CitationGround-1M: A Citation-Locked QA Dataset for RAG

CitationGround-1M is a citation-locked grounded QA/RAG dataset requiring answers to be based solely on provided contexts. It includes span-level citations and answerable=false hard negatives for training abstention behavior. The dataset was created by WithinUsAI and published on Hugging Face in December 2025.

Use Cases

• Train models to answer questions based solely on provided contexts, as described in the dataset's citation-locked QA feature.
• Evaluate model abstention behavior using the answerable=false hard negatives included in the dataset.
• Benchmark span-level citation accuracy in QA systems, leveraging the dataset's doc_id and offset citation format.

Strengths

• Includes hard negatives labeled answerable=false specifically for training abstention behavior.
• Requires span-level citations (doc_id + offsets), providing granular attribution for answers.

Limitations

• Row count, file formats, and column-level documentation are unknown, which may limit suitability assessment.
• License information is unknown, which may restrict usage.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

WithinUsAI

Freshness

Last updated 2025-12-31 09:04:03; freshness should be verified.