AntoineBlanot/snli-contrast

---
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: 283196540
num_examples: 1098734
- name: test
num_bytes: 5199496
num_examples: 19684
download_size: 23437414
dataset_size: 288396036
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---
# Dataset Card for "snli-contrast"
This dataset is the [snli-3way](https://huggingface.co/datasets/AntoineBlanot/snli-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)

原始信息汇总

数据集概述

数据集信息

• 特征:

  • premise: 字符串类型
  • hypothesis: 字符串类型
  • instruction: 字符串类型
  • label_name: 字符串类型

• 数据划分:

  • train: 字节数为283196540，样本数为1098734
  • test: 字节数为5199496，样本数为19684

• 数据大小:

  • 下载大小: 23437414字节
  • 数据集大小: 288396036字节

配置信息

• 默认配置:
  • train 数据文件路径: data/train-*
  • test 数据文件路径: data/test-*

数据集描述

该数据集是snli-3way数据集的扩展，新增了instruction特征。该特征及其相关的label_name表达了premise和hypothesis特征在原始数据集中的关系。

• 映射规则:

  • 如果原始样本属于entailment类别:

    • 正例 (label_name == "positive"): 指令为 "The meaning of the hypothesis is logically inferred from the meaning of the premise."
    • 负例 (label_name == "negative"): 指令为 "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."

  • 如果原始样本属于contradiction或neutral类别:

    • 正例 (label_name == "positive"): 指令为 "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."
    • 负例 (label_name == "negative"): 指令为 "The meaning of the hypothesis is logically inferred from the meaning of the premise."

该数据集是原始数据集的两倍大小，因为每个样本都关联了一个正例和一个负例的指令。