sileod/deberta-v3-base-tasksource-nli
Zero-Shot Classification • 0.2B • Updated • 13.2k • • 133
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Original data available here.
premise and hypothesis columns have been cleaned following common practices (1, 2), that is
<b>, <u>, </b>, </u>mean_human_score has been transformed into class labels following common practices (1, 2), that is
mean_human_score <= 3 -> "not-entailed" and mean_human_score >= 4 -> "entailed" (anything between 3 and 4 has been removed)mean_human_score < 3.5 -> "not-entailed" else "entailed"more details below.
import pandas as pd
from datasets import Features, Value, ClassLabel, Dataset, DatasetDict
def convert_label(score, is_test):
if is_test:
if score <= 3:
return "not-entailed"
elif score >= 4:
return "entailed"
return "REMOVE"
if score < 3.5:
return "not-entailed"
return "entailed"
ds = {}
for split in ("dev", "test", "train"):
# read data
df = pd.read_csv(f"<path to folder>/AN-composition/addone-entailment/splits/data.{split}", sep="\t", header=None)
df.columns = ["mean_human_score", "binary_label", "sentence_id", "adjective", "noun", "premise", "hypothesis"]
# clean text from html tags and useless spaces
for col in ("premise", "hypothesis"):
df[col] = (
df[col]
.str.replace("(<b>)|(<u>)|(</b>)|(</u>)", " ", regex=True)
.str.replace(" {2,}", " ", regex=True)
.str.strip()
)
# encode labels
if split == "test":
df["label"] = df["mean_human_score"].map(lambda x: convert_label(x, True))
df = df.loc[df["label"] != "REMOVE"]
else:
df["label"] = df["mean_human_score"].map(lambda x: convert_label(x, False))
assert df["label"].isna().sum() == 0
df["label"] = df["label"].map({"not-entailed": 0, "entailed": 1})
# cast to dataset
features = Features({
"mean_human_score": Value(dtype="float32"),
"binary_label": Value(dtype="string"),
"sentence_id": Value(dtype="string"),
"adjective": Value(dtype="string"),
"noun": Value(dtype="string"),
"premise": Value(dtype="string"),
"hypothesis": Value(dtype="string"),
"label": ClassLabel(num_classes=2, names=["not-entailed", "entailed"]),
})
ds[split] = Dataset.from_pandas(df, features=features)
ds = DatasetDict(ds)
ds.push_to_hub("add_one_rte", token="<token>")
# check overlap between splits
from itertools import combinations
for i, j in combinations(ds.keys(), 2):
print(
f"{i} - {j}: ",
pd.merge(
ds[i].to_pandas(),
ds[j].to_pandas(),
on=["premise", "hypothesis", "label"],
how="inner",
).shape[0],
)
#> dev - test: 0
#> dev - train: 0
#> test - train: 0