README.md

metadata

license: cc-by-4.0
task_categories:
  - token-classification
tags:
  - ner
  - dataset-mention
  - data-use
size_categories:
  - 1K-10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl
      - split: holdout
        path: data/holdout.jsonl

Dataset Card for Datause Dataset

Combined data-mention extraction dataset for the GLiNER2 data-use swarm, with three splits: train, validation, holdout.

Dataset Summary

The dataset is designed to teach Named Entity Recognition (NER) models to extract references to datasets, databases, and surveys from PDF-extracted text.

Splits

split	records	notes
train	1,779	v12-rerun training split (70% positive + 120 pinned hard negatives)
validation	415	v12-rerun validation split (early stopping)
holdout	1,149	canonical holdout_v10 prose-only evaluation benchmark

Schema (GLiNER2 Flat-NER Format)

Each record has two fields:

• input: The raw text paragraph containing potential data mentions.
• output:
  • entities: Dictionaries containing lists of span strings extracted for three categories:
    • named_data: Proper name of a dataset (e.g. National Education Outcomes Registry (NEOR)).
    • descriptive_data: Described data source (e.g. school enrolment and retention indicators).
    • vague_data: General data references (e.g. monitoring data).
  • entity_descriptions: Definitions for the three categories.

Example:

{
  "input": "We use data from the Demographic and Health Surveys (DHS) 2020 and MICS 2019 to analyze child nutrition.",
  "output": {
    "entities": {
      "named_data": ["Demographic and Health Surveys (DHS)", "MICS"],
      "descriptive_data": [],
      "vague_data": ["data"]
    },
    "entity_descriptions": {
      "named_data": "A data mention with a proper name: dataset, survey, database, index, census, registry, or information system.",
      "descriptive_data": "A data mention described by its characteristics or producer rather than a proper name.",
      "vague_data": "A data mention that only generally references data, information, or statistics."
    }
  }
}

Source Documents

• World Bank Group Project Appraisal Documents (PADs) — social protection, education, agriculture, water.
• UNHCR Refugee Operational Reports — livelihood surveys, MSNAs, Protection Briefs.
• Synthetic Target Sets — 27 contexts balancing underrepresented classes.
• Layout Hard Negatives — 120 curated tables/indices/footers with empty entities.

Loading

from datasets import load_dataset
ds = load_dataset("ai4data/datause-dataset")  # DatasetDict: train / validation / holdout