Datasets:

qualifire
/

grounding-benchmark

language: en
tags:
  - claim-grounding
  - natural-language-inference
  - reasoning
  - classification
  - grounding
  - hallucination
pretty_name: Grounding Claims Dataset
license: cc-by-nc-4.0
task_categories:
  - text-classification
dataset_info:
  features:
    - name: doc
      dtype: string
    - name: claim
      dtype: string
    - name: label
      dtype: string
    - name: dataset
      dtype: string
  splits:
    - name: test
      num_bytes: 1301919
      num_examples: 2700
  download_size: 766964
  dataset_size: 1301919
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

The Grounding Claims Dataset is a multi-domain dataset for evaluating whether a natural language claim is grounded (i.e., supported or entailed) by a document. The dataset is organized into four subsets, each requiring different types of reasoning:

general (1500 examples): Broad, everyday reasoning
logical (1000 examples): Logical consistency and inference
time_and_dates (100 examples): Temporal reasoning
prices_and_math (100 examples): Numerical and mathematical reasoning

Each entry consists of:

doc: A short context or passage
claim: A natural language statement to verify against the doc
label: A binary label indicating whether the claim is grounded in the document (1 for grounded, 0 for ungrounded)
dataset: The source subset name (e.g., "general")

📌 Features

Feature	Type	Description
`doc`	string	The document or passage providing the context
`claim`	string	A statement to verify against the document
`label`	string	grounded or ungrounded
`dataset`	string	The domain/subset the instance belongs to

📊 Usage

This dataset can be used to train and evaluate models on factual verification, natural language inference (NLI), and claim grounding tasks across multiple domains.

🏷️ Labels

grounded — The claim is grounded in the document.
ungrounded — The claim is ungrounded or contradicted by the document.