Med-HALT

Med-HALT: Evaluates clinical, biomedical, medical-exam, coding, or healthcare-document reasoning.

10rows
mean_hallucination_test_accuracyprimary metric
2026-05-27sampled

Metadata

ID
med_halt
Category
Healthcare
Release
Unknown

Metrics

Reasoning hallucination average accuracy, Reasoning hallucination average pointwise score, Memory hallucination average accuracy, Memory hallucination average pointwise score

Latest Results

Rows are transcribed from public Med-HALT CoNLL 2023 Tables 2 and 3. Primary score is a BenchmarkList-derived mean of reasoning and memory hallucination average accuracies.

Rank Subject Score Model Match Provenance Sampled
1 Falcon 40B 44.725% — Imported 2026-05-27
2 Falcon 40B-instruct 41.205% — Imported 2026-05-27
3 Llama-2 70B 40.185% — Imported 2026-05-27
4 Text-Davinci 37.105% — Imported 2026-05-27
5 Llama-2 13B 32.53% — Imported 2026-05-27
6 GPT-3.5 32.22% — Imported 2026-05-27
7 Llama-2-7B 21.945% — Imported 2026-05-27
8 Llama-2-13B-chat 20.51% — Imported 2026-05-27
9 Llama-2-7B-chat 14.38% — Imported 2026-05-27
10 Llama-2 70B Chat 12.155% — Imported 2026-05-27