civilcomments-wilds / processing_script.py
pietrolesci's picture
Create processing_script.py
94e0f09 verified
def download_civilcomments(data_path: str | Path) -> None:
print("Downloading CivilComments...")
civilcomments_dir = Path(data_path) / "civilcomments"
civilcomments_dir.mkdir(parents=True, exist_ok=True)
download_and_extract(
"https://worksheets.codalab.org/rest/bundles/0x8cd3de0634154aeaad2ee6eb96723c6e/contents/blob/",
str(civilcomments_dir / "civilcomments.tar.gz"),
)
def process_civilcomments(data_path: str | Path, dst: str | Path, keep_raw: bool = False) -> None:
print("Processing CivilComments...")
df = pd.read_csv(Path(data_path) / "civilcomments/all_data_with_identities.csv", index_col=0)
if keep_raw:
ds = Dataset.from_pandas(df, preserve_index=False)
# save locally
ds.save_to_disk(str(Path(dst) / "civilcomments-wilds" / "raw"))
# push to hub
ds.push_to_hub("<repo name>", "raw")
# extract labels, features, and metadata
input_output_vars = ["id", "split", "comment_text", "toxicity"]
identity_vars = ["male", "female", "LGBTQ", "christian", "muslim", "other_religions", "black", "white"]
auxiliary_vars = [
"identity_any",
"severe_toxicity",
"obscene",
"threat",
"insult",
"identity_attack",
"sexual_explicit",
]
# remove instances where label or text is missing
df = df.loc[df[input_output_vars[-2:]].isna().sum(1) == 0]
# remove instances where label < 0
df = df.loc[df[input_output_vars[-1]] >= 0]
# keep only columns we need
df = df.loc[:, input_output_vars + identity_vars + auxiliary_vars]
# encode label and identity attributes
cols = [input_output_vars[-1]] + identity_vars + auxiliary_vars
df[cols] = (df[cols] >= 0.5).astype(int)
# fmt: off
# deduplicate
gdf = df.groupby("comment_text")[identity_vars + ["split", "toxicity"]].agg("nunique")
gdf["multiple"] = (gdf != 1).sum(1)
print(f"""
There are {df["comment_text"].duplicated().sum()} exact duplicates (i.e., same `comment_text`).
Of these, only {len(gdf.query("multiple > 0"))} are unique `comment_text`.
Some duplicates appear with different attributes and labels, and some even in multiple splits.
In particular,
{(gdf[identity_vars + ["split", "toxicity"]] > 1).sum()}
""")
# if duplicates it keeps:
# - the occurrence in the validation set, or
# - the one with higher toxicity, or
# - the one with higher identity_vars (in order they appears in the list)
# - the one with higher auxiliary_vars (in order they appears in the list)
print(f"Length before deduplication: {len(df)}")
df = (
df.sort_values(["comment_text", "split", "toxicity"] + identity_vars + auxiliary_vars, ascending=False)
.drop_duplicates(subset="comment_text", keep="first")
)
print(f"Length after deduplication: {len(df)}")
# add column with all identity attributes
df = (
df.assign(active_attributes=lambda _df: _df[identity_vars].values.tolist())
.assign(
active_attributes=lambda _df: _df["active_attributes"].map(
lambda lst: [name for idx, name in zip(lst, identity_vars, strict=True) if idx == 1]
)
)
)
# fmt: on
# add column to flag whether any active attribute is present
assert ((df[identity_vars].sum(1) != 0) == (df["active_attributes"].map(len) > 0)).all() # simple check
df["has_active_attrs"] = df[identity_vars].sum(1) != 0
# add unique identifier as first column
df["uid"] = list(range(len(df)))
# reorder columns nicely
cols = df.columns.tolist()
df = df[["uid"] + input_output_vars + ["has_active_attrs", "active_attributes"] + identity_vars + auxiliary_vars]
# convert to DatasetDict
ds_dict = {}
for split in df["split"].unique().tolist():
ds = Dataset.from_pandas(df.query(f"split == '{split}'").drop(columns=["split"]), preserve_index=False)
ds = ds.cast_column("toxicity", ClassLabel(num_classes=2, names=["non-toxic", "toxic"]))
ds_dict[split if split != "val" else "validation"] = ds
ds_dict = DatasetDict(ds_dict)
# save locally
ds_dict.save_to_disk(str(Path(dst) / "civilcomments-wilds" / "texts"))
# push to hub
ds_dict.push_to_hub("<repo name>", "default")