README.md

Overview

Original dataset is available on the HuggingFace Hub here.

Dataset curation

This is the same as the snli_format split of the SciTail dataset available on the HuggingFace Hub (i.e., same data, same splits, etc). The only differences are the following:

• selecting only the columns ["sentence1", "sentence2", "gold_label"]
• renaming columns with the following mapping {"sentence1": "premise", "sentence2": "hypothesis", "gold_label": "label"}
• encoding labels with the following mapping {"entailment": 0, "neutral": 1, "contradiction": 2}

Note that there are 10 overlapping instances (as found by merging on columns "label", "premise", and "hypothesis") between train and test splits.

Code to create the dataset

from datasets import Features, Value, ClassLabel, Dataset, DatasetDict, load_dataset

# load datasets from the Hub
dd = load_dataset("scitail", "snli_format")

ds = {}
for name, df_ in dd.items():
    df = df_.to_pandas()

    # select important columns
    df = df[["sentence1", "sentence2", "gold_label"]]

    # rename columns
    df = df.rename(columns={"sentence1": "premise", "sentence2": "hypothesis", "gold_label": "label"})

    # encode labels
    df["label"] = df["label"].map({"entailment": 0, "neutral": 1, "contradiction": 2})

    # cast to dataset
    features = Features({
        "premise": Value(dtype="string", id=None),
        "hypothesis": Value(dtype="string", id=None),
        "label": ClassLabel(num_classes=3, names=["entailment", "neutral", "contradiction"]),
    })
    ds[name] = Dataset.from_pandas(df, features=features)

dataset = DatasetDict(ds)
dataset.push_to_hub("scitail", token="<token>")

# check overlap between splits
from itertools import combinations
for i, j in combinations(dataset.keys(), 2):
    print(
        f"{i} - {j}: ",
        pd.merge(
            dataset[i].to_pandas(), 
            dataset[j].to_pandas(), 
            on=["label", "premise", "hypothesis"], 
            how="inner",
        ).shape[0],
    )
#> train - test:  10
#> train - validation:  0
#> test - validation:  0