README.md

metadata

language: en
license: odc-by
tags:
  - text-classification
  - domain-classification
  - c4
task_categories:
  - text-classification
  - text-retrieval
configs:
  - config_name: default
    data_files:
      - split: train
        path: all.jsonl.zst
  - config_name: pristine
    data_files:
      - split: train
        path: pristine/train.jsonl.zst
      - split: validation
        path: pristine/validation.jsonl.zst
      - split: test
        path: pristine/test.jsonl.zst
  - config_name: classla_news
    data_files:
      - split: train
        path: classla_news/train.jsonl.zst
      - split: validation
        path: classla_news/validation.jsonl.zst
      - split: test
        path: classla_news/test.jsonl.zst
  - config_name: classla_ParlaCAP
    data_files:
      - split: train
        path: classla_ParlaCAP/train.jsonl.zst
      - split: validation
        path: classla_ParlaCAP/validation.jsonl.zst
      - split: test
        path: classla_ParlaCAP/test.jsonl.zst
  - config_name: doc_type_v1_primary
    data_files:
      - split: train
        path: doc_type_v1_primary/train.jsonl.zst
      - split: validation
        path: doc_type_v1_primary/validation.jsonl.zst
      - split: test
        path: doc_type_v1_primary/test.jsonl.zst
  - config_name: doc_type_v2_primary
    data_files:
      - split: train
        path: doc_type_v2_primary/train.jsonl.zst
      - split: validation
        path: doc_type_v2_primary/validation.jsonl.zst
      - split: test
        path: doc_type_v2_primary/test.jsonl.zst
  - config_name: fdc_label
    data_files:
      - split: train
        path: fdc_label/train.jsonl.zst
      - split: validation
        path: fdc_label/validation.jsonl.zst
      - split: test
        path: fdc_label/test.jsonl.zst
  - config_name: nvidia_domain
    data_files:
      - split: train
        path: nvidia_domain/train.jsonl.zst
      - split: validation
        path: nvidia_domain/validation.jsonl.zst
      - split: test
        path: nvidia_domain/test.jsonl.zst

English Document Classification Dataset

This dataset provides a curated subset of the first 1 million rows from the allenai/c4 (English configuration), enriched with multi-perspective topic annotations. It is designed for researchers exploring document classification, domain adaptation, and label noise in massive web-crawled corpora.

Dataset Summary

The dataset integrates predictions from five distinct classification models to provide a holistic view of each document’s content. To ensure utility, the data was stratified across these variables and split into 80/10/10% (Train/Validation/Test) sets, organized into specific configurations based on the source classifier.

Classifier Metadata & Schema

The following classifiers were used to generate the annotation columns:

Classifier	Column Name	Description
nvidia/domain-classifier	nvidia_domain	General web domain categorization.
classla/multilingual-IPTC-news-topic-classifier	classla_news	Standardized news industry topic codes.
classla/ParlaCAP-Topic-Classifier	classla_ParlaCAP	Legislative and parliamentary topic categories.
cardiffnlp/tweet-topic-latest-multi	cardiffnlp_tweet	Social-media style topic classification.
EssentialAI/eai-distill-0.5b	fdc_label and other columns	Categorical mapping derived from FDC data and content analysis.

FDC (Free Decimal Correspondence) Mapping: Labels were derived from predictions by EssentialAI/eai-distill-0.5b. Call numbers were truncated to the nearest multiple of 10 and mapped to official FDC categories.

Quality Indicators

• pristine column: A boolean flag based on data generated by the EAI classifier. It indicates whether the document is structurally complete or likely missing content/context.
• Configurations: The dataset is partitioned into configs (e.g., classla_ParlaCAP) to allow users to load data stratified by specific model outputs.

Limitations & Biases

Beware of the following caveats in automated labelling:

• Class Imbalance: Significant distribution shifts exist between categories, reflecting the natural frequency of topics in the C4 corpus.
• Silver-Standard Labels: All labels are model-generated (silver-standard). Errors or biases present in the source classifiers will be reflected in this dataset.
• Label Ambiguity: Web documents often overlap multiple domains (e.g., a personal blog discussing both geopolitics and cooking). Single-label assignment may oversimplify these documents.

Licensing

This dataset is released under the Open Data Commons Attribution License (ODC-BY). Please attribute the source models and the creators of the allenai/c4 dataset when using this resource.