MedHAM — Medical Hallucination, Accuracy & Misinformation Evaluation Dataset

Overview

MedHAM is a multi-layer evaluation dataset measuring hallucination, accuracy, and misinformation across four large language models answering evidence-based medical questions. It was produced by a 5-layer automated pipeline combining BioBERT semantic similarity, MedNLI inference, citation verification, FActScore factual grounding, and a 4-judge LLM panel.

The dataset supports the study:

"Evaluating Hallucination, Accuracy, and Misinformation in Large Language Models for Medical Question Answering: A Comparative Study of Retrieval-Augmented Generation and Citation Prompting"

---

Setup

pip install datasets pandas pyarrow

---

Dataset Structure

The dataset consists of 8 tables exported from a normalised SQLite database. Each file is a self-contained Parquet table; foreign keys are preserved as integer ID columns.

Table	Rows	Description
benchmark_items	1,000	Medical questions and ground truths from MedHallu (pqa_labeled)
models	5	LLMs used as tested models and/or judges
prompt_strategies	4	2×2 factorial conditions (zero-shot / citation × no-RAG / RAG)
rag_contexts	1,000	Retrieved knowledge context per question
model_responses	16,000	Raw LLM responses — 4 models × 4 strategies × 1,000 questions
evaluation_signals	16,000	Automated signals: BioBERT, NLI, citations, FActScore
judge_evaluations	64,000	Per-judge votes — 4 judges × 16,000 responses
evaluation_results	16,000	Final aggregated labels: accuracy, hallucination, misinformation

---

Experimental Design

Four prompt strategies were constructed from a 2×2 factorial design crossing two independent factors: retrieval augmentation (RAG vs. no RAG) and prompt format (zero-shot vs. citation-required).

Prompt Format	No RAG	RAG
Zero-Shot	ZS_NoRAG (baseline)	ZS_RAG
Citation-Required	CIT_NoRAG	CIT_RAG

Tested Models (4):

• GPT-4o (gpt-4o-2024-11-20) — OpenAI
• Claude Sonnet 4.6 (claude-sonnet-4-6) — Anthropic
• Gemini 2.5 Flash (gemini-2.5-flash) — Google
• Llama 3.1 8B (llama-3.1-8b-instant) — Meta via Groq

Judge Panel (4, blind evaluation):

• GPT-4o, Claude Sonnet 4.6, Gemini 2.5 Flash — tested models acting as blind judges
• DeepSeek V4 Flash (deepseek-v4-flash) — independent judge, no self-evaluation bias

All judges weighted equally (1.0) following Verga et al. (2024).

---

Evaluation Pipeline (5 Layers)

Layer	Method	Signal
1a	BioBERT F1 (dmis-lab/biobert-base-cased-v1.2)	Semantic similarity to ground truth
1b	MedNLI (PubMedBERT-MNLI-MedNLI)	Contradiction detection
2	Citation verification (CrossRef + PubMed)	Fake citation detection
3	FActScore (Min et al., 2023)	Atomic claim verification via PubMed
4–5	4-judge LLM panel + rule-based aggregation	Final verdict

---

Key Columns

evaluation_results — primary outcome table:

• accuracy_score — 0 (inaccurate), 1 (partly accurate), 2 (accurate)
• hallucination_label — 0/1 binary
• misinformation_label — 0/1 binary
• accuracy_consensus — weighted mean of judge votes (0–2, continuous)
• hallucination_consensus — weighted fraction of judges voting hallucination (0–1)
• misinformation_consensus — weighted fraction of judges voting misinformation (0–1)
• track_consensus — 1 if automated signals and judge panel agree directionally

evaluation_signals — automated layer outputs:

• biobert_precision, biobert_recall, biobert_f1
• misinfo_nli_signal, intrinsic_nli_signal — entailment / neutral / contradiction
• citation_real_count, citation_fake_count, citation_verdict
• factscore_f1, factscore_supported, factscore_total
• automated_track_verdict — likely_accurate / likely_hallucinated / uncertain

---

Loading the Data

import pandas as pd

results   = pd.read_parquet("hf://datasets/Hussam-q/MedHAM/evaluation_results.parquet")
signals   = pd.read_parquet("hf://datasets/Hussam-q/MedHAM/evaluation_signals.parquet")
responses = pd.read_parquet("hf://datasets/Hussam-q/MedHAM/model_responses.parquet")
judges    = pd.read_parquet("hf://datasets/Hussam-q/MedHAM/judge_evaluations.parquet")

---

Source Data

Questions, ground truths, and knowledge contexts are derived from the MedHallu benchmark:

UTAustin-AIHealth/MedHallu — pqa_labeled split (human-verified) https://huggingface.co/datasets/UTAustin-AIHealth/MedHallu

---

Reproducibility

Parameter	Value
Generation temperature	0.0
Max response tokens	600
Judge temperature	0.0
FActScore threshold	0.85
Judge majority threshold	0.50

Source code: github.com/hussam-q/MedHAM