README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - zh
pretty_name: ChineseWebText2.0-HighQuality
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*
    features:
      - name: text
        dtype: string
      - name: domain
        struct:
          - name: single_label
            dtype: string
          - name: multi_label
            sequence:
              dtype: string
      - name: toxicity
        struct:
          - name: label
            dtype: int32
          - name: score
            dtype: float64
      - name: quality_score
        dtype: float64

📘 ChineseWebText2.0-HighQuality

Overview

ChineseWebText2.0-HighQuality is a high-quality filtered subset of the original CASIA-LM/ChineseWebText2.0 dataset (Apache-2.0 License).

This subset retains only samples with:

• quality_score ≥ 0.9
• toxicity.score ≤ 0.01

The goal is to provide a cleaner and more reliable dataset suitable for language model pre-training, instruction tuning, and quality-sensitive downstream tasks.

This work is independent and not affiliated with the official CASIA-LM / ChineseWebText maintainers.

---

Key Features

• ✔ Derived from ChineseWebText2.0 (Apache-2.0 License)
• ✔ Quality-filtered using quality_score ≥ 0.9
• ✔ Safety-filtered using toxicity.score ≤ 0.01
• ✔ Light text cleaning: filters non-standard characters
• ✔ Retains original metadata schema (domain/toxicity/quality_score)
• ✔ Suitable for LLM pretraining and mixture construction

---

Dataset Construction

Source Dataset

• CASIA-LM/ChineseWebText2.0
• License: Apache-2.0

Filtering Rule

quality_score >= 0.9 and toxicity.score <= 0.01

Processing Notes

• The text is lightly cleaned by filtering non-standard characters.
• Only samples failing the thresholds are removed.
• No rewriting, deduplication, or translation has been applied in this derived subset.
• The dataset retains the original metadata schema.

---

Use Cases

Recommended

• Pretraining decoder/encoder–decoder language models
• Quality-sensitive mixture construction
• Safety-sensitive data distillation
• Domain-specific filtering based on the retained domain labels

Not Recommended

• Research requiring the unchanged distribution of the full ChineseWebText2.0 corpus
• Studies that explicitly need toxic/low-quality tails for robustness evaluation

---

License

This dataset is distributed under:

• Apache License 2.0 (same as the source dataset)

The original license and attribution to the source dataset creators are preserved.

---

Citation

If you use this dataset, please cite:

@misc{chinesewebtext2_highquality,
  title  = {ChineseWebText2.0-HighQuality: A High-Quality Subset of ChineseWebText2.0},
  author = {Morton Li},
  year   = {2026},
  note   = {Derived from CASIA-LM/ChineseWebText2.0 (Apache-2.0 License)}
}

---

Acknowledgements

Special thanks to the creators of:

• ChineseWebText2.0 (CASIA-LM)

Their open dataset and toolchain enable high-quality Chinese web corpora research for the community.