saibo/bookcorpus_deduplicated
收藏资源简介:
这是一个去重版本的原始Book Corpus数据集。Book Corpus(Zhu et al., 2015)曾被用于训练如BERT等流行模型,但根据Bandy和Vincent(2021)的研究,该数据集中存在大量完全重复的文档。去重后,数据集中的重复文本被删除,仅保留首次出现的文本,并保持文本出现的顺序。去重后的数据集包含38,832,894行文本,大小为2.91GB,而去重前为74,004,228行和4.63GB。去重训练数据具有多种优势,包括减少训练步骤、降低模型记忆文本的频率以及减少碳排放和能源消耗。
This is a deduplicated version of the original Book Corpus dataset. Book Corpus (Zhu et al., 2015) was previously utilized to train popular models such as BERT. However, according to the study by Bandy and Vincent (2021), the original dataset contained a large number of exact duplicate documents. After deduplication, duplicate texts within the dataset are removed, retaining only the texts that appear for the first time while maintaining their original occurrence order. The deduplicated dataset contains 38,832,894 rows of text with a total size of 2.91 GB, compared with 74,004,228 rows and 4.63 GB prior to deduplication. Deduplicated training data offers multiple advantages, including reducing training steps, lowering the frequency of model text memorization, as well as cutting carbon emissions and energy consumption.
数据集概述
数据集名称
- 名称: bookcorpus_deduplicated
数据集特征
- 特征:
- text: 数据类型为字符串(string)
数据集拆分
- 拆分:
- train:
- num_bytes: 2867856394
- num_examples: 38832894
- train:
数据集大小
- 下载大小: 1794567875
- 数据集大小: 2867856394
数据集描述
- 描述: 此数据集是原始Book Corpus数据集的去重版本。原始数据集存在大量重复文档,去重后仅保留文本首次出现的内容,并保持文本出现的顺序。
去重效果
- 行数对比: 38832894(去重后)VS 74004228(原始)
- 数据集大小对比: 2.91GB(去重后)VS 4.63GB(原始)
去重原因
- 优势:
- 减少训练步骤,达到相同或更高精度
- 减少模型输出记忆文本的频率
- 降低碳排放和能源消耗
去重脚本
- 脚本概述: 使用Python脚本进行去重,通过Pandas和datasets库实现,运行时间短,不到几分钟。
参考文献
- 文献1: Bandy, Jack and Vincent, Nicholas. "Addressing Documentation Debt in Machine Learning Research: A Retrospective Datasheet for BookCorpus". arXiv:2105.05241 [cs.CL, cs.CY, cs.LG].
- 文献2: Lee, Katherine et al. "Deduplicating Training Data Makes Language Models Better". arXiv:2107.06499 [cs.CL, cs.LG].




