---
dataset_info:
  features:
  - name: premise
    dtype: string
  - name: hypothesis
    dtype: string
  - name: instruction
    dtype: string
  - name: label_name
    dtype: string
  splits:
  - name: train
    num_bytes: 254483428
    num_examples: 785404
  - name: test
    num_bytes: 6297986
    num_examples: 19630
  download_size: 54354034
  dataset_size: 260781414
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---
# Dataset Card for "mnli-contrast"
This dataset is the [mnli-3way](https://huggingface.co/datasets/AntoineBlanot/mnli-3way) dataset with an additional `instruction` feature.
This new feature along with its related `label_name` expresses how the `premise` and `hypothesis` features are related in the original dataset.

The following explains how the mapping is done:

### If the original example was of class `entailment`
Two data points will be related to that example.

One is the positive example (i.e., `label_name` == "positive") which assign to it the folowing instruction: "The meaning of the hypothesis is logically inferred from the meaning of the premise."
The other is the negative example (i.e., `label_name` == "negative") which assign to it the folowing instruction: "The meaning of the hypothesis either contradicts the meaning of the premise, is unrelated to it, or does not provide sufficient information to infer the meaning of the premise."

### If the original example was of class `contradiction` or `neutral`
Two data points will be related to that example.

One is the positive example (i.e., `label_name` == "positive") which assign to it the folowing instruction: "The meaning of the hypothesis either contradicts the meaning of the premise, is unrelated to it, or does not provide sufficient information to infer the meaning of the premise."
The other is the negative example (i.e., `label_name` == "negative") which assign to it the folowing instruction: "The meaning of the hypothesis is logically inferred from the meaning of the premise."

This dataset is double the size of this original dataset because each is related to a positive and negative instruction.

[More Information needed](https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)