rafmacalaba's picture
Upload README.md with huggingface_hub
4e1dc85 verified
metadata
dataset_info:
  features:
    - name: input
      dtype: string
    - name: output
      list:
        - name: dataset_mention
          struct:
            - name: dataset_name
              dtype: string
            - name: dataset_tag
              struct:
                - name: value
                  dtype: string
                - name: confidence
                  dtype: string
            - name: data_type
              struct:
                - name: value
                  dtype: string
                - name: confidence
                  dtype: string
            - name: acronym
              dtype: string
            - name: author
              dtype: string
            - name: producer
              dtype: string
            - name: publication_year
              dtype: string
            - name: reference_year
              dtype: string
            - name: reference_population
              dtype: string
            - name: geography
              dtype: string
            - name: description
              dtype: string
            - name: is_used
              struct:
                - name: value
                  dtype: bool
                - name: confidence
                  dtype: string
            - name: usage_context
              struct:
                - name: value
                  dtype: string
                - name: confidence
                  dtype: string
  splits:
    - name: train
      num_bytes: 474751
      num_examples: 403
    - name: test
      num_bytes: 49767
      num_examples: 46
    - name: eval
      num_bytes: 54930
      num_examples: 51
  download_size: 154032
  dataset_size: 579448
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: eval
        path: data/eval-*
license: mit
task_categories:
  - token-classification
language:
  - en
tags:
  - GLiNER
  - data-mentions
  - World Bank
  - PRWP

World Bank PRWP - Refugee Data Manual Annotation

Dataset Description

This dataset consists of manually annotated excerpts from text describing data sources. It is intended for training and evaluating Named Entity Recognition (NER) models (specifically designed for the 13-field GLiNER2 data-mention schema) to extract mentions of datasets, databases, surveys, censuses, and other data sources.

The dataset focuses on the PRWP (Poverty and Equity Global Practice) refugee data contexts from the World Bank.

Organization

The dataset is divided into three splits:

  • train: 403 multi-mention records used for training models.
  • eval: 51 multi-mention records used for validation.
  • test: 46 multi-mention records used to benchmark model performance.

Data Instances

Each record corresponds to a text snippet (input) and contains a list of data mentions (output) complying with a strict 13-field JSON schema. The schema enforces verbatim grounding, where each dataset_name, along with non-classification metadata fields (acronym, author, producer, description, etc.), must be an exact substring of the input text.

Features:

  • input: The original text snippet.
  • output: A list of objects containing:
    • dataset_name: The verbatim mention of the data source.
    • dataset_tag: Classification (named, descriptive, vague).
    • data_type: Inferred type of data (survey, census, administrative, database, indicator, geospatial, microdata, report, other).
    • acronym, author, producer, description, geography, publication_year, reference_year, reference_population: Contextual entity string metadata fields representing properties of the dataset.
    • is_used: Indication if this data source was utilized in the research/analysis.
    • usage_context: Role of the data source (primary, supporting, etc.).

Quality Assurance

This ground-truth dataset underwent deep manual auditing and programmatic refinement:

  • Corrected categorization and unified data tags.
  • Verified 100% "verbatim" text grounding for spans.
  • Dropped false-positive non-data mentions, such as bare years, general organization references, or methodology fragments.
  • Detected and merged any duplicate entries or highly overlapping record snippets using Jaccard text similarity.

Usage

This dataset is structurally aligned to be used directly to fine-tune GLiNER2 adapter modules or to evaluate general data-mention entity extraction pipelines.