datajuicer/redpajama-pile-stackexchange-refined-by-data-juicer

这是一个由Data-Juicer工具精炼的RedPajama & The Pile中的StackExchange数据集版本。通过去除原始数据集中的一些低质量样本，提高了数据集的质量。该数据集通常用于预训练大型语言模型。完整数据集大小约为71GB，包含26,309,203个样本，保留了原始数据集的约57.89%。This is a refined StackExchange dataset variant sourced from RedPajama & The Pile, processed with the Data-Juicer tool. By eliminating low-quality samples from the original dataset, the overall quality of this dataset has been enhanced. This dataset is typically utilized for pre-training Large Language Models (LLMs). The complete dataset has a size of roughly 71 GB, contains 26,309,203 samples, and retains approximately 57.89% of the original dataset.

原始信息汇总

RedPajama & The Pile -- StackExchange (refined by Data-Juicer)

概述

这是一个经过Data-Juicer精炼的StackExchange数据集版本，属于RedPajama & The Pile项目。该数据集去除了一些“不良”样本，以提高整体质量。通常用于预训练大型语言模型。

数据集信息

• 样本数量: 26,309,203个样本（保留了原始数据集的约57.89%）

精炼配方

yaml

全局参数

project_name: Data-Juicer-stack-exchange dataset_path: /path/to/your/dataset # 数据集目录或文件路径 export_path: /path/to/your/dataset.jsonl

np: 50 # 处理数据集的子进程数量 open_tracer: true

处理流程

一系列处理操作及其参数

process:

• clean_email_mapper:

• clean_links_mapper:

• fix_unicode_mapper:

• punctuation_normalization_mapper:

• whitespace_normalization_mapper:

• alphanumeric_filter: tokenization: false min_ratio: 0.35 # <3sigma max_ratio: 0.943 # 3sigma

• average_line_length_filter: # 针对代码 min_len: 20 # >3sigma max_len: 400 # >3sigma

• character_repetition_filter: rep_len: 10 max_ratio: 0.4 # >3sigma (0.12)

• flagged_words_filter: lang: en tokenization: true max_ratio: 0.01 # >3sigma

• language_id_score_filter: # 移除语言过滤器 min_score: 0.1 # <3sigma

• maximum_line_length_filter: # 针对代码 min_len: 80

• perplexity_filter: lang: en max_ppl: 10000 # >3sigma

• special_characters_filter: min_ratio: 0.232 # 3sigma max_ratio: 0.7 # >3sigma

• text_length_filter: min_len: 200

• words_num_filter: lang: en tokenization: true min_num: 100

• word_repetition_filter: lang: en tokenization: true rep_len: 10 max_ratio: 0.8 # >3sigma

• document_simhash_deduplicator: #26309203 left tokenization: space window_size: 3 lowercase: true ignore_pattern:

  num_blocks: 9 hamming_distance: 7