MedQA-NLI: A Large-Scale Multi-Source Medical Question Answering and Natural Language Inference Dataset

MedQA-NLI is a comprehensive medical reasoning dataset comprising 42,889 instances designed for training and evaluating models on natural language inference (NLI) tasks in biomedical domains.

Each instance pairs a PubMed abstract with a question and a claim, enabling models to perform:

• Medical question answering
• Claim (answer) verification
• Textual entailment recognition

The dataset combines original content from PubMed, PubMedQA, and SciFact with synthetically generated text using multiple state-of-the-art language models via a soft-supervised generation approach.

---

Key Statistics

• Total Instances: 42,889

Label Distribution:

Label	Count	Percentage
Neutral	17,982	41.93%
Entailment	14,674	34.21%
Contradiction	10,233	23.86%

---

Dataset Composition

Abstract Sources

Source	Count	Percentage	Description
PubMed Original	24,907	58.07%	Authentic medical research abstracts
BioMistral-7B-DARE	9,207	21.47%	Synthetic medical abstracts
MedGemma-4B-IT	8,775	20.46%	Synthetic medical abstracts

Claim Sources

Source	Count	Percentage
OLMo-2-1124-7B-Instruct	6,323	14.74%
SOLAR-10.7B-Instruct-v1.0	6,236	14.54%
Mistral-7B-Instruct-v0.3	6,172	14.39%
MedGemma-4B-IT	6,145	14.33%
Llama-3.1-8B-Instruct	6,094	14.21%
PubMedQA (Original)	5,224	12.18%
GPT-OSS-20B	5,224	12.18%
SciFact	1,471	3.43%

Question Sources

Source	Count	Percentage
PubMedQA (Original)	10,448	24.36%
OLMo-2-1124-7B-Instruct	6,323	14.74%
SOLAR-10.7B-Instruct-v1.0	6,236	14.54%
Mistral-7B-Instruct-v0.3	6,172	14.39%
MedGemma-4B-IT	6,145	14.33%
Llama-3.1-8B-Instruct	6,094	14.21%
GPT-OSS-20B	1,471	3.43%

---

Generation Methodology

The dataset was built using a soft-supervised generation strategy:

1. Source Diversity: Multiple LLMs were used to generate abstracts, claims, and questions to ensure linguistic variety and reduce bias.
2. Foundation on Real Data: 58.07% of abstracts are sourced directly from PubMed for medical accuracy.
3. Hybrid Composition: Instances may mix original and synthetic components for balanced training.
4. Quality Control: Generated content was filtered for medical relevance and coherence with PubMed standards.

---

Data Format

Each instance follows this structure:

{
  "abstract": "Medical abstract text...",
  "question": "Question related to the abstract...",
  "claim": "Claim to be verified against the abstract...",
  "label": "entailment|contradiction|neutral",
  "metadata": {
    "abstract": "source information",
    "claim": "source information",
    "question": "source information"
  }
}

as an example:

{
  "abstract": "The hypothesis was tested that pectin content and methylation degree participate in regulation of cell wall mechanical properties and in this way may affect tissue growth and freezing resistance over the course of plant cold acclimation and de-acclimation. Experiments were carried on the leaves of two double-haploid lines of winter oil-seed rape (Brassica napus subsp. oleifera), differing in winter survival and resistance to blackleg fungus (Leptosphaeria maculans)...",
  "question": "Are pectins involved in cold acclimation and de-acclimation of winter oil-seed rape plants?",
  "claim": "The study demonstrates that alterations in temperature do not influence pectin concentration or its degree of methyl esterification within leaf tissues...",
  "label": "contradiction",
  "metadata": {
    "abstract": "original(from PubMedQA)",
    "claim": "generated by GPT-oss-20b",
    "question": "original(from PubMedQA)"
  }
}

• Curated by: Mehdi Nickzamir
• Language(s) (NLP): English
• License: Apache 2