Dataset Card for nli-zh-all

Dataset Description

• Repository: Chinese NLI dataset
• Dataset: zh NLI
• Size of downloaded dataset files: 4.7 GB
• Total amount of disk used: 4.7 GB

Dataset Summary

中文自然语言推理（NLI）数据合集（nli-zh-all）

整合了文本推理，相似，摘要，问答，指令微调等任务的820万高质量数据，并转化为匹配格式数据集。

Supported Tasks and Leaderboards

Supported Tasks: 支持中文文本匹配任务，文本相似度计算等相关任务。

中文匹配任务的结果目前在顶会paper上出现较少，我罗列一个我自己训练的结果：

Leaderboard: NLI_zh leaderboard

Languages

数据集均是简体中文文本。

Dataset Structure

Data Instances

An example of 'train' looks as follows.

{"text1":"借款后多长时间给打电话","text2":"借款后多久打电话啊","label":1}
{"text1":"没看到微粒贷","text2":"我借那么久也没有提升啊","label":0}

• label 有2个标签，1表示相似，0表示不相似。

Data Fields

The data fields are the same among all splits.

• text1: a string feature.
• text2: a string feature.
• label: a classification label, with possible values including entailment(1), contradiction(0)。

Data Splits

after remove None and len(text) < 1 data:

$ wc -l nli-zh-all/*
   48818 nli-zh-all/alpaca_gpt4-train.jsonl
    5000 nli-zh-all/amazon_reviews-train.jsonl
  519255 nli-zh-all/belle-train.jsonl
   16000 nli-zh-all/cblue_chip_sts-train.jsonl
  549326 nli-zh-all/chatmed_consult-train.jsonl
   10142 nli-zh-all/cmrc2018-train.jsonl
  395927 nli-zh-all/csl-train.jsonl
   50000 nli-zh-all/dureader_robust-train.jsonl
  709761 nli-zh-all/firefly-train.jsonl
    9568 nli-zh-all/mlqa-train.jsonl
  455875 nli-zh-all/nli_zh-train.jsonl
   50486 nli-zh-all/ocnli-train.jsonl
 2678694 nli-zh-all/simclue-train.jsonl
  419402 nli-zh-all/snli_zh-train.jsonl
    3024 nli-zh-all/webqa-train.jsonl
 1213780 nli-zh-all/wiki_atomic_edits-train.jsonl
   93404 nli-zh-all/xlsum-train.jsonl
 1006218 nli-zh-all/zhihu_kol-train.jsonl
 8234680 total

Data Length

count text length script: https://github.com/shibing624/text2vec/blob/master/examples/data/count_text_length.py

Dataset Creation

Curation Rationale

受m3e-base启发，合并了中文高质量NLI(natural langauge inference)数据集， 这里把这个数据集上传到huggingface的datasets，方便大家使用。

Source Data

Initial Data Collection and Normalization

如果您想要查看数据集的构建方法，你可以在 https://github.com/shibing624/text2vec/blob/master/examples/data/build_zh_nli_dataset.py 中找到生成 nli-zh-all 数据集的脚本，所有数据均上传到 huggingface datasets。

Who are the source language producers?

数据集的版权归原作者所有，使用各数据集时请尊重原数据集的版权。

SNLI:

@inproceedings{snli:emnlp2015, Author = {Bowman, Samuel R. and Angeli, Gabor and Potts, Christopher, and Manning, Christopher D.}, Booktitle = {Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP)}, Publisher = {Association for Computational Linguistics}, Title = {A large annotated corpus for learning natural language inference}, Year = {2015} }

Who are the annotators?

原作者。

Social Impact of Dataset

This dataset was developed as a benchmark for evaluating representational systems for text, especially including those induced by representation learning methods, in the task of predicting truth conditions in a given context.

Systems that are successful at such a task may be more successful in modeling semantic representations.

Licensing Information

for reasearch

用于学术研究

Contributions

shibing624 add this dataset.