agentlans/BAAI-IndustryCorpus2

Unofficial BAAI/IndustryCorpus2 Subset

This dataset is an unofficial, curated subset of IndustryCorpus2, a massive collection of industry-specific text data compiled by the Beijing Academy of Artificial Intelligence (BAAI).

It is designed to provide a lightweight, high-quality, English-focused sample of the broader corpus for quick experimentation, prototyping, and analysis.

Dataset Details

• Total Rows: Approximately 1,000,000 (10,000 rows sampled per file across 100 files)
• Format: Compressed JSONLines
• Language: English (en)
• Processing: No additional text processing or cleaning was performed other than random sampling, shuffling, and adding a source tracking column.

Method

1. Filtering: From the original pool of IndustryCorpus2 files, 100 Parquet files explicitly flagged as English and categorized as High Quality were selected.
2. Sampling: 10,000 rows were randomly downloaded from each of the 100 chosen files.
3. Tracking: An extra column (file) was appended to every row to preserve metadata pointing back to its original source file path.
4. Shuffling: The combined pool of 1,000,000 rows was thoroughly shuffled to eliminate ordering bias.

Dataset Structure

Each entry in the dataset includes the raw text along with quality scores and category metadata inherited from the original BAAI corpus.

• text (String): The raw textual content.
• alnum_ratio (Float): Ratio of alphanumeric characters to total characters.
• avg_line_length (Float): Average character length per line.
• char_rep_ratio (Float): Repetition ratio of character sequences.
• max_line_length (Integer): Maximum character length of a single line.
• num_words (Integer): Total word count.
• quality_score (Float): The heuristic quality score calculated by the original BAAI pipeline.
• special_char_ratio (Float): Ratio of special characters in the text.
• industry_type (String): The broad industry/topic classification (e.g., 文学_情感 for Literature/Emotion).
• file (String): The source filename from the original repository.

Example:

{
  "text": "> Before the Dawn\nBefore the Dawn Analysis\nToson Shimazaki\nNavigate Study Guiderows Analysis Before the Dawn (Literary Masterpieces, Volume 8)\nDespite a formidable reputation as a premier novelist of prewar Japan, Tson Shimazaki has remained largely unknown to Western readers, as his novels, always lengthy and complex, have daunted the enthusiasm of potential translators. This fine English version of his most significant and longest work of fiction presents the author's compelling vision of the Westernization of his country. Before the Dawn has long assumed the status of a modern classic in Japan...\nBefore the Dawn Bibliography (Literary Masterpieces, Volume 8)...",
  "alnum_ratio": 0.8009302326,
  "avg_line_length": 268.75,
  "char_rep_ratio": 0.041482172,
  "max_line_length": 1083,
  "num_words": 820,
  "quality_score": 4.35546875,
  "special_char_ratio": 0.2237209302,
  "industry_type": "文学_情感",
  "file": "literature_emotion/english/high/rank_01070.parquet"
}

Limitations & Safety

• Domain Focus: The text heavily favours general world knowledge, culture, and industry text. It contains minimal coverage of raw code or complex mathematics.
• Temporal Accuracy: The data represents a snapshot in time; it may contain outdated facts or inaccurate general knowledge.
• Personally Identifiable Information (PII): Because this text remains unfiltered except for quality, it may inadvertently contain PII. Use caution before using this data to train public-facing generative models.
• Usage Restrictions: Per the original creators' mandates, this dataset must not be used for harmful human experimentation, toxic content generation, or any malicious behavior.

Licensing & Citation

This subset is distributed under the Apache License 2.0.

If you use this subset in your research, please make sure to credit the original authors of the core dataset:

@misc {beijing_academy_of_artificial_intelligence,
   author       = { Xiaofeng Shi and Lulu Zhao and Hua Zhou and Donglin Hao },
   title        = { IndustryCorpus2 },
   year         = 2024,
   url          = { https://huggingface.co/datasets/BAAI/IndustryCorpus2 },
   doi          = { 10.57967/hf/3488 },
   publisher    = { Hugging Face }
}