MedHallu: 10,000 QA Pairs for Evaluating LLM Factuality in Healthcare

MedHallu is a dataset of 10,000 question-answer pairs designed for evaluating the reliability and factuality of large language models in healthcare. It was sourced from Kaggle, but the author, organization, and specific creation details are not provided. The dataset's last update date and license information are unknown.

Use Cases

• Benchmarking LLM factuality in medical contexts based on the described QA pairs.
• Training or fine-tuning models to detect or reduce hallucinations in medical LLM outputs.
• Developing evaluation metrics for LLM reliability in healthcare question-answering.

Strengths

• Contains 10,000 QA pairs, providing a substantial corpus for evaluation.
• Specifically targets the critical domain of healthcare LLM reliability and factuality.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific content, source, and annotation methodology are not detailed.
• Last update date is unknown; freshness unverified.

Provenance

Source

Kaggle