$UTC$",
	rf"Unblock {IP}",
	r"2nd Unblock Request",
	r":Category:",
	r"File:[^\s]+",
	r"\{\\|.+\\|\}", # Embedded code
	# r"\{\{.+\s.+\}\}", # Embedded code
	r"^\s+", # Initial whitespace
	r"\s+$", # Trailing whitespace
	]

	# List of patterns to replaces
	REGEX_REPLACE = {
	"\n+": "\n",
	"\\'": "'",
	'""+': '"',
	"''+": "'",
	# r"(WP\|Wikipedia):[^\s]+": "URL", # Wikipedia internal links
	r"[^\s]+#[^\s]+": "URL", # Wikipedia internal links
	r"https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{2,256}\.[a-z]{2,4}\b([-a-zA-Z0-9@:%_\+.~#?&//=]*)": "URL", # ULRs
	r"([uU]ser_[tT]alk\|[tT]alk):[^\s]+": "URL", # Talk links
	}


	def clean_sentence(sentence):
	"""Preprocess a sentence using the regex rules"""
	for pattern in REGEX_REMOVE:
	sentence = re.sub(pattern, "", sentence)
	for pattern, repl in REGEX_REPLACE.items():
	sentence = re.sub(pattern, repl, sentence)
	return sentence


	def make_binary_label(row):
	"""Make a row label binary by combining all toxicity types"""
	for column in TOXIC_COLUMNS:
	if row[column] == 1:
	return 1
	return 0


	print("Loading original data...")

	# Load up the original data
	train_df = pd.read_csv("orig_train.csv").set_index("id")
	test_text_df = pd.read_csv("orig_test.csv").set_index("id")
	test_labels_df = pd.read_csv("orig_test_labels.csv").set_index("id")

	# Remove the datapoints which have no label
	test_text_df = test_text_df.loc[test_labels_df["toxic"] != -1]
	test_labels_df = test_labels_df.loc[test_labels_df["toxic"] != -1]

	# Join the test text and labels to make a complete dataset
	test_df = test_text_df.join(test_labels_df)

	print("Cleaning train split...")
	for index, row in tqdm(train_df.iterrows(), total=len(train_df)):
	row["comment_text"] = clean_sentence(row["comment_text"])

	print("Cleaning test split...")
	for index, row in tqdm(test_df.iterrows(), total=len(test_df)):
	row["comment_text"] = clean_sentence(row["comment_text"])


	# Some texts will get reduced to the empty string. Let's remove them first
	print("Removing empty texts...")
	train_df = train_df.loc[train_df["comment_text"] != ""]
	test_df = test_df.loc[test_df["comment_text"] != ""]

	# Get rid of any duplicates we made
	print("Removing duplicate entries...")
	train_df = train_df.drop_duplicates(subset=["comment_text"])
	test_df = test_df.drop_duplicates(subset=["comment_text"])

	print("Creating binary column...")

	# Make the new binary column
	train_df["label"] = train_df.apply(make_binary_label, axis=1)
	test_df["label"] = test_df.apply(make_binary_label, axis=1)

	# Remove all other classification columns
	train_df = train_df.drop(columns=TOXIC_COLUMNS)
	test_df = test_df.drop(columns=TOXIC_COLUMNS)

	print("Creating eval split...")

	# Shuffle the current train split
	train_df = train_df.sample(frac=1, random_state=seed)

	# The new size of the train split
	train_size = math.floor(len(train_df) * 0.8)

	# Separate into train and eval splits
	eval_df = train_df[train_size:]
	train_df = train_df[:train_size]

	# print("Saving to disk...")
	with open("train.csv", "w") as f:
	train_df.to_csv(f)
	with open("validation.csv", "w") as f:
	eval_df.to_csv(f)
	with open("test.csv", "w") as f:
	test_df.to_csv(f)